Single seed delineation of brain tumor using multi-thresholding

A novel two-stage region of interest (ROI) segmentation is proposed for detecting glioblastoma multiforme (GBM) tumors from brain magnetic resonance images (MRIs). The method involves multi-level thresholding followed by post-processing. Initially discrete curve evolution (DCE) identifies multiple intervals around the significant (or visually critical) points, with a threshold being selected in each such interval using Otsu’s method or Li and Lee’s entropy. Next a post-processing on the segmented image, based on connected-component analysis and flood-fill operation, helps to extract each refined ROI around a single seed inserted by the user. The segmented ROI is more accurate, both quantitatively and qualitatively, as compared to related methods – inspite of using only a single seed. This is evaluated (i) visually, (ii) in terms of the Jaccard and Dice indices (on the ROI), and (iii) over time complexity of the algorithm. The experimental results on contrast enhanced T1-weighted MRI slices of 25 patients, each having the corresponding ground truth about the tumor regions, establish the effectiveness of our algorithm.