Echocardiographic image segmentation using deep Res-U network. (February 2021)

• Record Type: Journal Article
• Authors: Ali, Yasser; Janabi-Sharifi, Farrokh; Beheshti, Soosan
• Is Part Of: Biomedical signal processing and control. Volume 64(2021)
• Publication Date: 2021-02
• Subjects: Echocardiography -- Segmentation -- Deep learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2020.102248
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A fast and fully automatic deep learning framework for cardiac boundary segmentation with the focus on left ventricle segmentation. Proposing a network Res-U consisting of a modified version of ResNet-50 to be used as an encoder in the U-net. Proposing a new strategy of training to improve the accuracy of the proposed model. The model outperforms other state-of-the-art methods in terms of accuracy with a Dice metric of 0.97 ± 0.01. Abstract: Cardiac function assessment using echocardiography is a crucial step in daily cardiology. However, cardiac boundary segmentation and in particular, ventricle segmentation is a challenging procedure due to shadows and speckle noise. Manual segmentation of the cardiac boundary is a time-consuming process which rules out conventional segmentation for many situations such as emergency cases and image-guided robotic interventions. Therefore, providing an efficient and robust autonomous segmentation method is crucial for such applications. In this paper, a fast and fully automatic deep learning framework for left ventricle segmentation is proposed. This model couples the advantages of ResNet and U-Net to provide reliable segmentation results. We propose a new encoder in the U-Net, defined as ResU which is a modified version of ResNet-50 and has a superiority over ResNet in data denoising. We trained this model on the dataset CAMUS (Cardiac Acquisitions for Multi-structure Ultrasound Segmentation) which is a large, publicly available and fully annotated dataset for 2D echocardiographic assessment. It is shown that this model outperforms other state-of-the-art methods in terms of accuracy with a Dice metric of 0.97 ± 0.01 .