This PDF file contains the front matter associated with SPIE Proceedings Volume 12031 including the Title Page, Copyright information, and Table of Contents.

Tissue imaging is a pivotal component of both biomedical research and clinical practice. In order to identify tissue structures down to the cellular level, it requires the capability to image mm-size unstained tissue specimens with micron to sub-micron resolution. Tissue imaging is normally performed either using x-rays or visible light. While the latter is limited by light scattering in relatively thick tissues, the former often suffers from poor contrast in absorption-based systems. Phase-contrast x-ray microscopes exist but they often lack the required quantitativeness, entail acquisition times of the order of tens of hours for 3-D imaging and are limited to narrow fields of view. We propose a novel multi-modal phase-based x-ray microscope capable of imaging mm-thick tissue samples on a mm-size field of view using intensity-modulation masks. They act as optical elements allowing the quantitative retrieval of tissue properties such as transmission, refraction and scattering. Additionally, given that the system’s spatial resolution depends only on the mask aperture size, a multi-resolution approach is possible by selecting masks with aperture size matching the resolution requirements (micron and sub-micron) of specific samples. The design and optimization of the x-ray microscope is presented in this paper together with exemplar images of a thin foam sample resulting from the retrieval of the three contrast channels. The final paper will include details of the system parameter optimization (e.g., propagation distance, mask aperture and period), their effect on the retrieval algorithm and imaging performance as well as the first images of biological samples.

While breast density is known as one of the critical risk factors of breast cancer, Digital breast tomosynthesis (DBT)-based diagnostic performance is known to have a strong dependence on breast density. As a potential solution to increase the diagnostic performance of DBT, we are investigating dual-energy DBT imaging techniques. We estimated partial path lengths of an x-ray through water, lipid, and protein from the measured dual-energy projection data and the object thickness information. We reconstructed material-selective DBT images for the material-decomposed projection. The feasibility of the proposed dual-energy DBT scheme has been demonstrated by using physical phantoms.

The mathematical underpinnings of a novel reconstruction algorithm are presented that can facilitate 4D tomosynthesis for the purpose of guiding needle breast biopsies in real-time. Conventional tomosynthesis reconstruction algorithms produce motion artifacts when applied to a continuous tomosynthesis acquisition of a moving biopsy needle. The novel algorithm proposed in this work successfully overcomes this by using differences in slow-scan data to identify variational regions in the reconstructed volume, and adaptively reconstruct those regions to eliminate motion. The algorithm has been tested using simulated images, where reconstructed images of a moving needle had significantly better clarity than the conventional algorithm.

Aim: To evaluate the impact of a high ratio antiscatter grid on imaging system dose efficiency using a spatial frequency domain figure of merit, for coronary angiography applications. Methods: Four imaging conditions were assessed: a source-image receptor distance (SID) of 105cm, used with and without the standard 15/80 grid, and an SID of 120cm used with and without the new 29/80 grid. Automatic dose rate control (ADRC) x-ray factors were generated for these conditions and used to image a composite phantom, with each slab made of 20 mm poly(methyl)methacrylate (PMMA), 2 mm Al and 7 mm expanded polystyrene. Total phantom thickness ranged from 9 cm to 44 cm. An Fe square was included in the phantom and imaged at the iso-centre. Signal difference to noise ratio (SDNR) was measured and multiplied by modulation transfer function (MTF) based weighting factors to give SDNR(u), along with additional factors to account for the influence of SID and DQE. A figure of merit (FOM) defined as SDNR²(u)/K was used to compare the four conditions, where K is the air kerma at the interventional reference point. Results: Below phantom thicknesses of ~20 cm, neither grid improved system dose efficiency compared to imaging without grid. For thicknesses of 30 cm up to 44 cm, dose reduction using the 29/80 grid compared to the 15/80 grid ranged from 10% to 48%. Conclusion: Evaluation using the SDNR²(u) based FOM showed that the 29/80 grid reduced dose by up to 48% compared to the current 15/80 grid, for phantom thicknesses above 24 cm.

Conventional x-ray imaging provides little quantitative information due to scatter, beam hardening, and overlaying tissues. A single-shot quantitative x-ray imaging (SSQI) method was previously developed to quantify material-specific densities in x-ray imaging by combining the use of a primary modulator (PM) and dual-layer (DL) detector. The feasibility of this concept was demonstrated with simulations using an iterative patch-based method. In this work, we propose a new algorithm pipeline for SSQI that enables accurate quantification and high computational efficiency. The DL images contain four measurements that are obtained behind the unattenuated and partially attenuated regions of the PM of each layer. Using the low-frequency property of scatter and a precalibrated material decomposition (MD), four unknowns (i.e., two scatter images and two material-specific images) are jointly recovered by directly solving four equations given by the four measurements. We tested this algorithm in simulations and further demonstrated its efficacy on chest phantom experiments. Through simulation, we show that the new method for MD is robust against scatter. Its performance improves with smaller PM pitch size and smaller focal spot blur. The RMSE in material-specific images compared to ground truth reduces by 52%-84% versus without scatter correction. For our experimental study, we successfully separated soft tissue and bone. The computational time for processing each view was ~8 s without optimization. The reported results further strengthen the potential of SSQI for widespread adoption, leading to quantitative imaging not only for x-ray imaging but also for real-time image guidance or cone-beam CT.

Purpose: Contrast dilution gradient (CDG) analysis is a technique used to extract velocimetric 2D information from digitally subtracted angiographic (DSA) acquisitions. This information may then be used by clinicians to quantitatively assess the effects of endovascular treatment on flow conditions surrounding pathologies of interest. The method assumes negligible diffusion conditions, making 1000 fps high speed angiography (HSA), in which diffusion between 1 ms frames may be neglected, a strong candidate for velocimetric analysis using CDG. Previous studies have demonstrated the success of CDG analysis in obtaining velocimetric one-dimensional data at the arterial centerline of simple vasculature. This study seeks to resolve velocity distributions across the entire vessel using 2D-CDG analysis with HSA acquisitions. Materials and Methods: HSA acquisitions for this study were obtained in vitro with a benchtop flow loop at 1000 fps using the XC-Actaeon (Direct Conversion Inc.) photon counting detector. 2D-CDG analyses were compared with computational fluid dynamics (CFD) via automatic co-registration of the results from each velocimetry method. This comparison was performed using mean absolute error between pixel values in each method (after temporal averaging). Results: CDG velocity magnitudes were slightly under approximated relative to CFD results (mean velocity: 27 cm/s, mean absolute error: 4.3 cm/s) as a result of incomplete contrast filling. Relative 2D spatial velocity distributions in CDG analysis agreed well with CFD distributions qualitatively. Conclusions: CDG may be used to obtain velocity distributions in and surrounding vascular pathologies provided diffusion is negligible relative to convection in the flow, given a continuous gradient of contrast.

Digital breast tomosynthesis (DBT) reconstruction is an ill-posed inverse problem due to the limited-angle acquisition geometry. DBT is also a low dose imaging technique and has very noisy projection views. In this study, we investigated the feasibility of improving image quality of DBT reconstruction by combining (1) a model-based iterative reconstruction (MBIR) method that models the detector blur and correlated noise (DBCN) of the DBT system, and (2) a deep convolutional neural network based DBT denoiser, DNGAN, that we developed in our previous work. DBCN is physics-based whereas DNGAN is data-driven. We followed the regularization by denoising (RED) framework to construct a regularizer from DNGAN and used the DBCN-modeled terms in the MBIR formulation. We solved the optimization problem using the proximal gradient method. The proposed approach, named DBCN+DNGAN, was tested on a set of human subject DBT data sets. The image quality was evaluated quantitatively with figures of merit (FOMs) including the contrast-to-noise ratio, full width at half maximum, and task-based detectability index of a set of microcalcifications individually marked in the human subject data set. We found that these FOMs were improved in the DBCN+DNGAN-reconstructed DBT volumes compared to those reconstructed with DBCN alone or with the simultaneous algebraic reconstruction technique. The soft tissue appearance was visually satisfactory and the background noise level was low in the DBCN+DNGAN reconstructed images.

Simultaneous Digital Breast Tomosynthesis (DBT) and mechanical imaging (MI), called DBTMI, is a novel breast imaging method aimed at improving sensitivity and specificity of breast cancer screening. DBTMI combines improved cancer detection by three-dimensional DBT imaging, with the analysis of local stress over the compressed breast by MI, which can reduce false positive findings. The MI signal is affected by various factors, e.g., breast size, composition, tumor depth, etc. Assessing the individual effect of those factors using clinical data is difficult, due to their interdependence. These open clinical questions can be addressed by virtual clinical trials. Our current work is focused on the effects of tumor depth on the DBTMI signal. We simulated the breast anatomy by a matrix of adipose and glandular tissue compartments. Spherical tumors were inserted at various depths. The MI sensor is modeled by a compound material of PMMA and Ag. We calculated the local stress on the compressed breast surface at the tumor location and simulated the MI sensor output. We also simulated the corresponding DBT images and calculated the signal-difference-to-noise ratio (SDNR) with and without pre-processing to analyze the reduction in artifacts. Our preliminary analysis of 24 simulated tumors has shown 16% reduction in the local stress, when increasing tumor depth by 15 mm (10-25 mm from the breast surface). The SDNR improvement was highest for tumors near the sensor and the effect of pre-processing decreased with increasing tumor depth.

We proposed a direct-indirect dual layer detector combination for CEDM and CEDBT to eliminate patient motion artifacts in dual energy images. Both physical experiments and Monte Carlo simulations were conducted to compare dual shot technique with the dual layer technique. The proposed direct-indirect dual layer detector combination uses a 200 µm direct a-Se as the front detector for LE images and a 400~1000 µm indirect CsI as the back detector for HE images. Results showed similar breast tissue cancellation between dual layer and dual shot techniques. Dual layer technique with direct-indirect detector has higher SDNR/MGD compared with dual shot technique.

There has recently been renewed interest in quantitative iodinated contrast-enhanced breast imaging, sometimes referred to as contrast-enhanced spectral mammography (CESM). Photon-counting detectors (PCDs) have a number of benefits for iodinated contrast-enhanced imaging over dual-energy systems (using energy integrating detectors), one of which is the capability of acquiring multi-bin data with one exposure. Most PCDs and prototype systems being developed are using CdTe or CZT sensor material which have non-optimal characteristic X-ray emission with energies in the range used for breast imaging. This increases charge sharing and, hence, spectral degradation. Recently, a new PCD has been developed using a GaAs sensor. Since GaAs has lower energy characteristic x-rays (lower than CdTe), it is expected that this new PCD detector might be beneficial for spectral x-ray breast imaging. In this work, we have theoretically compared the two detector materials in terms of iodine quantification using the Cramer-Rao lower bound (CRLB) as a figure of merit. Four different experiments were performed, each comparing the performance of CdTe and GaAs: 1) sensitivity of iodine quantification to charge cloud size and electronic noise, 2) effect of dose and photon spectrum on iodine quantification, 3) how the CRLB changes with the number of energy bins, and 4) a dose analysis study to assess any possible dose reduction offered by either detector. Simulations of both the CdTe and GaAs PCDs were performed using the Photon Counting Toolkit (PcTK) software. Three sets of 5000 noise realizations were used to calculate the CRLB of iodine quantification in each study. Results of these studies suggest that GaAs is a promising detector material for contrast-enhanced spectral mammography

Using anti-scatter grids in digital breast tomosynthesis (DBT) is challenging due to the relative motion between source and detector. Therefore, algorithmic scatter correction methods could be beneficial to compensate for the image quality loss due to scatter radiation. In this work, we present a deep learning model capable of predicting the scatter fraction (SF) image for simulations of realistically shaped homogeneous phantoms. The model was trained, validated, and tested on a dataset of 600 homogeneous phantoms representing the compressed breast with thicknesses ranging from 30 mm to 89 mm with randomly generated breast shapes. Monte Carlo simulations were performed for all cases and at different projection angles to obtain estimates of the DBT primary and scatter projections. The same procedure was performed for patient-based phantoms with realistic internal glandular and adipose texture to evaluate the generalizability of our model results. The median and interquartile range (IQR) of the mean relative difference (MRD) and mean absolute error (MAE) for the homogeneous phantoms between Monte Carlo SF ground truth and model predictions resulted in approximately 0.49% (IQR, 0.26−0.76%) and 2.06% (IQR, 1.83−2.26%), respectively, while the patient-based phantoms resulted in results with a MRD of -1.39% (IQR, -4.13−1.60%) and a MAE of 3.97% (IQR, 3.38−4.71%). This seems to indicate that the model trained on the homogeneous phantoms captures the average representation inside the homogeneous breast, with reasonable accuracy in breasts with texture variations. Therefore, the model has the potential to be used as an algorithmic scatter correction method in the future.

Photon counting detectors offer some unique features for x-ray imaging. If designed correctly, photon counting detectors have no readout noise and no dark counts. This is an important feature in for example low dose CT imaging where the total dose is distributed over a large number of projections from different angles. In addition to this it is also possible to incorporate pulse height discrimination of each photon event, thus enabling the recording of images from multiple energy intervals in a single exposure. We demonstrate the performance of a newly developed fast photon counting detector with six energy thresholds and 150µm pixel size that can be read out with up to 10000fps. It is a four side buttable device that is bump bonded to a CdTe converter. The very high conversion efficiency of CdTe makes the detector suitable for a wide range of applications requiring high spatial resolution at low doses. The efficiency of the detector is maintained all the way out to the edge of the chip which opens the possibility to build larger detectors still fulfilling medical requirements. The novel detector incorporates a charge sharing correction feature and the effect of this function is demonstrated using the DQE measurements as well as with spectrum reconstruction from Cd109 and Am241 radioactive sources. We show that this charge sharing correction feature affects the properties of NPS and MTF, and the energy resolution is greatly enhanced. Measurements are also compared to a simulation model for the detector system.

Photon counting detectors (PCDs) with energy discrimination capabilities provide spectral information through energy binning and higher spatial resolution than conventional energy integrating detectors (EIDs). However, the projection data transmission from the detector across the slip ring to the processing computer becomes more challenging due to the increased amount of data, including multiple (e.g., 8) energy bins. In this work, we propose a projection-domain energy bin weighting method that produces two energy-weighted measurements to provide comparable spectral information as the original binned counts for material decomposition and virtual monoenergetic imaging tasks. We obtain the optimal energy bin weights by minimizing the Cramér–Rao lower bound (CRLB) ratio between the weighted measurements and that of the original binned counts and evaluate their respective material decomposition performance using Monte Carlo simulation. The experiments were conducted with realistic photon counting detector energy responses, which were not assumed to be known. Instead, only an empirical calibration using a step-wedge phantom is required, making this process extensible to any photon counting detector without prior knowledge of its energy response or the incident spectrum. The results show that the two energy-weighted measurements generated with our method can provide comparable material decomposition results with low bias and less than 20% variance penalty to that of the original binned counts for a large range of patient size, with a data reduction of 75% for a silicon detector with 8 energy bins and 60% for a CdTe detector with 5 energy bins.

Photon counting x-ray detectors enable spectral imaging, which can be utilized for material decomposition and quantitative imaging tasks. This study investigates potential benefits of tube voltage optimization, including fast kV switching, on material decomposition noise. In simulation studies, single kV scans including 80/100/120/140 kV were tested as well as an 80/140 fast kV switching pair. Fluences of spectra were normalized with respect to CTDI to keep dose neutral. An open-source spectral response model of a realistic Si detector was used in the Cramér–Rao lower bound calculation, which estimates the lower bound of the noise in material decomposition. A simulated CT scan of an anthropomorphic head phantom was performed for noise analysis in the image domain. Simulation results showed that a single kV can be optimized for specific imaging tasks depending on the object size, while fast kV switching can substantially reduce the noise in material decomposition compared to single kV scan. Additionally, noise can be further reduced with a fixed Kedge (Gd) filter during fast kV switching acquisitions.

C-arm x-ray systems equipped with flat panel detectors (FPD) lack spectral and ultra-high-resolution (UHR) capabilities desired by physicians for image guided interventions (IGIs), for example to discriminate between and/or quantify different materials such as iodine and calcium, or in the visualization of very fine structures or devices used in interventional procedures. Photon counting detectors (PCDs) can introduce these capabilities to the interventional suite: In this work, we propose a new dagger-shaped PCD design tailored for IGIs to upgrade the imaging capabilities in the C-arm interventional system while preserving the functionality of the existing FPD and reducing the system cost compared to completely replacing the FPD with a large-area PCD. The design consists of two modules integrated together: One is a long-strip shape for narrow-beam spectral and UHR CT with full axial coverage, and one is rectangle-shaped for volume- and region-of-interest 2D and 3D spectral and UHR imaging. As a proof of concept, prototypes of each module were used to perform phantom and in vivo animal experiments. Results show the potential of the proposed design in discriminating between and quantifying iodine and calcium by leveraging the spectral information provided by PCDs. UHR 2D and 3D PCD images show the improved capabilities of the dagger PCD in delineating small blood vessels with improved contrast-to-noise ratios, as well as resolving fine structures such as stents commonly used in IGIs.

Amorphous selenium (a-Se) is a high gain, low dark current, large area compatible photoconductor that has received significant attention towards the development of UV and X-ray detectors for medical imaging. Indirect detectors utilizing a-Se often feature blue emitting scintillators due to the high attenuation coefficient of a-Se in that region. However, emission tails from the scintillators often fall out of the conversion range of a-Se, and scintillators with emission peaks outside the absorption of Se cannot be utilized. In order to improve the sensitivity and gain in a-Se indirect detectors, we propose doping a-Se with tellurium as a function of depth, where tail emission will be absorbed by the lower bandgap aSe/Te after primary absorption in the initial Se layer. In addition, we employ a lateral device structure to avoid any absorption at short wavelengths from a transparent electrode or blocking layer. In this work, we present the first steps towards fabricating these devices. Studies of charge transport in doped a-Se/Te devices are performed using the transient photocurrent time-of-flight technique. We report hole and electron mobilities for a-Se_1-xTe_x (x = 0, 0.01, 0.05, 0.10) as a function of applied voltage, along with band gaps and comparisons to previous studies. Fabrication of lateral devices, with and without optical slits, is demonstrated and discussed.

