An unsupervised random walk approach for the segmentation of brain MRI

The segmentation of magnetic resonance data is a challenging task, essential to several clinical and research applications. Since they do not require assistance from a human expert, unsupervised segmentation approaches are especially useful for this task. In this paper, we present two novel unsuper-vised segmentation methods based on random walks. The proposed methods find the probability mode in a local region around pixels, defined by a stochastic diffusion process. As the well-known Hidden Markov Random Field (HMRF) algorithm, these methods can also adapt dynamically to the distribution of pixel intensities by recomputing iteratively the parameters of these distributions. Experiments carried out on real 3D brain MRI from the Internet Brain Segmentation Repository (IBSR) show these methods to be computationally efficient and outperform approaches based on HMRF and Gaussian Mixture Models (GMM), in terms of mean segmentation agreement.