Effectiveness of Artificial Intelligence-Assisted Colposcopy in a Resource-Limited Population

The average increase in the incidence of cervical cancer in China has been estimated at 7.3% annually over the last 20 years. The most common cause of cervical cancer is human papillomavirus (HPV) infection, and thus screening for HPV is a cornerstone of cervical cancer detection and prevention. Standard treatment for individuals with HPV includes colposcopy and biopsy, if indicated, but the accuracy of colposcopy for diagnosis is dependent on the physician; a goal of advances in this field is to enhance the diagnostic accuracy of colposcopy to prevent misdiagnosis. A potential solution to this problem is the use of artificial intelligence (AI), which has been shown to enhance the accuracy, efficiency, and accessibility of cervical cancer screening in previous studies. Using AI in colposcopy could reduce challenges related to population differences, but previous AI algorithms have focused only on small training sets, which are limited in their application. This study was designed to assess the applicability of AI colposcopy in cervical cancer screening, comparing conventional, AI, and AI-assisted colposcopy in premenopausal and postmenopausal women. This was a cross-sectional, multicenter observational study of patients in China who were recruited for cervical cancer screening between March 2023 and May 2024. The inclusion criteria were an anatomically intact cervix, understanding of the screening procedure, and written informed consent. The final analysis included 825 patients, with 583 also undergoing a biopsy. Median age was 53, and a total of 451 patients were postmenopausal. Of 825 women screened, 675 were positive for some form of HPV. Of these, 299 tested positive for HPV 16 or 18, the types most highly correlated with cervical cancer. Conventional colposcopy identified 392 positive test results, and AI colposcopy identified 640 positive test results. The agreement between the 2 methods was 53%. The sensitivity of conventional colposcopy was 85.9% (95% CI: 77.1%-91.8%) and the specificity was 57.7% (95% CI: 54%-61.3%). Sensitivity and specificity of AI colposcopy were 96% (95% CI: 89.4%-98.7%) and 24.9% (95% CI: 21.9%-28.3%). AI-assisted colposcopy had a sensitivity of 100% (95% CI: 95.3%-100%) and a specificity of 15.8% (95% CI: 13.3%-18.7%). Results were similar in postmenopausal women, with AI and AI-assisted colposcopy having significantly greater sensitivity compared with conventional colposcopy (P=0.026, P<0.001, respectively). Sensitivity and specificity varied slightly depending on menopausal status, type of HPV, and visibility of the squamocolumnar junction, but results were similar across groups, showing an improved result for AI and AI-assisted procedures. These results indicate that AI colposcopy has advantages compared with conventional colposcopy, including increased sensitivity. AI-assisted colposcopy showed significantly improved diagnostic utility compared with conventional methods, especially in cases where diagnosis is more difficult, such as when the squamocolumnar junction is not visible. This represents opportunities for low-resource areas in particular, and for areas where skilled colposcopists are rare or unavailable. These findings are consistent with previous literature. Future research should focus on improving the accuracy of targeted biopsies as well as validating these results in larger samples of HPV subtypes.