Brain networks : classification and controllability

The human brain is one of the most complex and fascinating objects we could imagine. Composed of a large number of interconnected neural regions, its structure can be represented as a graph, called a connectome. This mathematical representation of human brain connectivity offers new tools to analyse and understand our brain, and opens the door to perspectives in health and disease. In this work, we begin by explaining how to generate structural connectomes, starting from Magnetic Resonance Imaging data. We build our own dataset composed of 204 structural connectomes and show that their topology significantly differ from that of random networks. Afterwards, motivated by potential applications such as the early diagnosis of neurological disorders, we adopt a machine learning framework to classify connectomes with respect to some label. We propose four classification strategies, along with two state-of-the-art methods, to try to discriminate connectomes based on the gender, the handedness, the alcohol consumption or the smoking habit of the corresponding individuals. Results based on our dataset show that (i) connectomes of men and women are significantly different, (ii) connectomes of smokers and non-smokers are slightly different and (iii) our strategies are often better than the considered state-of-the-art methods. Results related to handedness and alcohol consumption are inconclusive and highlight some limitations of our dataset. Finally, we introduce some notions about the controllability of connectomes, a concept that may impact future treatments of neuropathologies. In particular, we show that cognitive systems exhibit different behaviors in terms of controllability, depending on their function. This result highlights the important relationship between brain structure and function.