Document Type : Original Research
Authors
- Taha Pishro Dabaghiyan 1
- Mona Mohammad- Asghari 1
- Ramin Niknam 2
- Hossein Parsaei 3
- Mohammad Mehdi Movahedi 3
- Tahereh Mahmoudi 3
- Kamran Bagheri Lankarani 2
- Seyed Ali Malek-Hosseini 4
- Seyed Alireza Taghavi 5
- Fardad Ejtehadi 5
- Ebrahim Fallahzadeh Abarghooei 5
- Gholam Reza Sivandzadeh 5
1 Student Research Committee, Shiraz University of Medical Sciences, Shiraz, Iran
2 Health Policy Research Center, Shiraz University of Medical Sciences, Shiraz, Iran
3 Department of Medical Physics and Engineering, School of Medicine, Shiraz University of Medical Sciences, Shiraz, Iran
4 Shiraz Transplant Center, Abu Ali Sina Hospital, Shiraz University of Medical Sciences, Shiraz, Iran
5 Gastroenterohepatology Research Center, Shiraz University of Medical Sciences, Shiraz, Iran
Abstract
Background: Accurate segmentation of common bile duct (CBD) stones during endoscopic retrograde cholangiopancreatography (ERCP) is essential to reduce procedural complications and ensure complete stone removal. However, the high deformability of the CBD and the small size of stones make accurate identification challenging in fluoroscopic images.
Objective: To develop and validate a deep learning based model capable of segmenting CBD stones, the CBD, and the duodenoscope in ERCP fluoroscopic images.
Material and Methods: This retrospective study utilized 1,668 ERCP cholangiograms collected from a single tertiary center. A U-Net-based convolutional neural network was trained using various individual and hybrid loss functions. Model performance was evaluated using Intersection over Union (IoU), precision, and recall.
Results: The model trained with a hybrid loss function combining Dice Loss and Categorical Focal Loss achieved IoU scores of 96.93% for the duodenoscope, 89.76% for the CBD, and a mean IoU of 80.61% for CBD stones. These results reflect a 1.19% improvement in CBD segmentation and a 7.93% improvement in stone segmentation compared to existing approaches.
Conclusion: The proposed deep learning model significantly enhances segmentation accuracy in ERCP imaging and shows strong potential for supporting real-time clinical decision-making. Its integration into ERCP workflows could improve procedural safety, efficiency, and patient outcomes.
Keywords
)