多尺度非对称卷积的轻量级U2-Net医学影像语义分割模型  

Lightweight U2-Net semantic segmentation model for medical images with multiscale asymmetric convolution

作　　者：孙水发[1,2] 王清华 邹耀斌 唐庭龙[3] 侯斌[3] 吴义熔 崔文超[3] SUN Shuifa;WANG Qinghua;ZOU Yaobin;TANG Tinglong;HOU Bin;WU Yirong;CUI Wenchao(School of Electrical Engineering and New Energy,China Three Gorges University,Yichang 443002,China;School of Information Science and Technology,Hangzhou Normal University,Hangzhou 310036,China;School of Computer and Information Technology,China Three Gorges University,Yichang 443002,China;Institute of Advanced Studies in Humanities and Social Sciences,Beijing Normal University,Zhuhai 519087,China)

机构地区：[1]三峡大学电气与新能源学院,湖北宜昌443002 [2]杭州师范大学信息科学与技术学院,杭州310036 [3]三峡大学计算机与信息学院,湖北宜昌443002 [4]北京师范大学人文和社会科学高等研究院,广东珠海519087

出　　处：《重庆理工大学学报（自然科学）》2024年第11期138-146,共9页Journal of Chongqing University of Technology：Natural Science

基　　金：国家社会科学基金项目(20BTQ066)。

摘　　要：基于U2-Net网络结构,设计了一种运行更高效、分割更准确的医学影像语义分割模型。通过多尺度非对称卷积核代替传统的注意力机制以及降低原U2-Net网络的层数,减少模型参数量;通过改变U2-Net网络连接方式,使用U-Net++网络的跳跃连接方式,使模型在传递特征信息时保持完整性,减少信息损失,使分割边缘更加精确且连续;根据正负样本不均衡和难易不同的问题设计了避免在训练过程中大量简单负样本占据主导地位的二分类交叉熵损失函数(BCE Loss),更倾向于挖掘前景区域的骰子损失函数(Dice Loss)以及更偏向于两图的结构相似性的多层级结构相似性损失函数(MS-SSIM Loss)的组合损失函数,用来监督网络优化。实验结果表明:所提算法在DRIVE、STARE数据集上比现有最先进网络模型(SOTA)的F1分数分别提高2.6%、1.4%,在ISIC-2018数据集上比SOTA的DSC指标提高2.6%。对分割结果进行可视化后表明,网络在样本较小的情况下可以充分提取样本信息,提高语义分割效果。In clinical practice,semantic segmentation of medical images plays a vital role in detecting diseases,allowing doctors to accurately determine the patients’conditions and make more targeted treatment plans.Based on the U2-Net network structure,we designed a semantic segmentation model for medical images with more efficient operation and more accurate segmentation.The number of model parameters was reduced by replacing the traditional attention mechanism with a multi-scale asymmetric convolution kernel as well as by reducing the number of layers of the original U2-Net network.By changing the connection method of the U2-Net network and using the hopping connection of the U-Net++network,the model was made to pass the feature information to maintain integrity,reduce information loss,and make the segmentation edges more accurate and continuous.Considering the imbalance of positive and negative samples and other difficulties,we designed the binary cross entropy loss function(BCE Loss)to avoid the dominance of a large number of simple negative samples in the training process,the dice loss function(Dice Loss)to excavate foreground regions,and the multiple loss function to favor the structural similarity of the two graphs.Structural similarity of the two graphs(MS-SSIM Loss),a combined loss function of the multilevel structural similarity loss function(MS-SSIM Loss),is employed to supervise network optimization.Our experimental results show our algorithm improves the F1 score by 2.6%and 1.4%over the existing state-of-the-art network model(SOTA)on the DRIVE and STARE datasets and improves the DSC metric by 2.6%on the ISIC-2018 dataset.Visualization of the segmentation results indicate the network fully extracts the sample information and improves the semantic segmentation effect in the case of smaller samples.

关 键 词：语义分割 医学影像 非对称卷积核 U2-Net网络 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]