The abstract is read at the publisher; the summary is JClub's.
AD-trained nnU-Net PVS segmentation (ADNI-PVS-Net) reached overall Dice 0.58 ± 0.14 internally in ADNI-3, but degraded externally to Dice ≤ 0.11.
- Why it matters: Internal accuracy may not preserve biologically meaningful PVS–biomarker relationships when models are applied to heterogeneous AD cohorts and external scans.
- What they did: Five open-source nnU-Net-based models were benchmarked against manual labels in 60 multi-site ADNI-3 participants and 12 VALDO scans, using overall/voxel/object/lesion-wise Dice and stratified checks by region, diagnosis, scanner, PVS severity, and WMH burden.
- The result: Only AD-trained models reproduced the expected PVS–amyloid biomarker relationship, while other models gave null results despite comparable or higher Dice, implying cohort-specific validation is required.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.