This PDF file contains the front matter associated with SPIE Proceedings Volume 12036, including the Title Page, Copyright information, Table of Contents, and Conference Committee listings

Cardiovascular disease is the leading cause of death and a significant contributor of health care costs. Noninvasive imaging plays an essential role in the management of patients with cardiovascular disease. Cardiac magnetic resonance (MR) can non-invasively assess heart and vascular abnormalities, including biventricular structure/function, blood hemodynamics, myocardial tissue composition, microstructure, perfusion, metabolism, coronary microvascular function, and aortic distensibility/stiffness. Its ability to characterize myocardial tissue composition is unique among alternative imaging modalities in cardiovascular disease. Advances in cardiac MR techniques, particularly faster image acquisition, quantitative myocardial tissue characterization, and image analysis have been critical to its growth. With recent advances in artificial intelligence, there are significant potential to improve various aspects of cardiac MR. In this presentation, I will review potentials and challenges of use of artificial intelligence in cardiac MR.

Graph topology inference in networks with co-evolving and interacting time-series is crucial for network studies. Vector autoregressive models (VAR) are popular approaches for topology inference of directed graphs; however, topology estimation becomes ill-posed in large networks with short time series. The present paper proposes a novel topology inference method for analyzing directed networks with co-evolving nodal processes that solves the ill-posedness problem. The proposed method, large-scale kernelized Granger causality (lsKGC), uses kernel functions to transform data into a low-dimensional feature space, solves the autoregressive problem in the feature space, and then finds the pre-images in the input space to infer the topology. Extensive simulations on synthetic datasets with nonlinear and linear dependencies and known ground-truth demonstrate significant improvement in the Area Under the receiver operating characteristic Curve (AUC) for network recovery compared to existing methods. Furthermore, tests on synthetic semi-realistic datasets from functional magnetic resonance imaging (fMRI) demonstrate significant improvement in the AUC for topology inference, enhancing the prior results of the best competing methods by 15.4 percent, which confirms the benefits of the proposed method as compared to existing literature.

Both internal carotid arteries (ICA) distribute blood via the anterior (ACA) and middle (MCA) cerebral arteries to the anterior part of the brain. Asymmetry of the pre-communicating part (A1-segment) of the ACA is common and is related to intracranial aneurysm formation. It is unknown if A1 asymmetry is also related to blood flow changes in the dominant A1 segment and the ipsilateral ICA and MCA. This study aims to relate artery diameters of both ICAs, M1- segments MCA and A1-segments ACA to blood-flow distribution in 10 subjects with symmetric A1s versus 10 with asymmetric A1s. Diameter measurements of the ICA (C3 and C7 segments), M1 and A1-segments were performed manually on the time-of flight magnetic resonance angiography (TOF-MRA) using an in-house developed tool. 4D phase-contrast MR imaging (PC-MRI)datasets were analyzed using CAAS software. The asymmetric group had on average 43% (range 26–84) asymmetry of the A1-segments and this asymmetry was directly related to the right-left flow difference (R²=0.890). The asymmetric group also had increased diameters and blood-flow for all ICA, M1 and A1 segments on the dominant A1 side. In conclusion, this preliminary study showed that asymmetry in A1 diameter is directly associated with increased blood-flow in the ICA, MCA and ACA on the dominant side. Our findings should be confirmed in a larger population which will also help to find an ideal cut-off in asymmetry diameter measurement that reflects a statistically significant difference in blood flow and velocity.

Accurate alignment of longitudinal diffusion weighted imaging (DWI) scans of a subject is necessary to investi- gate longitudinal changes in DWI-derived diffusion measures such as fractional anisotropy (FA), mean diffusivity (MD), and quantitative anisotropy (QA). Currently, studies investigating these changes in the context of repet- itive non-concussive head injuries (RHIs) perform pairwise rigid registration of all scans of a subject to the first scan or any other reference scan or template. Prajapati et.al1 show that this strategy of performing pairwise rigid registration lead to a discrepancy in the rigid transformations. To eliminate this discrepancy, they propose performing transitive inverse consistent rigid registration of the longitudinal scans, and they analyze the impact of this approach on the mean values of the local/regional estimates of these diffusion measures. In this work, we further analyze the impact of transitive inverse consistent rigid registration on the distributions (CDFs) of the local/regional estimates of diffusion measures. We identify the regions (among the 48 anatomically defined regions by the JHU DTI-based white matter atlas2,3) that show significant differences in the CDFs obtained using pairwise inverse consistent and transitive inverse consistent rigid registration by performing the two sided Kolmogorov-Smirnov(KS) hypothesis test. We find that for MD and QA, there are certain subjects that have five or more regions with significant differences in the CDFs. Further, these are the same subjects for which Prajapati et.al1 found regions with 2%-4% differences in the mean values of these diffusion measures. Thus, our results further strengthen the recommendation made by Prajapati et.al1 to employ transitive inverse consistent rigid registration when investigating local/regional longitudinal changes in diffusion measures.

In this work, we present CONTINUITY, a novel, open-source interactive computation and visualization tool for brain connectome data. The connectome processing pipeline performs surface based processing as the main mode of operation. The automated processing includes structural-to-diffusion image co-registration, surface reconstruction for subcortical structures, as well as fiber tractography. The tool supports 3 different probabilistic methods of tractography offered by the tractography frameworks in FSL, MRtrix and DIPY. All methods employ brain and subcortical surfaces as seeds to initialize the tractography algorithms. CONTINUITY implements a friendly Graphical User Interface (GUI) to make the workflow accessible for nontechnical users. Additionally, it offers the possibility to visualize the results of the brain connectome in several interactive plot types such as a hierarchical edge bundling circle plot and over 2D/3D brain templates. This visualization tool can also be applied to connectome matrices computed with other tools and pipelines.

The quantification of cerebrospinal fluid (CSF), specifically the extra-axial cerebrospinal fluid (EA-CSF), which is the CSF in the subarachnoid space surrounding the cortical surface of the brain, has recently been shown to play an important role in the neuropathology of autism spectrum disorder (ASD) in infants. While prior work addressed measuring the global volume of EA-CSF, there was no available tool that quantifies the local, anatomical distribution of the EA-CSF. A localized EA-CSF quantification would provide more accurate and interpretable measurements. In our recent work, we proposed such a local EA-CSF extraction by using a pipeline that combines probabilistic brain tissue segmentation, cortical surface reconstruction and streamline-based local EA-CSF quantification. Yet, that system had several shortcomings, in particular a lack of available software tools, as well as a quantification where EA-CSF portions are counted multiple times. The purpose of this article is to present a novel, graphical user interface based, publicly available software tool, called LocalEACSF, which allows the user to easily run an adapted version of this pipeline and provide a set of straightforward quality control visualizations to assess the quality of the EA-CSF quantification. This tool further adds improvements and optimizations to the prior assessment. The LocalEACSF tool allows neuroimaging labs to compute a local extraction of extra-axial CSF in their neuroimaging studies in order to investigate its role in normal and atypical brain development, without the need for extensive technical knowledge.

Brain age prediction based on functional magnetic resonance imaging (fMRI) data has the potential to serve as a biomarker for quantifying brain health. To predict the brain age based on fMRI data robustly and accurately, we curated a large dataset (n = 4259) of fMRI scans from seven different data acquisition sites and computed personalized functional connectivity measures at multiple scales from each subject’s fMRI scan. Particularly, we computed personalized largescale functional networks and generated functional connectivity measures at multiple scales to characterize each fMRI scan. To account for inter-site effects on the functional connectivity measures, we harmonized the functional connectivity measures in their tangent space and then built brain age prediction models on the harmonized functional connectivity measures. We compared the brain age prediction models with alternatives that were built on the functional connectivity measures computed at a single scale and harmonized using different strategies. Comparison results have demonstrated that the best brain age prediction performance was achieved by the prediction model built on the multi-scale functional connectivity measures that were harmonized in tangent space, indicating that multi-scale functional connectivity measures provided richer information than those computed at any single scales and the harmonization of functional connectivity measures in tangent space improved the brain age prediction.

It is a challenging research endeavor to infer causal relationships in multivariate observational time-series. Such data may be represented by graphs, where nodes represent time-series, and edges directed causal influence scores between them. If the number of nodes exceeds the number of temporal observations, conventional methods, such as standard Granger causality, are of limited value, because estimating free parameters of time-series predictors lead to underdetermined problems. A typical example for this situation is functional Magnetic Resonance Imaging (fMRI), where the number of nodal observations is large, usually ranging from 10² to 10⁵ time-series, while the number of temporal observations is low, usually less than 10³. Hence, innovative approaches are required to address the challenges arising from such data sets. Recently, we have proposed the large-scale Extended Granger Causality (lsXGC) algorithm, which is based on augmenting a dimensionality-reduced representation of the system’s state-space by supplementing data from the conditional source time-series taken from the original input space. Here, we apply lsXGC on synthetic fMRI data with known ground truth and compare its performance to state-of-the-art methods by leveraging the benefits of information-theoretic approaches. Our results suggest that the proposed lsXGC method significantly outperforms existing methods, both in diagnostic accuracy with Area Under the Receiver Operating Characteristic (AUROC = 0.849 vs. [0.727, 0.762] for competing methods, p < 10^-8), and computation time (3.4 sec vs. [9.7, 4.8 x 10³] sec for competing methods) benchmarks, demonstrating the potential of lsXGC for analyzing large-scale networks in neuroimaging studies of the human brain.

Inferring relationships among the elements in multivariate observational time-series data is challenging. Representing the interactions as graphs with edges and nodes can describe such relations. While the number of nodal observations in resting-state functional Magnetic Resonance Imaging (rs-fMRI) can rise up to millions of points, such as representing each voxel in a neuroimaging study, the number of temporal observations may remain scarce, leading to ill-posed problems in large-scale data. Here, we recently proposed a novel method for network connectivity analysis, large-scale Nonlinear Granger Causality (lsNGC), which combines the principle of Granger causality and nonlinear dimensionality reduction using Gaussian kernels leading to radial basis function neural networks for time-series prediction. In this study, we apply lsNGC on synthetic rs-fMRI data with known ground truth and compare its performance to competing state-of-the-art methods. We find that the proposed lsNGC method significantly outperforms the existing methods in accuracy, as measured by the Area Under the Receiver Operating Characteristic (AUROC, 0.867 ± 0.028), with p <10^-9 as compared to competing methods, thus quantitatively alarming the merits of lsNGC for the analysis of large-scale brain networks in neuroimaging studies.

Purpose: Intracerebral Hemorrhage (ICH) is one of the most devastating types of strokes with mortality and morbidity rates ranging from about 51%-65% one year after diagnosis. Early hematoma expansion (HE) is a known cause of worsening neurological status of ICH patients. The goal of this study was to investigate whether non-contrast computed tomography imaging biomarkers (NCCT-IB) acquired at initial presentation can predict ICH growth in the acute stage. Materials and Methods: We retrospectively collected NCCT data from 326 patients with acute (<6 hours) ICH. Four NCCT-IBs (blending region, dark hole, island, and edema) were identified for each hematoma, respectively. HE status was recorded based on the clinical observation reported in the patient chart. Supervised machine learning models were developed, trained, and tested for 15 different input combinations of the NCCT-IBs to predict HE. Model performance was assessed using area under the receiver operating characteristic curve and probability for accurate diagnosis (P_AD) was calculated. A 20-fold Monte-Carlo cross validation was implemented to ensure model reliability on a limited sample size of data, by running a myriad of random training/testing splits. Results: The developed algorithm was able to predict expansion utilizing all four inputs with an accuracy of 70.17%. Further testing of all biomarker combinations yielded P_AD ranging from 0.57, to 0.70. Conclusion: Specific attributes of ICHs may influence the likelihood of HE and can be evaluated via a machine learning algorithm. However, certain parameters may differ in importance to reach accurate conclusions about potential expansion.

