TotalSegmentator · Jul 29, 2026
Anatomy Contextualized Adaption of CT Foundation Models
A paper titled 'Anatomy Contextualized Adaption of CT Foundation Models' was published on arXiv on 2026-07-29. It introduces ACA, a lightweight framework that adapts frozen CT foundation model representations for anatomy-level vision-language alignment while enhancing global contextualization. ACA uses TotalSegmentator to decompose CT volumes into anatomy-level embeddings, refined via a transformer capturing cross-anatomy relationships, and aligned to per-anatomy and scan-level text from radiology reports. Evaluated on Merlin and CT-RATE, ACA outperforms frozen foundation model baselines and existing fine-grained methods in zero-shot finding classification, requiring less than one hour of training.
What happened
A new framework called Anatomy Contextualized Adaptation (ACA) is proposed to improve CT foundation models by aligning anatomy-level visual features with text while preserving global context. ACA leverages TotalSegmentator for anatomy decomposition and a transformer for cross-anatomy relationships, achieving superior zero-shot finding classification on Merlin and CT-RATE datasets with minimal training time.
Technical significance
ACA addresses the trade-off between fine-grained anatomy-level alignment and global context by adapting frozen CT foundation model representations. It uses a transformer to model cross-anatomy interactions, enabling efficient vision-language alignment without training from scratch. The approach demonstrates that lightweight adaptation can outperform both whole-volume and fine-grained methods in zero-shot classification.
Industry impact
The method's low computational cost (under one hour of training) and use of frozen foundation models suggest potential for rapid deployment in medical imaging workflows. It could reduce barriers to adopting advanced AI in radiology by enabling efficient fine-tuning on existing models.
What to watch
Next signals include validation on larger and more diverse datasets, integration into clinical decision support systems, and extension to other imaging modalities. The approach may inspire similar adaptation techniques for other foundation models in healthcare.
Decision value
ACA offers a cost-effective way to enhance CT analysis accuracy without retraining large models, potentially improving diagnostic support and reducing time-to-insight in radiology practices. Its efficiency could lower computational and operational costs for healthcare providers.