Conclusion

We have presented a fully automatic approach for the segmentation of brain MR images. The method is based on a HMRF-EM framework, which is a combination of the hidden Markov random field (HMRF) model and the associated MRF-MAP estimation and the EM fitting procedures. The HMRF model is proposed in this paper as a substitute for the widely used FM model, which is considered as sensitive to noise and therefore not robust. As a general method, the HMRF-EM algorithm could be applied to many other image segmentation problems. We also show that the framework can easily be extended by incorporating other techniques in order to improve its performance on certain problems. As an example, we demonstrated how the bias field correction algorithm by Guillemaud and Brady [13] can be incorporated into this framework. As a result, a three-dimensional fully automatic approach for brain MR image segmentation is achieved and significant improvements have been observed in terms of both the bias field estimation and the tissue classification.