Automatic deep learning segmentation of the hippocampus on high-resolution diffusion magnetic resonance imaging and its application to the healthy lifespan.

Diffusion tensor imaging (DTI) can provide unique contrast and insight into microstructural changes with age or disease of the hippocampus, although it is difficult to measure the hippocampus because of its comparatively small size, location, and shape. This has been markedly improved by the advent of a clinically feasible 1-mm isotropic resolution 6-min DTI protocol at 3 T of the hippocampus with limited brain coverage of 20 axial-oblique slices aligned along its long axis. However, manual segmentation is too laborious for large population studies, and it cannot be automatically segmented directly on the diffusion images using traditional T1 or T2 image-based methods because of the limited brain coverage and different contrast. An automatic method is proposed here that segments the hippocampus directly on high-resolution diffusion images based on an extension of well-known deep learning architectures like UNet and UNet++ by including additional dense residual connections. The method was trained on 100 healthy participants with previously performed manual segmentation on the 1-mm DTI, then evaluated on typical healthy participants (n = 53), yielding an excellent voxel overlap with a Dice score of ~ 0.90 with manual segmentation; notably, this was comparable with the inter-rater reliability of manually delineating the hippocampus on diffusion magnetic resonance imaging (MRI) (Dice score of 0.86). This method also generalized to a different DTI protocol with 36% fewer acquisitions. It was further validated by showing similar age trajectories of volumes, fractional anisotropy, and mean diffusivity from manual segmentations in one cohort (n = 153, age 5-74 years) with automatic segmentations from a second cohort without manual segmentations (n = 354, age 5-90 years). Automated high-resolution diffusion MRI segmentation of the hippocampus will facilitate large cohort analyses and, in future research, needs to be evaluated on patient groups.

Developmental Hip Dysplasia Diagnosis at Three-dimensional US: A Multicenter Study.

Purpose To validate accuracy of diagnosis of developmental dysplasia of the hip (DDH) from geometric properties of acetabular shape extracted from three-dimensional (3D) ultrasonography (US). Materials and Methods In this retrospective multi-institutional study, 3D US was added to conventional two-dimensional (2D) US of 1728 infants (mean age, 67 days; age range, 3-238 days) evaluated for DDH from January 2013 to December 2016. Clinical diagnosis after more than 6 months follow-up was normal (n = 1347), borderline (Graf IIa, later normalizing spontaneously; n = 140) or dysplastic (Graf IIb or higher, n = 241). Custom software accessible through the institution's research portal automatically calculated indexes including 3D posterior and anterior alpha angle and osculating circle radius from hip surface models generated with less than 1 minute of user input. Logistic regression predicted clinical diagnosis (normal = 0, dysplastic = 1) from 3D indexes (ie, age and sex). Output represented probability of hip dysplasia from 0 to 1 (output: >0.9, dysplastic; 0.11-0.89, borderline; <0.1, normal). Software can be accessed through the research portal. Results Area under the receiver operating characteristic curve was equivalently high for 3D US indexes and 2D US alpha angle (0.996 vs 0.987). Three-dimensional US helped to correctly categorize 97.5% (235 of 241) dysplastic and 99.4% (1339 of 1347) normal hips. No dysplastic hips were categorized as normal. Correct diagnosis was provided at initial 3D US scan in 69.3% (97 of 140) of the studies diagnosed as borderline at initial 2D US scans. Conclusion Automatically calculated 3D indexes of acetabular shape performed equivalently to high-quality 2D US scans at tertiary medical centers to help diagnose DDH. Three-dimensional US reduced the number of borderline studies requiring follow-up imaging by over two-thirds.

Discriminative analysis of regional evolution of iron and myelin/calcium in deep gray matter of multiple sclerosis and healthy subjects.

