Automatic infant brain segmentation based on diffusion tensor imaging

，

(1.Department of Biomedical Engineering, School of Basic Medical Science, Central SouthUniversity,Changsha 410083, China; 2.School of Clinical Medicine, ShaoyangUniversity, Shaoyang 422000, China)

Automatic infant brain segmentation based on diffusion tensor imaging

WANGLin1,2，YUGang1*

(1.DepartmentofBiomedicalEngineering,SchoolofBasicMedicalScience,CentralSouthUniversity,Changsha410083,China; 2.SchoolofClinicalMedicine,ShaoyangUniversity,Shaoyang422000,China)

Objective To investigate the feasibility and value of automatic segmentation of infant brain images based on diffusion tensor image (DTI). Methods A method of segmentation of infant brain based on DTI images was proposed. The method was mainly included two stages： ①Extracting the cerebrospinal fluid (CSF) using the distribution of water; ②Extracting the white matter (WM) adopting the anisotropic diffusion of water in neurons, followed by distinguishing the gray matter (GM) component. Results Through the feature selection method designed in this study, the effective DTI feature combination was selected. The first step was to extract CSF with mean diffusity (MD) and the third eigenvalue (L3), and the second step was to extract WM and GM with fractional anisotropy (FA) and the L3. The highest average similarity was obtained by the two steps. The two-step segmentation could be successfully performed in infant brain image segmentation and satisfied with the split effect. Conclusion The automatic segmentation of infant brain based on DTI in this study is reasonable and feasible, and has high segmentation accuracy.

Infant； Child； Magnetic resonance imaging； Diffusion tensor imaging； Brain segmentation

精确、自动地大脑分割技术可应用于婴幼儿大脑损伤、脑功能重塑、大脑发育的相关认知科学研究。尽管常规MR图像在形态上的分辨率较高，但仍存在诸多缺陷，如偏场、噪声和容积效应等，对大脑分割提出很大挑战。大脑分割包括参数化及非参数化方法。多数参数化方法假设脑MR图像服从高斯分布，运用最大后验概率、最大似然函数或最大期望法进行算法优化。为降低噪声，这些方法经常与分割框架相结合，如局部统计[1]、水平集[2]、马尔科夫[3]、隐藏马尔科夫链[4]等。非参数方法，如模糊分割[5]，其优点是可收敛到稳定解。其中，隐马尔科夫随机场(hidden Markov random field， HMRF)与最大期望法具有良好的抗噪性能，能同步估计MR的偏场，是当前主流的大脑分割方法[6]。然而，现有的大脑图像分割方法所应用的先验知识几乎均基于成人大脑，且输入数据多采用T1或T2加权的结构图像。生长发育过程中婴幼儿大脑形态和功能时刻发生变化，常规MR图像对这些信息并不敏感。同时，由于缺少理想的婴幼儿大脑图谱，无法依据图谱改善分割，采用成人大脑分割方法分割婴幼儿大脑图像误差较大。近年来，Dai等[7]提出婴幼儿大脑的纵向分割方法，该方法对数据集要求高，需要不同时间点(如0～2岁)多次重复采集的大脑图像来改善分割结果，同时以结构图像为主，仅利用扩散张量成像(diffusion tensor imaging， DTI)的各向异性分数(fractional anisotropy， FA)图像，因此不适合通常的婴幼儿数据集。DTI能够跟踪大脑水分子的扩散，相对于单纯结构图像可更敏感地反映大脑连接的细微改变。……