Asymmetric U-shaped network with hybrid attention mechanism for kidney ultrasound images segmentation. (February 2023)

• Record Type: Journal Article
• Authors: Chen, Gong-Ping; Zhao, Yu; Dai, Yu; Zhang, Jian-Xun; Yin, Xiao-Tao; Cui, Liang; Qian, Jiang
• Is Part Of: Expert systems with applications. Volume 212(2023)
• Publication Date: 2023-02
• Subjects: Kidney ultrasound -- Images segmentation -- Deep learning -- Attention mechanism -- Deep supervision; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.118847
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: An asymmetric U-shaped network is developed to segment kidney automatically. The multi-step down-sampling strategy can improve the adaptability of the network. A hybrid attention module is designed to capture more robust kidney characteristics. The side-out module can to guide network learns to predict precise segmentations. Abstract: Kidney ultrasound (KUS) images segmentation is one of the key steps in computer-aided diagnosis. The perturbation of heterogeneous structure, similar intensity distribution, and kidney morphology pose challenges for the segmentation of KUS images. In this paper, we proposed an asymmetric U-shaped network based on the U-net core architecture to segment KUS images accurately and reliably. Specifically, the architecture mainly consists of a dense residual connection encoder, a multi-step up-sampling decoder with the hybrid attention module, and a side-out deep supervision module. The design of the dense residual connection encoder can capture sufficient kidney feature information to improve the representation ability of the network. The development of the hybrid attention module can further guide the network to pay more attention to the representation of the kidney. In addition, the introduction of the side-out deep supervision model can help the network obtain segmentation results that are closer to the ground-truth masks. Moreover, to reduce network parameters, we proposed a multi-step up-sampling optimization strategy to simplify the design of the network. We compare with several state-of-the-art medical image segmentation methods on the same KUS dataset using seven quantitative metrics. The results of our method on Jaccard, Dice, Accuracy, Recall, Precision, ASSD and AUC are 89.95%, 94.59%, 98.65%, 94.47%, 95.07%, 0.3006 and 0.9703, respectively. Experimental results demonstrate that the proposed method achieves the most competitive segmentation performance on KUS images.