Pathological changes in blood flow lead to altered hemodynamic forces, which are responsible for a number of conditions related to the remodeling and regeneration of the vasculature. More specifically, wall shear stress (WSS) has been shown to be a significant hemodynamic parameter with respect to aneurysm growth and rupture, as well as plaque activation leading to increased risk of stroke. In-vivo measurement of shear stress is difficult due to the stringent requirements on spatial resolution near the wall boundaries, as well as the deviation from the commonly assumed parabolic flow behavior at the wall. In this work, we propose an experimental method of in-vitro WSS calculations from high-temporal resolution velocity distributions, which are derived from 1000 fps high-speed angiography (HSA). The high-spatial and temporal resolution of our HSA detector makes such high-resolution velocity gradient measurements feasible. Presented here is the methodology for calculation of WSS in the imaging plane, as well as initial results for a variety of vascular geometries at physiologically realistic flow rates. Further, the effect of spatial resolution on the gradient calculation is explored using CFD-derived velocity data. Such angiographic-based analysis with HSA has the potential to provide critical hemodynamic feedback in an interventional setting, with the overarching objective of supporting clinical decision-making and improving patient outcomes.

Purpose: Data-driven methods based on x-ray angiographic parametric imaging (API) have been successfully used to provide prognosis for intracranial aneurysm (IA) treatment outcome. Previous studies have mainly focused on embolization devices where the flow pattern visualization is in the aneurysm dome; however, this is not possible in IAs treated with endovascular coils due to high x-ray attenuation of the devices. To circumvent this challenge, we propose to investigate whether flow changes in the parent artery distal to the coil-embolized IAs could be used to achieve the same accuracy of surgical outcome prognosis. Methods: Eighty digital subtraction angiography sequences were acquired from patients with IA embolized with coils. Five API parameters were recorded from a region of interest (ROI) placed distal to the IA neck in the main artery. Average API values were recorded and pre-treatment values. A supervised machine learning algorithm was trained to provide a sixmonth post procedure binary outcome (occluded/not occluded). Receiver operating characteristic (ROC) analysis was used to assess the accuracy of the method. Results: Use of API parameters with data driven methods yielded an area under the ROC curve of 0.77 ±0.11 and accuracy of 78.6%. Single parameter-based analysis yielded accuracies which were suboptimal for clinical acceptance. Conclusions: We determined that data-driven method based on API analysis of flow in the parent artery of IA treated with coils provide clinically acceptable accuracy for the prognosis of six months occlusion outcome.

The cerebral vascular system is constituted by all the arteries and veins irrigating the brain. This vascular tree starts from two pairs of arteries, the vertebral arteries and the internal carotid arteries. These latter divide into a circular shape being called the Circle of Willis (CoW). There is considerable variability in the structure of the CoW among patients. The CoW can host various vascular diseases, among which intracranial aneurysms are of particular importance because their occurrence, or more precisely their rupture, can be devastating. Intracranial aneurysms often occur at the bifurcations of the arterial tree (saccular aneurysms), as a bulge in the vessel wall. It is crucial to recognize and monitor such aneurysms. Anatomical identification of the bifurcations of the CoW can be of great help to establish a diagnosis or to plan a surgical operation. In this study, we propose an automatic solution to categorize the vascular anatomy of the CoW in 3D volumes by identifying its main constituting bifurcations. Our solution combines machine learning and a multivariate analysis (Linear Discriminant Analysis: LDA). The LDA works as a classifier and reduces the dimensionality of the dataset by transforming the selected features in a lower dimensional space. This work is a preliminary study prior to moving to human cerebrovascular images. We evaluate the proposed method using several machine learning techniques combined with a leave-one-out validation applied on a set of 30 synthetic vascular images as well as 30 mouse cerebral vasculatures.

Digital subtraction angiography (DSA) remains the clinical standard for detailed visualization of the neurovasculature due to its high-spatial resolution; however, detailed blood-flow quantification is impaired by its low-temporal resolution. 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. We have implemented a physics-based optical flow method to extract such information from HSA, but validation of the angiography-derived velocity distributions is not straightforward. Computational fluid dynamics (CFD) is widely regarded as the benchmark for hemodynamic analysis, as it provides a multitude of quantitative flow parameters throughout the volume of interest. However, there are several limitations with this method related to over-simplification of boundary conditions and suboptimal meshing (spatial resolution), that make CFD simulation results an inexact criterion for validation. To overcome this issue for HSA validation, CFD was used to generate both simulated high-speed angiograms and the corresponding ground-truth 3D flow fields to better understand the relationship between the 3D volumetric-flow distribution and the 2D projected-flow distribution as is obtained with angiography, and the subsequent 2D approximation of flow velocity. Several geometries were investigated, ranging from simple pipe models to complex patient-specific aneurysms. Simulated datasets were analyzed with the optical flow algorithm, and the effects of flow divergence, quantum mottle, and intensity gradient on the calculation were evaluated. From these simulations, we can evaluate whether flow fields reconstructed from HSA are representative of significant flow patterns in the 3D vasculature.

3D microstructural analysis of the human lung can provide unknown information about lung growth and progression of lung diseases. In this study, the alveolar walls of alveolar ducts and alveolar sacs of normal adult lungs were analyzed and classified into primary and secondary septa. The procedure is as follows. (1) 11 lung specimens were prepared by the Heitzman's method, (2) 3D micro-CT experiments were executed using BL20B2 at SPring-8, (3) Alveolar walls were extracted from 3D reconstructed images (12,429 × 12,429 × 4,800, pixel size 3 μm), and (4) Primary and secondary septa were classified by principal component analysis of alveolar walls. 3D structure of alveolar ducts and alveolar walls in normal adult lungs were clarified.

Three-dimensional (3D) microstructural analysis of the human lung is useful for accurate understanding of lung growth and disease progression. Counting of alveoli in alveolar ducts and alveolar sacs has not been sufficiently investigated. In this study, we present these results by counting of alveoli in alveolar ducts and alveolar sacs from synchrotron radiation 3D CT images using and 3D U-Net.

Retinal images have been used in the diagnosis of many ocular diseases such as glaucoma and diabetic retinopathy. Here, automatic detection of optic disk (OD) is essential in deriving clinical parameters to assist clinical diagnosis. In fact, detecting OD center and its boundary is the essential step of most vessel segmentation, disease diagnostic, and retinal recognition algorithms. In this study, we proposed a new approach for localizing OD by combining local histogram matching and the concept of deep learning. The algorithm is composed of 4 steps, Image partitioning, Local histogram matching and validation, Convolutional Neural Network (CNN) classification, and OD detection. Here, we used OD of the five reference retinal images in each dataset to extract the histograms of each color channel. Then, we calculated the mean of histograms for each channel as template for creating some OD candidates. An AlexNet-like CNN was applied to classify candidates as ODs or nonODs. The candidates used as an input to feed the CNN for final classification. In this study, we worked on three databases (one rural, MUMS-DB, and two publicly available databases, DRIVE, STARE) including 520 retinal images to evaluate the proposed method. The accuracy of our algorithm was 100%, 90%, and 95% for the DRIVE, STARE, and MUMS-DB respectively. It is shown that this method provides higher detection rates than the existing methods that have reported.

Minimally invasive visualization of follicles in the ovaries of adolescent and young adult (AYA) patients with cancer is a promising technique for maintaining fertility by cryopreserving follicle-rich ovarian tissues. Optical coherence tomography (OCT) using near-infrared light, which is less invasive to the follicles, is a promising imaging technique for quantifying follicle density before cryopreservation. A spectral-domain OCT system equipped with a superluminescent diode with a central wavelength of 1300 nm was applied to visualize primordial follicles with a diameter of about 20 μm, which are often found at depths of 100–1000 μm in ovarian tissue. We used a 20x immersion objective to improve lateral image resolution and reduce ghost images due to surface reflections. OCT images of 30 μm diameter glass beads mimicking primordial follicles embedded in a polyacrylamide gel containing scattering particles were visualized at the depth required for imaging the ovaries. Ovarian tissues of 3-day-old and 12-day-old mice were fixed between an originally designed hydrophobic silicone plate and a cover glass. The distribution of primordial follicles and the structure of mature follicles were visualized in ovaries of 3-day-old and 12-day-old mice, respectively. These results indicate that near-infrared OCT at a central wavelength of 1300 nm with an extended depth of light penetration can effectively visualize the structure of follicles and quantify their density at depths where follicles are abundant in ovarian tissues. Further improvements in spatial resolution and image processing are expected to accelerate the contribution to reproductive medicine in the future.

X-ray luminescence computed tomography (XLCT) is a hybrid molecular imaging modality combining the merits of both x-ray imaging (high spatial resolution) and optical imaging (high sensitivity to tracer nanophosphors). Narrow x-ray beam based XLCT imaging has shown promise for high spatial resolution imaging, but the slow acquisition speed limits its applications for in vivo imaging. We introduced a continuous scanning scheme to replace the selective excitation scheme to improve imaging speed in a previous study. Under the continuous scanning scheme, the main factor that limits the scanning speed is the data acquisition time at each interval position. In this work, we have used a gated photon counter (SR400, Stanford Research Systems) to replace the high-speed oscilloscope (MDO3104, Tektronix) to acquire measurement data. The gated photon counter only counts the photon peaks in each measurement interval, while the oscilloscope records the entire waveform including both background noise data and photon peak data. The photon counter records much less data without losing any relevant information, which makes it ideal for super-fast three-dimensional (3D) imaging. We have built prototype XLCT imaging systems of both types and performed both single target and multiple target phantom experiments in 3D. The results have verified the feasibility of our proposed photon counter based system and good 3D imaging capabilities of XLCT within a reasonable time, yielding a 14 times faster scanning time compared with the oscilloscope based XLCT system. Now, the total scan time is reduced to 27 seconds per transverse section.

Whole-brain high temporal resolution CT perfusion (CTP) is now feasible with wide-detector row CT scanners, but the optimal dose distribution of dynamic images remains unknown. In this study, we investigated the accuracy of perfusion parameters estimated in digital perfusion phantoms generated at various temporal resolutions with fixed scan dose. In accordance with CTP guidelines, simulated dose was set to a time-density curve (TDC) noise of 10 HU at a sampling interval of 2.0 s over 60 s, and higher temporal resolutions of 1.0 and 0.5 s intervals were investigated at 14 and 20 HU of noise, respectively. Monte Carlo simulations with known ground truth perfusion were conducted to test the performance of model-independent and model-dependent deconvolution algorithms as a function of temporal resolution at isodose. Tissue TDCs were simulated by convolving gamma-variate, linear or boxcar residue functions with a patient arterial TDC before adding Gaussian noise at the appropriate level then sampling at the investigated temporal resolutions. A digital brain perfusion phantom with physiological ground truth perfusion was similarly investigated. Only cerebral blood flow (CBF) estimates with the model-dependent algorithm marginally improved at higher temporal resolution as indicated by mean absolute error (MAE; 7.1±4.6 ml/min/100 g at 0.5 s, 9.6±6.0 ml/min/100 g at 2.0 s) but not with the modelindependent algorithm (MAE: 11.6±11.4 ml/min/100 g at 0.5 s, 11.3±11.7 ml/min/100 g at 2.0 s). Higher temporal resolution did not improve parameter estimation in the brain perfusion phantom. For the investigated temporal resolutions and simulated CTP dose, dose distribution appears negligible.

Narrow band imaging (NBI) bronchoscopy enables enhanced visualization of microvascular structures in the mucosal layer of the lungs (airway walls). Such vessels are potential indications of developing cancerous lesions. To find these vascular patterns, the bronchoscope is navigated through the airways, and the physician manually observes potential mucosal vessel structures. We propose an automated video analysis framework based on deep learning and spatial-temporal information in NBI video to find potential cancerous lesions. Using patient data, we demonstrate that our method enables 89% accuracy, 93% sensitivity, and 86% specificity for lesion detection at ~19fps speed. Furthermore, we utilize an upgraded Siamese tracker using kinematic motion modeling jointly with the detection network to isolate abnormalities, achieving 95%/90% accuracy, 90%/74% sensitivity, and 99%/99% specificity, with and without the tracker, respectively.

