P4564 - Artificial Intelligence in Endoanal Ultrasound: A Multicentric Study on a Multilesion Model for Benign Anorectal Lesion Detection and Differentiation

Miguel Mascarenhas Saraiva, MD, PhD¹, Maria João Almeida, MD², Miguel Martins, MD², Francisco Mendes, MD¹, Joana Mota, MD¹, Pedro Cardoso, MD¹, Tiago Ribeiro, MD¹, João Afonso, MD¹, Patrícia Andrade, ², António Miguel M P D. Martins Pinto da Costa, MD³, Pedro Diaz Denoso, MD⁴, Jorge Arias, ⁴, João Ferreira, PhD⁵, Bruno Mendes, ⁴, Matheus Ferreira de Carvalho, MD⁶, Guilherme Macedo, MD, PhD¹, Castro Poças, ⁷
1Centro Hospitalar Universitário São João, Porto, Porto, Portugal; 2Centro Hospitalar Universitário de São João, Porto, Porto, Portugal; 3Hospital Universitario Puerta de Hierro-Majadahonda, Madrid, Madrid, Spain; 4Private Center of Surgery and Proctology, Buenos Aires, Buenos Aires, Argentina; 5Faculdade de Engenharia da Universidade do Porto, Porto, Porto, Portugal; 6Hospital das Clínicas da Faculdade de Medicina da Universidade de Sao Paulo, São Paulo, Sao Paulo, Brazil; 7Faculty of Engineering of the University of Porto, Porto, Porto, Portugal

Introduction:

Endoanal Ultrasound (EAUS) is a key diagnostic tool for anorectal diseases, valued for its simplicity and tolerability. However, its interpretation is challenged by significant interobserver variability, a steep learning curve, and limited accessibility. The application of artificial intelligence (AI) to enhance EAUS remains in its early stages. The aim of this study was to develop and validate an interoperable AI model capable of differentiating multiple benign anorectal lesions using EAUS.

Methods:

An EfficientNet-based AI model was developed using 4722 EAUS frames, from 105 procedures performed across three centers using two different ultrasound devices. Two experts independently classified frames into 5 categories: external anal sphincter (EAS) laceration, internal anal sphincter (IAS) laceration, anal fissure (AF), intersphincteric fistula (ISF), and transsphincteric fistula (TSF). Frames with concordant classifications established the ground truth. The dataset was split into training/validation (90%, with 3-fold cross-validation) and test (10%) sets. The model with the best F1-score was tested Performance metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score.

Results:

Training/validation mean (±SD) accuracies were: EAS 0.85 (±0.03), IAS 0.85 (±0.003), AF 0.98 (±0.02), ISF 0.99 (±0.01), TSF 0.99 (±0.01). On the test set, respective accuracies were: EAS 0.88, IAS 0.88, AF 0.99, ISF 1.00, TSF 0.99.

Discussion:

This pioneering AI model demonstrates high accuracy in differentiating five benign anorectal lesions using EAUS. As the first multilesion solution in this area, it represents a significant step toward standardizing EAUS interpretation and expanding access to expert-level diagnostic support. With further prospective validation, this technology has the potential to transform clinical workflows and improve decision-making in anorectal care.

Disclosures:


Miguel Mascarenhas Saraiva, MD, PhD¹, Maria João Almeida, MD², Miguel Martins, MD², Francisco Mendes, MD¹, Joana Mota, MD¹, Pedro Cardoso, MD¹, Tiago Ribeiro, MD¹, João Afonso, MD¹, Patrícia Andrade, ², António Miguel M P D. Martins Pinto da Costa, MD³, Pedro Diaz Denoso, MD⁴, Jorge Arias, ⁴, João Ferreira, PhD⁵, Bruno Mendes, ⁴, Matheus Ferreira de Carvalho, MD⁶, Guilherme Macedo, MD, PhD¹, Castro Poças, ⁷. P4564 - Artificial Intelligence in Endoanal Ultrasound: A Multicentric Study on a Multilesion Model for Benign Anorectal Lesion Detection and Differentiation, ACG 2025 Annual Scientific Meeting Abstracts. Phoenix, AZ: American College of Gastroenterology.