Artificial Intelligence for Anal Cancer Screening: First Multicentric Validation of an Interoperable System for Simultaneous Detection and Differentiation of Squamous Cell Precursor Lesions

Aims Anal cancer incidence continues to rise, underscoring the urgent need for more effective and scalable screening strategies. High resolution anoscopy (HRA) remains the cornerstone of early detection, yet achieving precise identification and confident differentiation of lesions remains challenging. Artificial intelligence has emerged as a powerful tool to enhance recognition and stratification of squamous intraepithelial lesions, including LSIL (low grade squamous intraepithelial lesion) and HSIL (high grade squamous intraepithelial lesion), with clear advancement along the technology readiness level scale (TRL, a measure of technological maturity). CADe systems (computer aided detection) improve lesion localization, while CADx systems (computer aided diagnosis) deliver real time differentiation that can directly shape clinical decisions. This study set out to develop an interoperable CADex system capable of simultaneously detecting and distinguishing LSIL and HSIL during HRA, aiming to elevate diagnostic precision, reduce subjectivity, and accelerate progress toward more consistent and effective anal cancer prevention [ 1 ] [ 2 ] [ 3 ] [ 4 ]. Methods This multicenter study analyzed 156,246 frames extracted from 308 HRA procedures performed in 279 patients across eight centers. Data were obtained using six different devices, including four colposcopes, one proctoscope, and one endoscope. Frames were annotated as HSIL or LSIL based on corresponding histopathology, which served as the ground truth. A YOLO-tiny object detection model was trained to perform simultaneous lesion detection and differentiation using bounding-box annotations. The dataset was randomly divided into training, validation, and test sets. Model performance was evaluated at two levels. Object detection metrics were mean average precision (mAP) and mAP at an Intersection over Union threshold of≥0.50 (mAP50). Classification metrics were recall (sensitivity), precision (PPV), and accuracy. Results Precision was 0.8332 for LSIL and 0.8165 for HSIL, with an overall precision of 0.8247. Recall was 0.9988 for LSIL and 0.9994 for HSIL, yielding an overall recall of 0.9991. Overall classification accuracy was 0.9432. For combined LSIL and HSIL detection, the model achieved a mAP of 0.7488 and mAP50 of 0.9738. Conclusions This study introduces the first interoperable CADex model capable of simultaneous detection and classification of anal canal precursor lesions, using an explainable bounding box approach. These findings indicate that CADex can provide a more robust pathway for integrating AI into anal cancer screening and mark a meaningful step forward in the TRL of AI applied to HRA. Its consistent performance across diverse imaging devices reinforces this potential and highlights interoperability as a prerequisite for broad clinical adoption. This interoperability also paves the way to secondary applications, such as opportunistic assessment of the anal canal during colonoscopy, where incidental visualization could ultimately complement and extend dedicated HRA based screening programs. Publication History Article published online: 05 June 2026 © 2026. European Society of Gastrointestinal Endoscopy. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany