3D Medical Image Segmentation with Anatomy-Guided Conditioning from Surrounding Structures

ICLR 2026 Conference Submission14694 Authors

Keywords: Anatomy-guided conditioning, Signed distance maps, Spatial context integration, Anatomical priors

TL;DR: We propose a novel anatomy-guided conditioning that injects signed distance maps of surrounding structures into segmentation networks, improving structural fidelity and boundary accuracy across diverse 3D medical imaging tasks.

Abstract:

Accurate segmentation of complex anatomical structures in 3D medical images is challenged by low contrast, unclear features, and complex topology. We propose Anatomy-Guided Conditioning (AGC), which integrates signed distance maps of surrounding anatomy into segmentation networks via feature modulation in the decoder.
Anatomical priors are obtained from automatic tools such as TotalSegmentator, requiring no additional training and enabling use in both multi-class and single-target tasks. We evaluate AGC on CTA coronary arteries, PET/CT visceral fat, CT head-and-neck organs, and CBCT dental canals.
Across CNN, Transformer, and hybrid backbones, AGC improves Dice, HD95, and topology-aware metrics (clDice, Betti error), reducing boundary errors and fragmentation. These results demonstrate that conditioning on surrounding anatomy provides a simple and broadly applicable inductive bias for anatomically constrained 3D segmentation.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 14694