Rationale: Artificial intelligence (AI)-assisted diagnosis imparts high accuracy to chest radiography (CXR) interpretation; however, its benefit for nonradiologist physicians in detecting lung lesions on CXR remains unclear. Objectives: To investigate whether AI assistance improves the diagnostic performance of physicians for CXR interpretation and affects their clinical decisions in clinical practice. Methods: We randomly allocated eligible patients who visited an outpatient clinic to the intervention (with AI-assisted interpretation) and control (without AI-assisted interpretation) groups. Lung lesions on CXR were recorded by seven nonradiologists with or without AI assistance. The reference standard for lung lesions was established by three radiologists. The primary and secondary endpoints were the physicians' diagnostic accuracy and clinical decision, respectively. Results: Between October 2020 and May 2021, 162 and 161 patients were assigned to the intervention and control groups, respectively. The area under the receiver operating characteristic curve was significantly larger in the intervention group than in the control group for the CXR level (0.840 [95% confidence interval (CI), 0.778-0.903] vs. 0.718 [95% CI, 0.640-0.796]; P = 0.017) and lung lesion level (0.800 [95% CI, 0.740-0.861] vs. 0.677 [95% CI, 0.605-0.750]; P = 0.011). The intervention group had higher sensitivity in terms of both CXR and lung lesion level and a lower false referral rate for the lung lesion level. AI-assisted CXR interpretation did not affect the physicians' clinical decisions. Conclusions: AI-assisted CXR interpretation improves the diagnostic performance of nonradiologist physicians in detecting abnormal lung findings. Clinical trial registered with Clinical Research Information Service of the Republic of Korea (KCT 0005466).
Keywords: AI; diagnosis; prospective study; radiography; thorax.