The abstract is read at the publisher; the summary is JClub's.
AnatCL and y-Aware produced the most reliable cross-scanner brain MRI embeddings, with dbICC of 0.91–0.92 (up to ~0.95 for FreeSurfer-derived anatomical measurements) in ON-Harmony travelling-heads data.
- Why it matters: Scanner-to-scanner variability limits trust in foundation-model MRI analyses, and it is unclear which model designs keep embeddings and brain-age predictions consistent across acquisition sites.
- What they did: Seven FMs were tested on 120 acquisitions from 20 participants in ON-Harmony, evaluating age prediction accuracy plus cross-scanner consistency of embeddings and predictions; brain-age predictors were trained on 1,411 healthy adults via Ridge on frozen embeddings or fine-tuning.
- The result: Ridge predictions had higher dbICC than full embeddings for NeuroVFM (0.88 vs 0.19) and Curia-2 (0.84 vs 0.16), while fine-tuning yielded ~2-year cross-scanner differences for AnatCL/y-Aware versus 4–6 years for NeuroVFM/Curia-2, improving internal-test accuracy in 3/7 models and reducing cross-scanner gaps in 2.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.