P0524 - Deep Learning and High Resolution Anoscopy: First Trinary Interoperable Model for Detecting and Differentiating Clinically Relevant Anal Lesions

Award: ACG Presidential Poster Award

Miguel Martins, MD¹, Miguel Mascarenhas Saraiva, MD, PhD¹, Luis Barroso, MD², Thiago Manzione, MD, PhD³, Pedro Diaz Denoso, MD⁴, Amine Alan, MD⁵, Lucas Spindler, MD⁵, Ahsan Javed, MD, PhD⁶, João Afonso, MD¹, Tiago Ribeiro, MD¹, Pedro Cardoso, MD¹, Francisco Mendes, MD¹, Maria João Almeida, MD⁷, Joana Mota, MD¹, André Santos, MD⁸, António Costa, MD⁹, Nadia Fathallah, MD⁵, Sidney Nadal, MD¹⁰, Luciana La Rosa, MD¹¹, João Ferreira, PhD¹², Guilherme Macedo, MD, PhD¹, Vincent De Parades, MD, PhD⁵
1Centro Hospitalar Universitário São João, Porto, Porto, Portugal; 2Wake Forest University Health Sciences, Winston-Salem, NC; 3Instituto de Infectologia Emílio. Ribas, São Paulo, Sao Paulo, Brazil; 4Private Center of Surgery and Proctology, Buenos Aires, Buenos Aires, Argentina; 5Hospital Paris Saint Joseph, Paris, Ile-de-France, France; 6Royal Liverpool University Hospital, Liverpool, England, United Kingdom; 7Centro Hospitalar Universitário de São João, Porto, Porto, Portugal; 8Faculdade de Medicina Universidade do Porto, Porto, Porto, Portugal; 9Hospital Universitario Puerta de Hierro Majadahonda, Madrid, Madrid, Spain; 10Instituto Infectologia Emilio Ribas, São Paulo, Sao Paulo, Brazil; 11Centro Privado de Cirurgía e Coloproctología, Buenos Aires, Buenos Aires, Argentina; 12Faculdade de Engenharia da Universidade do Porto, Porto, Porto, Portugal

Introduction: High-Resolution Anoscopy (HRA) is the gold standard for anal cancer screening. Early detection of premalignant lesions, such as high-grade squamous intraepithelial lesions (HSIL), is essential for reducing anal cancer mortality. However, differentiating HSIL from other HPV-related lesions—such as low-grade squamous intraepithelial lesions (LSIL) and non-dysplastic mimickers (e.g., inflammation)—remains suboptimal, even among experienced clinicians. This study aimed to develop the first Artificial Intelligence (AI) model capable of detecting and differentiating anal canal lesions across these three categories.

Methods: A multicenter, interoperable deep learning model was developed across five clinical sites using a YOLOv11-based object detection architecture using bounding boxes. The model was trained on 191,951 frames extracted from 107 High-Resolution Anoscopy (HRA) procedures. Each frame was annotated based on corresponding histopathological analysis. The dataset was divided into training, validation, and testing subsets. Lesion detection was defined by accurate placement of bounding boxes. The model’s recall (sensitivity), precision (positive predictive value), and accuracy were calculated as weighted averages across five confidence thresholds (0.45, 0.47, 0.50, 0.52, 0.55).

Results: Overall, the accuracy of the model was 95.2% (94.7- 95.7). For HSIL, recall and precision were 97.9% (97.6–98.2) and 93.9% (92.8–95.1), respectively. For LSIL, recall and precision were 98.9% (98.8–99.0) and 98.5% (98.3–98.7), respectively. For non-dysplastic lesions, recall and precision were 98.7% (98.5–98.8) and 96.0% (94.9–97.2), respectively.

Discussion: This pioneering trinary classification model, capable of accurately distinguishing HSIL, LSIL, and non-dysplastic mimickers, marks a significant breakthrough in the path toward clinical integration of AI-assisted HRA. By enabling precise and consistent identification of clinically relevant anal lesions, AI may enhance diagnostic accuracy, reduce interobserver variability, and pave the way for more targeted, precision-based screening and management strategies.


Figure: Illustrative examples of lesion detection and classification, supported by explainable Al through bounding box overlays.

Disclosures:
Miguel Martins indicated no relevant financial relationships.
Miguel Mascarenhas Saraiva indicated no relevant financial relationships.
Luis Barroso indicated no relevant financial relationships.
Thiago Manzione indicated no relevant financial relationships.
Pedro Diaz Denoso indicated no relevant financial relationships.
Amine Alan indicated no relevant financial relationships.
Lucas Spindler indicated no relevant financial relationships.
Ahsan Javed indicated no relevant financial relationships.
João Afonso indicated no relevant financial relationships.
Tiago Ribeiro indicated no relevant financial relationships.
Pedro Cardoso indicated no relevant financial relationships.
Francisco Mendes indicated no relevant financial relationships.
Maria João Almeida indicated no relevant financial relationships.
Joana Mota indicated no relevant financial relationships.
André Santos indicated no relevant financial relationships.
António Costa indicated no relevant financial relationships.
Nadia Fathallah indicated no relevant financial relationships.
Sidney Nadal indicated no relevant financial relationships.
Luciana La Rosa indicated no relevant financial relationships.
João Ferreira indicated no relevant financial relationships.
Guilherme Macedo indicated no relevant financial relationships.
Vincent De Parades indicated no relevant financial relationships.

Miguel Martins, MD¹, Miguel Mascarenhas Saraiva, MD, PhD¹, Luis Barroso, MD², Thiago Manzione, MD, PhD³, Pedro Diaz Denoso, MD⁴, Amine Alan, MD⁵, Lucas Spindler, MD⁵, Ahsan Javed, MD, PhD⁶, João Afonso, MD¹, Tiago Ribeiro, MD¹, Pedro Cardoso, MD¹, Francisco Mendes, MD¹, Maria João Almeida, MD⁷, Joana Mota, MD¹, André Santos, MD⁸, António Costa, MD⁹, Nadia Fathallah, MD⁵, Sidney Nadal, MD¹⁰, Luciana La Rosa, MD¹¹, João Ferreira, PhD¹², Guilherme Macedo, MD, PhD¹, Vincent De Parades, MD, PhD⁵. P0524 - Deep Learning and High Resolution Anoscopy: First Trinary Interoperable Model for Detecting and Differentiating Clinically Relevant Anal Lesions, ACG 2025 Annual Scientific Meeting Abstracts. Phoenix, AZ: American College of Gastroenterology.