Fuzzy volumetric delineation of brain tumor and survival prediction

Abstract

A novel three-dimensional detailed delineation algorithm is introduced for Glioblastoma multiforme tumors in MRI. It efficiently delineates the whole tumor, enhancing core, edema and necrosis volumes using fuzzy connectivity and multi-thresholding, based on a single seed voxel. While the whole tumor volume delineation uses FLAIR and T2 MRI channels, the outlining of the enhancing core, necrosis and edema volumes employs the T1C channel. Discrete curve evolution is initially applied for multi-thresholding, to determine intervals around significant (visually critical) points, and a threshold is determined in each interval using bi-level Otsu’s method or Li and Lee’s entropy. This is followed by an interactive whole tumor volume delineation using FLAIR and T2 MRI sequences, requiring a single user-defined seed. An efficient and robust whole tumor extraction is executed using fuzzy connectedness and dynamic thresholding. Finally, the segmented whole tumor volume in T1C MRI channel is again subjected to multi-level segmentation, to delineate its sub-parts, encompassing enhancing core, necrosis and edema. This was followed by survival prediction of patients using the concept of habitats. Qualitative and quantitative evaluation, on FLAIR, T2 and T1C MR sequences of 29 GBM patients, establish its superiority over related methods, visually as well as in terms of Dice scores, Sensitivity and Hausdorff distance.