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

Abstract

OBJECTIVE This study evaluates the performance of artificial intelligence (AI) colposcopy in detecting cervical cancer and precancerous lesions in real-world scenarios within resource-limited areas. METHODS This is a cross-sectional study. Participants with positive human papilloma virus results or who were cytologic positive were referred for colposcopy, during which AI colposcopy was implemented. Biopsies were performed for positive findings suggested by either the colposcopist or the AI system. For the analysis, we calculated the sensitivity, specificity, positive predictive value, negative predictive value, and area under the curve for detecting cervical intraepithelial neoplasia (CIN) 2+ and CIN 3+. Histopathology was the gold standard for disease diagnosis. RESULTS A total of 825 women underwent colposcopy, with 99 (12.0%) diagnosed with CIN 2+ and 53 (6.4%) with CIN 3+. Positive findings were reported in 392 women (47.5%) under conventional colposcopy and 640 (77.6%) with AI colposcopy. The sensitivity for detecting CIN 2+ was significantly higher for AI colposcopy (96.0%) and AI-assisted colposcopy (100%) than for conventional colposcopy (85.9%, P=.026, P<.001, respectively). In postmenopausal women, the sensitivities of AI colposcopy (94.3%) and AI-assisted colposcopy (100%) surpassed that of conventional colposcopy (77.4%, P=.026, P<.001, respectively). Artificial intelligence-assisted colposcopy also significantly enhanced the sensitivity of junior colposcopists with less than 10 years of clinical experience, achieving 100% compared with 84.6% by conventional colposcopy (P=.001), and improved detection in women with a squamocolumnar junction that was not visible (100% vs 70.4%, P=.004). For CIN 3+, the sensitivity of AI-assisted colposcopy was superior to that of conventional colposcopy (100% vs 86.8%, P=.013). In postmenopausal women, the sensitivities of both AI colposcopy and AI-assisted colposcopy were 100%; however, the sensitivity of conventional colposcopy was 77.8% (P=.023). CONCLUSION Artificial intelligence-assisted colposcopy enhances sensitivity in detecting CIN 2+ and CIN 3+, particularly among postmenopausal women. Moreover, it improves the diagnostic performance of junior colposcopists and improves detection in women with a squamocolumnar junction that is not visible.