We investigate X-ray photon-counting scintillation detectors with silicon photomultiplier (SiPM) readout. These circumvent some drawbacks of direct-conversion detectors. We measured observed count rate (OCR) versus X-ray tube current for single-pixel detectors consisting of LYSO:Ce and YAP:Ce scintillators coupled to ultrafast SiPMs. For a 30 keV threshold, the maximum OCRs equal 4.5 Mcps/pixel (LYSO:Ce) and 5.5 Mcps/pixel (YAP:Ce) for paralyzable-like counting and 10 Mcps/pixel (LYSO:Ce) and 12.5 Mcps/pixel (YAP:Ce) for nonparalyzable-like counting. We estimate that the twice as fast LaBr3:Ce scintillator yields OCRs approaching those of CdTe/CZT-based photon-counting CT detectors. We also show energy response data and discuss dose-efficient pixel miniaturization.

The next generation medical imaging will benefit significantly from artificial intelligence – therefore, not only advances in computing power are required, but also further materials and technology improvements will lead to better image quality. The strong X-ray absorption, high charge carrier mobility and lifetime recommend perovskites like methylammonium lead triiodide (MAPbI3) as novel direct X-ray converting materials. A major obstacle to the commercialization is the limited stability and lifetime, as reported until now. Here, we show for the first time that degradation is limited in our X-ray detector after 1.5 years of storage in ambient conditions.

High Speed Angiography (HSA) requires imaging detectors with both high-temporal and high-spatial resolution. Both the Aries and Acteon detectors by Direct Conversion (Stockholm, Sweden) are CdTe direct photon-counting detectors (PCD) that have acquisition frame rates of up to 1000-fps and a 100-micrometer pixel pitch; however, the new Aries detector offers a larger field of view (512 x 768 pixels) compared to the smaller Actaeon detector (256 x 256 pixels). An expanded field of view is required for imaging of larger vasculature, thus the Aries offers this advantage. Evaluations were performed of both detectors under Anti-Coincidence Circuitry (ACC-ON) mode, which corrects for charge sharing between pixels. Initial evaluations of instrumentation noise and detector energy-threshold calibration using Am-241 gamma spectroscopy were performed for the new Aries detector. Linearity was also evaluated for the Aries for each of the 12 individual modules that compose the detector field to check for homogeneity in response to exposure throughout the detector. Finally, Normalized Noise Power Spectrum (NNPS), Modulation Transfer Function (MTF) and Detective Quantum Efficiency (DQE) were then compared between the Aries and Actaeon detectors at two different exposures and detector energy thresholds. The detectors are linear up to approximately 1000 μR and have no instrumentation noise above a threshold of 15 keV. As expected, the MTF’s and DQE’s are similar between the Aries and Actaeon detectors, and there are thus no tradeoff’s in image quality to achieve the larger FOV.

While active matrix flat-panel imagers (AMFPIs) are commonly used in low dose-per-frame applications such as digital breast tomosynthesis, DQE performance under such irradiation conditions is constrained by the large electronic additive noise relative to imaging signal. Replacement of the CsI:Tl and a-Se converters typically used in AMFPIs with polycrystalline mercuric iodide fabricated using particle-in-binder techniques (PIB-HgI2) would largely overcome this limitation due to the 3 to 10 times higher signal per interacting x-ray offered by this photoconductive material. However, practical implementation of PIB-HgI2 converters in clinical systems requires significant reduction of the high levels of image lag associated with the material. A promising strategy for significantly reducing lag is to diminish the contribution of signal induced by hole transport (which is believed to be a principal source of lag) through incorporation of a grid structure (referred to as a Frisch grid) into the converter material – with previous theoretical studies indicating hole signal reductions of up to ~95%. In this paper, an initial examination of the direct influence of the grid on image lag is reported. This involved modeling the trapping and release of holes in the converter for grids supported by insulating pillars. Signal properties, including line spread function, MTF, and image lag, were examined as a function of grid design parameters and operational conditions. The modeling shows that, when charge trapping and release are accounted for, favorably high suppression of image lag as well as good MTF is possible for some grid designs.

A high spatial resolution and cost-effective amorphous selenium multi-layer detector architecture was previously proposed for high-energy applications including single-photon emission computed tomography (SPECT). To overcome the complexity of reading out signal from the multi-layer detector, we propose a resistive microstrip detector readout using charge division. The charge induced on the readout electrode undergoes charge division along the electrode and is proportional to the physical location of absorbed photons, enabling position sensitivity. A 1 mm collimated blue LED is employed to illuminate the electrode, which has a resistance of 242 kΩ. To improve the linearity of the position sensing, the microstrips were deposited underneath a-Se in order to move bonding pads far away from the active area. A low deposition temperature evaporated Cs-doped and As-stabilized a-Se layer is used as a hole blocking layer to ensure that the resistivity of microstrips is not affected and to prevent crystallization of the temperature sensitive a-Se layer. The detector structure was fabricated on a flexible substrate that can potentially be incorporated, in the future, into a curved detector structure for large area SPECT systems. The results presented demonstrate that the detected position is linearly proportional to the location of absorbed photons with a spatial resolution of 1 mm.

Phantoms are essential tools for assessing and verifying performance in computed tomography (CT). Realistic patientbased lung phantoms that accurately represent textures and densities are essential in developing and evaluating novel CT hardware and software. This study introduces PixelPrint, a 3D-printing solution to create patient-specific lung phantoms with accurate contrast and textures. PixelPrint converts patient images directly into printer instructions, where density is modeled as the ratio of filament to voxel volume to emulate local attenuation values. For evaluation of PixelPrint, phantoms based on four COVID-19 pneumonia patients were manufactured and scanned with the original (clinical) CT scanners and protocols. Density and geometrical accuracies between phantom and patient images were evaluated for various anatomical features in the lung, and a radiomic feature comparison was performed for mild, moderate, and severe COVID-19 pneumonia patient-based phantoms. Qualitatively, CT images of the patient-based phantoms closely resemble the original CT images, both in texture and contrast levels, with clearly visible vascular and parenchymal structures. Regions-ofinterest (ROIs) comparing attenuation demonstrated differences below 15 HU. Manual size measurements performed by an experienced thoracic radiologist revealed a high degree of geometrical correlation between identical patient and phantom features, with differences smaller than the intrinsic spatial resolution of the images. Radiomic feature analysis revealed high correspondence, with correlations of 0.95-0.99 between patient and phantom images. Our study demonstrates the feasibility of 3D-printed patient-based lung phantoms with accurate geometry, texture, and contrast that will enable protocol optimization, CT research and development advancements, and generation of ground-truth datasets for radiomic evaluations.

Fluence field modulation (FFM) using dynamic pre-patient attenuators could reduce radiation dose while preserving image quality by optimizing the distribution of x-ray photons incident on the patient. However, past dynamic attenuators require mechanical action that is challenging to implement in diagnostic CT scanners and generate a large variety of system states that makes calibration complex. To circumvent these difficulties, we propose a motion-free mechanism for FFM that uses electromagnetic deflection of the focal spot, also called flying focal spot (FFS), together with interference patterns generated from fixed metal gratings. Our proposed design is digital in that only a limited number of fluence fields is possible, but the fluence field from each focal spot position is stable with respect to source perturbation, simplifying calibration. Intermediate fluence fields can be virtually achieved during reconstruction by using either statistical reconstruction or rebinning. We describe the geometric constraints for our FFM mechanism and illustrate some of the possible fluence fields that can be achieved.

Optimization of CT image quality typically involves balancing noise and bias. In filtered back-projection, this trade-off is controlled by the particular filter and cutoff frequency. In penalized-likelihood iterative reconstruction, the penalty weight serves the same function. Deep neural networks typically do not provide this tuneable control over output image properties. Models are often trained to minimize mean squared error which penalizes both variance and bias in image outputs but does not offer any control over the trade-off between the two. In this work, we propose a method for controlling the output image properties of neural networks with a new loss function called weighted covariance and bias (WCB). Our proposed method includes separate weighting parameters to control the relative importance of noise or bias reduction. Moreover, we show that tuning these weights enables targeted penalization of specific image features (e.g. spatial frequencies). To evaluate our method, we present a simulation study using digital anthropormorphic phantoms, physical simulation of non-ideal CT data, and image formation with various algorithms. We show that WCB offers a greater degree of control over trade-offs between variance and bias whereas MSE has only one configuration. We also show that WCB can be used to control specific image properties including variance, bias, spatial resolution, and the noise correlation of neural network outputs. Finally, we present a method to optimize the proposed weights for stimulus detectability. Our results demonstrate the potential for this new capability to control the image properties of DNN outputs and optimize image quality for the task-specific applications.

Eye-tracking techniques can be used to understand the visual search process in diagnostic radiology. Nonetheless, most prior eye-tracking studies in CT only involved single cross-sectional images or video playback of the reconstructed volume and meanwhile applied strong constraints to reader-image interactivity, yielding a disconnection between the corresponding experimental setup and clinical reality. To overcome this limitation, we developed an eye-tracking system that integrates eye-tracking hardware with in-house-built image viewing software. This system enabled recording of radiologists’ real-time eye-movement and interactivity with the displayed images in clinically relevant tasks. In this work, the system implementation was demonstrated, and the spatial accuracy of eye-tracking data was evaluated using digital phantom images and patient CT angiography exam. The measured offset between targets and gaze points was comparable to that of many prior eye-tracking systems (The median offset: phantom – visual angle ~0.8°; patient CTA – visual angle ~0.7 – 1.3°). Further, the eye-tracking system was used to record radiologists’ visual search in a liver lesion detection task with contrast-enhanced abdominal CT. From the measured data, several variables were found to correlate with radiologists’ sensitivity, e.g., mean sensitivity of readers with longer interpretation time was higher than that of the others (88 ± 3% vs 78 ± 10%; p < 0.001). In summary, the proposed eye-tracking system has the potential of providing high-quality data to characterize radiologists’ visual-search process in clinical CT tasks.

Considering the potential radiation effect on patients in computed tomography (CT) imaging, it is desirable to reduce the radiation dose. Reduction in dose incurs degradation in image quality and possible reduced diagnostic performance. CT image quality needs to be maintained at standards sufficient for effective clinical reading. Therefore, the dose reduction should be guided by the potential image quality. Unfortunately, currently available CT image quality assessment (IQA) tools are based on maintaining a uniform image quality or use the CT exams themselves to retrospectively determine image quality. A robust and comprehensive IQA metric should represent the image quality of each patient at the organ level, and before the CT exams. Towards this objective, we devise a fully-automated, end-to-end deep learning-based solution to perform real-time, patient-specific, organ-level image quality prediction of CT scans. Leveraging the 2D scout (frontal and lateral) images of the actual patients, which are routinely acquired prior to the CT scan, our proposed Scout-IQA model estimates the patient-specific mean noise in real-time for six different organs. Our experimental evaluation on real patient data demonstrates the effectiveness of our Scout model not only in real-time noise estimation (only 6 ms on average per scan), but also as a potential tool for optimizing CT radiation dose in individual patients.

Purpose: The purpose of this study was to evaluate a novel stationary head CT system (sHCT) enabled by the carbon nanotube (CNT) field emission x-ray source array for volumetric head imaging. A data processing and image reconstruction package was developed and demonstrated for the system. Methods: The experimental sHCT system consisted of three CNT x-ray source arrays placed in parallel imaging planes separately. During imaging, the sources stayed stationary while only the bed advancing through a tunnel formed by the parallel planes. In each imaging cycle, 135 projections were acquired over 232 degrees of view. A CT ACR 464 phantom and a Kyoto head phantom were imaged to evaluate the system performance. Image uniformity, signal to noise ratio, spatial resolution, CT number and detectability were investigated for the proposed system. All images were processed and reconstructed with an iterative reconstruction-based package. Total-variation (TV) regularization methods such as the adaptive steepest descent projection onto convex set (ASDPOCS) algorithm were implemented to reduce noises and artifacts caused by the reduced projection views. Results: Volumetric data with good uniformity, high spatial resolution, and detectability for both high- and low-resolution features were demonstrated for the proposed sHCT scanner. Two scan protocols, step-and-shoot and continuous mode were compared and proved to provide similar image quality while the latter increased the total scan speed. The SIRTASDPOCS algorithm effectively suppressed the sparse-view and limited-view artifact and enhance the contrast noise ratio. Conclusion: We demonstrated the feasibility of sHCT for volumetric head imaging using multiple CNT x-ray source arrays. The device provides 3D images with high fidelity. The prototype sHCT system is being installed at the UNC hospital for a patient imaging study.

We estimate the minimum SNR necessary for object detection in the projection domain. We assume there is a set of objects O and we study an ideal observer that sequentially compares each member of O to the null hypothesis. This reduces to one-dimensional signal detection between two Gaussians. We find that for a search task of a circular 6 mm lesion in a region of interest 60 mm by 60 mm by 10 slices, and for a required sensitivity of 80% and specificity of 80%, the minimum required projection SNR is 5.1, a finding reminiscent of the Rose criterion.

Many published studies use deep learning models to predict COVID-19 from chest x-ray (CXR) images, often reporting high performances. However, the models do not generalize well on independent external testing. Common limitations include the lack of medical imaging data and disease labels, leading to training on small datasets or drawing classes from different institutions. To address these concerns, we designed an external validation study of deep learning classifiers for COVID-19 in CXR images including XCAT phantoms as well. We hypothesize that a simulated CXR image dataset obtained from the XCAT phantom allows for better control of the dataset including pixel-level ground truth. This setup allows for multiple advantages: First, we can validate the publicly available models using simulated chest x-rays. Secondly, we can also address clinically relevant questions with this setup such as effect of dose levels and sizeof COVID-19 pneumonia in performance of deep learning classifier. We

Characterization of microcalcification clusters in the breast and differentiation between benign and malignant structures on (contrast-enhanced) mammography (CEM) images is of great importance to determine cancerous lesions. Computer algorithms may help performing these tasks, but typically need large sets of data for model training. Therefore this paper develops a method to create synthetic microcalcification clusters that can later be used to overcome data sparsity problems. Starting from descriptors of the shape and size, both benign and malignant microcalcifications were created and then combined into 3-dimensional cluster models given realistic geometric properties. The distributions of the largest diameter and the number of microcalcifications per cluster in a set of 500 simulated clusters were set such that they agreed with those of real clusters. An existing simulation tool was then extended to insert the clusters into processed, low-energy CEM background images with appropriate contrast values. In a validation study comprised of 40 real and 40 synthetic cases, radiologists were asked to evaluate realism and malignancy. It was found that the shape and the structure of the individual microcalcifications as well as the complete clusters were realistic. Thus the descriptors were chosen correctly and enabled a good classification between benign and malignant cases. The realistic brightness and boundary smoothness proved the simulation tool can correctly insert the 3D clusters into real background images and is suitable of creating a large set of realistic microcalcification clusters simulated in existing (contrast-enhanced) mammography images. With improvements on the correspondence of insertion location in craniocaudal and mediolateral oblique view, which proved more challenging to simulate realistically, this promising method is expected to be applicable for modeling complete synthetic cases. Such a dataset can be used for data enrichment where data sources are limited and for development and training purposes.

The aim of this study is to evaluate the accuracy of one of the commercial breast density quantization software. The estimation is performed by means of mammograms of five different 3D printed breast phantoms obtained from clinical patient images. Mammograms of these phantoms were acquired and analyzed using the density software. In addition, the spectra used by the automatic exposure control for each mammogram was accurately characterized and modeled using a previously published spectral model. The result is that the amount of estimated dense tissue is accurate to within an intra-phantom mean of 10% (std. dev. <4.4%), with negligible bias.

Dynamic imaging systems monitor physiological processes that evolve or change over time. However, image reconstruction from dynamic data is made difficult by data incompleteness and significant computational burden. Data incompleteness, in particular, arises from severe undersampling often necessary to increase frame rate by reducing data acquisition time and leads to ill-posedness of the reconstruction problem. Computational cost and memory requirements are particularly burdensome for three-dimensional problems, especially for applications in which high-resolution in space and time is needed. Two main approaches exist for dynamic image reconstruction. Frame-by-frame approaches solve a sequence of image reconstruction problems (one for each frame). Spatiotemporal approaches instead directly reconstruct the dynamic object using data from all imaging frames at once. Although statistically suboptimal, frame-by-frame approaches have often been advocated because of the ease of implementation and lower memory requirements. This work explores a new spatiotemporal dynamic reconstruction approach that uses neural fields, a special class of neural networks, to drastically reduce the computational complexity and memory requirements while exploiting the object’s spatiotemporal redundancies. As a feasibility study, a simple dynamic image reconstruction problem whose forward operator is given by the circular Radon transform is considered. Numerical results demonstrate that the proposed approach is more accurate and uses less memory than the classical frame-by-frame approach.

Photon counting detectors in x-ray computed tomography (CT) are a major technological advancement that provides additional energy information and improves the decomposition of the CT image into material images. One important challenge in this new modality is how best to perform tomographic reconstruction. As a result of measuring multiple projections from different energy bins, more complex reconstruction algorithms are required. These are computationally demanding, due to the their large number of degrees of freedom. Also, the reconstruction algorithm needs to output multi-material image solutions. Reconstruction algorithms for spectral CT divide into two paradigms: two-step and one-step. Most typical solution is the two-step approach, where a first step consists of a material decomposition in projection domain, and a second step on tomographic reconstruction of each projection. This solution is computationally tractable but can cause a loss of information and it is difficult to regularize. The one-step solution, on the other hand, solves a joint optimization material decomposition and reconstruction, it is solved iteratively and can be, however, very time consuming. We present a deep learning-based solution to the one-step problem, with an architecture that mimics the updates of a primal-dual solver, and has demonstrated much greater computational efficiency than model-based iterative reconstruction. We have studied a proof-of-concept on a set of 700 Shepp-Logan phantoms. Our approach has shown enhanced performance compared to a model-based two-step approach, as well as compared to considering deep learning only in the first step of the two-step solution.

Computed medical imaging systems require a computational reconstruction procedure for image formation. In order to recover a useful estimate of the object to-be-imaged when the recorded measurements are incomplete, prior knowledge about the nature of object must be utilized. In order to improve the conditioning of an ill-posed imaging inverse problem, deep learning approaches are being actively investigated for better representing object priors and constraints. This work proposes to use a style-based generative adversarial network (StyleGAN) to constrain an image reconstruction problem in the case where additional information in the form of a prior image of the sought-after object is available. An optimization problem is formulated in the intermediate latent-space of a StyleGAN, that is disentangled with respect to meaningful image attributes or “styles”, such as the contrast used in magnetic resonance imaging (MRI). Discrepancy between the sought-after and prior images is measured in the disentangled latent-space, and is used to regularize the inverse problem in the form of constraints on specific styles of the disentangled latent-space. A stylized numerical study inspired by MR imaging is designed, where the sought-after and the prior image are structurally similar, but belong to different contrast mechanisms. The presented numerical studies demonstrate the superiority of the proposed approach as compared to classical approaches in the form of traditional metrics.

Cardiac CT imaging is invaluable in a variety of applications including non-invasive assessment of coronary artery calcifications, the ruling out of acute coronary artery syndrome, and the planning of valve replacement procedures. Newer applications include CT-derived fractional flow reserve and cardiac perfusion with new research opportunities through photon counting CT and deep learning. To facilitate advancements in research and clinical practice, we developed a vendor-neutral pipeline for virtual imaging trials involving cardiac CT simulation and reconstruction. Specifically, we developed a reconstruction method to be combined with dynamic virtual patients (XCAT) and a CT simulator (DukeSim) to generate realistic, retrospectively gated, helical cardiac CT projection data sets. Reconstruction was performed using a multi-channel GPU-based reconstruction toolkit. In this study, CT simulation experiments were performed using a standard adult male XCAT phantom (50th percentile in height and weight) with heart rates of 60, 90, and 120 bpm. The simulated acquisitions were done with a tube voltage of 120 kV, a pitch of 0.14, and a rotation time of 280 msec. Reconstruction was performed analytically, with rebinning and weighted filtered backprojection, and iteratively, with the split Bregman optimization method and low rank and gradient sparsity regularizers. Image quality (SSIM, RMSE) and temporal resolution were evaluated and compared with the ground truth phantoms. Iterative reconstruction was shown to reduce noise power by one order of magnitude without significantly impacting image resolution (NPS, MTF). Analytical reconstructions lead to less than 8% error in the measurement of standard cardiac functional metrics relative to ground truth values.

This work proposes a robust method of identifying the optimal cancellation parameters for dual-energy imaging. We designed the proposed method in a simple and practical way for dual-energy subtraction by utilizing histogram information rather than spatial information of images. A classification predictive modeling based on XGBoost was employed to identify cancellation parameters for soft-tissue and bone selected images. We verified the robust performance of the proposed method for 500 chest x-ray examinations by comparing predicted cancellation parameters with the optimal values determined by well-trained radiologists. The value of the proposed work may contribute to advancements in chest x-ray imaging technologies.

Dual-energy subtraction angiography (DESA) in the interventional setting promises to reduce motion-induced DSA artifacts and support iodine quantitation. A prototype dual-energy C-arm system with 30 Hz kV-switching capability is currently under development. The purpose of this work was to investigate methods of noise reduction in 2D iodine/boneonly images generated on this platform. The methods investigated were anti-correlated noise reduction (ACNR), which exploits noise correlations in iodine/bone and tissue-only images, and a hybrid algorithm (ACNR-ML) that employs machine learning to achieve further reduction in low-frequency noise compared to ACNR. An anthropomorphic chest phantom and a porcine angiographic study were used to evaluate iodine signal-difference-to-noise-ratio, vessel profile and noise texture in DESA images with and without denoising. ACNR iodine images were formed by subtracting a weighted high-pass filtered tissue image from the iodine image, where in practice the high-pass tissue image was the original minus a low-pass-filtered version. The ACNR-ML algorithm replaced the low-pass-filtered tissue image with a denoised tissue image produced by a convolutional neural network (U-Net). When compared to standard weighted log subtraction, the ACNR algorithm yielded an iodine SDNR improvement factor ranging from 2.58 to 2.91 and 3.66 to 4.08 in the phantom and porcine studies, respectively. The ACNR-ML algorithm yielded SDNR improvement from 2.82 to 3.44 in the phantom study and 5.41 to 6.08 in the porcine study. Both noise reduction techniques resulted in vessel full-width half-maximum measurements that were within 3% of the measurements from non-denoised images.

In recent years, much effort has been committed to lowering radiation dose in CT. However, when the radiation dose is lowered, not only is image noise elevated, but the CT number also becomes more inaccurate. Note that the CT number bias issue is intrinsically rooted in the statistical nature of photons and the standard image formation process that has been used for the past 50 years in medical CT practices: after CT data are acquired, a log-transform is applied to generate the sinogram projection data, then an image reconstruction algorithm is applied to reconstruct the CT images. However, there is a fundamental flaw in this image formation process: the log-transform itself is a statistically biased estimator since the statistical mean of the log-transformed data is different from the log-transform of the statistical mean of the data. In medical CT applications, we are forced to take the log transform of a single sample of the measured CT data and then images are reconstructed from the log-transformed data. Consequently, CT images will then have inaccurate CT numbers. In this work, we investigated the imaging physics foundation of the CT number inaccuracy issue in low dose CT and developed a simple, yet extremely effective correction method to address this long-standing issue in CT imaging. This correction scheme was experimentally validated in the context of photon counting detector CT (PCD-CT). Our experimental results demonstrated that the correction scheme addresses the CT number bias problem and improves material quantification accuracy of spectral PCD-CT images.

The proliferation of deep learning image processing calls for a quantitative image quality assessment framework that is suitable for nonlinear, data-dependent algorithms. In this work, we propose a method to systematically evaluate the system and noise responses such that the nonlinear transfer properties can be mapped out. The method involves sampling of lesion perturbations as a function of size, contrast, as well as clinically relevant features such as shape and texture that may be important for diagnosis. We embed the perturbations in backgrounds of varying attenuation levels, noise magnitude and correlation that are associated with different patient anatomies and imaging protocols. The range of system and noise response are further used to evaluate performance for clinical tasks such as signal detection and classification. We performed the assessment for an example CNN-denoising algorithm for low does lung CT screening. The system response of the CNN-denoising algorithm exhibits highly nonlinear behavior where both contrast and higher order lesion features such as spiculated boundaries are not reliably represented for lesions perturbations with small size and low contrast. The noise properties are potentially highly nonstationary, and should be assumed to be the same between the signal-present and signal-absent images. Furthermore, we observer a high degree dependency of both system and noise response on the background attenuation levels. Inputs around zeros are effectively imposed a non-negativity constraint; transfer properties for higher background levels are highly variable. For a detection task, CNN-denoised images improved detectability index by 16-18% compared to low dose CT inputs. For classification task between spiculated and smooth lesions, CNN-denoised images result in a much larger improvement up to 50%. The performance assessment framework propose in this work can systematically map out the nonlinear transfer functions for deep learning algorithms and can potentially enable robust deployment of such algorithms in medical imaging applications.

Chronic obstructive pulmonary disease (COPD) is a leading cause of patient mortality and morbidity. Measuring the amount of emphysema via a non-contrast CT scan plays a major role in the assessment of COPD and possible treatment options. We are interested in evaluating variability in lung CT attenuation values, which directly impacts emphysema scoring. Our investigations are focused on bias in these values due to patient size and positioning effects, as well as a method to reduce this bias. An anthropomorphic lung phantom was scanned in both centered and off-centered positions. Furthermore, patient size variation was simulated using a home-made fat belt. The obtained projection data was first pre-processed and reconstructed as conventionally done on the scanner, then pre-processed with a proposed correction pipeline that is designed to reduce scatter and beam hardening errors in a patient-specific manner. The images from the scanner show that off-centering the phantom changes the difference in mean value between the left and right lung, from 1.5 to 6.5 HU. Moreover, with the fat belt, the mean value in both lungs increases by more than 13 HU and the inhomogeneity of the lung tissue, measured with the standard deviation, nearly doubles with presence of a strong cupping artifact. Using the proposed method to reduce bias, the mean values of the left and right lung are generally closer to each other, the inhomogeneity is either reduced or unchanged, and the cupping artifact is much less pronounced. This is expected to allow more robust emphysema scoring.

There is substantial variability in the performance of radiologist readers. We hypothesized that certain readers may have idiosyncratic weaknesses towards certain types of lesions, and unsupervised learning techniques might identify these patterns. After IRB approval, 25 radiologist readers (9 abdominal subspecialists and 16 non-specialists or trainees) read 40 portal phase liver CT exams, marking all metastases and providing a confidence rating on a scale of 1 to 100. We formed a matrix of reader confidence ratings, with rows corresponding to readers, and columns corresponding to metastases, and each matrix entry providing the confidence rating that a reader gave to the metastasis, with zero confidence used for lesions that were not marked. A clustergram was used to permute the rows and columns of this matrix to group similar readers and metastases together. This clustergram was manually interpreted. We found a cluster of lesions with atypical presentation that were missed by several readers, including subspecialists, and a separate cluster of small, subtle lesions where subspecialists were more confident of their diagnosis than trainees. These and other observations from unsupervised learning could inform targeted training and education of future radiologists.

Recent years have seen the increasing application of deep learning methods in medical imaging formation, processing, and analysis. These methods take advantage of the flexibility of nonlinear neural network models to encode information and features in ways that can outperform conventional approaches. However, because of the nonlinear nature of this processing, images formed by neural networks have properties that are highly datadependent and difficult to analyze. In particular, the generalizability and robustness of these approaches can be difficult to ascertain. In this work, we analyze a class of neural networks that use only piece-wise linear activation functions. This class of networks can be represented by locally linear systems where the linear properties are highly data-dependent - allowing, for example, estimation of noise in image output via standard propagation methods. We propose a nonlinearity index metric that quantifies the fidelity of a local linear approximation of trained models based on specific input data. We applied this analysis to three example CT denoising CNNs to analytically predict the noise properties in the output images. We found that the proposed nonlinearity metric highly correlates with the accuracy of noise predictions. The analysis proposed in this work provides theoretical understanding of the nonlinear behavior of neural networks and enables performance prediction and quantitation under certain conditions.

Channelized Hotelling observer (CHO), which has been shown to be well correlated with human observer performance in many clinical CT tasks, has a great potential to become the method of choice for objective image quality assessment. However, its use has been quite limited in routine CT practice due to lack of efficient implementation. In this work, a CHO model optimized for the most widely used ACR CT accreditation phantom was applied to evaluate the low-contrast detectability of a deep-learning based reconstruction (DLIR) equipped on a GE Revolution scanner. The commercially available DLIR reconstruction method showed consistent increase in low-contrast detectability over the FBP and the IR method at routine dose levels, which suggests potential dose reduction to the FBP reconstruction by up to 27.5%.

Accuracy in proton range prediction is critical in proton therapy to ensure conformal tumor dose. Our lab proposed a joint statistical image reconstruction method (JSIR) based on a basis vector model (BVM) for estimation of stopping power ratio maps and demonstrated that it outperforms competing Dual Energy CT (DECT) methods. However, no study has been performed on the clinical utility of our method. Here, we study the resulting dose prediction error, the difference between the dose delivered to tissue based on the more accurate JSIR-BVM method and the planned dose based on Single Energy CT (SECT).

Coronary CT Angiography (cCTA) is commonly used to detect and quantify luminal stenoses in patients with coronary artery disease (CAD). However, its use is limited in patients with heavy coronary calcifications due to calcium blooming, which is caused by insufficient spatial resolution. This study evaluated the ability of a photon-counting-detector (PCD) CT in quantifying of luminal stenosis in the presence of heavy calcifications relative to an energy-integrating-detector (EID) CT. The phantom results indicate that with PCD-CT, luminal stenoses that were previously considered non-assessable due to the presence of heavily-calcified plaques can be assessed using cCTA.

Hepatocellular carcinoma, the fastest rising cause of cancer-related deaths, is commonly treated with percutaneous ablative therapies where tumor cells are destroyed once tissue temperatures reach a lethal threshold. However, high progression and recurrence rates post ablation suggest the need for intraprocedural temperature monitoring to ensure the lethal threshold (>60°C) is reached and a sufficient safety margin is obtained. A previously developed model generates physical density maps from spectral CT data. These spectral physical density quantifications enable thermometry by taking advantage of the thermal volumetric expansion equation that relates the change in temperature to physical density changes. To validate the physical density model, an ex vivo bovine muscle was weighed and scanned on a clinical spectral CT scanner with different scanning parameter combinations (collimation, dose, helical/axial scans). Calculated mass from physical density maps and volume demonstrated high accuracy with a maximum percent error of 0.34% (<1.1 grams for a345 gram sample) and minimal effects of scanning parameters. After validating the accuracy of the physical density maps, the muscle was subjected to heating and cooling while scanning to evaluate the relationship between physical density and temperature. Spectral results were continuously generated to calculate physical density maps at different temperatures. A linear relationship between change in temperature and change in physical density was established with strong correlation (R = 0.9781). The reflection of thermal volumetric expansion in physical density quantifications indicate its potential utility for providing real-time temperature feedback to interventional radiologists during ablative procedures for not only hepatocellular carcinoma, but also other types of malignancies.

This study introduces a framework to approximate the bias inflicted by CNN noise reduction of CT exams. First, CNN noise reduction was used to approximate the noise-free image and noise-only image of a CT scan. The noise and signal were then recombined with spatial decoupling to simulate an ensemble of 100 images. CNN noise reduction was applied to the simulated ensemble and pixel-wise bias calculated. This bias approximation technique was validated within natural images and phantoms. The technique was then tested on ten whole-body-low-dose CT (WBLD-CT) patient exams. Bias correction led to improved contrast of lung and bone structures.

Ring artifacts are a well-known problem in computed tomography (CT) and in particular in cone-beam CT (CBCT). This work addresses the reduction of ring artifacts in CT acquisitions using a data-driven approach. Deep convolutional neural networks (CNNs) of different dimensionalities are trained to estimate the ring artifacts directly from an uncorrected volume. This approach has the advantage that neither raw-data has to be available, nor any kind of resampling of the data is necessary. In addition to ring artifacts, our networks are also trained to correct for partial ring artifacts as they may occur in spiral CT or CBCT. This study shows that ring artifacts can be reduced in image domain by these neural networks. Our results suggest that a three-dimensional network is most suitable for this task.

Motion artifact is a major challenge in cardiac CT which hampers accurate delineation of key anatomic (e.g. coronary lumen) and pathological features (e.g. stenosis). Conventional motion correction techniques are limited on patients with high / irregular heart rate, due to simplified modeling of CT systems and cardiac motion. Emerging deep learning based cardiac motion correction techniques have demonstrated the potential of further quality improvement. Yet, many methods require CT projection data or advanced motion simulation tools that are not readily available. We aim to develop an image-domain motion-correction method, using convolutional neural network (CNN) integrated with customized attention and spatial transformer techniques. Forty cardiac CT exams acquired from a clinical dual-source CT system were retrospectively collected to generate training (n=26) and testing (n=14) sets. Dual-source data uniquely allow image reconstruction with different temporal resolutions from the same patient scan. Slow temporal resolution (140ms; equivalent to single-source CT (SSCT) half scan) and fast temporal resolution (75ms; dual source) images were reconstructed to generate paired samples of motion-corrupted and reference images. The combinations of 2 training inference strategies and 3 CNNs were evaluated: strategy #1 – whole-heart images in training / inference; strategy #2 – vessel patches in training / inference; CNN #1 – attention only; CNN #2 – spatial-transformer (STN) only; CNN #3 – attention and STN synergy. Testing data showed that CNN #3 with strategy #2 provided relatively better performance: improving vessel delineation, increasing structural similarity index from 0.85 to 0.91, and reducing mean CT number error of lumen by 71.0%. Our method could improve the image quality in cardiac exams with SSCT.

Preclinical PET imaging is widely used to quantify in vivo biological and metabolic process at molecular level in small animal imaging. In preclinical PET, low-count acquisition has numerous benefits in terms of animal logistics, maintaining integrity in longitudinal multi-tracer studies, and increased throughput. Low-count acquisition can be realized by either decreasing the injected dose or by shortening the acquisition time. However, both these approaches lead to reduced photons, generating PET images with low signal-to-noise ratio (SNR) exhibiting poor image quality, lesion contrast, and quantitative accuracy. This study is aimed at developing a deep-learning (DL) based framework to generate high-count PET (HC-PET) from low-count PET (LC-PET) images using Residual U-Net (RU-Net) and Dilated U-Net (D-Net)-based architectures. Preclinical PET images at different photon count levels were simulated using a stochastic and physics-based method and fed into the framework. The integration of residual learning in the U-Net architecture enhanced feature propagation while the dilated kernels enlarged receptive field-of-view to incorporate multiscale context. Both DL methods exhibited significantly (p≤0.05) better performance in terms of Structural Similarity Index Metric (SSIM), Peak Signal-to-Noise Ratio (PSNR) and Normalized Root Mean Square Error (NRMSE) when compared to existing non-DL denoising techniques such as Non-Local Means (NLM) and BM3D filtering. In objective evaluation of quantification task, the DL-based approaches yielded significantly lower bias in determining the mean standardized uptake value (SUVmean) of liver and tumor lesion than the non-DL approaches. Of the DL frameworks, D-Net based generation of HC-PET had the least bias and coefficient of variation at all photon count levels. Our study suggests that DL can predict HC-PET images with improved visual quality and quantitative accuracy from LC-PET (preclinical) images.

The tissue specific MRF type texture prior (MRFt) proposed in our previous work has been demonstrated to be advantageous in various clinical tasks. However, this MRFt model requires a previous full-dose CT (FdCT) scan of the same patient to extract the texture information for LdCT reconstructions. This requirement may not be met in practice. To alleviate this limitation, we propose to build a MRFt generator by internalizing a database with paired FdCT and LdCT scans using a (conditional) encoder-decoder structure model. We denote this method as the MRFtG-ConED. This generation model depends only on physiological features thus is robust for ultra-low dose CT scans (i.e., dosage < 10mAs). When the dosage is not extremely low (i.e., dosage > 10mAs), some texture information from LdCT images reconstructed by filtered back projection (FBP) can be also used to provide extra information.

