Automatic segmentation of choroid plexus using deep learning across neurodegenerative diagnoses in the multi-site COMPASS-ND Study

The choroid plexus (ChP) plays a central role in cerebrospinal fluid production, immune signaling, and metabolic clearance, and has emerged as a potential imaging biomarker of neurodegeneration. However, accurate and scalable quantification of ChP volume remains challenging due to its complex morphology and low contrast on conventional MRI. The Automatic Segmentation of Choroid Plexus (ASCHOPLEX), a deep learning framework originally trained on healthy controls and multiple sclerosis cohorts, has not been systematically evaluated in neurodegenerative populations. Using T1-weighted MRI from the multi-center COMPASS-ND study, we assessed standard ASCHOPLEX performance in cognitively unimpaired (CU), Alzheimer's disease (AD), and Parkinson's disease (PD) participants (N = 30), followed by fine-tuning using expert manual segmentations (N = 60). Segmentation accuracy was evaluated using Dice, Jaccard, precision, and recall. The fine-tuned model was then applied to a larger cohort (N = 277) to derive normalized ChP volumes, which were compared across diagnostic groups using linear regression models. Fine-tuning significantly improved segmentation accuracy across all metrics (Dice: 0.45 to 0.84; Jaccard: 0.32 to 0.73; all p < 0.0001), enabling robust ChP delineation across sites and conditions. In the full cohort, normalized ChP volume was significantly higher in AD compared with CU and PD (p < 0.0001), while PD did not differ from CU (p = 0.31). These findings demonstrate that dataset-specific adaptation is essential for deploying deep learning segmentation models in heterogeneous neuroimaging cohorts. The refined ASCHOPLEX framework enables scalable ChP quantification and supports its use as a structural imaging marker in neurodegenerative disease.