融合多尺度特征和注意力机制的超声甲状腺结节分割  

Ultrasound image segmentation of thyroid nodule combining multi-scale features and attention mechanism

作　　者：赵欣[1] 黎红豆 王洪凯[2] ZHAO Xin;LI Hongdou;WANG Hongkai(School of Information Engineering,Dalian University,Dalian 116622,Liaoning,China;School of Biomedical Engineering,Dalian University of Technology,Dalian 116604,Liaoning,China)

机构地区：[1]大连大学信息工程学院,辽宁大连116622 [2]大连理工大学生物医学工程学院,辽宁大连116024

出　　处：《声学技术》2024年第5期668-676,共9页Technical Acoustics

基　　金：国家自然科学基金(81971693)。

摘　　要：针对目前超声影像下甲状腺结节分割不够精准的问题,提出一种融合多尺度特征和注意力机制的超声甲状腺结节分割方法。该模型编码设计了多感受野通道选择模块,通过核心选择注意力对多个不同感受野的特征进行自适应加权组合,使包含目标的感受野通道占据主导。同时,设计自适应全局上下文模块自适应地提取瓶颈层多个尺度的全局上下文特征,以实现对瓶颈层高级语义的有效编码。此外,设计双注意力引导模块增强编解码器对等层之间的特征融合,以减少上采样过程中的信息损失。在公开的超声甲状腺结节数据集上进行实验,结果表明,文中所提方法优于其他对比网络,能更加精准地分割出甲状腺结节,有效提升了甲状腺结节的分割性能。A new ultrasound thyroid nodule segmentation model with multi-scale features and attention mechanisms is proposed to address the problem of inaccurate segmentation of thyroid nodules in ultrasound images.In the model,a multi receptive field channel selection module designed in the encoding phase is used to adaptively weight and combine features from different receptive fields with a focus on channels containing the target object.Additionally,an adaptive global context module is designed to extract global context features from multiple scales of the bottleneck layer to effectively encode high-level semantic information.Furthermore,a dual attention-guided module is used to enhance feature fusion between peer layers of the encoder and decoder for reducing information loss during upsampling.Experimental results on a publicly available ultrasound thyroid nodule dataset show that the proposed method outperforms other state-of-the-art networks and achieves more accurate segmentation of thyroid nodules,effectively improving segmentation performance.

关 键 词：深度学习 甲状腺结节 超声图像分割 多尺度特征提取 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]