Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Anatomy-Aware Masked Image Modeling for Self-Supervised Learning on 3D Brain MRI

Yeonwoo Kim¹, Won Hee Lee¹; ¹Kyung Hee University

Presenter: Yeonwoo Kim

Self-supervised pre-training on structural brain MRI can learn transferable representations without diagnostic labels, but standard masked image modeling (MIM) masks patches uniformly at random, ignoring the brain's anatomical organization. Scattered patch reconstruction can be solved from local intensity cues alone, bypassing the inter-regional reasoning most relevant to neuropsychiatric conditions. We propose anatomy-aware MIM, which replaces random patch masking with region-level masking derived from cortical parcellations (aparc+aseg). By masking entire anatomical regions at a time, the encoder is forced to reconstruct missing brain structures from long-range inter-regional context rather than local neighborhoods, with no additional supervision or parameters. Integrated into a multi-task pre-training framework based on 3D Swin Transformer and evaluated via linear probing, anatomy-aware MIM improves schizophrenia classification over random MIM: balanced accuracy from 64.7% to 66.9%, AUC from 0.767 to 0.771, and recall from 0.398 to 0.451 (+13.3% relative). These results support a broader principle that domain-appropriate masking is a meaningful inductive bias for self-supervised learning in structured biological systems.

Topic Area: Methods, Tools, Theory & Neural Coding