三维水平集智能分割算法在梅尼埃病患者内耳道磁共振图像分割中的应用效果  

Segmentation effect of IS3DLS algorithm on MRI images of internal auditory canal in patients with Meniere's disease

作　　者：滕腾 陀芳 葛洪洲 TENG Teng;TUO Fang;GE Hongzhou(Department of Otolaryngology,Qingdao Hospital of Traditional Chinese Medicine,Qingdao 266034,China;不详)

机构地区：[1]青岛市中医医院耳鼻咽喉科,山东青岛266034 [2]青岛市中医医院影像科

出　　处：《山东医药》2021年第26期25-29,共5页Shandong Medical Journal

摘　　要：目的探讨三维水平集智能分割(IS3DLS)算法对梅尼埃病患者内耳道磁共振成像(MRI)图像的分割效果。方法选择68例梅尼埃病患者,采用MRI系统进行内耳检查,对内耳MRI图像分别采取专家手工分割算法、区域生长水平集分割算法、IS3DLS算法进行图像分割处理。①手工分割算法:由专家直接在MRI原始图像上画出感兴趣区(ROI)的边缘,分割目标。②区域生长水平集分割方法:在内耳ROI内设置种子点,调至最优参数,获得最佳区域生长分割结果,去除无关的组织,获得预分割结果作为初始轮廓掩膜,水平集演化分割内耳。③任选28组内耳图像数据,由耳科专家手动分割获得三维内耳体素模型,用3D-Slicer软件获得形状训练样本。另取6组内耳图像数据作为测试样本,建立内耳统计形状模型,计算平均形状模型。对图像进行加权,采用配准技术自动定位和获取内耳ROI,对内耳ROI和平均形状模型进行配准,得到初始轮廓。采用IS3DLS对初始轮廓和特征图像进行分割。以专家手工分割算法为标准,采用马修斯相关系数(MCC)、戴斯相似系数(DSC)、假阳性率(FPR)和假阴性率(FNR)计算IS3DLS算法、区域生长水平集分割算法与专家手工分割算法的相似性和准确性。以真位置(TP)和假位置(FP)进行IS3DLS算法、区域生长水平集分割算法两种分割算法的误差比较,记录IS3DLS和区域生长集分割算法的运行时间并比较。结果IS3DLS算法的MCC、DSC、FPR、FNR分别为0.9599、0.9594、0.0325、0.0365,与专家手工分割算法相当(P<0.05);与专家手工分割算法相比,IS3DLS算法的MCC高于区域生长集分割算法,提示其与专家手工分割算法的结果更接近;IS3DLS算法的FPR、FNR均低于区域生长集分割算法;IS3DLS算法和区域生长集分割算法均具有较高的TP值;IS3DLS算法的FP值、HD值均小于区域生长集分割算法(P<0.05);IS3DLS的运行时间明显少于区域生长集分割(Objective To investigate the segmentation effect of three-dimensional level set intelligent segmentation(IS3DLS)algorithm on magnetic resonance imaging(MRI)images of internal auditory canal in patients with Meniere's dis⁃ease.Methods Sixty-eight patients with Meniere's disease were selected and the inner ear was examined by MRI sys⁃tem.The inner ear MRI images were segmented by expert manual segmentation algorithm,region growth level set segmen⁃tation algorithm,and IS3DLS algorithm,respectively.①Manual segmentation algorithm:the experts directly drew the edge of the region of interest(ROI)on the original MRI image and segmented the target.②Region growing level set seg⁃mentation method:seed points were set in the inner ear ROI,and were adjusted to the optimal parameters to obtain the best region growth segmentation results;the irrelevant tissues were removed,the pre-segmentation results were obtained as the initial contour mask,and the level set evolution was used to segment the inner ear.③Twenty-eight groups of inner ear image data were selected,and the three-dimensional inner ear voxel model was manually segmented by otologists,and the shape training samples were obtained by 3D-Slicer software.Another 6 groups of inner ear image data were taken as test samples,the statistical shape model of inner ear was established,and the average shape model was calculated.The image is weighted,and the inner ear ROI was automatically located and obtained by registration technology.The inner ear ROI and the average shape model were registered to get the initial contour.IS3DLS was used to segment the initial contour and feature image.Taking the expert manual segmentation algorithm as the standard,Matthews correlation coefficient(MCC),Deiss similarity coefficient(DSC),false positive rate(FPR)and false negative rate(FNR)were used to calculate the sim⁃ilarity and accuracy of IS3DLS algorithm,region growing level set segmentation algorithm and expert manual segmentation algorithm.Real position(TP)and false position(F

关 键 词：梅尼埃病 磁共振成像 图像分割 三维水平集智能分割算法 专家手工分割算法 区域生长水平集分割算法 

分 类 号：R764.3[医药卫生—耳鼻咽喉科]