Abstract

This PhD work focuses on the automatic segmentation of the guidewire—an endoscopic instrument—in 2D fluoroscopic X-ray images. Several methodological challenges must be addressed: limited imaging data, poor image quality, and the absence of annotated datasets. Furthermore, no existing methods are specifically dedicated to the segmentation of guidewires in ERCP (Endoscopic Retrograde Cholangiopancreatography). To overcome these challenges, several contributions are proposed. First, a harmonization and annotation of the collected dataset was carried out to build a usable reference for future supervised learning experiments. Due to the very fine structure of the guidewire, standard segmentation methods and metrics prove insufficient and often inadequate for proper model evaluation and comparison. Therefore, this work explores evaluation approaches used in other fields, particularly vascular segmentation, and considers clinical evaluation methods to complement data science techniques and provide physicians with better tools to assess model performance. For the segmentation task, techniques developed for endovascular interventions are adapted—especially U-Net architectures, which are widely used in medical imaging. The approach includes a comparison of various loss functions tailored to guidewire segmentation, resulting in accurate segmentation of clinically relevant portions. Additionally, the endoscope at the biliary duct entrance is also segmented to assess whether its position in the image can help improve guidewire segmentation. In conclusion, this work provides a first approach to guidewire segmentation with the goal of improving ERCP procedures.

Additional Information

The seminar will be given in 🇫🇷 French 🇫🇷, and is open to everyone, either in person in room APP3 (Paris campus) or online via Teams.