基於視覺語言模型與放射科關聯圖譜之胸腔 X 光報告生成校正系統

Abstract

視覺語言模型（Vision-Language Models, VLMs）在自動化胸腔 X 光（Chest X-ray, CXR）報告生成任務上已展現出良好潛力，然而其臨床可靠性仍受限於關鍵發現遺漏與幻覺生成等問題。本文提出一種輕量化的報告校正框架，透過基於圖結構的檢索機制與問題導向的推理流程，以提升胸腔 X 光報告的準確性與完整性。首先，將由 VLM 生成的初始報告轉換為結構化圖表示，並據此自放射科關聯圖譜知識圖譜中檢索相關臨床關係。接著，利用檢索所得之關係生成具針對性的問題，引導視覺語言模型進行視覺推理，並逐步精煉原始報告內容，以降低遺漏現象。

Vision-Language Models (VLMs) have demonstrated promising potential in automating Chest X-ray (CXR) report generation tasks; however, their clinical reliability remains limited by issues such as missing key findings and hallucination generation. This paper proposes a lightweight report correction framework that enhances the accuracy and completeness of chest X-ray reports through a graph-based retrieval mechanism and a question-oriented reasoning process. First, the initial report generated by the VLM is converted into a structured graph representation, and relevant clinical relationships are then retrieved from Radiology Association Graph. Subsequently, targeted questions are generated using the retrieved relationships to guide the VLM in visual reasoning and progressively refine the original report content to reduce omissions.