Stroke is a representative hemiplegia disease. Using a gerbils model (n = 6) divided into three groups, a control group without stroke (cont, n = 2), and 7-day (7 days, n = 2) and 14-day (14 days, n = 2) groups with right cerebral ischemia, we extracted the soleus muscle of the paralyzed side (left). We evaluated the mechanism underlying the stroke-induced muscle injury by using synchrotron radiation phase-contrast imaging (SR-PCI). We succeeded in quantifying the degree of injury by dividing the muscle space and fiber region. The analysis of the space volumes of cont and 7 days revealed an increase in volume (p<.05). The space region increased according to the period of evaluation and a very large difference was found both between 7 days and 14 days after stroke (p<.01) and cont and 14 days (p<.01). In addition, regression analysis showed a positive correlation with increasing space according to the evaluation period (r² = .97). We found that the fiber region had increased damage at 7 days than in the control (p<.05). Subsequently, the characteristic of muscle tissue recovery after stroke was observed at 14 days instead of 7 days (p<.05). Therefore, by establishing the possibility of animal study using SR-PCI, we expect that it will be possible to present a protocol for gait training of clinical patients that can improve their qualitative exercise ability by synthesizing the recovery period.

Modeling brain tissue mechanics is important for understanding the pathogenesis of traumatic brain injury, with models often including brain tissue geometry and microstructural features like white matter fiber orientation. Recently, the cerebral vasculature has been included in models, however the effect of cerebral vessels on the mechanical response of the brain is unclear. A dataset of 23 subjects that includes structural MRI, angiography, and mechanical neuroimaging using magnetic resonance elastography (MRE) was collected to determine if there is a dependence of vasculature on in vivo brain mechanical properties. A pipeline was implemented using existing methods for processing anatomical, angiography, and MRE images; all images were co-registered for each subject and transformed to a common space. The regional mean stiffness and damping ratio of the brain, by anatomical segmentation, showed no dependence on vessel density but showed heterogeneity across the brain. A sub-regional analysis after stratifying by MRE stiffness showed a strong positive correlation in the cortical gray matter (R²=0.69) and a strong negative correlation in the deep gray matter (R²=0.76). Other regions showed similar trends with R² values below 0.54. The opposite trends could be a result of regional microstructure difference, or a dependence on vessel type and size. A similar analysis using the brain damping ratio showed no dependence of vasculature on brain viscous properties. Quantifying the dependence of brain mechanical properties on vasculature will aid in understanding the biomechanics of the brain and inform their use in computational models of brain injury.

We propose a learning-based method to automatically segment extraprostatic nodal lesion from 18F-fluciclovine (anti-1- amino-3-[18F] fluorocyclobutane-1-carboxylic acid) PET images. Our proposed method, named hierarchical activation network, consists of three main subnetworks: a fully convolutional one-stage object detection (FCOS) network and a mask module, and a hierarchical convolutional block. While FCOS is employed to detect the view-of-interests (VOIs) of extraprostatic nodal lesion. Hierarchical convolutional block is used to derive activation map to boost the classification accuracy around lesion boundary. This is followed by the binary segmentation of extraprostatic nodal lesion within the detected VOI by mask module. To evaluate the proposed method, we retrospectively investigated 92 lesions with 18F- fluciclovine PET acquired. On each dataset, the extraprostatic lesions 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 DSC among all lesions is close to 0.7. The proposed method has great potential in improving the efficiency and mitigating the observer-dependence in extraprostatic lesion contouring for radiation therapy.

Multiparameter magnetic resonance imaging (mp-MRI) commonly used modality for segmentation of glioma and its subregions, whereas RT therapy is a commonly used treatment modality. Current workflow includes manual segmentation of brain tumor sub-regions, which is a very lengthy and laborious process given that different sets of MR images have to be analyzed for an appropriate diagnosis. This work focuses on implementing and testing feasibility of a new deep learning model for an automatic segmentation of brain tumor sub-regions. Our proposed method, named hierarchical substructural activation network, consists of three main parts: a detection and segmentation module, and a hierarchical convolutional block. While the detection module is employed to detect the view-of-interests (VOIs) of brain tumor, which include all tumor sub-structures, the hierarchical convolutional block is used to derive structural activation map (SAM) to boost the classification accuracy between different structures. This is followed by the semantic segmentation of each substructure within the detected VOI by segmentation module. Brain tumor segmentation challenge (BraTS) 2020 dataset was used for evaluating our proposed framework. We performed five-fold cross validation experiments on 100 BraTS datasets. Three substructures, i.e., necrosis and non-enhancing, edema, enhancing tumor (ET), tumor core (TC), were segmented and compared with manual contours using the Dice similarity coefficient (DSC) and mean surface distance (MSD). In terms of segmentation of necrosis and non-enhancing subregions, edema, ET and TC, our method yielded DSC of 0.69±0.24, 0.89±0.09, 0.80±0.14, and 0.88±0.11, respectively, and MSD of 2.01±2.75, 0.63±0.58, 0.98±1.10 and 1.74±1.06 mm, respectively. Preliminary results of this work show promise to both accurately and automatically segment brain tumor subregions by our proposed method, providing motivation for its clinical implementation to improve clinical workflow.

Deterioration of the overall musculoskeletal system with aging is a universal phenomenon influenced by different demographic and lifestyle factors. Often, pectoral muscle metrics are used to describe overall muscle health, and CTbased studies have demonstrated their associations with various diseases, lung function, and mortality. However, these studies use extremely laborious manual means to segment pectoral muscles limiting both study size and scope. Here, we present a CT-based automated method for segmentation of the pectoral muscle using deep learning and computation of pectoral muscle area (PMA). We examined the extent of change in PMA with aging and sex using retrospective chest CT scans (n = 260) from COPDGene Iowa cohort at baseline visits. A two-dimensional U-Net was developed, optimized, and trained (n = 60) to generate a pixel-wise pectoral muscle probability map from chest CT scans, which was followed by an image post-processing cascade to segment the muscle area. Preliminary results (n = 200) show that our CT-based automated segmentation method is accurate (Dice score = 0.93), and it detects muscle wasting with aging. Males had significantly greater PMA as compared to females (effect size: 0.84; p < 0.001). A five-year loss in PMA of 4.8% was observed in the study population with losses of 4.3% and 5.1% for females and males, respectively. Chest CT-based automated methods for pectoral muscle segmentation are suitable for large population studies exploring broader scientific knowledge under various diseases.

Cerebral aneurysms (CA) affect nearly 6% of the US population and its rupture is one of the major causes of hemorrhagic stroke. Neurointerventionalists performing endovascular therapy (ET) to treat CA rely on qualitative image sequences obtained under fluoroscopy guidance alone, and do not have access to crucial quantitative information regarding blood flow before, during and after treatment – partially contributing to a failure rate of up to 30%. Computational fluid dynamics (CFD) is a powerful tool that can provide a wealth of quantitative data; however, CFD has found limited utility in the clinic due to the challenges in obtaining hemodynamic boundary conditions for each patient. In this work, we present a novel CFD-based simulated angiogram approach (SAA) that resolves the blood flow physics and interaction between blood and injected contrast agent to extract quantitative hemodynamic parameters which can be used to design real-time parametric imaging analysis. The SAA enables correlating contrast agent transport to the underlying hemodynamic conditions via time-density curves (TDC) obtained at several points in the region of interest. The ability of the TDC and the SAA to provide critical hemodynamic parameters in and around CA anatomies, such as washout and local flow changes is explored and presented. This provides invaluable quantitative data to the clinician at the time of intervention, since it incorporates the physics of blood flow and correlates the contrast transport to hemodynamic parameters quantitatively – thereby enabling the clinician to take informed decisions that improve treatment outcomes.

Acoustic neuroma (AN) is a noncancerous and slow-growing tumor that influences the human hearing system. Magnetic resonance images (MRIs) are routinely utilized to monitor tumor progression. Quantifying tumor growth in an automated manner would allow more precise studies, both at the population level and for the clini- cal management of individual patients. In recent years, deep learning methods have shown excellent performance for many medical image segmentation tasks. However, most current methods do not work well on heterogeneous datasets where MRIs are acquired with vastly different protocols. In this paper, we propose a deep learning framework with ensembled convolutional neural networks (CNNs) to segment acoustic neuromas even in hetero- geneous datasets. We ensemble a 2.5D CNN model and a 3D CNN model together, with augmentations added to the model for better inter-dataset segmentation performance. We test our methods on two datasets: the publicly available dataset from the crossMoDA challenge and an in-house dataset. We examine our method with supervised learning on the crossMoDA dataset and directly apply the trained model to the in-house dataset. We use the Dice score, average surface distance (ASD), and 95-percent Hausdorff distance (95HD) as evaluation metrics. Our method has better performance than the baseline methods, not only on intra-dataset segmentation accuracy but also on inter-dataset generalizability.

Cerebral atrophy is characterized by a shrinking of the brain and consequently an enlargement of fluid-filled spaces within the cranium. It is a hallmark of normal aging and a sequelae following brain injury, and is of relevance in other brain diseases. There has been conflicting evidence of the effect of ventricle enlargement on the biomechanics of the brain during head impact. Computational simulations of brain biomechanics were used to investigate enlargement of the ventricles and subarachnoid space (SAS). These models are summarized as 1) a simplified 2D phantom, 2) an axial 2D brain model, and 3) subject-specific 3D brain model. Our preliminary results with the 2D models show minimal effect of enlarged ventricles on brain deformation, and shows decreasing brain strain with a thicker SAS layer. The 3D models show a general decrease in strain metrics for head motion about all three axes of rotation. Investigating the effect of the size of the fluid-filled spaces within the cranium on brain deformation will aid in the understanding of subject-specific brain injury risk, especially during aging and disease.

Purpose: To investigate the relation between delayed ischemic stroke and the intracranial atherosclerotic disease (ICAD) hemodynamics as determined by Non-invasive Optimal Vessel Analysis (NOVA) MRI measurements. Materials and Methods: Thirty-three patients with ICAD were enrolled in this study. All patients underwent clinically indicated angioplasty followed by 2-dimensional phase contrast MR (2D PCMR) performed on a 3.0 Tesla MRI scanner using either a 16-channel neurovascular coil or 32-channel head coil. The volumetric flow rate measurements were calculated from 2D PCMR with Non-invasive Optimal Vessel Analysis (NOVA) software (VasSol, Chicago, IL, USA). Flow rate measurements were obtained in 20 major arteries distal, proximal and within the Circle of Willis. Patients were followed up for six month, and ischemia reoccurrence and location were recorded. Receiver operating characteristic (ROC) analysis was performed using flow rates measurements in the ipsilateral side of the ischemic event occurrence. Results Complete set of measurements was achieved in n=34. Left and right hemisphere ischemia recurrence was observed in seven and three cases respectively. Best predictor of ischemic event reoccurrence was flow rate in the middle cerebral artery with area under the ROC of 0.821±0.109. Conclusions: This is an effectiveness study to determine whether blood flow measurements in the intracranial vasculature may be predictive of future ischemic events. Our results demonstrated significant correlation between the blood flow measurements using 2D PCMR processed with the NOVA software and the reoccurrence of ischemia. These results support further investigation for using this method for risk stratification of ICAD patients.

X-ray dark-field measured on laboratory sources with large focal spots and detector apertures is sensitive to intra-pixel phase gradients abundant in the lungs due to its hierarchical structure of subdividing airways terminating in thin-walled alveoli. This work leverages this sensitivity to exploit complementary information from x-ray dark-field and attenuation computed tomography (CT) images to improve quantification of morphology in pulmonary fibrosis. Specifically, a darkfield enhanced attenuation technique is developed to restore edges and small features in the attenuation image lost to blurring by appropriately scaling and subtracting the dark-field image. An intratracheally treated bleomycin mouse model of pulmonary fibrosis was used to evaluate the impact of the proposed dark-field enhanced attenuation technique on quantifying fibrosis extent. The mouse model was fixated ex vivo to be imaged with a Talbot-Lau grating interferometer micro-CT to generate x-ray dark field and attenuation volumes of 60 µm voxels. Then the specimen was imaged with a reference micro-CT scanner at 5 μm voxel resolution to get a ground truth approximation of local structure. The volumes were co-registered for visual and pixelwise comparisons. Qualitative image comparisons were used to assess visual sharpness while Bland-Altman plots were used to assess agreement with the reference scan at quantifying fibrosis in terms of tissue area fraction measured in 80 randomly sampled nonoverlapping 2 mm square patches. Visual comparisons demonstrated enhanced sharpness and retention of small lung structures while BlandAltman analysis revealed an improved agreement ratio of 0.544 compared to 0.374 in the original attenuation image with a reduction in variance. These results demonstrate that dark-field and attenuation images can be used together to improve resolution of small structures and aid in quantification of pulmonary fibrosis in a mouse model.

Gold nanoparticles (GNPs) as promising radiation sensitizers have been increasingly studied in a wide range of radiotherapy applications. By detecting the characteristic x-ray fluorescence (XRF) photons, x-ray fluorescence computed tomography (XFCT) can simultaneously determine both the spatial distribution and concentration of GNPs in vivo, affording for cancer diagnosis and irradiation guidance. However, the long scanning time of current single-pixel detectorbased configuration hinders the translation of XFCT to preclinical and clinical applications. This study presents a conebeam XFCT system using pixelated photon-counting detector with pinhole collimator to acquire XRF projection image in one motion, eliminating the previously step-by-step translation of objects, which allows fast whole-body GNP imaging. We have 3D printed a heat-resistance mold kit to cast a cone-beam x-ray source collimator using Cerrobend alloy. We selected HECITEC (High Energy X-ray Imaging Technology) as the XRF detector, in view of its high spatial resolution (0.25 mm of pitch) and energy resolution (800 eV FWHM at 60 keV). We have customized a 2-mm pinhole collimator to provide spatial information of XRF signals. We have also evaluated the roles of pixel binning and spectrum denoising in aspects of XRF peal extraction. Phantom experiments with GNP of different concentrations (0.078~2.5wt.%) were used to evaluate the sensitivity of GNP detection. In vivo experiments on mouse intravenously administered GNPs were used to validate the feasibility of the proposed system in terms of GNP biodistribution imaging. The results of this study will be helpful to guide XFCT development for routine in vivo GNP imaging

We developed the second-generation photoacoustic dual scan mammoscope (DSM) as a safe and effective modality for the breast imaging of patients with high breast density. Besides being a portable system with high resolution, DSM-2 has several improvements compared to the previous version: a larger field of view, better system stability, higher ultrasound imaging quality, and additional quasi-static elastography capability. The performance of the new system was demonstrated through clinical studies. The experiment result confirmed the capability of the second-generation DSM system as a powerful tool for breast imaging.

Non-invasive imaging strategies are essential in understanding the factors associated with the ongoing rise in cardiovascular disease (CVD) and Alzheimer’s disease (AD), and the possible interactions between these diseases. Our scientific premise is based on the role of the APOE gene where APOE4 is considered a risk factor for CVD and AD. This study incorporates the use of novel apolipoprotein E mouse models with a humanized innate immune system, through mNos2 KO and presence of the human NOS2 gene (APOE4/HN and APOE3/HN), together with the more traditional APOE (-/-) mice. We have imaged these models for CVD and AD with a cardiac photon-counting CT (PCCT) imaging pipeline to characterize their cardiac anatomy and function. The pipeline, consisting of contrast enhanced in vivo PCCT imaging, accurate intrinsic cardiac gating, temporally resolved multi-energy iterative reconstruction, spectral decomposition, 3D cardiac segmentation, and ex vivo, high-resolution PCCT imaging allowed for quantitative analysis and comparison of cardiac function as well as identification of anatomical irregularities such as calcified aortic plaques. Our analysis finds no statistically significant differences in cardiac functional metrics or the aortic diameter in these APOE mouse models. Future work will focus on optimizing the image reconstruction to reduce the computation time and on using staining for ex vivo PCCT imaging of the vascular system in these models.

Clinical stratification of rupture risk is limited to criteria based on geometry (diameter) which is not always accurate. We propose an image transformer approach applying neural networks for focused attention on abdominal aortic aneurysms (AAAs), which doesn’t require explicit segmentation, for predicting rupture risk, starting with CT angiography images. Our image dataset consisted of 16 cases with high rupture risk and 14 cases with low rupture risk. Our study reveals that 3D ResNet classifiers trained with neural embeddings from a 3D U-Net trained on images of any one rupture risk class produced an accuracy of 90% (83% sensitivity, 100% specificity). Our representation learning pipeline, AAA-Net, could be adapted to reduce the amount of time and clinical expertise required to identify AAA rupture risk, enabling efficient and automated aneurysm monitoring.

Magnetic resonance imaging (MRI) is useful for the detection of abnormalities affecting maternal and fetal health. In this study, we used a fully convolutional neural network for simultaneous segmentation of the uterine cavity and placenta on MR images. We trained the network with MR images of 181 patients, with 157 for training and 24 for validation. The segmentation performance of the algorithm was evaluated using MR images of 60 additional patients that were not involved in training. The average Dice similarity coefficients achieved for the uterine cavity and placenta were 92% and 80%, respectively. The algorithm could estimate the volume of the uterine cavity and placenta with average errors of less than 1.1% compared to manual estimations. Automated segmentation, when incorporated into clinical use, has the potential to quantify, standardize, and improve placental assessment, resulting in improved outcomes for mothers and fetuses.

A novel phenotype guided interpretable graph convolutional network (PGI-GCN) for the analysis of fMRI data is proposed. We utilize PGI-GCN to predict the ages of children and young adults based on multi-paradigm fMRI data of the Philadelphia Neurodevelopmental Cohort (PNC) dataset. We show PGI-GCN to have superior predictive capability compared to a simpler deep model that uses functional connectivity plus gender without the population-level graph. A learnable mask identifies 3 important intra-network (Memory Retrieval, Dorsal Attention, and Subcortical) and 3 important inter-network (Visual-Cerebellar, Visual-Dorsal Attention, and Subcortical-Cerebellar) connectivity differences between children and young adults.

Deep learning techniques specifically deep convolutional neural network (CNN) models, the latest core model of artificial neural networks, provide computer-vision capabilities, including medical and biomedical image analysis and classification. In this paper, we imaged ex vivo human tooth specimens using OCT imaging systems to classify and clarify the accuracy of different tooth samples with and without carious lesions via deep CNN models with transfer- learning and fine-tuning strategies. Collecting a large amount of OCT image data from dental samples can be difficult, and not providing sufficient data for CNN models can lead to overfitting. For these reasons, transfer learning and fine- tuning techniques were utilized in this study. OCT images of human extracted premolar and molar teeth were categorized into three classes. Five deep CNN models, specifically, a basic CNN with three convolutional and max pooling layers, VGG16 and VGG19 transfer-learning models, and finely tuned VGG16 and VGG19 models were developed and evaluated for OCT image classification of dental caries. In transfer learning, an existing learned model was employed as a feature extractor without changing the weight data, while in fine tuning, an existing learned model was utilized as a feature extractor by relearning some of the weight data. These methods are powerful methods for training deep CNN models without overfitting. This study highlights the performance of various deep learning models for OCT image classification of carious lesions.

Accurate prognostic stratification of Head-and-Neck-Squamous-Cell-Carcinoma (HNSCC) patients can be an important clinical reference when designing therapeutic strategies. We set to predict 4 outcomes: overall survival (OS), distant metastasis (DM), locoregional recurrence (LR), and progression-free survival (SP). We studied Hybrid Machine Learning Systems (HMLS), applied to datasets with radiomics features. In this multicenter study, 408 HNSCC patients were extracted from The Cancer Imaging Archive (TCIA) database. PET images were registered to CT, enhanced, and cropped. 215 radiomics features were extracted from each region of interest via our standardized SERA radiomics package. We employed multiple HMLSs: 12 feature extraction (FEA) or 9 feature selection algorithms (FSA) linked with 9 survival-prediction-algorithms (SPA) optimized by 5-fold cross-validation, applied to PET only, CT only and 4 PET-CT datasets generated by image-level fusion strategies. Datasets were normalized by z-score-technique, and cindices were reported to compare the models. For OS prediction, the highest c-index 0.73 ± 0.10 was obtained for HMLS with Ratio of low-pass pyramid (RP) fusion technique + gaussian process latent variable model (GPLVM) + causal structure learning-based feature modification method (CSFM). For DM prediction, we achieved 0.80±0.06 via Dual-tree complex wavelet transform (DTCWT) fusion + Laplacian Score (LAP) + Logistic regression hazards (LH). For LR prediction, we arrived at a c-index of 0.73 ± 0.13 using PET + Sammon Mapping Algorithm (SM)+ deep neural network to distribute first hitting times (DHS). For SP prediction, the performance of 0.68 ± 0.02 was obtained via PET + SM + Relative risk model-depend on time (CoxTime). When no dimensionality reduction (FEA/FSA) was employed, the above 4 performances decreased to 0.69 ± 0. 10, 0.74 ± 0.13, 0.66 ± 0.15, and 0.68 ± 0.04 for OS, DM, LR and SP prediction. We demonstrated that using fusion techniques followed by appropriate HMLSs, including FEAs/FSAs and SPAs, improved prediction performance.

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. Graph theory has been the standard tool to provide biomarkers in imaging connectomics showing the Alzheimer’s disease (AD). We propose a novel concept - graph signal processing - to analyze the evolution of disease graphs leading from mild cognitive impairment (MCI) to AD and derive frequency-based biomarkers representative for this disease. We show that high oscillations derived from the graph Fourier decomposition can provide important discriminatory information. To quantify the qualitative intuition of high oscillations, we use two concepts from signal theory: (1) zero crossings and (2) total variations. We apply these concepts on functional and structural brain connectivity networks for control (CN), mild cognitive impairment (MCI) and Alzheimer’s disease (AD) subjects. Our results applied to functional brain networks suggest that graph signal processing can accurately describe the frequencies of brain networks, and explain how AD is associated with low frequency and localized averaging confirmed by clinical results.

Spinal degeneration and vertebral fractures are common among the elderly adversely impacting mobility, quality of life, lung function, fracture risk, and mortality. Segmentation of individual vertebrae from computed tomography (CT) imaging is crucial for studying spine degeneration, vertebral fractures, and bone density with aging and their mechanistic links with demographics, lifestyle factors, and comorbidities. We present an automated method to segment individual vertebral bodies (T1-L1) and compute the kyphotic angle of the spine from chest CT images. A three-dimensional U-Net was developed, optimized, and trained to generate a voxellevel vertebral probability map from a chest CT scan. Multi-parametric thresholding was applied on the probability map to segment individual vertebrae by iteratively relaxing the probability threshold value, while avoiding fusion among adjacent vertebrae. The kyphotic angle was computed using two orthogonal planes on the spine centerline at the inter-vertebral spaces T3-T4 and T12-L1 and a common sagittal plane. Total lung capacity (TLC) chest CT scans from baseline visits of the COPDGene Iowa cohort were used for our experiments. The U-Net method was trained and validated using 40 scans and tested on a separate set of 100 scans. Segmentation of individual vertebrae achieved a mean Dice score of 0.93 as compared to manual segmentation, and the kyphotic angle computation method produced a linear correlation of 0.88 (r-value) with manual measurements. This method provides a fully automated tool to study different mechanistic pathways of age-related spine modeling and vertebral fractures in retrospective datasets available from large multi-site chest related studies.

