Ph.D. Thesis Defense | Decrypting Cryptogenic Epilepsy: Machine Learning Methods for Detecting Cortical Malformations

May 2, 2016
1:15 pm - 2:15 pm
196 Boston Avenue, 4th Floor
Speaker: Bilal Ahmed, Tufts University
Host: Carla E. Brodley


Epilepsy is a common neurological disorder, affecting approximately 1% of the world's population. Uncontrolled epilepsy can have harmful effects on the brain and increases the risk of injuries and sudden death. Cortical malformations, particularly focal cortical dysplasia (FCD) is recognized as one of the most common source of treatment resistant epilepsy (TRE). Surgical resection of the abnormal tissue is the only treatment for TRE patients, and a successful outcome results in complete seizure freedom. Chances of success when the lesion is visually detected on the MRI (MRI-positive) are 66\%, and only 29% for cases with undetected lesions (MRI-negative). Approximately 45%-60% of histologically confirmed FCD lesions are missed by expert neuroradiologists. This dissertation develops automated methods of detecting cortical malformations in MRI-negative patients using surface-based morphometry. Using data from MRI-negative patients to train machine learning (ML) algorithms has a number of confounding factors that limit their applicability to the lesion detection task. These include, label noise arising from subjectivity in determining the cortical region to resect without a visible abnormality. Similarly, inter-subject and intra-subject variations in brain morphology limit the generalization of ML methods trained on data aggregated from different individuals. To address these issues we develop two novel ML methods. We propose a multitask learning (MTL) method that models each patient as a separate learning task, and uses the results of intra-cranial EEG exam as added supervision to mitigate label noise. Next, we develop hierarchical conditional random fields (HCRF) for outlier-detection, which is a semi-supervised learning method that does not require labeled training data. By correcting for all three factors (i.e., label noise, intra-subject and inter-subject variation) HCRF outperforms the baseline methods and the MTL method. The high detection rate (75% for HCRF) of the proposed methods for MRI-negative patients shows that some electrophysiologically and histologically abnormal cortical regions are not visually apparent to the human eye but can be detected using ML methods. Incorporating such ML methods in the pre-surgical evaluation protocol have the potential to enhance the chances of detecting the lesion prior to surgery, leading to an increased number of patients being referred to resective surgery.