Deep learning techniques for optic nerve evaluation on multiple sclerosis patients using MRI

Abstract

Optic neuritis is a condition in which the optic nerve is lesioned, provoking visual loss, and it is one of the first manifestations of multiple sclerosis. A correct evaluation of the nerve could help in the diagnosis of the disease. The increasing prevalence of multiple sclerosis together with the improvement of deep learning techniques in the healthcare field creates a huge motivation for developing automatic tools for optic nerve assessment. Therefore, this thesis aims to develop two convolutional networks: a first 3D architecture for classifying axial T2-weighted fat-saturated MRI images between lesioned or non lesioned nerves and a second 2D one that automatically segments the optic nerves for longitudinal studies. Two datasets have been used for validating the system (N=107 and 62) and the labels and the ground truth masks have been manually created by analyzing all the scans and their clinical records. Two additional machine learning classifiers have been also implemented to do proper comparisons with our approach. Saliency maps have been applied for a better interpretation of the network’s performance. Both models showed good performance: the classification algorithm has an accuracy of 69.68%, which generalizes to unseen data (68.21%), and overcomes the performances of the simple classifiers. The segmentation algorithm has a dice score of 85.61% on unseen data, where only a few pixels differed from the ground truth masks. The saliency maps showed that it is possible to demonstrate that the first network focuses its attention on the areas that correspond to lesions in the optic nerve. The two methods showed satisfactory results and robustness in their tasks when performing with unseen data. Given its speed and performance, the developed methodologies could be translated into a clinical setting to assist radiologists in the scanner assessment by helping with the nerve identification and evaluation.