Overcoming Data Scarcity for Coronary Vessel Segmentation Through Self-supervised Pre-training

Cardiovascular diseases affect a significant part of the population, leading to deterioration in life quality, health degradation, and even premature death. One of the most effective diagnostic methods for the disease is based on medical imaging, specifically Computed Tomography Angiography, from which the complete 3D image of the coronary vessels can be reconstructed. Manual annotation and reconstruction is a tedious process, so a range of automated methods have been proposed over the years, with the methods based on deep neural networks achieving the best results recently. On the downside, such methods require extensive datasets for training. To overcome the problems with data scarcity, we propose a method for self-supervised pre-training of neural networks performing the task of coronary vessel segmentation. The method is based on a vesselness filter and significantly boosts performance, reducing the training time and boosting the accuracy without additional annotated data.