Meta-analysis of Artificial Intelligence-Assisted Pathology for the Detection of Early Cervical Cancer

Abstract

AbstractThe objective of this study is to evaluate the accuracy of AI in the diagnosis of early cervical cancer using a systematic evaluation/meta-analysis approach and a comprehensive search of published literature. A comprehensive computer search of foreign language databases such as PubMed/MEDLINE, Embase, Cochrane Library, and IEEE; and Chinese databases such as China Knowledge Network, Wan fang Data Knowledge Platform, and Wipu.com (VIP) was conducted to retrieve reports on diagnostic accuracy of AI in early cervical cancer included between 1946 and December 2022. The literature was screened according to inclusion and exclusion criteria, and the quality of the included literature was evaluated using the QUADAS-2 quality evaluation chart. 2 × 2 diagnostic data in text were extracted and complete data were calculated using Review Manager 5.3. Heterogeneity between studies was analyzed using Stata SE 15.0 software with Meta Di Sc 1.4 and causes of heterogeneity were sought. A total of 42 data sets were included in the study of AI for the identification of benign and malignant cervical vitreous nodules, with a combined Sen value of 0.90; a combined Spe value of 0.90; a combined + LR value of 9.0; a combined −LR combined value was 0.11; DOR combined value was 83; and AUC was 0.96. The Fagan plot suggested a 50% pre-test probability and a 90% post-test probability of confirming diagnosis when the AI model diagnosed a glassy nodule positively, and a 10% probability of misdiagnosing the nodule when the result was negative. A total of 34 data sets were included in the study to determine benignity and malignancy of solid cervical nodules by AI, showing a combined Sen value of 0.92; a combined Spe value of 0.93; a combined + LR value of 13.37; a combined −LR combined value of 0.08; DOR combined value of 164; AUC of 0.97. The Fagan plot suggested a 50% pre-test probability and a 93% post-test probability of confirming the diagnosis of a solid cervical nodule when the AI model was positive, and an 8% probability of misdiagnosing the nodule when the result was negative. The results of likelihood ratio dot plots suggest that the use of an AI model for cervical detection in the clinical setting has a good exclusionary diagnostic power. Summing up the accuracy and specificity of the A1 model for diagnosis of early cervical cancer, accuracy for diagnosis of solid cervical nodules (0.90) > diagnosis of cervical nodules (0.92), and specificity for diagnosis of solid cervical nodules (0.90) > diagnosis of cervical nodules (0.93). The AI model is highly accurate in diagnosing early cervical cancer and has high clinical diagnostic value. The accuracy of the AI model in diagnosing solid nodules in the cervical region was higher than diagnosing ground glass nodules in the cervical region. The labeling method, image pre-processing method, and feature learning method affected the accuracy of the AI model in diagnosing early cervical cancer, while the choice of learning image library and validation database did not usually affect the accuracy of the model.