Hybrid GA Variational Bayes inference of finite mixture models for voxel classification in brain images

This paper proposes a hybrid Genetic Algorithm (GA) and Variational Bayes (VB) inference Gaussian mixture model parameters estimation methodology for unsupervised voxel classification. Unlike Expectation-Maximization algorithm (EM), the mixture model distribution parameters are modeled by a set of hyper-parameters in VB inference framework. These hyper-parameters which characterize the mixture model distributions are estimated by a Variational Expectation-Maximization (VEM) learning algorithm. However, it is difficult to initialize the hyper-parameters for VEM algorithm without prior knowledge. This study introduces a hybrid GA and VEM methodology to estimate the finite mixture models fully automatically without any specific initialization steps for voxel classification in brain images. The proposed VEM, GAVEM algorithms are validated and compared with GA and EM algorithms on real three-dimensional human brain MRI images. The voxel tissue classification results demonstrate that VEM, and GAVEM algorithms achieve competitive segmentation results compared to other parameters estimation algorithms, and fully automatically.