Deep Learning Segmentation of the Right Ventricle in Cardiac MRI: The M&Ms Challenge.

In recent years, several deep learning models have been proposed to accurately quantify and diagnose cardiac pathologies. These automated tools heavily rely on the accurate segmentation of cardiac structures in MRI images. However, segmentation of the right ventricle is challenging due to its highly complex shape and ill-defined borders. Hence, there is a need for new methods to handle such structure's geometrical and textural complexities, notably in the presence of pathologies such as Dilated Right Ventricle, Tricuspid Regurgitation, Arrhythmogenesis, Tetralogy of Fallot, and Inter-atrial Communication. The last MICCAI challenge on right ventricle segmentation was held in 2012 and included only 48 cases from a single clinical center. As part of the 12th Workshop on Statistical Atlases and Computational Models of the Heart (STACOM 2021), the M&Ms-2 challenge was organized to promote the interest of the research community around right ventricle segmentation in multi-disease, multi-view, and multi-center cardiac MRI. Three hundred sixty CMR cases, including short-axis and long-axis 4-chamber views, were collected from three Spanish hospitals using nine different scanners from three different vendors, and included a diverse set of right and left ventricle pathologies. The solutions provided by the participants show that nnU-Net achieved the best results overall. However, multi-view approaches were able to capture additional information, highlighting the need to integrate multiple cardiac diseases, views, scanners, and acquisition protocols to produce reliable automatic cardiac segmentation algorithms.

Overall Survival Prediction of Glioma Patients With Multiregional Radiomics.

Radiomics-guided prediction of overall survival (OS) in brain gliomas is seen as a significant problem in Neuro-oncology. The ultimate goal is to develop a robust MRI-based approach (i.e., a radiomics model) that can accurately classify a novel subject as a short-term survivor, a medium-term survivor, or a long-term survivor. The BraTS 2020 challenge provides radiological imaging and clinical data (178 subjects) to develop and validate radiomics-based methods for OS classification in brain gliomas. In this study, we empirically evaluated the efficacy of four multiregional radiomic models, for OS classification, and quantified the robustness of predictions to variations in automatic segmentation of brain tumor volume. More specifically, we evaluated four radiomic models, namely, the Whole Tumor (WT) radiomics model, the 3-subregions radiomics model, the 6-subregions radiomics model, and the 21-subregions radiomics model. The 3-subregions radiomics model is based on a physiological segmentation of whole tumor volume (WT) into three non-overlapping subregions. The 6-subregions and 21-subregions radiomic models are based on an anatomical segmentation of the brain tumor into 6 and 21 anatomical regions, respectively. Moreover, we employed six segmentation schemes - five CNNs and one STAPLE-fusion method - to quantify the robustness of radiomic models. Our experiments revealed that the 3-subregions radiomics model had the best predictive performance (mean AUC = 0.73) but poor robustness (RSD = 1.99) and the 6-subregions and 21-subregions radiomics models were more robust (RSD  1.39) with lower predictive performance (mean AUC  0.71). The poor robustness of the 3-subregions radiomics model was associated with highly variable and inferior segmentation of tumor core and active tumor subregions as quantified by the Hausdorff distance metric (4.4-6.5mm) across six segmentation schemes. Failure analysis revealed that the WT radiomics model, the 6-subregions radiomics model, and the 21-subregions radiomics model failed for the same subjects which is attributed to the common requirement of accurate segmentation of the WT volume. Moreover, short-term survivors were largely misclassified by the radiomic models and had large segmentation errors (average Hausdorff distance of 7.09mm). Lastly, we concluded that while STAPLE-fusion can reduce segmentation errors, it is not a solution to learning accurate and robust radiomic models.

A systematic evaluation of learning rate policies in training CNNs for brain tumor segmentation.

