SAIL: Structure-Aware Interpretable Learning for Anatomy-Aligned Post-hoc Explanations in OCT

Anatomical knowledge improves AI trustworthiness in medical imaging: The SAIL framework demonstrates that integrating anatomical priors into deep learning models produces sharper, more clinically meaningful explanations for retinal disease detection in OCT scans, addressing a key barrier to clinical adoption where explainability is essential for regulatory approval.

Post-hoc explanations alone are insufficient: Existing explainability methods struggle to respect anatomical boundaries and suppress noise in medical images. The paper shows that combining structural priors with semantic features at the representation level—rather than applying explanations after the fact—is critical for producing trustworthy AI outputs.

Structure and semantics must work together: Ablation studies reveal that strong interpretability requires both anatomical structural information and semantic features; neither alone is sufficient, highlighting the importance of thoughtful system design rather than treating explainability as an add-on.

Better explanations enable clinical deployment: By producing anatomy-aware attribution maps that clinicians can understand and verify, structure-aware interpretable learning helps bridge the gap between achieving expert-level accuracy and gaining the clinical trust necessary for real-world adoption of AI diagnostic tools.