This PDF file contains the front matter associated with SPIE Proceedings Volume 11595, including the Title Page, Copyright Information, and Table of Contents.

This Conference Presentation, “Radiomics: transforming standard imaging into mineable data for diagnostic and theragnostic applications,” was recorded for the Medical Imaging 2021 Digital Forum.

Welcome and Introduction to SPIE Medical Imaging conference 11595: Physics of Medical Imaging

The accuracy and variability of quantifications in Computed Tomography Angiography (CTA) are affected by imaging parameters and patient attributes. While patient attributes cannot usually be altered for a scan, imaging parameters can be optimized to improve the accuracy and precision of the procedure. This study developed a mathematical approach to find the optimal controllable parameters that maximize an ideal estimator, namely the Estimability index (e′), for quantifying coronary stenosis in cardiac CTA. We applied one-hot encoding to the categorical features and normalized the numerical features within 0 to 1. We applied a ridge regression model to the transformed data with polynomial feature transforms with degrees of 1 to 5. A grid search identified the polynomial model with the highest accuracy to predict the e′ value. We then evaluated the influence of each parameter and its permutation on the accuracy of the model. We formulated the corresponding optimization problem as maximization of e′ where the decision parameters are subject to linear constraints defining the upper bound and lower bound of each decision variable. We mathematically calculated the deterministic and probabilistic optimal controllable parameters across a range of deterministic and probabilistic uncontrollable parameters. The results showed that a reduced noise (less than 17 HU) and sharper MTFf50 (greater than 0.45 𝑚𝑚−1) maximizes e′. Moreover, the cardiac motion velocity had a higher impact on the deviation of the optimal decision variables compared to percent stenosis, vessel radius, plaque material, and lumen contrast.

In previous works, we have presented concepts for dynamic beam attenuators (DBAs) allowing for substantial dose reductions. So far, we have used a tube current modulation (TCM) scheme where the tube current is proportional to the square root of the maximum object attenuation in each projection. As DBAs show particularly beneficial behavior in region-of-interest (ROI) imaging, the question arose whether the employed TCM method would still be the most meaningful in this case. We present a simulation framework calculating i) the dose distribution in the object and ii) the image quality in the ROI of the reconstructed images for a given primary fluence: The dose to every voxel was calculated from individual Monte- Carlo simulations for every beamlet. From this we can approximate the effective dose, that incorporates tissue-specific weighting factors describing the stochastic health risk, for any fluence distribution. Using the same fluence distribution, the image variance according to the propagated fluence can be calculated for a given ROI. We employed a homogeneous phantom with a centered ROI and a female thorax phantom where the ROI is described by either the heart or the spine. Eventually, we optimized the tube current according to the product of mean image variance in the ROI and patient dose. Furthermore, the tube current obtained was compared with a heuristic square root TCM (hsqTCM) method. In result, the optimized TCM matches the hsqTCM well for the idealized case of a central ROI in an elliptical, homogeneous phantom. For a more complex case with an off-center ROI, the optimized TCM differs substantially from the hsqTCM rule, offering an additional dose reduction of up to 30%.

Tube current modulation (TCM) is routinely implemented in clinical CT imaging. By modifying the x-ray tube current as a function of patient attenuation, image quality can be made more consistent and radiation dose can be better managed. Optimal TCM settings depend on the scan protocol and the physical characteristics of the patient. This study was undertaken to develop a realistic TCM model and integrate it into a scanner-specific CT simulation platform, allowing for faithful emulation of CT scans with TCM and optimization of TCM methods. The developed model adjusts the mAs for each projection based on attenuation estimated from two localizers, a strength parameter similar to the implementation of one major CT manufacturer (Siemens CARE Dose 4D), and imposed tube power limits. To demonstrate the utility of this framework, a virtual imaging trial was conducted to characterize image quality as a function of TCM strength. Three XCAT phantoms (BMIs at 25th, 50th, and 75th percentile) were imaged at five TCM strengths, using a CT configuration based on the geometry and physics of a commercial scanner (Siemens Definition Flash). Noise magnitude was measured in each reconstructed CT slice, average noise magnitude was calculated in lungs and bones, and RMSE, PSNR, SSIM, and contrast were computed in the lungs. The TCM strength which minimized noise increased with phantom BMI, suggesting that optimal TCM settings will differ for patients of different sizes. This study demonstrates the first incorporation of TCM into a scanner-specific virtual imaging trial platform and its application to patient-specific optimization.

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, the use of CHO in clinical CT is still quite limited, mainly due to its complexity in measurement and calculation in practice, and the lack of access to an efficient and validated software tool for most clinical users. In this work, a web-based software platform for CT image quality assessment and protocol optimization (CTPro) was introduced. A validated CHO tool, along with other common image quality assessment tools, was made readily accessible through this web platform for clinical users and researchers without the need of installing additional software. An example of its application to evaluation of convolutional-neural-network (CNN)-based denoising was demonstrated.

Perfusion CT imaging is commonly used for the rapid assessment of patients presenting with symptoms of acute stroke. Maps of perfusion parameters such as cerebral blood volume (CBV), cerebral blood flow (CBF), and mean transit time (MTT) derived from the scan data provide crucial information for stroke diagnosis and treatment decisions. Most vendors implement singular value decomposition (SVD)-based methods on their scanners to calculate these parameters. However, SVD-based method is known to have issues of improperly handling the imperfect scan. For example, increasing the acquisition interval or decreasing the scan duration may introduce a bias in the estimated perfusion parameters. In this work, we propose a Bayesian inference algorithm, which can tolerate the imperfect scan conditions better than conventional method and is able to derive the uncertainty of a given perfusion parameter. We apply the variational technique to the inference problem, which becomes an expectation-maximization problem. The probability distribution (with Gaussian mean and variance) of each estimated parameter can be obtained. We perform evaluations in simulation studies both with full and incomplete data. The proposed method can obtain much less bias in estimation than the conventional method, and additionally providing the degree of the uncertainty in measurement.

While radiomics models are finding increased use in computer-aided diagnostics and as imaging biomarkers for inference and discovery, their utility in computed tomography (CT) is limited by variability of the image properties produced by different CT scanners, imaging protocols, patient anatomy, and an increasingly diverse range of reconstruction and post-processing software. While these effects can be mitigated with careful data curation and standardization of protocols, this is impractical for diverse sources of image data. In this work, we propose to generalize traditional end-to-end imaging system models to include radiomics calculation as an explicit stage. Such a model permits both prediction of the undesirable variability of radiomics, but also forms a basis for inverting the process to estimate the true underlying radiomics. This framework has the potential to provide for standardization of radiomics across imaging conditions permitting more widespread application of radiomics models; larger, more diverse image databases; and improved diagnoses and inferences based on those standardized metrics. We apply this framework to a large class of popular radiomics based on the Gray Level Co-occurrence matrix under conditions of imaging system that are well describe by traditional linear systems approaches as well as nonlinear systems for which traditional analytic models do not apply.

This work reports the development of a C-arm photon counting detectors (PCD)-CT system for evaluating the potential clinical utility of PCD-CT for intraoperative imaging. A dual-threshold CdTe-based strip PCD was mounted on a Siemens Artis Zee C-arm interventional system. A new geometric calibration method was developed to correct for the geometric distortion of the C-arm system during rotation. Experimental results show that, under clinically relevant conditions, the C-arm PCD-CT system can reliably and reproducibly generate high quality MDCT-like images without any noticeable geometric distortion or banding artifact.

Owing to poor characterization of implant and adjacent human tissues, the presence of metal implants has been shown to be a risk factor for clinical results for proton therapy. In this project we have developed a way of characterizing implant and human materials in terms of water-equivalent thicknesses (WET) and relative stopping power (RSP) using a novel proton counting detector. We tracked each proton using a fast spectral imaging camera AdvaPIX-TPX3 which operated in energy mode measures collected energy per-voxel to derive the deposited energy along the particle track across the voxelated sensor. We considered three scenarios: sampling of WET of a CIRS M701 Adult Phantom (CMAP) at different locations; measurements of energy perturbations in the CMAP implanted with metal rods; sampling of WET of a more complex spine phantom. WET and RSP information were extracted from energy spectra at position along the central axis by using the shift in the most probable energy (MPE) from the reference energy (either initial incident energy or energy without a metal implant). Measurements were compared to TOPAS simulation results. Measured WET of the CMAP ranged from 18.63 to 25.23 cm depending on the location of the sampling which agreed with TOPAS simulation results within 1.6%. The RSPs of metals from CMAP perturbation measurements were determined as 1.97, 2.98, and 5.44 for Al, Ti and CoCr, respectively, which agreed with TOPAS within 2.3%. RSPs for material composition of a more complex spine phantom yielded 1.096, 1.309 and 1.001 for Acrylic, PEEK and PVC, respectively. In summary, this work has shown a method to accurately characterize RSPs of metal and human materials of CMAP implanted with metals and a complex spine phantom. Using the data obtained by the proposed method, it may be possible to validate RSP maps provided by conventional photon computed tomography techniques. Owing to poor characterization of implant and adjacent human tissues, the presence of metal implants has been shown to be a risk factor for clinical results for proton therapy. In this project we have developed a way of characterizing implant and human materials in terms of water-equivalent thicknesses (WET) and relative stopping power (RSP) using a novel proton counting detector. We tracked each proton using a fast spectral imaging camera AdvaPIX-TPX3 which operated in energy mode measures collected energy per-voxel to derive the deposited energy along the particle track across the voxelated sensor. We considered three scenarios: sampling of WET of a CIRS M701 Adult Phantom (CMAP) at different locations; measurements of energy perturbations in the CMAP implanted with metal rods; sampling of WET of a more complex spine phantom. WET and RSP information were extracted from energy spectra at position along the central axis by using the shift in the most probable energy (MPE) from the reference energy (either initial incident energy or energy without a metal implant). Measurements were compared to TOPAS simulation results. Measured WET of the CMAP ranged from 18.63 to 25.23 cm depending on the location of the sampling which agreed with TOPAS simulation results within 1.6%. The RSPs of metals from CMAP perturbation measurements were determined as 1.97, 2.98, and 5.44 for Al, Ti and CoCr, respectively, which agreed with TOPAS within 2.3%. RSPs for material composition of a more complex spine phantom yielded 1.096, 1.309 and 1.001 for Acrylic, PEEK and PVC, respectively. In summary, this work has shown a method to accurately characterize RSPs of metal and human materials of CMAP implanted with metals and a complex spine phantom. Using the data obtained by the proposed method, it may be possible to validate RSP maps provided by conventional photon computed tomography techniques.

Compton interactions amount to a significant fraction of the registered counts in a silicon detector. In a Compton interaction, only a part of the photon energy is deposited and a single incident photon can result in multiple counts unless tungsten shielding is used. Silicon has proved to be a competitive material for photon-counting CT detectors but to improve the performance further, it is desirable to use coincidence techniques to identify Compton scattered photons and reconstruct their energies. In a detector with no tungsten shielding, incident photons can interact through a series of interactions. By using information about the position and energy of each interaction, probability-based methods can be used to estimate the incident photon energy. In this work we present a framework of likelihood functions that can be used to estimate the incident photon energy in a silicon detector. For a low count-rate case with one incident photon per time frame, we show that the proposed likelihood framework can estimate the incident photon energy with a mean error of -0.26 keV and an RMS error of 1.14 keV for an ideal case with a perfect detector. For a non-ideal case including spatial and energy resolution, the corresponding results were -0.08 keV and 0.66 keV, respectively. The fraction of correctly identified interaction chains, with respect to the interaction order, were 99.2% in the ideal case and 97.3% in the non-ideal case.

Computed tomography (CT) using photon-counting detectors (PCD) offers dose-efficient ultra-high-resolution imaging, high iodine contrast-to-noise ratio, multi-energy and material decomposition capabilities. We have previously demonstrated the potential benefits of PCD-CT using phantoms, cadavers, and human studies on a prototype PCD-CT system. This system, however, had several limitations in terms of scan field-of-view (FOV) and longitudinal coverage. Recently, a full FOV (50 cm) PCD-CT system with wider longitudinal coverage and higher spatial resolution (0.15 mm detector pixels) has been installed in our lab capable of human scanning at clinical dose and dose rate. In this work, we share our initial experience of the new PCD-CT system and compare its performance with a state-of-the-art 3rd generation dual-source CT scanner. Basic image quality was assessed using an ACR CT accreditation phantom, high-resolution performance using an anthropomorphic head phantom, and multi-energy and material decomposition performance using a multi-energy CT phantom containing various concentrations of iodine and hydroxyapatite. Finally, we demonstrate the feasibility of high-resolution, full FOV PCD-CT imaging for improved delineation of anatomical and pathological features in a patient with pulmonary nodules.

The purpose of this study was to evaluate the performance of mammography images acquired with a prototype Gallium Arsenide (GaAs) photon counting detector. The contrast to noise ratio was measured using different aluminum/PMMA phantoms of different thicknesses. In addition, microcalcification detection accuracy was evaluated using a receiver operating characteristic (ROC) study with three observers reading an ensemble of images for each case. For both studies, comparisons were made to a commercial mammography system. The contrast to noise ratio was estimated by imaging 18, 36 and 110 μm thick aluminum targets placed on top of 6 cm of PMMA plates and was found to be similar or better over a range of exposures. Similar task performance in detecting microcalcifactions was observed between the systems over a range of clinically applicable dose levels with a small increase in GaAs’s system at lower dose levels. The GaAs system was evaluated using a typical mammography X-ray spectrum provided by the Automatic Exposure Settings (AEC) of the commercial mammography system and using only one energy threshold (i.e, one energy window). Operating the GaAs detector with multiple energy windows (i.e., two energy windows) may provide improved performance for a given dose.

