Artificial Intelligence for Automatic Detection and Differentiation of Anal HPV-related Dysplastic Lesions: An Interoperable Multicentric Study

Aims High Resolution Anoscopy (HRA) is used to detect and treat Anal Squamous Cell Carcinoma (ASCC) precursors. While Low-Grade Squamous Intraepithelial Lesions (LSIL) can be monitored, High-Grade Squamous Intraepithelial Lesions (HSIL) require treatment to prevent ASCC. This study aims to develop an interoperable Convolutional Neural Network (CNN) to differentiate HSIL from LSIL in HRA still frames

Methods A multicentric retrospective study (n=4 centers) included 66860 lesion-containing frames from 262 HRA procedures (n=4 devices). Frames were classified as HSIL or LSIL based on histopathological findings, including those from non-stained, stained (acetic acid or lugol), and post-manipulation phases of the HRA procedure. The dataset was split into 90% training/validation and 10% testing, with the testing set used to evaluate the CNN's sensitivity, specificity, positive and negative predictive value (PPV and NPV, respectively), accuracy and area under the ROC curve (AUC-ROC).

Results In the testing set, the model demonstrated 97.1% sensitivity, 99.1% specificity, 99.2% PPV, 97.0% NPV, and an overall accuracy of 98.1% in distinguishing HSIL from LSIL. The AUC-ROC were 1.00.

Conclusions Developing AI models for reliable real-world clinical use requires ensuring interoperability and dataset heterogeneity. This multicentric multidevice CNN addresses selection bias, achieving robust performance metrics and advancing its readiness for practical deployment.

Publication History

Article published online:
27 March 2025