Purpose: Cone-beam CT has become commonplace for 3D guidance in interventional radiology (IR), especially for vascular procedures in which identification of small vascular structures is crucial. However, its long image acquisition time poses a limit to image quality due to soft-tissue deformable motion that hampers visibility of small vessels. Autofocus motion compensation has shown promising potential for soft-tissue deformable motion compensation, but it lacks specific target to the imaging task. This work presents an approach for deformable motion compensation targeted at imaging of vascular structures. Methods: The proposed method consists on a two-stage framework for: i) identification of contrast-enhanced blood vessels in 2D projection data and delineation of an approximate region covering the vascular target in the volume space, and, ii) a novel autofocus approach including a metric designed to promote the presence of vascular structures acting solely in the region of interest. The vesselness of the image is quantified via evaluation of the properties of the 3D image Hessian, yielding a vesselness filter that gives larger values to voxels candidate to be part of a tubular structure. A cost metric is designed to promote large vesselness values and spatial sparsity, as expected in regions of fine vascularity. A targeted autofocus method was designed by combining the presented metric with a conventional autofocus term acting outside of the region of interest. The resulting method was evaluated on simulated data including synthetic vascularity merged with real anatomical features obtained from MDCT data. Further evaluation was obtained in two clinical datasets obtained during TACE procedures with a robotic C-arm (Artis Zeego, Siemens Healthineers). Results: The targeted vascular autofocus effectively restored the shape and contrast of the contrast-enhanced vascularity in the simulation cases, resulting in improved visibility and reduced artifacts. Segmentations performed with a single threshold value on the target vascular regions yielded a net increase of up to 42% in DICE coefficient computed against the static reference. Motion compensation in clinical datasets resulted in improved visibility of vascular structures, observed in maximum intensity projections of the contrast-enhanced liver vessel tree. Conclusion: Targeted motion compensation for vascular imaging showed promising performance for increased identification of small vascular structures in presence of motion. The development of autofocus metrics and methods tailored to vascular imaging opens the way for reliable compensation of deformable motion while preserving the integrity of anatomical structures in the image.

Cone-beam CT (CBCT) with non-circular acquisition orbits has the potential to improve image quality, increase the field-of view, and facilitate minimal interference within an interventional imaging setting. Because time is of the essence in interventional imaging scenarios, rapid reconstruction methods are advantageous. Model-Based Iterative Reconstruction (MBIR) techniques implicitly handle arbitrary geometries; however, the computational burden for these approaches is particularly high. The aim of this work is to extend a previously proposed framework for fast reconstruction of non-circular CBCT trajectories. The pipeline combines a deconvolution operation on the backprojected measurements using an approximate, shift-invariant system response prior to processing with a Convolutional Neural Network (CNN). We trained and evaluated the CNN for this approach using 1800 randomized arbitrary orbits. Noisy projection data were formed from 1000 procedurally generated tetrahedral phantoms as well as anthropomorphic data in the form of 800 CT and CBCT images from the Lung Image Database Consortium Image Collection (LIDC). Using this proposed reconstruction pipeline, computation time was reduced by 90% as compared to MBIR with only minor differences in performance. Quantitative comparisons of nRMSE, FSIM and SSIM are reported. Performance was consistent for projection data simulated with acquisition orbits the network has not previously been trained on. These results suggest the potential for fast processing of arbitrary CBCT trajectory data with reconstruction times that are clinically relevant and applicable - facilitating the application of non-circular orbits in CT image-guided interventions and intraoperative imaging.

Purpose: We investigated the feasibility of detection and quantification of bone marrow edema (BME) using dual-energy (DE) Cone-Beam CT (CBCT) with a dual-layer flat panel detector (FPD) and three-material decomposition. Methods: A realistic CBCT system simulator was applied to study the impact of detector quantization, scatter, and spectral calibration errors on the accuracy of fat-water-bone decompositions of dual-layer projections. The CBCT system featured 975 mm source-axis distance, 1,362 mm source-detector distance and a 430 × 430 mm² dual-layer FPD (top layer: 0.20 mm CsI:Tl, bottom layer: 0.55 mm CsI:Tl; a 1 mm Cu filter between the layers to improve spectral separation). Tube settings were 120 kV (+2 mm Al, +0.2 mm Cu) and 10 mAs per exposure. The digital phantom consisted of a 160 mm water cylinder with inserts containing mixtures of water (volume fraction ranging 0.18 to 0.46) - fat (0.5 to 0.7) - Ca (0.04 to 0.12); decreasing fractions of fat indicated increasing degrees of BME. A two-stage three-material DE decomposition was applied to DE CBCT projections: first, projection-domain decomposition (PDD) into fat-aluminum basis, followed by CBCT reconstruction of intermediate base images, followed by image-domain change of basis into fat, water and bone. Sensitivity to scatter was evaluated by i) adjusting source collimation (12 to 400 mm width) and ii) subtracting various fractions of the true scatter from the projections at 400 mm collimation. The impact of spectral calibration was studied by shifting the effective beam energy (± 2 keV) when creating the PDD lookup table. We further simulated a realistic BME imaging framework, where the scatter was estimated using a fast Monte Carlo (MC) simulation from a preliminary decomposition of the object; the object was a realistic wrist phantom with an 0.85 mL BME stimulus in the radius. Results: The decomposition is sensitive to scatter: approx. <20 mm collimation width or <10% error of scatter correction in a full field-of-view setting is needed to resolve BME. A mismatch in PDD decomposition calibration of ± 1 keV results in ~25% error in fat fraction estimates. In the wrist phantom study with MC scatter corrections, we were able to achieve ~0.79 mL true positive and ~0.06 mL false positive BME detection (compared to 0.85 mL true BME volume). Conclusions: Detection of BME using DE CBCT with dual-layer FPD is feasible, but requires scatter mitigation, accurate scatter estimation, and robust spectral calibration.

Simultaneous use of kilovoltage (kV) and megavoltage (MV) beams has numerous potential applications in cone beam computed tomography (CBCT)-guided radiotherapy, such as fast MV+kV CBCT for single breath-hold scan, tumor localization with kV CBCT imaging during MV therapy delivery, and metal artifact suppression. However, the introduction of MV beams results in a large MV-cross scatter fluence incident on the kV Flat Panel Detector (FPD), and thus, deteriorating the low contrast visualization and Hounsfield Unit (HU) accuracy. In this work, we introduced a novel and robust method for reducing the effects of MV cross scatter. First, we implemented a 2D antiscatter grid atop the detector which rejects a large section of MV cross scatter. This hardware-based approach, while effective, allows a fraction of MV cross scatter to be transmitted to the FPD, resulting in artifacts and degraded HU accuracy in CBCT images. We thus introduced a data correction step, which aimed to estimate and correct the remaining MV cross scatter. This approach, referred to as Grid-Based Scatter Sampling, utilized 2D antiscatter grid itself to measure and correct remaining MV cross scatter in projections. We investigated the performance of the proposed approach in experiments by simultaneously acquiring kV CBCT and delivering MV beams with a clinical linac. The results show that the proposed method can substantially reduce HU inaccuracy and increase contrast-to-noise ratio (CNR). Our method does not require synchronization of kV and MV beam pulses, reduction of kV frame acquisition rate, or MV dose rate, and therefore, it is more practical to implement in radiation therapy clinical setting.

Quantitative digital subtraction angiography (qDSA) aims to standardize clinical endpoints for embolization procedures by measuring arterial blood velocity from 2D image sequences intra-procedurally until a predetermined reduction in velocity is achieved. qDSA determines velocity by tracking oscillations in iodine contrast induced from the cardiac cycle as they propagate along a vessel centerline. The purpose of this work was to investigate the influence of vessel motion on the accuracy of qDSA and to develop a motion compensated qDSA approach (MC-qDSA). MC-qDSA uses a deeplearning-based frame-by-frame vessel segmentation approach followed by vessel registration. This allows automatic recalculation of centerlines for each frame and thus provides a dynamic centerline that moves with the vasculature. The approach was tested in a phantom study simulating physiologic blood flow, respiratory and cardiac motion. Errors in qDSA velocity measurements relative to the ideal case of a stationary phantom were quantified using the mean absolute percent difference (MAD). In the respiratory motion cases, no motion correction resulted in 160.0 ± 283.0% MAD and the proposed MC-qDSA approach improved the MAD to 6.5 ± 5.9%. In the cardiac motion cases, MAD was 11.9 ± 12.8% without motion correction and 5.4 ± 4.1% using MC-qDSA. A retrospective swine study was performed in motion corrupted image sequencesto test the ability of MC-qDSA to correctly determine vessel centerlines with more anatomically realistic backgrounds and vessels. The average Hausdorff distance between non-motion-compensated vessel centerlines and manually annotated centerlines was 1.8 ± 1.5mm and for the motion compensated centerlines it was 0.5 ± 0.2mm. The proposed method provides a potential means of utilizing qDSA in motion corrupted image sequences.

Today, 2D+T fluoroscopy is usually used for image guidance in interventional radiology. For challenging procedures, 4D (3D+T) image guidance would be advantageous. The difficulty in realizing X-ray-based 4D interventional guidance lies in the development of a very dose efficient reconstruction algorithm. To this end, we improve on a previously presented algorithm for the reconstruction of interventional tools. By incorporating temporal information into a 3D convolutional neural network, we reduce the number of X-ray projections that need to be acquired for the 3D reconstruction of guidewires from four to two, thereby halving dose and decreasing the demands put on imaging devices implementing the algorithm. In experiments with two moving guidewires in an anthropomorphic phantom, we observe little deviation of our 3D reconstructions from the ground truth.

Intraoperative imaging using C-arm X-ray systems enables percutaneous management of fractures by providing real-time visualization of tool to tissue relationships. However, estimating appropriate positioning of surgical instruments, such as K-wires, relative to safe bony corridors is challenging due to the projective nature of X-ray images: tool pose in the plane containing the principal ray is difficult to assess, necessitating the acquisition of numerous views onto the anatomy. This task is especially demanding in complex anatomy, such as the superior pubic ramus of the pelvis, and results in high cognitive load and repeat attempts even in experienced trauma surgeons. A perception-based algorithm that interprets interventional radiographs during internal fixation to infer the likelihood of cortical breach – especially early on, when the wire has not been advanced – might reduce both the amount of X-rays acquired for verification and the likelihood of repeat attempts. In this manuscript, we present first steps towards developing such an algorithm. We devise a strategy for in silico collection and annotation of X-ray images suitable for detecting cortical breach of a K-wire in the superior pubic ramus, including those with visible fractures. Beginning with minimal manual annotations of correct trajectories, we randomly perturb entry and exit points and project the 3D scene using a physics-based forward model to obtain a large number of 2D X-ray images with and without cortical breach. We report baseline results for anticipating cortical breach at various K-wire insertion depths, achieving an AUROC score of 0.68 for 50% insertion.

Stereotactic radiosurgery (SRS) is frequently employed to treat brain metastases. However, <50% of patients treated with this method develop distant brain metastases (DBMs). As a result, these patients are followed using Magnetic Resonance Imaging (MRI) to identify DBM development. There is no current pre-treatment risk metric to identify which patients might be likely to develop DBMs. In this study, pre-treatment MRIs and radiotherapy planning data including structure sets and radiation dose maps were obtained for 81 SRS brain metastases treatment courses. Clinical variables including performance status, age, number of tumors, and primary tumor type were also collected. Pre-treatment MRIs were skull-stripped and normalized. 3D radiomic features from grey-intensity, Laws Energy, Gabor, Haralick, and CoLlAGe feature families were extracted from T1, T1 contrast-enhanced (T1w), T2, and FLAIR pre-treatment MRI sequences in brain regions receiving 0-25%, 25-50%, 50-75%, and 75-100% of prescribed radiation dose. A baseline classification model for DBM was created using clinical variables. Ablation studies were performed to determine which dose region and MRI sequence contained radiomic features most predictive for DBM development using machine learning (ML) classifiers. An ML classifier trained on 3D radiomic features from the 50-75% dose region of pre-treatment T1w MRI (AUC: 0.71, 95% CI: 0.68-0.74) outperformed the baseline model (AUC: 0.50, 95% CI: 0.47-0.53) for DBM prediction. In conclusion, we leverage radiotherapy dose regions to identify subcompartments for radiomic feature extraction from multi-parametric pre-treatment MRI data. We demonstrate that radiomic features from these dose regions can be used to predict DBM for SRS-treated brain metastases.

Dual-source CT can provide high temporal resolution single energy imaging, but this is sacrificed in dual-energy mode (DS-DE). The current study evaluates a dual-source photon-counting-detector (DS-PCD) CT capable of high temporal resolution (66 ms) dual-energy imaging. A rod simulating coronary artery motion was scanned on both DS-DE and DS-PCD CT scanners and image quality was evaluated. Spatial registration—quantified by dice coefficient—between the high and low energy images was better with DS-PCD than DS-DE. Furthermore, the rod had a more circular appearance and its diameter was more accurate in iodine maps generated by DS-PCD whereas DS-DE suffered from motion artifacts.

The purpose of this study was to develop a virtual imaging framework that simulates a new photon-counting CT (PCCT) system (NAEOTOM Alpha, Siemens). The PCCT simulator was built upon the DukeSim platform, which generates projection images of computational phantoms given the geometry and physics of the scanner and imaging parameters. DukeSim was adapted to account for the geometry of the PCCT prototype. To model the photon-counting detection process, we utilized a Monte Carlo-based detector model with the known properties of the detectors. We validated the simulation platform against experimental measurements. The images were acquired at four dose levels (CTDIvol of 1.5, 3.0, 6.0, and 12.0 mGy) and reconstructed with three kernels (Br36, Br40, Br48). The experimental acquisitions were replicated using our developed simulation platform. The real and simulated images were quantitatively compared in terms of image quality metrics (HU values, noise magnitude, noise power spectrum, and modulation transfer function). The clinical utility of our framework was demonstrated by conducting two clinical applications (COPD quantifications and lung nodule radiomics). The phantoms with relevant pathologies were imaged with DukeSim modeling the PCCT systems. Different imaging parameters (e.g., dose, reconstruction techniques, pixel size, and slice thickness) were altered to investigate their effects on task-based quantifications. We successfully implemented the acquisition and physics attributes of the PCCT prototype into the DukeSim platform. The discrepancy between the real and simulated data was on average about 2 HU in terms of noise magnitude, 0.002 mm^-1 in terms of noise power spectrum peak frequency and 0.005 mm^-1 in terms of the frequency at 50% MTF. Analysis suggested that lung lesion radiomics to be more accurate with reduced pixel size and slice thickness. For COPD quantifications, higher doses, thinner slices, and softer kernels yielded more accurate quantification of density-based biomarkers. Our developed virtual imaging platform enables systematic comparison of new PCCT technologies as well as optimization of the imaging parameters for specific clinical tasks.

Coronary artery calcification is an important indicator of coronary disease. Accurate volume quantification of coronary calcification using computed tomography (CT) is challenging due to calcium blooming. In this study, ex-vivo coronary specimens were scanned on an investigational photon-counting detector (PCD) CT scanner and the estimated coronary calcification volume were compared with a conventional energy-integrating detector (EID) CT. An image-based denoising algorithm was applied to the PCD-CT images to achieve similar noise levels as EID-CT. Calcifications were segmented to estimate the volume, with micro-CT images of the same calcifications serving as reference. PCD-CT images showed reduced calcium blooming artifacts.

Aim: To develop and subsequently perform a systematic study on the impact of parameter settings on the biological reproducibility and sensitivity of extracted radiomic features from Full Field Digital Mammography (FFDM) images for the task of Breast Cancer Risk assessment. Methods: Cranio-caudal (CC) ”FOR PRESENTATION” images (88 in total, two centers: Slovenia and Belgium) were used for this study. Biological reproducibility of radiomic features was evaluated with two tests: reproducibility of extracted features between left and right breasts and by reproducibility of extracted features between the original and 4 perturbed images. The quantification was done using the intra-class correlation (ICC) coefficient between values of extracted radiomic features. To determine biological sensitivity, AUC between groups with low and high breast cancer risk was calculated. For the selection of optimal radiomic feature parameters, thresholds of 0.75 and 0.7 were defined for ICC and AUC, respectively. Results: Parameters bin Count and distances highly influenced biological reproducibility and sensitivity of specific radiomic features. Parameters weightingNorm and symmetricalGLCM had no effect. Overall, only 12/93 radiomic features passed the reproducibility and sensitivity tests in both centers. For five of these features, parameter ranges were crucial. Reproducibility varied greatly between the centers of Belgium and Slovenia. Conclusions: Rather than single radiomic parameters, parameter ranges were found to be a reasonable description for acceptable biological reproducibility and sensitivity. Overall, 12/93 radiomic features were found to be potential candidates for breast cancer risk prediction tasks, however further analysis is needed before definitive recommendations can be made.

Digital Breast Tomosynthesis (DBT) is becoming increasingly popular for breast cancer screening because of its high depth resolution. It uses a set of low-dose x-ray images called raw projections to reconstruct an arbitrary number of planes. These are typically used in further processing steps like backprojection to generate DBT slices or synthetic mammography images. Because of their low x-ray dose, a high amount of noise is present in the projections. In this study, the possibility of using deep learning for the removal of noise in raw projections is investigated. The impact of loss functions on the detail preservation is analized in particular. For that purpose, training data is augmented following the physics driven approach of Eckert et al.1 In this method, an x-ray dose reduction is simulated. First pixel intensities are converted to the number of photons at the detector. Secondly, Poisson noise is enhanced in the x-ray image by simulating a decrease in the mean photon arrival rate. The Anscombe Transformation2 is then applied to construct signal independent white Gaussian noise. The augmented data is then used to train a neural network to estimate the noise. For training several loss functions are considered including the mean square error (MSE), the structural similarity index (SSIM)3 and the perceptual loss.4 Furthermore the ReLU-Loss1 is investigated, which is especially designed for mammogram denoising and prevents the network from noise overestimation. The denoising performance is then compared with respect to the preservation of small microcalcifications. Based on our current measurements, we demonstrate that the ReLU-Loss in combination with SSIM improves the denoising results.