A peculiar edge enhancement effect was observed when a PCD was operated under anti-coincidence mode and irradiated by high-flux x-rays to cause pulse pileups. The severity of this effect increases with the input flux level and completely disappears when the anti-coincidence mode is turned off. A theoretical analysis shows that this edge enhancement effect was jointly caused by the pulse pileup-induced count loss and the arbitration process used the anti-coincidence logic. Compared with pixels blocked by an edge, pixels immediately outside the edge has a much higher likelihood to win the arbitration when there are coincident x-rays.

A novel physics-based data augmentation (PBDA) is introduced, to provide a representative approach to introducing variance during training of a deep-learning model. Compared to traditional geometric-based data augmentation (GBDA), we hypothesize that PBDA will provide more realistic variation representative of potential imaging conditions that may be seen beyond the initial training data, and thereby train a more robust model (particularly in the scope of medical imaging). PBDA is tested in the context of false-positive reduction in nodule detection in low-dose lung CT and is shown to exhibit superior performance and robustness across a wide range of imaging conditions.

Most of the Deep Neural Networks (DNNs) based CT image denoising literature shows that DNNs outperform traditional iterative methods in terms of metrics such as the RMSE, the PSNR and the SSIM. In many instances, using the same metrics, the DNN results from low-dose inputs are also shown to be comparable to their high-dose counterparts. However, these metrics do not reveal if the DNN results preserve the visibility of subtle lesions or if they alter the CT image properties such as the noise texture. Accordingly, in this work, we seek to examine the image quality of the DNN results from a holistic viewpoint for low-dose CT image denoising. First, we build a library of advanced DNN denoising architectures. This library is comprised of denoising architectures such as the DnCNN, U-Net, Red-Net, GAN, etc. Next, each network is modeled, as well as trained, such that it yields its best performance in terms of the PSNR and SSIM. As such, data inputs (e.g. training patch-size, reconstruction kernel) and numeric-optimizer inputs (e.g. minibatch size, learning rate, loss function) are accordingly tuned. Finally, outputs from thus trained networks are further subjected to a series of CT bench testing metrics such as the contrast-dependent MTF, the NPS and the HU accuracy. These metrics are employed to perform a more nuanced study of the resolution of the DNN outputs' low-contrast features, their noise textures, and their CT number accuracy to better understand the impact each DNN algorithm has on these underlying attributes of image quality.

The radiation risk of X-ray CT gained increasing concern in the past decades. Lowering CT scan dose leads to noisy raw data as well as streak artifacts after reconstruction. Extensive studies have been conducted to reduce noise and artifacts for low-dose CT (LDCT). As deep learning has achieved great success in computer vision tasks, it also become a powerful tool in LDCT denoising. Commonly used deep learning methods such as supervised learning and generative adversarial learning have a strong dependence on large normal-dose CT (NDCT) dataset. While in real cases, the NDCT dataset is often expensive or not accessible, which limits the implementation of deep learning. In recent studies, multiple deep learning methods have been proposed for LDCT denoising without NDCT data. Among them, a popular type of methods is noisy label training (NLT) which use LDCT data as labels for network supervised training. Noise2Void is an easily implementable and representative method of NLT and has achieved great results in pixel-independent noise denoising. Another type is distribution learning methods which reduce LDCT noise-level by learning NDCT distribution. Deep distribution learning from noisy samples (DDLN) learns the NDCT distribution from LDCT data only and adopts MAP estimation for LDCT denoising with the learned NDCT distribution prior. It is effective for LDCT projection data denoising. In this work, the two representative methods are compared for LDCT projection data denoising under different noise-levels to seek for their suitable application scenarios.

Deep learning (DL) approaches to image denoising have advanced real-world performance in one of the most thoroughly studied areas of digital signal processing. Keys to the success of DL are data-driven approximation of an underlying signal model and non-linear domain mapping. By contrast, “classic” denoising approaches rely on prior assumptions about the structure of data and noise and generally penalize deviations through convex optimization. Despite these apparent differences, denoising with CNNs can be viewed as an extension of older sparse coding models, and even state-of-the-art DL models commonly employ cost terms like mean-squared-error and total variation. In this work, we adapt the split Bregman optimization method (SBM) for use with CNN-based regularizers. We show how the operator splitting of the SBM lends itself to splitting cost terms for CNN-based denoising, simplifying the selection of regularization hyperparameters and the hybridization of classic and DL-specific regularizers (bilateral TV, patch rank, discriminator loss). Furthermore, we demonstrate a simple, data-driven method to rebalance cost terms during network training, while limiting the potential to “invent” features. Using preclinical photon-counting micro-CT data of the mouse heart, our proposed framework performs 2D denoising of the time (PSNR, 27.7; SSIM, 0.62) and energy (PSNR, 35.0; SSIM, 0.81) dimensions which compares favorably with both 3D, multi-channel iterative reconstruction (reference) and the BM3D denoising algorithm (time: PSNR, 26.6; SSIM, 0.58. energy: PSNR, 27.2; SSIM, 0.76). The time and energy networks evaluate in under 2 minutes each vs. ~1-2 hours for our iterative reconstructions using similar hardware.

A wide range of dose reduction strategies for x-ray computed tomography (CT) have been investigated. Recently, denoising strategies based on machine learning have been widely applied, often with impressive results, and breaking free from traditional noise-resolution trade-offs. However, since typical machine learning strategies provide a single denoised image volume, there is no user-tunable control of a particular trade-off between noise reduction and image properties (biases) of the denoised image. This is in contrast to traditional filtering and model-based processing that permits tuning of parameters for a level of noise control appropriate for the specific diagnostic task. In this work, we propose a novel neural network that includes a spatial-resolution parameter as additional input permits explicit control of the noise-bias trade-off. Preliminary results show the ability to control image properties through such parameterization as well as the possibility to tune such parameters for increased detectability in task-based evaluation.

X-ray scatter is a major limit for good CT image quality. Apart from using hardware approach (e.g. anti-scatter grid), computational algorithms based on Monte-Carlo simulation or convolution kernels have been proven to be valid for compensating scatter effect. However, computational algorithms always have to take care about the balance between complexity and efficiency, so the performance has some limitation when scatter contribution is large. In this paper we proposed a deep learning based approach by adopting a convolutional neuro-network (CNN) to predict the scatter distribution on projection domain. The performance of the CNN-based model is validated in both projection domain as well as reconstructed image domain. The result shows that the scatter correction algorithm with learning approach is able to compensate the artifact from scatter radiations under various complicated scenarios, resulting in equivalent or even better image quality than commercially used kernel-based scatter correction algorithm.

Staff-dose management in fluoroscopic procedures is a continuing concern due to insufficient awareness of radiation dose levels. To maintain dose as low as reasonably achievable (ALARA), we have developed a software system capable of monitoring the procedure room scattered radiation and the dose to staff members in real-time during fluoroscopic procedures. The scattered-radiation display system (SDS) acquires imaging-system signal inputs to update technique and geometric parameters used to provide a color-coded mapping of room scatter. We have calculated a discrete look-up-table (LUT) of scatter distributions using Monte-Carlo (MC) software and developed an interpolation technique for the multiple parameters known to alter the spatial shape of the distribution. However, the file size for the LUT’s can be large (~2GB), leading to long SDS installation times in the clinic. Instead, this work investigated the speed and accuracy of a regressional neural network (RNN) that we developed for predicting the scatter distribution from imaging-system inputs without the need for the LUT and interpolation. This method greatly reduces installation time while maintaining real-time performance. Results using error maps derived from the structural similarity index indicate high visual accuracy of predicted matrices when compared to the MC-calculated distributions. Dose error is also acceptable with a matrix element-averaged percent error of 31%. This dose-monitoring system for staff members can lead to improved radiation safety due to immediate visual feedback of high-dose regions in the room during the procedure as well as enhanced reporting of individual doses post-procedure.

COPD is the fourth leading cause of death in the United States. The structural and functional attributes of this disease can be assessed in vivo using computed tomography (CT). The value of quantitative CT has been demonstrated towards characterization and treatment evaluation of COPD. Although promising, the influence of imaging protocols on the accuracy and precision of these quantifications remains unknown. The purpose of this study was to build a novel and realistic virtual imaging platform and investigate the effects of CT imaging parameters on COPD quantifications. Ten COPD human models were created on the platform of the state-of-the-art XCAT human models. The shape, size, and the distribution of emphysema along with its material density were modeled separately in each XCAT phantom, based on the CT images of confirmed COPD patients. The developed phantoms were imaged using an established scanner-specific CT simulator (DukeSim) at various dose levels with and without tube current modulation. The projection images were then reconstructed using six different reconstructions from a commercial reconstruction toolbox. Established COPD imaging biomarkers were extracted from the simulated images and compared against their corresponding digital ground truth values. The relative bias ranged from -2.1% to 66.0% and the variability ranged from 0.9% to 29.5%. The results showed that with careful choice of smooth reconstruction kernel, CT can be used as a reliable quantification tool for the COPD, with higher mAs, fixed tube currents, and iterative reconstructions further enhancing the accurate and consistent quantification of the disease. This study demonstrates the first development of a suite of computational COPD phantoms, enabling virtual imaging trials in the context of COPD imaging.

Realistic lesion generation is a useful tool for system evaluation and optimization. Generated lesions can serve as realistic imaging tasks for task-base image quality assessment, as well as targets in virtual clinical trials. In this work, we investigate a data-driven approach for categorical lung lesion synthesis using public lung CT databases. We propose a generative adversarial network with a Wasserstein discriminator and gradient penalty to stabilize training. We further included conditional inputs such that the network can generate user-specified lesion categories. Novel to our network, we directly incorporated radiomic features in an intermediate supervision step to encourage similar textures between generated and real lesions. We calculated the network using lung lesions from the Lung Image Database Consortium (LIDC) database. Lesions are divided into two categories: solid vs. non-solid. We performed quantitative evaluation of network performance based on four criteria: 1) overfitting, in terms of structural and morphological similarity to the training data, 2) diversity of generated lesions, in terms of similarity to other generated data, 3) similarity to real lesions, in terms of distribution of example radiomics features, and 4) conditional consistency, in terms of classification accuracy using a classifier trained on the training lesions. We imposed a quantitative threshold for similarity based on visual inspection. The percentage of non-solid and solid lesions that satisfy low overfitting and high diversity is 87.1% and 70.2% of non-solid and solid lesions respectively. The distribution of example radiomics features are similar in the generated and real lesions indicated by low Kullback–Leibler divergence scores: 1.62 for non-solid lesions and 1.13 for solid lesions. Classification accuracy for the generated lesions are comparable with that for the real lesions. The proposed network presents a promising approach for generating realistic lesions with clinically relevant features crucial for the comprehensive assessment of novel medical imaging systems.

Aim: To develop, validate and apply a pipeline for breast cancer voxel model generation from patient digital breast tomosynthesis (DBT) cases for cancer type specific virtual clinical trials (VCT). Methods: Input cancer cases were retrieved from wide-angle DBT systems. Three aspects of the creation process were investigated: (1) The impact of the limited z-resolution of DBT on the shape of the voxel model using circularity measurements (i.e. ratio of diameters between input and result after simulation test), DICE coefficient and artefact spread function. (2) The possibility to speed up and automate lesion segmentation with a deep learning network. (3) The ultimate realism of the voxel models in a VCT application, visually scored by a radiologist and a medical physicist. Results: Deviations between ground truth and segmented voxel models due to the pseudo-3D characteristics of DBT were limited, with circularity changes smaller than 8%. A 4-layer U-net deep learning network with a multiplication of the DICE loss and the implemented boundary loss as loss function is capable to produce segmentations within the variability of manual segmentations (DICE coefficient = 0.80). A reader study of the VCT application showed an average realism score of 3.4 on a scale of 1 to 5 for the simulated lesion manually segmented, compared to an average of 4.3 for the real lesions. An initial total of 25 invasive cancer models (9 non-spiculated, 16 spiculated masses) was successfully created and validated. Conclusion: Segmentation from an object with limited z-resolution induces an acceptable deformation. Voxel models created from DBT images can be used to mimic realistic DBT cancer cases. The use of AI techniques has facilitated the cumbersome manual segmentation task.

Malignant breast tumours can be distinguished from benign lesions and normal tissue based on their mechanical properties. Our pilot studies have demonstrated the potential of using Mechanical Imaging (MI) combined with mammography to reduce recalls and false positives in breast cancer screening by more accurately identifying benign lesions. To enable further optimization of MI we propose a computer simulation of the MI acquisition, for use in a Virtual Clinical Trial (VCT) framework. VCTs are computer simulated clinical trials used to efficiently evaluate clinical imaging systems. A linear elastic finite element (FE) model of the breast under dynamic compression was implemented using an open-source FE solver. A spherical tumour (15 mm in diameter) was inserted into the simulated predominantly adipose breast. The location and stiffness of the tumour was varied. The average stress on the compressed breast surface was calculated and compared with the local average stress at the tumour location and the Relative Mean Pressure over lesion Area (RMPA) was calculated. Preliminary results were within a realistic range with an average stress on the breast (tumour) of 5.9-16.6 kPa which is in agreement with published values between 1.0 – 22.5 kPa. This corresponds to RMPA values of 0.96-2.15 depending on stiffness and location of the tumour. This can lead to more detailed validation of various MI acquisition schemes through VCTs before their use in clinical studies.

A 3D printed heart chamber phantom was developed to work combined with other commercially available phantom kit. Based on CT image combined with traditional phantom and anatomic structures, the 3D model was generated and input for printing with selected materials. The 3D printed phantom could realize multi-dimensional motion and deformation similarity, improved HU behavior as contrast enhanced tissue mimic, biological closed anthropomorphic structure including cardiac chambers and coronary arteries with contrast agents, as well as inserts for anatomic or functional abnormalities simulation. With those properties the proposed 3D printed phantom could potentially be used for either CCTA imaging performance or CCTA scan strategy verifications combined with scanner and patient properties.

