Adaptive FCM with contextual constrains for segmentation of multi-spectral MRI

An adaptive fuzzy c-means (FCM) clustering algorithm is explored for segmentation of three-dimensional (3D) multi-spectral MR images. This algorithm takes into consideration of both noise and 3D intensity non-uniformity. This algorithm models the intensity nonuniformity of MR images as a gain field or bias field that slowly varies in space, which is approximated by a linear combination of smooth basis functions made up of polynomials with different orders. The contextual constraints are included by introducing a regularization term into the cost function of FCM. The regularization term is a measure of aggregation of local voxels that tend to overcome the noise in voxel labeling. We present our scheme both for bias and gain fields, with special attention is paid to robust estimation of the bias field.