Multivariate Analysis of MR Images in Temporal Lobe Epilepsy

Epilepsy stands aside from other neurological diseases because clinical patterns of progression are unknown: The etiology of each epilepsy case is unique and so it is the individual prognosis. Temporal lobe epilepsy (TLE) is the most frequent type of focal epilepsy and the surgical excision of the hippocampus and the surrounding tissue is an accepted treatment in refractory cases, specially when seizures become frequent increasingly affecting the performance of daily tasks and significantly decreasing the quality of life of the patient. The sensitivity of clinical imaging is poor for patients with no hippocampal involvement and invasive procedures such as the Wada test and intracranial EEG are required to detect and lateralize epileptogenic tissue. This thesis develops imaging processing techniques using quantitative relaxometry and diffusion tensor imaging with the aiming to provide a less invasive alternative when detectability is low. Chapter 2 develops the concept of individual feature maps on regions of interest. A laterality score on these maps correctly distinguished left TLE from right TLE in 12 out of 15 patients. Chapter 3 explores machine learning models to detect TLE, obtaining perfect classification for left patients, and 88.9% accuracy for right TLE patients. Chapter 4 focuses on temporal lobe asymmetry developing a voxel-based method for assessing asymmetry and verifying its applicability to individual predictions (92% accuracy) and group-wise statistical analyses. Informative ROI and voxel-based informative features are described for each experiment, demonstrating the relative importance of mean diffusivity over other MR imaging alternatives in identification and lateralization of TLE patients. Finally, the conclusion chapter discuss contributions, main limitations and outlining options for future research