Accurate and objective image quality assessment is essential for the task of radiation dose optimization in clinical CT. Standard method relies on multi-reader multi-case (MRMC) studies in which radiologists are tasked to interpret diagnostic image quality of many carefully-collected positive and negative cases. The efficiency of such MRMC studies is frequently challenged by the lengthy and expensive procedure of case collection and the establishment of clinical reference standard. To address this challenge, multiple methods of virtual clinical trial to synthesize patient cases at different conditions have been proposed. Projection-domain lesion- / noise-insertion methods require the access to patient raw data and vendor-specific proprietary tools which are frequently not accessible to most users. The conventional image-domain noise-insertion methods are often challenged by the over-simplified lesion models and CT system models which may not represent conditions in real scans. In this work, we developed deep-learning lesion and noise insertion techniques that can synthesize chest CT images at different dose levels with and without lung nodules using existing patient cases. The proposed method involved a nodule-insertion convolutional neural network (CNN) and a noise-insertion CNN. Both CNNs demonstrated comparable quality to our previously-validated projection domain lesion- / noise-insertion techniques: mean structural similarity index (SSIM) of inserted nodules 0.94 (routine dose), and mean percent noise difference ~5% (50% of routine dose). The proposed deep-learning techniques for chest CT virtual clinical trial operate directly on image domain, which is more widely applicable than projection-domain techniques.

Digital Subtraction Angiography (DSA) is considered the gold standard for imaging and guiding treatment of neurovascular lesions, such as cerebral aneurysms and carotid stenoses. Though DSA can show high-resolution morphology, it remains difficult to extract temporal physiological information, because higher frame-rates are necessary to accurately quantify neurovascular flow details. Recent advances in photon-counting detector technology have led us to develop High-Speed Angiography (HSA), where X-ray images are acquired at 1000 fps for more accurate visualization and quantification of blood flow. Blood flow was imaged using HSA under constant flow conditions within various 3D printed patient-specific phantoms. Blood velocity was quantified using an open source Optical Flow algorithm, OpenOpticalFlow, to perform velocity estimation based on the spatio-temporal intensity changes of iodinated contrast wavefronts. The results of these algorithms are then compared with Computational Fluid Dynamics (CFD) simulations, using the same inlet boundary conditions and model geometries. The performance of these algorithms at lower temporal resolution was then also assessed by simulating lower frame rates from the acquired 1000 fps data. It is important to ascertain the hemodynamic effect of abnormal neurovascular conditions, as well as their effect on treatment of such conditions during the actual clinical interventional procedure. While theoretical CFD results requiring considerable computer capability are delayed for hours or more, it is expected that clinical results from multiple HSA sequences will be available almost immediately while the patient is still under treatment, and even right after flow conditions are changed beneficially by the intervention.

Endovascular procedures performed in the angio suite have gained considerable popularity for treatment of ischemic stroke as well as aneurysms. However, new intracranial hemorrhage (ICH) may develop during these procedures, and it is highly desirable to arm the angio suite with real-time and reliable ICH monitoring tools. Currently, angio suites are equipped with scintillator-based flat panel detector (FPD) imaging systems for both planar and cone beam CT (CBCT) imaging applications. However, the reliability of CBCT for ICH imaging is insufficient due to its poor low-contrast detectability compared with MDCT and lack of spectral imaging capability for differentiating between ICH, calcifications, and iodine staining from periprocedural contrast-enhanced imaging sequences. To preserve the benefits of the FPD for 2D imaging and certain high-contrast 3D imaging tasks while adding a high quality, quantitative, and affordable CT imaging capability to the angio room for intraoperative ICH monitoring, a hybrid detector system was developed that includes the existing FPD on the C-arm gantry and a strip photon-counting detector (PCD) that can be translated into the field-of-view for high quality PCD-CT imaging at a given brain section-of-interest. The hybrid system maintains the openness and ease of use of the C-arm system without the need to remodel the angio room and without installing a slidinggantry MDCT (aka Angio CT) with orders of magnitude higher costs. Additionally, the cost of the strip PCD is much less than the cost of a large-area PCD. To demonstrate the feasibility and potential benefits of the hybrid PCD-FPD system, a series of physical phantom studies, and human cadaver studies were performed at a gantry rotation speed (7 s) and radiation dose level that closely match those of clinical CBCT acquisitions. The experimental images of C-arm PCD-CT demonstrated MDCT-equivalent low-contrast detectability of PCD-CT and significantly reduced artifacts compared with FPD-based CBCT.

Purpose: Medical management after endovascular treatment (EVT) for acute ischemic stroke involves blood pressure management and antithrombotic medications which are strongly influenced by hemorrhagic complications of the procedure. In this study our aim is to modify a commercially available c-arm angiographic system to assess feasibility of distinguishing between blood and iodinated contrast using dual-energy CT technology. Methods: An angiographic C-arm system was manipulated with the addition of a flat sheet of tin (Sn) filtration with corresponding calibration. Two flat-panel cone-beam CT (FP-CBCT) were acquired at 70 kV and 125Sn kV. A multienergy CT phantom with iodine inserts at various concentrations ranging from 0.2 to 15 mg/cc was used to characterize the behavior (i.e. dual-energy ratios DER) of the materials as a function of the kVp pair. DER values were needed to adjust an existing material decomposition software. The accuracy of iodine quantification was assessed with inserts including solid water, iodine, blood and mixed iodine and blood; all with concentrations ranges relevant for stroke imaging. Results: Imaging indicated the ability to differentiate between blood and iodine for the use with FP-CBCT. Pure blood and iodine inserts fell within the 95% confidence interval for precision of contrast concentration measured. For iodine inserts, the error from expected measurement was 7.5% or under for inserts with concentrations above 2 mg/cc. Inserts containing a combination of blood and iodine were consistently reading 1 mg/ml or less deviations relative to the value of the material specified by the manufacturer, which represented a 25-35% error difference from the expected value. Conclusion: These results establish the reproducibility of the phantom values for dual energy calculations and suggest that this technology may be clinically useful after EVT for stroke. Improvements in the accuracy of the iodine/blood combination under dual energy evaluation is needed for other applications outside of the use for differentiating between hemorrhage and contrast staining.

Purpose: Cone-beam CT is a common approach for 3D guidance in interventional radiology (IR), but its long image acquisition time poses a limit to image quality due to complex deformable motion of soft-tissue structures. Multi-region autofocus optimization showed successful compensation of deformable motion with gains in image quality. However, the use of sub-optimal conventional autofocus metrics, accompanied by the high-dimensionality and non-convexity of the optimization problem result in challenged convergence. This work presents a learning-based approach to the design of autofocus metrics tailored to the specific problem of motion compensation in soft-tissue CBCT. Methods: A deep convolutional neural network (CNN) was used to automatically extract image features quantifying the local motion amplitude and principal direction in volumetric regions of interest (128 x 128x 128 voxels) of a motion contaminated volume. The estimated motion amplitude is then used as the core component of the cost function of the deformable autofocus method, complemented by a regularization term favoring similar motion for regions close in space. The network consists of a siamese arrangement of three branches acting on the three central orthogonal planes of the volumetric ROI. The network was trained with simulated data generated from publicly available CT datasets, including deformable motion fields from which volumetric ROIs with locally rigid motion were extracted. The performance of motion amplitude estimation and of the final CNN-based deformable autofocus were assessed on synthetic CBCT data generated similarly to the training dataset and featuring deformable motion fields with 1 to 5 components with random direction and random amplitude ranging from 0 mm to 50 mm. Results: Predicted local motion amplitude showed good agreement with the true values, showing a liner relationship (R2 = 0.96, slope 0.95), and slight underestimation of the motion amplitude. Absolute errors in total motion amplitude and individual components remained < 2 mm throughout the explored range. Relative errors were higher for low amplitude motion, pointing to need of a larger training cohort, more focused on low motion amplitude scenarios. Motion compensation with the learning-based metric showed successful compensation of motion artifacts in single ROI environments, with 40% reduction in median RMSE using the motion static image as reference. Deformable motion compensation resulted in better visualization of soft-tissue structures, and overall sharper image details, with slight residual motion artifacts for regions combining moderate motion amplitude with complex anatomical structures. Conclusion: Deformable motion compensation with automatically learned autofocus metrics was proved feasible, opening the way to the design of metrics with potential for more reliable performance and easier optimization than conventional autofocus metrics.

Purpose: Intraoperative cone-beam CT (CBCT) plays an important role in neurosurgical guidance but is conventionally limited to high-contrast bone visualization. This work reports a high-fidelity artifacts correction pipeline to advance image quality beyond conventional limits and achieve soft-tissue contrast resolution even in the presence of multiple metal objects – specifically, a stereotactic head frame. Methods: A new metal artifact reduction (MAR) method was developed based on a convolutional neural network (CNN) that simultaneously estimates metal-induced bias and metal path length in the projection domain. To improve generalizability of the network, a physics-based method was developed to generate highly accurate simulated, metalcontaminated projection training data. The MAR method was integrated with previously proposed artifacts correction methods (lag, glare, scatter, and beam-hardening) to form a high-fidelity artifacts correction pipeline. The proposed methods were tested using an intraoperative CBCT system (O-arm, Medtronic) emulating a realistic setup in stereotactic neurosurgery, including nominal (20 cm) and extended (40 cm) field of view (FOV) protocols. Results: The physics-based data generation method provided accurate simulation of metal in projection data, including scatter, polyenergetic, quantum noise, and electronic noise effects. The artifacts correction pipeline was able to accommodate both 20 cm and 40 cm FOV protocols and demonstrated ~80% improvement in image uniformity and ~20% increase in contrast-to-noise ratio (CNR). Fully corrected images in the smaller FOV mode exhibited ~32% increase in CNR compared to the 40 cm FOV mode, showing the method’s ability to handle truncated metal objects outside the FOV. Conclusion: The image quality of intraoperative CBCT was greatly improved with the proposed artifacts correction pipeline, with clear improvement in soft-tissue contrast resolution (e.g., cerebral ventricles) even in the presence of a complex metal stereotactic frame. Such capability gives clearer visualization of structures of interest for intracranial neurosurgery, and it provides an important basis for future work aiming to deformably register preoperative MRI to intraoperative CBCT. Ongoing work includes clinical studies now underway.

Many interventional procedures aim at changing soft tissue perfusion or blood flow. One problem at present is that soft tissue perfusion and its changes cannot be assessed in an interventional suite because cone-beam computed tomography is too slow (it takes 4–10 s per volume scan). In order to address the problem, we propose a novel method called IPEN for Intra-operative 4-dimensional soft tissue PErfusion using a standard x-ray system with No gantry rotation. In this simulation study, we provide a proof-of-concept that the core ideas work.

Coupling computed tomography with positron emission tomography (PET/CT) supplements tracer uptake with anatomical information for localization and improves PET quantification by using CT images for attenuation correction. Iodinated contrast agents in CT scans are used to enhance vascularity, organs, and abnormalities and for characterize lesions. However, current attenuation correction methodologies generate overestimates in standardized uptake values (SUV) in the presence of materials with high atomic numbers such as iodinated contrast agents. Utilizing electron density (ED) from dual energy CT (DECT) may result in less biased attenuation correction as ED is proportional to attenuation at PET emission energy of 511 keV. To evaluate different methods of attenuation correction, five phantom configurations with varying iodine concentrations and constant concentrations of Fluorine-18 were scanned using PET/CT and DECT at similar scanning parameters. Phantom configurations were scanned at CTDI_vol 2, 4, 6, and 8 mGy with DECT to evaluate the effect of dose on ED and SUV. For attenuation correction, ED was transformed into attenuation at 511 keV through reported material compositions and ED. SUV demonstrated less biased behavior in the presence of iodinated contrast media with ED-based correction (-1.3% to 1.4%, p=0.271) compared to nominal correction (1.5% to 8.6%, p=0.000). No interaction effect between dose and phantom configuration or effect of dose on SUV was present, which was also reflected in ED stability in different tissue mimics. Use of ED-based attenuation correction from DECT allowed for less biased SUV when increasing concentrations of iodinated contrast agents, indicating quantitative advantages of DECT coupled with PET.

Cadmium zinc telluride (CZT) radiation detectors are suitable for various applications, due to good energy performance at room temperature and simple pixilation to achieve high spatial resolution. Our group is developing a two-panel head-and-neck dedicated positron emission tomography (PET) system with CZT detectors. In this work, we present the back-end readout electronic design and the initial electronic noise results of our system. The back-end readout electronic incorporates RENA boards (a total 150 RENA boards for each panel), 30:1 fan-in boards connecting 30 RENA boards to the PicoZed 7010/7020 board (a total 5 fan-in boards for each panel). In each panel, 5:1 intermediate boards are used for biasing CZT detectors. The RENA board and the Picozed board are capable of data transmission of 50 Mbps and 6.6 Gbps, respectively. Electronic noise was also quantified using a square wave test pulse that provides charge injection into each channel of the RENA chip in the amount of 75fC/Volt. The pulse amplitude was chosen to generate approximately the same amount of charges as 511 keV photon would provide for each channels. The FWHM electronic noise at 511 keV was measured to be less than 1% (FWHM of 7.80 ± 1.47 ADC units or 4.89 ± 0.92 keV after conversion).

Abdominal MRI is susceptible to respiratory motion artifacts. The existing clinical solution is using breathing belt to track the movement of the abdomen and trigger MRI acquisition during the end-expiration phase. Attaching respiratory belt to patients often slows down clinical workflow and affects patient comfort especially for those with surgical wounds and respiratory disorders. Herein we, for the first, propose a novel MRI compatible frequency modulated continuous wave (FMCW) radar to track respiratory motion within MRI bore in a non-contact fashion. The electromagnetic wave from FMCW radar can penetrate clothing and MRI RF coils to achieve continuous monitoring of patient’s vital signs. The system consists of a front-end FMCW radar sensor and a FPGA based power management/communication board that interface with a clinical MRI scanner. This design fully integrates the FMCW radar signal with MRI control console to enable real time respiratory triggered MRI acquisition. Consistent respiratory waveform was validated by comparing FMCW signal with traditional breathing belt measurement. Superior image quality from clinical MRI pulse sequence was achieved using the developed system in healthy volunteers.

