Introduction

Accurate diagnosis and monitoring of inflammatory bowel diseases (IBD) are essential for effective management and reduction of complications. Intestinal ultrasound (IUS) is a non-invasive and cost-effective imaging modality for IBD assessment, yet its widespread adoption is limited by operator dependency and interobserver variability. Artificial intelligence (AI) offers the potential to overcome these limitations by improving diagnostic accuracy and standardizing interpretation.

Aims & Methods

This study aimed to develop an AI-assisted model for the automatic detection of active inflammation in IUS images, using bowel wall thickness (BWT) as a key marker. We further aimed to create an automatic tool for detecting bowel segments within ultrasound images and to extend its application from single images to cine loops. Grayscale ultrasound images and cine-loops of the terminal ileum were collected from Crohn’s disease patients undergoing IUS at a tertiary center. The dataset included both normal and abnormal BWT and was annotated by an experienced operator. Preprocessing involved anonymization and normalization. Automated segmentation algorithms were used to isolate bowel wall structures, and deep learning techniques, including convolutional neural networks (CNNs) and YOLO (You Only Look Once) object detection, were employed to analyze the images.

Results

The study included 2,735 grayscale ultrasound images and cine-loops of the terminal ileum from patients with Crohn's disease. The dataset was divided into three sets: a training set of 1980 images, a validation set of 220 images, and a test set of 515 images. Our model achieved robust performance metrics, including sensitivity, specificity, and accuracy of 99.1%, 94.6%, and 91.5%, respectively, for BWT detection, with an area under the receiver operating characteristic curve (AUC) of 0.85. Our model was validated on 10 independent cine-loops, comprising 7 abnormal and 3 normal cases. It accurately identified bowel wall thickening (BWT) in all seven abnormal cases, correctly classifying them as abnormal. Additionally, the model successfully confirmed the absence of BWT in all three normal cine-loops, categorizing them appropriately as normal.

Conclusion

This study represents a significant advancement in the application of artificial intelligence to intestinal ultrasound interpretation by integrating automated segmentation, automatic analysis capabilities, and cine-loop evaluation. Our model extends beyond traditional static image analysis, paving the way for real-time and dynamic assessments of bowel inflammation. This approach highlights the potential of our AI tools to reduce operator dependency, improve diagnostic accuracy, and enhance clinical workflows, making IUS more accessible and reliable for clinicians with varying levels of expertise.

Disclosure

Prof. Dan Carter has received speaker’s fees, research support, or consultancy fees from Takeda, Janssen, AbbVie, and Taro