Quantitative angiography is a 2D/3D x-ray imaging modality that summarizes hemodynamic information using time density curve (TDC) based parameters. Estimation of the TDC parameters are susceptible to errors due to various factors including, patient motion, incomplete temporal data, imaging trigger errors etc. In this study, we tested the feasibility of using recurrent neural networks (RNN) to recover complete TDC temporal information from incomplete sequences and evaluate quantitative parameters generated from the corrected TDCs. Digital subtraction angiograms (DSAs) were collected from patients undergoing endovascular treatments and angiographic parametric imaging (API) parameters were calculated from each DSA. Each set of API parameters was used to simulate a TDC resulting in a dataset of 760 TDCs. One-third of each TDC was continuously masked from pseudo-random points past the peak height (PH) point to simulate missing/artifact information. An RNN was developed, trained and tested to generate completed/corrected TDCs. The RNN recovered complete TDC temporal information with an average mean squared error of 0.0086±0.002. Average mean absolute errors were calculated between each API parameter generated from the ground truth TDCs and RNN corrected TDCs, these were 11.02%±0.91 for time to peak, 10.97%±0.69 for mean transit time, 5.65%±0.76 for PH, and 15.08%±0.98 for area under the TDC. The change in API parameters was not clinically significant and the predictive power of the API parameters was retained. This study proved the feasibility of using RNNs to mitigate motion artifacts and incomplete angiographic acquisitions to extract accurate quantitative parameters.

Cerenkov luminescence tomography (CLT) is a highly sensitive and promising imaging modality for three-dimensional visualization of radiopharmaceuticals. However, the approximate error generated by the simplified radiation transfer equation and the ill-posedness of the inverse problem limit the improvement of CLT reconstruction. In this research, a residual learning network (RLN) was proposed to improve morphological restorability. By learning the relationship between surface photon intensity and internal source, the errors from the inverse process could be avoided. RLN comprised two fully connected sub-networks: one was used to provide the coarse reconstruction result. The other optimized the final reconstruction result by learning the residual between the coarse reconstruction result and the true source. Monte Carlo method was used to generate the dataset. Furthermore, multilayer fully connected neural network (MFCNN) was used as baselines and compared. Single-source simulation and robustness experiments were conducted to evaluate the reconstruction performance. The experimental results show RLN achieved accurate localization and morphological reconstruction, which will promote the application of machine learning in optical tomography reconstruction.

Various concerns related to health, diagnostic outcomes and environmental impact have been raised recently on the wide usage of Gadolinium based contrast agents (GBCAs) in MR imaging. The purpose of this work is to propose a deep learning-based method to predict contrast enhanced MR images from the unenhanced counterpart. The proposed workflow consists of two cascade networks: the first network is trained to derive semantic features to identify the tumor regions under supervision of the tumor contours; the second network is trained to generate the synthetic contrast enhanced MR images with the input of the concatenation of the semantic features and non-contrast MR images. Qualitative and quantitative evaluations on the performance of the proposed method were conducted with MR images in the multimodal brain tumor segmentation challenge 2020 (BraTS2020) dataset. Preliminary results show that the synthetic contrast enhanced MR images were undifferentiable from the ground truth. Mean values and standard deviations of the normalized mean absolute error (NMAE), structural similarity index measurement (SSIM) and Pearson correlation coefficient (PCC) were 0.061±0.018, 0.993±0.005 and 0.996±0.005, respectively, for the whole brain; and were 0.049±0.022, 0.995±0.006 and 0.999±0.002, respectively, for the tumor regions. Utilizing cascade networks and supervising the training with tumor contours are novel for deep learning-based contrast enhanced MR image synthesis. It is expected to bypass the contrast agent usage in MR scans for diagnosis and treatment planning in radiotherapy if applied in the practice. I

Positron projection imaging (PPI) of tumor-bearing mice under certain circumstances can provide accurate in vivo estimates of total tumor radioactivity, an important pharmacokinetic measurement. However, the number of images generated in these studies is typically very large and many 2D tumor regions-of-interest (ROIs) must be manually defined to obtain accurate radioactivity estimates. In this study, we compared several methods that might allow automatic quantification of tumor radioactivity content. In total, 120 images (n = 81 mice) were acquired in pairs during two separate experiments. The first experimental batch was used for development, and the second as an independent testing cohort. Four methodologies were evaluated, including deep-learning (U-net), region-growing (Level-Set), and thresholding (Otsu, mean value). For all methodologies, preprocessing of the images included uptake normalization to fixed window. Tumor radioactivity is defined as total uptake within a tumor region minus a background estimate. Performance metrics were evaluated for both segmentation results (Sorenson-Dice Coefficient) and radioactivity calculation results (Bland-Altman). Using the test batch data, DICE score for U-net segmentation was 0.82, vs. 0.5-0.6 for the other three methods. Bland-Altman plots showed a mean difference of -0.26 for U-net based calculations vs. -0.5 to -0.8 for the other methods. The U-net approach had the highest accuracy in both segmentation and subsequent radioactivity calculation.

Numerous deep learning approaches have been proposed to automatically classify Alzheimer’s disease (AD) from medical images. However, common approaches, such as convolutional neural networks (CNNs), lack interpretability and are prone to over-fitting when trained on small datasets. As an alternative, significantly less work has explored applying deep learning approaches to region-based features commonly obtained from atlas partitions of known regions of interest (ROI). In this paper, we propose a self-attention mechanism to jointly learn a graph of ROI connectivity as a prior for learning meaningful features for AD prediction. We apply our method to both the classification of AD subjects from healthy controls and to predict whether mild cognitive impaired (MCI) subjects will progress to AD (pMCI) or not (sMCI). We systematically show that our model’s performance compares well with other common ML approaches for ROI-based methods, such as neural networks and support vector machines. Finally, we perform exploratory graph analysis to illustrate the interpretability properties of the attention graphs and how they can provide insight for scientific discovery.

Bone skeleton segmentation is a fundamental step in medical image analysis applications, such as computer-aided orthopedic surgery, fracture detection, detecting and diagnosing bone pathology and degenerative diseases. The extraction of bones in CT scans is a challenging task, which done manually by experts is a time-consuming process. In this work, a deep learning (DL) based solution for automatic segmentation of skeletal structure in conventional CT images is presented. To address the task of creating a diverse, high quality training dataset, an iterative data annotation process is utilized. A small training dataset is created using human annotation effort and used to train a segmentation model. The model is then inferred to initialize the ground truths for new cases. The new ground truths are reviewed and edited as necessary by the human annotators and added to the training dataset. The process is repeated until the model performance no longer improves on a held-out validation dataset. Within a few iterations the model generalization and prediction performance are observed to improve as a function of training dataset size and variety. Human effort in the dataset labeling process is also noted to reduce significantly for every interaction. The final DL segmentation models perform well across anatomy, scan and reconstruction settings and achieve a mean dice score of 0.988 on a held out, independent validation dataset.

Dental cone beam computed tomography (CBCT) offers a variety of fields of view (FOV), ranging from the entire craniofacial complex down to an individual tooth. To manage the radiation dose to the patient, partial rotation scanning is often used. However, the images suffer from non-uniformities related to attenuation of the beam as it crosses the patient anatomy and the uneven radiation field due in part to scattered radiation. We have modelled the cupping artifact in the maxillofacial field of view and applied a correction to improve uniformity in the image. We applied this correction to smaller fields of view. Images were obtained using a Carestream 9300 CBCT system of the SEDENTEX image quality phantom and 2 anthropomorphic skull phantoms. The maxillofacial FOV (17x11 cm² ) was used with clinical settings for an average adult (90 kVp, 4 mA, 6.4 s) and reconstructed with 0.25 mm voxel spacing. In MATLAB, we modeled the 2D surface across a uniform section of the SEDENTEX phantom with a polynomial fit to find an offset that could be added to reduce the cupping artefact. This offset was then applied to the images of the anthropomorphic and SEDENTEX phantoms for 17x11 cm² , 10x10 cm² , 10x5 cm² and 8x8 cm² FOVs. The offset was cropped for the 10x10 cm² , 10x5 cm² and 8x8 cm² FOVs. For all images, the uniformity was improved. From this study, we conclude that a single correction matrix can be used for all 4 FOVs on this machine to improve image quality.

Osteoporosis is an age-related bone disease causing increased bone loss and enhanced bone fragility and fracture-risk. Osteoporotic imaging plays important roles in quantitative assessment of bone quality, strength, and fracture-risk, and plays important roles in evaluating disease severity and treatment planning. High-resolution CT imaging on dedicated scanners is used for finite element (FE) analysis (FEA) of trabecular bone (Tb) microstructure. However, Tb micro FEA on clinical CT imaging is challenging and yet to be established due to difficulties with binary segmentation of Tb at relatively low-resolution. Here, we present a CT-based material density adjusted nonlinear FEA method for computing Tb shear modulus, while avoiding explicit segmentation of Tb micro-network. FE meshes were constructed over upright cylindrical VOIs derived from CT scans after alignment of tibia axes with the image axes. Image voxels were modelled as cubical mesh elements, and their mechanical properties were derived from their CT-derived ash-density. Tibiofemoral direction was used to define shear loading directions. The method was optimized and evaluated using clinical CT and micro-CT scans of cadaveric ankle specimens (n = 10). FEA stress propagation along Tb microstructures and nominal leakages over marrow space was confirmed. CT-derived shear modulus values were highly reproducible (ICC = 0.98) and high linear correlation (r ≥ 0.83) was observed with micro-CT-derived reference values. Nonlinear FEA using clinical CT imaging will broaden the scope of micro-mechanical analysis of Tb network at relatively low in vivo resolution alleviating the need for binary segmentation of Tb, while accounting for microdistribution of bone minerals.

Cadaveric computed tomography (CT) data are incredibly valuable resources that can support a variety of biomedical education and research endeavors.1–6 These tremendous sources of anatomical information can be used to generate cadaver-specific 3D anatomical models to greatly enhance the learning outcomes for gross anatomy students.7 Unfortunately, cadaveric CT data are also inherently challenging to properly utilize. The main obstacles to successfully extracting accurate anatomical structures for this objective include: the presence of encompassing preserved soft tissue, immediately adjacent anatomical structures of the same tissue type as that of the structure of interest, and CT imaging artifacts that arise from metal medical implants (see Fig. 1). Consequently, it is unfeasible for anatomy faculty or gross anatomy students to obtain cadaveric-specific 3D anatomical models. Furthermore, the challenge of extracting accurate 3D anatomical morphology from cadaveric CT data obstructs many biomedical research avenues. The current investigation establishes a protocol for extracting and analyzing the morphology of osteological structures from cadaveric CT data, which can be employed to augment gross anatomy curricula and galvanize biomedical research.

Knee osteoarthritis (OA) is typically well diagnosed with a clinical evaluation and confirmed by conventional radiographic imaging. While this disease is associated with pain and functional impairments, there is a well-documented discordance between radiographic severity and symptoms in knee OA patients.

In order to update a technical literature review from 2012 on the knee kinesiography, a comprehensive review was carried out to identify materials published since which used this technology to improve the understanding of the relationships between biomechanical dysfunctions and OA severity and progression (clinically and radiographically). This innovative exam, which can be performed with a KneeKG™ system, quickly assesses and quantifies knee joint function in the sagittal (flexion-extension), frontal (varus-valgus), and transverse (internal-external rotation) planes while the patient is walking on a commercial treadmill.

This review showed that biomechanical dysfunctions assessed through a knee kinesiography exam were most strongly associated with pain and function than OA radiographic severity. Furthermore, the added value of this assessment tool was highlighted in the primary care and total knee arthroplasty (TKA) populations. Objective data from this exam showed to be clinically relevant in conservative treatment care, as an input measure to identify patients deemed appropriate for surgery, and helped assessing how function is restored post-TKA while acquiring new insights on the choice of implant and surgical techniques.

This study suggests that the knee kinesiography can act as an add-on to conventional imaging to gather relevant and objective functional data to help clinicians better understand knee OA, its progression and impact of TKA.

The purpose of this study was to investigate if radiomic analysis based on spectral micro-CT with nanoparticle contrastenhancement can differentiate tumors based on tumor-infiltrating lymphocyte (TIL) burden. High mutational load transplant soft tissue sarcomas were initiated in Rag2+/- and Rag2-/- mice to model varying TIL burden. Mice received radiation therapy (20 Gy) to the tumor-bearing hind limb and were injected with a liposomal iodinated contrast agent. Five days later, animals underwent conventional micro-CT imaging using an energy integrating detector (EID) and spectral micro-CT imaging using a photon-counting detector (PCD). Tumor volumes, and iodine uptakes were measured. The radiomic features (RF) were grouped into feature-spaces corresponding to EID, PCD, and spectral decomposition images. RFs were ranked to reduce redundancy and increase relevance based on TIL burden. A leave one out strategy was used to assess separation using a neural network classifier. Tumor iodine concentration was the only significantly different conventional tumor metric between Rag2+/- (TILs present) and Rag2-/- (TIL-deficient) tumors. RFs further enabled differentiation between Rag2+/- and Rag2-/- tumors. The PCD-derived RFs provided the highest accuracy (0.84) followed by decomposition-derived RFs (0.78) and the EID-derived RFs (0.65). Such non-invasive approaches could aid in tumor stratification for cancer therapy studies.

