AI-Driven Automated Differentiation of Benign Anal Lesions via Endoanal Ultrasound
Overview
This study developed a convolutional neural network (CNN) using 3D endoanal ultrasound (EAUS) images to automatically classify benign anal lesions, including fissures and sphincter lacerations. The AI model demonstrated high accuracy, particularly achieving 100% sensitivity, specificity, and accuracy for anal fissures, highlighting its potential to support clinical diagnosis.
Background
The anal canal's sphincters are critical for fecal continence, and benign anorectal disorders such as fissures and sphincter lacerations are common conditions that significantly affect quality of life. Endoanal ultrasonography (EAUS) is the gold standard for assessing sphincter integrity but is limited by operator dependency and accessibility. Artificial intelligence offers a promising approach to overcome these limitations by enhancing diagnostic accuracy and reducing the learning curve associated with EAUS interpretation.
Data Highlights
Lesion Type
Sensitivity (%)
Specificity (%)
Accuracy (%)
External Lacerations
82.5
93.5
88.2
Internal Lacerations
91.7
85.9
88.2
Anal Fissures
100
100
100
Key Findings
A CNN based on the YOLOv11 model was trained on 4528 EAUS frames to classify anal fissures and sphincter lacerations.
The dataset included 516 fissure frames, 2174 external laceration frames, and 1838 internal laceration frames.
The model achieved perfect classification metrics (100% sensitivity, specificity, accuracy) for anal fissures.
For external lacerations, the model showed 82.5% sensitivity and 93.5% specificity.
For internal lacerations, sensitivity was 91.7% and specificity 85.9%, with overall accuracy of 88.2% for both laceration types.
The AI model can potentially reduce the dependency on expert operators and improve reproducibility in EAUS interpretation.
Clinical Implications
The integration of AI into EAUS interpretation could facilitate earlier and more accurate diagnosis of benign anal lesions, particularly in settings lacking specialized expertise. This technology may streamline clinical workflows, reduce variability in image interpretation, and ultimately improve patient management by enabling timely and precise identification of sphincter injuries and fissures.
Conclusion
This proof-of-concept study demonstrates that AI-driven analysis of EAUS images can accurately differentiate benign anal lesions, offering a promising adjunct to clinical practice. Further validation and integration into routine diagnostics could enhance proctologic care and accessibility.
References
ManopH Gastroenterology Clinic, Porto, Portugal, 2022-2024 -- Automated Differentiation of Benign Anal and Sphincter Lesions Using AI and EAUS
