摘要
为提高乳腺超声图像中肿瘤区域的分割精度并降低计算复杂度,提出一种融合十字形窗口Transformer(CSWin Transformer)与聚焦线性注意力(FLA)的分割模型FLA-CSWin-U-Net。该方法以U-Net为基本架构,编码器采用改进的聚焦线性注意力十字形窗口Transformer(FLA-CSWin Transformer)模块,增强全局上下文建模能力;引入聚焦线性注意力机制,强化关键区域特征交互,同时保持线性计算复杂度;解码器通过动态上采样(DySample)算子来提升细节还原效率。在公共数据集--BUSI数据集上的实验表明,所提模型Dice系数达到94.3%,较传统U-Net提升11.07%,参数量仅为23.06 M,计算量降至4.09 GFLOPs,使模糊边界与小病灶的分割效果得到显著改善,具有较高的临床实用价值和部署可行性。
To improve the segmentation accuracy of tumor regions and reduce computational complexity in breast ultrasound images,a segmentation model FLA-CSWin-U-Net that integrates cross-shaped window Transformer(CSWin Transformer)with focused linear attention(FLA)is proposed.This method adopts U-Net as its basic architecture.The encoder utilizes an improved focused linear attention cross-shaped window Transformer(FLA-CSWin Transformer)module,which enhances the global context modeling capability.A focused linear attention mechanism is introduced to enhance feature interaction in key regions while maintaining linear computational complexity.The decoder employs dynamic upsampling(DySample)operator to improve the efficiency of detail restoration.Experimental results on the public BUSI dataset demonstrate that the proposed model achieves a Dice coefficient of 94.3%,which is an improvement of 11.07%over the traditional U-Net.With a parameter count of 23.06 M and a computational load reduced to 4.09 GFLOPs,the model significantly enhances segmentation effect of blurred boundaries and small lesions,and has high clinical practical value and deployment feasibility.
作者
王道荣
杨录
刘康驰
WANG Daorong;YANG Lu;LIU Kangchi(State Key Laboratory of Electrical Testing,School of Information and Communication Engineering,North University of China,Taiyuan 030051,China)
出处
《传感器与微系统》
北大核心
2026年第4期29-33,共5页
Transducer and Microsystem Technologies
基金
国家自然科学基金青年科学基金资助项目(62401525)
山西省高等学校科技创新项目(2024L209)。
