Deep learning-based biliary stent classification and transfer learning adaptation to an additional stent type

OBJECTIVE: Accurate classification of previously deployed biliary stents is crucial for planning reintervention, yet conventional image interpretation is challenging due to diverse vendor-specific designs. This study aimed to develop a deep learning model for biliary stent classification and to evaluate transfer learning as an adaptation strategy for a new stent type. MATERIALS AND METHODS: This single-center study included 185 patients who underwent biliary stent placement. The primary dataset (412 images from 151 patients) included four stent types: Epic™, EGIS, Niti-S, and Bonastent® uncovered. The augmented dataset (488 images from 185 patients) incorporated additional images of Bonastent® partially covered stents. A ResNet-50 model was trained to classify stents by number (single versus multiple), vendor, and specific type using 5-fold cross-validation. Transfer learning was applied to incorporate the additional stent type into the model. Accuracy, precision, recall, and F1 score were calculated as mean ± standard deviation across folds. RESULTS: For the primary dataset, the model achieved an F1 score of 57.03 ± 6.77 for single/multiple detection and 94.78 ± 4.07 for vendor classification. Stent-specific F1 scores ranged from 91.43 ± 3.43 (Bonastent® uncovered) to 97.91 ± 2.59 (Epic™). After augmentation, the model yielded similar or improved scores. Transfer learning achieved comparable results (e.g., 59.06 ± 9.08 for single/multiple, 83.6 ± 5.8 to 97.73 ± 2.42 for stent-specific). CONCLUSION: The ResNet-50 model demonstrated reliable performance in classifying biliary stents on radiographic and fluoroscopic images. Transfer learning allowed the model to be updated to incorporate newly introduced stent types with minimal degradation in performance. RELEVANCE STATEMENT: Deep learning-based biliary stent classification may assist clinicians by rapidly classifying previously placed stents on radiographic and fluoroscopic images, thereby supporting procedural planning and device confirmation during biliary interventions. KEY POINTS: Accurate classification of previously deployed biliary stents remains difficult due to diverse vendor-specific designs and similar radiographic appearances. A ResNet-50 deep learning model reliably classified biliary stents, and transfer learning enabled the model to incorporate newly introduced stent types while maintaining performance comparable to the baseline model. This artificial intelligence-based approach provides rapid classification of previously placed biliary stents on routine images, which may assist with device verification during biliary interventions.