Attenuation compensation (AC) is a pre-requisite for reliable quantification and beneficial for visual interpretation tasks in single-photon emission computed tomography (SPECT). Typical AC methods require the availability of an attenuation map, which is obtained using a transmission scan, such as a CT scan. This has several disadvantages such as increased radiation dose, higher costs, and possible misalignment between SPECT and CT scans. Also, often a CT scan is unavailable. In this context, we and others are showing that scattered photons in SPECT contain information to estimate the attenuation distribution. To exploit this observation, we propose a physics and learning-based method that uses the SPECT emission data in the photopeak and scatter windows to perform transmission-less AC in SPECT. The proposed method uses data acquired in the scatter window to reconstruct an initial estimate of the attenuation map using a physicsbased approach. A convolutional neural network is then trained to segment this initial estimate into different regions. Predefined attenuation coefficients are assigned to these regions, yielding the reconstructed attenuation map, which is then used to reconstruct the activity distribution using an ordered subsets expectation maximization (OSEM)-based reconstruction approach. We objectively evaluated the performance of this method using highly realistic simulation studies conducted on the clinically relevant task of detecting perfusion defects in myocardial perfusion SPECT. Our results showed no statistically significant differences between the performance achieved using the proposed method and that with the true attenuation maps. Visually, the images reconstructed using the proposed method looked similar to those with the true attenuation map. Overall, these results provide evidence of the capability of the proposed method to perform transmissionless AC and motivate further evaluation.

Background: Compressed sensing magnetic resonance imaging (CS-MRI) is an important technique of accel- erating the acquisition process of magnetic resonance (MR) images by undersampling. It has the potential of reducing MR scanning time and costs, thus minimising patient discomfort. Motivation: One of the successful CS-MRI techniques to recover the original image from undersampled images is generative adversarial network (GAN). However, GAN-based techniques suffer from three key limitations: training instability, slow convergence and input size constraints. Method and Result: In this study, we propose a novel GAN-based CS-MRI technique: WPD-DAGAN (Wavelet Packet Decomposition Improved de-aliaising GAN). We incorporate Wasserstein loss function and a novel structure based on wavelet packet decomposition (WPD) into the de-aliaising GAN (DAGAN) architecture, which is a well established GAN-based CS-MRI technique. We show that the proposed network architecture achieves a significant performance improvement over the state-of-the-art CS-MRI techniques.

Motion artefacts created by patient motion during an MRI scan occur frequently in practice, often rendering the scans clinically unusable and requiring a re-scan. While many methods have been employed to ameliorate the effects of patient motion, these often fall short in practice. In this paper we propose a novel method for detecting and timing patient motion during an MR scan and correcting for the motion artefacts using a deep neural network. The deep neural network contains two input branches that discriminate between patient poses using the motion’s timing. The first branch receives a subset of the k-space data collected during a single dominant patient pose, and the second branch receives the remaining part of the collected k-space data. The proposed method can be applied to artefacts generated by multiple movements of the patient. Furthermore, it can be used to correct motion for the case where k-space has been under-sampled to shorten the scan time, as is common when using methods such as parallel imaging or compressed sensing. Experimental results on both simulated and real MRI data show the efficacy of our approach.

In this work, we have used Geant4 Monte Carlo simulations to evaluate the spatial resolution and detection efficiency metrics of an amorphous selenium (a-Se) photon-counting detector for high energy gamma-ray detection with 1mm and 50µm pixel pitches and a varying number of detector layers. A noiseless monochromatic particle gun with an energy of 140 keV is used to resemble the typical energy of single-photon emission computed tomography (SPECT). At high energies like 140 keV, a-Se detection efficiency significantly drops due to its absorption coefficient. By using this novel multilayer a-Se detector, the drop in detection efficiency can be compensated. The point spread function (PSF) is obtained by illuminating the detector with 106 photons. The modulation transfer function (MTF) is calculated from the one-dimensional integral of the PSF, known as line spread function (LSF), and is compared to the ideal pixel MTF. Spatial resolution is considered as the spatial frequency at which the MTF is equal to 0.5. The simulation results indicate that by increasing the number of layers, the MTF was degraded slightly due to Compton scattering, however, it did not degrade spatial resolution for the 1mm pixel size. At the same time, by using more layers, the detection efficiency is increased to 80% for 10 layers. This detection efficiency includes the noise counts (error counts) caused by Compton scattering. Leveraging the photon-counting energy threshold enables partial compensation for noise counts. Using an energy threshold of 110 keV results in 52% efficiency for 10 layers and reduces the noise counts significantly. This increase in detection efficiency along with the high intrinsic spatial resolution makes a-Se a cost-effective candidate for large area SPECT applications.

Active matrix flat panel imagers (AMFPIs) with thin film transistor (TFT) arrays are becoming the standard for digital x-ray imaging due to their high image quality and real time readout capabilities. However, in low dose applications their performance is degraded by electronic noise. A promising solution to this limitation is the Scintillator High-Gain Avalanche Rushing Photoconductor AMFPI (SHARP-AMFPI), an indirect detector that utilizes avalanche amorphous selenium (a-Se) to amplify optical signal from the scintillator prior to readout. We previously demonstrated the feasibility of a large area SHARP-AMFPI, however there are several areas of desired improvement. In this work, we present a newly fabricated SHARP-AMFPI prototype detector with the following developments: metal oxide hole blocking layer (HBL) with improved electron transport, transparent bias electrode for increased optical coupling, and detector assembly allowing for a back-irradiation (BI) geometry to improve detective quantum efficiency of scintillators. Our measurements showed that the new prototype has improved temporal performance, with lag and ghosting below 1%. We also show an improvement in optical coupling from 25% to 90% for cesium iodide (CsI) scintillator emissions. The remaining challenge of the SHARP-AMFPI is to reduce the dark current to prevent dielectric breakdown under high bias and further increase optical quantum efficiency (OQE) to CsI scintillators. We are proposing to use a newly developed quantum dot (QD) oxide layer, which shows to reduce the dark current by an order of magnitude, and tellurium doping, which could increase OQE to 85% to CsI at avalanche fields, in future prototype detectors.

The DQE performance of active matrix flat-panel imagers (AMFPIs) degrades under low dose conditions – such as those encountered in digital breast tomosynthesis where electronic additive noise becomes comparable to imaging signal. Compared to commercially available x-ray converter materials such as CsI:Tl and a-Se, particle-in-binder polycrystalline mercuric iodide (PIB HgI₂) offers imaging signal 3 to 10 times larger. However, PIB HgI₂ exhibits an unacceptably high degree of image lag, believed to originate from the trapping of holes. To suppress hole signal contribution (with the expectation of decreasing image lag), a Frisch grid structure embedded within a PIB HgI₂ detector is under investigation. In this theoretical study involving finite element analysis modeling, the effects of grid design and charge carrier lifetimes on the line spread function and modulation transfer function (MTF) were investigated. Two design parameters were considered: grid pitch (defined as the center-to-center distance between adjacent grid elements) and R_GRID (defined as the ratio of grid element width to grid pitch). Results show that for a grid pitch comparable to the pixel pitch (i.e., 100 μm) and high R_GRID, MTF is significantly degraded compared to a detector without a grid while, for a grid pitch of 20 μm, MTF is largely maintained and is almost independent of R_GRID. The best identified design is a grid pitch of 20 μm and R_GRID of 45% – providing an MTF similar to that of a detector without a grid while suppressing hole signal by 78%. This work is supported by NIH grant R01-EB022028.

Indirect flat panel imager (I-FPI) performance is limited by noise in the scintillator’s response to single x-rays, i.e. random variation of gain and blur. Single x-ray imaging (SXI) is an experimental methodology to record the response of scintillators to single x-rays using an image intensifier that is lens coupled to an EMCCD camera. In this work we developed a framework for using ensembles of SXI images to compute the spatial-frequency-dependent pulse height spectrum, Pr[g(f)], a novel detector performance metric whose moments may be used to compute scintillator MTF(f), qNNPS(f), and DQE(f). The approach is demonstrated using cesium iodide (CsI:Tl) samples of various thicknesses and screen-optical properties. SXI was conducted on each sample using ~60 keV γ-rays emitted by Am-241. Each image in an SXI ensemble was summed in 1D, and Fourier transformed to form an ensemble of frequencydependent responses. The histogram at each frequency f forms the Pr[g(f)]. To validate this approach against standard methods, each sample was also evaluated using the standard method by coupling it to a photodiode-TFT array as in a conventional I-FPI. MTF(f) and qNNPS(f) were measured using a slanted edge and flood field images, respectively, using an 85 kVp x-ray spectrum filtered to achieve an average energy of 60 keV. MTF(f) and qNNPS(f) results agreed well between both methods, and discrepancies were consistent with the differing experimental conditions. The Pr[g(f)] may be used to help optimize CsI:Tl in I-FPIs, and to improve the computational efficiency of Monte Carlo simulations of I-FPI performance.

The MTF and DQE of an x-ray detector are important metrics of performance. They are generally measured following IEC guidelines which requires the use of raw (unprocessed) images. However, many detectors now incorporate processing algorithms that are deeply embedded in the imaging chain and may be difficult or impossible to disable, blurring the line between processed and unprocessed image data, and questioning the relevance of MTF and DQE testing. This study develops a framework to represent the spectral performance of linear and shift-invariant digital-processing algorithms and examines their effects on MTF and DQE measurements with examples. Processing algorithms are represented as a cascade of operations where images are represented using an integrable impulse-sampled notation. This allows the use of Fourier transform theorems and relationships, which differ to a discrete Fourier transform notation, including a specific representation of signal and noise aliasing. It is concluded that: i) digital convolution of image data gives the same result, with the same aliasing artifacts, as a true convolution integral of presampling image data followed by sampling; ii) the slanted-edge method to measure MTF provides the presampling MTF even when processing algorithms operate on aliased image data; iii) the DQE is largely unaffected by LSI post processing, however spectral zero-crossings can suppress specific frequency content in both the MTF and DQE, and unsharp masking algorithms can decrease the DQE at low frequencies.

Physical density assessments may provide valuable insights for a range of diagnostic purposes in abdominal, pulmonary, and breast imaging. These purposes include differentiating iso-attenuating cysts from lesions, aiding in function tests of viral pneumonia, and quantifying breast cancer risk. Physical density assessments also provide a natural and intuitive measurement of human tissue that may be useful for measuring global body mass distributions and comparing measurement techniques. In this work, we present a spectral physical density map generated from clinical dual-energy computed tomography (DECT) datasets. We utilized available DECT estimates for effective atomic number, monoenergetic attenuation, and electron density as inputs into the Alvarez-Macovski model and the relation between a material’s physical and electron densities. To achieve higher accuracy assessments, these underlying equations were parametrized and fit to published tissue composition and density data that had been supplemented with simulated iodine enhancements. Validating the fits with phantom experiments, we observed measurements that are within ± 0.02 g/ml of their nominal values. Our assessments thus fall inside the margin-of-error for the ground-truth densities declared by the phantom manufacturer. Though we validated our density maps only on a dual-layer DECT implementation, the development and optimization of this new spectral result were independent of any other spectral CT technology. Because of this independence and the high accuracies of the maps, we encourage future clinical trials testing the potential applications of this new result for diagnostic imaging.

Purpose: To mitigate effects of metal artifacts in Dual-Energy (DE) CT imaging, we introduce a constrained optimization algorithm to enable simultaneous reconstruction-decomposition of three materials: two tissues-of-interest and the metal. Methods: The volume conservation principle and nonnegativity of volume fractions were incorporated as a pair of linear constraints into the Model-Based Material Decomposition (MBMD) algorithm. This enabled solving for three unknown material concentrations from DE projection data. A primal-dual Ordered Subsets Predictor-Corrector Interior-Point (OSPCIP) algorithm was derived to perform the optimization in the proposed constrained-MBMD (CMBMD). To improve computational efficiency and monotonicity of CMBMD, we investigated an approach where the constraint was applied locally onto a small region containing the metal (identified from a preliminary reconstruction) during initial iterations, followed by final iterations with the constraint applied globally. Validation studies involved simulations and test bench experiments to assess the quantitative accuracy of bone concentration measurements in the presence of fracture fixation hardware. In all studies, DE data was acquired using a kVp-switching protocol with the 60 kVp low-energy beam and the 140 kVp high-energy beam. The system geometry emulated the extremity Cone-Beam CT (CBCT). Simulation studies included: i) a cylindrical phantom (80 mm diameter) with a 30 mm long Ti screw and an insert of varying cortical bone concentrations (3 – 13%), and ii) a realistic tibia phantom created from patient CBCT data with Ti fixation hardware of increasing complexity. The test bench experiment involved a 100 mm diameter water bath containing four Ca inserts (6.5 – 39.1% bone concentration) and a Ti plate. Results: CMBMD substantially reduced artifacts in DE decompositions in the presence of metal. The sequentially localglobal constraint strategy resulted in more monotonic convergence than using a global constraint for all iterations. In the simulation studies, CMBMD achieved quantitative accuracy within ~12% of nominal bone concentration in areas adjacent to metal, and within ~5% in areas further away from the metal, compared to ~80% error for the two-material MBMD. In the test bench study, CMBMD generated ~40% reduction in the error of bone concentration estimates compared to MBMD for nominal insert concentrations of <250 mg/mL, and ~12% reduction for concentrations <250 mg/mL. Conclusion: Proposed CMBMD enables accurate DE decomposition in the presence of metal implants by incorporating the metal as an additional base material. Proposed method will be particularly useful in quantitative orthopedic imaging, which is often challenged by metal fracture fixation and joint replacement hardware.

