A feature selection for multispectral image segmentation algorithm is proposed to segment brain Magnetic Resonance Imaging (MRI) in the native coordinate space, with no need of transformation to a standard coordinate space or the probability maps for initiation of segmentation. The proposed method is a simple voxel-based algorithm which derived from multispectral remote sensing techniques. The proposed method requires minimal in-teractive input which manually depicts three 3x3 pixels of Gray Metter (GM), White Matter (WM), and Cerebral Spinal Fluid (CSF) tissue clusters to segment 3D high-resolution mul-tislice-multispectral MRI's.
In this study, Sequential floating forward selection (SFFS) was employed to select the input feature vectors for classification and segmentation. Support vector machines (SVM's) were employed to classify the brain tissue by type. The similarity indexes, expressing over-lap between segmentation results and the ground truth, was used as the performance evalua-tion. The ground truth images were acquired from Brainweb. For the features not selected by the SFFS, the segmentation performance for GM, WM and CSF is 95.4%, 96.6% and 95.7%, respectively. After SFFS feature selection, the segmentation performance for GM, WM and CSF is raised to 99.9%, 99.9% and 99.8%, respectively. Statistical t-test analysis shows that the input feature images selected by the SFFS significantly improve the segmen-tation performance. The output of this research provides an essential segmentation and quantification tool for other types of MRI modalities, such as Diffusion Tensor Imaging (DTI) and Dynamic Contract-Enhanced MRI (DCE-MRI).