结合K-Means与移动最小二乘法的脑部纤维追踪方法

Brain Fiber Tracking Method Based on K-Means and Moving Least Squares

摘要: 脑部神经纤维追踪整合单个体素内纤维的方向信息，描绘全局的纤维走向分布，因此单个体素的方向信息决定了纤维追踪结果的准确性和完整性，而传统方法往往直接将体素主特征向量方向作为体素纤维方向，却没有考虑体素张量特性，会产生较大的误差。对此，提出一种结合K-means和移动最小二乘法的纤维追踪方法。首先，根据体素的椭球模型，利用K-means聚类算法将体素聚为3类；接着，用移动最小二乘法拟合非细长椭球状体素的张量信息；最后，采用STT算法对更新后的张量场进行纤维追踪。采用两份数据进行实验结果分析：在真实临床数据中，该算法比STT能追踪到更完整的纤维束；在模拟人脑数据中，该算法能追踪到最多的23束正确纤维束，正确纤维连接比达到45%，错误纤维束相较UKF减少16束，错误纤维连接比降低到28%。

Abstract: Brain nerve fiber tracking refers to the integration of fiber direction information in a single voxel to depict the global fiber direction distribution, so, the direction information of a single voxel determines the accuracy and completeness of the fiber tracking results. Traditional methods often directly use the direction of the main feature vector of the voxel as the direction of the voxel fiber, but do not consider the characteristics of the voxel tensor, which will cause large errors. Therefore, a fiber tracking method combining K-means and moving least square method is proposed. Firstly, according to the ellipsoid model of voxels, the K-means clustering algorithm was used to cluster the voxels into three categories. Then, the tensor information of non-slender ellipsoid voxels was fitted by moving least square method. Finally, STT algorithm is used to track the fiber of the updated tensor field. Two data sets were used to analyze the experimental results. In real clinical data, the algorithm could trace more complete fiber bundles than STT. In simulated human brain data, the algorithm can track the maximum number of correct fiber bundles of 23, the correct fiber connection ratio reached 45%, and the number of wrong fiber bundles decreased by 16 compared to UKF, the wrong fiber connection ratio reduced to 28%.