Identification of latent biomarkers in brain imaging of Parkinson’s disease using explainable artificial intelligence

Abstract

Parkinson’s disease (PD) is the second most common neurodegenerative disorder worldwide, characterized by the progressive degeneration of dopaminergic neurons. Early diagnosis remains challenging due to the lack of specific clinical tests. Although imaging techniques such as SPECT and MRI are commonly used to support diagnosis, their analysis is most often limited to striatal regions. In this study, we introduce a deep learning-based method for PD detection, while also exploring the role of non-striatal brain regions, which are often overlooked. Using DaTSCAN volumes, we trained a three-dimensional convolutional neural network (3D CNN) to distinguish control subjects from PD patients at different stages (1 to 3). To interpret the predictions, we applied the Grad-CAM technique to localize the regions influencing the model’s decision. Our network achieved remarkable accuracy (>97%) across all stages of the disease, and the Grad-CAM maps revealed a significant involvement of cortical and subcortical regions beyond the striatum. These findings suggest the existence of early or complementary biomarkers, highlighting the value of explainable artificial intelligence in brain imaging to refine PD diagnosis and broaden our understanding of how the disease develops and affects the brain