Assessment of the efficiency and accuracy of an artificial intelligence assistive system in the diagnosis of Pap cervical atypical glandular cell cytology
Abstract
Background
A diagnosis of atypical glandular cells (AGC) on Papanicolaou (Pap) slides is rare but has clinically significant findings associated with high‐risk cervical and endometrial lesions. The authors evaluated the efficiency and diagnostic performance of an artificial intelligence (AI)‐assisted platform (Riuqian WSI‐2400; with the registered trademark AICyte) in identifying AGCs on Pap slides.
Methods
A retrospective analysis of 485 Pap cases was conducted, including 185 cases with AGCs, 50 cases with high‐grade squamous intraepithelial lesions, 50 cases with low‐grade squamous intraepithelial lesions, and 200 negative cases; of these, 264 cases had histologic correlations. An experienced cytopathologist reviewed all slides using conventional microscopy and AICyte. Then, the same cases were evaluated by two other pathologists using the AICyte system.
Results
The initial study demonstrated a kappa value of 0.744, which indicated strong agreement of the Pap interpretation from the same pathologist between using microscopy and AICyte methods, whereas the average interpretation time was significantly reduced with AICyte (137 vs. 44 seconds). Diagnostic consensus among three pathologists using the AICyte system was strong, with a Kendall W coefficient of 0.802. The AICyte‐pathologist consensus reached an exact match with original interpretations in 95.1% of cases. AICyte‐assisted interpretations demonstrated improved specificity and diagnostic accuracy for glandular lesions compared with original interpretations while maintaining 100% sensitivity and negative predictive value.
Conclusions
To the authors' knowledge, this is the first study focusing on assessment of AGCs on an artificial intelligence system. The findings demonstrated that the AICyte system offers substantial improvements in efficiency and diagnostic consistency for the interpretation of AGCs and significantly reduces slide reading time. These results support the potential of AI to augment performance, especially in resource‐limited settings or high‐volume screening environments.