Convolutional neural networks (CNNs) have recently emerged as a powerful approach for automatic segmentation of brain tumor subregions on 3D multi-parametric MRI scans. Learning rate is a crucial hyperparameter in the training of CNNs, impacting the performance of the learned model. Different learning rate policies trace unique trajectories in the optimization landscape that converge to local minima with varying generalization properties. In this work, we empirically evaluated nine learning rate policy-optimizer pairs with two state-of-the-art architectures, namely 2D slice-based U-Net and 3D DeepMedicRes, on an augmented brain tumor dataset of 534 subjects. Segmentation performance was quantified in terms of Dice similarity coefficient and Hausdorff distance metrics. The policies were ranked based on the final ranking score (FRS) employed by the BraTS challenge, with the statistical significance of the rankings evaluated by random permutation test. For 2D slice-based U-Net architecture, an overall ranking of learning rate policies showed that the polynomial decay policy with Adam optimizer significantly outperformed other policies for the task of individual and hierarchical segmentation of tumor subregions (p< 10-4). For 3D segment-based DeepMedicRes architecture, polynomial decay policy with Adam optimizer performed significantly better than all other policies, with the exception of polynomial decay with SGD optimizer for the same task (p< 10-4). Based on the FRS, polynomial decay policy with Adam and SGD optimizer occupied the top two positions respectively, but the difference was not statistically significant (p> 0.3). These findings were also validated on the BraTS 2019 Validation dataset which comprised of an additional 125 subjects.

Enhancement of dynamic myocardial perfusion PET images based on low-rank plus sparse decomposition.

BACKGROUND AND OBJECTIVE: The absolute quantification of dynamic myocardial perfusion (MP) PET imaging is challenged by the limited spatial resolution of individual frame images due to division of the data into shorter frames. This study aims to develop a method for restoration and enhancement of dynamic PET images. METHODS: We propose that the image restoration model should be based on multiple constraints rather than a single constraint, given the fact that the image characteristic is hardly described by a single constraint alone. At the same time, it may be possible, but not optimal, to regularize the image with multiple constraints simultaneously. Fortunately, MP PET images can be decomposed into a superposition of background vs. dynamic components via low-rank plus sparse (L + S) decomposition. Thus, we propose an L + S decomposition based MP PET image restoration model and express it as a convex optimization problem. An iterative soft thresholding algorithm was developed to solve the problem. Using realistic dynamic 82Rb MP PET scan data, we optimized and compared its performance with other restoration methods. RESULTS: The proposed method resulted in substantial visual as well as quantitative accuracy improvements in terms of noise versus bias performance, as demonstrated in extensive 82Rb MP PET simulations. In particular, the myocardium defect in the MP PET images had improved visual as well as contrast versus noise tradeoff. The proposed algorithm was also applied on an 8-min clinical cardiac 82Rb MP PET study performed on the GE Discovery PET/CT, and demonstrated improved quantitative accuracy (CNR and SNR) compared to other algorithms. CONCLUSIONS: The proposed method is effective for restoration and enhancement of dynamic PET images.

Quantification of intramyocardial blood volume with 99mTc-RBC SPECT-CT imaging: A preclinical study.

BACKGROUND: Currently, there is no established non-invasive imaging approach to directly evaluate myocardial microcirculatory function in order to diagnose microvascular disease independent of co-existing epicardial disease. In this work, we developed a methodological framework for quantification of intramyocardial blood volume (IMBV) as a novel index of microcirculatory function with SPECT/CT imaging of 99mTc-labeled red blood cells (RBCs). METHODS: Dual-gated myocardial SPECT/CT equilibrium imaging of 99mTc-RBCs was performed on twelve canines under resting conditions. Five correction schemes were studied: cardiac gating with no other corrections (CG), CG with attenuation correction (CG + AC), CG + AC with scatter correction (CG + AC + SC), dual cardiorespiratory gating with AC + SC (DG + AC + SC), and DG + AC + SC with partial volume correction (DG + AC + SC + PVC). Quantification of IMBV using each approach was evaluated in comparison to those obtained from all corrections. The in vivo SPECT estimates of IMBV values were validated against those obtained from ex vivo microCT imaging of the casted hearts. RESULTS: The estimated IMBV with all corrections was 0.15 ± 0.03 for the end-diastolic phase and 0.11 ± 0.03 for the end-systolic phase. The cycle-dependent change in IMBV (ΔIMBV) with all corrections was 23.9 ± 8.6%. Schemes that applied no correction or partial correction resulted in significant over-estimation of IMBV and significant under-underestimation of ΔIMBV. Estimates of IMBV and ΔIMBV using all corrections were consistent with values reported in the literature using invasive techniques. In vivo SPECT estimates of IMBV strongly correlated (R2 ≥ 0.70) with ex vivo measures for the various correction schemes, while the fully corrected scheme yielded the smallest bias. CONCLUSIONS: Non-invasive quantification of IMBV is feasible using 99mTc-RBCs SPECT/CT imaging, however, requires full compensation of physical degradation factors.

