---
title: "Anatomy Contextualized Adaptation of CT Foundation Models: A Lightweight Framework ACA"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.27154v1"
permalink: "https://feed.astralor.com/items/itm_cc506801364be66f303535e1"
published: "2026-07-29T17:32:57.000Z"
score: 81
channel: "ai"
tags: ["ACA","CT Vision-Language Foundation Model","TotalSegmentator","Zhejiang University","Zero-Shot Finding Classification"]
---

# Anatomy Contextualized Adaptation of CT Foundation Models: A Lightweight Framework ACA

## AI commentary

> This paper presents a highly practical lightweight framework ACA, especially suitable for medical AI teams with limited compute resources. It elegantly combines global representation of CT foundation models with anatomy-level fine-grained alignment, achieving remarkable training efficiency. For researchers in medical image analysis, ACA offers a deployable baseline, particularly worth noting for its improvement in zero-shot finding classification.

## At a glance

CT vision-language foundation models show promising performance in downstream tasks but typically use whole-volume representations that dilute fine-grained anatomical signals. Existing fine-grained approaches align anatomy-level features with text but discard global context and are computationally expensive. This paper proposes Anatomy Contextualized Adaptation (ACA), a lightweight framework that decomposes CT volumes into anatomy-level embeddings via TotalSegmentator, refines them with a cross-anatomy transformer, and aligns with per-anatomy and scan-level text from radiology reports. On Merlin and CT-RATE, ACA consistently outperforms frozen baselines and fine-grained methods in zero-shot finding classification, requiring under one hour of training, with attention weights indicating plausible cross-anatomy context routing.

## Key points

- ACA decomposes CT volumes into anatomy-level embeddings via TotalSegmentator, preserving fine-grained signals.
- It uses a cross-anatomy transformer to capture inter-anatomy relationships and enhance global context.
- Outperforms frozen baselines and existing fine-grained methods in zero-shot finding classification with under one hour of training.
- Attention weights reveal plausible cross-anatomy context routing mechanisms.

## Full text

CT vision-language foundation models have demonstrated promising performance across downstream tasks, but are typically trained with whole-volume representations that dilute fine-grained anatomical signals. Fine-grained vision-language pre-training addresses this by aligning anatomy-level visual features with anatomy-specific text, but in doing so discards the global context that whole-volume models provide. Furthermore, existing fine-grained approaches train from scratch, making them computationally expensive. We introduce Anatomy Contextualized Adaptation (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, which are refined via a transformer that captures cross-anatomy relationships, and aligned to both per-anatomy and scan-level text extracted from radiology reports. Evaluated on Merlin and CT-RATE, ACA consistently outperforms both the frozen foundation model baselines and existing fine-grained methods in zero-shot finding classification, while requiring less than one hour of training once embeddings are cached. The attention weights learned by ACA's inter-anatomy transformer additionally indicate plausible cross-anatomy context routing. Altogether, these results support ACA as a lightweight approach for adapting CT foundation models to anatomically grounded vision-language alignment while preserving and enhancing global anatomical context.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_cc506801364be66f303535e1