This PDF file contains the front matter associated with SPIE Proceedings Volume 11600 including the Title Page, Copyright information, and Table of Contents.

Welcome and Introduction to SPIE Medical Imaging conference 11600: Biomedical Applications in Molecular, Structural, and Functional Imaging

The Alzheimer’s Disease Neuroimaging Initiative (ADNI) seeks to detect Alzheimer’s Disease (AD) as early as possible in the preclinical phase of AD through the use of Neuroimaging based biomarkers. Earlier research found that the subarachnoid space is impacted in Alzheimer’s Disease. This investigation seeks to establish a cross-sectional or longitudinal relationship between extra-axial cerebrospinal fluid (EACSF) volume and an AD diagnosis, as EACSF has been overlooked thus far. An automated approach was employed to compute the segmentation of MRI images in AutoEACSF; subsequent manual quality control filtered for adequate brain mask and EACSF segmentations. Crosssectional data analysis compared EACSF volumes for the cognitively normal (CN) and AD diagnosis and found no significant difference (p = 0.7827). In the cross-sectional general linear model, an association was found between EACSF and ventricle volume (p = 0.0005), and entorhinal volume (p = 0.040). Fusiform volume, mid temporal volume, and hippocampal volume were deemed non-significant contributors to the model. The longitudinal model included subjects with more than 2 successful scans and excluded those with implausible negative changes in EACSF volume across time. This analysis included two additional categories: mild cognitive impairment (MCI) and patients who changed diagnosis (change). Between the AD, CN, MCI, and change groups, there was a statistically significant difference between stable MCI and change subjects longitudinally after age and sex were added to the model; no difference was found between any combination of the other groups. Overall, we found minimal evidence that EACSF relates cross-sectionally or longitudinally to AD or progression.

Although volumetric assessment of intracerebral hemorrhage (ICH) plays a key role for clinicians to make optimal treatment decisions and predict prognosis of ICH patients, qualitative assessment of neuroradiologists in reading brain CT images is not accurate and has large interreader variability. To overcome this clinical challenge, this study develops and tests a new interactive computer-aided detection (ICAD) tool to quantitatively assess hemorrhage volumes. A retrospectively assembled dataset including 200 patients with ICH was collected for this study. After loading each case, the ICAD tool first segments intracranial brain volume, performs CT labelling of each voxel, then contour-guided image-thresholding techniques based on CT Hounsfield Unit is used to estimate and segment hemorrhage-associated voxels (ICH). Next, two experienced neurology residents examine and corrects the markings of ICH categorized into either intraparenchymal hemorrhage (IPH) or intraventricular hemorrhage (IVH) to obtain the true markings. Additionally, volumes and maximum two-dimensional diameter of each sub-type of hemorrhage were also computed for understanding ICH prognosis. The performance to segment hemorrhage regions between semi-automated ICAD and the verified neurology residents’ true markings was evaluated using dice similarity coefficient (DSC). Data analysis results show that median and [interquartile range] of DSC are 0.96 [0.91, 0.98], 0.97 [0.93, 0.99], 0.92 [0.83, 0.97] for ICH, IPH and IVH, respectively. Thus, this study demonstrates that the new ICAD tool enables to segment and quantify ICH and other hemorrhage volume with higher DSC, which has potential to quantify ICH in future clinical practice.

Magnetic resonance spectroscopy (MRS) is a non-invasive imaging tool for detecting and quantifying metabolites in vivo. However, the spectral data often suffer from poor signal-to-noise ratio (SNR) due to low concentrations of metabolites and limited acquisition time. In this work, we introduce a deep learning-based method for denoising MRS spectra. By identifying and characterizing the sparse representations in the feature vector domain of the noise in MRS data using the deep learning network, the non-signal components can be removed to improve the SNR of spectral data. We used a stack auto-encoder (SAE) network to train the deep learning-based model based on high SNR data collected from a brain phantom using a high number of signal average (NSA=192). To overcome overfitting, the SAE network is trained in a patch-based manner. The network and denoising method were then tested using the noisy data collected in the same sample volume but using low NSA (NSA=8) with much shorter acquisition time. The denoising performance of the reported method was then evaluated using statistical comparison analysis of spectra before and after being denoised and the paired “ground truth” high NSA spectra based on the measurements of SNR, and mean squared error (MSE) and known metabolite levels. for the results of testing and evaluating the reported method showed that the SNR improved by 40% and the MSE reduced by 72% in the data collected from the brain phantom, while, the SNR improved by 47% and the MSE reduced by 27% in the data collected from the human subjects. In summary, the reported deep-learning method demonstrated a model-free approach to enhance SNR of noisy MRS data that otherwise cannot be used for quantitative analysis. The potential applications of include using low NSA or short data collection time to accelerate MRS exams while maintaining adequate spectroscopic information for detection and quantification of the metabolites of interest when a little time is available for an MRS exam in the clinical setting.

Amyloid-beta imaging with 11C-PiB shows a robust increase in cortical binding in Alzheimer’s Disease (AD). Standardized uptake value ratio (SUVR) and distribution volume ratio (DVR) were shown to be effective quantitative measures in discriminating healthy and AD subjects. However, these measures typically require long wait/scan time and optimal selection of the reference region. The aim of this study is to propose a novel measure, which has superior diagnostic performance than conventional measures and has a less complicate scan workflow. We used ¹¹C-PiB datasets that is publicly available from OASIS project. A total number of 50 dynamic scans are included in this study, of which 25 are from AD and 25 are from heathy control (HC) subjects. For all subjects, we first quantified standardized uptake value ratio (SUVR, 50-60 mins) and distribution volume ratio (DVR, T*=30 min) in frontal cortex region. Reference region was selected as whole cerebellum. We then defined a new measure, Peak-Clearance-Rate (PCR) which calculated the uptake decay rate in early stage of the scan (from the peak uptake time to 20 min). The new measure, SUVR and DVR all can differentiate the HC and AD subjects (p<0.05). ROC analysis was performed and indicated that PCR has better performance (AUC=0.912) than DVR (AUC=0.846) and SUVR (AUC=0.843). Cohen’s size for PCR is 2.53, while the ones for SUVR and DVR are 1.79 and 1.75. The proposed measure is an alternate measure of diagnosis in AD with ¹¹C-PiB imaging. Additionally, it only requires first 20-min scan and does not require the procedure of delineation of reference tissue region. Further investigation will focus on staging, specifically in differentiating mild cognitive impairment (MCI) subjects from aged healthy and AD subjects.

We propose a method for exploring non-linear directed information transfer in complex systems, which we expect to be useful for analyzing functional MRI (fMRI) neuroimaging data. In contrast to existing approaches that attempt to analyze complex systems by simplification and subsequent analysis of a simplified system, we propose to retain the original system complexity during the analysis. To this end, we introduce largescale Non-Linear Granger Causality (lsNGC) as a method for effective connectivity network analysis in high-dimensional systems with short time-series. By introducing a dimension reduction step into a non-linear timeseries prediction approach, lsNGC aims at directed, non-linear, multivariate time-series causality analysis in large complex networks, such as brain activity in fMRI analysis. We quantitatively evaluate the performance of lsNGC in computer simulations on structural recovery of synthetic networks with known ground truth. We find that our method performs better than a widely used non-linear network analysis method (Convergent Cross Mapping – CCM) with high statistical significance (p<10⁻⁶ ). In addition, as an outlook to possible clinical application, we perform a preliminary qualitative analysis of connectivity matrices for fMRI data of Autism Spectrum Disorder (ASD) patients and typical controls, using a subset of 59 subjects of the Autism Brain Imaging Data Exchange II (ABIDE II) data repository. Our results suggest that lsNGC, by extracting sparse connectivity matrices, may be useful for network analysis in complex systems, and may be applicable to clinical fMRI analysis in future research, such as targeting disease-related classification or regression tasks on clinical data.