Iodine contrast-enhanced spectral mammography (CEM) combines an iodinated contrast agent, such as one used for a typical CT scan, with mammography imaging. The contrast enhancement improves the ability to visualize some cancers, and so it has been proposed as a costeffective and robust alternative to magnetic resonance imaging (MRI) for breast cancer imaging, especially in dense breasts. However, one drawback is poor quantification of contrast agent due to the two-dimensional projection in mammogram images. Digital breast tomosynthesis (DBT) is a pseudo-three-dimensional (3D) imaging modality that uses limited angle tomography. DBT typically exhibits high in-plane resolution, with poor out-of-plane resolution. This out-of-plane blur in DBT distorts the reconstructed lesion and can degrade lesion quantification and volume estimation. This work will explore whether convolutional neural networks (CNN) can be trained to predict a full angle CT reconstruction of the lesion from a limited angle DBT input image of the lesion. Various networks were trained to perform this image restoration using a large number of Monte-Carlo simulated lesion volumes-of-interest (VOI) from DBT and breast CT reconstructions. Our preliminary results show that the output images from the trained neural networks yield a more accurate values in terms of lesion quantification and volume estimation than those estimated from their DBT counterparts.

We investigate the effects of training strategies, such as the dataset size, early stop and cross-validation, on the performance of deep neural network. We used the residual network architecture, dedicated to restore low-dose (LD) digital breast tomosynthesis (DBT) raw projections. For this assessment, we generated 500 synthetic breast phantoms through virtual clinical trials software (OpenVCT) and validated them in terms of noise and signal properties. We acquired real raw DBT projections using a physical anthropomorphic breast phantom and restored the LD projections after training the convolutional neural networks. We found that early stop can be applied in the training process depending on the denoising strength desired for the network. Also, different training realizations are necessary to achieve good results. Furthermore, the training sample size may be smaller compared to other computer vision tasks using deep learning algorithms.

Noiseless digital mammograms (DM) are unobtainable in clinical screening environments, limiting the development of deep learning-based (DL) denoising applications. Virtual clinical trials (VCTs) allow the precise simulation of noise levels in DM images for controlled training of DL models. We evaluated a set of DL denoising models, trained using VCT data, that showcases the trade-offs between denoising strength and fine structure preservation. Our results show that metrics, such as peak signal-to-noise ratio (PSNR), are improved with the use of our trained residual convolutional neural network. This quantifiable improvement indicates that our proposed DL methodology can accurately denoise simulated mammograms.

We compare simulated mammograms with real mammogram images using radiomics texture features. Simulated mammogram images are from imaging VICTRE breast models using Monte Carlo x-ray transport simulation. CBIS-DDSM dataset is used for the real mammograms. Texture features are calculated on the entire breast to compare the overall texture. Our results indicate that the simulated images can be seen as a subset of the CBIS-DDSM dataset in terms of textural variability. This preliminary work shows that radiomics can be a useful tool for comparing breast characteristics between different datasets, including simulated versus real datasets.

The objective of this work was to create a MLEM software-based scatter correction algorithm for removing the effect of Compton Scatter from mammography images acquired without scatter grid or with analyzer-less interferometry. We developed an MLEM algorithm with an efficient linear scatter model to estimate the thickness of compressed breast and evaluated the algorithm with breast images acquired with the GEANT4 Monte Carlo software. The thicknesses estimated from the algorithm on the GEANT4 images were compared to the true geometric thicknesses of the ellipsoid for each pixel of the detector and matched to within ~2mm RMS error.

Virtual clinical trials (VCTs) in medical imaging can be used as a supplement to clinical trials, as they are less costly and less time-consuming and do not depend on study participants. This study aims at modeling the variety in volumetric breast density (VBD) in a virtual population of women attending mammography and breast tomosynthesis screening. Probability distributions (PDs) were fitted to available clinical data on VBD from Malmö, for women in different age groups. The PDs were then used for sampling VBD values for virtual populations. Each age group had two virtual populations, one containing N=30 samples and the other N>1500 samples. The sampled VBD and the clinical VBD were compared both for the large and small populations. Results showed that the VBD probability distributions vary between different age groups in the clinical population and that the mean VBD decreases with aging. It also showed that the simulated VBD (both the large and small populations) does not significantly differ from the clinical data (>0.07) and that the mean VBD also decreases for the simulated populations. The larger population (N>1500) had a smaller percent error than the smaller virtual population (N=30) when compared with the clinical mean VBD values. The study suggests that the method could be applied when simulating VBD for virtual patients in VCTs.

In many cone-beam CT (CBCT) designs (linac-mounted and C-arms), source and detector trajectories are reproducible but deviate from their nominal paths due to mechanical flex. Since accurate knowledge of the scan geometry is crucial for CBCT image quality, angle-dependent geometric calibration is required, generally using phantoms with bearing balls (BBs) at known positions. However, when tangentially offsetting the detector to increase the field-of-view, multi-BB calibration is not ideal because: (a) the same BBs are not visible in opposite views, negatively affecting the robustness in matching opposing projections, and (b) the calibration becomes more sensitive to deviations from nominal BB positions, e.g. from manufacturing tolerances or post-manufacturing deformation. We present a novel calibration method for scanners with significantly offset detector. Inspired by checkerboard camera calibration, it uses circuit boards with etched copper patterns, readily available to micrometre precision. The method offers full nine-degrees-of-freedom calibration, is robust to phantom assembly and positioning inaccuracies, and uses patterns that cover a large portion of the detector and extend across the overlap region to improve matching of opposing views. For quantitative evaluation was used the smallest sphere intersected by rays back-projected from a 360-degree scan of BBs 100 mm from the isocentre. For a linac-mounted CBCT, the average radius was decreased from 0.48 mm (with single-BB calibration) to 0.14 mm. Corresponding improvements were seen in off-centre slices of Catphan phantom reconstructions. We propose the method as an easy-to-manufacture, simple-to-use, robust alternative to multi-BB calibration for CBCT systems with offset detector, suitable also for noncircular orbits.

A carbon nanotube (CNT) emitter-based high-speed micro-computed tomography (CT) system is designed herein. CT scans result in prolonged exposure to radiation, thus, pulsed-field emission is performed to generate pulsed X-rays to reduce radiation dose. The radiation dose rates of continuous X-rays and X-rays generated by different pulses were compared. X-rays were stably generated 40 times per second through a pulsed-field emission current with MOSFET circuit. The CNT field emitter characteristic was maintained for more than 35 hours continuously, which means more than 5 million X-ray shots can be taken when sufficient current flows on pulse mode. X-ray images were also successfully acquired through such pulse driving.

We developed a region-of-interest (ROI) image reconstruction method that effectively reduces truncation artifacts in CBCT. By using U-Net-based deep learning (DL) methods, we devised a method to reduce truncation artifacts for ROI imaging. A total of 16294 image slices from 49 patient cases were used to generate projection data. The center of the projected image was cropped to a width of 150 mm. Then, the outer part of the truncation image was filled with each outermost pixel value for the initial correction. After the filtering process, the truncation area was cut off and used as input data in the DL model. Finally, inference images were reconstructed by use of the FDK algorithm. SSIM values for the test set of 14 patients were calculated as 0.541, 0.709 and 0.979 for FBP, Extension and the proposed ROI method, respectively. We have achieved promising results and believe that the proposed ROI image reconstruction method can help reduce radiation dose while preserving image quality

The cone-beam CT (CBCT) imaging systems that based on flat panel detectors have been widely implemented in image-guided intervention and radiation therapy applications. However, the imaging performance of CBCT is strongly limited. One of such limitations is the lack of quantitative imaging capability, which is important for material recognition, image contrast enhancement, and dose reduction. Over the past decade, dual-energy computed tomography (DECT) has become a promising imaging technique in generating quantitative material information, whereas, multiple (>2) basis images with high quality and accuracy are hard to be obtained from the conventional DECT image reconstruction algorithms. In this work, an innovative deep learning technique is presented to realize three materials decomposition from the dual-energy CBCT scans. In this strategy, a dedicated end-to-end convolutional neural network (CNN) is developed. It accepts the low and high energy CBCT projections, and automatically outputs three different basis image volumes (water basis, iodine basis, CaCl₂ basis) with high accuracy. Training data was synthesized numerically from the photos downloaded from ImageNet. Dual-energy projections of the Iodine/CaCl2 phantom with ground truth were acquired from our in-house benchtop CBCT system to validate the proposed method. Results demonstrate that this novel network is able to generate three different material bases with high accuracy (decomposition errors less than 5%). In conclusion, the proposed CNN based multi-material (≥ 3) decomposition approach shows promising benefits in high quality dual-energy CBCT imaging applications.

Deep learning (DL) applications strongly depend on the training dataset size; however, it is unclear how to objectively select this parameter. This work investigates the dependence of the classification performance of a DL model on the dataset size for the anatomic classification of cone-beam computed tomography (CBCT) projections. CBCT scans from 1055 patients were collected and manually classified into 5 anatomic classes and used to develop a DL model to predict the anatomic class from 0º x-ray projections. A VGG-16 convolutional neural network (CNN) was trained with 75% of the data, while the remaining 25% was used for testing and evaluation. The training dataset was downsampled to various sizes. The training dataset size could be reduced to ~50% of its initial size, without compromising the classification performance.

Cone-beam computed tomography (CBCT) provides volumetric scans while keeping a relatively low dose and cost. This characteristic makes CBCT favorable in the field of dentistry, but the frequent presence of metallic objects in patients causes metal artifact that severely degrades the resulting images. Dual-energy CT provides more information about the materials which can be used to synthesize virtual monoenergetic images (VMI) to reduce metal artifacts. In this work, we investigated and optimized a dual-energy cone-beam CT (DE-CBCT) system using a carbon-nanotube CNT x-ray source with dual focal spots and spectral filtrations. We used two x-ray spectra generated by applying spectral filters at a constant x-ray tube voltage. We imaged an anthropomorphic head phantom with metal beads. The projection images were reconstructed separately using an iterative CT reconstruction algorithm and bilateral filtering is applied to the reconstructed images for denoising. The VMIs were synthesized from the denoised reconstructed images using an image domain decomposition method and displayed noise comparable to the images obtained from a single energy. The resulting VMIs displayed fewer metal artifacts compared to the single energy images of the same object from a clinical CBCT scanner. To optimize the spectral separation, we implemented a filter selection algorithm and improved the mean energy separation from 13keV to 17.5keV while maintaining the same x-ray output. We also demonstrated the ability of this system to independently control the imaging dose of the two energy scans by varying the x-ray exposure current and time using the dual focal spot CNT source.

Dynamic dose modulation techniques are applied in CT acquisitions to ensure optimal working conditions of the detector unit and to reduce overall radiation exposure of the patient. In C-arm CT systems, large variations in the desired irradiation may require tube voltage modulation (TVM). Recent studies showed that TVM does not affect the quality of perfusion images obtained by clinical CT. Here, we investigate the impact of TVM in a C-arm cone beam CT perfusion imaging setting. We conduct a simulation study based on a real perfusion acquisition (incl. tube modulation) to directly compare results from acquisitions with and without modulation. Using two different reconstruction techniques, we analyze the influence of TVM on the extracted perfusion parameters and quantify the similarity by their correlation coefficients. Our results demonstrate that high correlation (r < 0.99) between the results with and without TVM are achieved for all perfusion parameters using a straightforward and model-based reconstruction technique. These findings suggest that dose modulation techniques, incl. TVM, can be used in C-arm CT perfusion scans without the need for additional correction methods to retain image quality of constant voltage scans.

In many interventional settings it would be beneficial to perform a final CBCT acquisition for outcome control after the intervention is done. However, due to high patient dose this is often omitted. Volume-of-interest acquisitions offer considerable dose reduction, but image reconstruction typically suffers from cupping artifacts and offsets in radiodensity due to the truncated projection data. In a previous work we presented a method which allows to incorporate available prior volume data into the reconstruction of volume-of-interest acquisitions in CBCT. The method works by making use of the fluoroscopic positioning images typically acquired before CBCT acquisitions in a 3D Radon space-based registration method registering the prior volume to the volumeof-interest scenario. Here, we demonstrate the application of this method on real clinical data or the first time.

Cone-beam CT (CBCT) plays a crucial role in modern image-guided radiotherapy for patient setup and verification. However, the image quality of CBCT is inferior to that of CT in terms of HU accuracy, image artifact, and tissue contrast, which impedes the potential of CBCT in further radiotherapy applications, such as online contouring and dose calculation for adaptive radiotherapy. In this work, we propose an optimization model for material decomposition of spectral CBCT, which innovatively incorporates CT as guidance to simultaneously improve the image quality and material composition of CBCT images. Both phantom and patient studies demonstrate the effectiveness and superiority of the proposed method in noise and artifact removal, decomposition-accuracy maintenance, etc.

Respiratory-Correlated cone beam computed tomography (4D-CBCT) is an emerging image-guided radiation therapy (IGRT) technique that is used to account for the uncertainties caused by respiratory-induced motion in the radiotherapy treatment of tumors in thoracic and upper-abdomen regions. In 4D-CBCT, projections are sorted into bins based on their respiratory phase and a 3D image is reconstructed from each bin. However, the quality of the resulting 4D-CBCT images is limited by the streaking artifacts that result from having an insufficient number of projections in each bin. In this work, an interpolation method based on Convolutional Neural Networks (CNN) is proposed to generate new in-between projections to increase the overall number of projections used in 4D-CBCT reconstruction. Projections simulated using XCAT phantom were used to assess the proposed method. The interpolated projections using the proposed method were compared to the corresponding original projections by calculating the peak-signal-to-noise ratio (PSNR), root mean square error (RMSE), and structural similarity index measurement (SSIM). Moreover, the results of the proposed method were compared to the results of existing standard interpolation methods, namely, linear, spline, and registration-based methods. The interpolated projections using the proposed method had an average PSNR, RMSE, and SSIM of 35.939, 4.115, and 0.968, respectively. Moreover, the results achieved by the proposed method surpassed the results achieved by the existing interpolation methods tested on the same dataset. In summary, this work demonstrates the feasibility of using CNN-based methods in generating in-between projections and shows a potential advantage to 4D-CBCT reconstruction.

Compared to energy-integration, photon-counting detection of x-ray facilitates spectral channelization (energy binning) in spectral CT and thus offers the opportunity of implementing sophisticated data acquisition schemes, e.g., gapping or interleaving in spectral channels. In this work, we investigate the feasibility and performance of material decomposition based spectral imaging in photon-counting CT via such data acquisition schemes. Using a deliberately designed anthropomorphic head phantom, a simulation study is carried out with the focus on two-material decomposition-based spectral imaging, under both ideal and realistic detector spectral responses. The projection data are acquired in four spectral channels and are sorted to implement the schemes of gapping ((ch₁, ch₃); (ch₂, ch₄); (ch₁, ch₄)) and interleaving ((ch₁, ch₃)+(ch₂, ch₄); (ch₁, ch₄)+(ch₂, ch₃); ((ch₁+ch₃), (ch₂+ch₄)); ((ch₁+ch₄), (ch₂+ch₃))) in spectral channels, in addition to the benchmark scheme ((ch₁+ch₂), (ch₃+ch₄)) and other conventional schemes (ch₁, ch₂), (ch₂, ch₃) and (ch₃, ch₄), where ‘ch’ denotes channel, ‘+’ denote addition, and (·,·) the operation of material decomposition and image reconstruction. The contrast-to-noise ratio between targeted regions of interest as the figure of merit in the study. It is observed that under ideal detector spectral response, the scheme (ch₁, ch₄) outperforms the benchmark scheme ((ch₁+ch₂), (ch₃+ch₄)) and others in gapped and/or interleaved spectral channelization in material specific imaging, while the interleaved scheme (ch₁, ch₄)+(ch₂, ch₃) performs the best in virtual monochromatic imaging. Under realistic detector spectral response, the difference in imaging performance over all spectral channelization schemes diminishes, along with degradation in each scheme’s individual performance. The spectral channelization schemes and associated imaging performance reported herein are novel and informative to the community, which may further the understanding of fundamental physics and design principles for spectral imaging in photon-counting CT and other x-ray related imaging modalities.

Deep convolutional neural network (CNN) based methods have become popular choices for reducing image noise in CT. Some of these methods showed promising results, especially in terms of preserving natural CT noise texture. Early attempts of CNN denoising were based on 2D CNN models with either single-slice or 3-slice input. The 3-slice input was mainly to utilize the existing network architecture that were proposed for natural images with 3 input channels. Multi-slice input has the potential to incorporate spatial information from adjacent slices. However, it remains unknown if this strategy indeed improves the denoising performance compared to a 2D model with a single-slice input and what is the best network architecture to utilize the multi-slice input. Two categories of network architectures can be used for multi-slice input. First, multi-slice low-dose images can be stacked channelwise as multi-channel input to a 2D CNN model. Second, multi-slice images can be employed as the 3D volumetric input to a 3D CNN model, in which the 3D convolution layers are adopted. In this study, we compare the performance of multiple CNN models with 1, 3, and 7 input slices. For the 7-slice input, we also include a comparison between 2D and 3D CNN models. When the input channels of the 2D CNN model increases from 1 to 3 to 7, a trend of improved performance was observed. Comparing the two models with 7-slice input, the 3D model slightly outperforms the 2D model in terms of noise texture and homogeneity in liver parenchyma as well as better subjective visualization of vessels such as intrahepatic portal vein and jejunal artery.

CT imaging provides physicians valuable insights when diagnosing disease in a clinical setting. In order to provide an accurate diagnosis, is it important to have a high accuracy with controlled variability across CT scans from different scanners and imaging parameters. The purpose of this study was to analyze variability of lung imaging biomarkers across various scanners and parameters using a customized version of a commercially available anthropomorphic chest Phantom (Kyoto Kagaku) with several experimental sample inserts. The phantom was across 10 different CT scanners with a total of 209 imaging conditions. An algorithm was developed to compute different imaging biomarkers. Variability across images from the same scanner and from different scanners was analyzed by computing coefficients of variation (CV) and standard deviations of HU values. LAA -950 and LAA -856 biomarkers had the highest levels of variability, while the majority of other biomarkers had variability less than 10 HU or 10% CV in both inter and intrascan measurements. There was no clear trend present between the biomarker measurements and CTDIvol. The results of this study demonstrates the existing variability in CT quantifications for lung imaging, which prompt further studies on how to reduce such variation.

