Researchers have developed a new framework called Anatomy Contextualized Adaptation (ACA) that improves the performance of CT foundation models in vision-language tasks. These models typically use whole-volume representations, which can dilute fine-grained anatomical signals. ACA addresses this by decomposing CT volumes into anatomy-level embeddings and refining them with a transformer that captures cross-anatomy relationships. The framework is lightweight and requires less than one hour of training once the embeddings are cached. It outperforms both frozen foundation model baselines and existing fine-grained methods in zero-shot finding classification. --- Why it matters: This matters to engineers working on vision-language tasks because ACA provides a more effective way to adapt CT foundation models, which can improve performance in downstream applications such as medical imaging analysis. The framework's ability to preserve and enhance global anatomical context is particularly relevant for tasks that require fine-grained understanding of anatomy. Source: https://arxiv.org/abs/2607.27154