# Deep Learning and High-Resolution Anoscopy: Development of an Interoperable Algorithm for the Detection and Differentiation of Anal Squamous Cell Carcinoma Precursors—A Multicentric Study

## TL;DR

This paper introduces an AI system that accurately detects early signs of anal cancer using high-resolution anoscopy images, with high accuracy across different imaging methods.

## Contribution

A deep learning algorithm was developed and validated for detecting anal cancer precursors using HRA images with high accuracy.

## Key findings

- The AI system achieved 94.6% accuracy in distinguishing high-grade from low-grade lesions in HRA images.
- The algorithm performed well across different staining methods and after treatment, with accuracies up to 99.3%.
- The system is compatible with both conventional and digital HRA systems, showing strong interoperability.

## Abstract

High-resolution anoscopy (HRA) is crucial for spotting and treating early signs of anal cancer. The researchers created an artificial intelligence (AI) system to analyze HRA images and identify high-grade and low-grade lesions accurately. They trained a computer program with thousands of images, achieving a remarkable accuracy of 94.6%. The AI system proved effective across different examination methods, such as using acetic acid or lugol iodine, and even after treatment. This advancement could improve the early detection of anal cancer precursors, potentially saving lives.

High-resolution anoscopy (HRA) plays a central role in the detection and treatment of precursors of anal squamous cell carcinoma (ASCC). Artificial intelligence (AI) algorithms have shown high levels of efficiency in detecting and differentiating HSIL from low-grade squamous intraepithelial lesions (LSIL) in HRA images. Our aim was to develop a deep learning system for the automatic detection and differentiation of HSIL versus LSIL using HRA images from both conventional and digital proctoscopes. A convolutional neural network (CNN) was developed based on 151 HRA exams performed at two volume centers using conventional and digital HRA systems. A total of 57,822 images were included, 28,874 images containing HSIL and 28,948 LSIL. Partial subanalyses were performed to evaluate the performance of the CNN in the subset of images acetic acid and lugol iodine staining and after treatment of the anal canal. The overall accuracy of the CNN in distinguishing HSIL from LSIL during the testing stage was 94.6%. The algorithm had an overall sensitivity and specificity of 93.6% and 95.7%, respectively (AUC 0.97). For staining with acetic acid, HSIL was differentiated from LSIL with an overall accuracy of 96.4%, while for lugol and after therapeutic manipulation, these values were 96.6% and 99.3%, respectively. The introduction of AI algorithms to HRA may enhance the early diagnosis of ASCC precursors, and this system was shown to perform adequately across conventional and digital HRA interfaces.