Coordinates encoding networks: an image segmentation architecture for side-viewing catheters

Side-viewing catheter-based medical imaging modalities are used to produce cross-sectional images underneath tissue surfaces. Mainstream side-viewing catheters are based on Optical Coherence Tomography (OCT) or Ultrasound, and they are often applied to the luminal environment. Automatic lumen segmentation provides geometry information for tasks like robotic control and lumen assessment for real-time diagnosis task with side-viewing catheters. In this work, we propose a novel lumen segmentation deep neural networks based on explicit coordinates encoding, which is named CE-net. CE-net is computationally efficient and produces and produces clean segmentation by explicitly encoding the boundaries coordinates in one shot. The experimental evaluation shows a processing time of approximately 8ms per frame while maintaining robustness. We propose a data generation method to improve CE-net generalization, which shows considerable performance by just training with a small dataset.