Abstract
Quality control (QC) is a mandatory yet resource-intensive step in clinical radiology, where erroneous reports can directly affect treatment decisions. Although vision-language models (VLMs) are increasingly used to generate radiology reports, their ability to inspect existing reports—a prerequisite for automated QC—remains poorly understood. We present RadQC-Bench, a benchmark of 10,362 image–report pairs with clinically realistic synthetic errors, built from two public chest X-ray datasets (MIMIC-CXR and IU X-Ray) via LLM-based error injection covering five error types: omission, fabrication, laterality inversion, severity distortion, and typographical defects. We formulate QC as a detection-and-correction task: given an image and a report, the model must identify and fix all injected errors. Evaluating three general VLMs (GPT-5.4, Qwen3.8-Max, Gemini-3.5-Flash) and one medical VLM (MedGemma), we find that even the best model correctly repairs only 39.1% of erroneous reports on MIMIC-CXR and 42.7% on IU X-Ray, while the medical VLM almost completely fails (≤3.3%). These results indicate that current VLMs, including domain-specialized ones, are not yet reliable report inspectors, and that medical-domain pretraining alone does not confer QC ability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CREATE_019.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=6KsHDE0Ldz
BibTex
@InProceedings{FanYux_RadQCBench_MICCAISAT2026,
author = { Fang, Yuxiang AND Jin, Haibo},
title = { { RadQC-Bench: Benchmarking Vision-Language Models for Quality Control in Radiology Reporting } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17275},
month = {pending},
page = {pending}
}
