Diagnostic Accuracy of Artificial Intelligence vs. Oncologist Interpretation in Digital Cervicography for Abnormal...

Objective: We compared the diagnostic performance of artificial intelligence (AI) with that of a gynecologic oncologist during digital cervicography. Methods: Women with abnormal cytology who underwent cervicography between January 2019 and December 2023 were included. A gynecologic oncologist interpreted the digital cervicography and the results were compared with those of the AI system. Diagnostic performances were assessed using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and diagnostic accuracy for low-grade squamous intraepithelial lesions (LSILs) and high-grade squamous intraepithelial lesions (HSILs)/cancer. Cohen's kappa quantified agreement. Results: This study included 449 women (mean age, 41.0 years). A Cohen's kappa of 0.511 (p < 0.0001) indicated moderate agreement between the oncologist and AI. Among 226 cases of HSILs/cancer, the oncologist's sensitivity was 62.8%, compared to 47.8% for AI, with similar specificity (81.2% vs. 83.5%). The oncologist's PPV and NPV were 85.0% and 56.3%, respectively, whereas AI's were 83.1% and 48.5%, respectively. For LSILs/HSILs/cancer (n = 283), the oncologist achieved 98.2% sensitivity and 44.7% specificity, compared to AI's 93.3% sensitivity and 46.1% specificity. Both had a similar PPV (86.9% vs. 86.6%); however, the oncologist's NPV (87.2%) exceeded AI's 64.8%. Diagnostic accuracy for LSILs/HSILs/cancer was 86.9% for the oncologist and 82.3% for AI, whereas for HSILs/cancer, it was 69.6% and 61.0%, respectively. Conclusions: Moderate agreement was observed between the oncologist and AI. Although AI demonstrated similar performance in diagnosing cervical lesions, the oncologist achieved higher diagnostic accuracy. AI is a complementary tool and future research should refine AI algorithms to align with clinical performance.