← News

ResearchSources: arXiv · 16 Nov 2026

AG-JEPA adds anatomy-aware training to brain MRI models for Alzheimer's detection

Researchers describe a self-supervised method that focuses MRI representation learning on brain regions tied to neurodegeneration, reporting gains in early diagnosis and progression prediction tasks.

In plain words

The authors of a new arXiv paper propose AG-JEPA, a way of training AI on brain MRI scans so it pays attention to the specific anatomy affected by Alzheimer's disease. In tests, their approach classified patients better than existing methods and, in a variant, forecast how disease would progress. The results come from the paper's own experiments and have not been independently verified.

What was announced

The paper argues that standard JEPA models, which learn general structural features from images, miss the subtle region-specific changes that mark the shift from healthy cognition to mild impairment. AG-JEPA addresses this by introducing anatomical tokens that interact with image patch tokens during pretraining, steering the learned representations toward areas relevant to neurodegeneration. A companion variant, Prog-AG-JEPA, represents disease progression as a conditioned latent displacement within the same representation space, according to the abstract.

The authors report evaluation across five downstream tasks covering Alzheimer's diagnosis and progression prediction. They state AG-JEPA beat the strongest baseline by 8.4 percentage points in AUROC on cognitively normal versus mild cognitive impairment classification, and by 7.1 points on out-of-distribution CN/AD classification using the MR-RATE dataset. For progression modeling, they say Prog-AG-JEPA outperformed all baselines on future diagnosis and brain volume change prediction. These figures are the authors' own reported results.

What it does not show

  • Evidence is limited to the paper's abstract; the full article was not reviewed, so methods and experimental details could not be checked.
  • All performance numbers are author-reported on their chosen tasks and datasets; no independent replication or external validation is described in the supplied material.
  • The abstract does not specify dataset sizes, training data, or clinical readiness, so no conclusions about real-world deployment can be drawn.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.