Optimal Coronary Artery Segmentation Based on Transfer Learning and UNet Architecture

Recent results demonstrated that the use of AI to perform complicated segmentation of medical images becomes very useful when the coronary arteries are considered. Nevertheless, the different segments of the coronary arteries (distal, middle and proximal) exhibit singularities, mostly linked to section changes and image visibility, that point in the direction to consider each in a singular way. In the present contribution we thoroughly analyse the quality of the segmentation obtained using different neural networks, based on the UNet architecture, applied to the three segments of the coronary arteries. We observe that for proximal segments any of the AI considered provides acceptable segmentations while for distal segments the 3D UNet is not able to recognise the coronary structures. In addition, in the distal region there is a noticeable improvement in the 2D UNet without pre-training compared to the 2D networks with pre-training.