Deep learning-based detection of suboptimal arterial phase timing in abdominal CT: prevalence and associated patient factors

To develop a deep learning method for automated contrast phase identification in abdominal CT and characterize suboptimal arterial phase timing in patients with altered hemodynamics. A 3D-ResNet18 model was trained on 3,129 CT series from 1,160 patients to classify four contrast phases: noncontrast, early arterial phase (AP), late arterial phase, and portal venous phase (PVP). Performance was evaluated on internal test data (710 patients) and external validation (WAW-TACE dataset, 233 patients; three-category task combining early and late AP) using precision, sensitivity, specificity, F1 score, and AUC. Gradient-weighted class activation mapping enhanced interpretability. Logistic regression evaluated patient factors associated with early arterial phase capture when late arterial phase was intended. Internal test F1 scores were 1.00 for noncontrast, 0.92 for early AP, 0.88 for late AP, and 0.98 for PVP. External validation showed comparable three-category performance, with F1 scores ranging from 0.96 to 0.99. Among the five evaluated models, 3D-ResNet18 achieved the highest macro-averaged F1 score on the internal test set (0.943; 95% CI, 0.930–0.954). Grad-CAM showed clinically appropriate focus on vascular and parenchymal enhancement. Older age (> 60 years; OR, 1.854; P = 0.001), liver cirrhosis (OR, 4.461; P < 0.001), ascites (OR, 2.598; P = 0.001), and pancreatic cancer (OR, 7.103; P < 0.001) were independently associated with early AP capture. This deep learning algorithm achieved excellent four-phase classification performance on internal testing and robust three-category performance on external validation; early-versus-late arterial phase discrimination itself was not externally validated. The observed associations suggest that fixed timing protocols may not fully accommodate interpatient variation in arterial enhancement. The model may support protocol auditing and research applications in multiphase radiomics and deep learning, while individualized timing strategies require prospective validation.