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

TL;DR

CoCa-CXR addresses temporal chest X-ray understanding by curating a temporally structured dataset (CXR-4) via an LLM-assisted pipeline and Chest ImaGenome, and by extending the CoCa framework with a regional cross-attention module to compare current and prior images. The method follows a three-stage training regime to learn robust multimodal representations and localized temporal changes, enabling both accurate progression classification and radiology report generation. It achieves state-of-the-art performance on MS-CXR-T temporal classification (65.0% average test accuracy, +4.8% over prior SOTA) and competitive RadGraph F1 on MIMIC-CXR, demonstrating improved temporal reasoning and localization for longitudinal CXR analysis. This approach provides a scalable pathway for temporally aware CXR interpretation and richer, localized abnormality understanding through cross-attention across time and structured VL supervision.

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.

Figures (4)

• Figure 1: CoCa-CXR adds a regional cross-attention module to CoCa and is trained with three stages. We utilize an LLM, Gemini, to parse MIMIC-CXR reports and get image description and pair comparison. Chest ImaGenome scene graphs provide us with the local comparison between CXR image pairs. To leverage the locality of disease conditions, we apply cross-attention to emphasize the correlation between neighboring tokens of two images.
• Figure 2: Report generation of CoCa-CXR on MIMIC-CXR validation set. If we swap the order of the image pair, the comparison prediction changes accordingly.
• Figure 3: Visualization of the text-based condition progression detection.
• Figure 4: The illustration of the four sub-datasets of CXR-4.