Abstracts Category : Other

Probabilistic patient grouping and prediction of Multiple Sclerosis disease activity based on a Bag of Lesions brain image representation

by Andrew Doyle

This thesis presents the first application of the Bag of Words probabilistic model to the context of Multiple Sclerosis (MS). A probabilistic framework was adopted to represent lesions visible in patients' Magnetic Resonance Imaging (MRI) scans as codewords in a novel Bag of Lesions representation. In this context, lesions' appearance can vary widely in terms of size, location, shape, intensity and texture, and the Bag of Lesions encodes these characteristics using automatic clustering of an arbitrary number of computer vision features, such as RIFT and LBP. These are probabilistically combined to find types of lesions. This work uses the assumption that patients with similar distributions of lesions in their brains will experience MS similarly to each other, and a method for the automatic prediction of future disease activity is described. Once patients are represented by a distribution of lesions, the types of lesions that are most predictive of future lesion activity are automatically selected using a backwards elimination method. A method is presented to automatically determine groups of patients using the Expectation Maximization (EM) algorithm to find clusters of patients that experience the disease similarly. The model is learned on data from a large multi-centre clinical trial consisting of 1048 patients who were each scanned in four different MRI modalities, and tested using 50-fold cross validation. A Random Forest classifier is trained to automatically predict future MS activity up to two years in the future based on patients' lesions' appearance, and achieves a sensitivity of 94% when only the most certain predictions are counted. This classifier is compared to Support Vector Machines (SVM) using a variety of kernels, Nearest Neighbour methods using several distance metrics, and a simpler Naive Bayesian classifier that takes the number of lesions of different sizes as input, and shown to make better predictions. Finally, a method combining two classifiers, trained on untreated and treated MS patients, is presented to automatically identify responders to specific MS treatments. Ce mmoire prsente la premire application du modle probabiliste de Bag of Words au contexte de la sclrose en plaques (SEP). Un modle probabiliste a t utilis pour reprsenter les lsions visibles dans des imageries par rsonance magntique (IRM) des patients comme des mots-cls dans une nouvelle reprsentation de Bag of Lesions. Dans ce contexte, l'apparence des lsions peut varier grandement en terme de taille, emplacement, forme, intensit et texture. La mthode du Bag of Lesions encode ces caractristiques en utilisant un regroupement automatique d'un nombre arbitraire de traits de vision artificielles, tel que RIFT et LBP. Ces donnes sont combines de manire probabiliste an de trouver diffrents types de lsions. Cette mthode prsume que les patients avec des distributions de lsions similaires dans leurs cerveaux vont vivre des expriences de SEP similaires les uns aux autres et que ceci serait une mthode pourAdvisors/Committee Members: Tal Arbel (Internal/Supervisor).