Unsupervised and Adaptive Segmentation of Multispectral 3D Magnetic Resonance Images of Human Brain: A Generic Approach

A generic algorithm is presented for the segmentation of threedimensional multispectral magnetic resonance images. The algorithm is unsupervised and adaptive, does not require initialization, classifies the data in any number of tissue classes and suggests an optimal number of classes. It uses a statistical model including Bayesian distributions for brain tissues intensities and Gibbs Random Fields (GRF)-based spatial contiguity constraints. The classification is unsupervised, that is to say the intensity-based signatures of brain tissues and the spatial hyperparameters of the underlying GRF are derived from the data. Adaptivity is achieved through the variation of the size of the neighborhoods used for the estimation of the intensity characteristics. This allows slow variations of signal intensity in space to account for MRI intensity nonuniformity. Segmentation results with proton density, T2 and T1-weighted data are provided. The algorithm can be used as an independent segmentation module within a brain MRI data processing pipeline.