基于相似注意力的医学超声影像协同分割  

Medical Ultrasound Image Co-Segmentation Based on Similar Attention

作　　者：叶浩然 陈芳 谢彦廷 万鹏[1] YE Haoran;CHEN Fang;XIE Yanting;WAN Peng(Department of Computer Science and Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing Jiangsu 210016,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京210016

出　　处：《中国医疗设备》2023年第3期14-20,共7页China Medical Devices

基　　金：国家自然科学基金(U20A20389,61901214);中国博士后科学基金(2021T140322,2020M671484)。

摘　　要：目的探究医学超声影像研究可准确分割组织图像的自动分割算法。方法提出一种端到端协同分割网络模型,其由孪生编解码器、相似性学习模块和注意力学习模块组成。对输入的一对超声图片先经过编码器提取特征,由相似性学习模块优化特征,再由注意力模块增强特征,最后通过解码器输出对组织前景的自动预测结果,并在包含了3类组织(胎儿、甲状腺、乳腺)的多类别超声图像数据集上对本文提出的方法进行了验证。结果针对像素准确度、精确率和Jaccard相似系数这3种评价指标,分别取得了97.25%的像素准确度,94.51%的精确率和0.90的Jaccard相似系数,并进一步通过对比实验和消融实验验证了本文所提出方法的准确性和有效性。结论提出的协同分割模型对多类别医学超声影像均有较高的分割精度,能够快速准确获得组织区域,有助于计算机辅助诊断。Objective To explore the automatic segmentation algorithm for tissue image segmentation in medical ultrasonography.Methods This paper proposed an an end-to-end co-segmentation model,which consisted of encoder-decoder,similarity learning module and attention learning module.The features of a pair of input ultrasound images were extracted by an encoder,optimized by a similarity learning module,enhanced by an attention module.Finally,the decoder output the automatic prediction of the tissue prospect.The proposed method was validated on a multi-category ultrasound image dataset containing three types of medical ultrasound images(fetus,thyroid and breast).Results For pixel accuracy,accuracy and Jaccard similarity coefficient,the pixel accuracy of 97.25%,accuracy of 94.51%and Jaccard similarity coefficient of 0.90 were obtained,respectively.Further,the accuracy and effectiveness of the proposed method were verified by comparison and ablation experiments.Conclusion The proposed cosegmentation model has high segmentation accuracy for multiple types of medical ultrasound images,and can quickly and accurately obtain tissue regions,which is helpful for computer-aided diagnosis.

关 键 词：医学超声影像 协同分割 相似注意力 自动分割算法 

分 类 号：R197.39[医药卫生—卫生事业管理] TP391.4[医药卫生—公共卫生与预防医学]