High resolution anoscopy (HRA) is the gold standard for anal cancer screening, yet its interpretation is limited by suboptimal accuracy and marked interobserver variability. Artificial intelligence (AI) has the potential to improve diagnostic yield, but existing studies have focused on differentiating low-grade and high-grade squamous intraepithelial lesions (LSIL and HSIL) in still images. To date, no real-time clinical application has been reported.

We describe the first real-time deep learning application implemented during HRA. The system was evaluated in three patients undergoing anal cancer screening at a high-volume American center. To enable real-time lesion recognition, a YOLO-based object detection model was integrated into the live video stream, allowing continuous frame analysis. When a region displayed visual characteristics suggestive of HSIL, the algorithm generated an on-screen bounding box to highlight the suspected lesion in real time. 

Three patients undergoing HRA were evaluated with concurrent real-time AI analysis. Case 1: A 53-year-old HIV-negative male with prior LSIL had three lesions identified on HRA: two acetowhite, lugol-positive LSIL-type lesions and one acetowhite, lugol-negative flat lesion with a mosaic pattern suggestive of HSIL. The AI model generated a bounding box exclusively for the HSIL-suspected anterior lesion, which histopathology confirmed, while the remaining lesions proved LSIL. Case 2: A 38-year-old HIV-negative male with incidentally detected LSIL presented with two LSIL-compatible lesions. The AI system generated no HSIL predictions, and biopsies confirmed LSIL. Case 3: A 40-year-old HIV-positive male with previous HSIL showed no suspicious lesions on HRA, and the AI system generated no HSIL predictions. Across all procedures, the model detected only histologically confirmed HSIL and showed no activation when no lesions or only LSIL were present. 

This first demonstration of real time AI enhanced HRA shows that deep learning algorithms can deliver accurate predictions and perform consistently during live procedures, expanding their value far beyond still image analysis. This technology can reduce interobserver variability, accelerate skill acquisition, and support broader adoption among clinicians with varying experience, while strengthening diagnostic performance in expert settings. These advances signal a meaningful step toward improving anal cancer screening by enabling earlier detection and more timely intervention. Larger multicenter studies are now essential to validate these findings and confirm their generalizability.