分层Transformer和多尺度并行聚合的乳腺肿块分割方法  

Hierarchical Transformer with Multi-Scale Parallel Aggregation for Breast Tumor Segmentation

作　　者：夏平[1,2] 王雨蝶 雷帮军[1,2] 彭程[1,2] 张光一 唐庭龙 Xia Ping;Wang Yudie;Lei Bangjun;Peng Cheng;Zhang Guangyi;Tang Tinglong(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,Three Gorges University,Yichang 443002,Hubei,China;College of Computer and Information Technology,Three Gorges University,Yichang 443002,Hubei,China)

机构地区：[1]三峡大学水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《激光与光电子学进展》2025年第2期252-263,共12页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(U1401252);湖北省重点实验室开放基金(2018SDSJ07)。

摘　　要：针对乳腺超声图像分割中肿块与正常组织间对比度低、边界模糊、肿块形状和位置复杂、图像噪声大等问题,提出了一种分层Transformer与多尺度并行聚合的网络用于乳腺肿块分割。编码器采用MiT-B2,建立长距离特征依赖关系,有效提取不同分辨率下的特征信息;编码器与解码器间的跳跃连接构建多层分支感受野与置换注意力机制的级联模块;构建多层分支感受野用以捕获肿块多尺度局部信息,解决肿块与周围正常组织间高度相似的问题;置换注意力机制准确识别、定位肿块,并抑制噪声干扰;解码端构建聚合模块逐层融合并行分支特征,提升分割精度。在BUSI数据集上实验结果表明,相较于基线TransFuse网络,本文模型的Dice、交并比指标分别提升了3.21%、3.19%。在其他2个数据集上,本文模型的分割性能也有优良的表现。The problems of breast tumor segmentation from ultrasound images,such as low contrast between the tumor and the normal tissue,blurred boundaries,complex shapes and positions of tumors,and high noise in images,are a concern for researchers.This paper presents a hierarchical transformer with a multiscale parallel aggregation network for breast tumor segmentation.The encoder uses MiT-B2 to establish long-range dependencies and effectively extract features at different resolutions.At the skip connection between the encoder and the decoder,a cascaded module incorporating a multi-scale receptive field block and shuffle attention(SA)mechanism is constructed.receptive field block is used to capture multi-scale local information of the tumor,addressing the problem of high similarity between the lesion and surrounding normal tissue.The SA mechanism accurately identifies and localizes tumors while suppressing noise interference.In the decoder,an aggregation module is constructed to progressively fuse features from parallel branches to enhance segmentation accuracy.The experimental results on the BUSI dataset show that,compared to TransFuse,the proposed model achieves improvements of 3.21%and 3.19%in the Dice and intersection over union metrics,respectively.The model also shows excellent results for the other two datasets.

关 键 词：深度学习 乳腺肿块分割 分层Transformer 多尺度并行聚合模块 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]