Segmenting Multiple Sclerosis Lesions Using a Spatially Constrained K-Nearest Neighbour Approach

We propose a method for the segmentation of Multiple Sclerosis lesions. The method is based on probability maps derived from a K-Nearest Neighbours classification. These are used as a non parametric likelihood in a Bayesian formulation with a prior that assumes connectivity of neighbouring voxels. The formulation is solved using the method of Iterated Conditional Modes (ICM). The parameters of the method are found through leave-one-out cross validation on training data after which it is evaluated on previously unseen test data. The multi modal features investigated are 3 structural MRI modalities, the diffusion MRI measures of Fractional Anisotropy (FA), Mean Diffusivity (MD) and several spatial features. Results show a benefit from the inclusion of diffusion primarily to the most difficult cases. Results shows that combining probabilistic K-Nearest Neighbour with a Markov Random Field formulation leads to a slight improvement of segmentations.