Dottorato in PLASTICITÀ NEUROSENSORIALE

Titolo della tesi: CLUSTERING AND SEGMENTATION OF MULTIPLE SCLEROSIS LESIONS: A DEEP LEARNING EXPLORATION

Introduction: Multiple Sclerosis (MS) is an autoimmune and inflammatory disease characterised by typical demyelinating ‘lesions’ of brain and spinal cord. Magnetic Resonance Imaging (MRI) is the key tool to diagnose and classify MS in phenotypes. The widespread diffusion of MRI for clinical and research purposes, is leading to a steady increase of MS imaging data. Artificial Intelligence (AI) can process this amount of information in a controlled and faster manner than humans. Moreover, the Deep Learning (DL) approach takes advantage from unsupervised representations to recognized pattern using multimodal information. Starting from a visionary idea to overcome unsolved issues in MS, as aetiology and pathology of the disease, we tried to develop an exploratory and unsupervised DL approach to evaluate MRI lesions distribution. Objectives: I) to cluster MS patients applying an unsupervised approach on lesions distribution maps, correlating with clinical and radiological features; II) to evaluate the DL approach’s performance in lesion segmentation of in-house MRI images, compared to manually segmented maps. Method: we decided to use a DL approach called U-Net for our purpose. U-net is developed to segment the lesions. Moreover, we extracted a low dimensional embedding of imaging characteristics, resulting in a vector. We applied a principal components analysis (PCA) to select the directions of maximum variance in the data. Ultimately, a classic K-means algorithm was used for clustering. We chosed to explore K={2,3,4,5}. Preliminary results: 299 patients were evaluated for an amount of 707 exams sessions. No significant correlations between clustering and activity, disability and MS phenotypes characteristics were found. Significant associations were found with age, k=2 (OR 0.932, 95%CI [0.894, 0.970]) and k=4 (OR 1.059, 95%CI[1.009, 1.117]), sex, k=3(OR 28.875, 95%CI [7.935, 148.531]) and k=4 (OR 21.625, 95%CI[6.065, 108.357]) and disease duration, K=5 (OR 1.119, 95%CI[1.045,1.216]). Compared to manually, U-Net confirmed its high capacity in lesions segmentation, with accuracy (99.98±0.02), sensitivity (89.14±6.11) and specificity (99.99±0.01). Conclusions: Our study demonstrate that U-Net and k-means are useful for lesion segmentations and clustering of patients’ lesions map using an unsupervised approach. However, this study should be considered as an initial step of a figured wider project, since it reported very preliminary results. Our initial idea was to use an exploratory DL approach to solve the complex and unreached issues in MS, e.g. the etiology of the disease. We are aware that this goal is very far from being achieved and that many research efforts, advanced imaging techniques and biological information are still needed to do that.