Assessment of Efficacy and Accuracy of Cervical Cytology Screening With Artificial Intelligence Assistive System

Abstract

The role of artificial intelligence (AI) in pathology offers many exciting new possibilities for improving patient care. This study contributes to this development by identifying the viability of the AICyte assistive system for cervical screening, and investigating the utility of the system in assisting with workflow and diagnostic capability. In this study, a novel scanner was developed using a Ruiqian WSI-2400, trademarked AICyte assistive system, to create an AI-generated gallery of the most diagnostically relevant images, objects of interest (OOI), and provide categorical assessment, according to Bethesda category, for cervical ThinPrep Pap slides. For validation purposes, 2 pathologists reviewed OOIs from 32,451 cases of ThinPrep Paps independently, and their interpretations were correlated with the original ThinPrep interpretations (OTPI). The analysis was focused on the comparison of reporting rates, correlation between cytological results and histologic follow-up findings, and the assessment of independent AICyte screening utility. Pathologists using the AICyte system had a mean reading time of 55.14 seconds for the first 3000 cases trending down to 12.90 seconds in the last 6000 cases. Overall average reading time was 22.23 seconds per case compared with a manual reading time approximation of 180 seconds. Usage of AICyte compared with OTPI had similar sensitivity (97.89% vs 97.89%) and a statistically significant increase in specificity (16.19% vs 6.77%) for the detection of cervical intraepithelial neoplsia 2 and above lesions. When AICyte was run alone at a 50% negative cutoff value, it was able to read slides with a sensitivity of 99.30% and a specificity of 9.87%. When AICyte was run independently at this cutoff value, no sole case of high-grade squamous intraepithelial lesions/squamous cell carcinoma squamous lesion was missed. AICyte can provide a potential tool to help pathologists in both diagnostic capability and efficiency, which remained reliable compared with the baseline standard. Also unique for AICyte is the development of a negative cutoff value for which AICyte can categorize cases as “not needed for review” to triage cases and lower pathologist workload. This is the largest case number study that pathologists reviewed OOI with an AI-assistive system. The study demonstrates that AI-assistive system can be broadly applied for cervical cancer screening.