We introduce large-scale Augmented Granger Causality (lsAGC) as a method for connectivity analysis in complex systems. The lsAGC algorithm combines dimension reduction with source time-series augmentation and uses predictive time-series modeling for estimating directed causal relationships among time-series. This method is a multivariate approach, since it is capable of identifying the influence of each time-series on any other time-series in the presence of all other time-series of the underlying dynamic system. We quantitatively evaluate the performance of lsAGC on synthetic directional time-series networks with known ground truth. As a reference method, we compare our results with cross-correlation, which is typically used as a standard measure of connectivity in the functional MRI (fMRI) literature. Using extensive simulations for a wide range of time-series lengths and two different signal-to-noise ratios of 5 and 15 dB, lsAGC consistently outperforms cross-correlation at accurately detecting network connections, using Receiver Operator Characteristic Curve (ROC) analysis, across all tested time-series lengths and noise levels. In addition, as an outlook to possible clinical application, we perform a preliminary qualitative analysis of connectivity matrices for fMRI data of Autism Spectrum Disorder (ASD) patients and typical controls, using a subset of 59 subjects of the Autism Brain Imaging Data Exchange II (ABIDE II) data repository. Our results suggest that lsAGC, by extracting sparse connectivity matrices, may be useful for network analysis in complex systems, and may be applicable to clinical fMRI analysis in future research, such as targeting disease-related classification or regression tasks on clinical data.

Normal and aberrant cognitive functions are the result of the dynamic interplay between large-scale neural circuits. Describing the nature of these interactions has been a challenging task yet important for neurodegenerative disease evolution. The origin of Alzheimer’s disease lies in the hippocampus and subsequently diffuses to the temporal, parietal and prefrontal cortices. Determining the sources of dementia is crucial to the prediction of the disease evolution and choice of treatment. State-of-the-art method for determining dementia progression are network diffusion models derived from the heat equation without diffusion sources. We propose a different research avenue based on epidemic modeling to localize the disease sources. These models may better characterize the empirical spread of dementia through brain regions. We explore an estimation algorithm based on a susceptible-infected (SI) epidemic algorithm and a network diffusion model for comparison purposes emulating the disease evolution from sources (susceptible) to non-recovered (atrophy, infected) areas. The goal is to identify the probable disease diffusion sources, which we accomplish via a ranking heuristic based upon steady-state convergence times. Graph centrality measures are employed to provide a baseline for further comparison. Our results applied on structural brain networks in dementia suggest that epidemic models are able to accurately describe the different node roles in controlling trajectories of brain networks comparably to the existing diffusion approach.

There is a need to lower mass of scintillators in PET imagers. However, the tradeoff between scintillator thickness and axial field of view is not obvious, particularly if the total mass of the scintillator is a limitation. In this work, we developed fast analytical methods to assess performance of PET as a function of scintillator thickness. Calculation of the photopeak detection efficiency (PDE) is complicated by the fact that most incident 511 keV gamma rays first undergo Compton scattering in the scintillator resulting in a partial deposition of energy, as well as the production of a lower energy secondary gamma ray. The PDE is dependent on scintillator geometry and source position and must be recalculated when either changes. We compare Monte Carlo and our analytical lower and upper bound estimation of PDE for thin slab LSO scintillators as a function of its thickness assuming normal incidence. We show that that our analytical method achieves excellent agreement with the results obtained by time-consuming Monte Carlo approach. This is important because application of our fast method enables preliminary system optimization prior to time-consuming, complex Monte Carlo modeling.

Magnetic resonance spectroscopic imaging (MRSI) has been widely used for studying metabolic alterations in brain-related pathologies, especially due to its non-invasiveness. Even though some software for MRSI data analysis have been developed, only a few are used by biomedical researchers and in a clinical setting routine, as their use still poses a challenge in keeping the trade-off between information content and ease of implementation. Aiming to increase MRSI analysis automation, our study proposes an open-source toolbox for analysis, automatic spectra quality control and MRSI data visualization. The proposed toolbox allows the visual inspection of all spectra. It makes possible the automated selection of spectra of interest using two different approaches: by clustering them using Pearson’s correlation coefficient or by discarding spectra based on spectral quality metrics. Using the magnetic resonance imaging (MRI) content, the toolbox provides information about the region from which the MRSI grid were acquired, such as the brain tissue ratio in each voxel: white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF). Once spatial and spectral information is combined, spatial averaging over anatomically defined regions of interest (ROIs) can be applied, for instance, by averaging the spectra and fitting the result. The proposed toolbox aims to simplify and automate MRSI analysis, being easy to install and to use.

It has been shown in the literature that marijuana use is associated with changes in brain network connectivity. We investigate whether large-scale Extended Granger Causality (lsXGC) can capture such changes using restingstate fMRI. This method combines dimension reduction with source time-series augmentation and uses predictive time-series modeling for estimating directed causal relationships among fMRI time-series. It is a multivariate approach, since it is capable of identifying the interdependence of time-series in the presence of all other timeseries of the underlying dynamic system. Here, we investigate whether this model can serve as a biomarker for classifying marijuana users from typical controls using 126 adult subjects with a childhood diagnosis of ADHD from the Addiction Connectome Preprocessed Initiative (ACPI) database. We use brain connections estimated by lsXGC as features for classification. After feature extraction, we perform feature selection by Kendall’s-tau rank correlation coefficient followed by classification using a support vector machine. As a reference method, we compare our results with cross-correlation, which is typically used in the literature as a standard measure of functional connectivity. Within a cross-validation scheme of 100 different training/test (90%/10%) data splits, we obtain a mean accuracy range of [0.714, 0.985] and a mean Area Under the receiver operating characteristic Curve (AUC) range of [0.779, 0.999] across all tested numbers of features for lsXGC, which is significantly better than results obtained with cross-correlation, namely mean accuracy of [0.728, 0.912] and mean AUC of [0.825, 0.969]. Our results suggest the applicability of lsXGC as a potential biomarker for marijuana use.

Objective: Our objective was to train machine-learning algorithms on hyperpolarized ³He magnetic resonance imaging (MRI) datasets to generate models of accelerated lung function decline in participants with chronic obstructive pulmonary disease (COPD). We hypothesized that hyperpolarized gas MRI ventilation, machine-learning and multivariate modelling could be combined to explain clinically relevant changes in forced expiratory volume in 1 sec (FEV₁) over a relatively short, three year time period. Methods: Hyperpolarized ³He MRI was acquired using a coronal Cartesian FGRE sequence with a partial echo and segmented using a k-means cluster algorithm. A maximum entropy mask was used to generate a region of interest for texture feature extraction using a custom-built algorithm and PyRadiomics platform. Forward logistic-regression and principal-component-analysis were used for feature selection. Ensemble-based and single machine-learning classifiers were utilized; accuracies were evaluated using a confusion-matrix and area under the curve (AUC) of a sensitivityspecificity plot. Results: We evaluated 42 COPD patients with three year follow-up data, 27 of whom (9 Females/18 Males, 66±7 years) reported negligible changes in FEV₁ and 15 participants (5 Females/10 Males, 71±8 years) reported worsening FEV₁ greater than -5%pred, 30±8 months later. We generated a predictive model to explain FEV₁ decline using bagged-trees trained on four texture features which correlated with FEV₁ and FEV₁/FVC (r=0.2-0.5; p<0.05) and yielded a classification accuracy of 85%. Conclusion: For the first time, we have employed hyperpolarized ³He MRI ventilation texture features and machine learning to identify COPD patients with accelerated decline in FEV₁ with 84% accuracy.

Peripheral Artery Diseases (PAD) are caused by the occlusions of arteries in the peripheral locations of the circulatory system. The severity of PAD is usually assessed using the Ankle Brachial Index (ABI) and the Ultrasound Doppler. Non-contact Photoplethysmography (PPG) imaging is a recent emerging technology capable of monitoring skin perfusion. Using an off-the-shelf camera and a light source, is possible to remotely detect the dynamic changes in blood volume in the skin and derive a map correlated to the blood perfusion. The aim of this study is the evaluation of a PPG imaging system (iPPG) for the assessment of Peripheral Arterial Diseases. Reduced blood flow is simulated on 21 volunteers by increasing the pressure in a pressure cuff. For each volunteer, measurements with iPPG, ultrasound, Laser Speckle Contrast Analysis (LASCA) and ABI were acquired. Our experiments show that iPPG can detect reduced perfusion levels, and correlates well with the other measurement systems.

