Abstract Glioma sub-region segmentation plays a critical role in providing tumor information such as appearance and location that is essential for better prognosis and treatment. Several factors limits to create an automatic glioma segmentation technique like intensity ambiguity, data scarcity, class imbalance, etc. For developing such a technique, we have proposed a stage 3 hybrid glioma sub-region segmentation method that modifies traditional techniques in this paper. Our technique develops a system that can select a) the best modality, and b) seed values at stage 1 and stage 2. A hybrid segmentation technique is introduced at stage 3 that includes graph cut (GC) and a novel Location Based Region Growing (LBRG) method. The complete system can extract intricate small and disjoint regions of brain glioma tumors. Our work is evaluated on publically accessible datasets BRATS 2017, BRATS 2019, BRATS 2020, and BRATS 2021, and a competitive performance comparison is provided with state-of-the-art methods. Some of these methods have competed in the BRATS MICCAI challenge and got the best results. The dice score values evaluated for BRATS 2017 dataset are 0.874 and 0.891 for Enhancing Tumor (ET) and Tumor Core (TC), respectively. Similarly, Dice Score for BRATS 2020 dataset is 0.868 and 0.873 for ET and TC regions, respectively. Sensitivity of 0.93 and 0.99 is obtained for ET and TC regions for BRATS 2017. For similar region, 0.98 and 0.99 values are obtained on BRATS 2020 dataset. Specificity of 0.906 and 0.873 is obtained for ET and TC regions for BRATS 2017. For similar region, 0.912 and 0.8744 values are obtained on BRATS 2020 dataset. An improvement percentage of 9%, 7.2% in comparison to state-of-the-art method for TC and, an improvement of 19%, 18.4% for ET region as well is observed.
Dogra et al. (Sat,) studied this question.