深度学习结合区域生长对MR图像胶质瘤的分割  

Deep Learning Combined with Region Growth for Segmentation of Glioma in MR Images

作　　者：李阳 宋悦 穆伟斌[1] 张淑丽[1] LI Yang;SONG Yue;MU Weibin;ZHANG Shuli(School of Medical Technology,Qiqihar Medical University,Qiqihar Heilongjiang 161003,China)

机构地区：[1]齐齐哈尔医学院医学技术学院,黑龙江齐齐哈尔161003

出　　处：《中国医疗设备》2022年第12期35-39,48,共6页China Medical Devices

基　　金：齐齐哈尔市科技计划联合引导项目(LHYD-202018);齐齐哈尔医学科学院临床科研基金项目(QMSI2019L-03)。

摘　　要：目的MR图像胶质瘤的精准分割是判定肿瘤范围和制定治疗方案的前提。为解决传统胶质瘤分割方法的过程中存在的复杂度高和精度低的问题,本文提出一种改进的U-Net网络与区域生长算法相结合的方法来分割MR图像胶质瘤。方法从公开数据库中下载胶质瘤的MR图像和手动分割标签。在U-Net网络的各层和桥中的2个卷积层间加入残差模块来改进网络,然后对网络分割结果做适度的区域生长操作来描述肿瘤的边界。使用Dice相似系数(Dice Similarly Coefficient,DSC)和边界F1(Boundary F1,BF)轮廓匹配分数(BF Score)等指标来评价本文方法的分割性能。结果在区域生长参数优化集中,区域生长的最大强度差异和种子点的灰度阈值为0.01和86时,分割结果达到最优。在包含了肿瘤所有层面的测试集中,DSC和BF Score分别达到了0.8332和0.7283。DSC得分相较于传统的FCN-8s和DeepLab v3+网络分别提高了7.43%和4.56%。结论改进的U-Net网络结合区域生长操作能很好地描述胶质瘤的位置、范围和边界信息,可用于辅助医生对胶质瘤进行定量分析。Objective Accurate segmentation of glioma in MR images is the premise of determining the scope of tumor and formulating treatment plans.In order to solve the problems of high complexity and low accuracy in the process of traditional glioma segmentation methods,this paper proposes a method that improves U-Net network combined with region growth algorithm to segment gliomas in MR images.Methods MR images and manual segmentation labels of glioma were downloaded from public database.The residual module was added between the two convolutional layers of each stage and bridge of the U-Net network to improve the network,and then a moderate region growth operation was performed on the network segmentation result to describe the boundaries of the tumor.Use indicators such as the dice similarly coefficient(DSC)and the boundary F1(BF)contour matching score(BF score)were used to evaluate the segmentation performance of the proposed method.Results In the dataset of optimized regional growth parameters,the segmentation results were optimal when the maximum intensity distance of regional growth and the gray threshold of the seed points were 0.01 and 86.In the test set,which included all layers of the tumor,the DSC and BF scores reached 0.8332 and 0.7283,respectively.Compared with the traditional FCN-8s and DeepLab v3+networks,the DSC score were improved by 7.43%and 4.56%,respectively.Conclusion The improved U-Net network combined with the region growing operation can well describe the location,scope and boundary information of glioma,which can be used to assist doctors in quantitative analysis of glioma.

关 键 词：磁共振图像 深度学习 区域生长 胶质瘤 分割 

分 类 号：R739.41[医药卫生—肿瘤] R197.39[医药卫生—临床医学]