CFNet: A medical image segmentation method using the multi-view attention mechanism and adaptive fusion strategy. (January 2023)

• Record Type: Journal Article
• Authors: Zhan, Bangcheng; Song, Enmin; Liu, Hong; Gong, Zhenyu; Ma, Guangzhi; Hung, Chih-Cheng
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 1
• Publication Date: 2023-01
• Subjects: Medical image segmentation -- Deep symmetric architecture -- Attention mechanism -- Deep learning -- Neural network; 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.2022.104112
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: The image feature extraction method based on the attention mechanism has contributed significantly to the accuracy of medical image segmentation. However, the current attention mechanism is based on the single-view information for feature extraction which imposes certain limitations for extracting efficient features. In this study, we propose the encoder-decoder structure of the U-Net as the basic network structure to construct a medical image segmentation method based on the multi-view attention mechanism and adaptive fusion strategy. We will refer to this new network as CFNet. The first component of CFNet is a cross-scale feature fusion method (CFF) employing a new multi-view attention mechanism (MAM) for feature extraction. This can effectively extract features in the multi-receptive field space and obtain more effective cross-scale fusion features in skip-connection. The second component is a fusion weight adaptive allocation strategy (FAS), which can guide the cross-scale fusion features to effectively connect to the decoder features for solving the semantic gap. We evaluated the CFNet using two publicly available medical image segmentation datasets: MoNuSeg and LGG. The experimental results show that the CFNet can achieve better performance compared with the current state-of-the-art methods in medical image segmentation. We then perform extensive ablation studies to validate our method.