Context-aware network fusing transformer and V-Net for semi-supervised segmentation of 3D left atrium. (15th March 2023)

• Record Type: Journal Article
• Authors: Zhao, Chenji; Xiang, Shun; Wang, Yuanquan; Cai, Zhaoxi; Shen, Jun; Zhou, Shoujun; Zhao, Di; Su, Weihua; Guo, Shijie; Li, Shuo
• Is Part Of: Expert systems with applications. Volume 214(2023)
• Publication Date: 2023-03-15
• Subjects: 3D MRI -- Contextual information -- Image segmentation -- Semi-supervised learning -- Transformers; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.119105
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: A novel network is proposed for the semi-supervised segmentation of Left Atrium. The proposed network benefits from the Transformer and the V-Net. A discriminator with attention mechanism introduces strong prior information. Only a small amount of label data is required to achieve high performance. Performance for segmenting the left atrium outperforms some SOTA methods. Abstract: Accurate, robust and automatic segmentation of the left atrium (LA) in magnetic resonance images (MRI) is of great significance for studying the LA structure and facilitating the diagnosis and treatment of atrial fibrillation. Semi-supervised learning has attracted great attention in medical image segmentation, since it alleviates the heavy burden of annotating training data. In this paper, we propose a context-aware network called CA-Net for semi-supervised LA segmentation from 3D MRI. The information of 3D MRI to be learned is not only the contextual information in each slice, but also the spatial information among different slices of the data, which is not sufficiently exploit by existing methods. In the proposed CA-Net, a Trans-V module is coined from both Transformers and V-Net, which is able to learn contextual information in 3D MRI. In the training processing, the discriminator with attention mechanisms is introduced to calculate an adversarial loss so that a large amount of unlabeled data can be utilized. Experimental results on the Atrial Segmentation Challenge dataset show that the contextual information is helpful to extract more accurate atrial structures, and the proposed CA-Net achieves better performance than some SOTA semi-supervised networks. Our method achieves dice scores of 88.14% and 90.09% in segmentation results when trained with 10% and 20% of labeled data, respectively. Code will be available at: https://github.com/RhythmI/CA-Net-master .