UNIRG trains X-ray report generation to reduce clinical errors
Microsoft researchers have combined supervised training with reward-based learning to improve automated chest X-ray reports. Their UNIRG framework evaluates both report similarity and clinical mistakes, with radiologist assessments supplementing automated scores. Tests use existing imaging datasets, including a separate external collection. The authors report improvements over comparison systems while retaining human oversight because generated reports still contain errors.
Artificial Intelligence··Midday
Training for clinically useful reports
Microsoft Research scientists have introduced UNIRG, a framework for producing medical imaging reports. Their Nature Communications paper applies it to chest radiograph reports. The model first learns from existing examples, then receives rewards for generating reports that align with clinical quality criteria. This second step aims to move training beyond matching the wording of a reference report. The chest-imaging model uses an eight-billion-parameter base model and is evaluated against other report-generation systems.[1]
Scores are checked against expert readings
The reward combines language similarity with clinical information and errors. Its stages target a composite report-quality measure and then add a reward for reducing mistakes. Evaluation uses held-out images and an external collection of 11,815 studies, alongside assessments by 4 radiologists. The specialists examine completeness, factual accuracy and clinically meaningful errors. The researchers report better ratings and fewer false findings and omissions than the comparison systems, while acknowledging that incorrect reports remain possible.[1]
Existing images, continuing oversight
The study draws on existing imaging datasets and collects no new patient data. Its tests assess report generation rather than prospective patient outcomes. The authors say radiologist oversight remains necessary because the model still makes errors. They release model weights and inference code, while licensing terms prevent redistribution of the proprietary images and radiologist evaluation data. The work is confined to chest X-ray reports; applying the same framework to other imaging methods remains future research.[1]
Related columns
For more information on this topic, you can read the related columns.