A blind deconvolution method incorporated with anatomical-based filtering for partial volume correction: Validations with 123 I-mIBG cardiac SPECT/CT.

PURPOSE: Segmentation of contrast-enhanced CT and measurement of SPECT point spread function (PSF) are usually required for conventional partial volume correction (PVC). This study was to develop a segmentation-free method with blind deconvolution (BD) and anatomical-based filtering for SPECT PVC. METHODS: The proposed method was implemented using an iterative BD algorithm to estimate the restored image and the PSF simultaneously. An anatomical-based filtering was implemented at each iteration to reduce Gibbs artifact and suppress noise amplification in the deconvolution process. The proposed method was validated with 123 I-metaiodobenzylguanidine (123 I-mIBG) SPECT/CT imaging of NCAT phantoms with and without myocardial perfusion defect and a physical cardiac phantom. Fifteen heart-to-mediastinum ratios (HMRs) were configured in the NCAT and physical phantoms. Correlations between SPECT-quantified and true HMRs were calculated from images without PVC as well as from BD restored images. The proposed method was also performed on a human 123 I-mIBG study. RESULTS: Relative bias and standard deviation images of NCAT phantoms showed that the proposed method reduced both bias and noise. Mean relative bias in the simulated normal myocardium was markedly improved (-16.8% ± 0.4% versus -0.8% ± 0.6% for low noise level; -16.7% ± 0.7% versus -2.3% ± 0.9% for high noise level). Mean relative bias in the simulated myocardial defect was also noticeably improved (-12.7% ± 1.2% versus 1.2% ± 1.6% for low noise level; -13.5% ± 2.4% versus -0.9% ± 2.8% for high noise level). The signal to noise ratio (SNR) of the defect was improved from 2.95 ± 0.09 to 4.07 ± 0.16 for low noise level (38% increase of mean), and from 2.56 ± 0.15 to 3.62 ± 0.22 for high noise level (41% increase of mean). For both NCAT and physical phantoms, HMRs calculated from images without PVC were underestimated (correlations between SPECT-quantified and true HMRs: y = 0.81x + 0.1 for NCAT phantom; y = 0.82x + 0.14 for physical phantom). HMRs from BD restored images were markedly improved (correlations between SPECT-quantified and true HMRs: y = x + 0.05 for NCAT phantom; y = 0.97x - 0.12 for physical phantom). After applying the proposed PVC method, the estimation error between the SPECT-quantified and true HMRs was significantly reduced from -0.75 ± 0.57 to 0.04 ± 0.17 for NCAT phantom (P = 8e-05), and from -0.68 ± 0.67 to -0.26 ± 0.42 for physical phantom (P = 0.005). The human study demonstrated that the HMR increased by 8% with PVC. CONCLUSIONS: The proposed segmentation-free PVC method has the potential of improving SPECT quantification accuracy and reducing noise without the need for premeasuring the image PSF.

Fully automatic multi-atlas segmentation of CTA for partial volume correction in cardiac SPECT/CT.

