Comparison of Loss Functions for Guidewire Automatic Detection on a Limited Dataset of Fluoroscopic Images for Endoscopic Retrograde Cholangiopancreatography

Guidewire detection is an important yet challenging task to improve Endoscopic Retrograde CholangioPancreatography (ERCP). Knowing the accurate position of instruments within the bile ducts is helpful to the endoscopist and reduces the risk of procedure failure or adverse event. This article is the first to address guidewire detection for ERCP. We propose a method based on a well-known and simple framework for biomedical image segmentation, called U-Net, and improve its performance for thin structures such as guidewires, with adapted pre-processing and loss functions. We also use transfer learning to improve performance on a second dataset. The contributions are i) a comparison of loss functions for guidewire detection on 2D X-Ray fluoroscopy images of ERCP, and ii) a proposition to complete metrics for evaluation. We obtained Dice, clDice and Tversky scores of 0.66, 0.74 and 0.64 on our first dataset. On our second dataset we obtained Dice, clDice and Tversky scores of 0.65, 0.75 and 0.63. This study demonstrates the impact of loss functions and transfer learning to improve performances on small and noisy datasets.