CoCa-CXR: Contrastive Captioners Learn Strong Temporal Structures for Chest X-Ray Vision-Language Understanding

2025-02-27

AI Summary

To address the semantic misalignment between textual descriptions of disease progression in longitudinal chest X-ray (CXR) reports and their corresponding images, this paper introduces the first vision-language joint modeling framework specifically designed for CXR temporal understanding. Methodologically: (1) we propose a CXR report temporal structure extraction pipeline; (2) we design a region-level cross-image cross-attention module to explicitly model local anatomical changes between paired CXRs; and (3) we integrate LLM-driven report parsing, contrastive image-text encoding, and multi-task contrastive learning jointly trained with image captioning. Our method achieves an average accuracy of 65.0% on MS-CXR-T—outperforming BioViL-T by 4.8%—and attains a RadGraph F1 score of 24.2% on MIMIC-CXR, matching Med-Gemini’s performance. The core contribution lies in the first end-to-end co-modeling of progression description generation and radiographic change perception.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Vision-language models have proven to be of great benefit for medical image analysis since they learn rich semantics from both images and reports. Prior efforts have focused on better alignment of image and text representations to enhance image understanding. However, though explicit reference to a prior image is common in Chest X-Ray (CXR) reports, aligning progression descriptions with the semantics differences in image pairs remains under-explored. In this work, we propose two components to address this issue. (1) A CXR report processing pipeline to extract temporal structure. It processes reports with a large language model (LLM) to separate the description and comparison contexts, and extracts fine-grained annotations from reports. (2) A contrastive captioner model for CXR, namely CoCa-CXR, to learn how to both describe images and their temporal progressions. CoCa-CXR incorporates a novel regional cross-attention module to identify local differences between paired CXR images. Extensive experiments show the superiority of CoCa-CXR on both progression analysis and report generation compared to previous methods. Notably, on MS-CXR-T progression classification, CoCa-CXR obtains 65.0% average testing accuracy on five pulmonary conditions, outperforming the previous state-of-the-art (SOTA) model BioViL-T by 4.8%. It also achieves a RadGraph F1 of 24.2% on MIMIC-CXR, which is comparable to the Med-Gemini foundation model.

Problem

• Aligning progression descriptions with image pair semantics.
• Extracting temporal structure from CXR reports using LLM.
• Improving CXR image understanding and report generation accuracy.

Innovation

• CXR report processing pipeline extracts temporal structure
• Contrastive captioner model CoCa-CXR for image progression
• Regional cross-attention module identifies local image differences