Segmentation-Guided Radiology Report Generation for Pneumothorax Detection in Chest X-Rays

Recent developments on chest radiographs has primarily focused on developing multi-disease frameworks that aim to diagnose a wide range of thoracic abnormalities from the Chest X-ray datasets. In contrast, this study specifically targets pneumothorax, a life-threatening condition commonly referred to as a collapsed lung, which requires timely detection and accurate clinical reporting. Existing automated report generation Vision-Language Models (VLMs) mainly rely on image-level features and often fail to fully leverage the rich structural information embedded in medical image segmentation. To address this limitation, we propose a distinct strategy to incorporate pneumothorax segmentation masks, which delineate affected regions and provide precise localization guidance to enhance the accuracy of medical image interpretation. Experimental results demonstrate that the proposed segmentation-guided approach integrates visual and textual understanding more effectively for pneumothorax diagnosis from chest radiographs. By employing segmentation masks as guidance, VLMs can accurately localize pathological regions while preserving anatomical context, thereby improving both interpretability and diagnostic precision. Quantitative evaluations across multiple metrics further confirm the effectiveness of the proposed methods in bridging the gap between image-level localization and report-level reasoning.