Simple anatomical features match complex AI models on brain MRIs

In a preprint analyzing structural brain MRI scans from approximately 80,000 human participants across 18 datasets, researchers evaluated three AI feature-extraction strategies across seven clinical tasks. They compared explicit anatomical surface and volume measurements against supervised convolutional neural networks and pretrained vision transformer foundation models. A linear model built on basic anatomical features matched the diagnostic performance of complex AI frameworks trained on thousands of scans. The authors also found that deep learning networks implicitly learn relevant anatomical details on their own. Using this insight, the team introduced Anatomy Segmentation Pretraining, an approach that incorporates anatomical segmentation directly during model pretraining and outperforms existing models in biological age estimation.