A recent study published in Radiology on March 25, 2025, evaluated the diagnostic accuracy of a multimodal generative AI model, utilizing 8.8 million radiograph-report pairs for training and 2,145 x-ray exams for testing, collected from 42 hospitals over an 18-year period.

Researchers assessed the quality of the AI-generated reports using a five-point scoring system, revealing that these reports received statistically higher median agreement and quality scores compared to those generated by ChatGPT-4Vision.

Brent Little, MD, in an accompanying editorial, emphasized the need for generative AI systems to innovate reporting formats rather than simply replicating existing ones to enhance the effectiveness of radiological reporting.

The findings suggest a promising integration of generative AI into radiological workflows, sparking ongoing discussions about its potential impact on the field.

Future research will focus on prospective study designs that incorporate more diverse case complexities to further improve the interpretability and usability of AI-generated reports in clinical settings.

The AI model demonstrated impressive sensitivity rates, achieving 95.3% for pneumothorax detection and 92.6% for subcutaneous emphysema, with a report acceptance rate of 70.5%, which is comparable to the 73.3% acceptance rate for human-generated reports.

In a ranking analysis, the model-generated reports were ranked highest 60% of the time, while reports generated by GPT-4Vision ranked lowest 73.6% of the time.

Led by Dr. Eun Kyoung Hong from Brigham & Women’s Hospital in Boston, the research team found the AI model effective in detecting critical conditions, contributing to a high acceptance rate of its reports among radiologists.