Self-attentional microvessel segmentation via squeeze-excitation transformer Unet. (April 2022)

• Record Type: Journal Article
• Authors: Shen, Xiongri; Xu, Jingjiang; Jia, Haibo; Fan, Pan; Dong, Feng; Yu, Bo; Ren, Shangjie
• Is Part Of: Computerized medical imaging and graphics. Volume 97(2022)
• Publication Date: 2022-04
• Subjects: Microvessel segmentation -- Deep neural network -- Self attention mechanism -- Fundus image -- Optical coherence tomography; Diagnostic imaging -- Periodicals; Imaging systems in medicine -- Periodicals; Diagnosis, Radioscopic -- Data processing -- Periodicals; Diagnostic Imaging -- Periodicals; Imagerie pour le diagnostic -- Périodiques; Diagnostic imaging; Periodicals; Electronic journals; Electronic journals; 616.0754
• DOI: 10.1016/j.compmedimag.2022.102055
• Languages: English
• ISSNs: 0895-6111

Abstract

Abstract: Automatic vessel segmentation is a key step of clinical or pre-clinical vessel bio-markers for clinical diagnosis. In previous research, the segmentation architectures are mainly based on Convolutional Neural Networks (CNN). However, due to the limitation of the receipt of field (ROF) of convolution operation, it is difficult to further improve the accuracy of the CNN-based methods. To solve this problem, a Squeeze-Excitation Transformer U-net (SETUnet) is proposed to break the ROF limitation of CNN. The proposed squeeze-excitation Transformer can introduce the self attention mechanism into the vessel segmentation task by generating a global attention mapping according to the entire vessel image. To test the performance of the proposed SETUnet, the SETUnet is trained and tested on several public vessel data-sets. The results show that the SETUnet outperforms several state-of-the-art vessel segmentation neural networks, especially on the connectivity of the segmented vessels. Highlights: A CNN-Transformer Hybrid network is developed for micro-vessel segmentation. Residual attention block is used to improve feature extraction performance. Squeeze-Excitation operation is adopted for more efficient self-attention mechanism. The proposed network achieved state-of-the-art performance on five public datasets.