A deep learning-driven image processing pipeline for differential classification of atypical Parkinsonian disorders

Abstract

Early and accurate diagnosis of atypical Parkinsonian disorders (APD), including progressive supranuclear palsy (PSP) and multiple system atrophy (MSA), remains a major clinical challenge due to overlapping symptoms with Parkinson’s disease (PD), particularly in early disease stages. Three dimensional (3D) segmentation of brain structures from magnetic resonance imaging (MRI) has the potential to provide supportive imaging biomarkers; however, manual segmentation of relevant brain structures is impractical for large cohorts and routine clinical workflows. In this study, we present a fully automated, deep learning–based image processing pipeline designed to support the differential diagnosis of APD. A region-based U-net model was developed to segment 12 deep-brain structures implicated in APD. By dividing the MRI volumes into anatomically targeted regions around the brainstem, ventricular system, and striatum, this approach reduces GPU memory demands and training time while maintaining competitive segmentation accuracy. The segmentation outputs, including masks and volumetric features, are then incorporated together with the T1-weighted MRI into a hybrid classification framework that combines convolutional neural networks with machine learning (ML) models to explore subtype differentiation in APD. By integrating spatial image features with quantitative structural biomarkers, we demonstrate the potential of this end-to-end pipeline to serve as a computational framework that may contribute to improved differential classification of APD.