Deep learning models for cervical cancer subtyping using whole slide images.

Objective: This study aims to develop and evaluate an artificial intelligence-based model for cervical cancer subtyping using whole-slide images (WSI), incorporating both patch-level and WSI-level analyses to enhance diagnostic accuracy. Methods: A total of 438 whole slide images were retrieved from three databases, one public dataset for model training and two independent private datasets for evaluation of generalization. It is comprised of two consecutive stages: a patch-level prediction and a WSI-level prediction. Patch-level predictions were performed using the four convolutional neural networks model, while WSI-level predictions were based on five machine learning algorithms with three different aggregation methods. We compared the models in terms of discrimination (accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUROC)) and calibration. Results: At the patch level, the Inception-v3 model achieved an AUROC of 0.960 (95% confidence interval(95%CI): 0.943, 0.978) in private dataset one and an AUROC of 0.942 (95% CI: 0.929, 0.956) in private dataset one. For WSI-level predictions, the support vector machine algorithm based on Term Frequency-Inverse Document Frequency (TF-IDF) features performed the best, with an AUROC of 0.964 (95% CI: 0.916, 0.996) in private dataset one and 0.947 (95% CI: 0.879, 0.996) in private dataset two. The decision curve analysis and calibration curves further validated the clinical potential of the model. Discussion: This study demonstrates the potential of using AI models for cervical cancer subtyping, with strong generalization across multiple datasets and clinical settings.