Head Tissue Template

MethodThe method of Brudfors et. al. (2020) was used to construct a tissue probability map from T2-weighted and PD-weighted scans of the first 64 subjects from the IXI dataset, along with the T1-weighted scans of the next 64 IXI subjects. The 15 training subjects' scans from the MICCAI Challenge Dataset were also included in the template construction. Default settings were used throughout, except for the regularisation for the diffeomorphic registration, which was set to be higher than the default settings (``Shape Regularisation'' on the user interface was set to [0.0001 0.5 0.5 0.0 1.0]). After merging several of the automatically identified tissue classes, the tissue probability map has 1 mm isotropic resolution, dimensions of 191x243x229 voxels and consists of 11 tissue types, three of which approximately correspond with brain tissues.This image (roughly) contains the logarithms of the tissue probabilities, which can be recovered using a softmax (exp(mu)/(sum(exp(mu)+1)).ReferenceBrudfors M, Balbastre Y, Flandin G, Nachev P, Ashburner J. Flexible Bayesian Modelling for Nonlinear Image Registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention 2020 Oct 4 (pp. 253-263). Springer, Cham.方法：本研究采用Brudfors等人（2020）提出的方法，基于IXI数据集（IXI dataset）前64名受试者的T2加权扫描图像（T2-weighted scans）与PD加权扫描图像（PD-weighted scans），以及后续64名IXI受试者的T1加权扫描图像（T1-weighted scans），构建组织概率图谱。同时将MICCAI挑战赛数据集（MICCAI Challenge Dataset）的15名训练受试者的扫描图像纳入模板构建流程。实验全程采用默认参数，仅针对同胚配准（diffeomorphic registration）的正则化项进行调整：将用户界面中的「形状正则化（Shape Regularisation）」参数设置为[0.0001 0.5 0.5 0.0 1.0]，该值高于默认配置。经合并若干自动识别的组织类别后，所得组织概率图谱分辨率为1mm各向同性，维度为191×243×229体素，共包含11种组织类型，其中3种大致对应脑组织。该图谱近似存储了组织概率的对数值，可通过softmax函数（exp(μ)/(sum(exp(μ)) + 1)）还原得到原始组织概率。参考文献：Brudfors M, Balbastre Y, Flandin G, Nachev P, Ashburner J. 面向非线性图像配准的柔性贝叶斯建模. 见：国际医学图像计算与计算机辅助干预大会（International Conference on Medical Image Computing and Computer-Assisted Intervention）2020年10月4日会议论文集（第253-263页）. Cham：Springer出版社.