Liver lesion detection and characterization presents a longstanding challenge for radiologists. Since liver lesions are mainly characterized from information obtained at both arterial and portal venous circulatory phases, current hepatic Computed tomography (CT) protocols involve intravenous contrast injection and subsequent multiple CT acquisitions. Because detection of lesions by CT often requires further investigation with MRI, improved differentiation CT capabilities are highly desirable. Recently developed imaging protocols for spectral photon-counting CT enable simultaneous mapping of arterial and portal-venous enhancements by injecting two different contrast agents sequentially, allowing robust pixel-to- pixel spatial alignment between the different contrast phases with a reduction of radiation exposure. Here we propose a method that allows to quantitatively and reliably distinguish between two contrast agents in a single dual-energy CT (DECT) acquisition by taking advantage of the unique abilities of modern self-learning algorithms for non-linear mapping, feature extraction, and feature representation. For this purpose, we designed a U-net architecture convolutional neural network (CNN). To overcome training data requirements, we utilizing clinical DECT images to simulate dual-contrast spectral datasets. With the unique network architecture and training datasets, we demonstrate reliable dual-contrast quantifications from DECT. Our results demonstrate an ability to quantify densities of water, iodine and gadolinium, with root mean square errors of 0.2 g/ml, 1.32 mg/ml and 1.04 mg/ml, respectively. While observing some material-cross artifacts, our model demonstrated a high robustness to noise. With the rapid increase in DECT usage, our results pave the way for improved diagnostics and better patient outcome with available hardware implementations.

Dual energy CT acquisitions can produce two material basis images with high accuracy. From dual energy CT scans with contrast enhancement, one often attempts to generate the corresponding non-enhanced CT images for calcium scoring or for renal stone detection. However, to accurately generate the virtual non-contrast (VNC) images, three-material decomposition is needed. With dual energy CT images, one needs to introduce additional constraints such as the mass or volume conservation condition to approximately perform three-material decomposition. In this work, we present a deep learning strategy to generate VNC image from DECT material basis images. In this strategy, the needed constraint for three-material decomposition is accomplished by learning from the large amount of available training data. In practice, the supervised learning strategy requires matched training data, but this requirement is hard to satisfy in practice since the DECT and non-contrast CT images are acquired in two different CT scans and thus mis-registration often plagues the ordinary supervised learning strategies. In this paper, a new strategy was developed to enable the training of the proposed VNC-Net without using matched training data.

Simultaneous estimation of spectra and basis images in multispectral CT reconstruction employs a data model with unknown spectra. One approach is based on the linearization of the data model, which leads to two linear terms, with regards to the basis image and to the spectrum. The latter one, i.e., the linearized matrix of spectral contribution, is new, to the best of our knowledge, and warrants investigations. In this work, we have characterized the conditioning of the linearized matrix of spectral contribution using singular value decomposition (SVD). We have also proposed a SVD-based preconditioner for the matrix and incorporated it in a constrained optimization problem for recovering the spectrum. The results have showed improved conditioning of the matrix and accurate recovery of the spectrum by use of the SVD-based preconditioner.

The PET-enabled dual-energy CT method allows dual-energy CT imaging on PET/CT scanners without the need for a second x-ray CT scan. A 511 keV γ-ray attenuation image can be reconstructed from time-of-flight PET emission data using the maximum-likelihood attenuation and activity (MLAA) algorithm. However, the attenuation image reconstructed by standard MLAA is commonly noisy. To suppress noise, we propose a neuralnetwork approach for MLAA reconstruction. The PET attenuation image is described as a function of kernelized neural networks with the flexibility of incorporating the available x-ray CT image as anatomical prior. By applying optimization transfer, the complex optimization problem of the proposed neural MLAA reconstruction is solved by a modularized iterative algorithm with each iteration decomposed into three steps: PET activity image update, attenuation image update, and neural-network learning using a weighed mean squared-error loss. The optimization algorithm is guaranteed to monotonically increase the data likelihood. The results from computer simulations demonstrated the neural MLAA algorithm can achieve a significant improvement on the γ-ray CT image quality as compared to other algorithms.

Cancerous masses are more conspicuous in wide-angle digital breast tomosynthesis (DBT) due to better depth resolution and tissue separation, while the detection of subtle microcalcifications (MC) is challenging. This study aims at providing guidance for a new DBT system design to enhance lesion detection through variable angular dose image acquisition and improved detector performance. Digital breast phantoms were generated, and projection images were simulated using the FDA VICTRE tool ¹. Simulated masses and clusters of MC were inserted at different locations of the digital breast phantom. Projection images were simulated with 28 kVp W/Rh energy spectrum, 200 μm thick amorphous selenium (a- Se) direct active matrix flat panel imager (AMFPI), and 50° angular range with 25 views. The impact of a-Se detector performance, i.e., complementary metal-oxide-semiconductor (CMOS) active pixel sensor (APS) versus AMFPI, with different electronic noise and pixel pitches, was investigated. Even and uneven angular dose distribution schemes (ADS) were designed to test the combined effect of image acquisition setting and detector performance on lesion detection. 2D Filtered Channel observer (FCO) and 3D Channelized Hotelling observer (CHO) were employed to evaluate lesion detectability under different scenarios. The results demonstrated the improvement of direct conversion CMOS APS on the detectability of small MC clusters by reducing FSM, improving DQE at a lower dose, and improve the sharpness of reconstructed MC. The uneven dose distribution benefits the detection of MC without compromising mass detection in FBP reconstructed volumes with slice thickness filter for wide angle DBT.

Aim: Compare an in-house developed hybrid simulation framework and the FDA’s total simulation framework VICTRE in an exercise to simulate realistic DBT images. Methods: Three different set-ups were investigated in increasing order of difficulty: (1) A simple object insert simulated in homogeneous backgrounds. (2) The same simple test object in a breast phantom to investigate the impact of a non-homogeneous background. (3) The simple test object replaced with clinically relevant lesions (spiculated and non-spiculated masses, calcification clusters) to test the frameworks in their entirety. Next to a visual analysis, a quantitative comparison based on contrast and signal-difference-to-noise (SDNR) measurements was performed. Results: Similar contrast and SDNR values in the ‘for processing’ images are obtained for a glandular test object when simulated with VICTRE and the Leuven platform, for both homogeneous as well as structured backgrounds (e.g. structured background; contrast: 0.13 and 0.12, SDNR: 8.99 and 8.56 for the Leuven platform and VICTRE respectively). The reconstruction algorithms of both frameworks differ, but the input of VICTRE images in an offline reconstruction tool like in the Leuven framework leads to similar results. In DBT reconstructed slices, the simulated mass models looked similar to the real lesions from which the model was derived. The simulated calcification clusters are more subtle when using the VICTRE framework, while all clusters appear to be realistic. This illustrates the need of full characterization of the methods. Conclusion: The step-by-step comparison of two very different frameworks was successful. Both frameworks are able to simulate objects with the same characteristics (contrast, SDNR, shape) and can create images with realistic lesions.

Our previous work showed that digital breast tomosynthesis (DBT) supports super-resolution (SR). Clinical systems are not yet designed to optimize SR; this can be demonstrated with a high-frequency line-resolution pattern. SR is achieved if frequencies are oriented laterally, but not if frequencies are oriented in the perpendicular direction; i.e., the posteroanterior (PA) direction. We are developing a next-generation tomosynthesis (NGT) prototype with new trajectories for the x-ray source. This system is being designed to optimize SR not just for screening, but also for diagnostic mammography; specifically, for magnification DBT (M-DBT). SR is not achieved clinically in magnification mammography, since the acquisition is 2D. The aim of this study is to investigate SR in M-DBT, and analyze how anisotropies differ from screening DBT (S-DBT). We have a theoretical model of a high-frequency sinusoidal test object. First, a conventional scanning motion (directed laterally) was simulated. In the PA direction, SR was not achieved in either S-DBT or M-DBT. Next, the scanning motion was angled relative to the lateral direction. This motion introduces submillimeter offsets in source positions in the PA direction. Theoretical modeling demonstrated that SR was achieved in M-DBT, but not in S-DBT, in the PA direction. This work shows that, with the use of magnification, anisotropies in SR are more sensitive to small offsets in the source motion, leading to insights into how to design M-DBT systems.

This paper investigates the training of deep convolutional neural networks (DCNNs) to denoise digital breast tomosynthesis (DBT) images. In our approach, the DCNN was trained with high dose (HD)/low dose (LD) image pairs, in which the HD image was used as reference to guide the DCNN to learn to reduce noise of a corresponding input LD image. In the current work, we studied the effect of the dose level of the HD target images on the effectiveness of the trained denoiser. We generated a set of simulated DBT data with different dose levels using CatSim, a dedicated x-ray imaging simulator, in combination with virtual anthropomorphic breast phantoms simulated by VICTRE. We found that a DCNN trained with higher dose level of HD target images led to less noisy denoised images. We also acquired DBT images of real physical phantoms containing simulated microcalcifications (MCs) for training and validation. The denoisers trained with either simulated or physical phantom data improved significantly (𝑝𝑝 < 0.0001) the contrast-tonoise ratio of MCs in the validation phantom images. In an independent test set of human subject DBTs, the MCs became more conspicuous, and the mass margins and spiculations were well preserved. The study showed that the denoising DCNN was robust in that the denoiser could be trained with either simulation or physical phantom data. Moreover, the denoiser trained with CatSim simulation data was directly applicable to human DBTs, allowing flexibility in terms of the training data preparation, especially the HD images.

It is standard for the x-ray source in conventional digital breast tomosynthesis (DBT) acquisitions to move strictly along the chest wall of the patient. A prototype, next-generation tomosynthesis (NGT) system has been developed that is capable of acquiring customized geometries with source motion parallel and perpendicular to the chest wall. One well-known consequence of acquiring projections with the x-ray source anterior to the chest wall is that a small volume of tissue adjacent to the chest wall is missed. Here we evaluate strategies in DBT to avoid missing tissue while improving overall image quality. Acquisition geometries tested in this study include the conventional (control), “T-shape,” and “bowtie” geometries. To evaluate the impact of moving the x-ray source away from the chest wall, the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were measured as a function of location within the reconstructed volume. Using simulations and physical experiments, the SNR and CNR were compared with conventional DBT. Simulations of two different phantoms were performed: a “tube” phantom and a “lattice” phantom. Experiments with uniform and textured phantoms were also conducted. While the image quality was slightly reduced immediately adjacent to the chest wall, there was no missed tissue and both the T-shape and Bowtie geometries exhibited SNR and CNR improvement over the vast majority of the reconstruction volume; the overall result being an improvement in image quality with both the T-shape and bowtie geometries.

X-ray phase imaging has found limited clinical use due to requirements on x-ray coherence that may not be easily translated to clinical practice. Instead, this work employs a conventional source to create structured illumination with a simple wire mesh. The system has been employed to produce high contrast absorption images with simultaneous differential phase contrast images. In previous work we have demonstrated accurate quantitative phase extraction. In this work, we have incorporated the dark field information to successfully reveal additional structure in low contrast objects.

The aim of this study is to investigate and reveal the potential of employing a direct conversion amorphous selenium (a-Se) CMOS based high resolution x-ray detector in both propagation-based (PB) and edge illumination (EI) x-ray phase contrast imaging (XPCi) techniques. Both PB-XPCi and EI-XPCi modalities are evaluated through a numerical model and are compared based on their contrast, edge-enhancement, visibility, and dose efficiency characteristics. It is demonstrated how EI-XPCi configuration outperforms the PB-XPCi one considering using the same x-ray source and detector. After highlighting the strength of EI-XPCi system and reviewing today’s XPCi technologies, absorption mask grating fabrication is addressed as the main challenge to upgrade and improve EI-XPCi setups to higher resolution detectors. Mammography is considered as a case study to elucidate the importance of employing a high resolution EI-XPCi technique for microcalcification detection through numerical simulation of a breast phantom.

X-ray fluorescence computed tomography (XFCT) with nanoparticles (NPs) as contrast agents has reached technical maturity allowing for in vivo preclinical imaging in the laboratory setting. We present the first in vivo longitudinal study with XFCT where mice were 5 times each during an 8-week period. Imaging is performed with low radiation dose (<25 mGy) and high signal-to-background for high-spatial-resolution imaging (200-400 µm) of molybdenum NP accumulations (down to ~50 µg/ml Mo). We further discuss our ongoing development of protein-coated NPs for actively targeting molecular markers (e.g., cancer), as well as potential clinical applications.

Gout is the most prevalent inflammatory arthritis found in men. A prompt diagnosis and early treatment of gout are crucial in preventing eventual functional impairment and reduction in comorbidities. In this work, the quantitative material information provided by a multi-contrast x-ray (MCXR) imaging acquisition is leveraged to develop a rapid, non-invasive, and low dose diagnostic method for gout detection and gout-pseudogout differentiation. This work establishes a theoretical foundation to demonstrate how a single-kV MCXR acquisition is capable of differentiating gout from pseudogout via a projection domain two-material decomposition. Experimental results from a benchtop MCXR system are presented. The imaging performance of the proposed MCXR technique is compared to dual-energy radiography to further validate the method.

The interior tomographic reconstruction problem concerns reconstruction of a local region of interest (ROI) from projection data that are conformally collimated to only illuminate the target ROI. In the past decade or so, it has been proven that a stable solution exists to address the interior tomographic reconstruction problem provided that a) the Tuy data sufficiency condition is satisfied and b) image values of some pixels inside the ROI are known(not all pixels to void triviality), although there seems to be no analytical reconstruction method available to reconstruct the image. In this work, we present a new pathway to address the interior tomographic reconstruction problem, referred to as DeepInterior. The new scheme consists of two steps: 1) Direct backprojection of the acquired fully truncated divergent beam projection data to form a backprojection image B₀ without the conventional differentiation operations and 2) the true image is then reconstructed from the blurred image B₀ using a trained deep neural network architecture. We demonstrated that the trained DeepInterior is shift-invariant and can accurately reconstruct ROIs at arbitrary locations.

