文章摘要

文章信息/Info

Title:: Automatic region segmentation method of breast tumor based on deep neural network

Author(s):: HOU Pei1; QI Yali2; Second People′s Hospital of Jiaozuo City，Jiaozuo 454000,China；2. Department of Medicine, Jiyuan Vocational and Technical College, Jiaozuo 459000

Keywords:: Deep neural network; Sobel operator; Breast tumor; Region segmentation; Multi-scale features

摘要:: 为提升乳腺肿瘤影像的检测效果，本研究提出一种基于深度神经网络的乳腺肿瘤自动区域分割方法。采用直方图自适应均衡化算法与Sobel算子，对乳腺肿瘤成像的灰度单通道图像进行预处理。基于深度神经网络构建乳腺肿瘤自动区域分割算法网络，其主要构成包括上下文特征多尺度提取单元、3D与2D混合卷积单元和主路网络三个单元。将广义骰子损失当作网络模型训练时的损失函数。训练分割网络后，利用该网络对乳腺肿瘤区域自动分割。使用由专业放射科医生标记与收集的乳腺肿瘤核磁共振成像图像数据集，测试本研究方法的分割性能。结果表明，该方法能够达到很好地分割效果，DSC指标、PPV指标、PPV指标分别达到了76.48%、82.40%、75.93%，实现了乳腺肿瘤区域自动化分割。

Abstract:: In order to improve the detection effect of breast tumor images,we proposed an automatic breast tumor region segmentation method based on deep neural network. The histogram adaptive equalization algorithm and Sobel operator were used to preprocess the gray-scale single-channel images of breast tumor images. Based on deep neural network, an algorithm network for automatic region segmentation of breast tumors was constructed. There were three main components included of context feature multi-scale extraction unit, 3D and 2D mixed convolution unit and main road network. In order to train segmentation network more effectively, generalized dice loss was regarded as the loss function of network model training. After trained the segmentation network, this network was used to segment breast tumors automatically. In the experiment, the segmentation performance was tested by using the breast tumor MRI images data set marked and collected by professional radiologists. The test results showed that this method could achieve a good segmentation effect, with DSC, PPV and PPV reaching 76.48%, 82.40% and 75.93% respectively. It realizes the automatic segmentation of breast tumor region.