Unsupervised MRI segmentation with spatial connectivity

Magnetic Resonance Imaging (MRI) offers a wealth of information for medical examination. Fast, accurate and reproducible segmentation of MRI is desirable in many applications. We have developed a new unsupervised MRI segmentation method based on k-means and fuzzy c-means (FCM) algorithms, which uses spatial constraints. Spatial constraints are included by the use of a Markov Random Field model. The result of segmentation with a four-neighbor Markov Random Field model applied to multi-spectral MRI (5 images including one T1-weighted image, one Proton Density image and three T2-weighted images) in different noise levels is compared to the segmentation results of standard k-means and FCM algorithms. This comparison shows that the proposed method outperforms previous methods.