Original title: RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching
Charles Corbière, Léo Machado, Aubin Charley, Baptiste Callard +2 · September 2026
Artificial intelligence tools can now write draft medical reports for radiology scans. Checking if these drafts are safe and accurate is hard. Current AI evaluators give just one mysterious score, making it tough for doctors to see mistakes. Researchers created RadMatch to fix this. Instead of scoring the whole page at once, RadMatch breaks reports down into single findings. It matches each finding, checks its importance, and grades it across seven clinical details, such as body location and severity. It labels findings as correct, partial, incorrect, missed, or hallucinated. Its main score simply counts critical, actionable errors. Across two expert test sets, RadMatch matched the agreement rate of human radiologists and beat previous tools.
Imagine you have a recipe for a cake and someone gives you a version they wrote. Normally, you might just taste it and say if it's good. But to check it properly, you'd compare each ingredient and step, like how much sugar or baking time. In this paper, the system compares medical reports by matching each detail one by one and pointing out exactly where they differ.
This matters because radiologists can now audit AI-generated reports to ensure they are clinically accurate. It improves patient safety by catching errors in automated diagnoses.
The abstract doesn't specify limitations, but it focuses on radiology reports and doesn't confirm testing on other imaging modalities or anatomies.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching — Charles Corbière, Léo Machado, Aubin Charley et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.