The photon-counting (PC) detector technology promises to enhance the number of CT applications due to the spectral information. Of high interest for the cancer research community is imaging the tumor delivery of Cisplatin (CisPt), a chemotherapeutic agent utilized for treatment of numerous malignancies. CisPt contains platinum (Pt), a high-Z element material with a K-edge (78.4 keV) in the diagnostic spectrum. Our group has developed a preclinical prototype photon counting (PC) CT and applied it in cancer studies using nanoparticles. This study aims to investigate if CisPt can be imaged by K-edge spectral PCCT. Simulations and phantom experiments were performed to investigate CisPt detection using PCCT. We have selected scanning parameters (125 kVp) and energy thresholds (28, 34, 70, 78 keV) to enable K-edge separation of Pt and iodine (I) from calcium (Ca). The simulations include modeling of the polychromatic spectrum, and the PC detector response with spectral distortions. Two digital phantoms were used with maximum concentrations corresponding to low (2 mg/mL) and high (10 mg/mL) concentrations of I and Pt. A physical phantom with CisPt, I and Ca solutions was imaged both on our PC micro-CT and a novel clinical PCCT system. Material decompositions confirm the separation of Pt, Ca and I. However, low concentrations (<1 mg/mL) of CisPt are unlikely to be separated. Nevertheless, a liposomal nanoparticle-based CisPt formulation can enhance tumor delivery, via enhanced permeability and retention (EPR) and benefit from PCCT monitoring. Thus, depending on the levels of tumor accumulation, PCCT imaging of nanoparticles containing CisPt may become possible.

Gold nanoparticles (GNPs) are widely studied in medical research due to their favorable biocompatibility, variety in shape and size, and simple surface modification. Nanoparticles are particularly valuable in cancer research due to their enhanced permeation and retention effect property whereby nanoparticles accumulate in tumors. However, imaging GNPs are a challenge for most imaging modalities. Therefore, recent studies using x-ray fluorescence (XRF) imaging offer a potential for precise quantification and localization of GNPs without single endpoint studies such as immunohistochemistry and mass spectrometry. This study aims to accurately quantify and localize GNPs in ex-vivo tissue from GNP injected mice. 15-nm PEGylated GNPs were conjugated to anti-PSMA antibodies using 1-Ethyl-3-(3- dimethylaminopropyl) carbodiimide and N-hydroxy sulfosuccinimide. 3 SCID mice per group bearing subcutaneous LNCaP xenografts were intravenously injected with either anti-PSMA antibody conjugated GNPs (15mg/mL, 200μL) or Mouse IgG GNPs 24hrs prior to dissection. 11 organs along with the tumor were collected from the mice. Each organs’ GNP content was measured for quantification and localization via XRF on an in-house-developed dual-modality computed tomography and XRF system. Following imaging, organs were dehydrated and dissolved for quantification with inductively coupled plasma mass spectrometry analysis. XRF imaging quantified GNPs in tissue down to 25ng/g. Quantification with XRF imaging showed ~2x times greater accumulation of GNPs in the tumor with anti-PSMA targeted GNPs compared to Mouse IgG control GNPs. Additionally, XRF imaging of anti-PSMA targeted GNPs in all organs showed accurate quantification when compared to ICPMS analysis. XRF computed tomography further confirmed quantification of GNPs in tumors and spleen. This study confirmed the viability of XRF imaging for accurate quantification of anti-PSMA targeted GNPs in ex-vivo tissues.

Detection of left ventricular (LV) wall motion abnormalities (WMA) from 4DCT by visual interpretation is challenging. Quantitative assessment requires complex computation on multiple frames with large data sizes. Volume Rendering (VR) of the LV in CT across the cardiac cycle can enable the evaluation of 3D wall motion with significantly reduced data size. We propose a deep-learning (DL) framework to automate WMA detection in volume-rendered videos of clinical 4DCT studies. For 253 cardiac 4DCT studies, 6 VR videos depicting the LV were automatically generated corresponding to views rotated every 60 degrees around the long axis. Ground-truth WMA classification was performed for each video of the LV views by evaluating the extent of impaired regional shortening that was visible in that view. For DL prediction, videos were first processed by a pre-trained CNN “Inception V3” to extract image features. Then, extracted features from multiple frames were concatenated into a matrix and then input into a long short-term memory for the binary classification of WMA presence in the video. Studies were classified as abnormal if ≥2 out of 6 videos were abnormal. Studies were split chronologically so that the first 174 patients were used in 5-fold cross-validation and the final 79 studies were used in testing. VR significantly compressed data size (~800-fold). DL classification of WMA had high (<=89%) per-video and perstudy accuracy, sensitivity, and specificity for both cross-validation and testing cohorts. This novel method may offer a simple and accurate way to screen CT cases for WMA from highly compressed data.

Lung SBRT patients could have daily setup variations that lead to suboptimal treatment delivery. While currently a cone- beam CT (CBCT) is captured prior to each fraction for patient alignment, no organ contours or dosimetric calculations are routinely done to verify radiation therapy (RT) treatment delivery quality. Organ contours on CBCT are challenging because of the inferior image quality and low image contrast of CBCT. Besides, manual contouring is labor-intensive that prohibits clinical implementation. Therefore, if organ contours could be obtained automatically on fractional CBCT with limited human interventions, it would pave the way to obtain dosimetry for coverage evaluation and toxicity assessment. In this study, we developed a deep learning-based method for automated segmentation of multiple organs on CBCT images which simultaneously performs detection, classification, and segmentation of multi-organ. Our proposed hierarchical network method consists of four subnetworks: feature extractor, detector, hierarchical block, and mask module. The feature extractor subnetwork is used to extract informative features from CBCT. The detector subnetwork is used to locate the volume-of-interest (VOIs) of multiple organs. The hierarchical block network is used to enhance the feature contrast around organ boundaries and improve the organ classification. The mask module subnetwork then segments organ from the refined feature map within the VOIs. We conducted a five-fold cross-validation on 30 CBCTs. Five organs (esophagus, heart, spinal cord, left lung, and right lung) were segmented and compared with manual contours using several evaluation metrics. The Dice similarity coefficient (DSC) is 0.68, 0.87, 0.80, 0.91 and 0.93 for esophagus, heart, spinal cord, left lung, and right lung, respectively. These results demonstrate the feasibility and efficacy of our proposed hierarchical network method for CBCT lung segmentation, which could be used for fractional delivery dose evaluations in the future.

Multi-parametric magnetic resonance imaging (mp-MRI) is a promising tool for diagnosis of renal masses and may outperform computed tomography (CT) to differentiate between benign and malignant renal masses due to superior soft tissue contrast. Deep learning (DL)-based methods for kidney segmentation are under-explored in mp-MRI which consists of several pulse sequences, including primarily T2-weighted (T2W) and contrast-enhanced (CE) images. Multi-parametric MRI images have domain shift due to differences in acquisition systems and image protocols, leading to lack of generalizability of methods for image segmentation. To perform similar automated kidney segmentation on another mp- MRI sequence, the model needs a large dataset with manual segmentations to train a model from scratch, which is labor intensive and time consuming. In this paper, we first trained a DL-based method using 108 cases of labeled data to contour kidneys using T1 weighted-Nephrographic Phase CE-MRI (T1W-NG). We then applied a transfer learning approach to other mp-MRI images using pre-trained weights from the source domain, thus eliminating the need for large manually annotated datasets in target domain. The fully automated 2D U-Net for kidney segmentation in source domain containing total 108 3D images of T1W-NG, yielded Dice-similarity coefficient (DSC) of 0.91 ± 0.07 on test cases. The transfer learning of pretrained weights of T1W-NG model on the smaller target domain T2W dataset containing total 50 3D images for automated kidney segmentation generated DSC of 0.90 ± 0.06 (p<0.05), which was an improvement of 3.43% in DSC by compared to the without transfer learning approach (T2W-UNet model).

Computed tomography (CT) is commonly used for the characterization and tracking of abdominal muscle mass in surgical patients for both pre-surgical outcome predictions and post-surgical monitoring of response to therapy. In order to accurately track changes of abdominal muscle mass, radiologists must manually segment CT slices of patients, a time-consuming task with potential for variability. In this work, we combined a fully convolutional neural network (CNN) with high levels of preprocessing to improve segmentation quality. We utilized a CNN based approach to remove patients’ arms and fat from each slice and then applied a series of registrations with a diverse set of abdominal muscle segmentations to identify a best fit mask. Using this best fit mask, we were able to remove many parts of the abdominal cavity, such as the liver, kidneys, and intestines. This preprocessing was able to achieve a mean Dice similarity coefficient (DSC) of 0.53 on our validation set and 0.50 on our test set by only using traditional computer vision techniques and no artificial intelligence. The preprocessed images were then fed into a similar CNN previously presented in a hybrid computer vision-artificial intelligence approach and was able to achieve a mean DSC of 0.94 on testing data. The preprocessing and deep learning-based method is able to accurately segment and quantify abdominal muscle mass on CT images.

Magnetic resonance imaging (MRI) is a widely used modality for visualizing patient anatomy due to its non-invasiveness and superior soft tissue contrast. MRI can be used to quantify anatomical changes before and after therapy to assess treatment outcome and efficacy. However, longitudinal tracking depends on the accurate alignment of multiple image sets, which is challenged by rigid and deformable displacements. Deformable image registration is a promising tool to account for these changes and enable accurate longitudinal assessments. In this study, we aim to develop a deep learning-based method for automatic deformable registration to align post-treatment and pre-treatment head and neck (HN) MRIs. Our proposed method, named dual-feasible framework, is implemented by a mutual network that consists of a forward module and a backward module. The two modules alternate in generating a deformation vector field (DVF) for image registration. First the pre-treatment MRI is registered to the post-treatment MRI and then the post-treatment MRI is registered to the pre-treatment MRI, and the process repeats under a mutual enhancing strategy. Dual feasible loss is used to optimize the mutual network. We conducted longitudinal experiments on 4 public patient datasets (40 MRI scans), each with 2 head and neck (HN) MRI sequences (T1-weighted and T2-weighted) across 5 timepoints: one pre-treatment MRI and four posttreatment MRIs. To evaluate the proposed method, the pre-treatment MRIs were used as the target, and we calculated the peak-signal-to-noise ratio (PSNR), structural similarity index (SSIM) and mean absolute error (MAE) between the deformed post-treatment MRI and the pre-treatment MRI. The PSNR, SSIM and MAE are 29.3±0.3 dB, 0.89±0.02 and 52.4±2.7 for the T1-weighted MRI, and are 27.0±0.8 dB, 0.87±0.03 and 97.5±14.1 for the T2-weighted MRI. These results demonstrate the feasibility and efficacy of our proposed method for MRI deformable image registration.

