Automated acquisition of explainable knowledge from unannotated colposcopic images for predicting the natural course of CIN2.

OBJECTIVE

Cervical intraepithelial neoplasia grade 2 (CIN2) represents a critical turning point in cervical cancer progression. While its natural regression rate reaches 50%-57%, CIN2 can also progress to higher-grade lesions or cervical cancer. Accurately predicting CIN2 outcomes is essential for individualized management and avoiding overtreatment. The present study aimed to develop a deep learning-based prognostic model using colposcopic images to predict CIN2 outcomes: regression, persistence, and progression, and to provide clinically meaningful risk stratification.

METHODS

Colposcopic images from 212 patients diagnosed with CIN2 were retrospectively collected, including unstained, acetic acid-stained, and Lugol's iodine-stained images. Region of interest extraction and cluster analysis were employed for feature quantification. Four machine learning models (logistic regression, XGBoost, Random Forest and Extra Trees) were constructed to predict outcomes. Performance was evaluated using confusion matrices, receiver operating characteristic (ROC) curves, precision-recall curves, and decision curve analysis.

RESULTS

Among 212 patients, 52.4% regressed, 31.1% persisted, and 16.5% progressed. Automated pipeline generated 42-45 informative image clusters per staining type. For unstained images, logistic regression performed best (macro-AUC 0.884); for acetic acid-stained images, XGBoost achieved the highest accuracy (macro-AUC 0.933) with 70.0% sensitivity for progression; for iodine-stained images, Extra Trees showed the highest regression sensitivity (93.9%). Decision curve analysis confirmed clinical utility, and SHAP analysis highlighted prognostic features.

CONCLUSION

Machine learning models based on colposcopic images can effectively predict CIN2 prognosis, offering an objective tool for risk assessment and individualized clinical management. Different staining modalities provide complementary information, aiding in balancing progression risk reduction with overtreatment avoidance.