Artificial intelligence and high-resolution anoscopy: automatic identification of anal squamous cell carcinoma precursors using a convolutional neural network

Background High-resolution anoscopy (HRA) is the gold standard for detecting anal squamous cell cancer (ASCC) precursors. Although it is superior to other diagnostic methods, particularly cytology, the visual identification of areas suspected of having high-grade squamous intraepithelial lesions remains difficult. Convolutional neural networks (CNNs) have shown great potential for assessing endoscopic images. The aim of the present study was to develop a CNN-based system for automatic detection and differentiation of HSIL versus LSIL in HRA images. Methods A CNN was developed based on 78 HRA exams from a total of 71 patients who underwent HRA at a single high-volume center (GH Paris Saint-Joseph, Paris, France) between January 2021 and January 2022. A total of 5026 images were included, 1517 images containing HSIL and 3509 LSIL. A training dataset comprising 90% of the total pool of images was defined for the development of the network. The performance of the CNN was evaluated using an independent testing dataset comprising the remaining 10%. The sensitivity, specificity, accuracy, positive and negative predictive values, and area under the curve (AUC) were calculated. Results The algorithm was optimized for the automatic detection of HSIL and its differentiation from LSIL. Our model had an overall accuracy of 90.3%. The CNN had sensitivity, specificity, positive and negative predictive values of 91.4%, 89.7%, 80.9%, and 95.6%, respectively. The area under the curve was 0.97. Conclusions The CNN architecture for application to HRA accurately detected precursors of squamous anal cancer. Further development and implementation of these tools in clinical practice may significantly modify the management of these patients.

Automated summary

A CNN-based system was developed to automatically detect and differentiate HSIL versus LSIL in HRA images. The system achieved an overall accuracy of 90.3%, with sensitivity, specificity, positive and negative predictive values of 91.4%, 89.7%, 80.9%, and 95.6%, respectively. The area under the curve was 0.97.