AI-clinic fusion: Multimodal deep learning redefines IBD diagnostics via synergistic CT-clinical feature integration

Traditional methods like endoscopy and imaging are important for IBD management, which are limited by operator dependency, complexity, and high costs, making standardization and routine monitoring difficult. In this study, seven machine learning algorithms were applied to the UK Biobank database to optimize the selection of clinical features and identify key variables associated with IBD. The nnUNetv2 model was then used for CT gut segmentation and validated by a local dataset. Next, the YOLOv10 algorithm was employed to accurately detect IBD lesion regions in the intestine. Finally, an IBD diagnostic model was developed by integrating the selected clinical features with the YOLOv10 identification results, thereby demonstrating the advantages of multimodal data integration. A machine learning-driven feature selection identified chronic abdominal pain, LDL-C, gender, TC, RDW, WBC, ALB, HDL-C, and neutrophil counts as clinical predictors. The nnUNetv2 model achieved high-precision intestinal segmentation, while YOLOv10 demonstrated robust lesion detection. A multimodal nomogram combining clinical features with YOLOv10-derived imaging biomarkers significantly enhanced IBD diagnosis, achieving an AUC of 0.967 that surpassed the clinical-only model (AUC=0.730). This improvement highlights the synergistic value of deep learning-augmented diagnostics for classification accuracy. Integrating clinical features with AI-driven image analysis improves IBD diagnostic precision, enabling better clinical decision-making and patient outcomes. To our knowledge, this is the first multimodal nomogram that combines the predictive clinical features of the UK Biobank scale with CT-based YOLOv10 lesion detection for IBD diagnosis. Different from prior models that solely rely on clinical data or radiomics, our approach achieves integration of population-level epidemiology and deep learning-based imaging, and was validated through external datasets.