AI-Generated Radiology Reports Show Promise in Reducing Reading Times
New research suggests that artificial intelligence could help radiologists work more efficiently, though important questions remain about accuracy and clinical impact. A study published in the Journal of the American College of Radiology reports that AI-generated chest X-ray reports reduced radiologists’ interpretation times by about one-quarter while also gaining greater acceptance among physicians over the course of the trial.
The project, led by Eun Kyoung Hong, MD, PhD, from the Department of Radiology at Mass General Brigham in Boston, investigated whether a generative AI tool could streamline the process of radiograph reporting. The researchers tested Kakao Brain’s KARA-CXR model, which is designed for research use, to create “findings-only” reports without impressions or treatment recommendations. These preliminary reports were then reviewed by radiologists to measure efficiency, quality, and acceptance.
The study relied on a dataset of 756 publicly available chest X-rays. Five radiologists used the AI tool to generate reports, and two additional radiologists assessed the quality of the outputs. Over the course of the experiment, the average time required for reading and reporting decreased from 25.8 seconds to 19.3 seconds—an improvement of roughly 25%. This time savings, while modest per case, could scale meaningfully in high-volume clinical environments.
Equally notable was the change in how frequently radiologists approved AI-generated reports without making edits. Acceptability rose from 54.6% at the start to 60.2% by the end, suggesting growing trust in the technology as radiologists became more familiar with it. However, the AI system demonstrated stronger performance in straightforward cases: reports were deemed acceptable 68.9% of the time when no abnormalities were present, compared with just 52.6% for scans that showed disease.
Despite these encouraging signs, the authors stressed that AI should not be seen as a replacement for radiologists, particularly in complex or ambiguous cases. Agreement and quality scores for scans with abnormalities varied significantly, underlining the importance of human oversight in ensuring diagnostic accuracy.
The study also raised questions about whether these measured improvements translate into real-world gains. While the reductions in reading time and the uptick in acceptance were statistically significant, the researchers noted they might not yet represent clinically meaningful changes in workflow or patient outcomes. The team emphasized the need for prospective trials in live clinical settings to determine how radiologists’ growing comfort with AI influences their decision-making and diagnostic performance over time.
In their conclusion, Hong and colleagues called for further research into strategies that could enhance the accuracy of AI-generated preliminary reports. They also recommended evaluating how radiologists’ confidence in these tools affects their workflow and clinical decisions.
For now, the findings add to a growing body of evidence showing that AI can complement radiologists by boosting efficiency and reducing repetitive tasks. Still, the path forward requires careful study to ensure that time savings do not come at the expense of diagnostic rigor or patient safety.