数据驱动的三叉神经纤维束自动分割算法

目前三叉神经的纤维跟踪成像过程中普遍存在人工依赖性问题,主要包括人工绘制感兴趣区域(ROI)及手动筛选目标纤维束,导致结果的不确定性和数据误差。针对此类问题,提出一种数据驱动的三叉神经纤维自动分割算法。利用多组大脑样本的纤维数据,建立数据驱动的纤维聚类图谱,实现新样本纤维数据的自动分割,直接得到三叉神经纤维束。在实验中,选择25组青年健康人的数据作为样本数据。首先,利用FSL软件分割工具提取脑干作为ROI,进行确定性纤维跟踪。其次,通过对20组纤维数据进行多样本配准和谱聚类,创建数据驱动的纤维聚类图谱。根据三叉神经细小的特点,在建立纤维图谱过程中,通过对脑干纤维束进行二次分类来标注三叉神经纤维束。最后,选择5组青年健康人的新样本数据,将其脑干纤维数据应用纤维图谱自动分割得到三叉神经纤维束,并计算同一样本数据的自动分割结果与手动分割结果之间的加权Dice系数。结果显示,所提出的方法成功分割5组数据的三叉神经纤维束,而传统人工方法成功识别4组三叉神经纤维束,两者结果之间的加权Dice系数分别为0.865,0.939,0.824,0.942。该方法可以有效避免人为因素的影响,提高神经外科医生与颅神经研究者的工作效率。

Data-Driven Automatic Segmentation Algorithm for Trigeminal Nerve Fiber

A common problem in the process of trigeminal nerve fiber tractography is artificial dependence,mainly including artificial rendering of region of interest (ROI) and manual screening of target fibers,which generally results in uncertainty and data errors. To ovecome this problem,a data-driven automatic trigeminal nerve fiber segmentation algorithm was proposed in this paper. A data-driven fiber clustering atlas was established based on the fiber data of several groups of brain samples,which automatically segmented the fiber data of new samples and directly obtained the trigeminal nerve fibers. In experiments,25 groups of healthy young individuals were selected as samples. Firstly,the brainstem was extracted by FSL software segmentation tool as ROI for deterministic fiber tracking. Secondly,a data-driven clustering atlas of fibers was created by multi-sample registration and spectral clustering of 20 groups of fibers. According to the tiny characteristics of trigeminal nerve,the trigeminal nerve fibers were labeled by secondary classification of brainstem fibers in the process of establishing fiber atlas. Finally,new sample data of 5 groups of healthy young people were selected,and their brainstem fiber data were automatically segmented using fiber atlas to obtain trigeminal nerve fiber bundles,and theweighted Dice coefficient between the results of automatic segmentation and manual segmentation of the same sample data was calculated. Results showed that the proposed method successfully segmented 5 sets of trigeminal nerve fiber bundles while the conventional manual method successfully identified 4 sets. The weighted Dice coefficients between the two results were 0.865,0.939,0.824,and 0.942. These results showed that this method can effectively avoid the influence of human factors,and greatly improve the work efficiency of neurosurgeons and cranial nerve researchers.