Mass General Brigham / Brigham and Women's Hospital

Radiologists face high time burdens interpreting chest X-rays, with wide inter-reader variability in sensitivity (54.2%–80.7%) and specificity (84.9%–93.4%) across individual readers. This variability and workload contribute to workflow inefficiencies and inconsistent diagnostic quality.

Researchers at Mass General Brigham evaluated AIRead (Soombit.ai), a multimodal generative AI system that produces preliminary CXR reports. Five radiologists reviewed 758 chest X-rays with and without AI-generated preliminary reports in a retrospective study, assessing both efficiency and diagnostic accuracy.

AI-assisted reading reduced average CXR interpretation time by 42% (19.8 seconds vs. 34.2 seconds). Sensitivity for pleural lesions improved by 9.7% and widened mediastinum by 6.5%, while inter-reader variability narrowed significantly. Sensitivity for consolidation rose 17.9% and lung opacity 22.5%, though lung nodule sensitivity was slightly lower with AI assistance.

• Generative AI preliminary reports can substantially reduce radiologist reading time while improving detection of certain findings, but performance is not uniformly better across all pathology types.
• AI-assisted reporting homogenizes diagnostic performance across radiologists, reducing variability and potentially raising the floor of quality.
• Lung nodule sensitivity was lower with AI assistance (80% vs. 86.7%), highlighting the need to evaluate AI tools on a finding-by-finding basis rather than assuming universal improvement.