基于扩散磁共振成像的概率纤维追踪算法研究

Diffusion Magnetic Resonance Imaging (dMRI) provides a noninvasive tool for the investigation of white matter tracts in the human brain. Using tractography, white matter tracts can be represented as 3D streamlines. Due to the noise in dMRI acquisition, uncertainty is introduced into the fiber tracking process, and researchers have developed probabilistic tractography methods to account for the uncertainty. Bootstrap provides a nonparametric approach to uncertainty characterization, and has been used to develop probabilistic tractography algorithms. A bootstrap-based tractography method involves three steps: 1) synthesis of diffusion signals using bootstrap to estimate the distributions of diffusion signals; 2) estimation of fiber orientations (FOs) and thus FO distributions, which are crucial features for fiber tracking; 3) fiber tracking by randomly drawing samples from the FO distribution. Sparse dMRI signal models have recently gained popularity because they decrease the number of diffusion gradients required for resolving crossing fibers and thus imaging time. However, existing bootstrap tractography methods have not considered the sparsity assumption in the signal model. In addition, in these existing probabilistic tractography methods, FOs are estimated in a voxelwise manner, which can be improved by introducing spatial consistency. In this proposal, we plan to study probabilistic tractography algorithms based on bootstrap and sparse dMRI signal models. The proposal has four steps. 1) We plan to explore bootstrap strategies that are compatible with sparse dMRI signal models, where distributions of diffusion signals can be estimated. 2) From the diffusion signals synthesized by bootstrap, we plan to develop FO estimation algorithms that incorporate spatial consistency of FOs to achieve accurate estimation of FO distributions, and perform probabilistic fiber tracking using the estimated distributions. 3) The proposed tractography (including the FO estimation) algorithm will be validated on both simulated and real dMRI data. 4) To demonstrate the scientific application of the proposed method, we will apply it to a study on the cerebellar connectome. We hope the proposed work can provide fundamental tools for the investigation of white matter tracts and brain wiring, and eventually advance basic and applied neuroscience using medical imaging.

利用扩散磁共振成像（diffusion Magnetic Resonance Imaging，dMRI）和纤维追踪技术，神经束可以重建为三维流线。研究者们设计了基于bootstrap的概率纤维追踪算法，以非参数的估计噪声引起的纤维追踪的不确定性。为了缩短成像时间，稀疏信号模型近年来被广泛应用于dMRI图像处理中。但现有的基于bootstrap的概率纤维追踪算法没有考虑信号的稀疏性。因此，我们计划研究基于bootstrap和稀疏信号模型的概率纤维追踪算法。本项目分为四个部分。1）探索适用于稀疏信号模型的bootstrap方法，估计dMRI信号分布。2）利用dMRI信号分布，设计改进的纤维方向估计算法，得到准确的纤维方向分布，进行概率纤维追踪。3）利用模拟和真实数据验证本方法。4）利用本方法研究小脑神经网络。我们希望本工作为神经束和脑网络的研究提供基本工具，并推动医学成像在神经科学中的应用。

脑研究是当前重要的前沿科学研究内容。脑研究计划已经成为世界性的研究计划，包括美国、欧盟、日本等国家和地区都开展了脑研究计划。在脑研究中，白质神经束分析是重要的研究内容，可以揭示脑连接和脑网络的相关信息，在脑疾病、脑发育等问题中具有重要应用。弥散磁共振成像提供了非侵入式研究白质神经束的重要手段。利用纤维追踪技术，神经束可以重建为三维流线。但是在临床条件下，由于弥散梯度数目有限，神经束重建容易受到噪声和复杂纤维结构的影响，准确的神经束重建及其不确定性量化具有挑战。本项目利用弥散磁共振信号的稀疏性和空间一致性，改进了纤维方向估计算法，并且提出了相应的基于Lasso Bootstrap的概率性纤维追踪方法，描述神经束重建的不确定性。此外，本项目探索了基于深度学习的纤维方向估计方法，将弥散磁共振信号的稀疏性以及其他先验知识引入了网络设计中。这些针对纤维方向计算的方法还被扩展到其他基于弥散磁共振图像的结构信息计算中。本项目对所设计的算法通过模拟和真实数据进行了全面的定性和定量的验证。在临床可行的弥散梯度数目下，本项目设计的方法可以准确的重建白质神经束并描述重建结果的不确定性，并且其性能优于现有的其他方法。此外，这些算法已初步应用于脑连接的研究中。本项目的研究成果为脑科学中白质神经束的研究提供了更加先进的计算工具。