BACKGROUND: Combined R2* and quantitative susceptibility (QS) has been previously used in cross-sectional multiple sclerosis (MS) studies to distinguish deep gray matter (DGM) iron accumulation and demyelination. PURPOSE: We propose and apply discriminative analysis of regional evolution (DARE) to define specific changes in MS and healthy DGM. STUDY TYPE: Longitudinal (baseline and 2-year follow-up) retrospective study. SUBJECTS: Twenty-seven relapsing-remitting MS (RRMS), 17 progressive MS (PMS), and corresponding age-matched healthy subjects. FIELD STRENGTH/SEQUENCE: 4.7T 10-echo gradient-echo acquisition. ASSESSMENT: Automatically segmented caudate nucleus (CN), thalamus (TH), putamen (PU), globus pallidus, red nucleus (RN), substantia nigra, and dentate nucleus were retrospectively analyzed to quantify regional volumes, bulk mean R2*, and bulk mean QS. DARE utilized combined R2* and QS localized changes to compute spatial extent, mean intensity, and total changes of DGM iron and myelin/calcium over 2 years. STATISTICAL TESTS: We used mixed factorial analysis for bulk analysis, nonparametric tests for DARE (α = 0.05), and multiple regression analysis using backward elimination of DGM structures (α = 0.05, P = 0.1) to regress bulk and DARE measures with the follow-up Multiple Sclerosis Severity Score (MSSS). False detection rate correction was applied to all tests. RESULTS: Bulk analysis only detected significant (Q ≤ 0.05) interaction effects in RRMS CN QS (η = 0.45; Q = 0.004) and PU volume (η = 0.38; Q = 0.034). DARE demonstrated significant group differences in all RRMS structures, and in all PMS structures except the RN. The largest RRMS effect size was CN total R2* iron decrease (r = 0.74; Q = 0.00002), and TH mean QS myelin/calcium decrease for PMS (r = 0.70; Q = 0.002). DARE iron increase using total QS demonstrated the highest correlation with MSSS (r = 0.68; Q = 0.0005). DATA CONCLUSION: DARE enabled discriminative assessment of specific DGM changes over 2 years, where iron and myelin/calcium changes were the primary drivers in RRMS and PMS compared to age-matched controls, respectively. Specific DARE measures of MS DGM correlated with follow-up MSSS, and may reflect complex disease pathology. LEVEL OF EVIDENCE: 3 Technical Efficacy: Stage 1 J. Magn. Reson. Imaging 2018.

Progressive iron accumulation across multiple sclerosis phenotypes revealed by sparse classification of deep gray matter.

PURPOSE: To create an automated framework for localized analysis of deep gray matter (DGM) iron accumulation and demyelination using sparse classification by combining quantitative susceptibility (QS) and transverse relaxation rate (R2*) maps, for evaluation of DGM in multiple sclerosis (MS) phenotypes relative to healthy controls. MATERIALS AND METHODS: R2*/QS maps were computed using a 4.7T 10-echo gradient echo acquisition from 16 clinically isolated syndrome (CIS), 41 relapsing-remitting (RR), 40 secondary-progressive (SP), 13 primary-progressive (PP) MS patients, and 75 controls. Sparse classification for R2*/QS maps of segmented caudate nucleus (CN), putamen (PU), thalamus (TH), and globus pallidus (GP) structures produced localized maps of iron/myelin in MS patients relative to controls. Paired t-tests, with age as a covariate, were used to test for statistical significance (P ≤ 0.05). RESULTS: In addition to DGM structures found significantly different in patients compared to controls using whole region analysis, singular sparse analysis found significant results in RRMS PU R2* (P = 0.03), TH R2* (P = 0.04), CN QS (P = 0.04); in SPMS CN R2* (P = 0.04), GP R2* (P = 0.05); and in PPMS CN R2* (P = 0.04), TH QS (P = 0.04). All sparse regions were found to conform to an iron accumulation pattern of changes in R2*/QS, while none conformed to demyelination. Intersection of sparse R2*/QS regions also resulted in RRMS CN R2* becoming significant, while RRMS R2* TH and PPMS QS TH becoming insignificant. Common iron-associated volumes in MS patients and their effect size progressively increased with advanced phenotypes. CONCLUSION: A localized technique for identifying sparse regions indicative of iron or myelin in the DGM was developed. Progressive iron accumulation with advanced MS phenotypes was demonstrated, as indicated by iron-associated sparsity and effect size. LEVEL OF EVIDENCE: 1 Technical Efficacy: Stage 1 J. Magn. Reson. Imaging 2017;46:1464-1473.

Subcortical gray matter segmentation and voxel-based analysis using transverse relaxation and quantitative susceptibility mapping with application to multiple sclerosis.

