Automatic Segmentation of Brain Parenchyma and Tumor in MRI

Abstract

Orthogonal subspace projection (OSP) approach has shown success in Magnetic Resonance image classification. The proposed approach of OSP can be divided into two main steps, removal of the unwanted signature an undesired signature annihilator followed by the use of matched filter. An undesired signature annihilator is used to separate the desired signature from the unwanted signatures so that the unwanted signatures can be eliminated via orthogonal subspace projection. Therefore, it requires a complete knowledge of the desired signature and the unwanted signatures present in images. In this paper, an unsupervised orthogonal subspace projection (UOSP) approach is proposed where the only knowledge of the desired signature to be classified is required. UOSP comprises two processes. Target Generation Process (TGP) and Target Classification Process (TCP). The objective of TGP is to generate a set of potential target signatures from an unknown background, which will be subsequently classified by TCP. As a result, UOSP can be used to search for a specific target in unknown scenes. Finally, the effectiveness of UOSP in target detection and classification is evaluated by several MRI experiments. All experiments were under supervision of the expert radiologist. Results show that the cerebral tissue was segmented accurately into four images, tumor, gray matter, white matter and cerebral spinal fluid indicating the possible usefulness of this method.