The Centiloid (CL) scale represents an attempt to standardize amyloid positron emission tomography (PET) imaging to create a measure of amyloid-β (Aβ) deposition in the brain. CL converts the standard uptake value ratio (SUVR) from brain PET into values from 0 to 100 using paired magnetic resonance (MR) images in the processing pipeline. This study aims to evaluate the CL variability with processing pipelines without MR (MR-less). Image data of 79 individuals (34 young controls; 45 Alzheimer's disease patients) from the GAAIN database were processed with three pipelines: one MRbased, and two MR-less, using an MR template smoothed with Gaussian filters (4 mm/8 mm). Using the whole cerebellum as a reference, we used PMOD - PNEURO tool to convert SUVR from the global cortical target region. We find a strong agreement between the MR-based pipeline and previous studies (R² = 0.997), exceeding the minimum acceptance criteria (slope = 0.99; intercept = 0.92). For the 4 mm Gaussian filter MR-less pipeline, we reach a strong agreement (R² = 0.981) and minimum acceptance criteria (slope = 0.98; intercept = 1.29). However, results for the 8 mm filter Gaussian MR-less pipeline are below the acceptance criteria (R² = 0.969, slope = 0.97; intercept = 1.52), mainly for higher CL values. No statistical differences were found comparing MR-based vs. MR-less and GAAIN vs. MR-less for both filters. Concluding, CL implementation using 4 mm Gaussian filter MR-less pipeline is comparable to the paired MR-based, simplifying the clinical practice to quantify Aβ deposition in the brain. However, MR-based processing for brain quantification is indicated mainly for clinical research in aging studies.

Purpose: Subarachnoid Hemorrhage (SAH) is a lethal hemorrhagic stroke that account for 25% of cerebrovascular deaths. As a result of the initial bleed, a chain of physiological events are initiated which may lead to Delayed Cerebral Ischemia (DCI). As of now we have no diagnostic capability to identify patients which may present DCI a few weeks after initial presentation. We propose to investigate whether a data driven approach using angiographic parametric imaging (API) may predict occurrence of the DCI. Materials and Methods: Digital Subtraction Angiographic (DSA) sequences from 125 SAH patients were used retrospectively to perform API assessment of the entire brain hemisphere where the hemorrhage was detected. Four Regions of Interests (ROIs) were placed to extract five average API biomarkers in the lateral and AP DSAs. Data driven analysis using Logistic Regression was performed for various API parameters and ROIs to find the optimal configuration to maximize the prognosis accuracy. Each model performance was evaluated using area under the curve of the receiver operator characteristic (AUROC). Results: Data driven approach with API has a 60% accuracy predicting DCI occurrence. We determined that location of the ROI for extraction of the API parameters is very important for the data driven model performance. Normalizing the values using the inlet velocities for each patient yield higher and more consistent results. Single API biomarkers models had poor prediction accuracies, barely better than chance. Conclusions: This effectiveness exploratory study demonstrates for the first time, that prognosis of the DCI in SAH patients, is feasible and warrants a more in-depth investigation.

Obstructive sleep apnea syndrome (OSAS) is one of the most common sleep disorders that endangers human health, which is associated with episodes of apnea or hypopnea during sleep. In children, OSAS is associated with cardiovascular morbidity, neurobehavioral deficits, and poor quality of life, which highlights the importance for early diagnosis and treatment. Recent studies using dynamic magnetic resonance imaging (dMRI) have shown that adults with OSAS exhibit airway narrowing in specific regions that display increased variability in diameter during sleep as compared with controls. In this paper, we propose a novel method to compare OSAS patients with control subjects during awake and asleep states to assess the regional dynamic changes that occur in specific locations of the upper airway. Firstly, we segment the 3D upper airway with a previously developed fully automatic method. Then, different types of breathing cycles are selected by experts based on polysomnography. For each cycle, we calculate the distance of each point on the surface of the upper airway from end-expiration (EE) to end-inspiration (EI), which is then utilized for subsequent motion analysis. The 3D upper airway is subsequently divided into 4 anatomical parts manually. Lastly, comparisons of the dynamic upper airway motion measurements from different cycle groups are performed between OSAS patients and control subjects. Comparisons of different types of cycles within the same anatomical part demonstrated significant differences between control subjects and OSAS patients in all anatomic parts with some exceptions. These novel observations may provide some insights into the pathophysiology of OSAS.

Lipedema is a painful connective tissue disease involving excessive subcutaneous adipose tissue (SAT) accumulation in the lower extremities. Lipedema remains poorly recognized as a unique clinical entity and is often misdiagnosed as obesity. Whole-body magnetic resonance imaging (MRI) acquisitions could provide insight into the unique body composition of lipedema, yet methodologies for multi-slice analyses are lacking. In this work, a semi-automated processing workflow was developed to segment and quantify adiposity from whole-leg chemical-shift encoded (CSE) MRI to distinguish lipedema. Patients with lipedema (N=15) and controls (N=13) matched for age and body mass index underwent a CSE MRI exam in eight stacks from the head-to-ankles. Slices from thighs-to-ankles were segmented via Chan-Vese segmentation, clustering, and morphological techniques to separate SAT and skeletal muscle. SAT and muscle volume per slice and the SAT-to-muscle volume ratio were recorded in decades of slices and compared between groups using Mann-Whitney U test with two-sided significance criteria p<0.05. SAT volume was significantly elevated in participants with lipedema in all decades (p<0.001), while muscle volume was not significantly different. SAT-to-muscle volume ratio was elevated in lipedema compared to controls (p<0.001), with the greatest effect size (rrb = 0.74) observed in the eighth decade corresponding to the mid-thigh region. These findings reveal SAT distribution is uniquely elevated throughout the legs of participants with lipedema as discerned from whole-leg CSE MRI. CSE MRI and analysis methods developed herein for SAT quantification could inform the diagnosis of lipedema, which suffers from few objective strategies to differentiate the disease from obesity.

Radiomic studies utilize AI and quantitative features from medical images to create models that can predict patient outcomes. An integral step in these radiomic studies is the delineation of the regions of interest where the features are extracted. Manual segmentation is labor intensive and time-consuming for large studies. Semi-automatic segmentation tools have been used in recent radiomic studies to achieve more reproducible segmentations and robust radiomics features. However, for the segmentation of lung tumors on CT images, tools in the literature are difficult to find publicly and require extensive user interaction. Therefore, we aimed to build a semi-automatic segmentation tool which was intuitive, fast, and required minimal user interaction. We used one dataset to develop the segmentation algorithm on (n=49), and another to test its performance (n=144). All 144 tumors were segmented on the CT images using the semiautomatic tool by three separate users. A gold standard tumor delineation was determined by a trained radiologist. The segmentation robustness was assessed using the Dice, mean absolute boundary distance (MAD) and volume difference (VD). A total of 408 radiomic features were extracted and feature robustness was determined using an intra-class correlation coefficient (ICC) greater than 0.8. The developed tool achieved an average Dice of 0.90, MAD of 0.62 mm and a VD of 0.97 ml between the three users. A total of 181 (76%) of the extracted features displayed excellent reliability. This tool has the potential to augment the reliability of radiomic studies by making segmentations and feature sets more reproducible.

Image co-registration is an important tool that is commonly used to quantitatively or qualitatively compare information from images or data sets that vary in time, origin, etc. This research proposes a method for the semi-automatic coregistration of the 3D vascular geometry of an intracranial aneurysm to novel high-speed angiographic (HSA) 1000 fps projection images. Using the software Tecplot 360, 3D velocimetry data generated from computational fluid dynamics (CFD) for patient-specific vasculature models can be extracted and uploaded into Python. Dilation, translation, and angular rotation of the 3D velocimetry data can then be performed in order to co-register its geometry to corresponding 2D HSA projection images of the 3D printed vascular model. Once the 3D CFD velocimetry data is geometrically aligned, a 2D velocimetry plot can be generated and the Sørensen–Dice coefficient can be calculated in order to determine the success of the co-registration process. The co-registration process was performed ten times for two different vascular models and had an average Sørensen–Dice coefficient of 0.84 ± 0.02. The method presented in this research allows for a direct comparison between 3D CFD velocimetry data and in-vitro 2D velocimetry methods. From the 3D CFD, we can compare various flow characteristics in addition to velocimetry data with HSAderived flow metrics. The method is robust to other vascular geometries as well.

Severe arterial stenosis encompasses complex flow structures especially when the blood flow rate exceeds the critical Reynolds number (Re ≥ 2000), resulting in ow instability and turbulence. Uncovering reduced-order ow characteristics in blood flow data facilitate understanding flow physics and efficient data-driven modeling. In this paper, we used Computational Fluid Dynamics (CFD) and 4D flow MRI data in a phantom model of arterial stenosis with 87% degree of narrowing for performing Proper Orthogonal Decomposition (POD) and Dynamic Mode Decomposition (DMD) on the velocity and pressure data. We found the required modes to reconstruct the CFD and 4D flow MRI velocity and pressure data in the phantom model and identified the most energetic modes with temporal dynamics of coherent structures. In addition, we evaluated the compromise between the simplicity and accuracy of the reconstructed data. These data-driven modeling techniques have the potential to reduce the complexity of 4D flow MRI data. We envisage that it can ultimately be applied to enhancing the resolution, denoising 4D flow MRI data, and impacting data collection requirements.

A fundamental understanding of sex differences that exist in healthy individuals is critical for the study of neurological illnesses that exhibit phenotypic differences between both genders. Functional magnetic resonance imaging(fMRI) is a useful way to study this problem since it provides a non-invasive and high-resolution tool for observing the fluctuation in blood oxygenation level dependent (BOLD) signals to characterize the metabolism of the human brain. In the meantime, graph neural networks (GNNs) can be applied to fMRI data to effectively discover novel biomarkers underlying brain development. We propose a multi-modal graph isomorphism network (MGIN) to analyze the sex differences based on fMRI task data. Our method is able to integrate all the available connectivity data into graphs for deep learning, and it can be applied to multigraphs with different nodes to learn local graph information without binding to the entire graph. MGIN model can identify important subnetworks between and within multi-task data. In addition, it is interpretable by using GNNExplainer to provide important domain insights to identify graph structures and node features that contribute significantly to the classification results. Our MGIN model can achieve better classification accuracy compared to competing models. We applied the model to a cohort of brain development study to classify sex during different stages of adolescence and experimental results showed that our model can improve classification accuracy and help in our understanding of neurodevelopment during adolescence.

In this work, we propose a novel GCN based Residual connected (GCN-RC) network to improve the quality of Fluorescence molecular tomography (FMT) morphological reconstruction. Instead of using a simplified linear model of photon propagation for FMT reconstruction, the method can directly construct a nonlinear mapping relationship between the surface photon density and internal fluorescent source. In order to validate the reconstruction performance of GCN-RC, we performed numerical simulation experiments and in vivo experiments based on tumor-bearing mice. Both numerical simulated and in vivo experimental results demonstrated that GCN-RC achieved improved reconstruction in terms of both source localization and morphology recovery.

Structural substrates of sex differences in human function and behavior have been elucidated in previous studies. Diffusion weighted magnetic resonance imaging (DW-MRI) is a widely used non-invasive imaging technique in studying human brain white matter structural organization. While many DW-MRI studies reporting sex differences in WM structure are based on diffusion tensor imaging (DTI) measures, tract specific microstructural differences require further investigation. In this study, we aim to investigate sex differences and sex-specific hemispheric differences in white matter microstructural development in healthy 8-year-old children based on novel track weighted imaging (TWI) analysis. Average pathlength map (APM) is a TWI contrast in which the average length of fibers passing through a voxel is utilized. In this study, we employed tract specific APM measures to evaluate sex differences in WM microstructural development. A total of 37 WM tracts were analyzed including 7 commissural tracts, 9 bilateral association tracts and 6 bilateral projection tracts. APM maps were generated for each tract. Tract-wise group tests were done using the mean values of APM maps. Sex differences were tested using general linear model based group comparisons. Age and total brain volume were included as covariates in the group analysis. Sex specific hemispheric differences were performed for the 15 bilateral tracts. One sample t-tests were done independently for left<right and right<left cases and the APM measures were controlled for age and total cerebral hemispheric volume. P-values<0.05 were considered significant after correcting for multiple comparisons accounting for the total number of tracts. Significant sex differences were revealed in APM measures between boys and girls in 11 WM tracts including rostral body of corpus callosum (CC), left inferior fronto-occipital fasciculus (IFOF), right cingulum, bilateral first and second segments of superior longitudinal fasciculus (SLF), right middle longitudinal fasciculus (MLF), bilateral fronto-pontine (FPT) and right parieto-occipital pontine tracts (POPT). The sex differences showed higher APM values for these 11 tracts in boys as compared to that of girls. In hemispheric differences analysis for both boys and girls, 2 tracts, arcuate fasciculus and optic radiation showed higher APM in left tracts as compared right; 5 tracts, IFOF, MLF, third segment of SLF, FPT and superior thalamic radiation showed higher APM in right tracts as compared to left. This indicates that boys and girls possess similar lateral asymmetries in these 7 tracts. Additionally, anterior thalamic radiation (ATR) showed higher APM in left tract and 4 tracts, first segment of SLF, POPT, inferior longitudinal fasciculus and cortico-spinal tract showed higher APM in right for boys. In girls, second segment of SLF and uncinate fasciculus showed higher APM in right hemisphere. These results indicate different lateral asymmetries between boys and girls for 7 tracts. Overall, boys showed higher average fiber length in most of the tracts, even after controlling for total brain volume.

