Evaluation of Fiber Clustering Methods for Diffusion Tensor Imaging

Paper (pdf) to appear IEEE Visualization 2005. This article is based on the Master's thesis of Bart Moberts.

Abstract

Fiber tracking is a standard approach for the visualization of the results of Diffusion Tensor Imaging (DTI). If fibers are reconstructed and visualized individually through the complete white matter, the display gets easily cluttered making it difficult to get insight in the data. Various clustering techniques have been proposed to automatically obtain bundles that should represent anatomical structures, but it is unclear which clustering methods and parameter settings give the best results.

We propose a framework to validate clustering methods for white-matter fibers. Clusters are compared with a manual classi- fication which is used as a gold standard. For the quantitative evaluation of the methods, we developed a new measure to assess the difference between the gold standard and the clusterings. The measure was validated and calibrated by presenting different clusterings to physicians and asking them for their judgement. We found that the values of our new measure for different clusterings match well with the opinions of physicians.

Using this framework, we have evaluated different clustering algorithms, including shared nearest neighbor clustering, which has not been used before for this purpose. We found that the use of hierarchical clustering using single-link and a fiber similarity measure based on the mean distance between fibers gave the best results.

Figures

	Cluttered image showing the fibers in a healthy brain by seeding in the whole volume. The color coding shows main eigenvalue.
	Hierarchical clustering with single-link and mean distance between fibers (left) and closest point (right).
	Overview of the validation process.
	Gold standard.
	Incomplete (left) and incorrect clustering (right).

---