基于字典基函数框架的纤维方向分布模型重建

A Novel Fiber Orientation Distribution Reconstruction Method Based on Dictionary Basis Function Framework

脑白质纤维方向分布模型重建是脑纤维成像的重要过程之一,为纤维束跟踪提供精确的纤维方向估计。传统方法的约束条件往往依赖于先验的纤维方向信息,限制了计算效率与精度的提高。提出一种基于字典基函数框架的纤维方向分布函数(f ODF)估计方法,在单位球面上建立均匀的字典基函数分布,构造基于字典基函数的球面反卷积模型,直接通过非负最小二乘方法求得基函数系数,避免伪峰与复杂的算法以及更高阶的大规模病态逆问题的计算,使得纤维方向分布的稀疏性和非负性可以通过几个基函数的加权和很容易地表示。模拟磁共振数据与临床脑部数据的实验结果显示,该方法在同等条件下与目前流行的Super-CSD和QBI方法相比,计算复杂度大大降低,计算时间仅为Super-CSD的1/3;角度分辨率扩大到20°以内(其中,Super-CSD为40°,QBI为60°)。由于该方法简单高效,角度误差不会随角度等参数的变化而跳变,也因为角度分辨率和角度误差性能的提升,使得纤维方向的错误重构概率降低到8%以下,可为纤维跟踪技术提供精确的局部纤维走向。

Brain WM fiber orientation reconstruction is the throat of fiber imaging that set the premise of fiber tractgraphy. In conventional approaches, constraint conditions are dependent on prior information of fiber orientation, limiting the improvement of computation efficiency and precision. In this work, we presented a dictionary basic function based fiber orientation distribution function （fODF） estimation method. First, the uniform distribution of dictionary basis functions on unit sphere was established; second, spherical deconvolution （SD） model based on the dictionary basis functions was established; third, the coefficients of basic functions were directly obtained using non-negative least square method. The proposed method successfully avoided the spurious peak, leases computation burden and eliminates higher order ill-posed inverse problem, which makes the sparsity and non-negativity needed for the representation straightforwardly represented by weighted sums of basic functions. Experiments on simulated and clinical brain data both verified that computational complexity of the proposed method is much less than that of Super-CSD and QBI under the same conditions, such as, computational time of the method is a third as long as that of Super-CSD ; the angular resolution of the method increases to a range of 20°, while Super-CSD is 40° and QBI is 60°. Because the method is simple and efficient, angular error does not change with such parameters as angle, and with the addition of the improvement of angular resolution and angular error performance, the probability of error fiber orientation reconstruction decreases to less than 8% , which will provide accurate local fiber orientation for fibertractgraphy.