An implementation of the EM algorithm in white matter fiber tract clustering

Automatic unsupervised clustering of white matter fiber tracts is necessary for the group analysis of brain neural network integrity using diffusion tensor imaging (DTI) and DTI tractography techniques. In this paper, we present an implementation of the expectation-maximization (EM) algorithm to conduct the automatic unsupervised clustering of reconstructed white matter fiber tracts in DTI. The statistical model is the multivariate and multi-mode Gaussian mixture which depicts the probability distribution of white matter fiber tracts. Issues related to the parameter estimation, initialization, and convergence for applying the EM algorithm in the context of the white matter fiber tract clustering are discussed in detail. Comparisons of the EM algorithm with the K-means approach are also performed. The difference between the fixed and variant prior probabilities on the clustering result as the EM algorithm proceeds is demonstrated by experiments. Real DTI datasets are used to evaluate the performance of the proposed method. Experimental results show that the proposed approach is feasible and may be useful in the automatic unsupervised clustering of white matter fiber tracts.