基于多层融合注意力的乳腺肿瘤图像分割方法-《计算机技术与发展》

文章信息/Info

Title:: Breast Tumor Images Segmentation Method Based on Multi-layer Fusion Attention

作者:: 王宇昕¹; 2 ; 付晓薇¹; 2 ; 赵思宇¹; 2 ; 陈芳³; 1. 武汉科技大学计算机科学与技术学院,湖北武汉 430065;
2. 智能信息处理与实时工业系统湖北省重点实验室,湖北武汉 430065;
3. 武汉科技大学校医院超声影像科,湖北武汉 430065

Author(s):: WANG Yu-xin1; 2 ; FU Xiao-wei1; 2 ; ZHAO Si-yu1; 2 ; CHEN Fang3; 1. School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;
2. Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan 430065,China;
3. Department of Ultrasound and Imaging of Hospital,Wuhan University of Science and Technology,Wuhan 430065,China

Keywords:: breast ultrasound images; tumors segmentation; U-Net; convolutional neural network; attention mechanism

摘要:: 针对超声乳腺肿瘤图像中存在的高散斑噪声较多、肿瘤边缘模糊以及形状复杂多样等问题,提出了一种基于多层融合注意力的超声乳腺肿瘤图像分割方法。首先,在保持 U-Net 编-解码结构的基础上,采用经过预训练的 ResNet-34 模型,用于在编码部分提取更深层次的特征;然后,在跳跃连接部分对相邻的浅层特征与深层特征分别进行空间与通道维度上的增强;其次,将经过注意力增强后的不同层次特征进行融合,重点关注肿瘤区域的位置,以避免散斑噪声干扰下的错误分割;最后,利用普通卷积层进行特征还原,得到分割结果。实验结果表明,所提方法对噪声干扰较大的超声乳腺肿瘤图像鲁棒性更强,Dice 系数、IoU　Recall 和 Precision 分别能够达到 0. 852 2、0. 768 2、0. 877 3 和 0. 863 0。同时,所提方法在模型复杂度上也有较好的表现,较对比方法具有更优的分割性能。

Abstract:: In view of the more speckle noise,blurred tumor edges and diverse tumor shapes in ultrasound breast ultrasound images,an ultrasound breast tumor images segmentation method based on multi-layer fusion attention is proposed. Firstly,on the basis of maintainingthe encoding and decoding structure of U-Net,pre-trained ResNet-34 model is used in the encoding part to extract deeper features. Secondly,in the skip connection, the adjacent shallow features and deep features are enhanced in the spatial and channel dimensionsrespectively, Then,the features of different layers after attention enhancement are fused,focusing on the location of the tumor area toavoid false segmentation under speckle noise interference. Finally,the ordinary convolutional layer is used for feature restoration to obtainthe segmentation results. The experiment results shows that the proposed method is more robust to ultrasound breast tumor images withlarge noise interference, with the Dice coefficient, IoU, Recall, and Precision reaching 0. 852 2, 0. 768 2, 0. 877 3, and 0. 863 0,respectively. Meanwhile,the proposed method also has better performance in model complexity and better segmentation performance thanthe comparison method.