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%.
CITATION STYLE
Hamdi, R., Kerkeni, A., Bedoui, M. H., & Ben Abdallah, A. (2022). Res-GAN: Residual Generative Adversarial Network for Coronary Artery Segmentation. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 13756 LNCS, pp. 391–398). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-21753-1_38
Mendeley helps you to discover research relevant for your work.