Noninvasive biomarkers that capture the adenocarcinoma aggressiveness could provide crucial quantitative information for precision medicine to aid clinical decision making. Texture features are known to measure tumor heterogeneity and have been identified as the features having a potential correlation to outcomes in lung adenocarcinomas. Nevertheless, current methods for analyzing texture patterns that arise from local intensity variation are limited to reveal a spatial configuration of the texture structures in 3D thoracic CT images. This lack of an intuitive visualization of the texture of nodules makes understanding the meanings underlying the tumor heterogeneity a challenging problem. In this study, we propose an approach combining a structure-texture image decomposition with a topological data analysis to represent a spatial configuration of the texture of lung adenocarcinoma. The image decomposition aims to split the 3D thoracic CT image into two components, namely, the structure component with the piecewise-smooth part having the global structural information of nodule and the texture component with the locally-patterned oscillating part. We demonstrate the use of topological data analysis to capture architectural features that arise from the texture component. Specifically, using persistent homology of texture components, we compute topological representations of lung adenocarcinomas with the appearance of consolidation on CT images. Applying the method to an example of early-stage lung adenocarcinomas graded with texture features based on the popular algorithm such as gray-level co-occurrence matrix (GLCM), we present that the structure-texture image decomposition model with topological data analysis might be a promising tool in analyzing the tumor heterogeneity in 3D thoracic CT images.

Objective: We aimed to develop a linear-systems model to evaluate the impact of mathematical assumptions related to multi-compartment ¹²⁹Xe MRI methods. We employed a linear-systems approach to fill a knowledge gap associated with a single-point Dixon approach currently used for magnetization calculations and physiologic estimates. Methods: We generated a linear-systems model for magnetization calculations needed to generate multi-compartment ¹²⁹Xe MRI data. 1Lung tissue and red-blood-cell compartments were isolated by acquiring data 90° out of phase and aligned with perpendicular quadrature channels. Results: In our linear-systems model, a single-lobe sinc-pulse in the time domain was used to excite ¹²⁹Xe atoms in the tissue and red blood-cell compartments. We assumed that T2* exponential decay in the time domain was equivalent to convolution with a complex Lorentzian function in the frequency domain, and the gradient echo envelope was equivalent to convolution with a rectangular pulse with phase shifting. A rectangular window function and Dirac comb sampling in the time domain modeled analog-to-digital recording. Fast Fourier transforms were modelled by Dirac combs in both time and frequency domains. To account for non-uniform sampling, k-space was re-sampled using a unique sampling function convolution followed by Cartesian sampling. Phase inhomogeneities were corrected using reference gas and spectroscopy data. Conclusion: The proposed linear-system analysis provides a framework for modelling decay constants, peak overlap, and magnetization evolution common in multi-compartment ¹²⁹Xe MRI. Understanding the spectral properties of ¹²⁹Xe MRI will provide a way to identify novel compartments and their abnormalities in diverse pulmonary diseases.

We are developing imaging methods for the preclinical arm of a co-clinical trial investigating synergy between immunotherapy and radiotherapy. We perform in vivo micro-CT of mouse lungs to detect lung metastasis after treatment. This work explores deep learning (DL) as a fast and accurate approach to lung nodule segmentation. We examine the performance of DL lung tumor detection using realistically simulated nodules inserted into temporally-resolved real micro-CT datasets. Our simulations suggest that DL is a promising approach for fast, accurate segmentation of lung nodules in mice.

Upper airway segmentation in static and dynamic MRI is a prerequisite step for quantitative analysis in patients with disorders such as obstructive sleep apnea. Recently, some semi-automatic methods have been proposed with high segmentation accuracy. However, the low efficiency of such methods makes it difficult to implement for the processing of large numbers of MRI datasets. Therefore, a fully automatic upper airway segmentation approach is needed. In this paper, we present a novel automatic upper airway segmentation approach based on convolutional neural networks. Firstly, we utilize the U-Net network as the basic model for learning the multi-scale feature from adjacent image slices and predicting the pixel-wise label in MRI. In particular, we train three networks with the same structure for segmenting the pharynx/larynx and nasal cavity separately in axial static 3D MRI and axial dynamic 2D MRI. The visualization and quantitative results demonstrate that our approach can be applied to various MRI acquisition protocols with high accuracy and stability.

Shape analysis is an important and powerful tool in a wide variety of medical applications. Many shape analysis techniques require shape representations which are in correspondence. Unfortunately, popular techniques for generating shape representations do not handle objects with complex geometry or topology well, and those that do are not typically readily available for non-expert users. We describe a method for generating correspondences across a population of objects using a given template. We also describe its implementation and distribution via SlicerSALT, an open-source platform for making powerful shape analysis techniques more widely available and usable. Finally, we show results of this implementation on mouse femur data.

Non-small cell lung cancer (NSCLC) is one of the leading causes of death worldwide. Medical imaging is used to determine cancer staging; however, these images may hold additional information which could be utilized to aid in outcome prediction. A multi-modality radiomics approach incorporating quantitative and qualitative features from the tumor and its surrounding regions, along with clinical features, has yet to be explored. Therefore, we hypothesize that a model containing CT and PET radiomic features, in addition to clinical and qualitative features, has the potential improve risk-stratification of NSCLC patients better than cancer stage alone. Our dataset consisted of 135 NSCLC patients (training: n=94, testing: n=41) who underwent surgical resection. Each region of interest was segmented using a semi-automatic approach on both the pre-treatment CT and PET images. Radiomic features were extracted using the Quantitative Image Feature Engine. A total of 1030 features were extracted including clinical, qualitative, and radiomic features. LASSO regression was used to identify the top features to predict time to recurrence in the training cohort and the model was evaluated in the testing cohort. A total of nine features were selected, including two clinical, one CT, and six PET radiomic features. The model achieved a concordance of 0.81 in the training cohort, which was validated in the testing cohort (concordance=0.79) and outperformed stage alone (concordances=0.68-0.69). This model has the potential to assist physicians in risk-stratifying patients with NSCLC and could be used to identify patients that may benefit from more aggressive or personalized treatment options.

Motion management is of utmost importance to ensure the effectiveness of radiation treatment for lung cancer patients. The emerging 4D radiotherapy technique aims to actively manage lung tumor motion by providing a time-dependent treatment. This requires the delineation of tumors on all the 3D phase CT images in 4D CT, which is labor intensive and time consuming. In this study, we propose a novel deep learning-based method to automatically localize and segment lung tumor on 4D CT to facilitate the clinical workflow of 4D radiotherapy. The proposed network, named motion Regionbased Convolutional Neural Networks (R-CNN), consists of four stages, i.e., feature extraction, rough tumor location, fine tumor location, and segmentation within tumor region-of-interest (ROI). To aid the network to get rid of unreasonable detected ROIs, different from traditional mask R-CNN, our proposed method first fed 4D CT with consecutive phases into the backbone to extract tumor motion information and then utilized these motion information to estimate global and local deformation vector fields (DVFs), which is useful for measuring the movement of the tumor ROI between each two phases. Our method was tested on 20 patients’ lung 4D CT images in this study. Five metrics were used for quantitative evaluation, i.e., center-of-mass distance (COMD), Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD₉₅), mean surface distance (MSD), and volume difference (VD) calculated between the manual tumor contour and the contour obtained by our method. Averaged over the 20 cases, our method yielded 1.14±0.95 mm on COMD, 0.76±0.25 on DSC, 2.29±1.31 mm on HD95, 0.75±0.44 mm on MSD, and 0.51±0.51 cc on VD. These results have demonstrated the feasibility and efficacy of our motion R-CNN method for automatic localization and segmentation of lung tumors on 4D CT images. Our method is also expected to be applicable to in-treatment real-time volumetric imaging to provide 3D markerless tumor tracking during treatment delivery.

Cellular spheroid is a complex aggregate of cells, and imaging is an important step in understanding how cell behavior operates and give references about possibilities according to the cells in the spheroid formation. Identifying structures in these images is manual and time-consuming and have a high rate of variability inter experts. An automated identification can solve these problems. The aim of this work is to present a study of Convolutional Neural Network (CNN) applied to live cells identification in NIH-3T3 spheroid. Four different CNN architectures are exploited in this paper: AlexNet, ResNet18, GoogLeNet, and VGG 16 with batch normalization. Many experiments were performed to get the best architecture involving data augmentation, hyperparameter tuning, and transfer learning using ImageNet. The experiments identify up to five different structures in a spheroid image, where the AlexNet achieved the best performance considering the F1-score as the evaluation metric. The use of CNN for this kind of identification opens the possibility of following the spheroid’s behaviour when cultured in more complex images.

