文章摘要

文章信息/Info

Title:: Feature guided segmentation algorithm of mammograms fusion with self attention

作者:: 申文静1; 2; 丛金玉1; 2; 班楷第1; 2; 王苹苹1; 2; 刘坤孟1; 2; 司兴勇1△; 魏本征1; 2△; （1.山东中医药大学医学人工智能研究中心，青岛266112；2.山东中医药大学青岛中医药科学院，青岛 266112）

Author(s):: SHEN Wenjing1; 2; CONG Jinyu1; 2; BAN Kaidi1; 2; WANG Pingping1; 2; LIU Kunmeng1; 2; SI Xingyong1; WEI Benzheng1; 2; （1.Center for Medical Artificial Intelligence,Shandong University of Traditional Chinese Medicine,Qingdao 266112,China;2.Qingdao Academy of Chinese Medicine Sciences,Shandong University of Traditional Chinese Medicine,Qingdao 266112）

Keywords:: Breast cancer; Mammograms; Image segmentation; Self attention; Feature guided

摘要:: 为提高对乳腺癌钼靶图像中病灶区域的识别精度，本研究设计了一种面向乳腺肿块和钙化区域分割的特征引导注意网络。首先，该网络通过特征提取模块学习乳腺组织的语义特征；其次，利用融合自校正注意力的解码模块，增强对病灶区域边缘信息的关注度，提高边界的清晰度；最后，采用特征引导注意模块增强通道的依赖关系，进一步还原病灶区域边缘细节，提高分割精度。实验结果表明，本研究网络在扩充后的INBreast1数据库中肿块和钙化分割的平均骰子系数（mDice）分别达到了0.971和0.888，在DDSM数据库肿块分割的mDice达到了0.911，优于其他常规的分割模型，对乳腺癌的早期诊断和治疗具有重要意义。

Abstract:: In order to enhance the recognition accuracy of breast cancer mammography，we designed a feature guided attention network for the segmentation of breast mass and calcification areas. Firstly, the feature extraction module was designed to learn semantic features of breast tissue.Then, the decoding module integrating self correcting attention was used to enhance attention to the edge information of the lesion area, and improve the clarity of the boundary. Finally, feature guided attention module was used to enhance channel dependencies, further restore edge details of the lesion area, and improve segmentation accuracy. The experimental results showed that the average Dice coefficient (mDice) of mass and calcification segmentation on the expanded INBreast1 reached 0.971 and 0.888 respectively,the mDice of mass segmentation on DDSM reached 0.911, which was better than that of other conventional segmentation models.The research is of great significance for early diagnosis and treatment of breast cancer.