Attention-inception-based U-Net for retinal vessel segmentation with advanced residual. (March 2022)

• Record Type: Journal Article
• Authors: Wang, Huadeng; Xu, Guang; Pan, Xipeng; Liu, Zhenbing; Tang, Ningning; Lan, Rushi; Luo, Xiaonan
• Is Part Of: Computers & electrical engineering. Volume 98(2022)
• Publication Date: 2022-03
• Subjects: Retinal vessel segmentation -- U-Net -- Attention mechanism -- Inception module -- Residual block; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2021.107670
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: Constructing a lightweight network with far fewer parameters than the classic U-Net. Proposing a new residual block with dilated convolution of different dilation rates. Combining inception and attention module to capture deeper semantic features. Good generalization ability for different medical image segmentation applications. Abstract: U-Net based methods have been widely used in retinal vessel segmentation tasks. But there are still some obstacles needing to be crossed, such as the loss of microvasculature details at the end of vessels, as well as the interference of the hard exudate produced by lesions. To address these issues, this paper proposes a novel AR-SA U-Net model that integrates residual block with dilated convolution, inception module and an scSE-based attention mechanism. The model also optimizes the upsampling of original U-Net by combining bilinear interpolation and transpose convolution to update their weights dynamically. The experimental results on three retinal vessel image datasets show that the proposed model can eliminate the influence of hard exudate produced by the lesions in the segmentation results, and the segmentation results are clear in detail with high performance. The accuracy of the proposed model on DRIVE, STARE and CHASE_DB1 is 96.11%, 97.78% and 96.79% respectively, which is 1.07%, 1.06% and 1.29% higher than that of original U-Net. The proposed model also shows strong generalization ability in non-fundus medical image datasets. Graphical abstract: Image, graphical abstract