基于RMAU-Net的DBT图像肿块自动分割方法  

Automatic segmentation of breast masses in DBT images based on RMAU-Net

作　　者：茅瑜 孙浩天 吴俊[1] 郑健[2] MAO Yu;SUN Haotian;WU Jun;ZHENG Jian(Yangzhou University School of Information Engineering,Yangzhou 225127,China;Medical Imaging Laboratory,Suzhou Institute of Biomedical Engineering and Technology of Chinese Academy of Sciences,Suzhou 215263,China)

机构地区：[1]扬州大学信息工程学院,扬州225127 [2]中国科学院苏州生物医学工程技术研究所,苏州215163

出　　处：《中国体视学与图像分析》2024年第2期148-159,共12页Chinese Journal of Stereology and Image Analysis

基　　金：国家自然基金(62371449);苏州市基础研究计划(SJC2021023)。

摘　　要：准确的乳腺肿块分割对于早期乳腺癌的诊断和治疗具有重要的意义。目前数字乳腺断层摄影(DBT)已广泛应用于乳腺癌的检查诊断,具有较高的病变检出率。但是DBT图像中乳腺致密度较高、对比差异度较低,使得乳腺肿块的自动分割更具挑战性。为了高效、准确的对DBT图像中的肿块进行分割,本文提出了一种残差多注意U形分割网络(RMAU-Net),利用残差结构避免了梯度消失而导致的模型性能下降。同时,在网络中采用深层注意特征融合模块和多路径深层特征融合模块,提高了网络的特征提取能力以及对可疑区域边界的识别能力。RMAU-Net在一个私有的DBT图像数据集(DBT_SZ)中对乳腺肿块进行分割,Dice达到86.77%,敏感性达到87.84%、IOU达到80.15%。此外,本文还将RMAU-Net与一些先进的分割网络进行了比较,实验结果表明,RMAU-Net可以提取到更精确的乳腺肿块边缘,提高了分割精度。Accurate breast mass segmentation is important for the diagnosis and treatment of early breast cancer.Digital breast tomosynthesis(DBT)has been widely used for breast cancer screening with a high detection rate for lesions.However,the high breast densities and low contrast in DBT images make the automatic segmentation of breast masses very challenging.In order to efficiently and accurately segment the masses in DBT images,this paper proposes a residual multi-attention U-shaped segmentation network(RMAU-Net),which utilizes a residual structure to avoid performance degradation caused by gradient vanishing.Meanwhile,a deep attention feature fusion module and a multipath high-level feature fusion module are used in the network to improve the feature extraction ability of the network as well as the ability to recognize the boundary of suspicious regions.The RMAU-Net performs segmentationon a private DBT image dataset(DBT_SZ)and achieves a Dice of 8677%,a sensitivity of 8784%,and an IOU of 8015%.In addition,this paper compares RMAU-Net with some advanced segmentation networks.Experimental results show that RMAU-Net can extract mass edges more accurately so that improve the segmentation accuracy.

关 键 词：乳腺肿块分割 数字乳腺断层摄影图像 深度学习 残差结构 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]