Deep learning-based artificial intelligence can improve the diagnosis of small bowel obstruction: stratified comparison study and hierarchical Bayesian model

Accurate diagnosis of small bowel obstruction (SBO) is critical to patient outcomes, particularly in the emergency department (ED). To enhance diagnostic precision, we developed an artificial intelligence (AI) technology that automatically extracts dilated intestinal segments from contrast-enhanced computed tomography (CT) images. 158 contrast-enhanced CT examinations containing 5,200 annotated images were used for deep learning, and the potential utility of AI in improving SBO diagnosis was subsequently evaluated by residents and surgeons. CT images from 30 patients with suspected SBO in the ED were used as a test set. Seventeen residents and ten surgeons were divided into two groups, one interpreting images with AI support and the other without AI. Participants assessed the presence of SBO and identified the obstruction location, and diagnostic time was recorded. A hierarchical Bayesian model was applied for analysis. The median precision, recall, and Dice score of the AI model were 0.98, 0.63, and 0.77, respectively. For both residents and surgeons, the correct diagnosis rate of the obstruction location was significantly higher in the AI-assisted group compared with the non-AI group (74.1% vs. 56.7% and 88.2% vs. 66.2%, respectively; P < 0.0001 and P = 0.0001). Among residents, AI support significantly improved the diagnosis of obstruction location (odds ratio: 4.20; 95% credible interval: 2.14–8.26) and reduced reading time by 26.84 s per case (95% credible interval: −50.37 to − 2.83). These findings indicate that AI technology is clinically feasible and can improve diagnostic accuracy while reducing the time required for diagnosing bowel obstruction.