陈翔1　蔡延光1,2　龚国俊3　张瑞湖4　蔡颢5 

（1.广东工业大学自动化学院，广东 广州 510006　

2.广州理工学院智能制造与电气工程学院，广东 广州 510540 

3.广东省岭南工商第一技师学院，广东 广州 510000

4.广东省粤东技师学院，广东 汕头 515000

5.东莞实业投资控股集团有限公司，广东 东莞 523000）

摘要：乳腺癌是危害女性健康的重要疾病，及时准确地诊断是降低乳腺癌死亡率的关键。为提高乳腺癌诊断的准确性，提出一种改进U-Net的X射线乳腺肿瘤区域分割方法。首先，融合空间注意力和通道注意力机制，构建平行注意力模块，以加强网络对关键特征的提取和识别能力；然后，将平行注意力模块与残差模块相结合，设计了残差平行注意力模块，提高U-Net模型的深层特征提取与高效聚焦能力；最后，将残差平行注意力模块引入U-Net模型的编码器部分，提高了U-Net模型对X射线乳腺肿瘤区域的分割精度。在CBIS-DDSM上的实验结果表明，改进U-Net模型的Dice系数和平均交并比分别达到94.20%和90.76%，有效提升了X射线乳腺肿瘤区域的分割精度。

关键词：图像分割；注意力机制；乳腺肿瘤；残差网络；U-Net

中图分类号：TP391.41           文献标志码：A          文章编号：1674-2605(2025)04-0004-07

DOI：10.12475/aie.20250404                                  开放获取

X-ray Breast Tumor Region Segmentation Method Using an Improved U-Net

CHEN Xiang1  CAI Yanguang1,2  GONG Guojun3  ZHANG Ruihu4  CAI Hao5

(1.College of Automation, Guangdong University of Technology, Guangzhou 510006, China  

2.School of Intelligent Manufacturing and Electrical Engineering, Guangzhou Institute of Science and Technology, Guangzhou 510540, China  3.Guangdong Lingnan First Technical College of Industry and Commerce, Guangzhou 510000, China  4.Guangdong Yuedong Technician College, Shantou 515000, China

5.Dongguan Industrial Investment Holding Group Co., Ltd., Dongguan 523000, China) 

Abstract: Breast cancer is a serious disease that threatens women's health, and timely and accurate diagnosis is crucial for reducing its mortality rate. To improve the accuracy of breast cancer diagnosis, this study proposes an enhanced U-Net-based method for segmenting breast tumor regions in X-ray images. First, a parallel attention module is constructed by integrating spatial and channel attention mechanisms to strengthen the network's ability to extract and identify key features. Next, the parallel attention module is combined with a residual module to design a residual parallel attention module, enhancing the U-Net model's deep feature extraction and high-efficiency focusing capabilities. Finally, the residual parallel attention module is incorporated into the encoder part of the U-Net model, improving its segmentation accuracy for X-ray breast tumor regions. Experimental results on the CBIS-DDSM dataset demonstrate that the improved U-Net model achieves Dice coefficients and mean intersection over union (mIoU) of 94.20% and 90.76%, respectively, significantly enhancing the segmentation accuracy of X-ray breast tumor regions.

Keywords: image segmentation; attention mechanism; breast tumor; residual network; U-Net