A Hybrid Approach to MR Imaging Segmentation Using Unsupervised Clustering and Approximate Reducts

Abstract. We introduce a hybrid approach to magnetic resonance image segmentation using unsupervised clustering and the rules derived from approximate decision reducts. We utilize the MRI phantoms from the Simulated Brain Database. We run experiments on randomly selected slices from a volumetric set of multi-modal MR images (T1, T2, PD). Segmentation accuracy reaches 96% for the highest resolution images and 89% for the noisiest image volume. We also tested the resultant classifier on real clinical data, which yielded an accuracy of approximately 84%.

• Type: Conference
• Year: 2005
• Where: RSFDGRC
• Authors: Sebastian Widz, Kenneth Revett, Dominik Slezak
• Keywords: Approximate Decision Reducts, Resonance Image Segmentation, RSFDGRC 2005, Segmentation Accuracy