In radiation therapy, regional lung function assessment can help physicians make decision to spare healthy and functional lung tissues while maintaining target coverage. Fluorodeoxyglucose (F-18 FDG) PET/CT is usually acquired once before treatment to evaluate nodal involvement and enhance the tumor region for target definition. The F-18 FDG PET/CT could highlight the FDG-avid tumor regions, however, it does not provide regional lung function information, such as ventilation, perfusion, and tissue elasticity. 4DCT is routinely acquired in CT simulation to assess the tumor motion for respiratory motion management and ITV contouring. 4DCT which bins the whole respiratory cycle into ten phases can be used to model lung motion and evaluate regional lung function. In this study, we proposed a deep learning (DL) based image registration method to derive lung volumetric strain from the deformation vector field (DVF). A total of 20 4DCT datasets were used to train the network and another 5 datasets including both 4DCT and F-18 FDG PET/CT from lung cancer patients were collected to evaluate the performance of the proposed method. The proposed DL-based registration network was trained to predict the DVF to register a pair of 3DCT images. The resultant DVF was used to calculate the volumetric strain for regional lung function evaluation. To assess the accuracy of volumetric strain map in lung function evaluation, the strain map was used to contour regions with low absolute volumetric strain, indicating stiff tissue regions, and compared to the tumor region segmented on the CT images. Dice similarity coefficient (DSC) was reported for 5 testing patients. The average DSC is 0.65±0.07, indicating the accuracy of volumetric strain in regional lung stiffness evaluation. The proposed method has the potential to be used in regional lung function evaluation and lung tissue stiffness-based lung tumor staging.

As of 14 December 2021, severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the virus that causes coronavirus disease 2019 (COVID-19), caused nearly 269 million confirmed cases and almost 5.3 million deaths worldwide. Chest computed tomography (CT) has high diagnostic sensitivity for the detection of pulmonary disease in COVID-19 patients. Toward timely and accurate clinical evaluation and prognostication, radiomic analyses of CT images have been explored to investigate the correlation of imaging and non-imaging clinical manifestations and outcomes. Delta (∆) radiomics optimally performed from pre-infection to the post-critical phase, requires baseline data typically not obtained in clinical settings; additionally, their robustness is affected by differences in acquisition protocols. In this work, we investigated the reliability, sensitivity, and stability of whole-lung radiomic features of CT images of nonhuman primates either mock-exposed or exposed to SARS-CoV-2 to study imaging biomarkers of SARS-CoV-2 infection. Images were acquired at a pre-exposure baseline and post-exposure days, and lung fields were segmented. The reliability of radiomic features was assessed, and the dynamic range of each feature was compared to the maximum normal intra-subject variation and ranked.

Laser speckle flowgraphy (LSFG) is a non-invasive imaging technology for quantifying microvascular blood flow. In the eye, LSFG quantifies the relative dynamics of blood flow of the retina, choroid and optic nerve head on a continuous scale. Currently, LSFG analysis requires the placement of “rubber bands” (defining regions of interest) to measure blood flow at desired locations. However, the placement of rubber bands requires knowledge of which regions are likely to be affected by disease. Here, we demonstrate a fully automated superpixel-histogram method without rubber band placement to determine regional blood flow abnormalities. Regional blood flow patterns were quantified via superpixel distributions of mean blur rate (MBR, linearly proportional to blood flow) and percentages of total superpixels at five pre-defined ranges of blood flow. We applied the proposed method to help diagnose acute arteritic anterior ischemic optic neuropathy (AAION) and found that compared to normal eyes, acute AAION eyes showed a significant blood flow reduction of the choroid due to the effect of giant cell arteritis on the posterior ciliary arteries (supplying the choroid and optic nerve circulation). The proposed method demonstrated a novel approach for quantifying abnormal regions of blood flow in different vascular beds caused by disease.

In vivo imaging, particularly micro-computed tomography (micro-CT), uses contrast agents to visualize the subject's vasculature. Small rodents have small blood volumes which limit the injectable dosage of contrast agent into their blood. To increase the contrast enhancement of the reconstructed images, iodine concentration is increased, and the injectable volume is decreased from 0.01 mL/g of 50 mg/mL Fenestra VC to 0.005 mL/g of 100 mg/mL Fenestra HDVC. Fenestra VC and Fenestra HDVC were diluted into 9 iodine concentrations ranging between 0-50 mg/mL for Fenestra VC and 0- 100 mg/mL for Fenestra HDVC. The dilution vials were imaged using the in vivo micro-CT scanner and measured for the mean CT numbers and standard deviations. The results graphed on a dilution graph shows that Fenestra HDVC with half the dosage volume has the same contrast enhancement as Fenestra VC for the same iodine concentration. The micro- CT scans for ten C57BL/6 female mice were used to determine the mean CT numbers using MicroView for the right atrium, liver, air, bladder, leg muscle, kidney, spleen, and vena cava. These regions were measured using 1mm3 region of interest at 8 time points: pre-contrast, post-contrast at t = 0, 8, 24, 48, 72 hours, 1 and 2 weeks. These in vivo values were compared between mice injected with Fenestra VC and Fenestra HDVC to show that half the dosage volume using Fenestra HDVC enhances like the full dosage volume using Fenestra VC. Thus, Fenestra HDVC may be a safer choice for in vivo rodent imaging.

White matter lesion (WML) segmentation applied to magnetic resonance imaging (MRI) scans of people with multiple sclerosis has been an area of extensive research in recent years. As with most tasks in medical imaging, deep learning (DL) methods have proven very effective and have quickly replaced existing methods. Despite the improvement offered by these networks, there are still shortcomings with these DL approaches. In this work, we compare several DL algorithms, as well as methods for ensembling the results of those algorithms, for performing MS lesion segmentation. An ensemble approach is shown to best estimate total WML and has the highest agreement with manual delineations.

Absolute myocardial perfusion imaging (MPI) can be beneficial in the diagnosis and prognosis of patients with coronary artery disease. However, validation and standardization of perfusion estimates across centers is needed to ensure safe and adequate integration into clinical routine. MPI phantoms can contribute to this clinical need as these models can provide ground truth evaluation of absolute MPI in a simplified, though controlled setup. This work presents verification of phantom design choices, including the justification for using sorbents in mimicking contrast kinetics (i.e., tracer uptake and retention). Moreover, we compare preliminary phantom results obtained with SPECT-MPI with a patient example. Finally, we applied a general two-tissue compartment model to describe the obtained phantom time activity curve data. These evaluation steps support shaping of a suitable verification and validation strategy for the multimodal myocardial perfusion phantom design and realization.

Human papillomavirus (HPV) remains a leading cause of virus-induced cancers. Hence early detection of HPV plays a crucial role in providing timely, optimal and effective intervention before such a cancer develops. While conventional light microscopy constitutes one of inseparable tools applied for studying biological cell structures, its low resolution at ~100nm per pixel falls short of detecting HPV that typically has a size of 52 to 55nm in diameter, giving rise to visualisation of HPV and subsequent evaluation of the efficacy of anti-HPV drugs at such sub-pixel level a challenging task if not overwhelmingly. This study employs an explainable deep learning network of texture transformer (TTSR) to up sample by four folds (×4). In comparison with other super resolution approaches, TTSR appears to perform the best with PSNR and SSIM being 28.70 and 0.8778 respectively whereas 25.80/0.7910, 18.35/0.5059. 30.31/0.8013, and 28.07/0.6074 are observed for the methods of RCAN, Pix2Pix, CycleGAN, and ESRGAN respectively. Significantly, the training pairs of TTSR does not need to be precisely match between low (LR) and high resolution (HR) images since the LR and HR images, which are required by many other super resolution approaches. This work constitutes one of the first to detect HPV applying explainable deep learning network, which will lead to the real world implementation to evaluate the efficacy of the developed anti-HPV drugs.

Extra-cellular volume (ECV) mapping cardiac magnetic resonance (CMR) imaging allows for the characterization of expanded myocardial extracellular space, a common feature of myocardial fibrosis (MF). Quantification of MF is feasible using ECV mapping techniques; however, prior manual delineation of the endocardial and epicardial borders is required. In this study, we propose a method for automated myocardial delineation of ECV maps using convolutional neural networks (CNNs). We compare two methods based on the standard U-Net and the U-Net++ architectures using a five-fold cross validation on basal, mid, and apical short-axis ECV maps of the left ventricle (LV) in 73 patients with ischemic (n=38) or dilated (n=35) cardiomyopathies. The standard U-Net and U-Net++ -based architectures yielded DSC metrics of 87.61% and 87.89%, respectively, against manual contours derived by an expert. Precision and recall were reported >85% and relative error <12% for both CNNs. The U-Net++ architecture outperformed the standard U-Net on the order of 1-2% for all metrics. An inter-operator variability analysis was performed on a subset of myocardial contours derived by three operators. The inter-operator analysis demonstrated significant differences in the distribution of myocardial ECV values among three operators as per the Kruskal-Wallis H-test (average pair-wise P-value = 0.040), but operator differences failed to show significance against U-Net++ or standard U-Net (average pair-wise P-value 0.055 and 0.060, respectively). Correlation of global ECV improved for operators against U-Net++ (𝜌=0.88) and against standard U-Net (𝜌=0.877) compared to correlation of global ECV values between all operators (ρ=0.828).

Automatic multi-organ segmentation is a cost-effective tool for generating organ contours using computed tomography (CT) images. This work proposes a deep-learning algorithm for multi-organ (bladder, prostate, rectum, left and right femoral heads) segmentation in pelvic CT images for prostate radiation treatment planning. We propose an encoder-decoder network with a V-net backbone for local feature extraction and contour reconstruction. Novel to our network, we utilize a token-based transformer, which encourages long-range dependency, to forward more informative high-resolution feature maps from the encoder to the decoder. In addition, a knowledge distillation strategy was applied to improve the network’s generalization. We evaluate the network using a dataset collected from 50 patients with prostate cancer. A quantitative evaluation of the proposed network’s performance was performed on each organ based on: 1) volume similarity between the segmented contours and ground truth using Dice score, segmentation sensitivity, precision, and absolute percentage volume difference (AVD), 2) surface similarity evaluated by Hausdorff distance (HD), mean surface distance (MSD) and residual mean square distance (RMSD). The performance was then evaluated against other state-of-art methods. The average volume similarities achieved by the network over all organs were: Dice score = 0.83, sensitivity = 0.84, and precision = 0.83; the average surface similarities were HD = 5.77mm, MSD = 0.93mm, RMSD = 2.77mm, and AVD =12.85%. The proposed methods performed significantly better than competing methods in most evaluation metrics. The proposed network may be a promising segmentation approach for use in routine prostate radiation treatment planning.

Aortic dissection is an acute condition of the aorta. It typically starts with an intimal tear and continues with the separation of the aortic wall layers. This situation typically leads to the creation of a second lumen, i.e., the false lumen, where blood can flow into. For diagnosis of this pathology, computed tomography angiography (CTA) is usually used. To have a better understanding of its causes and for measuring cross-sectional caliber at onset and at each follow-up, segmentation of true and false lumen is important in clinical use. In this work, a pipeline for aortic dissection segmentation is evaluated to obtain the correct visualization of true and false lumen separated by the dissection flap that characterizes this pathology. We provide an evaluation of three different vessel enhancement filters, used as a preprocessing step, through both a qualitative and quantitative evaluation.