Anatomical-based partial volume correction (PVC) has been shown to improve image quality and quantitative accuracy in cardiac SPECT/CT. However, this method requires manual segmentation of various organs from contrast-enhanced computed tomography angiography (CTA) data. In order to achieve fully automatic CTA segmentation for clinical translation, we investigated the most common multi-atlas segmentation methods. We also modified the multi-atlas segmentation method by introducing a novel label fusion algorithm for multiple organ segmentation to eliminate overlap and gap voxels. To evaluate our proposed automatic segmentation, eight canine 99mTc-labeled red blood cell SPECT/CT datasets that incorporated PVC were analyzed, using the leave-one-out approach. The Dice similarity coefficient of each organ was computed. Compared to the conventional label fusion method, our proposed label fusion method effectively eliminated gaps and overlaps and improved the CTA segmentation accuracy. The anatomical-based PVC of cardiac SPECT images with automatic multi-atlas segmentation provided consistent image quality and quantitative estimation of intramyocardial blood volume, as compared to those derived using manual segmentation. In conclusion, our proposed automatic multi-atlas segmentation method of CTAs is feasible, practical, and facilitates anatomical-based PVC of cardiac SPECT/CT images.

Generalized PSF modeling for optimized quantitation in PET imaging.

Point-spread function (PSF) modeling offers the ability to account for resolution degrading phenomena within the PET image generation framework. PSF modeling improves resolution and enhances contrast, but at the same time significantly alters image noise properties and induces edge overshoot effect. Thus, studying the effect of PSF modeling on quantitation task performance can be very important. Frameworks explored in the past involved a dichotomy of PSF versus no-PSF modeling. By contrast, the present work focuses on quantitative performance evaluation of standard uptake value (SUV) PET images, while incorporating a wide spectrum of PSF models, including those that under- and over-estimate the true PSF, for the potential of enhanced quantitation of SUVs. The developed framework first analytically models the true PSF, considering a range of resolution degradation phenomena (including photon non-collinearity, inter-crystal penetration and scattering) as present in data acquisitions with modern commercial PET systems. In the context of oncologic liver FDG PET imaging, we generated 200 noisy datasets per image-set (with clinically realistic noise levels) using an XCAT anthropomorphic phantom with liver tumours of varying sizes. These were subsequently reconstructed using the OS-EM algorithm with varying PSF modelled kernels. We focused on quantitation of both SUVmean and SUVmax, including assessment of contrast recovery coefficients, as well as noise-bias characteristics (including both image roughness and coefficient of-variability), for different tumours/iterations/PSF kernels. It was observed that overestimated PSF yielded more accurate contrast recovery for a range of tumours, and typically improved quantitative performance. For a clinically reasonable number of iterations, edge enhancement due to PSF modeling (especially due to over-estimated PSF) was in fact seen to lower SUVmean bias in small tumours. Overall, the results indicate that exactly matched PSF modeling does not offer optimized PET quantitation, and that PSF overestimation may provide enhanced SUV quantitation. Furthermore, generalized PSF modeling may provide a valuable approach for quantitative tasks such as treatment-response assessment and prognostication.

Quantitative myocardial perfusion PET parametric imaging at the voxel-level.

Quantitative myocardial perfusion (MP) PET has the potential to enhance detection of early stages of atherosclerosis or microvascular dysfunction, characterization of flow-limiting effects of coronary artery disease (CAD), and identification of balanced reduction of flow due to multivessel stenosis. We aim to enable quantitative MP-PET at the individual voxel level, which has the potential to allow enhanced visualization and quantification of myocardial blood flow (MBF) and flow reserve (MFR) as computed from uptake parametric images. This framework is especially challenging for the (82)Rb radiotracer. The short half-life enables fast serial imaging and high patient throughput; yet, the acquired dynamic PET images suffer from high noise-levels introducing large variability in uptake parametric images and, therefore, in the estimates of MBF and MFR. Robust estimation requires substantial post-smoothing of noisy data, degrading valuable functional information of physiological and pathological importance. We present a feasible and robust approach to generate parametric images at the voxel-level that substantially reduces noise without significant loss of spatial resolution. The proposed methodology, denoted physiological clustering, makes use of the functional similarity of voxels to penalize deviation of voxel kinetics from physiological partners. The results were validated using extensive simulations (with transmural and non-transmural perfusion defects) and clinical studies. Compared to post-smoothing, physiological clustering depicted enhanced quantitative noise versus bias performance as well as superior recovery of perfusion defects (as quantified by CNR) with minimal increase in bias. Overall, parametric images obtained from the proposed methodology were robust in the presence of high-noise levels as manifested in the voxel time-activity-curves.