Research
Predictive Connectomics
Functional connectomics is an emerging tool in computational neuroscience that provides a unique glimpse into the steady-state organization of the brain. It is based on the underlying assumption that two brain regions, which reliably co-activate are more likely to participate in similar neural processes than two uncorrelated or anti-correlated regions. These pairwise relationships are typically assessed via resting-state fMRI (rs-fMRI) and provide the foundation for studying network architectures in the brain across different spatial and temporal scales. Over the past decade, functional connectomics has become ubiquitous in clinical neuroscience, where group-level changes in the functional organization of the brain are treated as biomarkers of a particular neuropsychiatric disorder. However, most studies focus on a simple case/control differentiation, which ignores the heterogeneity of neuropsychiatric disorders.
This project develops novel mathematical frameworks to extract functional subnetworks in the brain that are simultaneously predictive of clinical deficits across multiple cognitive and behavioral domains. Our techniques fuse classical optimization strategies with deep learning to extract both interpretable and generalizable information. We have also developed extensions that use structural connectivity as an implicit regularizer on the functional subnetworks to identify sparser and more stable network patterns. Likewise, we have explored dynamic models that capitalize on changing brain states to better characterize patient-level manifestations.
Our primary application testbed is Autism Spectrum Disorder (ASD), which is a notoriously heterogeneous developmental disorder. We have demonstrated that, unlike current methods, our frameworks can simultaneously predict ADOS, SRS, and Praxis (i.e., gesture and imitation) scores, each of which captures a different facet of ASD.
Moving forward, another application domain is predicting treatment response in acute Spinal Cord Injury (SCI). SCI is often caused by physical trauma and can result in permanent loss of movement and sensation. Physical rehabilitation is known to aid in SCI recovery by retraining neural processes in the brain, a phenomenon known as functional reorganization. We are using our framework to understand these functional reorganization patterns and identify which patients may benefit the most from therapy.
SELECTED PUBLICATIONS
Deep sr-DDL: Deep Structurally Regularized Dynamic Dictionary Learning to Integrate Multimodal and Dynamic Functional Connectivity for Multidimensional Clinical Characterizations.
N.S. D’Souza, M.B. Nebel, D. Crocetti, N. Wymbs, J. Robinson, S. Mostofsky, A. Venkataraman.
NeuroImage, 241:118388, 2021.
M-GCN: A Multimodal Graph Convolutional Network to Integrate Functional and Structural Connectomics Data to Predict Multidimensional Phenotypic Characterizations.
N.S. D’Souza, M.B. Nebel, D. Crocetti, N. Wymbs, J. Robinson, S. Mostofsky, A. Venkataraman.
In Proc. MIDL: Medical Imaging with Deep Learning, MLR:1-12, 2021. Selected for a Long Oral Pres (<15% of Papers)
A Deep-Generative Hybrid Model to Integrate Multimodal and Dynamic Connectivity for Predicting Spectrum-Level Deficits in Autism.
N.S. D’Souza, M.B. Nebel, D. Crocetti, N. Wymbs, J. Robinson, S. Mostofsky, A. Venkataraman.
In Proc. MICCAI: Medical Image Computing and Computer Assisted Intervention, LNCS 12267:437-447, 2020. [Acceptance Rate ≈ 30%]
A Joint Network Optimization Framework to Predict Clinical Severity from Resting State fMRI Data.
N.S. D’Souza, N. Wymbs, M.B. Nebel, S. Mostofsky, A. Venkataraman.
NeuroImage, 206:116314, 2020.
FUNDING