Fiber segmentation using a density-peaks clustering algorithm

Automatic segmentation of fiber bundles can be beneficial to quantitative analysis on neuropsychiatric diseases. Previous clustering methods for fiber segmentation typically specify the number of clusters in advance or rely on prior knowledge. In this paper, we develop a new segmentation algorithm based on density-peaks clustering, in which the number of clusters can arise intuitively. This clustering algorithm finds bundle centers by formulating two properties of a center: 1) its density is higher than neighbors, and 2) it has to be far away from the other fibers with higher density. 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 for each bundle, yielding robust segmentation. Visualization and overlap values between segmented and delineated bundles are used to evaluate performance on JHU-DTI data set. Experimental results show that the clustered bundles have higher consistency compared with those from classical clustering methods.