Generative Adversarial Networks as a Data Augmentation Tool for CNN-Based Parkinson's Disease Diagnostics

Abstract

Growing research interest has arisen towards automated neurodegenerative disease diagnostics based on the information extracted from the digital drawing tests. Since the performance of modern modelling techniques (machine learning, deep learning) relies heavily on the size of training data available, data scarcity is one of the most significant problems in computer-aided diagnostics. This paper proposes using Generative Adversarial Networks to synthesise digital drawing tests acquired from Parkinson's patients and healthy controls. Four different architectures (StyleGAN2-ADA, StyleGAN2-ADA + LeCam, StyleGAN3 and ProjectedGAN) are evaluated and compared with the traditional data augmentation methods. Convolutional neural networks are utilised for Parkinson's disease diagnostics. Our results indicate that GAN-generated images’ addition outperforms the standard augmentation methods in classifying Parkinson's disease in some settings. Therefore, the proposed framework could serve as a potential decision support tool for clinicians in computer-aided fine-motor analysis for neurodegenerative disease diagnostics.