Automated segmentation of caudate nucleus in MR brain images with voxel classification Yulia Arzhaeva, Eva van Rikxoort, and Bram van Ginneken Image Sciences Institute, University Medical Center Utrecht, the Netherlands {yulia,eva,bram}@isi.uu.nl Abstract. This paper presents a supervised voxel classification method for segmentation of the caudate nucleus from brain MRI images. Su- pervised voxel classification is a general pattern recognition technique. In this application general spatial and local structure features extracted from image voxels were used together with a k-nearest neighbor classifier. The trained classifier has been applied to different groups of test data. On test data that originated from the same population as the training images, the method yielded segmentations that correlated very well with human segmentations (Pearson correlation coefficient (PCC) of 0.82 and volumetric overlap (VO) of 74.2%). On data from a different source that exhibits intensity ranges similar to the training data, the method per- formed slightly worse (PCC of 0.52, VO of 64%), and the method failed on data with different intensity ranges. 1 Introduction Magnetic resonance imaging (MRI) provides detailed three-dimensional (3D) im- ages of living tissues and is widely used for brain studies. Quantitative analysis of brain MRI images often requires precise segmentation of specific neuroanatom- ical structures. One of these is the caudate nucleus (CN), a subcortical compo- nent of the basal ganglia that is involved in sensory-motor control, cognition, language, emotion and other important brain functions. The caudate is a periventricular gray matter structure that shows up lighter compared to the majority of cortical gray structures in T1-weighted MRI im- ages. The caudate has a rather homogeneous intensity but is difficult to segment because it is attached to other gray structures at multiple locations. It consists of a relatively large head, a cone-shaped body and a thin tail. Manual 3D segmentation of the caudate is difficult and time consuming. For that reason various fully automated and computer-assisted segmentation methods have been developed including atlas-based registration techniques, de- formable models, knowledge-driven and histogram-driven approaches, and sta- tistical modelling (see for a survey [1] and a number of references from [2]). However, an automated or semi-automated method for CN segmentation has not yet become the technique of choice in cognitive and neuroscience laborato- ries. That suggests that an optimal automated technique is yet to be found. T. Heimann, M. Styner, B. van Ginneken (Eds.): 3D Segmentation in The Clinic: A Grand Challenge, pp. 65-72, 2007.