Photon-counting CT (PCCT) is an emerging CT technology that uses photon-counting detectors (PCDs) to offer better spatial resolution, higher contrast, lower noise, and material-specific imaging as compared to conventional energy-integrating CT. To study the efficiency and performance of PCCT technologies in clinical use, virtual imaging trials (VITs) can be used. VITs use computational human phantoms to generate scanner-specific CT images. The integration of PCCT into VITs requires modeling the signal generation and signal processing in the detector and electronics, which includes incorporating the effects of nonidealities in PCDs such as crosstalk, charge sharing, and pulse pileup. These non-idealities adversely affect the image quality of PCCT systems, and their inclusion is important in accurate and realistic modeling of the PCDs. The existing scanner simulators model either charge sharing or pulse pileup but not their combined effects. The purpose of this study was to develop an experimentally validated modular detector response model that accounted for the combined effects of crosstalk, charge sharing, and pulse pileup in CdTe- and Si-based PCDs. It can be used to simulate variety of PCCT designs, including different detector materials and geometry, facilitating the evaluation and study of present and future PCCTs. The validation showed a close agreement with the experimental data acquired using Pixirad-1/Pixie-III PCDs. The platform was used to generate spatio-energetic covariance correlation matrices that integrated with a VIT framework called DukeSim to simulate scanner specific PCCT images.

X-ray Computed Tomography (CT) is an indispensable imaging modality in the medical field, notably in the diagnosis of traumatic brain injury and brain hemorrhage. Existing clinical CT systems, all of which use rotating x-ray sources, are too complex for practical deployment in resource-poor environments. A stationary array of sources and detectors with sufficient angular coverage and focal spot density could potentially simplify the image acquisition hardware and generate adequate projection data for CT reconstruction. In this study, we present a stationary head CT (s-HCT) prototype which combines projection data from three separate but parallel imaging planes into a complete volumetric iterative CT reconstruction. The fully operational scanner features three carbon nanotube (CNT) x-ray source arrays with 45 distributed focal spots each, and an Electronic Control System (ECS) for high speed control of the x-ray exposure from each focal spot. Projection data is acquired by translating the object along the z-axis at constant speed to expose all three imaging planes for the full dataset. As the object is moving at constant speed, 135 views are collected every 0.44 s (cycle time). The 3D iterative reconstruction designed specifically for the s-HCT configuration has been used to produce phantom images for initial assessment of the spatial resolution. A linear collimator was designed and constructed for the reduction of cross-plane scatter. Finally, the prototype has been able to acquire these images in scan times comparable to those of commercial scanners (<1min), indicating the CNT x-ray and s-HCT technologies are developed enough for clinical trials.

The temporal resolution of x-ray computed tomography (CT) is limited by the scanner rotation speed and detector readout time. One way to reduce the detector readout time is to acquire fewer number of projections. However, reconstruction using sparse-view data could result in spatial resolution loss and reconstruction artifacts that may negatively affect the clinical diagnoses. Therefore, improving the spatial resolution of sparse-view CT (SVCT) is of great practical value. In this study, we proposed a deep learning-based approach for SVCT spatial resolution enhancement. The proposed method utilizes a densely connected convolutional neural network (CNN) that is further aided by a radial location map to recover the radially dependent blurring caused by the continuous rotation of an x-ray source. The proposed method was evaluated using sparse-view data synthesized from full-view projection data of real patients. The results showed that the proposed CNN was able to recover the resolution loss and improve the image quality. Compared with the network using the same main structure but without a radial location map, the proposed method achieved better image quality in terms of the mean absolute error and structure similarity.

Many advanced reconstruction and image processing methods are being developed with the aim of improving image quality in CT. Development and testing of these methods is aided by the ability to simulate realistic images, to both control the acquisition process and be able to use digital phantoms with known ground truth. Therefore, in this work, we present a method to simulate realistic scanner-specific sinograms from digital phantoms. For this, a series of measurements was conducted on a clinical CT system to characterize resolution loss, noise characteristics, and the exposure-to-detector output relationship. These measurements were used to develop a simulation pipeline, which involves raytracing of a digital phantom, taking into account the focal spot size and gantry rotation, followed by the use of Lambert’s Law to determine the amount of energy arriving at each detector element. The spectrum for the specific tube voltage and current was modeled using previously published spectral models. The resulting sinogram was then corrupted, by applying the measured detector Modulation Transfer Function (MTF), and adding noise based on the Noise Power Spectrum (NPS) and mean-variance relationship. Simulator results were compared to those acquired with the CT system in our clinic, showing an average difference of 2.1% in the off-center MTF magnitude, 0.048 in normalized NPS magnitude and only 6 and 5 Hounsfield Units (HU) difference in the voxel values for respectively water and air. The developed simulator seems capable of generating realistic CT images, which can help researchers develop and test their algorithms.

Proton therapy requires accurate dose calculation for treatment planning to ensure conformal doses are precisely delivered to tumor target. This study proposes a fully connected neural network (FCNN) to figure out the underlying correlation between dual-energy computed tomography (DECT) parametric maps, material mass density and stopping power ratio (SPR) maps, which are essential for proton analytical and Monte Carlo dose calculations. A Siemens SOMATOM Definition Edge scanner (Siemens Medical Solutions, Germany) was used to acquire the DECT images with TwinBeam protocols. The effective atomic number maps, electron density maps, and virtual mono-energetic images were derived by Siemens Syngo.Via. The proposed FCNN includes 9 hidden layers, 200 hidden units, and nonlinear activation functions with layer normalization to prevent gradient vanishing issues of deep neural networks. The model was trained using multiple scanned data from a CIRS electron density phantom 062M with different inserts (Computerized Imaging Reference Systems, Inc., Norfolk, VA) and tested using a CIRS M701 anthropomorphic phantom. For the anthropomorphic phantom, the relative mean absolute errors of the density map were 0.45%, 0.77%, 0.23%, and 1.15% for lung, soft tissue, brain, and bone, while the SPR errors were 0.7%, 1.59%, 1.56%, and 1.23%, respectively. The results indicated that FCNN-based DECT parametric mapping generated robust and accurate mass density and SPR maps with minimum impacts by CT noise due to the usage of millions of training data. The proposed FCNN-based method potentially can improve proton range uncertainty by offering accurate material properties converted from DECT, especially for lung and bone.

CT imaging has proved to be an effective tool in evaluating COPD patients. The development of photon-counting detectors has offered better spatial resolution and reduced image noise, overcoming one major limitation of conventional scanners. However, this promising technology needs to be compared against conventional scanners and carefully optimized for its specific application. The purpose of this study was to investigate the potential benefits of PCCT in the context of COPD imaging. Our analysis demonstrated that PCCT achieved more accurate results compared to conventional scanners and that lower dose levels, sharp kernels, and thicker reconstruction slices would negatively affect the accuracy.

CT continues to be one of the most widely used medical imaging modalities. Concerns about long term effect of x-ray radiation on patients have led to efforts to reduce the x-ray dose imparted during CT exams. Lowering CT dose results in a lower signal to noise ratio in CT data which lowers CT Image Quality (IQ). Deep learning algorithms have shown competitive denoising results against the state-of-art image-based denoising approaches. Among these deep learning algorithms, deep residual networks have demonstrated effectiveness for edge-preserving noise reduction and imaging performance improvement compared to traditional edge-preserving filters. Previously published Residual Encoder- Decoder Convolutional Neural Network (RED-CNN) showed significant achievement for noise suppression, structural preservation, and lesion detection. However, its 2D architecture makes it unsuitable for thin slice and reformatted (sagittal, coronal) imaging. In this work, we present a novel 3D RED-CNN architecture, evaluate the effect of model parameters on performance and IQ, and show steps to improve optimization convergence. We use standard imaging metrics (SSIM, PSNR) to assess imaging performance and compare to previously published algorithms. Compared to 2D RED-CNN, our proposed 3D RED CNN produces higher quality 3D results, as shown by reformatted (sagittal, coronal) views, while maintaining all advantages of the original RED-CNN in axial imaging.

The Zoom-In Partial Scans (ZIPS) method is a recently introduced high-resolution CT technique that utilizes the high geometric magnification in off-center regions in the CT scanner’s field-of-view to boost the intrinsic spatial resolution of existing clinical multi-slice CT. ZIPS performs two off-center partial high-resolution scans of a region of interest (ROI), then an image reconstruction algorithm merges the partial scan data to produce a final high-resolution reconstructed image. In this study, we illustrate the feasibility of ZIPS image reconstruction with simultaneous estimation of inter-scan rigid ROI motion between the two partial scans. A total-variation and an entropy-based image alignment loss function was introduced for registration of the ROI between the two partial scans. Optional denoising filters were also introduced to stabilize the image alignment loss function. The feasibility of the ZIPS reconstruction framework is evaluated in a Catsim simulation environment. Results show that the proposed algorithmic compensation of ROI motion produced images visually indistinguishable from the images reconstructed with the ground truth positions of the ROI. The residual errors of the estimated ROI positions were no greater than 0.12 mm in x- and y- translation and 0.26 degree in rotation. Up to twofold improvement in the modulation transfer function (MTF) was achieved by ZIPS CT relative to a conventional centered scan.

Patient specific dose estimation is traditionally calculated though Monte-Carlo methods but not performed during CT image acquisition due to long computation times. We propose the implementation of a NN to perform patient specific dose estimation that can be performed alongside the CT acquisition due to the reduction in computation time. This is achieved by first performing MC simulations and then training the NN to replicate these predictions. The NN shows promise with a high degree of total cranial dose accuracy between the predictions and ground truth with a standard deviation of less than ±0.5 mGy.

Low-dose CT has been investigated and employed in various forms for clinical practices. One of the viable options for low-dose imaging is using a multi-slit beam collimator to achieve a sparse sampling. We have earlier demonstrated the feasibility of such technique, and extended the method for dual-energy imaging from a single scan using a multi-slit beam-filter in the circular CBCT system. Multi-detector-row helical CT is indeed in wide use in the clinics. In this work, we continue to explore the beam-filter based imaging technique in the multi-row helical CT scans. We conducted a simulation study by applying a virtual filter to the real data acquired from a helical CT scanner. We separated the sinograms of each row from the multi-row helical CT data, and removed the streaks using a notch filter in the Fourier domain of each sinogram. We then reconstructed the image by use of the filtered-backprojection algorithm and reduced the image noise by applying l0-norm based smoothing.

Sparse-view computed tomography (CT) has been attracting attention for its reduced radiation dose and scanning time. However, analytical image reconstruction methods such as filtered back-projection (FBP) suffer from streak artifacts due to sparse-view sampling. Because the streak artifacts are deterministic errors, we argue that the same artifacts can be reasonably estimated using a prior image (i.e., smooth image of the same patient) with known imaging system parameters. Based on this idea, we reconstruct an FBP image with sparse-view projection data, regenerate the streak artifacts by forward and back-projection of a prior image with sparse views, and then subtract them from the original FBP image. For the success of this approach, the prior image needs to be patient specific and easily obtained from given sparse-view projection data. Therefore, we introduce a new concept of implicit neural representations for modeling attenuation coefficients. In the implicit neural representations, neural networks output a patient-specific attenuation coefficient value for an input pixel coordinate. In this way, network’s parameters serve as an implicit representation of a CT image. Unlike conventional deep learning approaches that utilize a large, labeled dataset, an implicit neural representation is optimized using only sparse-view projection data of a single patient. This avoids having a bias toward a group of patients in the dataset and helps to capture unique characteristics of the individual properly. We validated the proposed method using fan-beam CT simulation data of an extended cardiac-torso phantom and compared the results with total variation-based iterative reconstruction and an image-based convolutional neural network.

Wide-coverage detector CT and ultra-high-resolution (UHR) detector CT are two important features for current cardiac imaging modalities. The former one enables the scanner to finish a whole heart image scan in one bed position; the latter one gives superior resolution in fine structures such as stenoses, calcifications, implemented stents, and small vessel boundaries. However, no commercially available scanner has both these features in one scanner as of today. Herein, we propose to use existing UHR-CT data to train a super resolution (SR) neural network and apply the network in a widecoverage detector CT system. The purpose of the network is to enhance the system resolution performance and reduce the noise while maintaining its wide-coverage feature without additional hardware changes. Thirteen UHR-CT patient datasets and their simulated-normal-resolution pairs were used for training a 3D residual-block U-Net. The modulation transfer function (MTF) measured from Catphan phantom scans showed the proposed super-resolution aided deep learning-based reconstruction (SR-DLR) improved the MTF resolution by relative ~30% and ~10% as compared to filtered-back projection and model-based iterative reconstruction approaches. In real patient cases, the SR-DLR images show better noise texture and enhanced spatial resolution along with better aortic valve, stent, calcification, and soft tissue features as compared to other reconstruction approaches.

Organ segmentation of medical images is a key step in virtual imaging trials. However, organ segmentation datasets are limited in in terms of quality (because labels cover only a few organs) and quantity (since case numbers are limited). In this study, we explored the tradeoffs between quality and quantity. Our goal is to create a unified approach for multi-organ segmentation of body CT, which will facilitate the creation of large numbers of accurate virtual phantoms. Initially, we compared two segmentation architectures, 3D-Unet and DenseVNet, which were trained using XCAT data that is fully labeled with 22 organs, and chose the 3D-Unet as the better performing model. We used the XCAT-trained model to generate pseudo-labels for the CT-ORG dataset that has only 7 organs segmented. We performed two experiments: First, we trained 3DUNet model on the XCAT dataset, representing quality data, and tested it on both XCAT and CT-ORG datasets. Second, we trained 3D-UNet after including the CT-ORG dataset into the training set to have more quantity. Performance improved for segmentation in the organs where we have true labels in both datasets and degraded when relying on pseudo-labels. When organs were labeled in both datasets, Exp-2 improved Average DSC in XCAT and CT-ORG by 1. This demonstrates that quality data is the key to improving the model’s performance.

As one of the most advanced CT imaging modalities, spectral CT plays important roles in generating materialspecific information and adding vital clinical values for disease diagnosis and therapy. To obtain the spectral CT images, currently, advanced X-ray source or detector assembly is often required, which significantly increases the hardware cost and the patient burden. As a consequence, the accessibility of spectral CT is strongly limited and has not been widely implemented in daily clinics. To solve such difficulty, this work attempts to investigate a new CT data acquisition protocol and spectral CT image reconstruction algorithm. In particular, the X-ray tube voltage is slowly modulated during the gantry rotation. By doing so, spectral information that varies from one projection view to another can be acquired in one single CT scan. Afterwards, a model-based material decomposition algorithm that reconstructs the CT image from the acquired projections is utilized to perform multi-material decomposition. To evaluate the performance of this novel spectral CT imaging approach, a numerical phantom containing iodine and gadolinium solutions is imaged with different kVp modulations, i.e., different number of modulation periods per rotation. Results demonstrate that the proposed spectral CT image reconstruction algorithm can be used to accurately decompose the water, iodine and gadolinium basis images for different tube voltage modulation rates. Moreover, high-quality monochromatic images can be synthesized as well. In conclusion, a low-cost multi-material spectral CT imaging approach is developed based on the slow tube voltage modulation method.

We proposed the optimization of a more generic X-ray tomography model. By imaging an object with a range of source spectra, we showed that spectral images could nevertheless be obtained, even if we have energy blind detectors. For observations to be integrated over energy levels, we tried to solve the problem by jointly estimating the X-ray measurement and X-ray absorption spectrum under the constraint. The research includes using DL (deep-learning) to learn a low-dimensional model of absorption spectra, using DL constraint to recover absorption spectra from X-ray measurements, and using DL to decompose absorption spectra estimates into material density estimates.

Photon-counting detectors (PCD) are the most recent advancement in computed tomography (CT). PCDs allow, among other things, for material decomposition, which decomposes the imaged object into a set of basis materials. Another field that is gaining attention, is the use of deep learning to improve the image reconstruction process in CT. In this work, we study the use of deep learning, specifically convolutional neural networks trained on the KiTS19 Challenge kidney data set, to improve the image quality of basis images resulting from three-material decomposition, a problem that is difficult due to its high sensitivity to noise. Our objective is to compare different network architectures and investigate whether these are best implemented in the projection domain or in the image domain. We study three different network architectures: U-Net, Dilated U-Net and ResNet, each applied in either the image domain or in the projection domain. The resulting image quality is evaluated in terms of contrast-to-noise ratio, task transfer function and noise power spectrum. Results show that for the type of phantoms the networks were trained on, the most effective option is to implement the network in the image domain and to use either the U-Net or Dilated U-Net architectures. However, when applying the networks to other phantoms, it seems that the networks in the sinogram generalize better, and produce better results. We also discuss why this might be the case, compare it with previous research, and consider what further improvements can be made.

In our previous work, we proposed to reduce motion artifacts in computed tomography (CT) using an image-based convolutional neural network (CNN). However, its motion compensation performance was limited when the degree of motion was large. We note that a fast scan mode can reduce the degree of motion but also cause streak artifacts due to sparse view sampling. In this study, we aim to initially reduce motion artifacts using a fast scan mode, and to reduce both the streak and motion artifacts using a CNN-based two-phase approach. In the first phase, we focus on reducing streak artifacts caused by sparse projection views. To effectively reduce streak artifacts in the presence of motion artifacts, a CNN with the U-net architecture and residual learning scheme was used. In the second phase, we focus on compensating motion artifacts in output image of the first phase. For this task, the attention blocks with global average pooling were used. To generate datasets, we used extended cardiac-torso phantoms and simulated sparse-view CT using half, quarter, and one-eighth of full projection views with corresponding 6-degree of freedom rigid motions. The results showed that the proposed two-phase approach effectively reduced both the motion and streak artifacts and taking fewer projection views down to one-eighth views (thus improving a scanning speed) provided the better image quality in our simulation study.

The objective of this research was to develop and validate a computed tomography image-based noise addition tool for producing simulated reduced dose CT images by adding statistical noise to the images. The method relies only on an input CT series. It works by first estimating the noise power spectrum (NPS) of the input images, creating synthetic forward projections, generating random white noise in the projection domain in proportion to synthetic projection values, backprojecting to the image, filtering by the NPS, and scaling according to the desired dose reduction level. This results in simulated noise with similar magnitude, texture, streak, and non-stationarity characteristics compared to true CT noise. The utility of the noise addition tool was evaluated using clinical images of 74 patients of various CT examinations including brain, renal, liver, and lung scans. These images were obtained at high and low doses and reconstructed by filtered back projection, iterative reconstruction (ASIR; GE Healthcare), model-based iterative (Veo; GE Healthcare) reconstruction algorithms. The simulated low dose images were compared to the actual low dose images in terms of noise magnitude, which showed good agreement between them for different scan types and reconstruction methods for a relative difference of 2.38%. The noise texture between the simulated and actual low dose images had similar visual appearance. The developed CT noise addition tool produces simulated reduced dose images with realistic noise properties and has the potential to simplify dose reduction studies for CT protocol optimization.