COVID-19 is a highly contagious infectious disease that has infected millions of people worldwide. Polymerase Chain Reaction (PCR) is the gold standard diagnostic test available for COVID-19 detection. Alternatively, medical imaging techniques, including chest X-ray (CXR), has been instrumental in diagnosis and prognosis of patients with COVID-19. Enabling the CXR with machine learning-based automated diagnosis will be important for rapid diagnosis of the disease by minimizing manual assessment of images by the radiologists. In this work, we developed a deep learning model that utilizes the transfer learning approach using a pre-trained Residual Network model. The Residual Network 50 (ResNet50) is trained from scratch by utilizing the initial architecture and pre-trained weights to provide the classification results. Two types of classification (two-class and three-class) is performed using the developed model. A cascaded approach is adopted for two-class classification where the classification is performed in two phases. The dataset used for training and evaluating the model comprises of 8,254 images in total out of which 1651 images were considered for testing the cascaded model (15 COVID-19) and three-class classification (51 COVID-19). The model was evaluated using accuracy, sensitivity, specificity, and F1-score metrics. Our cascaded model yielded an accuracy of 91.8% for classification of abnormal and normal cases and 97.9% for the classification of pneumonia and COVID-19 images. In the three-class classification, our model reported an accuracy of 92% in classifying normal, pneumonia (bacterial and viral) and COVID-19 cases.

As a mainstay of metabolic studies, Positron Emission Tomography (PET) has aroused remarkable attention in the clinical arena and the translational realm. The amount of radiotracer dosage is amongst the major problems in PET imaging, creating ongoing challenges for both the clinical community and the preclinical researchers. In quest of generating diagnostic quality PET images in extremely low-dose conditions, several deep-learning(DL)-inspired methods have sprung up in human imaging over the past few years. Propelled by the successful application of DL techniques in human studies and the unique advantages of deep neural networks in learning specific features, we have investigated a fully 3D U-Netlike model which enables reconstructing standard-dose PET dataset directly from its low-dose equivalent. We verified the performance of the method both in mice and rat PET scans through calculating image evaluation metrics such as RMSE, PSNR, and SSIM. Our measurements revealed that the proposed method could provide high-quality PET scans with improved noise properties in low-dose rodent studies.

In comparison to two-dimensional (2D) ultrasound (US), three-dimensional (3D) US imaging is a more sensitive alternative for monitoring the size and shape of neonatal cerebral lateral ventricles. It can be used when following posthemorrhagic ventricular dilatation, after intraventricular hemorrhaging (IVH), which is bleeding inside the lateral ventricles of the brain in preterm infants. Tracking ventricular dilatation is important in neonates as it can cause increased intracranial pressure, leading to neurological damage. However, manually segmenting 3D US images is time-consuming and tedious due to poor image contrast and the complex shape of cerebral ventricles. In this paper, we describe an automated segmentation method based on the U-Net model for the segmentation of 3D US images that may contain one or both ventricle(s). We trained and tested two models, a 3D U-Net and slice-based 2D U-Net, on a total of 193 3D US images (105 one ventricle and 88 two ventricle images). To mitigate the class imbalance of the object vs. background, we augmented the images through rotation and translation. As a benchmark comparison, we also trained a U-Net++ model and compared the results with the original U-Net. When all the images were used in a single U-Net model, the 3D U-Net and 2D U-Net yielded a Dice similarity coefficient (DSC) of 0.67±0.16 and 0.76±0.09 respectively. When two 2D U-Net models were trained separately, they yielded a DSC of 0.82±0.09 and 0.74±0.07 for one ventricle and two ventricle images, respectively. Compared to the best previous fully automated method, the proposed 2D U-Net method reported a comparable DSC when using all images but an increased DSC of 0.05 when using only one ventricle image.

We propose a learning-based method to automatically segment prostate and its dominant intraprostatic lesions (DILs) from positron emission tomography (PET)/computed tomography (CT) images. A dual attention mask R-CNN is introduced to enable end-to-end segmentation. To avoid the effect of useless region, mask R-CNN is used to get rid of non-tumor regions via fist locate the tumor region-of-interest (ROI) and then segment tumor within that ROI. Dual attention networks are used as backbone in mask R-CNN to extract comprehensive features from both CT and PET images. The binary mask of tumor of an arrival patient’s CT and PET image is generated by the well-trained network. To evaluate the proposed method, we retrospectively investigate 25 PET/CT datasets. On each dataset, prostate and DILs were delineated by physicians and was served as ground truth and training target. The proposed method was trained and evaluated by a Five-fold cross validation strategy. The average centroid distance, volume difference and DSC value for prostate/DIL among all 25 patients are 0.83±0.91mm, -0.01±0.79 and 0.84±0.09, respectively. The proposed method has great potential in improving the efficiency and mitigating the observer-dependence in prostate and DIL contouring for DIL focal boost radiation therapy

We introduce a method for tracking results and utilization of Artificial Intelligence (tru-AI) based on machine learning applications in medical imaging, for analyzing pandemic-induced effects on healthcare systems. By tracking both large-scale utilization and AI results data, the tru-AI approach can establish surrogates for measuring the amount of care provided and estimate the prevalence of certain disease conditions under unusual circumstances, such as pandemic outbreaks. To quantitatively evaluate our approach, we analyzed service requests for automatically identifying intracranial hemorrhage (ICH) on head CT using a commercial AI solution (Aidoc, Tel Aviv, Israel). This software is typically used for AI-based prioritization of radiologists’ reading lists for reducing turnaround times in patients with emergent clinical findings, such as ICH or pulmonary embolism. Imaging data is anonymized, uploaded to a cloud-based inference machine in real time, and AI-based ICH detection results are returned. We recorded N = 3,084 emergency-setting non-contrast head CT studies at a major US healthcare system during two observation periods, namely (i) a pre-pandemic epoch (January 1–31, 2020) and (ii) after the Covid-19 outbreak (March 15 – April 30, 2020). Although daily counts of unique imaged patients were significantly lower during (37.9 ± 7.6) than before (42.0 ± 6.2) the Covid-19 outbreak, we found that ICH was more likely to be observed during than before the Covid-19 outbreak (p<0.05). Our results suggest that, by tracking both large-scale utilization and AI results data, the tru-AI approach can contribute clinical value as an exploratory tool, aiming at a better understanding of pandemic-related effects on healthcare.

Attenuation correction (AC) is important for an accurate interpretation and quantitative analysis of SPECT myocardial perfusion imaging. Dedicated cardiac SPECT systems have invaluable efficacy in the evaluation and risk stratification of patients with known or suspected cardiovascular disease. However, most dedicated cardiac SPECT systems are standalone, not combined with a transmission imaging capability such as computed tomography (CT) for generating attenuation maps for AC. To address this problem, we propose to apply a conditional generative adversarial network (cGAN) for generating attenuation-corrected SPECT images (SPECT_GAN ) directly from non-corrected SPECT images (SPECT_NC) in image domain as a one-step process without requiring additional intermediate step. The proposed network was trained and tested for 100 cardiac SPECT/CT data from a GE Discovery NM 570c SPECT/CT, collected retrospectively at Yale New Haven Hospital.The generated images were evaluated quantitatively through the normalized root mean square error (NRMSE), peak signal to noise ratio (PSNR), and structural similarity index (SSIM) and statistically through joint histogram and error maps. In comparison to the reference CT-based correction (SPECT_CTAC ), NRMSEs were 0.2258±0.0777 and 0.1410±0.0768 (37.5% reduction of errors); PSNRs 31.7712±2.9965 and 36.3823±3.7424 (14.5% improvement in signal to noise ratio); SSIMs 0.9877±0.0075 and 0.9949±0.0043 (0.7% improvement in structural similarity) for SPECT_NC and SPECT_GAN , respectively. This work demonstrates that the conditional adversarial training can achieve accurate CT-less attenuation correction for SPECT MPI, that is quantitatively comparable to CTAC. Standalone dedicated cardiac SPECT scanners can benefit from the proposed GAN to reduce attenuation artifacts efficiently.

Our ultimate goal is to predict adverse cardiovascular events from CT calcium score exams using radiomics. Unlike the traditional whole heart Agatston score, our approach analyzes individual calcifications, creating a challenge for accurate/reproducible assessments. Optional preprocessing consisted of 3D blind deconvolution optimized on masked sharp structures, giving PSFs similar to measured ones, followed by noise reduction. In our experiment, we found that blind deconvolution improves the accuracy and reproducibility of CT calcium score evaluation. It is beneficial to use this to predict adverse cardiovascular event since it provides a robust and accurate evaluation of clinically relevant and pre-clinical calcifications. We conclude that for archived images where PSFs are unavailable, 3D blind deconvolution is a useful preprocessing step for improved radiomics assessments of CT-calcium-score images.

