A Bayesian approach to inferring fiber tract bundle labels in Diffusion Tensor Imaging

Abstract

Aiming for achieving anatomically meaningful results in automatic unsupervised clustering of reconstructed white matter (WM) fiber tracts in Diffusion Tensor Imaging (DTI), we present a Bayesian approach enabling to incorporate prior anatomical knowledge and handle outliers easily. Particularly, the distribution of WM fiber tracts is described by a Gaussian mixture model. By applying the Bayesian theorem, we are able to evaluate the posterior probability that a fiber tract belongs to each target bundle based on both the probability distribution and prior anatomical information. A fiber tract is labeled into a target bundle with which the maximum posterior probability occurs. If all calculated posterior probabilities are smaller than a user defined threshold, the fiber tract is labeled as an outlier. The prior anatomical information is represented by the target fiber tract bundles' prior distribution which can be obtained by anatomical studies and may be amended by further researches. The fact that this type of prior information is less dependent on individual brain structures than that in some existing methods makes this approach useful in group studies. Real DTI datasets are used to assess the performance of the method. Experimental results show that this technique is feasible and may have potential applications in group analysis of WM fiber tracts in DTI.