Research Interests:
Research Interests:
Deep learning models performed very well in many medical image analysis tasks. However, the majority of these results had been obtained on carefully selected datasets. At the same time, the real clinical flow of Computed Tomography... more
Deep learning models performed very well in many medical image analysis tasks. However, the majority of these results had been obtained on carefully selected datasets. At the same time, the real clinical flow of Computed Tomography studies often contains series with different properties. We address a particular discrepancy related to a much larger scanning interval, e.g., a single series for thorax, abdomen, and pelvis. We propose to use 1D body organ detection for coarse organ localization on thorax-abdomen CT scans. Localized segments, containing volumes of interests, could be further processed by a heavier task-specific network. We convert 3D CT images into multi-channel 2D coronal images, thus drastically decreasing the dimensionality of the data. We next train a conventional U-net like architecture to solve the task of body part regression and build simple threshold rules to localize lungs along the coronal plane. Additionally, this approach allows for the detection of organs only partially presented in the image. Our network was trained on 20 thousand thorax-abdomen volume segments and validated on three separate datasets. It shows high localization accuracy, stability across datasets and processes a high-resolution CT volume in no more than 200 ms.
Research Interests:
We present a method to simultaneously learn several linear discriminative models with explicit information sharing. We use TV-Ll-regularized Logistic Regression in conjunction with a Tikhonov regularization term expressing shared... more
We present a method to simultaneously learn several linear discriminative models with explicit information sharing. We use TV-Ll-regularized Logistic Regression in conjunction with a Tikhonov regularization term expressing shared information across disorders. The weighting of the crossdisorder term is spatially adapted based on the local mutual information between linear models. We apply the model to mesh-based morphometric features from 14 subcortical structures in Parkinson’s and Alzheimer’s disease datasets, PPMI and ADNI. To assess the overall improvement in model performance with cross-disorder information sharing over the baseline TV-LI model, we sample out-of-fold ROC AUC scores using a shuffle-split procedure. Beyond improved prediction, the procedure can be used to formally test for the presence of shared morphometric signatures across diseases in specific regions of interest. Significantly higher ROC AUC scores were found for Parkinson’s prediction accuracy when regularized with the Alzheimer’s model based on putamen and caudate morphometry.
Research Interests:
Research Interests:
Alzheimer's disease (AD) and Parkinson's disease (PD) are the two of the most common neurodegenerative diseases [1]. In previous studies, AD and PD demonstrate overlap in clinical syndromes, including presence of cognitive... more
Alzheimer's disease (AD) and Parkinson's disease (PD) are the two of the most common neurodegenerative diseases [1]. In previous studies, AD and PD demonstrate overlap in clinical syndromes, including presence of cognitive impairment, dementia in PD, and post‐mortem studies with regional atrophy and amyloid and tau deposition [2][3]. Here, we propose a novel morphometric‐feature based machine learning model to explore shared degenerative patterns in AD and PD on neuroimaging.
Research Interests:
We propose a simple yet powerful extension for event-based progression disease model by exploiting the Network Diffusion Hypothesis. Our approach allows incorporating connectivity information derived from diffusion MRI data in the form of... more
We propose a simple yet powerful extension for event-based progression disease model by exploiting the Network Diffusion Hypothesis. Our approach allows incorporating connectivity information derived from diffusion MRI data in the form of an informative prior on event ordering. This simple extension using a definition of transition probability based on network path length leads to improved reproducibility and discriminative power. We report experimental results on a subset of the Alzheimer’s Disease Neuroimaging Initiative data set (ADNI 2). Though trained solely on cross-sectional data, our model successfully assigns higher progression scores to patients converting to more severe stages of dementia.
Research Interests:
We consider a task of classifying normal and pathological brain networks. These networks (called connectomes) represent macroscale connections between predefined brain regions, hence, the nodes of connectomes are uniquely labeled and the... more
We consider a task of classifying normal and pathological brain networks. These networks (called connectomes) represent macroscale connections between predefined brain regions, hence, the nodes of connectomes are uniquely labeled and the set of labels (brain regions) is the same across different brains. We make use of this property and hypothesize that connectomes obtained from normal and pathological brains differ in how brain regions cluster into communities. We develop an algorithm that computes distances between brain networks based on similarity in their partitions and uses these distances to produce a kernel for a support vector machine (SVM) classifier. We demonstrate how the proposed model classifies brain networks of carriers and non-carriers of an allele associated with an increased risk of Alzheimer's disease. The obtained classification quality is ROC AUC 0.7 which is higher than that of the baseline.
Research Interests:
We present a method for metric optimization and template construction in the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework. The construction treats the Riemannian metric on the space of diffeomorphisms as a... more
We present a method for metric optimization and template construction in the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework. The construction treats the Riemannian metric on the space of diffeomorphisms as a data-embedding kernel in the context of predictive modeling, here Kernel Logistic Regression (KLR). The task is then to optimize kernel parameters, including the LDDMM metric parameters as well as the registration template, resulting in a parameterized argminimum optimization. In practice, this leads to a group-wise registration problem with the goal of improving predictive performance, for example by focusing the metric and template on discriminating patient and control populations. We validate our algorithm using two discriminative problems on a synthetic data set as well as 3D subcortical shapes from the SchizConnect cohort. Though secondary to the template and kernel optimization, accuracy of schizophrenia classification is improved by LDDMM-KLR compared to...
