Machine Learning-Based Cervical Cancer Screening Using Cervigrams During Visual Inspection With Acetic Acid: a Systematic Review.

Abstract Background: The World Health Organization (WHO) recommendations for promoting effective management of cervical cancer screening in low- and medium-income countries (LMIC) include human papillomavirus (HPV) testing as primary screening followed by visual inspection with acetic acid (VIA) and, if required, treatment. The application of acetic acid induces a transient whitening effect which appears and disappears differently in precancerous lesions and cancer than in benign conditions. However, this assessment by human observers is generally subjective and accuracy is limited. This study presents a systematic review of the automated algorithms for cervical (pre)cancer screening based on images taken during VIA with the objective of assessing their potential as screening tool.Methods: We performed a systematic literature search in PubMed, Google Scholar and Scopus. The selected studies introduce automated algorithms for the classification of cervical intraepithelial neoplasia grade 2 or higher (CIN2+) with respect to benign conditions, based only on images taken during VIA. We included studies that use, as gold standard, histopathology for CIN2+ cases and, histopathology or normal cytology and colposcopy for benign conditions. The selected studies were analysed in terms of specificity and sensitivity. From each study, the algorithm with the highest accuracy was further studied considering key features such as type of algorithms, acquisition devices, the number of images used per patient, or its performance in comparison to the experts’ classification. The quality and risk of the studies was assessed following the QUADAS-2 guidelines.Results: Of the 1519 studies identified, nine met the inclusion criteria. The algorithms with the highest accuracy from each study reported a sensitivity and specificity values ranging from 0.60 to 0.93 and 0.67 to 0.95, respectively. Conclusion: Machine learning-based cervical cancer screening algorithms have the potential to become a key tool for cervical cancer screening in countries that suffer from a lack of healthcare infrastructure and personnel. Nevertheless, the selected studies assess their algorithms using small datasets made of highly selected images without reflecting real screened populations. Large-scale and real conditions testing is required to assess the potential of these algorithms as the future of cervical cancer screening.Systematic review registration: PROSPERO CRD42021270745