Research On Reproducible Comparison Pipeline For Registration And Fiber Tracts Segmentation Algorithm

Registration algorithm is an essential procedure for brain imaging analysis. The registration effects can directly determine the feasibility and reliability of the analysis. In the past 20 years, researchers propose numerous nonlinear registration algorithms and provide corresponding software packages. While no single algorithm can fit well on all imaging data. How to select an appropriate algorithm becomes an urgent issue.Diffusion Tensor Imaging (DTI) can infer white matter fiber tracts pathway through measuring diffusion anisotropy of water molecule, and thus providing human brain’s connecting information. The number of fiber tracts through Tractography is more than 10, 000. Fiber tracts show high variations among individuals. Boundaries between fiber bundles are not evident. These factors lead to fiber tracts segmentation a great challenge.As for the selection of registration algorithm, we propose a registration comparison pipeline with reproducibility in this paper. We take completely the same preprocessing procedure on raw data, including resampling, skull extraction, deletion of distorted image. Then different registration algorithms are applied to these images, different templates also are deployed to analyze the effect of template. After registration, we segment subject’s image and templates, and calculate the overlap value between segmented tissue, including white matter and gray matter, to evaluate registration’s performance. We took experiments on this pipeline on 1-2 years-old children, and find out that IRTK and Nifty can achieve superior results.We employed a density-peaks clustering method to segment fiber tracts. This clustering algorithm is based on two assumptions:(1) density of cluster center is higher than its neighbors, and (2) cluster center has to be far away from the other fibers with higher density. Those fibers satisfied assumptions are set as cluster centers. Remaining fibers are assigned to the same cluster as their nearest neighbor with higher density. Moreover, outliers can be detected via a border density threshold, yielding robust segmentation. We used visualization and calculating overlap method to test the performance of this clustering algorithm on JHU-DTI dataset. Experiments show that visualization and performances metric have higher consistency with manually segmented bundles compared with classical clustering methods.