PURPOSE: To investigate subcortical gray matter segmentation using transverse relaxation rate (R2 *) and quantitative susceptibility mapping (QSM) and apply it to voxel-based analysis in multiple sclerosis (MS). MATERIALS AND METHODS: Voxel-based variation in R2 * and QSM within deep gray matter was examined and compared to standard whole-structure analysis using 37 MS subjects and 37 matched controls. Deep gray matter nuclei (caudate, putamen, globus pallidus, and thalamus) were automatically segmented and morphed onto a custom atlas based on QSM and standard T1 -weighted images. Segmentation accuracy and scan-rescan reliability were tested. RESULTS: When considering only significant regions as returned by the multivariate voxel-based analysis, increased R2 * and QSM was found in MS subjects compared to controls in portions of all four nuclei studied (P < 0.002). For R2 *, regional analysis yielded at least 66-fold improved P-value significance in all nuclei over standard whole-structure analysis, while for QSM only thalamus benefited, with 5-fold improvement in significance. Improved segmentation over standard methods, particularly for globus pallidus (2.8 times higher Dice score), was achieved by incorporating high-contrast QSM into the atlas. Voxel-based reliability was highest for QSM (<1% variation). CONCLUSION: Automatic segmentation of iron-rich deep gray matter can be improved by incorporating QSM. Voxel-based evaluation yielded increased R2 * and QSM in MS subjects in all four nuclei studied with R2 *, benefiting the most from localized analysis over whole-structure measures.

Low-Rank Plus Sparse Decomposition of fMRI Data With Application to Alzheimer's Disease.

Studying functional brain connectivity plays an important role in understanding how human brain functions and neuropsychological diseases such as autism, attention-deficit hyperactivity disorder, and Alzheimer's disease (AD). Functional magnetic resonance imaging (fMRI) is one of the most popularly used tool to construct functional brain connectivity. However, the presence of noises and outliers in fMRI blood oxygen level dependent (BOLD) signals might lead to unreliable and unstable results in the construction of connectivity matrix. In this paper, we propose a pipeline that enables us to estimate robust and stable connectivity matrix, which increases the detectability of group differences. In particular, a low-rank plus sparse (L + S) matrix decomposition technique is adopted to decompose the original signals, where the low-rank matrix L recovers the essential common features from regions of interest, and the sparse matrix S catches the sparse individual variability and potential outliers. On the basis of decomposed signals, we construct connectivity matrix using the proposed novel concentration inequality-based sparse estimator. In order to facilitate the comparisons, we also consider correlation, partial correlation, and graphical Lasso-based methods. Hypothesis testing is then conducted to detect group differences. The proposed pipeline is applied to rs-fMRI data in Alzheimer's disease neuroimaging initiative to detect AD-related biomarkers, and we show that the proposed pipeline provides accurate yet more stable results than using the original BOLD signals.

Stable Anatomy Detection in Multimodal Imaging Through Sparse Group Regularization: A Comparative Study of Iron Accumulation in the Aging Brain.

Multimodal neuroimaging provides a rich source of data for identifying brain regions associated with disease progression and aging. However, present studies still typically analyze modalities separately or aggregate voxel-wise measurements and analyses to the structural level, thus reducing statistical power. As a central example, previous works have used two quantitative MRI parameters-R2* and quantitative susceptibility (QS)-to study changes in iron associated with aging in healthy and multiple sclerosis subjects, but failed to simultaneously account for both. In this article, we propose a unified framework that combines information from multiple imaging modalities and regularizes estimates for increased interpretability, generalizability, and stability. Our work focuses on joint region detection problems where overlap between effect supports across modalities is encouraged but not strictly enforced. To achieve this, we combine L 1 (lasso), total variation (TV), and L 2 group lasso penalties. While the TV penalty encourages geometric regularization by controlling estimate variability and support boundary geometry, the group lasso penalty accounts for similarities in the support between imaging modalities. We address the computational difficulty in this regularization scheme with an alternating direction method of multipliers (ADMM) optimizer. In a neuroimaging application, we compare our method against independent sparse and joint sparse models using a dataset of R2* and QS maps derived from MRI scans of 113 healthy controls: our method produces clinically-interpretable regions where specific iron changes are associated with healthy aging. Together with results across multiple simulation studies, we conclude that our approach identifies regions that are more strongly associated with the variable of interest (e.g., age), more accurate, and more stable with respect to training data variability. This work makes progress toward a stable and interpretable multimodal imaging analysis framework for studying disease-related changes in brain structure and can be extended for classification and disease prediction tasks.

