Abstract
To develop a deep learning model using abdominal noncontrast computed tomography (CT) for diagnosing acute cholecystitis (AC) and predicting progression to acute suppurative cholecystitis (ASC). A total of 641 patients from three medical centers were retrospectively analyzed. Deep learning models were constructed for AC diagnosis and ASC prediction. Model interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM) and t-distributed stochastic neighbor embedding. Performance was compared with radiomics models and radiologist assessments. The deep learning model achieved accuracies of 89.81% (internal) and 81.83% (external) for AC and 84.52% (internal) and 85.60% (external) for ASC prediction. Grad-CAM visualizations showed focus on gallbladder regions and surrounding areas. The multimodal models integrating clinical information outperformed imaging-only models. Deep learning significantly surpassed radiomics models and radiologist assessments, with high computational efficiency (segmentation: 10.5 ± 3.8 s, inference: 1.3 ± 0.4 s). This efficient deep learning system accurately identifies AC and ASC from noncontrast CT, offering robust tools for time-sensitive clinical workflows.