Tumor extraction from multimodal MRI

Abstract

Manual segmentation of brain tumor from 3D multimodal magnetic resonance images (MRI) is time-consuming task and leading to human errors. In this paper, two automated approaches has been developed for brain tumor segmentation to discuss which one will provide accurate segmentation that is close to the manual results. The MR feature images used for the segmentation consist of three weighted images (enhanced T1, proton density(PD) and T2) for each axial slice through the head. The first approach is based on multi-features Fuzzy-c-means (FCM) algorithm followed by a post-processing step based on prior knowledge to refine the tumor region. The second approach is three-pass step. First, each single modality MRI is classified separately with FCM algorithm. Second, classified images are fused by Dempster-Shafer evidence theory to get the final brain tissue labeling. Finally, prior knowledge are used to refine the tumor region. For validation, ten tumor cases of different size, shape and location in the brain are used with a total of 200 multimodals MRI.The brain tumor segmentation results are compared against manual segmentation carried out by two independent medical experts and used as the ground truth. Our experimental results suggest that the second approach produces results with comparable accuracy to those of the manual tracing compared to the first approach.