Diagnóstico asistido de la enfermedad de Parkinson mediante aprendizaje profundo a partir de datos fMRI

Abstract

This project addresses the need for early and objective diagnosis of Parkinson’s disease by proposing the development of an Artificial Intelligence model for automatic patient classification using functional magnetic resonance imaging (fMRI) data. The main objective is to develop a deep learning model that speeds up diagnostic times and reduces the inherent influence of specialist subjectivity. Two fMRI datasets were collected and standardized: one with subjects in a resting state (DR) and another performing attention tasks (DA). Two preprocessing flows were applied: SPM for DR data and Fmriprep for DA data. Subsequently, two deep learning architectures, Vision Transformer and a hybrid 3DCNN+LSTM model were trained and evaluated with both datasets to compare their performance. The Vision Transformer model performed better with a classification accuracy of 93%. It is concluded that resting state fMRI data are the most suitable for this task and that the pre-processing flow with SPM together with the Vision Transformer model constitutes the most effective and efficient approach for assisted diagnosis of the disease. Keywords: Deep learning, Neuroimaging, Vision Transformer, SPM.