The cerebral atlas of diffusion tensor magnetic resonance image (DT-MRI, shorted as DTI) is one of the key issues in neuroimaging research. It is crucial for comparisons of neuronal structural integrity and connectivity across populations. Usually, the atlas is constructed by iteratively averaging the registered individual image. In tradition, the fuzzy group average image is easily generated in the initial stage, which is harmful to providing clear guidance for subsequent registration, to the performance of the final atlas. To solve this problem, an improved unbiased DTI atlas construction algorithm based on adaptive weights is proposed in this paper. The adaptive weighted strategy based on diffeomorphic deformable tensor registration is introduced. At the same time, the distance measure for tensors is used as a constraint condition, which ensures the unbiasedness of the atlas. Then, using 77 DTIs from the dataset in http://www.brain-development.org, three study-specific atlases, i.e. the constructed atlases of the proposed algorithm and two open-sourced algorithms (DTIAtlasBuilder and DTI-TK), are compared with two standardized atlases (IIT v. 4.1 and NTU-DSI-122-DTI). The performances of the atlases were evaluated in spatial normalization way with six region-based criteria (including Euclidean distances between diffusion tensors, Euclidean distances of the deviatoric tensors, standard deviation, overlaps of eigenvalue-eigenvector, cross-correlations and three sets angles of eigenvalue-eigenvector pairs between diffusion tensors) and three fiber-based criteria (including distances between fiber bundles, angles between fiber bundles and fiber property profile-based criteria). The experimental results showed that the overall performances of the study-specific atlases are better than those of the standardized atlases for specific datasets, and the comprehensive performance of the improved algorithm proposed in this paper is the best.

弥散张量磁共振图像脑图谱(DT-MRI，简称DTI)是神经影像学研究的关键问题之一。这对于比较不同人群的神经元结构完整性和连通性至关重要。通常，图集是通过对配准的单个图像进行迭代平均来构建的。传统方法在初始阶段容易产生模糊的组平均图像，不利于为后续配准提供清晰的指导，不利于最终图集的表现。针对该问题，本文提出了一种改进的基于自适应权值的无偏DTI图谱构建算法。介绍了基于微分同胚可变形张量配准的自适应加权策略。同时，以张量的距离度量作为约束条件，这保证了图谱的无偏性。然后，使用来自 http://www.brain-development.org 数据集的 77 个 DTI，三个研究特定的图集，即所提出算法的构建图集和两个开源算法（DTIAtlasBuilder 和 DTI-TK），是与两个标准化地图集（IIT v. 4.1 和 NTU-DSI-122-DTI）相比。使用六个基于区域的标准（包括扩散张量之间的欧几里德距离、偏张量的欧几里德距离、标准差、特征值-特征向量的重叠、互相关和特征值的三组角度）以空间归一化的方式评估图集的性能-扩散张量之间的特征向量对）和三个基于纤维的标准（包括纤维束之间的距离、纤维束之间的角度和基于纤维特性曲线的标准）。