Multiclass Segmentation of Brain Tumor from MRI Images

A brain tumor is a fatal disease which takes thousands of lives each year. Thus, timely and accurate treatment planning is a critical stage to improve the quality of life. MRI is a very novel method of diagnosis of the brain which shows a fine level of details of the brain tumor. For the treatment of brain tumor, accurate segmentation of the tumor part is highly desirable. The manual tumor segmentation is a challenging and time-consuming process. Thus, automatic tumor segmentation can play a very important role in the process to speed up the treatment process. There exist different types of MRI sequences, each with its own merits and showing varying levels of information. We experimented with T2-weighted (T2), T1 with enhancing contrast (T1c), and FLAIR MRI images of the BRATS 2013 dataset and try to show that one particular MRI sequence is very useful for segmentation of one particular class of tumor than other. We have used thresholding and K-means algorithms along with a set of preprocessing and postprocessing methods for the segmentation of tumor regions. We extracted three classes of tumor: whole tumor, tumor core (whole tumor except edema), and active tumor region which is unique to high-grade (HG) cases from FLAIR, T2, and T1c, respectively. The obtained result according to Dice coefficient matrix is 0.81, 0.54, and 0.61 for the whole tumor, the tumor core, and the active tumor, respectively.