MvRNA: A New Multi-view Deep Neural Network for Predicting Parkinson's Disease

Lin Chen, Yuxin Zhou, Xiaobo Zhang, Zhehao Zhang, Hailong Zheng

Abstract: Magnetic Resonance Imaging (MRI) is a critical medical diagnostic tool that assists experts in precisely identifying lesions. However, due to its high-dimensional nature, it requires substantial storage resources. If only one MRI slice were to be used, a significant amount of information might be lost. To address these issues, we propose segmenting 3D MRI data and training these slices separately. We propose a new Multi-view Learning neural network based on ResNet and an Attention mechanism, called MvRNA. ResNet18 is selected as the backbone network, and the Squeeze-and-Excitation network is applied between blocks to extract features from slices. Additionally, we propose a new BWH (Basic Block with Hybrid Dilated Convolution) module to capture a broader range of receptive fields, thus acquiring additional spatial features. We obtained data from Parkinson’s Progression Markers Initiative (PPMI) and applied our method to distinguish between Healthy Control, Prodromal, and Parkinson’s disease patients. The experimental results demonstrate that our method achieved an accuracy of 81.84%.