Fractional Flow Reserve (FFR) is a widely used metric to quantify the functional significance of stenoses in coronary arteries. FFR is the ratio of pressure before and after a stenosis and is measured using a transducer during coronary catheterization. To avoid unnecessary catheterization, many analytical and data-driven models based on non-invasive imaging have been proposed for FFR estimation. In this study, we construct physics-informed analytical models and datadriven machine learning models for FFR estimation based on CT-derived information. All four models require simple information about suspect stenoses, offering rapid, practical approaches for functional assessment. The four models we study are: (1) a patient-specific blood flow informed pressure drop model based on Navier-Stokes equations, (2) a purely geometric model based on stenosis area reduction, (3) a Gaussian process regression model trained on patient specific stenosis geometry and blood flow data, and (4) a Gaussian process regression model trained only on patient specific stenosis geometry. The models were developed and tested using a simulation study with ground truth FFR values from computational fluid dynamics analysis of blood flow through a population of stenosed arteries. In total, 60 different stenosis conditions based on known prevalence were simulated with a range of measurement errors leading to 10000 data sets. The RMSE of the model estimates for approach 1 through 4 are, 0.19, 0.45, 0.06, 0.14. The flow informed machine learning model leads to ~1% lower bias and ~12% lower variance than the flow informed analytical model. Considering the improved variance performance, the machine learning models likely outperform analytic expressions because they learn optimal regression associations robust to noise. This work suggests that machine learned approaches may be superior to conventional analytic expressions for FFR estimation, particularly when inputs contain realistic measurement error.

Cardiac motion quantification in magnetic resonance (MR) images provides vital information to diagnose and evaluate cardiovascular diseases. Motion quantification can be obtained from routinely acquired MR images. However, the methods available for motion estimation present many sources of inconsistencies, thus creating constraints to use it as a reliable diagnostic technique. Recently, convolutional neural networks (CNNs) have demonstrated to be a powerful tool for many different imaging tasks, including optical flow estimation, a technique widely used for motion estimation. In this work, we evaluate the suitability of a compact and powerful CNN architecture based on Pyramid, Warping, and Cost Volume (PWC) for motion estimation in synthetic cardiac resonance images. The synthetic images were generated using the extended cardiac-torso (XCAT) and MRXCAT software, which generates temporal series of highly detailed MR images and their corresponding ground-truth motion field, which would be impossible to obtain in real-life data. The CNN training was unsupervised, simulating real data. The ground-truth provided by the synthetic images was used to evaluate the PWC performance, determining its reliability. The CNN achieved an average end-point-error of 0.61 ± 0.25 pixel and a mean absolute error of 0.38 ± 0.15 pixel in the test set, surpassing state-of-the-art methods. The results obtained in this work indicate a high potential of the unsupervised PWC network for future applications in real cardiac images.

In this paper we propose a deep learning framework to estimate pressure from 4D flow MRI. Pressure drop is an important parameter to detect and diagnose different cardiovascular diseases. Accurate estimation of pressure from 4D flow MRI is hampered however due to noise and low resolution of 4D flow data. In the proposed method we consider the pressure estimation as a mapping function between velocity to pressure and employ an encoder-decoder based deep network for the mapping. A computational fluid dynamic model was designed which identically matched the geometry of a stenotic flow phantom used in 4D Flow MRI experiments and velocity and pressure data was simulated for 1000 different flow conditions to train the network. In addition, the proposed network was tested on real in -vitro 4D flow MRI in the same stenotic model for 3 different flow rates. Estimated pressures from the network showed excellent agreement with the reference CFD simulated pressures. As measure of fidelity, relative pressure drop across the stenosis was computed between the reference pressure and estimated pressure and were compared with the simplified Bernouli method. It was determined that the pressure drop estimation by the proposed method is more accurate than competing method.

Complex hemodynamics as those related to carotid stenosis are not always straightforward due to multifaceted challenges presented by the collateral flow in the Circle of Willis (CoW) and brain flow autoregulation. Advanced computational and benchtop methods to investigate hemodynamics aspects related to such complex flows are often used. In this study we investigated these aspects by performing correlated computational fluid dynamics (CFD) simulations and benchtop experiments in patient specific 3D printed phantoms. Initial results present challenges in the use of these methods. Ongoing work is being done to address phantom design and proper boundary conditions in CFD simulation.

Accurate landmark localization in 3D medical volumes is nowadays needed for many medical and technical applications such as segmentation, detection of anatomical structures, image registration, and more. Due to the time-consuming process of manual landmarking and the variety of landmark positioning by individual observers, a more uniform and automatic solution is needed. This paper examines a state-of-the-art approach based on a patch-based Convolutional Neural Network (CNN), that is able to predict multiple landmarks simultaneously, and its performance on ill-posed landmarks. It combines regression and classification in one framework to improve the localization accuracy of individual landmarks by taking into consideration the relationship between an image patch and an anatomical landmark position. During inference, image patches are passed to the CNN as input, which are then moved by an amount, specified by the regression output, in the direction of highest probability, which is specified by the classification output. One of the main challenges here was that there are no salient landmarks in the aorta, they are all just generic points in a tube. This adds variability in the manual positioning of the landmarks and even the predictions of the network for each volume. The network’s prediction accuracy was quantitatively evaluated and the best setup achieved an average landmark localization error of 16.04mm, with 99.0% of all predictions having a mean error smaller than 30 mm, while the best prediction during evaluation had a mean error of 3.53 mm over all landmarks.

Scintigraphy is a common nuclear medicine method to image molecular target’s bio-distribution and pharmacokinetics through the use of radiotracers and gamma cameras. The patient’s images are obtained by using a pair of opposing large flat gamma ray detectors equipped with parallel-hole lead or tungsten collimators that preferentially detect gamma-rays that are emitted perpendicular to the plane of the detector. The resulting images form an anterior/posterior (A/P) planar image pairs. The obtained images are contaminated by noise and contain artifacts caused by gamma-ray attenuation, collimator penetration, scatter and other detrimental factors. Post-filtering of the images can reduce the noise, but at the cost of spatial resolution loss, and cannot remove any of the aforementioned artifacts. In this study, we introduced a new image reconstruction-based method to recover a single corrected planar scintigraphic patient image corrected for attenuation, system spatial resolution and collimator penetration, using the A/P image pair (two conjugated views) as data. To accomplish this task, we used a system model based on the gamma camera detectors physical properties and applied regularization method based on sparse image representation to control noise while preserving spatial resolution. In this proof-of-concept study, we evaluated the proposed approach using simple numerical phantoms. The images were evaluated for simulated lesions images contrast and background variability. Our initial results indicate that the proposed method outperforms the conventional methods. We conclude, that the proposed approach is a promising methodology for improved planar scintigraphic image quality and warrants further exploration.

Dental caries remains the most prevalent chronic disease in both children and adults. Optical coherence tomography (OCT) is a noninvasive optical imaging modality utilized to image oral samples to diagnose carious lesions, but detecting early stage dental caries with high-level accuracy remains challenging. Deep learning models have been employed to classify OCT images for various healthcare applications. In this paper, human tooth specimens were imaged ex vivo using OCT imaging systems, and a three-class grading system based on deep learning model for detection and classification of carious lesions was developed. This study is a step forward in the development of automated deep learning/OCT imaging system for early dental caries diagnosis.

X-ray fluorescence computed tomography (XFCT) is a molecular imaging technique of x-ray photons, which can be used to sense different elements or nanoparticle (NP) agents inside deep samples or tissues. XFCT has been an active research topic for many years. However, XFCT has not been a popular molecular imaging tool because it has limited molecular sensitivity and spatial resolution. To further investigate XFCT imaging, we present a benchtop XFCT imaging system, in in which a unique pencil beam x-ray source and a ring of x-ray spectrometers were simulated using GATE (Geant4 Application for Tomographic Emission) software. An accelerated majorization minimization (MM) algorithm with an L¹ regularization scheme was used to reconstruct the XRF image of Molybdenum (Mo) NP targets from the numerical measurements of GATE simulations. With a low x-ray source output rate, good target localization was achieved with a DICE coefficient of 83.681%. The reconstructed signal intensity of the targets was found to be relatively proportional to the target concentrations if detector number and placement is optimized. The MM algorithm performance was compared with maximum likelihood expectation maximization (ML-EM) and filtered back projection (FBP) algorithms. In the future, the benchtop XFCT imaging system will be tested experimentally.

