基于SRL-UNet的肾脏病变图像分割算法  

Renal lesion image segmentation algorithm based on SRL-UNet

作　　者：史健婷 李雪瑶 李志军[3] SHI Jianting;LI Xueyao;LI Zhijun(School of Computer and Information Engineering,Heilongjiang University of Science and Technology,Harbin 150022,China;Graduate College,Heilongjiang University of Science and Technology,Harbin 150022,China;Guangxi Key Laboratory of Machine Vision and Intelligent Control,Wuzhou University,Wuzhou 543002,Guangxi,China)

机构地区：[1]黑龙江科技大学计算机与信息工程学院,哈尔滨150022 [2]黑龙江科技大学研究生学院,哈尔滨150022 [3]梧州学院广西机器视觉与智能控制重点实验室,广西梧州543002

出　　处：《智能计算机与应用》2024年第11期80-87,共8页Intelligent Computer and Applications

基　　金：2023年度黑龙江省属高校基本科研业务费项目(2023-KYYWF-0547)。

摘　　要：肾脏的病变通常隐匿无明显症状,随着时间推移,还可能导致并发症的出现。由于肾脏产生病变原因复杂多样,且造成病变因素众多,因此针对病变肾脏的准确诊断至关重要。精确的病变肾脏图像分割对临床诊断具有重要价值,为此提出一种基于改进UNet的肾脏病变图像分割算法SRL-UNet,用于自动检测和分割肾脏中的病变区域。通过在UNet模型中引入SE注意力机制,并且使用ResNet50作为UNet的主干网络,然后将UNet和水平集方法结合生成混合损失函数,实现网络端到端的训练来提高模型泛化性能,增强图像边缘的分割能力。仿真结果表明,SRL-UNet算法得到的准确率为0.9933,MIoU指数为0.8127,与UNet、DeepLabv3和HRNet三种方法比较,准确率分别提高了0.61%、0.35%和0.24%;MIoU指数分别提高了18.21%、4.04%和10.02%。同时,也分别解决了3种方法产生的空洞问题、分割缺失问题以及边缘不光滑问题,达到了良好的分割效果,体现出所提方法在分割算法性能上的优势。Renal lesions are usually asymptomatic and may lead to complications over time.Due to the complex and diverse causes of kidney disease,as well as the numerous factors that cause it,accurate diagnosis of diseased kidneys is crucial.Accurate segmentation of diseased kidney images is of great value for clinical diagnosis.Therefore,a kidney disease image segmentation algorithm SRLUNet based on improved UNet is proposed for automatic detection and segmentation of diseased areas in the kidney.By introducing SE attention mechanism into the UNet model and using ResNet50 as the backbone network,and combining UNet and level set methods to generate a mixed loss function,end-to-end training of the network is achieved to improve the model's generalization performance and enhance the segmentation ability of image edges.The results show that the accuracy obtained by the SRL-UNet algorithm is 0.9933,with a MIoU index of 0.8127.Compared with UNet,DeepLabv3,and HRNet methods,the accuracy has been improved by 0.61%,0.35%,and 0.24%,respectively;The MIoU index increased by 18.21%,4.04%,and 10.02%,respectively.At the same time,the three methods have also solved the problems of voids,missing segmentation,and uneven edges,achieving good segmentation results,reflecting the advantages of the proposed method in segmentation algorithm performance.

关 键 词：图像分割 UNet 深度学习 水平集方法 病变肾脏 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]