"RFJ-Bench: A Multimodal Benchmark Dataset for Rheumatic Finger Joint X-Ray Imaging and Automated Report Generation"

Abstract

"RFJ-Bench is a multimodal benchmark dataset for automated radiology report generation in rheumatology. It contains 2,239 finger joint posteroanterior (PA) X-ray examinations from 2,124 unique patients, retrospectively collected at Dazhou Central Hospital between October 2020 and March 2024 under an IRB-approved protocol with waived informed consent. Each case pairs a de-identified hand radiograph (8-bit grayscale PNG, longest side up to 1024 pixels, with Protected Health Information stripped from DICOM headers) with six structured clinical fields \u2014 chief complaint, history of present illness, past medical history, auxiliary examinations, primary diagnosis, and current diagnosis \u2014 plus patient sex and age at examination, and an expert-written reference diagnostic report produced by attending radiologists.Every clinical field and every reference report is provided in both the original Chinese and an English translation, and ready-to-train instruction files are supplied for two prompt scopes (six clinical fields, or chief complaint only) in both ms-swift and LLaVA-Rad formats.The cohort spans eight disease categories, with rheumatoid arthritis (n = 1,542) the largest, followed by undetermined arthralgia\/arthritis (n = 360), hand osteoarthritis (n = 253), and smaller groups of gouty arthritis, connective tissue disease-related arthritis, psoriatic arthritis, and juvenile idiopathic arthritis. The cohort comprises 1,709 female and 415 male patients (mean age 55.3 \u00b1 11.4 years, range 9\u201388).The dataset is partitioned at the patient level into training (70%), validation (10%), and test (20%) subsets by stratified random sampling over disease category (seed 42), yielding 1,572 \/ 221 \/ 446 images. RFJ-Bench supports multimodal report generation, together with text-only clinical auxiliary diagnosis and image-only disease identification task definitions, and All-X \/ Only-X ablation designs for quantifying the marginal contribution of each clinical field. Baseline results for twelve multimodal large language models, evaluated with standard NLG metrics and a rheumatology-adapted CheXprompt clinical consistency framework, are reported in the companion paper. "