In recent years, the segmentation and projection techniques of different structures of medical interest have had significant growth due to its usefulness; doctors have been using them as tools for the diagnosis and evolution of different diseases. The segmentation of the Inner Limiting Membrane (ILM) in retinal scans acquired using the Optical Coherence Tomography (OCT) imaging technique has generated particular interest in the medical area since it provides clinically relevant information about diseases such as Glaucoma, Diabetic Macular Edema (DME) or Multiple Sclerosis. Furthermore, the generation of a surface that shows the current morphological situation of the scanned retinal area is a tool that complements the medical analysis. In this paper, a new methodology for the ILM segmentation on OCT retinal images and a surface projection from different axially spaced scans acquired over the macular and peripapillary zone is presented. The proposed scheme consists of a wavelet-based denoising step and a contrast enhancement stage using Eigenvalues of Hessian matrix, while the segmentation process is based on the Canny edge detection algorithm; these stages are applied to each image of a C-scan for a later surface generation using Cubic spline interpolation. This method is applied to a publicly available OCT data-set composed of 22 patients with several retinal diseases obtaining a correct individual segmentation of each image, while the surface generation results demonstrate high performance in the visualization of the ILM morphology, which can be used for dimensional analysis of this membrane.

Tissue simulating optical multilayer phantoms that closely mimic optical properties of specific tissues are used to evaluate and calibrate spectral imaging studies in clinical diagnostics. Consistency in phantom layer thickness, fabricated through the spin coating method is crucial in using the methodology to create large datasets to facilitate artificial intelligence (AI) based optical diagnostic systems. These phantoms are usually imaged under a scanning electron microscope to verify layer dimensions, varying with the amount of used materials, changes in sample parameters, and spin speed. Imaging a considerable number of phantoms under SEM for different applications is challenging. This paper addresses the automated detection of layered skin phantom thickness acquired from light microscopy images. The thickness estimation reported in this work is also useful in optimizing the sample and spin coating parameters to achieve a predefined sample dimension during the fabrication procedure.

Non-invasive assessment of tissue perfusion has gained increased interest in the recent past for monitoring various physiological dysfunctions. Laser speckle imaging and associated processing techniques present useful and inexpensive methods of achieving this goal. However, differentiating the effects of static and dynamic scatterers from the captured speckle image is challenging and limits the sensitivity of such techniques. This paper presents a fractal analysis in the temporal domain to assess the various dynamic scattering features of tissue perfusion demonstrated on a perfusion model. We use detrended fluctuation analysis for the quantitative assessment of image fractal dimensions, which reflect the concentration and flow features of the moving scatterers in the perfusion model.

We developed a Bessel-beam photoacoustic microscopical simulation platform by using the k-Wave: MATLAB toolbox. The simulation platform uses the ring slit method to generate Bessel beam. By controlling the inner and outer radius of the ring slit, the depth-of-field (DoF) of Bessel beam can be controlled. And the large volumetric image is obtained by point scanning. The simulation experiments on blood vessels was carried out to demonstrate the feasibility of the simulation plat-form. This simulation work can be used as an auxiliary tool for the research of Bessel-beam photoacoustic microscopy.

As a preclinical imaging modality, bioluminescence tomography (BLT) is designed to locate and quantify threedimensional (3D) information of viable tumor cells in a living organism non-invasively. However, because of the ill-posedness of the inverse problem of reconstruction, BLT is hard to achieve the accurate recovery of the distribution of light sources. In this study, we proposed a Gaussian weighted block sparse Bayesian learning strategy based on K-means clustering algorithm (GBSBLK) for accurate BLT reconstruction. GBSBLK integrated the structured sparsity assumption, the K-means clustering strategy, and the block sparse Bayesian learning (BSBL) framework to overcome the over-smoothness and over-sparsity in BLT reconstructions, and without using the tumor segmentation from anatomical images as a priori. To better define the structured sparsity, we used the K-means clustering algorithm to directly cluster all the mesh points to get the K blocks. Furthermore, to prevent from over-smoothness of the light source, we applied Gaussian weighted distance prior to build the intra-block correlation matrix. At last, we used the BSBL framework to ensure the accuracy and robustness of the backward iterative computation. Results of both numerical simulations and in vivo experiments demonstrated that GBSBLK achieved the accurate quantitative analysis not only in tumor spatial positioning but also morphology recovery. We believe that GBSBLK can achieve great benefit in the application of BLT for quantitative analysis.

X-ray luminescence computed tomography (XLCT) imaging is a hybrid molecular imaging modality combining the merits of both conventional x-ray imaging (high spatial resolution) and optical imaging (high measurement sensitivity). The narrow x-ray beam based XLCT imaging has been shown to be promising. However due to the selective excitation scheme, the imaging speed is slow thus limiting its practical applications for in vivo imaging. In this work, we have introduced a continuous scanning scheme to acquire data for each angular projection in one motion, eliminating the previous stepping scheme and reducing the data acquisition time, which makes it feasible for multiple transverse scans for three-dimensional (3D) imaging. We have introduced a high accuracy vertical stage to our focused x-ray beam based XLCT imaging system to perform high-resolution and 3D XLCT imaging. We have also included a scintillator crystal coupled to a PMT to act as a single-pixel detector for boundary detection purposes to replace our previous flat panel x-ray detector. We have verified the feasibility of our proposed scanning scheme and imaging system by performing phantom experimental studies. A phantom was embedded with a set of cylindrical targets with 200 μm edge-to-edge distance and was scanned in our imaging system with the proposed method. To test the feasibility for 3D scanning, we took measurements from 4 transverse slices with a vertical step size of 1 mm. The results of the experiments verified the feasibility of our proposed method to perform 3D XLCT imaging using a narrow x-ray beam in a reasonable time.

Osteoporosis is a common age-related disease associated with increased bone loss causing reduced bone strength and enhanced fracture-risk. Finite element (FE) modelling is used to estimate bone strength from high-resolution threedimensional (3-D) imaging modalities including micro-CT, MRI, and HR-pQCT. Emerging technologies of multi-row detector CT (MDCT) imaging offer spatial image resolution comparable to human trabecular thickness. However, at the current MDCT resolution regime, FE modelling based on segmented trabecular bone (Tb) microstructure suffers from noise and other imaging artifacts. In this paper, we present a bone mineral density (BMD)-adjusted FE modeling method of Tb microstructure from MDCT imaging without requiring Tb segmentation. The method spatially varies mechanical stiffness based on local ash-density estimated from MDCT-derived calcium hydroxyapatite (CHA) density and, thus, models the hypothesis that stress-flow is primarily absorbed by Tb microstructure as compared to marrow space under mechanical compression. Specifically, an MDCT-based linear FE analysis method was developed using a voxel-mesh and the above model of space-varying stiffness, and the performance of the method was examined. For FE analysis, an axial cylindrical image core of 8mm diameter from 4-6% of distal tibia was extracted after aligning the tibial bone axis with the coordinate z-axis of the image space. Intra-class correlation coefficient (ICC) of 0.98 was observed in a repeat MDCT scan reproducibility experiment using cadaveric distal tibia specimens (n = 10). Also, high linear correlation (r = 0.87) was found between von Mises stress values and MDCT based CHA at individual voxels supporting the central hypothesis of our method.

Dual-energy subtraction (DES) is a technique that separates soft tissue from bones in a chest radiograph (CR). As DES requires specialized equipment, we propose an automatic method based on convolutional neural networks (CNNs) to generate virtual soft tissue images. A dataset comprising 35 pairs of CR and its soft-tissue version split in training (28 image pairs) and testing (7 image pairs) sets were used with data augmentation. We tested two types of images: the lung region’s cropped image and the segmented lung image. The ribs suppression was treated as a local problem, so each image was divided into 784 patches. The U-Net architecture was used to perform bone suppression. We tested two types of loss functions: mean squared error (^Lmse) and L^sm, which combines L^mse with the structural similarity index measure (SSIM). Due to the patches overlapping, it was necessary to interpolate the gray levels on the reconstructed image from the predicted patches. Evaluations were based on SSIM and root mean square error (RMSE) over the reconstructed lung area. The combination that presented the best results used the loss L^sm and the segmented lung image as input to the U-Net (SSIM of 0.858 and RMSE of 0.033). We observed that the U-Net has poor performance when trained with cropped images containing all information from the chest cavity and how the loss using local information can improve CR rib bone suppression. Our results suggest that it is possible removing the rib bones accurately in CR using CNN and a patch-based approach.

