Domain-specific AI outperforms general-purpose LLM in chest X-ray report consistency

Domain-specific AI (M4CXR) showed 55.8% complete concordance vs 19.8% for ChatGPT-4o on 500 CXRs, and cut reporting time from 179s to 16s (P<.001). Retrospective single-center; findings suggest specialized AI improves consistency.

• Retrospective study of 500 chest radiographs from a single tertiary center, reports evaluated by 4 board-certified radiologists.
• M4CXR-assisted reports had no significant RADPEER discrepancy vs unaided interpretation, with good agreement (ICC 0.70, weighted kappa 0.65).
• Single-center, retrospective design without external validation; generalizability unproven.

RadPigeon summaries are original and for information only. They are not clinical advice.