BoneMammo-RADS

RADS-Ref is an evaluation resource for measuring the factual accuracy of radiology reports generated by vision language models (VLMs). Unlike metrics built on lexical overlap or generic entity-relation pairs, RADS-Ref is anchored in RADS-defined key features — the expert-defined, task-specific elements that drive clinical decisions. The dataset covers two radiology tasks: Bone radiography (Bone-RADS): 221 cases with de-identified X-ray images, radiologist reference reports, and structured Bone-RADS key-feature annotations (margin, periosteal reaction, endosteal scalloping, pathological fracture, soft tissue mass). Mammography (BI-RADS): 30 cases with de-identified mammography images, radiologist reference reports, and structured BI-RADS key-feature annotations covering mass features, calcifications, architectural distortion, asymmetry, and associated findings for both breasts. Each case links one or more de-identified images to an expert reference report and a structured RADS key-feature annotation. The dataset was used to validate RADS-guided metrics (Element Alignment Score, EAS; Side-Aware Jaccard, SAJ) against radiologists' factual accuracy ratings and to benchmark VLMs against human readers. Intended use: developing and validating evaluation metrics for radiology report generation; benchmarking VLM-generated findings against expert reference reports; research on clinical factuality of medical text generation. This dataset is a research benchmark, not a clinical tool. Dataset structure: RADS-Ref/├── bone/│ ├── images/ # 433 de-identified radiographs (JPG, max 1024 px)│ ├── reports/ # 221 reference reports (TXT, Chinese)│ └── annotations/ # 221 Bone-RADS key-feature annotations (JSON)├── mammo/│ ├── images/ # 120 de-identified mammograms (PNG, max 1536 px)│ ├── reports/ # 30 reference reports (TXT, Chinese)│ └── annotations/ # 30 BI-RADS key-feature annotations (JSON)└── manifest.json # 251-entry dataset registry Ethics: The study was approved by the relevant institutional review boards (IRB no. <2026-KY-138[K]>) of Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, and the requirement for informed consent was waived for this retrospective, de-identified dataset. Access is gated. This dataset contains de-identified patient imaging and radiologist-authored reference reports. Access is granted for non-commercial academic research only and requires manual approval. By requesting access you agree to the Data Use Agreement . For reasonable access requests, please email [e-mail] Release schedule: v1.0 (current): Bone radiography (221 cases) + Mammography (30 cases) with images, reference reports, and RADS key-feature annotations. Post-acceptance: Remaining portions including VLM-generated outputs and radiologist factuality ratings from the reader study will be released after the associated manuscript is formally accepted for publication. Citation: If you use RADS-Ref, please cite: Wang M, Tu Y, et al. RADS-guided Metric for Factual Accuracy of VisionLanguage Model Reports in Bone Radiography and Mammography. [Journal], [Year].