基于DRA-UNet模型的超声图像分割方法  

Ultrasonic image segmentation method based on DRA-UNet model

作　　者：王雷[1] 郭新萍 王钰帏 李彬[2] WANG Lei;GUO Xinping;WANG Yuwei;LI Bin(School of Computer Science and Technology,Shandong University of Technology,Zibo 255000,Shangdong China;School of Automation Science and Engineering,South China University of Technology,Guangzhou 510641,China)

机构地区：[1]山东理工大学计算机科学与技术学院,山东淄博255000 [2]华南理工大学自动化科学与工程学院,广东广州510641

出　　处：《华中科技大学学报（自然科学版）》2024年第5期83-89,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金青年基金资助项目(61502282);国家自然科学基金面上资助项目(62273155);山东省自然科学基金面上资助项目(ZR2021MF017).

摘　　要：针对普通卷积运算无法关注重点区域、编码器无法有效提取全局上下文信息、简单的跳跃连接无法捕获显著特征,以及易导致分割图像分辨率降低、重要细节丢失、小物体信息无法被准确捕获等问题,提出基于膨胀率注意力机制的UNet(DRA-UNet)模型,并发展了基于此模型的超声图像分割方法.在UNet模型的基础上,引入膨胀率注意门和多尺度卷积(ConvMulti)模块.膨胀率注意门模块利用空洞卷积能得到更大的感受野,将编码器语义位置的局部区域像素联合到上采样区域,可以实现更加高效的跳跃连接.ConvMulti模块用来获取更加详细的高层特征信息,使编码器功能更强大.实验结果表明:本模型可以有效抑制图像噪声,大幅提高特征的表达能力,具有很强的鲁棒性,相比六种经典分割方法,所提出方法在交并比、F1分数和精度指标下分别达到72.25%,83.89%和97.47%.Aiming at the problems that the ordinary convolution operation cannot focus on the key areas,the encoder cannot effectively extract the global context information,and the simple skip connection cannot capture the salient features,so that it is easy to cause that the segmentation image resolution is reduced,the important details are lost,and the small object information cannot be accurately captured,a UNet(DRA-UNet)model based on the expansion rate attention mechanism was proposed,and an ultrasonic image segmentation method based on this model was developed.On the basis of the UNet model,the expansion rate attention gate and the multi-scale convolution(ConvMulti)module were introduced.The expansion rate attention gate module used dilated convolution to obtain a larger receptive field,and combined the local region pixels of the encoder's semantic position into the upsampling region to achieve more efficient skip connections.The ConvMulti module was used to obtain more detailed high-level feature information to make the encoder more powerful.Experimental results show that the proposed model can effectively suppress image noise,and greatly improve the expression ability of features with strong robustness.Compared with the six classical segmentation methods,the proposed method achieves 72.25%,83.89%and 97.47%under the intersection over union,F1-value and accuracy,respectively.

关 键 词：超声图像 图像分割 U-Net模型 空洞卷积 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]