Abstract 1136P
Background
Clinical application of artificial intelligence in cervical cancer cytology screening is still limited, so a Cervical Cancer Artificial Intelligence Screening System (CAISS) was aimed to established in this study.
Methods
The study consisted of a multicenter population-based study and randomized controlled trial (RCT) performed in China, enrolling 16,056 individuals aged over 18 who had liquid-based cytology pap test with eligible cervical cytological WSIs. 11,468 individuals’ WSIs from Sun Yat-sen Memorial Hospital (SYSMH) were randomly assigned (4:1) into training and internal validation dataset to train CAISS, and validated in SYSMH internal, Guangzhou Women and Children Medical Center (GWCMC), The Third Affiliated Hospital of Guangzhou Medical University (TAHGMU) validation datasets, and SYSMH prospective validation dataset. The RCT was conducted to compare the performance between CAISS, cytotechnicians, and CAISS-assisted. The sensitivity, specificity, accuracy, and AUC were used to assess CAISS's performance.
Results
The sensitivity of CAISS in identifying patients with abnormal cytology grades was 0.906, 0.902, 0.918 in the SYSMH internal, GWCMC external and TAHGMU external validation datasets, respectively. In prospective validation dataset, the CAISS showed similar sensitivity (0.946 vs 0.909, p= 0.304) and AUC (0.947 vs 0.948, p=0.952) to cytotechnician, and CAISS-assisted achieved better sensitivity, than cytotechnician alone (p= 0.024; p= 0.0006). In randomized controlled dataset, the specificity and accuracy of CAISS-assisted were significantly outperformed CAISS (0.989 vs 0.854, p< 0.001; 0.990 vs 0.861, p< 0.001), and no statistical difference in sensitivity between CAISS and cytotechnician (p= 0.552).
Conclusions
In this study, CAISS achieved high sensitivity for diagnosing cervical cytology grade that rivals cytotechnicians’ performance, and help cytotechnicians improve diagnostic sensitivity and accuracy to a higher level, which could improve the effectiveness of cervical cancer screening.
Clinical trial identification
NCT04551287.
Editorial acknowledgement
Legal entity responsible for the study
Yufang Yu.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.