改进BIRCH算法的MRI脑图像分割  

Research on MRI brain image segmentation based on improved BIRCH algorithm

作　　者：郑伟[1,2,3] 王洁 郝钰蓉 马泽鹏 ZHENG Wei;WANG Jie;HAO Yurong;MA Zepeng(College of Electronic Information Engineering,HeBei University,Baoding Hebei 071002,China;HeBei Key LaBoratory of Digital Medical Engineering,Baoding Hebei 071002,China;HeBei Machine Vision Engineering Technology Research Center,Baoding Hebei 071002,China;Affiliated Hospital of HeBei University,Baoding Hebei 071000,China Abstrac)

机构地区：[1]河北大学电子信息工程学院,保定071002 [2]河北省数字医疗工程重点实验室,保定071002 [3]河北省机器视觉工程技术研究中心,保定071002 [4]河北大学附属医院,保定071000

出　　处：《激光杂志》2022年第1期184-191,共8页Laser Journal

基　　金：河北省自然科学基金(No.F2020201025,No.H2020201021);河北省人力资源与社会保障厅留学回国人员项目(No.606999919029);河北大学高性能计算平台支持。

摘　　要：针对现有磁共振常规扫描序列对于颅脑白质、灰质信号相近分辨不清,解剖病变欠佳,难以达到临床高精准诊断的需求,选用改进的BIRCH算法,首先将3维MRI体数据经过预处理,由灰度与梯度组成特征向量,然后利用Cophenet相关系数,确定最优参数——分支因子B、阈值T,最后通过定义可调节线段L,改进原BIRCH算法仅将数据样本点到质心的平均距离作为半径R的局限性。仿真实验表明,提出的改进BIRCH算法,与已有BIRCH算法相比,聚类指标FMI值与RI值指数分别达到0.754 5与0.542 1,分别提升了2.79%与1.42%,并于其他聚类算法比较,所提算法性能表现仍为最优,脑WM、GM、CSF的组织分割精度Dice指数分别为0.939 4、0.834 2、0.853 1,Hausdorff距离分别为14.988 1、12.964 2、13.601 5,所提算法可为临床医学提供一定帮助。Aiming at the existing conventional MRI scan sequence for the close resolution of cranial white matter and gray matter signals, and poor anatomical lesions, it is difficult to meet the needs of high-precision clinical diagnosis, choose the improved BIRCH algorithm, first of all, the 3 D MRI voxel after pretreatment, consists of gray level and gradient feature vector, then using Cophenet correlation coefficient, to determine the optimal parameters-branch factor B and threshold T, finally by defining adjustable line L, improved the original BIRCH algorithm only data sample points to the average distance of the centroid as the limitations of radius R. Simulation experiments show that the proposed improved BIRCH algorithm, compared with the existing BIRCH algorithm, clustering index values of FMI and RI values were 0.754 5 and 0.542 1, respectively increased by 2.79% and 1.42%, and in other clustering algorithm, the proposed algorithm is still for optimal performance, brain tissue segmentation accuracy of the WM, GM and CSF Dice index were 93.94%, 83.42% and 85.31%, respectively, of 2.35%, 0.49% and 2.86%.The proposed algorithm can provide some help for clinical medicine.

关 键 词：MRI图像分割 层次聚类 BIRCH算法 Cophenet相关系数 

分 类 号：TN209[电子电信—物理电子学]