Osteoporosis is a common bone disease that occurs when the creation of new bone does not keep up with the loss of old bone, resulting in increased fracture risk. Adults over the age of 50 are especially at risk and see their quality of life diminished because of limited mobility, which can lead to isolation and depression. We are developing a robust screening method capable of identifying individuals predisposed to hip fracture to address this clinical challenge. The method uses finite element analysis and relies on segmented computed tomography (CT) images of the hip. Presently, the segmentation of the proximal femur requires manual input, which is a tedious task, prone to human error, and severely limits the practicality of the method in a clinical context. Here we present a novel approach for segmenting the proximal femur that uses a deep convolutional neural network to produce accurate, automated, robust, and fast segmentations of the femur from CT scans. The network architecture is based on the renowned u-net, which consists of a downsampling path to extract increasingly complex features of the input patch and an upsampling path to convert the acquired low resolution image into a high resolution one. Skipped connections allow us to recover critical spatial information lost during downsampling. The model was trained on 30 manually segmented CT images and was evaluated on 200 ground truth manual segmentations. Our method delivers a mean Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95) of 0.990 and 0.981 mm, respectively.

Osteoporosis is an age-related disease associated with reduced bone density and increased fracture-risk. It is known that bone microstructural quality is a significant determinant of trabecular bone strength and fracture-risk. Emerging CT technology allows high-resolution in vivo imaging at peripheral sites enabling assessment of bone microstructure at low radiation. Resolution dependence of bone microstructural measures together with varying technologies and rapid upgrades in CT scanners warrants data-harmonization in multi-site as well as longitudinal studies. This paper presents an unsupervised deep learning method for high-resolution reconstruction of bone microstructure from low-resolution CT scans using GAN-CIRCLE. The unsupervised training alleviates the need of registered low- and high-resolution images, which is often unavailable. Low- and high-resolution ankle CT scans of twenty volunteers were used for training, validation, and evaluation. Ten thousand unregistered low- and high-resolution patches of size 64×64 were randomly harvested from CT scans of ten volunteers for training and validation. Five thousand matched pairs of low- and highresolution patches were generated for evaluation after registering CT scan pairs from other ten volunteers. Quantitative comparison shows that predicted high-resolution scans have significantly improved structural similarity index (p < 0.01) with true high-resolution scans as compared to the same metric derived from low-resolution data. Also, trabecular bone microstructural measures such as thickness and network area measures computed from predicted high-resolution CT images showed higher (CCC = [0.90, 0.84]) agreement with the reference measures from true high-resolution scans compared to the same measures derived from low-resolution images (CCC = [0.66, 0.83]).

Two quantitative methods were used to determine high-temporal resolution velocity changes throughout HSA image sequences, one based on movements of discrete radiopaque microspheres and the other on iodinated contrast injections. The chosen pump settings encompassed a wide range of flow situations, all of which produced unique velocity measurements over the flow duration as captured by the detector. Variations in velocity are illustrated due to both the cyclic action of the pump and various phantom morphologies. These HSA methods can be applied to derive a wide range of velocity distributions, which may be used to further quantify important hemodynamic parameters.

We compared steady and pulsatile CFD simulations with 4D Flow MRI in a phantom model of arterial stenosis and investigated the hemodynamic features distal to the occlusion. With CFD, we characterized the flow structure to estimate both the magnitude and degree of anisotropy of the turbulence-related indices including vorticity, wall shear stress, helicity, and swirling strength using Large-Eddy Simulation (LES) and Reynolds-Averaged Navier Stocks (RANS) method. The results revealed that the LES approach captures well both the overall and the detailed ow features in comparison to RANS predictions with the standard k-epsilon model when compared with 4D Flow MRI. Furthermore, the Q-criterion was employed to reveal the temporal evolution of the vortex ring which was observed distal to the stenotic narrowing. It was found that both RANS and LES simulations were highly accurate for wall shear stress when validated against 4D Flow MR Imaging.

Estimating myocardial blood flow (MBF) is essential for diagnosing and risk stratifying myocardial ischemia. Currently, positron emission tomography (PET) is a gold standard for non-invasive, quantitative MBF measurements. In this work, we compare our machine learning derived MBF estimates to PET derived estimates, and 2-compartmental model derived MBF estimates. Our best performing model (ensemble regression tree) had a root mean squared error (RMSE) of 0.47 ml/min/g. Comparatively, the compartmental model achieved an RMSE of 0.80 ml/min/g. Including CAD risk factors improved flow estimation accuracy for models that trained on feature selected TAC data and worsened accuracy for models that trained on PCA data. Overall, our machine learning approach produces comparable MBF estimations to verified DCE-CT and PET estimates and can provide rapid assessments for myocardial ischemia.

Radiation dose escalation to the dominant intraprostatic lesions in prostate cancer could improve tumor control. However, it is difficult to identify the dominant intraprostatic lesions on CT images. Multiparametric MRI has superior soft tissue contrast and is often used to detect the intraprostatic lesions. In this study, we developed a deep learning-based point cloud matching network to register the multiparametric MRI to the CBCT images for dominant lesion identification for prostate cancer radiotherapy. Prostate in both the CBCT and MRI was first automatically contoured and then meshed in to point clouds. A point cloud matching network was trained using point cloud pairs that were generated using finite element analysis. The trained network was able to perform MRI-CBCT prostate image registration with inherent biomechanical constraints. The mean and standard deviation of our method were 0.93±0.01, 1.66±0.10mm and 2.68±1.91mm for Dice similarity coefficient, mean surface distance and target registration error, respectively.

We propose a learning-based method to register 4D CBCT images from different respiratory phases. An unsupervised spatial transformation network (STN)-based deformable registration method is introduced for registering phases of 4D CBCT. Namely, no ground truth deformation vector filed (DVF) is needed during training. To avoid the scatter artifact effect while learning the trainable parameters, a structural similarity-based loss is used for supervision. To evaluate the proposed method, we retrospectively investigate 20 lung 4D CBCT datasets. Five-fold cross-validation was used to evaluate the proposed method. During each experiment, 16 4D CBCT datasets were used for training and the rest 4 4D CBCTs were used as testing. During training, each two phases from one patient were used as moving and fixed images. The average TRE is 1.67 ±3.30 mm. The proposed method has great potential in quantifying tumor trajectory for making clinical decision during radiation therapy.

We propose a learning-based method to synthesize contrast-enhanced MR from non-contrast MR images. Attention network is integrated into a cycle-consistent adversarial network (CycleGAN) framework, called Atten-CycleGAN, to learn a mapping between non-contrast MRI and contrast MRI. Attentional U-Net was used as generator of CycleGAN to force the trained model to extract informative that can represent the specific difference between non-contrast MRI and contrast MRI, namely, the tissue contrast difference. To evaluate the proposed method, we retrospectively investigate 274 brain MR datasets. Each dataset contains a contrast-enhanced MR volume and a non-contrast MR volume. The contrast-enhanced one was served as ground truth and training target, and the non-contrast MR is the input. The proposed method was trained from 224 patients and evaluated by a hold-out test strategy on 50 patients. The average normalized mean absolute error of the synthesized contrast MR is 0.035±0.004. The proposed method has great potential in accurately generating contrast MR and therefore bypassing the contrast administration step during MR scan.

The delineation of target and organs-at-risk (OARs) is a necessary step in radiotherapy treatment planning. The accuracy of the target and OAR contours directly affects the quality of radiotherapy plans. Manual contouring of OARs is the routine procedure at present, which, however, is very time-consuming and requires significant expertise, especially for those head-and-neck (HN) cancer cases, where OARs densely distribute around tumors with complex anatomical structures. In this study, we propose a deep learning-based fully automated delineation method, namely, mask scoring regional convolutional neural network (MS-RCNN), to obtain consistent and reliable OAR contours in HN CT. In the model, MR images were synthesized by a cycle-consistent generative adversarial network given CT images. A backbone network was utilized to extract features from MRI and CT independently. The high bony-structure contrast in CT and soft-tissue contrast in MRI are complementary in nature. Through combining those complementary contrasts, the accuracy of OAR delineation is expected to be improved. Due to the ability of various object detection and classification, ResNet 101 was used as backbone in MS-RCNN. Quantities including Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), mean surface distance (MSD), and residual mean square distance (RMS) were calculated to evaluate the performance of the proposed method. An average DSC, HD95, MSD and RMS of 0.78 (0.58 - 0.89), 4.88 mm (2.79 mm - 7.46 mm), 1.39 mm (0.69 mm - 1.99 mm), and 2.23 mm (1.30 mm - 3.23 mm), were respectively achieved across all of the 12 OARs by our proposed method. The proposed method is promising in facilitating auto-contouring for radiotherapy treatment planning.

