Assessing the capability of deep-learning models in Parkinson’s disease diagnosis

Abstract

Parkinson’s Disease is one of the leading age-related neurological disorders affecting the general population. Current diagnostic techniques rely on patient symptoms rather than biomarkers. Symptomatic diagnoses are subjective and can vary highly. Our work aims to remedy this by presenting a novel approach to Parkinson’s Disease diagnosis. We propose and assess four deep-learning based models that classify patients based on biomarkers found in structural magnetic resonance images, and find that our 3D-Convolution-Neural-Network model demonstrates high efficacy in the task of diagnosing Parkinson’s disease, with an accuracy of 75% and 76% sensitivity. As well, our work highlights potential biomarkers for the disease found in the cerebellum and occipital lobe.