Automated detection and classification of cervical and anal squamous cancer precursors using deep learning and multidevice colposcopy

Human papillomavirus (HPV) infection presents neoplastic risks in both cervix and anus. High-resolution colposcopy/anoscopy is crucial for assessing these regions but has suboptimal accuracy. This study aims to develop a Convolutional Neural Network (CNN) to identify and differentiate low-grade (LSIL) and high grade (HSIL) squamous intraepithelial lesions, in the cervix and anus. A retrospective multicenter study was conducted to develop a CNN using 320 colposcopy and anoscopy examinations, from 3 device types. Dataset included 88,073 frames, categorized as LSIL or HSIL based on pathological analysis. The data was split into training/validation (90%, n = 79,265, including a threefold cross-validation) and test sets (10%, n = 8808). Diagnostic metrics including sensitivity, specificity, accuracy, positive and negative predictive values (PPV and NPV, respectively) and an area under the receiving operating and the precision-recall curves (AUC-ROC and AUC-PR) were calculated. During training/validation phase, the model achieved an average sensitivity for HSIL of 98.1% (IC95% 97.6-98.5%), specificity of 97.4% (IC95% 96.0-98.8%), PPV of 97.2% (IC95% 95.8-98.7%), NPV of 98.2% (IC95% 97.7-98.6%), and accuracy of 97.7% (IC95% 97.2-98.6%). The mean AUC-ROC and AUC-PR were both 0.98 ± 0.01. In the testing phase, performance metrics for HSIL were: sensitivity 99.0%, specificity 97.8%, PPV 97.6%, NPV 99.0%, and accuracy 98.3%. HPV infection impacts both cervical and anal region. This study developed a pioneering CNN to differentiate HSIL and LSIL in HPV-related dysplastic lesions, during cervical and anal examinations. This model achieved promising results, suggesting its potential to improve detection accuracy and cost-effectiveness in clinical practice.