Detection of basal nuclei on magnetic resonance images using support vector machines

Abstract

Basal nuclei are brain structures related to motion control and treatment of neurological pathologies, such as Parkinson's or Huntington's diseases; therefore, to determine their location has great relevance for neurosurgical procedures planning. These structures are difficult to detect on MR images because they have fuzzy borders and present few changes in gray level intensities with respect to neighboring anatomical structures. Because of that, traditional image processing techniques cannot segment the basal nuclei. A classification Support Vector Machine (SVM) is able to automatically identify the nuclei and other anatomical structures when is trained with textural information from the tissues to be detected. In this work, this technique was applied on MR images to detect two structures that compose the basal nuclei. The SVM used a Gaussian kernel and was trained with feature vectors, described by gray level mean and directional variances calculated within an observation window. The structures were properly detected and segmented; it was validated by a neuroanatomist by comparing the generated segmentation to the manual segmentation of the same structures. Results show that the technique can be used for surgical planning and could be potentially extended for tri-dimensional modeling of basal nuclei. © 2009 Springer Berlin Heidelberg.