Several methods have been introduced over the years to measure CT image noise magnitudes in vivo applying different image segmentation strategies, HU thresholds, and regions of interests in which the noise is calculated. Therefore, it is impossible to directly compare results obtained with different methods and to objectively evaluate image quality in CT. This study provided an unbiased comparison of two methods to estimate noise utilizing a Virtual Imaging Trials platform simulating 1800 image datasets representing different dose levels, anatomical areas, and reconstruction techniques. Only methods that measure noise magnitude in soft tissue can effectively inform protocol design and optimization.

Background: Coronary computed tomography angiography (CCTA) allows non-invasive assessment of luminal stenosis and coronary atherosclerotic plaque. We aimed to develop and externally validate an artificial intelligence-based deep learning (DL) network for CCTA-based measures of plaque volume and stenosis severity. Methods: This was an international multicenter study of 1,183 patients undergoing CCTA at 11 sites. A novel DL convolutional neural network was trained to segment coronary plaque in 921 patients (5,045 lesions). The DL architecture consisted of a novel hierarchical convolutional long short-term memory (ConvLSTM) Network. The training set was further split temporally into training (80%) and internal validation (20%) datasets. Each coronary lesion was assessed in a 3D slab about the vessel centrelines. Following training and internal validation, the model was applied to an independent test set of 262 patients (1,469 lesions), which included an external validation cohort of 162 patients Results: In the test set, there was excellent agreement between DL and clinician expert reader measurements of total plaque volume (intraclass correlation coefficient [ICC] 0.964) and percent diameter stenosis (ICC 0.879; both p<0.001, see tables and figure). The average per-patient DL plaque analysis time was 5.7 seconds versus 25-30 minutes taken by experts. There was significantly higher overlap measured by the Dice coefficient (DC) for ConvLSTM compared to UNet (DC for vessel 0.94 vs 0.83, p<0.0001; DC for lumen and plaque 0.90 vs 0.83, p<0.0001) or DeepLabv3 (DC for vessel both 0.94; DC for lumen and plaque 0.89 vs 0.84, p<0.0001). Conclusions: A novel externally validated artificial intelligence-based network provides rapid measurements of plaque volume and stenosis severity from CCTA which agree closely with clinician expert readers.

Metals included inside an object (such as clips, bolts, or artificial joints) cause streaks and dark band artifacts in computed tomography. Although several metal artifact reduction (MAR) methods have been reported, they require a large amount of processing time such as for iterative forward projection from reconstructed image, but they do not provide sufficient correction depending on the metal and object and are sometimes accompanied by image degradation due to new artifacts. To overcome these problems, MAR methods using artificial intelligence (AI-MAR) with deep learning are reported. We have developed new AI-MAR to directly predict artifacts using deep learning and subtract them from the original filtered back-projection (FBP) image to maintain the structure of the object and achieve a high artifact reduction effect. The proposed AI-MAR was compared with FBP, linear interpolation method (LI), and conventional MAR. In a metal simulation experiment, the proposed AI-MAR successfully reduced the metal artifacts, and the structural similarity indices (SSIMs) evaluated with the FBP, LI, conventional MAR, and proposed AI-MAR were 0.941, 0.986, 0.969, and 0.988, respectively, and an improvement rate of SSIM of more than 80% was demonstrated. The proposed AI-MAR can improve device performance by providing high-speed and highly accurate images with reduced metal artifacts.

Material decomposition algorithms enable discrimination and quantification of multiple contrast agent and tissue compositions in spectral image datasets acquired by photon-counting computed tomography (PCCT). Image denoising has been shown to improve PCCT image reconstruction quality and feature recognition while preserving fine image detail. Reduction of image artifacts and noise could also improve the accuracy of material decomposition but the effects of denoising on material decomposition have not been investigated. In particular, deep learning methods can reduce inherent PCCT image noise without using a system-based or assumed prior noise model. Therefore, the objective of this study was to investigate the effects of image denoising on quantitative material decomposition in the absence of an influence of spatial resolution on feature recognition. Phantoms comprising multiple pure and spatially uniform contrast agent (gadolinium, iodine) and tissue (calcium, water) compositions were imaged by PCCT with four energy thresholds chosen to normalize photon counts and leverage contrast agent k-edges. Image denoising was performed by the established blockmatching and 3D-filtering (BM3D) algorithm or deep learning using convolutional neural networks. Material decomposition was performed on as-acquired, BM3D-denoised, and deep-learning-denoised datasets using constrained maximum likelihood estimation and compared to known material concentrations in the phantom. Image denoising by BM3D and deep learning improved the quantitative accuracy of material concentrations determined by material decomposition compared to ground truth, as measured by the root-mean-squared error. Material classification was not improved by image denoising compared with as-acquired images, suggesting that material decomposition was robust against inherent acquisition noise when feature recognition was not challenged by the system spatial resolution. Deeplearning-denoised images balanced preservation of local detail compared to more aggressive smoothing with BM3D, as measured by line profiles across features.

Compared with commercial dual-energy computed tomography (DECT) solutions, beam modulation is an economical and efficient alternative to achieve single-scan DECT which can be enabled with conventional diagnostic or cone-beam CT machines. Existing beam modulation methods employ regularization strategy for DECT reconstruction, which cannot handle the half-fan mode due to inevitable data loss in each spectrum. Besides, all previous works employ primarymodulation scheme, leading to magnified penumbra area in the detector side. In this work, a dedicated beam modulator is attached to the detector surface to realize dual-energy projection acquisition during a single scan. Then the detectormodulated sinogram is fed into the pre-trained pix2pix GAN model to generate full-sampled sinograms at high- and lowenergy spectra. Analytical reconstruction and material decomposition can be subsequently implemented using the complete dual-energy sinogram data.

Photon Counting Detector CT (PCD-CT) is expected to provide new possibilities by its capability of energy decomposition. However, because of non-ideal phenomena like the finite widths of individual energy windows of the detector or the scattered radiation, PCD-CT can have artifacts of not only dark but bright bands in some images. Therefore, we developed a correction method to reduce both dark and bright bands and verified it with data acquired with a prototype machine. The correction is performed on the material decomposition images, and acrylic and iodine images were employed. The dark band observed in the iodine image was corrected with the method used in the Energy Integrated Detector CT system. To suppress the bright band seen in the acrylic image, several virtual monochromatic images were created from the material decomposition images, and the energy was sought at which the dark and bright bands were cancelled. The bright band correction was performed using this energy and the correction amount of the dark band in the iodine image. It was demonstrated with the prototype machine data that the dark and bright bands were reduced in the material decomposition image and the virtual monochromatic image, and it was shown that the artifacts of both dark and bright bands can be corrected by the proposed method.

In helical scan protocol, direct filtered back projection method introduces helical artifacts in the reconstructed volume. Especially, the helical artifacts become severer when a large two dimensional planar detector array is employed. In this work, we proposed a method to reduce the helical artifacts in the reconstructed volume. The proposed algorithm proceeds in two sequential parts: bone induced artifacts reduction part and soft tissue induced artifacts reduction part. Since artifacts induced by bones are more severe, the bone image are segmented from the initial reconstructed image, and then the bone induced artifacts are generated by conducting the forward projection and filtered back projection on the segmented bone image. The difference between the segmented bone image and reconstructed bone image is calculated, which is subtracted from the initial reconstructed image. In the same manner, the soft tissue induced artifacts are generated from the bone induced artifacts corrected image, and the final correction image is acquired by subtracting the soft tissue induced artifacts from the bone induced artifacts corrected image. The proposed algorithm was verified for extended cardiac-torso (XCAT) simulation. Performance of the proposed algorithm was quantitatively evaluated for uncorrected images, bone corrected images and bone-and-soft tissue corrected images by normalized mean square error (NMSE) and structure similarity index (SSIM). The results showed that the bone artifacts reduction part effectively reduced the helical artifacts, and soft tissue artifacts reduction part further reduced residual artifacts remained in the bone corrected images. The proposed algorithm can reduce helical artifacts without compromising the existing FBP’s computational advantages.

In this work, we applied the singular value decomposition (SVD) method to a set of monochromatic images to extract the dominant physical contributions to image formation. We showed that the first two principal components can be related to an arbitrary pair of basis material in mathematically enclosed expression. The later principal components are assumed to carry mostly sub-leading image formation effects, noise, and reconstruction artifact contribution. The proof of concept is shown on numerical (linear) images and later confirmed on physical spectral CT phantom images obtained with monochromatic x-ray radiation at Elettra synchrotron in Trieste, Italy. Following material decomposition, we also performed a quantitative description of tissue-equivalent phantom materials in terms of material density and effective atomic number.

CT perfusion imaging (CTP) plays an important role in decision making for the treatment of acute ischemic stroke with large vessel occlusion. Since the CT perfusion scan time is approximately one minute, the patient is exposed to a non-negligible dose of ionizing radiation. However, further dose reduction increases the level of noise in the data and the resulting perfusion maps. We present a method for reducing noise in perfusion data based on dimension reduction of time attenuation curves. For dimension reduction, we use either the fit of the first five terms of the trigonometric polynomial or the first five terms of the SVD decomposition of the time attenuation profiles. CTP data from four patients with large vessel occlusion and three control subjects were studied. To compare the noise level in the perfusion maps, we use the wavelet estimation of the noise standard deviation implemented in the scikit-image package. We show that both methods significantly reduce noise in the data while preserving important information about the perfusion deficits. These methods can be used to further reduce the dose in CT perfusion protocols or in perfusion studies using C-arm CT, which are burdened by high noise levels.

The circular Radon transform (CRT) is widely employed as an imaging model for wave-based tomographic bioimaging modalities like ultrasound reflectivity tomography. A complete set of CRT data function is known to have redundancies. However, no explicit non-iterative image reconstruction method is known for inverting temporally-truncated data. To address this, a learning-based approach is proposed to establish a filtered backprojection (FBP) method for use with the half-time CRT data function. The proposed method approximates a mapping that is known to exist in theory; therefore, it is fundamentally different than many deep-learning based reconstruction methods that seek to establish a non-existent mapping. Thus, the proposed method performs well on unforeseen data. The learned half-time FBP achieves image quality comparable to a conventional full-time FBP method although it uses half of the complete data.

Novel deep learning (DL) methods have produced state-of-the-art results in nearly every area of x-ray CT data processing. However, DL-driven, iterative reconstruction remains challenging because of the volumetric nature of many reconstruction problems: the system matrix relating the projection and image domains is too large to incorporate into network training. Past approaches for 2D reconstruction include consecutive projection and image domain processing in a single pass, employing a known analytical operator between domains, and unrolling an established iterative method, solving a series of sub-problems with separately trained networks. Here, we synergize these approaches within the split Bregman optimization framework. Specifically, we formulate a cost function and supervised training approach which yield an analytical reconstruction sub-step, to transform between domains, and a regularization sub-step, which is consistent between iterations. Combined with projection and image domain splitting, these properties reduce the number of free parameters which must be learned, making volumetric, dual-domain data processing more practical. Here, we simultaneously train 3D image and projection domain regularizers using supervised learning during iterative reconstruction with promising algorithm convergence results. Our trained reconstruction framework outperforms a more traditional iterative reconstruction method when starting from 90 noisy projections of the MOBY mouse phantom (image SSIM: iterative, 0.65; DL, 0.86). Furthermore, we successfully apply the model to similarly sampled in vivo mouse data acquired with micro-CT, reducing noise from 277 HU in an initial reconstruction to 41 HU in the DL reconstruction compared with 63 HU of noise in a fully sampled reference reconstruction.

In spite of its extensive adaptation in almost every medical diagnostic and examinatorial application, Magnetic Resonance Imaging (MRI) is still a slow imaging modality which limits its use for dynamic imaging. In recent years, Parallel Imaging (PI) and Compressed Sensing (CS) have been utilised to accelerate the MRI acquisition. In clinical settings, subsampling the k-space measurements during scanning time using Cartesian trajectories, such as rectilinear sampling, is currently the most conventional CS approach applied which however, is prone to producing aliased reconstructions. With the advent of the involvement of Deep Learning (DL) in accelerating the MRI, reconstructing faithful images from subsampled data became increasingly promising. Retrospectively applying a subsampling mask onto the k-space data is a way of simulating the accelerated acquisition of kspace data in real clinical setting. In this paper we compare and provide a review for the effect of applying either rectilinear or radial retrospective subsampling on the quality of the reconstructions outputted by trained deep neural networks. With the same choice of hyper-parameters we train and evaluate two distinct Recurrent Inference Machines (RIMs), one for each type of subsampling. The qualitative and quantitative results of our experiments indicate that the model trained on data with radial subsampling attains higher performance and learns to estimate reconstructions with higher fidelity paving the way for other DL approaches to involve radial subsampling.

In cone-beam computed tomography (CBCT) imaging, a scanning configuration with an offset-detector is often used for extending the field of view (FOV) of the system. Due to the truncation of data at certain views, data are required to be collected over a full angular range (FAR) of 360◦ for accurate reconstruction by use of existing analytical-based algorithms. However, there exist interests in practical applications for limited-angular-range (LAR) imaging because it may allow for the reduction of radiation dose and scanning time and for the avoidance of the collisions between the moving gantry and scanned objects. Under such imaging conditions, existing algorithms generally yield reconstructions with significant artifacts. In this work, we develop and investigate a directional-total-variation (DTV) algorithm for image reconstruction from partially truncated data collected over LARs. By using the DTV algorithm, we have performed numerical simulation studies with partially truncated data collected from a pelvic phantom over different LARs with an offset-detector CBCT system. The results of the numerical studies demonstrate that the proposed algorithm can yield, from partially truncated LAR data, images with significantly reduced artifacts that are observed otherwise in images obtained with existing analytical-based algorithms.

In computed tomography (CT) imaging, recent developments in reconstruction algorithm and scan configuration design have provided useful tools for image reconstruction from data collected over a limited-angular range (LAR). In this work, we aim to investigate the impact of angular sampling interval on the accuracy of reconstruction from LAR data. In specific, we employ a two-orthogonal-arc scan configuration, and collect data from a numerical chest phantom over an LAR with various angular intervals. We then investigate image reconstruction by using the directional-total-variation (DTV) algorithm and evaluate reconstructions qualitatively and quantitatively. Results show that increased angular sampling interval can degrade image quality. Results of the simulation study also indicate an appropriate interval for sufficient reconstruction accuracy under specific imaging conditions, which provides insights for upper-bound performance of reconstructions in practical use.

This work applies a back-projection approach to the reconstruction of images obtained with a 3D near-field coded aperture camera, a method derived from that used to analyse astronomical data collected by the INTEGRAL/IBIS instrument.1 The method used in this paper is a form of deconvolution that updates the image as the camera or the object under observation is moving, and as such could be applied to dynamic studies or as a real-time surgical probe. Multi-isotope parathyroid imaging has been identified as one of the optimal applications of this technique, where high sensitivity and high energy resolution would permit the differentiation between thyroid and parathyroid tissues on the surgical table, and aid in different types of thyroid surgery. The back-projection technique is combined with a 3D version of the iterative CLEAN algorithm, which reduces the effect of the systematic noise intrinsic to imaging with a non-perfect aperture.

Inherent to Computed tomography (CT) is image reconstruction, constructing 3D voxel values from noisy projection data. Unlike deriving forward operators, modeling the inverse operation remains challenging. Given the ill-posed nature of the inverse problem, data-driven methods still need additional regularization to enhance the accuracy of the results. Besides, the generalization of the results hinges upon the availability of training data and access to the ground truth. This paper offers a new strategy to reconstruct CT images with the advantage of ground truth accessible through a virtual imaging trial (VIT) platform. A learned primal-dual deep neural network (LPD-DNN) employed the forward model and its adjoint as a surrogate of the imaging’s geometry and physics. VIT offered paired projection and ground truth data from anthropomorphic human models without noise and resolution degradation. The models included a library of anthropomorphic, computational (XCAT) patient models. The DukeSim simulator was utilized to form realistic projection data emulating the impact of the physics and geometry of a commercial CT scanner (Flash, Siemens). The resultant noisy sinogram data associated with each slice was thus generated for training. Corresponding linear attenuation coefficients of phantoms’ materials at the x-ray spectrum effective energy were used as the labels. The LPD-DNN was then deployed, learning the complex operators and hyper-parameters in the proximal primal-dual optimization. The obtained validation results showed 12% normalized root mean square error with respect to the ground truth labels, peak signal-to-noise ratio of 32 dB, a signal-to-noise ratio of 1.5, and a structural similarity index of 96%. These results were highly favorable compared to standard filtered-back projection reconstruction (65%, 17 dB, 1.0, 26%).

Dual-energy CT (DECT) with limited-angular-range (LAR) data is of interest, as it could potentially reduce radiation dose and scanning time and avoid collision of the moving gantry with the imaged subject. In DECT with LAR data, images suffer from LAR and beam-hardening (BH) artifacts. In this work, we investigate the simultaneous correction of LAR and BH artifacts for DECT with LAR data. Under a scanning configuration with overlapping arcs of low- and high-kVp spectra, data are generated from a digital suitcase phantom. A data-domain decomposition method is used to correct for the BH artifacts first, while basis images are reconstructed from the decomposed basis sinograms of LAR by use of the previously developed directional-total-variation (DTV) algorithm to correct for the LAR artifacts. Visual inspection of the monochromatic images and quantitative analysis of estimated atomic numbers suggest that the simultaneous correction of BH and LAR artifacts in DECT can effectively reduce, and almost eliminate, BH and LAR artifacts in monochromatic images from data of LAR as low as 30◦ , and also yield accurately estimated atomic numbers that are almost numerically identical to the reference values from the full-angular-range data of 360° .

