Parkinson Net: Convolutional Neural Network Model for Parkinson Disease Detection from Image and Voice Data

Abstract

Parkinson's disease (PD) is a critical dopaminergic neuron problem that causes brain disorders. The early prediction of PD can save human lives. So, computer-aided detection (CAD) with artificial intelligence (AI) models can predict the PD in quick time as compared to manual prediction. Traditional machine learning (ML) methods, on the other hand, were identified PD using either voice or image datasets. However, they resulted in poor PD detection performance, which caused misclassification. So, this work focused on the implementation of a deep learning (DL) mechanism for PD identification from both voice and image datasets, which is named ParkinsonNet. Initially, a combined dataset is considered, which contains the voice and image samples. Then, a data processing operation is performed to normalise the images to a uniform size, which also performs the data balancing operation in the voice dataset. Then, a voice-image ensemble-based convolutional neural network (VIE-CNN) model is trained with the pre-processed voice-image data. Here, categorical cross entropy (CCE) is used to optimise the losses generated during the training. Then, the VIE-CNN model predicts the normal and abnormal classes from the test data. The simulation results show that the proposed ParkinsonNet achieved 99.67% accuracy on image data and 98.21% accuracy on voice data. The simulation results show that the proposed ParkinsonNet resulted in improved accuracy over conventional methods.