Learning anatomy's consistent patterns across brains and modalities

Medical imaging models typically learn from individual patients in isolation, missing the fact that anatomical structures maintain consistent spatial relationships across people. This work exploits cross-instance topological consistency as a supervisory signal through two alignment strategies: intra-instance alignment using pixel correspondences to preserve local topology across modalities, and inter-instance alignment using pseudo-correspondences to align anatomical neighborhoods without explicit supervision. Tested on 7 downstream tasks across multi-modal 3D medical imaging, the approach achieves 1.1% average improvement in segmentation and 5.94% in classification, with notably better robustness when modalities are missing at test time.