Principles of cone-beam sampling are considered in terms of an analytical figure of merit for data incompleteness [tan(ψ_𝑚𝑖𝑛)] and experimental measurements of cone-beam artifact magnitude in an image quality phantom. The phantom incorporates disk-pairs oriented parallel to the axial plane (perpendicular to the z-axis) at various locations throughout the field of view. Experiments were conducted in which the phantom was imaged using a mobile C-arm with various, tilted and untilted source-detector trajectories (orbits). In the current work, scan orbits were limited to circular orbits – viz., a semicircular arc spanning ~196°. The magnitude of cone-beam artifacts was assessed in terms of the signal modulation (z_mod) in the narrow region between disk-pairs, and the relationship of z_mod and tan(ψ_𝑚𝑖𝑛) was evaluated for each orbit. The results show visually and quantitatively the effect of scan orbit and object orientation on cone-beam sampling effects and demonstrate the relationship between analytical tan(Ψ_𝑚𝑖𝑛) and empirical z_mod figures of merit. The analysis suggests a means by which cone-beam sampling effects (incompleteness) can be quantifiably considered in the development of new cone-beam CT system geometries, multi-source configurations, and protocols involving novel source-detector trajectories.

Single-photon emission computed tomography (SPECT) performed with a pinhole collimator often suffers from parallax error due to depth-of-interaction uncertainty. One possible way to reduce the parallax error for a new generation of SPECT pinhole cameras would be to incorporate fiber optics to control the spread of light and improve 3D position estimation. In this work, we have developed a Monte Carlo simulation for an SiPMbased modular gamma camera that incorporates a fiber optic-plate as a light guide. We have created a custom photon transport code written in Swift and we perform the computationally taxing components on a GPU using Metal. This code includes refraction according to Snell’s law as well as reflection according to Fresnel’s laws at material boundaries. The plate is modeled as a hexagonally-packed array of individual fibers. We also include the scintillation statistics of NaI(Tl) and the detection efficiency of the silicon photomultipliers. We use the simulation code to create mean-detector-response functions (MDRFs) from which Fisher information on event positioning can be assessed. We compare planar detectors with different light guides to determine the effects of the fiber optics. We model three geometries; one that only uses a monolithic light guide, one that only has a fiber-optic plate, and one that has a monolithic light guide and a fiber-optic plate in combination. The spatial resolutions are compared by using Fisher Information Matrices to calculate the Cram´er-Rao Lower Bounds on position estimate variances.

Breathing-related movement analysis is important in the study of many disease processes. The analysis of diaphragmatic motion via thoracic imaging in particular is important in a variety of disorders. Compared to computed tomography (CT) and fluoroscopy, dynamic magnetic resonance imaging (dMRI) has several advantages, such as better soft tissue contrast, no ionizing radiation, and greater flexibility in selecting scanning planes. In this paper, we propose a novel method for full diaphragmatic motion analysis via free-breathing dMRI. Firstly, after 4D dMRI image construction in a cohort of 51 normal children, we manually delineated the diaphragm on sagittal plane dMRI images at end-inspiration and end-expiration. Then, 25 points were selected uniformly and homologously on each hemi-diaphragm surface. Based on the inferior-superior displacements of these 25 points between end-expiration (EE) and end-inspiration (EI) time points, we obtained their velocities. We then summarized 13 parameters from these velocities for each hemi-diaphragm to provide a quantitative regional analysis of diaphragmatic motion. We observed that the regional velocities of the right hemi-diaphragm were almost always statistically significantly greater than those of the left hemi-diaphragm in homologous locations. There was a significant difference for sagittal curvatures but not for coronal curvatures between the two hemi-diaphragms. Using this methodology, future larger scale prospective studies may be considered to confirm our findings in the normal state and to quantitatively assess regional diaphragmatic dysfunction when various disease conditions are present.

Image reconstruction is a key task in various applications related to medical scanning, geological prospecting, etc. Usually measured data for reconstruction is limited due to physical constraints in these practices. Consequently image reconstruction with limited data becomes an underdetermined inverse problem. Recently, using un-trained Deep Neural Networks (DNN) such as Deep Decoder and ConvDecoder has been proved as a plausible approach for image reconstruction. By leveraging the DNN structure itself as a prior, the un-trained neural networks have shown great advantages over the traditional reconstruction methods like total-variation (TV) in many aspects. Comparing with the stateof-art trained neural networks like U-net, the un-trained neural networks also alleviate the cost of robustness and scalability in out-of-distribution samples. In this work, we propose a new un-trained image reconstruction neural network called EffiDecoder. Our contribution is two-fold. First we introduce mobile inverted bottleneck (MBconv) at each layer with computational-friendly depthwise separable convolutions. Second, the feature channels are adaptively reduced according to the output. In accelerated MRI reconstruction tasks the experimental results demonstrate that the proposed method can achieve competitive performance with reduced computational cost. Therefore, with the notable improvements of EffiDecoder, we believe the un-trained neural networks have great potentials to be discovered in extensive reconstruction applications.

Full-waveform modelling serves as the basis for many emerging inversion techniques within ultrasound computed tomography. Being able to accurately depict strong material interfaces, such between soft tissue and bone, is particularly important for ensuring that these numerical methods produce physically correct results. We present a procedure for constructing digital twins of various parts of the human body through the use of conforming hexahedral meshes, which are used together with the spectral-element method to accurately model the interactions of the ultrasound wavefield at these sharp material boundaries. In silico cranial and knee phantoms are used as examples.

This paper presents new imaging systems for the estimation of physiological random processes in medical imaging. In this work, a physiological random process is a sequence of biochemical interactions taking place inside a living organism. These interactions involve things such as proteins and enzymes, that behave differently in response to external stimuli (such as nutrients or administered drugs). Understanding how these physiological processes interact and evolve is critical in the development of effective therapeutic approaches. The general setup of our imaging systems includes a fast detector for the measurement of visible light from which to estimate various parameters about the radiation emitted by the physiological process(es) of interest. Our setup is applicable to imaging with different kinds of radiation, including gamma rays (SPECT and PET), and charged particles, such as alpha and beta particles. Parameters we are interested in estimating for these photons/particles go beyond the 2D or 3D position typically measured in medical imaging applications, and include the direction of propagation and photon/particle energy. Recent work has shown the advantage of measuring direction of propagation and photon/particle energy, in addition to just position. It has been shown that if these additional photon/particle parameters are taken into account during reconstruction, the null space of the imaging system is strongly reduced or eliminated. This reduction in null space is critical to adequately characterize complicated physiological processes.

Among different techniques developed for x-ray phase-contrast imaging (XPCi) coded-aperture and Talbot-Lau can potentially provide higher contrast and higher sensitivity in imaging soft and hard tissues. Absorption grating is the central component that makes these techniques different from other XPCi techniques. For hard xray imaging, compatible x-ray absorption gratings are a must so x-ray can be stopped at these absorption gratings to create the desired wave-fronts. For hard x-ray imaging or soft x-ray imaging with high-resolution detectors, the aspect-ratio of gratings feature size goes beyond 10:1 (height:width) – in other words thickness of absorbing part increases – which is technologically challenging through conventional UV lithography techniques. Although SU-8 photoresist along with x-ray lithography is an alternative to fabricate such a high-aspect ratio grating, a compatible UV lithograph-based fabrication process is more desirable as it reduces the cost and increases the throughput. In this work, we have broken down a high-aspect ratio x-ray grating design into multiple layers of lower-aspect ratio structures and employed only the conventional UV lithography. SU-8 photoresists are known for their multi-layer coating specification, which is used in this study. The new fabrication process proposed in this work results in a final high-aspect ratio x-ray absorption grating through accessible UV lithography with lower cost and scaling-up compatibility, thus every research group and industry can benefit from it. To the best of our knowledge, this is the first time a multi-layer x-ray grating design is proposed and reported. The output of this work can be used to perform large field-of-view high-energy coded-aperture x-ray phase-contrast imaging.

High-aspect ratio x-ray absorption gratings are required for hard x-ray or high-resolution soft x-ray phasecontrast imaging (XPCi) applications. X-ray gratings play a central role in grating-based XPCi techniques, where visibility – image quality – is highly dependent on the quality of gratings. Although the fabrication process of x-ray gratings for soft x-ray imaging is well-developed, there are technological challenges in highresolution hard x-ray imaging that impede us from making high-aspect ratio large field-of-view fine x-ray gratings through conventional fabrication processes. Here we present a design and a fabrication process to fabricate highaspect ratio gratings that benefit from a self-aligned hard-mask – a patterned chromium-gold-chromium thin film deposited on a transparent ITO-on-glass substrate – which facilitates both lithography and electroplating processes. The repeatability of the proposed method makes it suitable for achieving high-aspect ratio fine structures as thick as desired through a multi-layer structure without any restraint or limitation on the aspect ratio of features. The multi-layer structure design overcomes the aspect ratio limitation associated with UVlithography. Although this method works well with X-ray lithography, the key advantage of this design is that it enables UV lithography for high-aspect ratio grating fabrication through a reliable yet simple process. To the best of our knowledge, self-aligned multi-layer SU-8 based grating design has not been previously reported. The proposed design and fabrication process help researchers further develop x-ray gratings performance to facilitate high-resolution coded-aperture and Talbot-Lau high-energy x-ray phase-contrast imaging.

We have previously shown in simulations that X-ray Interferometry using Modulated Phase Gratings can create an interference pattern in clinical detectors from which attenuation, differential phase, and dark-field contrast images can be formed. This interferometric technique is advantageous since it eliminates the need to use an absorption grating as compared to Talbot-Lau grating systems, providing better dose-efficiency. In this work we experimentally evaluated this modulated phase grating system using initial test gratings obtained from Microworks GmbH, Germany. Experiments with the MPG gratings were conducted at the monochromatic 8 keV beamline at the LSU Center for Advanced Microstructures and Devices (CAMD). After analyzing our fringe pattern and eliminating the effect of source grating (G0), we observed stable fringe patterns for our MPG system at different source grating (G0) to MPG distances. The fringe pattern results from these experiments that indicate the feasibility and potential of an X-ray MPG system that could be functional with only a single-phase grating (and source grating) as opposed to a standard interferometry system that additionally requires an absorption grating near the detector.

X-ray absorption gratings are the heart of coded-aperture and Talbot-Lau x-ray phase-contrast imaging (XPCi) techniques. The quality of imaging – visibility – in the aforementioned techniques is highly dependent on the quality of gratings. SU-8, an epoxy-based photoresist, is valued in Micro-Electro-Mechanical-Systems (MEMS) fabrication due to its excellent mechanical and optical properties. For fabricating x-ray absorption grating following the MEMS process, gold is the material most widely used to stop incident x-rays. One of the main challenges in fabricating a quality x-ray absorption grating is the adhesion between SU-8 molds and gold thin film, where due to poor adhesion of SU-8 to gold thin films and SU-8’s internal high level of stress, SU-8 molds delaminate, crack, or peel off during developing or electroplating steps. This problem becomes more challenging when a greater field of view gratings are required. Employing SU-8 on a gold substrate requires an adhesion layer to assure SU-8’s mechanical stability during fabrication steps. Without this layer, realizing a high-quality grating is challenging. In this work, we have investigated the adhesion quality of SU-8 to gold thin films for different adhesion promoter layers. We have employed a combination of a SU-8 thin film and a metallic-silane-based nanometer-thin film to improve the adhesion quality between SU-8 photoresist and a gold thin film substrate. We have tested the quality of adhesion using tape test and through a harsh (highly acidic) electroplating solution.

A phase contrast system with a modulated phase grating (MPG) eliminates the need for an analyzer when compared to a standard Talbot Lau X-ray interferometer. This can provide three scans with the same total dose to the object as a standard mammogram. In this work, a hybrid MPG phase contrast system was investigated where the system fringe period can be varied, by changing a single system parameter, allowing interrogation at different resolutions and scatter lengths to further the variety of scans possible with the phase contrast system. The system can be also used for screening in a default setting.

Purpose: This study aims at establishing the optimum x-ray energy for synchrotron acquired propagation-based computed tomography (PB-CT) images to obtain highest radiological image quality of breast mastectomy samples. It also examines the correlation between objective physical measures of image quality with subjective human observer scores to model factors impacting visual determinants of image quality. Approach: Thirty mastectomy samples were scanned at Australian Synchrotron’s Imaging and Medical Beamline. Samples were scanned at energies of 26, 28, 30, 32, 34, and 60 keV at a standard dose of 4mGy. Objective physical measures of image quality were assessed using signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), SNR/resolution (SNR/res), CNR/resolution (CNR/res) and visibility. Additional calculations for each measure were performed against reference absorption-based computer tomography (AB-CT) images scanned at 32 keV and 4mGy. This included differences in SNR (dSNR), CNR (dCNR), SNR/res (dSNR/res), CNR/res (dCNR/res), and visibility (dVis). Physical measures of image quality were also compared with visual grading analysis data to determine a correlation between observer scores and objective metrics. Results: For dSNR, dCNR, dSNR/res, dCNR/res, and dVis, a statistically significant difference was found between the energy levels. The peak x-ray energy for dSNR and dSNR/res was 60 keV. For dCNR and dCNR/res 34 keV produced the highest measure compared to 28 keV for dVis. Visibility and CNR correlate to 56.8% of observer scores. Conclusion: The optimal x-ray energy differs for different objective measures of image quality with 30-34 keV providing optimum image quality for breast PB-CT. Visibility and CNR correlate highest to medical imaging expert scores.

The purpose of this study is to evaluate and compare the quantitative imaging performance of the dual-energy CT (DECT) and differential phase contrast CT (DPCT). The electron density (ρ_e) and the effective atomic number (Z_eff) are selected as the two comparison bases for the DECT and DPCT imaging. From the numerically simulated data, image domain based decomposition algorithms are used to extract the ρ_e and Zeff information for three different spatial resolution levels (0.3 mm, 0.1 mm, and 0.03 mm). The contrast-to-noise-ratio (CNR) and modeled human observer studies have been investigated to compare the DECT and DPCT quantitative imaging performance. At low spatial resolution (0.3 mm), the DECT shows better quantitative imaging performance than DPCT. As a contrary, the DPCT outperforms the DECT for ultra high spatial resolution (0.03 mm) imaging. With the 0.1 mm spatial resolution, the DECT and DPCT shows similar quantitative imaging performance. In conclusion, the DECT is more favored for low spatial resolution applications, such as the diagnostic imaging tasks. However, the DPCT would be recommended for ultra high spatial resolution imaging tasks, such as the micro-CT imaging tasks.

Quantification of the strength and quality of the muscle, adipose tissue, and bone is important for the characterization of the effects of normal aging, age-related diseases, metabolic disorders and neuromuscular diseases. The cost, duration and risks of medical imaging trials may render the task of generating a sufficient high number of quality data for the training of systems for clinical trials and evaluations impracticable. In this work, we developed a model of the human mid-thigh with structural representation of its muscles, adipose tissue and bone anatomy, for use in virtual clinical studies. This is a simulation-based approach to optimizing medical imaging systems. We simulated thigh phantoms based on the OpenVCT software framework, originally designed for digital breast imaging studies. We designed and generated phantoms of mid-thigh anatomical structures representing normal anatomy. Exploiting the flexibility of the system, we were able to generate a controlled population for which we varied, separately, different anatomical structures. We simulated regional mid-thigh muscle area degeneration that is frequently observed in diabetes patients and quantified the structural changes relatively to a healthy anatomy. This framework also allows us to simulate variations of anatomical structures – hence serving as a system for advanced data augmentation that may be used for training machine learning-based diagnostic methods, simulating the effects of diseases, and designing clinical studies.

Traditional methods of quantitative analysis of CT images typically involve working with patient data, which is often expensive and limited in terms of ground truth. To counter these restrictions, quantitative assessments can instead be made through Virtual Imaging Trials (VITs) which simulate the CT imaging process. This study sought to validate DukeSim (a scanner-specific CT simulator) utilizing clinically relevant biomarkers for a customized anthropomorphic chest phantom. The physical phantom was imaged utilizing two commercial CT scanners (Siemens Somatom Force and Definition Flash) with varying imaging parameters. A computational version of the phantom was simulated utilizing DukeSim for each corresponding real acquisition. Biomarkers were computed and compared between the real and virtually acquired CT images to assess the validity of DukeSim. The simulated images closely matched the real images both qualitatively and quantitatively, with the average biomarker percent difference of 3.84% (range 0.19% to 18.27%). Results showed that DukeSim is reasonably well validated across various patient imaging conditions and scanners, which indicates the utility of DukeSim for further VIT studies where real patient data may not be feasible.

Virtual clinical trials (VCTs) have been used widely to evaluate digital breast tomosynthesis (DBT) systems. VCTs require realistic simulations of the breast anatomy (phantoms) to characterize lesions and to estimate risk of masking cancers. This study introduces the use of Perlin-based phantoms to optimize the acquisition geometry of a novel DBT prototype. These phantoms were developed using a GPU implementation of a novel library called Perlin-CuPy. The breast anatomy is simulated using 3D models under mammography cranio-caudal compression. In total, 240 phantoms were created using compressed breast thickness, chest-wall to nipple distance, and skin thickness that varied in a {[35, 75], [59, 130), [1.0, 2.0]} mm interval, respectively. DBT projections and reconstructions of the phantoms were simulated using two acquisition geometries of our DBT prototype. The performance of both acquisition geometries was compared using breast volume segmentations of the Perlin phantoms. Results show that breast volume estimates are improved with the introduction of posterior-anterior motion of the x-ray source in DBT acquisitions. The breast volume is overestimated in DBT, varying substantially with the acquisition geometry; segmentation errors are more evident for thicker and larger breasts. These results provide additional evidence and suggest that custom acquisition geometries can improve the performance and accuracy in DBT. Perlin phantoms help to identify limitations in acquisition geometries and to optimize the performance of the DBT prototypes.

3D models are frequently used for augmenting gross anatomy education, constructing 3D prints that are used in myriad applications, collecting data in biomedical research, and improving patient outcomes from surgery. The rapidity with which 3D modeling has been developed has resulted in a wide range of applications, but a dearth of information concerning optimal methodological practices and the influence of data processing variations on the results of computational analyses. One of the most important subjects that still needs to be examined in detail is the influence of the decimation process on 3D model morphology. Decimation (i.e., mesh complexity simplification) is a very common data processing step used to obtain more computationally efficient 3D models. Recent studies have revealed that requisite 3D mesh acquisition and processing practices can dramatically influence the quantitative data collected from 3D meshes.1, 2 However, the exact impact of the decimation process on quantitative morphological data is not well-understood. The current investigation evaluates the hypothesis that visually comparable 3D meshes of disparate mesh complexities may actually yield significantly different morphological data.