Automated Brain Tumor Segmentation from MRI Data Based on Local Region Analysis

Magnetic resonance imaging for its excellent contrast information about brain tissues from a variety of excitation sequences has become the standard modality of choice for brain tumor diagnosis. Accurate delineation of tumor region prior to invasive surgery is a must and yet very tedious to be done manually for medical practitioners and neurosurgeons. However, automating this task is very challenging due to the immense heterogeneity of tumor structure, severe partial volume effect and for the presence of different noises. In this work, a fully automated brain tumor segmentation method is proposed in which the local region of pixels is considered as a basic processing unit of classification instead of the traditional pixel-based classification scheme. The extracted local regions are then clustered into two groups using K means clustering algorithm for multiple feature points which effectively increases the clustering performance. The features used here are different statistical measures of intensity values and textural features obtained from the Gaussian filter bank. The segmented tumor volume is constructed by combining all the local tumor regions. Not only the dice similarity measure of our method is comparable with the state of art methods but also is more efficient than the other complex approaches in the perspective of simplicity, reproducibility and low computational cost.