Machine learning methods for multiple sclerosis classification and prediction using MRI brain connectivity

In this thesis, the power of Machine Learning (ML) algorithms is combined with brain connectivity patterns, using Magnetic Resonance Imaging (MRI for classification and prediction of Multiple Sclerosis (MS). White Matter (WM) as well as Grey Matter (GM) graphs are studied as connectome data types. The thesis addresses three main research objectives. The first objective aims to generate realistic brain connectomes data for improving the classification of MS clinical profiles in cases of data scarcity and class imbalance. To solve the problem of limited and imbalanced data, a Generative Adversarial Network (GAN) was developed for the generation of realistic and biologically meaningful connec- tomes. This network achieved a 10% better MS classification performance compared to classical approaches. As second research objective, we aim to improve classification of MS clinical profiles us- ing morphological features only extracted from GM brain tissue. First, regional thickness degeneration was investigated. Compared to the current literature, which mostly extracts information from the WM connectome, our approach requires only the T1-weighted im- age modality, which represents the most common modality in clinical practice. Hence, these images are easily and cheaply acquired, thereby justifying the lower scores in classification performance. Furthermore, as an extension, multiple morphological features of the GM connectome were combined using a multiview kernel-based tensor factorization approach. The use of multiple morphological biomarkers describing the degeneration of the GM tissue, induced by the MS disease, enabled a 10% better discrimination between the early MS stages and the progressive MS clinical profiles. The third objective tackles the problem of automatic and interpretable evaluation of MS progression based on cognitive and physical patients? disability status. Multiple machine learning models were combined together to generate ensemble models by following a hi- erarchical structure, known as stacking generalization procedure. However, due to the relevance as well as the impact on the patient?s daily activities, this disability estimation analysis was also performed by adding information from the diffusion signal obtained from tractography. In fact, although very informative, structural brain connectivity does not take into account important fiber tracks such as the CorticoSpinal Track and the Cor- pus Callosum which are known to be heavily involved in the process of tissue degener- ation. Additionally, in order to offer more robust guidance to the neurologist?s decision toward a better diagnosis and treatment of the patient?s status, a new interpretability model was implemented with the aim to pin down the most important connections between pairs of GM regions, deemed to be relevant by the machine learning model for disability estimation. In conclusion, the multiple contributions proposed in this doctoral thesis, represent one of the preliminary attempts to exploit the power of connectome data analysis together with advanced machine learning algorithms for the classification and prediction of MS disability using MRI information from the white and grey matter tissue. These contributions pave the way toward a better understanding of neurodegenerative pathways being formed inside the brain of MS patients.