Endoscopy 2025; 57(S 02): S185
DOI: 10.1055/s-0045-1805462
Abstracts | ESGE Days 2025
Moderated poster
ERCP and EUS: what's new? 03/04/2025, 09:00 – 10:00 Poster Dome 2 (P0)

Artificial Intelligence and Endoanal Ultrasound: Pioneering Automated Differentiation of Benign Anal and Sphincter Lesions

Authors

  • M Mascarenhas

    1   São João Universitary Hospital Center, Porto, Portugal

  • M J Almeida

    2   Department of Gastroenterology, São João University Hospital Center, Porto, Portugal

  • M Martins

    2   Department of Gastroenterology, São João University Hospital Center, Porto, Portugal

  • F Mendes

    2   Department of Gastroenterology, São João University Hospital Center, Porto, Portugal

  • J Mota

    2   Department of Gastroenterology, São João University Hospital Center, Porto, Portugal

  • C Pedro

    1   São João Universitary Hospital Center, Porto, Portugal

  • J Ferreira

    3   Faculty of Engineering – University of Porto, Porto, Portugal

  • B Mendes

    3   Faculty of Engineering – University of Porto, Porto, Portugal

  • M Guilherme

    1   São João Universitary Hospital Center, Porto, Portugal

  • F Poças

    4   Hospital Geral de Santo António, Porto, Portugal
Further Information
Also available at
 

Aims Anal injuries, such as lacerations and fissures, present diagnostic challenges due to the anatomical complexity of the anal canal and reliance on expert interpretation of Endoanal Ultrasound (EAUS) images. EAUS is a highly reliable tool for detailed visualization of anal structures, but its efficacy is limited by variability in expertise. Artificial intelligence (AI), particularly Convolutional Neural Networks (CNNs), may enable more accurate, consistent, and accessible diagnoses. This study aimed to develop and validate a pioneer CNN-based model for automatic classification of fissures and anal lacerations (external and internal) using EAUS frames.

Methods In this proof-of-concept study, 238 EAUS radial probe exams performed between April 2022 and January 2024 were analyzed, yielding 4528 frames categorized into fissures (516), external lacerations (2174), and internal lacerations (1838). Frame categorization was validated by three expert reviewers. Data was split 80% for training and 20% for testing. Performance metrics included sensitivity, specificity, and accuracy to evaluate model performance.

Results The CNN demonstrated promising diagnostic performance. For external lacerations, sensitivity, specificity, and accuracy were 82.5%, 93.5%, and 88.2%, respectively. For internal lacerations, sensitivity was 91.7%, specificity 85.9%, and accuracy 88.2%. Notably, for anal fissures, the model achieved 100% sensitivity, specificity, and accuracy.

Conclusions This first worldwide AI-assisted model for differentiating benign anal injuries on EAUS demonstrates excellent diagnostic performance, particularly for anal fissures. By reducing reliance on expertise and improving diagnostic consistency, the model has the potential to support broader clinical adoption of EAUS in proctology. Limitations, including the small dataset and single-center scope, underline the need for future studies to validate findings on larger, multi-center datasets



Publication History

Article published online:
27 March 2025

© 2025. European Society of Gastrointestinal Endoscopy. All rights reserved.

Georg Thieme Verlag KG
Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany