Commercial chest X-ray AI tools made radiologists faster but not more accurate in a prospective real-world study from Germany, and for some findings they made reads worse.
Researchers at the Technical University of Munich’s Rechts der Isar hospital tested four commercially available AI solutions in a crossover design with five radiology residents reviewing 1,200 consecutive patients. The results, published online Aug 21 in Academic Radiology, showed AI assistance did not improve diagnostic accuracy overall.
For pleural effusions and pulmonary nodules, accuracy fell in several reader-tool pairings because the systems produced more false positives. Workflow did improve: three of five residents cut interpretation time by a median of 6 to 17 seconds per case, and four of five reported higher diagnostic confidence. AI support also reduced senior consultations in some pairings and, in one case, avoided escalation to CT.
The authors warn that unchanged accuracy combined with higher confidence is a recipe for automation bias. They concluded that AI deployment requires careful local adaptation, and that clinical trials are still needed to show patient outcomes improve.
The findings land as hospitals adopt radiology AI at record pace, often before evidence of real-world benefit exists. The study suggests buying decisions should weigh not just speed but whether a tool shifts diagnostic thresholds in ways clinicians can audit.
