Diagnostic efficiency of artificial intelligence for pulmonary nodules based on CT scans

To explore the accuracy of artificial intelligence (AI) for the diagnosis of pulmonary nodules (PNs) on computerized tomography (CT) scans.

In this study, 360 PNs (251 malignant nodules and 109 benign nodules) were retrospectively analyzed in 309 participants examined for PNs, and CT images were reviewed both by radiologists and using AI technology. With postoperative pathologic results as the gold standard, the accuracy, misdiagnosis, missed diagnosis, and true negative rates of CT results (human and AI) were calculated by using 2×2 crosstabs. Data confirmed to be normally distributed by the Shapiro-Wilk test were compared by the independent sample t-test, and the reading time of AI and human radiologists was compared.

1) The accuracy rate of AI for diagnosing PNs was 81.94% (295/360), the missed diagnosis rate was 15.14% (38/251), the misdiagnosis rate was 24.77% (27/109), and the true negative rate was 75.23% (82/109). 2) The accuracy, missed diagnosis, misdiagnosis, and true negative rates of human radiologists in the diagnosis of PNs were 83.06% (299/360), 22.31% (56/251), 4.59% (5/109), and 95.41% (104/109), respectively. 3) The accuracy and missed diagnosis rates were comparable between AI and radiologists, but AI had a significantly higher misdiagnosis rate and a markedly lower true negative rate. 4) The image reading time required for AI (195.4±65.2 s) was statistically shorter than that required for manual examination (581.1±116.8 s). 5) The accuracy of AI for detecting low, moderately, and highly malignant PNs was 13.64% (9/66), 25.33% (19/75), and 48.61% (35/72), respectively.

AI demonstrates favorable accuracy for CT diagnosis of lung cancer and requires a shorter time for film reading. However, its diagnostic efficiency in identifying low- and moderate-grade PNs is relatively low, indicating a need for expansion of machine learning samples to improve its accuracy in identifying lower grade cancer nodules.