Manifold learning using deep neural networks been shown to be an effective tool for noise reduction in low-dose CT. We propose a new iterative CT reconstruction algorithm, called Manifold Reconstruction of Differences (MRoD), which combines physical models with a data-driven prior. The MRoD algorithm involves estimating a manifold component, approximating common features among all patients, and the difference component which has the freedom to fit the measured data. While the manifold captures typical patient features (e.g. healthy anatomy), the difference image highlights patient-specific elements (e.g. pathology). We present a framework which combines trained manifold-based modules with physical modules. We present a simulation study using anthropomorphic lung data showing that the MRoD algorithm can both isolate differences between a particular patient and the typical distribution, but also provide significant noise reduction with relatively low bias.

Many modern medical imaging systems are computed in nature and need computational reconstruction to form an image. When the acquired measurements are insufficient to uniquely determine the sought-after object, prior knowledge about the object needs to be utilized in order to successfully recover an image. An emerging line of research involves the use of deep generative models such as generative adversarial networks (GANs) as priors in the image reconstruction procedure. However, when GANs are employed, reconstruction of images outside the range of the GAN leads to errors that result in realistic but false image estimates. To address this issue, an image reconstruction framework is proposed that is formulated in the latent space of an invertible generative model. A novel regularization strategy is introduced that takes advantage of the architecture of certain invertible neural networks (INNs). To evaluate the performance of the proposed method, numerical studies of reconstructing images from stylized MRI measurements are performed. The method outperforms classical reconstruction methods in terms of traditional image quality metrics and is comparable to a state-of- the-art adaptive GAN based framework, while benefiting from a deterministic procedure and easier tuning of the regularization parameters.

Cone-Beam Computed Tomography (CBCT) usually suffers from motion blurring artifacts when scanning at the region of the thorax. Consequently, it may result in inaccuracy in localizing the target of treatment and verifying the delivered dose in radiation therapy. Despite that 4D-CBCT reconstruction technology could alleviate the motion blurring artifacts, it introduces severe streaking artifacts due to the under-sampled projections used for reconstruction. Aiming at improving the overall quality of 4D-CBCT images, we explored the performance of the deep learning-based technique on 4D-CBCT images. Inspired by the high correlation among these 4D-CBCT images, we proposed a spatial-temporal plus prior image-based CNN, which is a cascaded network of a spatial CNN and a temporal CNN. For the spatial CNN, it is in the manner of the encoder-decoder architecture that utilizes a pair of a prior image-based channel and an artifact-degraded channel in the encoder stage for feature representation and fuses the feature maps for image restoration in the decoder stage. Next, three consecutive phases of images that are predicted by N-net individually are stacked together for latent image restoration via the temporal CNN. By doing so, temporal CNN learns the correlation between these images and construct the residual map covering streaking artifacts and noise for further artifact reduction. Experimental results of both simulation data and patient data indicated that the proposed method has the capability not only reduces the streaking artifacts but also restore the original anatomic features while avoiding inducing error tomographic information.

In this work, we studied the adversarial attacks and their corresponding defense strategies specifically in x-ray computed tomography image reconstruction tasks. After a small amount of imperceptible noise was added to the input image, these barely noticeable additional noise to the input image resulted in artifactual false-positive structures into output images of the well referenced high performance deep learning reconstruction methods. Since the adversarial attacks often occur at a specific stage of the entire imaging chain, defense measures can be developed to incorporate the uncontaminated data in the imaging chain into the image reconstruction framework to eliminate hazardous effects of adversarial attacks.

Due to the subtle variations of energy response functions across photoconductive panels, photon counting detector CT are subject to severe banding artifacts. This work presents a physics based method to correct for these artifacts. It employs calibration objects with known thicknesses and composition to estimate the panel-specific response functions, which are used to concert the raw photon counting projection data of an arbitrary image object into acrylic- and aluminum-equivalent pathlengths. Experimental results show the method not only remove the banding artifacts but also address the beam hardening artifacts.

The performance of photon counting detectors (PCDs) are limited by the spectral distortion caused by charge sharing and pulse pileup. Recently we have proposed a new detector design called MEICC for multi-energy inter-pixel coincidence counter. MEICC records the counting signals in the same way as the current PCDs, while also recording the energy-dependent cross-talk with neighboring pixels. The previous studies have shown that MEICC decreases the effect of charge sharing significantly. Building on the previous success, the aim of this study was to assess the performance of MEICC when the pulse pileup also causes spectral distortion (in addition to charge sharing). Monte Carlo simulations have been performed to compare the performance of MEICC, the current PCD, and two different versions of real-time charge sharing correction schemes. The study found that MEICC had the best performance when count rates were moderate or high.

In a clinical setting, the accurate representation of pathologies and anatomical structures in CT reconstructions is important for diagnostic and therapeutical interventions. However, the presence of metal within the patient can cause significant streak artifacts within the reconstruction, possibly degrading or obscuring features within a region of interest to the imaging study. Spectral CT can be utilized to better resolve the attenuation properties of different materials and hence provide more accurate reconstructions of images. In addition, accurately modeling the polyenergetic transmission of x-ray photons, the noise properties of the data, and undersampling caused by materials with high photon attenuation can address some of the artifacts caused by metal. Previous work using a one-step image reconstruction algorithm (cOSSCIR), demonstrated the potential to remove metal artifacts caused by beam hardening and photon starvation, but streaks due to the high gradient around the boundary of metal remained. Within this framework, an accurate model of the detector geometry can be leveraged to further address metal artifacts in the reconstructed image. In this study, we investigate errors caused by “partial-volume” effects using photon-counting CT phantom simulations. Noiseless 2D fan-beam data was first generated from a pelvic phantom with bilateral hip prostheses, where each detector element was modeled as having a finite aperture using multiple ‘raylets.’ The cOSSCIR framework was modified to incorporate the nonlinear partial volume (NLPV) model based on raylets. Basis images were reconstructed by cOSSCIR with and without the NLPV model. The simulation results demonstrated that cOSSCIR with the NLPV model was able to accurately tom with metal, therefore removing metal artifacts caused by beam hardening and nonlinear partial volume effects.

This work extends the methodology for estimating spectral response of photon-counting detectors (PCDs) from transmission measurements through an object of known composition and geometry. The extension incorporates a non-linear model that relates the photon counts incident on the PCD with the photon counts registered by the PCD. The model calibration is performed on an actual step-wedge phantom and it used to predict Xray transmission through a different material than what comprises the phantom. In addition, stability of the determination of the model parameters is investigated, and this analysis is used to determine regularization strength and basis materials for the test object.

Photon counting x-ray detectors (PCDs) with energy discrimination capabilities provide spectral information, can eliminate electronic and Swank noise, and image multiple contrast agents simultaneously with high resolution. However, with high flux on the detectors, pulse pileup leads to count loss and spectral distortion. Accurate description of the output count rate and spectrum behavior of a PCD is crucial to the development of further applications, e.g., material decomposition. In this work, a fully analytical model of the pulse pileup spectrum and count statistics for a nonparalyzable detector with nonzero pulse length is proposed, which accounts for seminonparalyzable behavior, i.e., retriggering of counts by pulses incident during the dead time. We use a triangle to approximate a realistic pulse shape and recursively compute the output spectrum of different pulse pileup orders. A count-rate and spectrum dependent expression of count statistics is further derived based on renewal theory. Our analytical model can then be used to predict material decomposition noise using the Cramér–Rao lower bound (CRLB). We compared our model predictions with Monte Carlo simulations. The results show that our model can accurately predict the count loss and spectral distortion resulting from pulse pileup and can be used in predicting material decomposition noise over a large range of count rates.

While the benefits of photon counting (PC) CT are significant, the performance of current PCDs is limited by physical effects distorting the spectral response. In this work, we examine a deep learning (DL) approach for spectral correction and material decomposition of PCCT data using multi-energy CT data sets acquired with an energy integrating detector (EID) for spectral calibration. We use a convolutional neural network (CNN) with a U-net structure and compare image domain and projection domain approaches to spectral correction and material decomposition. We trained our networks using noisy, spectrally distorted PCD projections as input data while the labels were derived from multi-energy EID data. For this study, we have scanned: 1) phantoms containing vials of iodine and calcium in water and 2) mice injected with an iodine-based liposomal contrast agent. Our results show that the image-based approach corrects best for spectral distortions and provides the lowest errors in material maps (RMSE in iodine: 0.34 mg/mL) compared with the projection-based approach (0.74 mg/mL) and image domain decomposition without correction (2.67 mg/mL). Both DL methods are, however, affected by a loss of spatial resolution (from 3 lp/mm in EID labels to ~2.2 lp/mm in corrected reconstructions). In summary, we demonstrate that multi-energy EID acquisitions can provide training data for DL-based spectral distortion correction. This data-driven correction does not require intimate knowledge of the spectral response or spectral distortions of the PCD. While the benefits of photon counting (PC) CT are significant, the performance of current PCDs is limited by physical effects distorting the spectral response. In this work, we examine a deep learning (DL) approach for spectral correction and material decomposition of PCCT data using multi-energy CT data sets acquired with an energy integrating detector (EID) for spectral calibration. We use a convolutional neural network (CNN) with a U-net structure and compare image domain and projection domain approaches to spectral correction and material decomposition. We trained our networks using noisy, spectrally distorted PCD projections as input data while the labels were derived from multi-energy EID data. For this study, we have scanned: 1) phantoms containing vials of iodine and calcium in water and 2) mice injected with an iodine-based liposomal contrast agent. Our results show that the image-based approach corrects best for spectral distortions and provides the lowest errors in material maps (RMSE in iodine: 0.34 mg/mL) compared with the projection-based approach (0.74 mg/mL) and image domain decomposition without correction (2.67 mg/mL). Both DL methods are, however, affected by a loss of spatial resolution (from 3 lp/mm in EID labels to ~2.2 lp/mm in corrected reconstructions). In summary, we demonstrate that multi-energy EID acquisitions can provide training data for DL-based spectral distortion correction. This data-driven correction does not require intimate knowledge of the spectral response or spectral distortions of the PCD.

We propose an automated framework to generate 3D detailed person-specific computational phantoms directly from patient medical images. We investigate the feasibility of this framework in terms of accurately generating patient-specific phantoms and the clinical utility in estimating patient-specific organ dose for CT images. The proposed framework generates 3D volumetric phantoms with a comprehensive set of radiosensitive organs, by fusing patient image data with prior anatomical knowledge from a library of computational phantoms in a two-stage approach. In the first stage, the framework segments a selected set of organs from patient medical images as anchors. In the second stage, conditioned on the segmented organs, the framework generates unsegmented anatomies through mappings between anchor and nonanchor organs learned from libraries of phantoms with rich anatomy. We applied this framework to clinical CT images and demonstrated its utility for patient-specific organ dosimetry. The result showed the framework generates patientspecific phantoms in ~10 seconds and provides Monte Carlo based organ dose estimation in ~30 seconds with organ dose errors <10% for the majority of organs. The framework shows the potential for large scale and real-time clinic analysis, standardization, and optimization.

Skin dose is dependent on the incident beam angle and corrections are needed for accurate estimation of the risk of deterministic effects of the skin. Angular-correction factors (ACF) were calculated and incorporated into our skin-dosetracking system (DTS) and the results compared to Monte-Carlo simulations for a neuro-interventional procedure. To obtain the ACF’s, EGSnrc Monte-Carlo (MC) software was used to calculate the dose averaged over 0.5, 1, 2, 3, 4 and 5 mm depth into the entrance surface of a water phantom at the center of the field as a function of incident beam to skin angle from 90-10 degrees for beam field sizes from 5-15 cm and for beam energies from 60-120 kVp. These values were normalized to the incident primary dose to obtain the ACF. The angle of incidence at each mesh vertex in the beam on the surface of the DTS patient graphic was calculated as the complement of the angle between the normal vector and the vector of the intersecting ray from the tube focal spot; skin dose at that vertex was calculated using the corresponding ACF. The skin-dose values with angular correction were compared to those calculated using MC with a matching voxelized phantom. The results show the ACF decreases with decreasing incident angle and skin thickness, and increases with increasing field size and kVp. Good agreement was obtained between the skin dose calculated by the angular-corrected DTS and MC, while use of the ACF allows the real-time performance of the DTS to be maintained.

EGSnrc Monte-Carlo software was used to calculate the “patient’s skin dose” as a function of incident beam angle for cylindrical water phantoms with underlying subcutaneous fat and various thicknesses of bone. Simulations were done for incident angles from 90 to 10 degrees, entrance beam sizes from 5 to 15 cm, and energies from 60 to 120 kVp. The depth-averaged scatter-plus-primary to incident-primary dose ratio decreases with decreasing skin incident angle and increasing underlying bone thickness, and increases with increasing field size and energy. Corrections for these factors improve the accuracy of skin-dose estimation for neuro-interventional procedures with our Dose-Tracking-System.

Purpose: Scatter reduction remains a challenge for chest tomosynthesis. The purpose of this study was to validate a lowdose patient-specific method of scatter correction in a large animal model and implement the technique in a human imaging study in a population with known lung lesions. Method: The porcine and human subjects were imaged with an experimental stationary digital chest tomosynthesis system. Full field projection images were acquired, as well as with a customized primary sampling device for sparse sampling of the primary signal. A primary sampling scatter correction algorithm was used to compute scatter from the primary beam information. Sparse scatter was interpolated and used to correct projections prior to reconstruction. Reconstruction image quality was evaluated over multiple acquisitions in the animal subject to quantify the impact of lung volume discrepancies between scans. Results: Variations in lung volume between the full field and primary sample projection images induced mild variation in computed scatter maps, due to acquisitions during separate breath holds. Reconstruction slice images from scatter corrected datasets including both similar and dissimilar breath holds were compared and found to have minimal differences. Initial human images are included. Conclusions: We have evaluated the prototype low-dose, patient-specific scatter correction in an in-vivo porcine model currently incorporated into a human imaging study. The PSSC technique was found to tolerate some lung volume variation between scans, as it has a minimal impact on reconstruction image quality. A human imaging study has been initiated and a reader comparison will determine clinical efficacy.