Evaluation of automated computed tomography segmentation to assess body composition and mortality associations in cancer patients.

BACKGROUND: Body composition from computed tomography (CT) scans is associated with cancer outcomes including surgical complications, chemotoxicity, and survival. Most studies manually segment CT scans, but Automatic Body composition Analyser using Computed tomography image Segmentation (ABACS) software automatically segments muscle and adipose tissues to speed analysis. Here, we externally evaluate ABACS in an independent dataset. METHODS: Among patients with non-metastatic colorectal (n = 3102) and breast (n = 2888) cancer diagnosed from 2005 to 2013 at Kaiser Permanente, expert raters annotated tissue areas at the third lumbar vertebra (L3). To compare ABACS segmentation results to manual analysis, we quantified the proportion of pixel-level image overlap using Jaccard scores and agreement between methods using intra-class correlation coefficients for continuous tissue areas. We examined performance overall and among subgroups defined by patient and imaging characteristics. To compare the strength of the mortality associations obtained from ABACS's segmentations to manual analysis, we computed Cox proportional hazards ratios (HRs) and 95% confidence intervals (95% CI) by tertile of tissue area. RESULTS: Mean ± SD age was 63 ± 11 years for colorectal cancer patients and 56 ± 12 for breast cancer patients. There was strong agreement between manual and automatic segmentations overall and within subgroups of age, sex, body mass index, and cancer stage: average Jaccard scores and intra-class correlation coefficients exceeded 90% for all tissues. ABACS underestimated muscle and visceral and subcutaneous adipose tissue areas by 1-2% versus manual analysis: mean differences were small at -2.35, -1.97 and -2.38 cm2 , respectively. ABACS's performance was lowest for the <2% of patients who were underweight or had anatomic abnormalities. ABACS and manual analysis produced similar associations with mortality; comparing the lowest to highest tertile of skeletal muscle from ABACS versus manual analysis, the HRs were 1.23 (95% CI: 1.00-1.52) versus 1.38 (95% CI: 1.11-1.70) for colorectal cancer patients and 1.30 (95% CI: 1.01-1.66) versus 1.29 (95% CI: 1.00-1.65) for breast cancer patients. CONCLUSIONS: In the first study to externally evaluate a commercially available software to assess body composition, automated segmentation of muscle and adipose tissues using ABACS was similar to manual analysis and associated with mortality after non-metastatic cancer. Automated methods will accelerate body composition research and, eventually, facilitate integration of body composition measures into clinical care.

Five year iron changes in relapsing-remitting multiple sclerosis deep gray matter compared to healthy controls.

BACKGROUND: Relapsing-Remitting MS (RRMS) Deep Grey Matter (DGM) 5 year changes were examined using MRI measures of volume, transverse relaxation rate (R2*) and quantitative magnetic susceptibility (QS). By applying Discriminative Analysis of Regional Evolution (DARE), R2* and QS changes from iron and non-iron sources were separated. METHODS: 25 RRMS and 25 age-matched control subjects were studied at baseline and 5-year follow-up. Bulk DGM mean R2* and QS of the caudate nucleus, putamen, thalamus and globus pallidus were analyzed using mixed factorial analysis (α = 0.05) with sex as a covariate, while DARE employed non-parametric analysis to study regional changes. Regression/correlation analysis was performed with disease duration and MS Severity Score (MSSS). RESULTS: No significant change in Extended Disability Status Score was found over 5 years (baseline = 2.4 ±â€¯1.2; follow-up = 2.8 ±â€¯1.3). Significant time effects were found for R2* in the caudate (Q = 0.000008; η2 = 0.36), putamen (Q = 0.0000007; η2 = 0.43), and globus pallidus (Q = 0.0000007; η2 = 0.43), while significant longitudinal effects were only found for QS in the putamen (Q = 0.002; η2 = 0.22). Significant bulk interaction was only found for thalamus volume (Q = 0.02; η2 = 0.20). Iron decrease was the only detected significant effect using DARE, and the highest significant DARE effect size was mean thalamus R2* iron decrease (Q = 0.002; η2 = 0.26). No significant correlations or regressions were demonstrated with clinical measures. CONCLUSIONS: Thalamic atrophy was the only bulk effect that demonstrated different rates of changes over 5 years compared to age-matched controls. DARE Iron decrease in regions of the caudate, putamen, and thalamus were prominent features in stable RRMS over 5 years.

