Topological Network Analysis and Graph-Based Deep Learning of Multimodal MRI: An ENIGMA-Epilepsy Study
(3/16/2021 - 3/15/2022)
Carrie McDonald, UC San Diego
Leonardo Bonilha, MUSC
Brent Munsell, UNC Chapel Hill
Big Data Health Sciences Center (BDHSC) Pilot Project Award (2020), University of South Carolina.
Abstract: Epilepsy is marked by sudden recurrent episodes of sensory disturbance, loss of consciousness, or convulsions, affecting over 50 million people worldwide. Approximately one third of epilepsy patients are resistant to anti-epileptic drug treatment and require additional diagnostic procedures such as electroencephalographic (EEG) evaluation to localize the epileptogenic zone, neuronal network capable of generating seizures, for surgical resection. This approach, however, relies heavily on the expertise of the specialist clinicians reading the EEG. Neuroimaging techniques such as magnetic resonance imaging (MRI) thus play a critical role in the diagnosis of patients with focal epilepsy through identifying visible lesions. Yet, currently around 20 – 45% of focal epilepsy patients do not show lesions on MRI, let alone generalized epilepsy cases that are by default non-lesional. A data-driven approach has shown evidence in improving imaging diagnosis and prognosis of epilepsy, but existing studies tend to be limited in sample sizes and power. In this project, we overcome the limitation by leveraging the first and currently the largest international neuroimaging database on epilepsy provided by the Enhancing Neuroimaging Genetics through Meta-analysis (ENIGMA) Consortium and its Global Alliance for Worldwide Imaging in Epilepsy (ENIGMA-Epilepsy). The global initiative integrates neuroimaging data from over 2,100 epilepsy patients from 24 sites in 14 countries, thus providing unprecedented power of analysis and a unique opportunity to answer complex clinical questions in epilepsy. We will achieve two specific aims in this project using structural and diffusion MRI (sMRI and dMRI) in the ENIGMA-Epilepsy database: 1) detect subtle structural brain abnormalities associated with both focal and generalized epilepsy syndromes with topological network analysis on sMRI and dMRI; 2) predict epilepsy treatment outcomes by building graph-based deep learning algorithms on sMRI and dMRI.
Joint Brain Network Modeling of Multimodal MRI across Aphasia Types
(6/1/2019 - 8/31/2021)
Aspire I Award (2019), University of South Carolina
Abstract: Aphasia is an acquired language disorder that commonly develops after stroke and is estimated to affect approximately one million people in the United States. The long-term adverse effect of the disorder calls for basic and translational research to study its underlying mechanism for better clinical assessment. Multimodal magnetic resonance imaging (MRI) study is an integrative way of understanding post-stroke aphasia impairment through structural and functional brain connectivity captured by diffusion and functional MRI (dMRI and fMRI). Existing multimodal MRI models tend to independently model structural and functional connectivity from dMRI and fMRI and combine results for statistical analysis. A recently developed joint structural-functional brain network modeling approach simultaneously models structural and functional connectivity. But between-subject variability in lesion size and location makes it a challenge to apply existing joint structural-functional network models to multimodal post-stroke aphasia MRI. In this project, we will develop a joint structural-functional brain network model for multimodal MRI constrained by limited connections in lesion sites. Once validated against standard multimodal MRI models, we will use a Bayesian network approach to understand the effect of lesion size and location on aphasia impairment across aphasia types through classification of network features from the standard and joint models.