We investigate an image-based strategy to compensate for cardiac motion-induced artifacts in Digital Chest Tomosynthesis (DCT). We apply the compensation to conventional unidirectional vertical “↕” scan DCT and to a multidirectional circular trajectory "O" providing improved depth resolution. Propagation of heart motion into the lungs was simulated as a dynamic deformation. The studies investigated a range of motion propagation distances and scan times. Projection-domain retrospective gating was used to detect heart phases. Sparsely sampled reconstructions of each phase were deformably aligned to yield a motion compensated image with reduced sampling artifacts. The proposed motion compensation mitigates artifacts and blurring in DCT images both for “↕” and "O" scan trajectories. Overall, the “O” orbit achieved the same or better nodule structural similarity index in than the conventional “↕” orbit. Increasing the scan time improved the sampling of individual phase reconstructions.

We propose two-dimensional (2D) dual-energy (DE) x-ray imaging of lung structure and function for the assessment of COPD, and investigate the resulting image quality theoretically using the human observer detectability index (d') as a figure of merit. We modeled the ability of human observers to detect ventilation defects in xenon enhanced DE (XeDE) images and emphysema in unenhanced DE images. Our model of d' accounted for the extent of emphysematous destruction and functional impairment as a function of defect/lesion contrast, spatial resolution, x-ray scatter, quantum and background anatomical noise power spectrum (NPS), and the efficiency of human observers. The effect of x-ray spectrum and exposure allocation factor on d' was also explored. Our results suggest that, the detectability is maximized for exposure allocation factors that minimize quantum NPS. The optimal combination of tube voltage was found to be ~50/140 kV or 60/140 kV depending on the task and patient at an x-ray exposure equal to that of a standard chest x-ray. In 2D DE x-ray imaging of COPD, the detectability is primarily limited by low contrast, x-ray scatter, and anatomic noise, the latter two of which reduce the detectability of individual defects by 30% and ~>90%, respectively.

Conventional x-ray imaging struggles with quantitation due to several challenges, including scatter, beam hardening, and multiple overlaying materials. We propose a new single-shot quantitative imaging (SSQI) method enabled by the combination of a primary modulator and a dual-layer detector to quantify area density of specific materials. The primary modulator enables scatter correction for the dual-layer detector, and the dual-layer detector enables beam hardening correction for the primary modulator. The dual-layer detector further provides motion-free dual-energy imaging from a single shot. When combined, the scatter-corrected dual-energy images can be used to create materialspecific images, including of soft tissue, bone, or iodine. A simulation study was performed of SSQI for chest x-ray imaging. We simulated projections of a digital chest phantom using a 120 kV spectrum, primary modulator, and dual-layer detector, with the addition of scatter and noise. Using the low-frequency property of scatter, a pre-calibrated material decomposition, and the known primary modulator pattern, we jointly recovered the scatter images in the dual-layer images and the material decomposition of the phantom. The resulting material decomposition accurately separated soft tissue from bone, reducing the RMSE in material-specific images by 66-84% as compared to no scatter correction. Through this simulation study, we have demonstrated the potential of SSQI for material quantification that is robust against scatter. The simplicity and broad applicability of SSQI to x-ray systems has the potential for widespread adoption, leading to improved quantitative imaging not only for chest x-ray but also for real-time image guidance and for cone-beam CT.

Chest radiography (CXR) plays an important role in triage, management, and monitoring of patients with COVID-19. In this work, we use a dual-layer (DL) flat-panel detector to perform artifact-free single-exposure DE CXR for COVID-19 detection. A simulation study was performed to generate DE CXRs from CT volumes of 13 patients diagnosed with COVID-19, which were compared with simulated conventional CXR by two chest radiologists. A phantom study was also conducted using an anthropomorphic chest phantom with synthetic opacifications. Improved visualization of opacities was observed for DE CXRs in both studies, indicating that the proposed DL detector could be a powerful new tool for COVID-19 diagnosis and patient management.

Breast density (BD) has been shown to be an independent risk factor for breast cancer, and it can be quantified by measuring the volumetric breast density (VBD). The longitudinal change in VBD has been used as a biomarker to assess the effect of chemoprevention drug for breast cancer risk reduction. However, the reproducibility of VBD measurement is hindered by breast compression, which causes temporal changes in the distribution of breast tissues, making it difficult to assess small longitudinal changes in VBD. Dual energy digital breast tomosynthesis (DE DBT) may provide more reproducible VBD measurements over different breast compressions. In this study, we aim to evaluate the reproducibility of VBD measurement using DE DBT. Digital breast phantoms were generated using a virtual clinical trial software (VICTRE) and were compressed multiple times into different clinical-relevant compressed breast thicknesses to emulate the change in repeated imaging of the same breast. The compressed thickness difference of 10 mm exaggerates the variation seen in clinical practice. DE DBT projection images of the phantoms were generated by Monte-Carlo simulations and included quantum noise and scatter radiation. A previously developed scatter correction method and a DE material decomposition technique were applied to obtain the BD map and calculate VBD. Our results show that DE DBT can provide reproducible VBD measurements for breasts under different compressions with the average absolute discrepancy of 2.3% ± 1.1% between two repeated measurements for all phantoms. The largest discrepancy is 3.9% for the extreme case with 20 mm compressed thickness difference.

Purpose: We compare the effects of scatter on the accuracy of areal bone mineral density (BMD) measurements obtained using two flat-panel detector (FPD) dual-energy (DE) imaging configurations: a dual-kV acquisition and a dual-layer detector. Methods: Simulations of DE projection imaging were performed with realistic models of x-ray spectra, scatter, and detector response for dual-kV and dual-layer configurations. A digital body phantom with 4 cm Ca inserts in place of vertebrae (concentrations 50 - 400 mg/mL) was used. The dual-kV configuration involved an 80 kV low-energy (LE) and a 120 kV high-energy (HE) beam and a single-layer, 43x43 cm FPD with a 650 μm cesium iodide (CsI) scintillator. The dual-layer configuration involved a 120 kV beam and an FPD consisting of a 200 μm CsI layer (LE data), followed by a 1 mm Cu filter, and a 550 μm CsI layer (HE data). We investigated the effects of an anti-scatter grid (13:1 ratio) and scatter correction. For the correction, the sensitivity to scatter estimation error (varied ±10% of true scatter distribution) was evaluated. Areal BMD was estimated from projection-domain DE decomposition. Results: In the gridless dual-kV setup, the scatter-to-primary ratio (SPR) was similar for the LE and HE projections, whereas in the gridless dual layer setup, the SPR was ~26% higher in the LE channel (top CsI layer) than in the HE channel (bottom layer). Because of the resulting bias in LE measurements, the conventional projection-domain DE decomposition could not be directly applied to dual-layer data; this challenge persisted even in the presence of a grid. In contrast, DE decomposition of dual-kV data was possible both without and with the grid; the BMD error of the 400 mg/mL insert was -0.4 g/cm² without the grid and +0.3 g/cm² with the grid. The dual-layer FPD configuration required accurate scatter correction for DE decomposition: a -5% scatter estimation error resulted in -0.1 g/cm² BMD error for the 50 mg/mL insert and a -0.5 g/cm² BMD error for the 400 mg/mL with a grid, compared to <0.1 g/cm² for all inserts in a dual-kV setup with the same scatter estimation error. Conclusion: This comparative study of quantitative performance of dual-layer and dual-kV FPD-based DE imaging indicates the need for accurate scatter correction in the dual-layer setup due to increased susceptibility to scatter errors in the LE channel.

This study explores candidate metrics for use in a vendor-neutral Quality Control (QC) protocol for dual energy based (‘spectral’) contrast-enhanced digital mammography (DE-CEDM). The physical metrics were chosen for their anticipated link with clinical image quality. An extra criterion was that limiting values can be applied to these metrics. The following DE-CEDM characteristics were tested: (1) signal produced by test object inserts versus their iodine concentration for different breast tissue thicknesses and compositions, the possibility of iodine quantification and iodine detectability thresholds; (2) normal breast tissue cancellation for different types of simulated tissue compositions; (3) artefacts; (4) image uniformity; (5) exposure time and (6) mean glandular dose (MGD). The tests employed custom made and commercially available phantoms based on breast-tissue equivalent materials and iodine inserts. Initial test results on a specific DE-CEDM application (SenoBright HD, GE Healthcare) found a linear response to iodine concentration that was largely independent of background composition, with slopes in the range of 30 – 43 for iodine pixel value as a function of iodine concentration, making absolute iodine quantification possible. The residual signal due to normal tissue in the subtracted image was equivalent to less than 0.3 mg/cm² of iodine. Artefacts were quantified and considered acceptable. For a 5 cm thick breast, the system operates at a total MGD of 1.92 mGy with a total exposure time of 4.25 s. The preliminary QC test results are ready for discussion in the broader community of radiologists, medical physicists and manufacturers.

The presence of round cystic and solid mass lesions identified at mammogram screenings account for a large number of recalls. These recalls can cause undue patient anxiety and increased healthcare costs. Since cystic masses are nearly always benign, accurate classification of these lesions would be allow a significant reduction in recalls. This classification is very difficult using conventional mammogram screening data, but this study explores the possibility of performing the task on dual-energy full field digital mammography (FFDM) data. Since clinical data of this type is not readily available, realistic simulated data with different sources of variation are used. With this data, a deep convolutional neural network (CNN) was trained and evaluated. It achieved an AUC of 0.980 and 42% specificity at the 99% sensitivity level. These promising results should motivate further development of such imaging systems.

Low x-ray dose is desirable in x-ray computed tomographic (CT) imaging due to health concerns. But low dose comes with a cost of low signal artifacts such as streaks and low frequency bias in the reconstruction. As a result, low signal correction is needed to help reduce artifacts while retaining relevant anatomical structures. Low signal can be encountered in cases where sufficient number of photons do not reach the detector to have confidence in the recorded data. X-ray photons, assumed to have Poisson distribution, have signal to noise ratio proportional to the dose, with poorer SNR in low signal areas. Electronic noise added by the data acquisition system further reduces the signal quality. In this paper we will demonstrate a technique to combat low signal artifacts through adaptive filtration. It entails statistics-based filtering on the uncorrected data, correcting the lower signal areas more aggressively than the high signal ones. We look at local averages to decide how aggressive the filtering should be, and local standard deviation to decide how much detail preservation to apply. Implementation consists of a pre-correction step i.e. local linear minimum mean-squared error correction, followed by a variance stabilizing transform, and finally adaptive bilateral filtering. The coefficients of the bilateral filter are computed using local statistics. Results show that improvements were made in terms of low frequency bias, streaks, local average and standard deviation, modulation transfer function and noise power spectrum.

In this work, we investigate and develop a method for cross-section image reconstruction from data collected over limited-angular ranges in the context of human-limb imaging. We first design a convex optimization program with constraints on directional image total-variations (TVs), and then tailor a convex primal-dual algorithm, which is referred to as the directional TV (DTV) algorithm, for solving this program. By using the proposed DTV algorithm, we investigate image reconstructions in studies with data collected from numerical thigh phantoms over a limited-angular range of 60^◦. The results of the numerical studies demonstrate that the method proposed can yield, from limited-angular-range data, cross-section images with significantly reduced artifacts that are observed otherwise in images obtained with existing algorithms.

In digital breast tomosynthesis (DBT), in-plane images are of clinical utility, whereas images within transverse planes contain significant artifacts simply because the existing algorithms are not designed for reconstructing accurately images within transverse planes from extremely limited-angular-range data. In this work, we investigate and develop a convex primal-dual (CPD) algorithm that incorporates directional total-variation (DTV) constraints for yielding breast images within transverse planes with substantially reduced artifacts when images are reconstructed from DBT data. We have performed numerical studies to demonstrate that the algorithm proposed has the potential to yield breast images with substantially reduced artifacts within transverse planes observed in images obtained otherwise with existing algorithms.

In general X-ray radiography, inconsistency of brightness and contrast in initial presentation is a common complaint from radiologists. Inconsistencies, which may be a result of variations in patient positioning, dose, protocol selection and implant could lead to additional workflow by technologists and radiologists to adjust the images. To tackle this challenge posed by conventional histogram-based display approach, an AI Based Brightness Contrast (AI BC) algorithm is proposed to improve the consistency in presentation by using a residual neural network trained to classify X-ray images based on N by M grid of brightness and contrast combinations. More than 30,000 unique images from sites in US, Ireland and Sweden covering 31 anatomy/view combinations were used for training. The model achieved an average test accuracy of 99.2% on a set of 2700 images. AI BC algorithm uses the model to classify and adjust images to achieve a reference look and then further adjust to achieve user preference. Quantitative evaluation using ROI based metrics on a set of twelve wrist images showed a 53% reduction in mean pixel intensity variation and a 39% reduction in bone-tissue contrast variation. A study with application specialists adjusting image presentation of 30 images covering 3 anatomies (foot, abdomen and knee) was performed. On average, the application specialists took ~20 minutes to adjust the conventional set, whereas they took ~10 minutes for the AI BC set. The proposed approach demonstrates the feasibility of using deep learning technique to reduce inconsistency in initial display presentation and improve user workflow.

As the conventional method of geometric calibration not only needs extra-scan but also have the vulnerability to the dimensional accuracy of the dedicated phantom, we proposed a new geometric calibration scheme based on data redundancy which uses projection data of any arbitrary patient already scanned, mainly focused on a geometry of Varian’s OBI system. Using the fan-beam based data redundancy which can be applied to cone-beam projection data of either full-fan or half-fan geometry, one could successfully detect the misalignments of geometric parameters and find the values of them. The simulation studies using the XCAT numerical phantom were conducted to check the feasibility of data-redundancy-based method can detect the misalignment of geometric parameters and find true values.