Significant Anatomy Detection Through Sparse Classification: A Comparative Study.

We present a comparative study for discriminative anatomy detection in high dimensional neuroimaging data. While most studies solve this problem using mass univariate approaches, recent works show better accuracy and variable selection using a sparse classification model. Two types of image-based regularization methods have been proposed in the literature based on either a Graph Net (GN) model or a total variation (TV) model. These studies showed increased classification accuracy and interpretability of results when using image-based regularization, but did not look at the accuracy and quality of the recovered significant regions. In this paper, we theoretically prove bounds on the recovered sparse coefficients and the corresponding selected image regions in four models (two based on GN penalty and two based on TV penalty). Practically, we confirm the theoretical findings by measuring the accuracy of selected regions compared with ground truth on simulated data. We also evaluate the stability of recovered regions over cross-validation folds using real MRI data. Our findings show that the TV penalty is superior to the GN model. In addition, we showed that adding an l2 penalty improves the accuracy of estimated coefficients and selected significant regions for the both types of models.

Body Composition Assessment in Axial CT Images Using FEM-Based Automatic Segmentation of Skeletal Muscle.

The proportions of muscle and fat tissues in the human body, referred to as body composition is a vital measurement for cancer patients. Body composition has been recently linked to patient survival and the onset/recurrence of several types of cancers in numerous cancer research studies. This paper introduces a fully automatic framework for the segmentation of muscle and fat tissues from CT images to estimate body composition. We developed a novel finite element method (FEM) deformable model that incorporates a priori shape information via a statistical deformation model (SDM) within the template-based segmentation framework. The proposed method was validated on 1000 abdominal and 530 thoracic CT images and we obtained very good segmentation results with Jaccard scores in excess of 90% for both the muscle and fat regions.

Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the VESSEL12 study.

The VESSEL12 (VESsel SEgmentation in the Lung) challenge objectively compares the performance of different algorithms to identify vessels in thoracic computed tomography (CT) scans. Vessel segmentation is fundamental in computer aided processing of data generated by 3D imaging modalities. As manual vessel segmentation is prohibitively time consuming, any real world application requires some form of automation. Several approaches exist for automated vessel segmentation, but judging their relative merits is difficult due to a lack of standardized evaluation. We present an annotated reference dataset containing 20 CT scans and propose nine categories to perform a comprehensive evaluation of vessel segmentation algorithms from both academia and industry. Twenty algorithms participated in the VESSEL12 challenge, held at International Symposium on Biomedical Imaging (ISBI) 2012. All results have been published at the VESSEL12 website http://vessel12.grand-challenge.org. The challenge remains ongoing and open to new participants. Our three contributions are: (1) an annotated reference dataset available online for evaluation of new algorithms; (2) a quantitative scoring system for objective comparison of algorithms; and (3) performance analysis of the strengths and weaknesses of the various vessel segmentation methods in the presence of various lung diseases.

A variational formulation for discrete registration.

We present a novel variational formulation of discrete deformable registration as the minimization of a convex energy functional that involves diffusion regularization. We show that a finite difference solution (FD) of the variational formulation is equivalent to a continuous-valued Gaussian Markov random field (MRF) energy minimization formulation previously proposed as the random walker deformable registration method. A computationally efficient solution using the finite element method (FEM) method has been proposed to solve the variational minimization problem. Our proposed method obtained competitive results when compared with 14 other deformable registration methods on the CUMC12 MRI dataset.

Tumor invasion margin on the Riemannian space of brain fibers.