The Data Storage for Computation and Integration (DSCI) proposes management innovations for web-based secure data storage, algorithms deployment, and task execution. Its architecture allows uploading, browsing, sharing, and task execution in remote computing grids. Here, we demonstrate the DSCI implementation and the deployment of Image processing tools (TMJSeg), machine learning algorithms (MandSeg, DentalModelSeg), and advanced statistical packages (Multivariate Functional Shape Data Analysis, MFSDA), with data transfer and task execution handled by the clusterpost plug-in. Due to its comprehensive web-based design, local software installation is no longer required. The DSCI aims to enable and maintain distributed computing and collaboration environment across multi-site clinical centers for the data processing of multisource features such as clinical, biological markers, volumetric images, and 3D surface models.

Accurately brain tumor segmentation is critical on treatment plan making and treatment outcome prediction. Manually segmenting tumor is tedious and time consuming. Therefore, developing a reliable and automatic brain tumor segmentation model is necessary. In this study, we developed a new multimodal weighted network (MW-Net), which fully utilizes the biological information from multiple modalities. Since the contribution from different modality is different, the relative weight in introduced into MW-Net and trained as the hyperparameter with other parameters in an end-to-end way. The 3D segmentation results can be directly obtained in testing stage. The experimental results showed MW-Net outperformed 3D-U-Net.

Lung cancer is the leading cause of cancer fatalities in the world. A recent trend has begun to focus on the idea of using bronchoscopy for early detection of suspect cancerous lesions developing along the airway walls. Because standard white-light bronchoscopy has insufficient sensitivity in locating suspect lesions, researchers are turning to the promising modality referred to as narrow band imaging (NBI). NBI bronchoscopy has the advantage of highlighting the blood vessels contained in the lung mucosa. Since cancer lesions tend to exhibit abnormal vessel growth, NBI bronchoscopy is able to highlight such lesions. Unfortunately, the task of locating lesions and their vessel patterns in an NBI bronchoscopy video stream proves to be tedious for the physician. We present automatic methods for enhancing and segmenting the major blood vessels depicted in NBI bronchoscopic video. Results with ground-truth data indicate that our methods can achieve superior results to a popular existing vessel-segmentation method. We also consider a preliminary application of deep learning to this task; while this approach gives low sensitivity compared to the other approaches, it achieves higher specificity and accuracy.

Post-traumatic Epilepsy is one of the common aftereffects of brain injury. This neurological disorder can persist throughout the lifetime of patients and impacts their quality of life significantly. Identification of markers that indicate the likelihood of developing PTE can help develop preventive care for subjects identified as at risk. Despite the relatively high prevalence of PTE, brain imaging-based biomarkers for the diagnosis of PTE are lacking. This is due in part to the heterogeneity of injury in traumatic brain injury patients. Here we investigate the use of structural and functional imaging features for training machine learning models. Recently, applying state of the art machine learning methods such as neural networks on clinical data to help diagnosis attract a considerable amount of attention. However, the choice of a good algorithm and subset of features is fundamental to achieve reliable classification. Our goal is to explore if different Machine Learning techniques could be leveraged for the early diagnosis of PTE. We compared four popular machine learning methods performance to predict PTE after brain injury (1) support vector machine and (2) random forest (3) fully connected neural network and (4) graph convolutional network. Our result demonstrates the advantage of using a combination of connectivity features (functional) and lesion volume (structural) in conjunction with a Kernel SVM approach in predicting PTE. We also shown using a feature reduction method such as principal component analysis (PCA) is more effective than penalizing the classifiers. This might be due to the limitation of penalized models for a framework where features are correlatedPost-traumatic Epilepsy is one of the common aftereffects of brain injury. This neurological disorder can persist throughout the lifetime of patients and impacts their quality of life significantly. Identification of markers that indicate the likelihood of developing PTE can help develop preventive care for subjects identified as at risk. Despite the relatively high prevalence of PTE, brain imaging-based biomarkers for the diagnosis of PTE are lacking. This is due in part to the heterogeneity of injury in traumatic brain injury patients. Here we investigate the use of structural and functional imaging features for training machine learning models. Recently, applying state of the art machine learning methods such as neural networks on clinical data to help diagnosis attract a considerable amount of attention. However, the choice of a good algorithm and subset of features is fundamental to achieve reliable classification. Our goal is to explore if different Machine Learning techniques could be leveraged for the early diagnosis of PTE. We compared four popular machine learning methods performance to predict PTE after brain injury (1) support vector machine and (2) random forest (3) fully connected neural network and (4) graph convolutional network. Our result demonstrates the advantage of using a combination of connectivity features (functional) and lesion volume (structural) in conjunction with a Kernel SVM approach in predicting PTE. We also shown using a feature reduction method such as principal component analysis (PCA) is more effective than penalizing the classifiers. This might be due to the limitation of penalized models for a framework where features are correlated.

Detecting and segmenting neoplasms are an important part of radiotherapy treatment planning, disease monitoring, and predicting patient outcome. In the brain, functional magnetic resonance imaging (MRI) like dynamic susceptibility contrast enhanced (DSC) MRI are important tools for disease monitoring post-surgical resection (post-op) and post-radiotherapy. However, the manual contouring of these neoplasms of the brain is time consuming, expensive, and introduces inter-observer variability. In this work, we propose to uDetecting and segmenting neoplasms are an important part of radiotherapy treatment planning, disease monitoring, and predicting patient outcome. In the brain, functional magnetic resonance imaging (MRI) like dynamic susceptibility contrast enhanced (DSC) MRI are important tools for disease monitoring post-surgical resection (post-op) and post-radiotherapy. However, the manual contouring of these neoplasms of the brain is time consuming, expensive, and introduces inter-observer variability. In this work, we propose to use a 3D Mask R-CNN method to automatically detect and segment the brain tumor bed contour for DSC MRI perfusion images. Sixteen patients’ perfusion sequence images, each with 60 time point volumes, were used in this study. Experimental results show that our proposed 3D Mask R-CNN method achieves an overall Dice similarity, precision, recall and volume of difference (VoD) were 82%±6%, 84%±3%, 89%±3% and 86%±3% respectively.

Current medical imaging uses MRI or CT images to diagnose tissue injuries. In addition to this classic procedure, there are also alternative technologies that have advantages against MRI or CT. These include electrical impedance tomography (EIT). With the help of EIT it is possible to obtain an initial screening of the body quickly and without a lot of hardware. Classical software-based methods of imaging reconstruction use a linear back projection or iterative approaches, such as Gauss-Newton algorithm. This paper introduces innovative approaches of artificial intelligence (AI) for imaging. For this purpose, extensive AI-based simulations with a Generative Adversial Network (GAN) are performed and the approaches are transferred to a gelatine phantom and to the human body within a small study.

We visualized the luminal structures generated by DCIS and UDH using a refraction-contrast X-ray CT technique, which provides a contrast close to stained tissue images, and showed that DCIS and UDH forms a bubble-like shape and a tube-like shape, respectively. To express the difference in the three-dimensional structure between these tissue clearly, the number of luminal spaces, luminal volume, luminal density, and path length of the luminal structures were introduced. As a result, it was found that DCIS has many smaller and shorter lumens than UDH, which can contribute to the development of 3D pathology.

Three-dimensional microstructural analysis of human infant, adult, and diseased lungs can help us understand lung growth and disease progression. Three-dimensional analysis of alveolar sacs and alveolar ducts is fundamental. However, it has not been done completely due to technical limitations. In this study, we will acquire ultra-high resolution 3D CT images by large-field synchrotron radiation micro-CT and reveal the 3D structure of alveolar septa and alveolar sacs by CT image analysis.

Aortic dissections (AD) are injuries of the inner vessel wall of the (human) aorta. As this disease poses a significant threat to a patient’s life, it is crucial to observe and analyze the progression of the dissection over the course of the disease. The clinical examinations are usually performed with the application of Computed Tomography (CT) or Computed Tomography Angiography (CTA), based on which, automated post-processing procedures would be beneficial for the management of critical pathologies. One of the main tasks during post-processing is aorta segmentation. Different methods have been developed for the segmentation of aorta, including the tracking methods, the active contour/surface methods and the deep learning methods. In this study, a method for the automatic segmentation of aorta and its branches from original thorax CT and CTA images is introduced. The aorta is segmented based on deep learning algorithm and afterwards the branches are tracked based on particle filter algorithm.