A recurrent skip deep learning network for accurate image segmentation. (April 2022)

• Record Type: Journal Article
• Authors: Shi, Ce; Zhang, Juan; Zhang, Xin; Shen, Meixiao; Chen, Hao; Wang, Lei
• Is Part Of: Biomedical signal processing and control. Volume 74(2022)
• Publication Date: 2022-04
• Subjects: Image segmentation -- OCT -- Color fundus photography -- Deep learning network -- Skip connection; 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.2022.103533
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A short backward skip connection was introduced to enable flexible aggregations of convolutional features. An attention-aware convolutional block was introduced to suppress irrelevant background. A novel deep learning network was developed by integrating the introduced components with the available BiO-Net for segmentation purposes. Extensive experiments suggested that the developed network can accurately extract different target objects and compete with several existing networks. Abstract: Accurate image segmentation plays a vital role in quantitatively assessing various diseases and their prognosis. In this study, we described a novel deep learning network termed Recurrent Skip Network (RS-Net) by integrating a backward skip connection and an attention-aware convolutional block with the available BiO-Net. To validate its performance and merits, we applied it (1) for the segmentation of three corneal layers (i.e., epithelium layer, Bowman's layer, and stroma layer) depicted on optical coherence tomography (OCT) images and (2) for the segmentation of the optic disc (OD) and cup (OC) depicted on color fundus photography (CFP). Our experiments showed that RS-Net achieved an average Dice score of 0.9327 and 0.8868, respectively for the two different segmentation tasks, demonstrating a unique performance as compared with BiO-Net and other networks.