Clustering of Major White Matter Bundles using Tract-specific Geometric Curve Features

The tract-based white matter analysis is an effective framework to study the structural brain using quantitative parameters of interest along with major fibre bundles. Accurate and consistent clustering of brain white matter bundles from the whole brain, therefore, is essential. In this paper, an efficient automated supervised curve-clustering algorithm is proposed for segmenting each of the major white matter bundles from the whole-brain tracts using geometrical curve features of each bundle. The proposed stepwise multi-feature matching approach is examined on the ADNI project dataset with 20 normal adults and 6 major bundles. To achieve the automatic segmentation of the major fibre bundle curves, curve-clustering algorithms performed at four levels. At the preprocessing level, geometrical methods work with 3D training curves of a major bundle to extract the features and to compute the best representative curve and average curve. In the second level, algorithms are utilized the pre-processing results for the whole-brain tracts to cluster a set of curves based on the location of the major bundle. At the third level, more clustering is performed on the previous-level output curves based on the multi-feature matching. At the final level, techniques try to extract the original curves of the major fibre bundle using the discrete Fréchet distance. The proposed method shows 87% in average accuracy for the 6 major fibre bundles. Our work explores the benefits of segmenting the complex whole-brain tracts into anatomically realistic major fibre bundles and provide resources for tract-based quantitative analysis.