Abstract

Parkinson’s disease (PD) exhibits substantial clinical and neuroanatomical heterogeneity, limiting robust patient stratification and clinically meaningful progression modeling from MRI. We propose a unified multimodal 3D generative representation-learning framework that learns an interpretable latent space from co-registered baseline T1/T2 MRI with an edge-aware channel. Confound-corrected latent embeddings support unsupervised subtype discovery, while disease duration provides weak supervision to orient a continuous progression axis. On an independent external cohort (PPMI, N=171), the discovery-trained subtype structure shows significant partial replication (ARI=0.35; permutation test p=0.001) and enables longitudinal clinical stratification: mixed-effects modeling reveals subtype-dependent MDS-UPDRS III progression, with the strongest effect in subtype 2 (p< 0.001) persisting after adjustment for levodopa equivalent daily dose (LEDD), alongside differences in cognitive trajectories. Progression-axis scores correlate with disease duration (ρ=0.248), and decoder-based latent traversal yields voxel-wise visualizations consistent with known PD patterns, supporting generative multimodal MRI embeddings as interpretable imaging biomarkers for translational PD research.

Committee Chair

Dr. Shinjini Kundu

Committee Members

Dr. Aimilia Gastounioti, Dr. Joseph O’Sullivan

Degree

Master of Science (MS)

Author's Department

Electrical & Systems Engineering

Author's School

McKelvey School of Engineering

Document Type

Thesis

Date of Award

Spring 5-2026

Language

English (en)

Share

COinS