Anatomy Contextualized Adaption of CT Foundation Models

TL;DR - Anatomy Contextualized Adaptation (ACA) efficiently adapts frozen CT foundation models for anatomy-level vision-language alignment while retaining whole-scan context. It improves zero-shot finding classification on Merlin and CT-RATE with under one hour of training after embedding caching.

• TotalSegmentator decomposes CT volumes into anatomy-level embeddings.
• A transformer models cross-anatomy relationships and aligns embeddings with anatomy-specific and scan-level report text.
• ACA outperforms frozen foundation-model baselines and existing fine-grained methods.
• Learned attention patterns suggest plausible routing of context across anatomical regions.