Microsoft unveils UniRG to improve accuracy in AI-generated radiology reports

Microsoft Research has developed a reinforcement learning framework, UniRG, to improve the clinical accuracy of automated radiology report generation, launching its first chest x-ray model, UniRG-CXR.

The company said existing supervised approaches to report generation tend to overfit, often replicating institution-specific phrasing and generating fluent but clinically inaccurate reports when applied to data from unfamiliar settings.

According to a Microsoft blog post, UniRG combines traditional supervised fine-tuning with reinforcement learning that optimises a composite reward.

This reward integrates rule-based metrics, model-based semantic scores and clinical-error signals derived from large language models to bring training in line with radiology best practices.

UniRG-CXR was trained on more than 560,000 imaging studies, 780,000 chest x-ray images and data from 226,000 patients across over 80 medical institutions.

Related reading

• Demis Hassabis: why AGI is closer than it looks, and further away than the hype suggests
• Apple’s rumoured January event: Refresh, or reset?
• The new AI tech that aims to beat electricity grid constraints

Microsoft said the model achieved state-of-the-art performance as of 22 January on the ReXrank leaderboard, a benchmark that assesses medical report generation across multiple datasets, diagnostic tasks, longitudinal scenarios and demographic subgroups.

The company added that UniRG-CXR produces reports with fewer clinically significant errors, maintains strong performance across unseen institutional data, and is capable of comparing current studies with prior exams in longitudinal settings.

The Recap

• UniRG applies reinforcement learning to medical report generation.
• UniRG-CXR trained on over 560,000 studies and 780,000 images.
• Framework can be extended to other modalities and tasks.