An Efficient Algorithm for Multiple Sclerosis Lesion Segmentation from Brain MRI

We propose a novel method for the segmentation of Multiple Sclerosis (MS) lesions in MRI. The method is based on a three-step approach: first a conventional k-NN classifier is applied to pre-classify gray matter (GM), white matter (WM), cerebro-spinal fluid (CSF) and MS lesions from a set of prototypes selected by an expert. Second, the classification of problematic patterns is resolved computing a fast distance transformation (DT) algorithm from the set of prototypes in the Euclidean space defined by the MRI dataset. Finally, a connected component filtering algorithm is used to remove lesion voxels not connected to the real lesions. This method uses distance information together with intensity information to improve the accuracy of lesion segmentation and, thus, it is specially useful when MS lesions have similar intensity values than other tissues. It is also well suited for interactive segmentations due to its efficiency. Results are shown on real MRI data as wall as on a standard database of synthetic images.