Optimal feature space for MRI

We present development and application of an optimal feature space for magnetic resonance imaging (MRI), without explicit calculation of tissue parameters. We generate a three-dimensional (3D) feature space representation of MRI using a linear transformation that optimizes clustering properties of data. The optimal transformation maximizes the ratio of interset to intraset Euclidean distance for normal tissues subject to the constraint that these tissues be clustered at prespecified target positions. In the resulting 3D feature space, abnormalities are clustered at positions different from the prespecified locations for normal tissues. From the 3D histogram (cluster plot), we identify clusters and define regions of interest (ROIs) for normal and abnormal tissues. These ROIs are used to estimate signature (feature) vectors for each tissue type which in turn are used to segment the image. The proposed feature space has been compared to those generated by tissue-parameter-weighted images, principal component images, and angle images, demonstrating its superiority for feature extraction [1]. Here, the method and its performance are illustrated using MR images of phantoms and human brains with tumor.