Grounding Radiology Report Findings into Medical Image Segmentation

Routine radiology practice documents abnormalities primarily in narrative reports rather than pixel-level annotations. However, many clinically important analyses require spatially localized evidence. This mismatch limits the reuse of large-scale image-report archives for spatially resolved quantitative analysis, as heterogeneous clinical findings cannot be directly translated into consistent spatial representations. We therefore present Clinical-Findings-Guided Segmentation (CF2Seg), a chest X-ray segmentation framework that learns spatial representations directly from full radiology findings. We further curate a multi-source benchmark comprising 53,386 examinations, each paired with expert spatial annotations and corresponding reports across diverse thoracic pathologies and institutions. CF2Seg uses a dual-stage text–image fusion strategy in which report findings guide feature formation and are revisited during decoding to improve boundary delineation across scales. On a unified multi-source benchmark, CF2Seg achieves strong overall performance and remains stable under clinically realistic variation, including distribution shift, annotation scarcity, and clinically realistic image and report perturbations. Beyond segmentation accuracy, CF2Seg-derived masks preserve proportional lesion-burden trends with strong agreement to expert annotations, supporting retrospective spatial annotation and quantitative reuse of large-scale clinical image–report archives.