A review of new science, in plain language

Nature Communications  ·  12 November 2026  ·  From issue 66

Scaling medical imaging report generation with multimodal reinforcement learning

An AI trained with a reward system beat older models at writing chest X-ray reports.

Qianchu Liu, Sheng Zhang et al.

An AI model wrote chest X-ray reports by learning to avoid medical mistakes, not just copy patterns from training data. It was trained on four public hospital datasets, then tested on a private dataset from a different source. It beat older AI models by more than 50% on two test sets. It also made fewer serious clinical errors and kept working well across different patient groups.

The number over 50% improvement

The gain over the best prior AI model on two chest X-ray report test sets.

Tested on public and one private hospital dataset, not real clinical use.

Integrity screen: passed (3 checks) Checked 13 November 2026. Retraction record: none. DOI resolves at doi.org. Metadata record found (Nature Communications). Read the source

This screen checks process signals. It does not tell you the findings are true.

Five vetted papers in plain language, free, every two days

Five papers free every issue, about 15 emails a month. No spam. Leave in one click.

Also in issue 66 15 papers
More from Nature Communications