Research Interests:
Disease progression models (DPM) of Alzheimer’s disease (AD) based on non‐invasive biomarkers have received significant attention in recent years. Crucially for drug development, DPM promises to identify pre‐clinical stages of AD.... more
Disease progression models (DPM) of Alzheimer’s disease (AD) based on non‐invasive biomarkers have received significant attention in recent years. Crucially for drug development, DPM promises to identify pre‐clinical stages of AD. However, traditional DPMs consider only a single canonical sequence of neurodegeneration, which may ignore important population clusters and individual variation. Addressing this, we propose a method to use individual dMRI‐based connectomes as an informative prior in AD DPM.
Research Interests:
We consider a task of predicting normal and pathological phenotypes from macroscale human brain networks. These networks (connectomes) represent aggregated neural pathways between brain regions. We point to properties of connectomes that... more
We consider a task of predicting normal and pathological phenotypes from macroscale human brain networks. These networks (connectomes) represent aggregated neural pathways between brain regions. We point to properties of connectomes that make them different from graphs arising in other application areas of network science. We discuss how machine learning can be organized on brain networks and focus on kernel classification methods. We describe different kernels on brain networks, including those that use information about similarity in spectral distributions of brain graphs and distances between optimal partitions of connectomes. We compare performance of the reviewed kernels in tasks of classifying autism spectrum disorder versus typical development and carriers versus noncarriers of an allele associated with an increased risk of Alzheimer’s disease.
Research Interests:
We present a method for metric optimization in the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework, by treating the induced Riemannian metric on the space of diffeomorphisms as a kernel in a machine learning context. For... more
We present a method for metric optimization in the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework, by treating the induced Riemannian metric on the space of diffeomorphisms as a kernel in a machine learning context. For simplicity, we choose the kernel Fischer Linear Discriminant Analysis (KLDA) as the framework. Optimizing the kernel parameters in an Expectation-Maximization framework, we define model fidelity via the hinge loss of the decision function. The resulting algorithm optimizes the parameters of the LDDMM norm-inducing differential operator as a solution to a group-wise registration and classification problem. In practice, this may lead to a biology-aware registration, focusing its attention on the predictive task at hand such as identifying the effects of disease. We first tested our algorithm on a synthetic dataset, showing that our parameter selection improves registration quality and classification accuracy. We then tested the algorithm on 3D subcort...
Research Interests:
In this article, we compare the performance of a state-of-the-art segmentation network (UNet) on two different glioblastoma (GB) segmentation datasets. Our experiments show that the same training procedure yields almost twice as bad... more
In this article, we compare the performance of a state-of-the-art segmentation network (UNet) on two different glioblastoma (GB) segmentation datasets. Our experiments show that the same training procedure yields almost twice as bad results on the retrospective clinical data compared to the BraTS challenge data (in terms of Dice score). We discuss possible reasons for such an outcome, including inter-rater variability and high variability in magnetic resonance imaging (MRI) scanners and scanner settings. The high performance of segmentation models, demonstrated on preselected imaging data, does not bring the community closer to using these algorithms in clinical settings. We believe that a clinically applicable deep learning architecture requires a shift from unified datasets to heterogeneous data.
Research Interests:
Human anatomical brain networks derived from the analysis of neuroimaging data are known to demonstrate modular organization. Modules, or communities, of cortical brain regions capture information about the structure of connections in the... more
Human anatomical brain networks derived from the analysis of neuroimaging data are known to demonstrate modular organization. Modules, or communities, of cortical brain regions capture information about the structure of connections in the entire network. Hence, anatomical changes in network connectivity (e.g., caused by a certain disease) should translate into changes in the community structure of brain regions. This means that essential structural differences between phenotypes (e.g., healthy and diseased) should be reflected in how brain networks cluster into communities. To test this hypothesis, we propose a pipeline to classify brain networks based on their underlying community structure. We consider network partitionings into both non-overlapping and overlapping communities and introduce a distance between connectomes based on whether or not they cluster into modules similarly. We next construct a classifier that uses partitioning-based kernels to predict a phenotype from brain networks. We demonstrate the performance of the proposed approach in a task of classifying structural connectomes of healthy subjects and those with mild cognitive impairment and Alzheimer’s disease.
Research Interests:
Research Interests:
BACKGROUND Despite advances in the characterization of genomic alterations in glioblastoma patient survival remains poor. The aim of this single center retrospective study was to estimate prognostic factors affecting to survival for... more
BACKGROUND Despite advances in the characterization of genomic alterations in glioblastoma patient survival remains poor. The aim of this single center retrospective study was to estimate prognostic factors affecting to survival for optimizing personalized treatment strategy in glioblastoma patients. MATERIAL AND METHODS 574 consecutive patients with primary glioblastoma treated at the Burdenko National Medical Research Center of Neurosurgery from 2014 to 2019 were included in this study (314 male /260 female). Survival data was analyzed using Kaplan-Meier with log-rank tests to assess statistical significance. A linear regression model was built to predict overall survival. RESULTS Older patients (> 45 years) had the worse prognosis than patients ≤ 45 (p = 2.37 * 10^-6). Female patients had advantages in overall survival (p = 0.014). Time interval from surgery to the starting of radiotherapy longer than 4 weeks was associated with worse overall survival (p = 0.062). Patients wit...