基于解剖区域先验的胸部X光报告生成方法
Anatomical Region Prior-Guided Chest X-ray Report Generation
陈泓博 1尹斯星1
作者信息
- 1. 北京邮电大学信息与通信工程学院,北京 100876
- 折叠
摘要
胸部X光报告生成旨在将胸片自动转化为符合临床表述习惯的放射学报告,对缓解放射科医师报告撰写压力具有现实意义。然而,现有方法多以全图视觉特征直接驱动下游生成模型,背景区域与关键解剖区域的视觉token在注意力计算中相互竞争,易稀释关键解剖区域的表示并削弱生成报告的临床相关性。为此,本文提出一种基于解剖区域先验的胸部X光报告生成方法：使用MedSAM提取肺野区域掩膜,将其映射至ViT-B/14的patch网格,依据重叠关系对视觉token进行筛选与加权增强,并经两层MLP投影层接入大语言模型作为视觉提示。在MIMIC-CXR上的实验表明,所提方法在CheXbert-14宏平均F1与RadGraph-F1两项临床一致性指标上相较于现有公开可比方法取得提升,同时在ROUGE-L等语言生成质量指标上保持有竞争力的水平。结果表明,以解剖区域先验引导视觉特征增强是改善胸部X光报告生成临床一致性的一条可行路径。
Abstract
Chest X-ray report generation aims to automatically transform chest radiographs into radiology reports that follow clinical conventions, easing the reporting workload of radiologists. However, existing methods typically feed full-image visual features into downstream generative models, where background tokens compete with tokens of key anatomical regions during attention computation, diluting anatomically relevant representations and weakening the clinical relevance of generated reports. To address this, we propose an anatomical region prior-guided chest X-ray report generation method: lung field masks extracted by MedSAM are mapped onto the patch grid of ViT-B/14, visual tokens are selected and weighted according to their overlap with the lung field, and the resulting features are projected through a two-layer MLP into Large Language model as visual prompts. Experiments on MIMIC-CXR show that the proposed method achieves improvements over existing publicly comparable methods on the two clinical consistency metrics CheXbert-14 macro-F1 and RadGraph-F1, while maintaining competitive language generation quality on ROUGE-L. The results indicate that guiding visual feature enhancement with anatomical region priors is a viable approach to improving the clinical consistency of chest X-ray report generation.关键词
医学图像处理/胸部X光/放射学报告生成/解剖区域先验/大语言模型Key words
chest X-ray/radiology report generation/anatomical region prior/MedSAM/large language model引用本文复制引用
陈泓博,尹斯星.基于解剖区域先验的胸部X光报告生成方法[EB/OL].(2026-05-08)[2026-05-10].http://www.paper.edu.cn/releasepaper/content/202605-1.学科分类
医学研究方法/临床医学
评论