UCLA

Epilepsy - a neurological disorder characterized by recurring seizures - affects lives of more than 3.4 million Americans nationwide. Typical treatment procedure for patients resistant to anti-seizure medication involves invasive surgery to correctly characterize abnormalities in the epileptic network and localize epileptogenic zones by intracranial electrodes. Intracranial Electroencephalogram (iEEG) measured by this method provides a comprehensive way to monitor propagation of seizures and test hypotheses regarding the epileptogenic zones. However, the electrode implantation procedure poses unavoidable risks to the patients. Functional Magnetic Resonance Imaging (fMRI), on the other hand, is a non-invasive method providing another perspective on the epileptic network but does not have defining features for epilepsy analysis. It is of particular interest to find a link bridging the two modalities on the quest to have a comprehensive view of the network in epileptic brain. In this thesis we present a data-driven approach to find the mapping from fMRI-derived epileptic network to iEEG-derived epileptic network. We propose U-BrainNet, a deep learning model with special architectural considerations for cross-modality learning employing convolution operations specifically designed for connectomic data. We evaluate the model together with three other baselines on a population of 43 patients having intractable epilepsy, and provide insights into their performance as well as their feasibility to become clinically applicable with future modifications.