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

The research team implemented a single-center design involving 185 patients who underwent stent placement. Review approach framing focuses on the use of a ResNet-50 neural network for image analysis. The investigators curated a primary dataset containing 412 images from 151 individuals. They subsequently expanded this collection to 488 images to include partially covered devices. A 5-fold cross-validation strategy ensured rigorous evaluation of the model across different data partitions. The team calculated mean values and standard deviations for accuracy, precision, recall, and F1 metrics. Transfer learning techniques allowed the system to adapt to new stent categories without full retraining. This methodology enabled the assessment of model performance stability during incremental updates.

Key findings from the literature indicate that the model achieved an F1 score of 94.78 ± 4.07 for vendor identification. For single versus multiple stent detection, the system yielded an F1 score of 57.03 ± 6.77. Stent-specific performance ranged from 91.43 ± 3.43 for uncovered devices to 97.91 ± 2.59 for Epic™ models. After incorporating additional images, the model maintained or improved these performance metrics. Transfer learning adaptation resulted in an F1 score of 59.06 ± 9.08 for single versus multiple detection. Stent-specific scores after adaptation ranged between 83.6 ± 5.8 and 97.73 ± 2.42. The data confirm that the model reliably classified devices on radiographic and fluoroscopic inputs. These results demonstrate that the system remains effective even when updated with new stent types.

The ResNet-50 architecture provides a reliable framework for identifying biliary stents on standard imaging. Synthesis and implications suggest that this approach supports procedural planning by confirming device identity. The authors propose that transfer learning effectively integrates new stent models into existing systems. This method maintains high performance levels despite the addition of novel device categories. The findings indicate that artificial intelligence tools can mitigate challenges posed by diverse manufacturer designs. Clinicians may utilize these automated systems to streamline device verification during biliary interventions. The evidence demonstrates that model adaptation is feasible without significant degradation in diagnostic accuracy. These results highlight the potential for integrating deep learning into routine clinical workflows for biliary care.