A deep learning-driven method for safe and effective ERCP cannulation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Surgical Technology

Background:

• Computer-assisted endoscopic retrograde cholangiopancreatography (ERCP) cannulation is challenged by complex anatomy and small, similar-looking anatomical landmarks.
• Accurate detection of the duodenal papilla and surgical cannula is critical for successful ERCP procedures.

Purpose of the Study:

• To develop a deep learning-driven graphical user interface (GUI) to assist ERCP cannulation.
• To improve the accuracy and robustness of duodenal papilla and surgical cannula detection in challenging endoscopic scenarios.

Main Methods:

• A novel deep learning method, four swin transformer decoupled heads (4STDH), was proposed for detecting objects of varying sizes.
• The model utilizes decoupled classification and regression networks and integrates swin transformer modules for enhanced attention.
• A new dataset, DPAC, comprising 1840 annotated endoscopic images, was created and will be publicly available.

Main Results:

• The 4STDH method achieved a mean Average Precision (mAP) of 93.2% on the DPAC dataset, outperforming state-of-the-art methods.
• The GUI provides real-time positional data for the papilla and cannula, including planar distance and direction for accurate targeting.
• The system demonstrated superior generalization performance.

Conclusions:

• The developed GUI, powered by deep learning, shows significant potential to enhance the safety and efficiency of clinical ERCP cannulation.
• Validation in human endoscopic videos confirms the practical applicability and effectiveness of the deep learning approach.