Consistent View Alignment Improves Foundation Models for 3D Medical Image Segmentation
Overview
Many self-supervised representation-learning methods assume that any two “views” of the same data point are automatically useful for learning — but this work shows that meaningful latent structure doesn’t emerge on its own from uncorrelated views; it has to be explicitly encouraged, especially in heterogeneous 3D medical volumes where mismatched views can hurt more than they help.
Method
The paper introduces Consistent View Alignment (CVA), a regularisation technique that constructs views using explicit region correspondences and aligns them at the patch/token level. This avoids the false positives that arise when contrastive learning is forced to match unrelated regions, while still enforcing local feature consistency.
Results
CVA improves downstream performance for 3D medical image segmentation and highlights the importance of structured view alignment in representation learning. The method placed 1st (Primus vision transformer track) and 2nd (ResEnc convolutional track) in the MICCAI 2025 SSL3D challenge.