Comparison of Radiomics and Deep Learning Using Intestinal Ultrasound Images for Differentiating Inflammatory and...

Intestinal strictures represent a significant complication in Crohn's disease (CD), and accurate differentiation between inflammatory and fibrotic strictures is critical for effective treatment planning. Intestinal ultrasound (IUS) assists in diagnosing and monitoring strictures in CD, yet it faces limitations in distinguishing stricture types. This study aimed to compare radiomics-based and deep learning-based approaches for classifying bowel strictures using IUS images.

We conducted a retrospective analysis of 64 CD patients who underwent surgery between 2018 and 2023. A total of 87 ultrasound images were evaluated using radiomics and deep learning methodologies. We used 5-fold cross-validation, where in each fold, the training set was used for feature extraction and diagnostic model construction, and the test set was used to evaluate model performance. Model performances were assessed through accuracy, sensitivity, specificity, and the area under the ROC curve. Additionally, two experienced radiologists independently evaluated the nature of the strictures, and their assessments were compared with the models. Histopathological analysis via Masson's trichrome staining served as the reference standard.

Results indicated the deep learning-based model had superior performance, achieving an accuracy of 83.3%, sensitivity of 88.9%, and specificity of 77.8%, outperforming the radiomics-based approach, which achieved an accuracy of 67.0%. Although expert evaluations provided the highest specificity (86.7%), the deep learning model demonstrated overall superior accuracy compared to both radiomics and expert predictions. Class Activation Mapping highlighted effective feature identification, enhancing discrimination between inflammatory and fibrotic strictures.

This pioneering research applies deep learning and radiomics to intestinal ultrasound for the first time to distinguish the types of bowel strictures. The real-time visualization of deep learning attention maps (class activation maps) effectively highlights discriminative pathological regions and enhances interpretability, demonstrating strong potential to support clinical decision-making in intestinal stricture classification.