Res-GAN: Residual Generative Adversarial Network for Coronary Artery Segmentation

Abstract

Segmentation of coronary arteries in X-ray angiograms is a crucial step in the assessment of coronary disease. Recently, many automatic approaches have been proposed to minimize time-consuming clinicians intervention. However, due to noise and complex vessel structure in this modality, most of those approaches fail to segment thin vessels. In this paper, we introduce a new generative adversarial network called Res-GAN to obtain accurate vessel segmentation of both thick and thin vessels. It consists of a Residual-UNet generator following the encoder-decoder structure; and a Residual CNN discriminator for more efficient segmentation. Besides, in order to improve the training process, we adopt a loss function combining both binary cross-entropy and Dice losses. For the experiment results, we used our private dataset to compare the proposed architecture with others state-of-the-art models. The results demonstrate that Res-GAN outperforms the others architectures. It achieves the highest accuracy of 96,55% and Dice metric of 81,18%.