3D Multimodal Brain Tumor Segmentation and Grading Scheme based on Machine, Deep, and Transfer Learning Approaches

Erdal  Tasci; Aybars  Ugur; Kevin  Camphausen; Ying Zhuge; Rachel Zhao; Andra Krauze

doi:10.61797/ijbic.v1i2.153

Erdal Tasci Computer Engineering Department, Ege University, Bornova, Izmir 35100, Turkey
Aybars Ugur Computer Engineering Department, Ege University, Bornova, Izmir 35100, Turkey
Kevin Camphausen Center for Cancer Research, National Cancer Institute, NIH, Building 10, Bethesda, MD, 20892, USA
Ying Zhuge Center for Cancer Research, National Cancer Institute, NIH, Building 10, Bethesda, MD, 20892, USA
Rachel Zhao University of British Columbia, Faculty of Medicine, 317 - 2194 Health Sciences Mall, Vancouver, BC V6T 1Z3, Canada
Andra Krauze Center for Cancer Research, National Cancer Institute, NIH, Building 10, Bethesda, MD, 20892, USA

DOI: https://doi.org/10.61797/ijbic.v1i2.153

Keywords: Glioma tumor, Glioma grades, Segmentation, Feature extraction, Deep learning, Ensemble learning, MRI classification

Abstract

Glioma is one of the most common tumors of the brain. The detection and grading of glioma at an early stage is very critical for increasing the survival rate of the patients. Computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems are essential and important tools that provide more accurate and systematic results to speed up the decision-making process of clinicians. In this paper, we introduce a method consisting of the variations of the machine, deep, and transfer learning approaches for the effective brain tumor (i.e., glioma) segmentation and grading on the multimodal brain tumor segmentation (BRATS) 2020 dataset. We apply popular and efficient 3D U-Net architecture for the brain tumor segmentation phase. We also utilize 23 different combinations of deep feature sets and machine learning/fine-tuned deep learning CNN models based on Xception, IncResNetv2, and EfficientNet by using 4 different feature sets and 6 learning models for the tumor grading phase. The experimental results demonstrate that the proposed method achieves 99.5% accuracy rate for slice-based tumor grading on BraTS 2020 dataset. Moreover, our method is found to have competitive performance with similar recent works.