Dermatoscopes are commonly used to evaluate skin lesions. There are a wide array of medical imaging devices entering the market, some of which allow patients to analyze skin lesions themselves. These devices usually come in the form of smartphone attachments that leverage smartphone optics to acquire images; and in some cases, even give a preliminary diagnosis. While these attachments are useful, smartphone sensors are small which limits the extent and detail of captured images as opposed to images from a professional camera. Our work focuses on the information lost due to the known limitations of smartphone sensors, and its effect on image appearance. This analysis has been performed using a virtual simulation pipeline for dermatology, VCT-Derma, which contains simulated skin and dermatoscope models, among others. We discuss the necessary sensor parameters to adapt the dermatoscope model to various sensors, and with the help of the skin model and a colorgauge chart, obtain images from the sensors simulated. Results indicate differences in image clarity as well as observed color fidelity between the reference dermatoscope and smartphone sensors. Results of imaging the skin model show improved feature clarity in the reference device image as compared to the two smartphone sensors. Results of imaging the colorgauge chart show average ΔE2000 values of ~12.5 across all color patches for the reference device, and smartphone sensors. Under the same lightning, smartphone sensors showed areas with saturated pixels, as opposed to the reference device. Research is ongoing on the influence of multispectral illumination on these sensors.

Volumetric segmentation using deep learning is a computationally expensive task, but one with great utility for medical image analysis in radiology. Deep learning uses the process of convolution to calculate voxel level relationships and predict class membership of each voxel (e.g. segmentation). We hypothesize that (1) kidney segmentation in cadaveric, non-contrast enhanced CT images is possible; (2) a volumetric UNet (VNet) architecture will out-perform a 2D UNet architecture in kidney segmentation; and (3) as increasing anatomically relevant information present within the volumes will increase the ability of the system to understand the relationship of anatomical structures, thus enabling more accurate segmentation. In this project we utilized a difficult dataset (cadaveric, non-contrast enhanced CT data) to determine how much anatomical information is necessary to obtain a quantifiable segmentation with the lowest Hausdorff Distance and highest Dice Coefficient values between the output and the ground truth mesh. We used a 70/20/10% training testing and validation split with a total N of 30 specimens. In order to test the anatomical context required to properly segment structures we evaluated and compared the performance of four separate segmentation models: (1) a 2D UNet model that pulled random cross sections from the volumes for training; (2) a 2D UNet model that had the training samples augmented with 3D perturbations for more anatomical context; (3) a 3D VNet with volumetric patching and a padded border to protect against edge artifacts; and (4) a 3D VNet with volumetric patching and image compression by ½ the volume with the padded border. Our results show that as anatomical context in the image or volume increases, segmentation performance also improves.

The X-ray source using ZnO nanowires as the cold cathode is first densely integrate a large number of X-ray sources into flat panel devices. Compared with traditional ray sources, X-ray sources device based on ZnO nanowires have small size, instantaneous response, and can be electronically controlled. We propose a new conformal transmission imaging system that uses this flat-panel ray source to irradiate objects with multiple points to complete extremely low-dose projection. In order to get the final image, it is necessary to process the original projection with an image restoration method to remove the projection aliasing caused by multiple point sources. Due to the addition of iteration in the image restoration process, the error caused by the geometric parameters will gradually become larger as the iteration progresses, resulting in distortion of the final image. Therefore, geometric correction of the imaging system is required. The new imaging system has a unique geometric relationship. Based on this geometric relationship, this paper proposes a geometric correction method applied to the new imaging scheme. In this method, the angle parameters in the new imaging system are estimated using a single point volume model. It only needs one projection of the phantom, and it can derive the analytical formula about the three angle errors, which can avoid falling into a local minimum. In this paper, we give the simulation results of the calibration method, which proves the validity and accuracy of the method.

In this study, we propose a method to reduce streak artifacts of sparse view CT images via convolutional neural network (CNN). The main idea of the proposed method is to utilize both image and sinogram domain data for CNN training. To generate datasets, projection data were acquired from 512 (128) views using Siddon’s ray-driven algorithm, and full (sparse) view CT images were reconstructed by filtered back projection with a Ram-Lak filter. We first trained U-net based CNN_img, which was designed to reduce the streak artifacts of sparse view CT in image domain. Then, the output images of CNN_img were used as prior images to conduct pseudo full view sinogram. Before upsampling, sparse view sinogram was normalized by the prior images, and then linear interpolation was employed to estimate the missing view data compared to full view sinogram. The upsampled data were denormalized using prior images. To reduce the residual errors in pseudo full view sinogram data, we trained CNN_hybrid with residual encoder-decoder CNN, which is known to be effective in reducing the residual errors while preserving structural details. In order to increase the learning efficiency, the dynamic range of the pseudo full view sinogram data was converted via exponential function. The results show that the CNN_hybrid provides better performance in streak artifacts reduction than CNN_img, which is also confirmed by quantitative assessment.

Cardiovascular disease is a leading cause of death worldwide. Among all cardiovascular diseases, coronary artery disease (CAD) is a major contributor to mortality, accounting for approximately 10% of deaths per annum. Imaging techniques such as computed tomography (CT) can provide information on the location, composition, and severity of disease within the coronary arteries. However, cardiac motion makes protocol optimization for coronary CT angiography (CCTA) image acquisition and reconstruction uniquely challenging. The clinical trials necessary to optimize these protocols require a cohort of diverse patients, clinical staff, and are time-consuming. Consequently, there is a need for safer, less expensive, and faster alternatives. A virtual imaging trial (VIT) enables studies on the optimization of CCTA protocol parameters in a controlled (simulated) environment. In this work, we developed a framework to integrate computational models of CAD with computational, anthropomorphic phantoms (XCAT) and a CT simulator (DukeSim) so that VITs can be performed with these tools to optimize CCTA protocols for clinical tasks. To demonstrate the framework, we included a calcification in the right coronary artery of a 50th percentile BMI male XCAT phantom. The calcification was subjected to cardiac motion (60, 90, 120 BPM) and underwent simulated imaging with and without contrast enhancement (simulating a 370 mgI/mL injection) and was reconstructed via filtered back-projection (FBP) with two kernels (B26f, B70f). Measurements of the right coronary artery under imaging indicated that the small calcification was readily detected below 90 BPM, with no iodinated contrast media, and reconstructed via FBP with the B70f kernel. As demonstrated, this new VIT framework can provide an efficient means or CCTA protocol optimization.

We propose a learning-based method to synthesize high-resolution (HR) CT images from low-resolution (LR) CT images. A self-super-resolution framework with cycle consistent generative adversarial network (CycleGAN) is proposed. As an ill-posed problem, recent super-resolution methods rely on the presence of external/training atlases to learn the transform LR images to HR images, which is often not available for CT imaging to have high resolution for slice thickness. To circumvent the lack of HR training data in z-axis, the network learns the mapping from LR 2D transverse plane slices to HR 2D transverse plane slices via CycleGAN and inference HR 2D sagittal and coronal plane slices by feeding these sagittal and coronal slices into the trained CycleGAN. The 3D HR CT image is then reconstructed by collecting these HR 2D sagittal and coronal slices and image fusion. In addition, in order to force the ill-posed LR to HR mapping to be close to a one-to-one mapping, CycleGAN is used to model the mapping. To force the network focusing on learning the difference between LR and HR image, residual network is integrated into the CycleGAN. To evaluate the proposed method, we retrospectively investigate 20 brain datasets. For each dataset, the original CT image volume was served as ground truth and training target. Low-resolution CT volumes were simulated by downsampling the original CT images at slice thickness direction. The MAE is 17.9±2.9 HU and 25.4±3.7 HU for our results at downsampling factor of 3 and 5, respectively. The proposed method has great potential in improving the image resolution for low pitch scan without hardware modification.

The feasibility of acquiring multi-energy CT data through slow modulation of the kVp as an alternative to photon-counting detectors (PCDs) is currently under exploration. A low kVp-switching rate can be enabled with a conventional CT system but raises challenges due to missing sinogram views. Our previous work used a CNN-based method for sinogram completion by generating full-sampled images from undersampled sinograms, providing an acceptable image quality at a 22°/kVp switching rate. The purpose of this study was to investigate a GAN-based spectral sinogram completion method for enabling a lower kVp switching rate. A Pix2Pix GAN model with paired undersampled sinogram of 45° or 120° projections/kVp and its corresponding full-sampled sinogram was implemented and trained. The completed data was subsequently used to perform sinogram domain material decomposition. Our results on a simulated FORBILD abdomen phantom dataset showed that the GAN-based method can further lower the kVp switching rate to 45° projections/kVp. The proposed GAN-based sinogram completion method facilitates slow-kVp switching acquisitions and thus further relaxes hardware requirements.

Image quality estimation is crucial in modern CT with noise magnitude playing a key role. Several methods have been proposed to estimate noise surrogates in vivo. This study aimed to ascertain the accuracy of three different noise-magnitude estimation methods. We used ensemble noise as the ground truth. The most accurate approach to assess ensemble noise is to scan a patient repeatedly and assess the noise for each pixel across the ensemble of images. This process is ethically undoable on actual patients. In this study, we surmounted this impasse using Virtual Imaging Trials (VITs) that simulate clinical scenarios using computer-based simulations. XCAT phantoms were imaged 47 times using a scanner-specific simulator (DukeSim) and reconstructed with filtered back projection (FBP) and iterative (IR) algorithms. Noise magnitudes were calculated in lung (ROIn), soft tissues (GNI), and air surrounding the patient (AIRn), applying different HU thresholds and techniques. The results were compared with the ensemble noise magnitudes within soft tissue (En). For the FBP-reconstructed images, median En was 30.6 HU; median ROIn was 46.6 HU (+52%), median GNI was 40.1 HU (+31%), and median AIRn 25.1 HU (-18%). For the IR images, median En was 19.5 HU; median ROIn was 31.2 HU (+60%), median GNI was 25.1 HU (+29%), and median AIRn 18.8 HU (-4%). Compared to ensemble noise, GNI and ROIn overestimate the tissue noise, while AIRn underestimates it. Air noise was least representative of variations in tissue noise due to imaging condition. These differences may be applied as adjustment or calibration factors to better represent clinical results.

Background: With new advances in CT energy-integrating (EID) and photon-counting (PCD) detector technology, the spatial resolution of CT has significantly improved from 0.50 mm to less than 0.25 mm isotropic. Our goal was to explore the image quality improvements that are associated with UHR CT, specifically in imaging and characterization of atherosclerotic plaques. Methods: We imaged seven excised human hearts, with a known history of atherosclerosis, on an EID-CT scanner equipped with a UHR comb. The scans were performed at 120 kVp and 211 mAs, with 1-sec gantry rotation time. The images were reconstructed twice; first with a standard resolution (SR) reconstruction kernel (Uq36u) at 0.5x0.5x3.0 mm³ voxel size, and second with a UHR kernel (Uq77u) at 0.2x0.2x1.5 mm³. We measured calcium volume, Agatston score, and number of lesions with dense calcification (HU>1000). We propose a multi-resolution visualization scheme in which smooth soft tissue features are derived from the SR image, and the sharp hard plaque features blended from the UHR images. Results: We detected a total of 105 lesions in the seven hearts. The UHR images had significantly reduced blooming artifacts which leads to higher average HU values and smaller lesion volumes. The calcium volume for UHR was 67% of the SR volumes (r²= 0.81). Agatston scores were also systematically lower in UHR compared to SR 41% (r²= 0.81). In addition, we found 102 lesions with dense calcification compared to 4 in SR. Conclusion: UHR-CT can significantly reduce blooming artifacts and partial volume effects in atherosclerotic plaque imaging.

In dental CT, the presence of metal objects introduces various artifacts caused by photon starvation and beam hardening. Although several metal artifacts reduction methods have been proposed, they still have limitations in terms of reducing the metal artifacts. In this work, we proposed a method to reduce the metal artifacts with convolutional neural network (CNN). The proposed method is comprised of two steps. In STEP 1, we acquired a more accurate prior image, which is used in normalized metal artifact reduction (NMAR) technique through the CNN. The metal artifacts in output image from STEP 1 are reduced by CNN training, which provides more accurate prior images. In STEP 2, the NMAR is conducted with the acquired prior image from CNN result. To validate the proposed method, we used dental CT images containing metals and without metal to evaluate that the proposed method could significantly reduce the metal artifacts compared to the NMAR method.

Pediatric imaging utilizes the quantitative capabilities of CT to guide clinical decision making and treatment, but image quality is heavily affected by variation in patient sizes and the need for lower dose scans. Dual energy CT generates spectral results such as virtual monoenergetic images (VMI), electron density (ED), and effective atomic number (Zeff) that enhance material characterization and quantification. Though it has not been extensively explored, application of DECT to pediatric imaging may allow increased stability in quantitative measures with varying patient size, dose, and tube voltage. To examine the dependency of size, dose, and tube voltage, a phantom with tissue-mimicking inserts was scanned with dual-layer spectral detector CT with different extension rings to simulate different pediatric patient sizes. Each size configuration was subsequently scanned at CTDIvol of 9, 6, and 3 mGy with 100 and 120 kVp to obtain conventional CT and spectral results. Overall, both VMI and ED values were accurately quantified. VMI at 70 keV and 9 mGy demonstrated smaller differences among patient size and kVp compared to conventional images. Low dose dependency relative to 9 mGy was also present for VMI. Similarly, ED and Zeff showed low dependency on patient size, dose, and kVp and maintained material differentiability. Stability of these spectral results with different patient sizes, doses, and tube voltages illustrates the potential application of spectral detector CT to pediatric patients not only to improve the consistency of quantitative measures across patient sizes but also to allow lower doses without impairing quantification.