Deep Learning and Graph Theory for Brain Connectivity Analysis in Multiple Sclerosis

Abstract

Multiple sclerosis (MS) is a chronic disease of the central nervous system, leading cause of nontraumatic disability in young adults. MS is characterized by inflammation, demyelination and neurodegenrative pathological processes which cause a wide range of symptoms, including cognitive deficits and irreversible disability. Concerning the diagnosis of the disease, the introduction of Magnetic Resonance Imaging (MRI) has constituted an important revolution in the last 30 years. Furthermore, advanced MRI techniques, such as brain volumetry, magnetization transfer imaging (MTI) and diffusion-tensor imaging (DTI) are nowadays the main tools for detecting alterations outside visible brain lesions and contributed to our understanding of the pathological mechanisms occurring in normal appearing white matter. In particular, new approaches based on the representation of MR images of the brain as graph have been used to study and quantify damages in the brain white matter network, achieving promising results. In the last decade, novel deep learning based approaches have been used for studying social networks, and recently opened new perspectives in neuroscience for the study of functional and structural brain connectivity. Due to their effectiveness in analyzing large amount of data, detecting latent patterns and establishing functional relationships between input and output, these artificial intelligence techniques have gained particular attention in the scientific community and is nowadays widely applied in many context, including computer vision, speech recognition, medical diagnosis, among others. In this work, deep learning methods were developed to support biomedical image analysis, in particular for the classification and the characterization of MS patients based on structural connectivity information. Graph theory, indeed, constitutes a sensitive tool to analyze the brain networks and can be combined with novel deep learning techniques to detect latent structural properties useful to investigate the progression of the disease. In the first part of this manuscript, an overview of the state of the art will be given. We will focus our analysis on studies showing the interest of DTI for WM characterization in MS. An overview of the main deep learning techniques will be also provided, along with examples of application in the biomedical domain. In a second part, two deep learning approaches will be proposed, for the generation of new, unseen, MRI slices of the human brain and for the automatic detection of the optic disc in retinal fundus images. In the third part, graph-based deep learning techniques will be applied to the study of brain structural connectivity of MS patients. Graph Neural Network methods to classify MS patients in their respective clinical profiles were proposed with particular attention to the model interpretation, the identification of potentially relevant brain substructures, and to the investigation of the importance of local graph-derived metrics for the classification task. Semisupervised and unsupervised approaches were also investigated with the aim of reducing the human intervention in the pipeline.