Glioma is one of the most challenging types of brain tumors to treat or control locally. One of the main problems is to determine which areas of the apparently normal brain contain glioma cells, as gliomas are known to infiltrate several centimeters beyond the clinically apparent lesion that is visualized on standard Computed Tomography scans (CT) or Magnetic Resonance Images (MRIs). To ensure that radiation treatment encompasses the whole tumor, including the cancerous cells not revealed by MRI, doctors treat the volume of brain that extends 2cm out from the margin of the visible tumor. This approach does not consider varying tumor-growth dynamics in different brain tissues, thus it may result in killing some healthy cells while leaving cancerous cells alive in the other areas. These cells may cause recurrence of the tumor later in time, which limits the effectiveness of the therapy. Knowing that glioma cells preferentially spread along nerve fibers, we propose the use of a geodesic distance on the Riemannian manifold of brain diffusion tensors to replace the Euclidean distance used in the clinical practice and to correctly identify the tumor invasion margin. This mathematical model results in a first-order Partial Differential Equation (PDE) that can be numerically solved in a stable and consistent way. To compute the geodesic distance, we use actual Diffusion Weighted Imaging (DWI) data from 11 patients with glioma and compare our predicted infiltration distance map with actual grwoth in follow-up MRI scans. Results show improvement in predicting the invasion margin when using the geodesic distance as opposed to the 2cm conventional Euclidean distance.

3D variational brain tumor segmentation using Dirichlet priors on a clustered feature set.

PURPOSE: Brain tumor segmentation is a required step before any radiation treatment or surgery. When performed manually, segmentation is time consuming and prone to human errors. Therefore, there have been significant efforts to automate the process. But, automatic tumor segmentation from MRI data is a particularly challenging task. Tumors have a large diversity in shape and appearance with intensities overlapping the normal brain tissues. In addition, an expanding tumor can also deflect and deform nearby tissue. In our work, we propose an automatic brain tumor segmentation method that addresses these last two difficult problems. METHODS: We use the available MRI modalities (T1, T1c, T2) and their texture characteristics to construct a multidimensional feature set. Then, we extract clusters which provide a compact representation of the essential information in these features. The main idea in this work is to incorporate these clustered features into the 3D variational segmentation framework. In contrast to previous variational approaches, we propose a segmentation method that evolves the contour in a supervised fashion. The segmentation boundary is driven by the learned region statistics in the cluster space. We incorporate prior knowledge about the normal brain tissue appearance during the estimation of these region statistics. In particular, we use a Dirichlet prior that discourages the clusters from the normal brain region to be in the tumor region. This leads to a better disambiguation of the tumor from brain tissue. RESULTS: We evaluated the performance of our automatic segmentation method on 15 real MRI scans of brain tumor patients, with tumors that are inhomogeneous in appearance, small in size and in proximity to the major structures in the brain. Validation with the expert segmentation labels yielded encouraging results: Jaccard (58%), Precision (81%), Recall (67%), Hausdorff distance (24 mm). CONCLUSIONS: Using priors on the brain/tumor appearance, our proposed automatic 3D variational segmentation method was able to better disambiguate the tumor from the surrounding tissue.

Random walks for deformable image registration.

We introduce a novel discrete optimization method for non-rigid image registration based on the random walker algorithm. We discretize the space of deformations and formulate registration using a Gaussian MRF where continuous labels correspond to the probability of a point having a certain discrete deformation. The interaction (regularization) term of the corresponding MRF energy is convex and image dependent, thus being able to accommodate different types of tissue elasticity. This formulation results in a fast algorithm that can easily accommodate a large number of displacement labels, has provable robustness to noise and a close to global solution. We experimentally demonstrate the validity of our formulation on synthetic and real medical data.

Tumor invasion margin on the Riemannian space of brain fibers.

Gliomas are one of the most challenging tumors to treat or control locally. One of the main challenges is determining which areas of the apparently normal brain contain glioma cells, as gliomas are known to infiltrate for several centimeters beyond the clinically apparent lesion visualized on standard CT or MRI. To ensure that radiation treatment encompasses the whole tumour, including the cancerous cells not revealed by MRI, doctors treat a volume of brain extending 2cm out from the margin of the visible tumour. This expanded volume often includes healthy, non-cancerous brain tissue. Knowing that glioma cells preferentially spread along nerve fibers, we propose the use of a geodesic distance on the Riemannian manifold of brain fibers to replace the Euclidean distance used in clinical practice and to correctly identify the tumor invasion margin. To compute the geodesic distance we use actual DTI data from patients with glioma and compare our predicted growth with follow-up MRI scans. Results show improvement in predicting the invasion margin when using the geodesic distance as opposed to the 2cm conventional Euclidean distance.