Glioma classification in MRI using a hybrid deep learning framework with majority vote ensemble

Abstract

Glioma diagnosis remains a critical challenge, often plagued by subjectivity and inconsistent grading. This study explores the potential of deep learning to overcome these limitations, proposing a novel hybrid convolutional neural network (CNN) approach in classifying low-grade (LGG) and high-grade (HGG) tumors on T2-weighted magnetic resonance imaging (T2-W MRI) data. Five pre-trained convolutional neural networks (AlexNet, VGG-16, SqueezeNet, GoogLeNet, and ResNet-50) were fine-tuned and combined through ensemble methods: Majority Voting (MJ), Weighted Voting (WV), and Stacked Ensemble (SE). On the BraTS 2018 dataset, the ensembles demonstrated excellent performance, with the SE method achieving up to 99.35 % accuracy, 99.50 % sensitivity, 99.45 % specificity, and 99.40 % AUC. Testing on the external BraTS 2020 dataset showed strong generalization, with SE achieving 97.90 % accuracy, 98.05 % sensitivity, 97.80 % specificity, and 97.85 % AUC.The proposed ensemble techniques outperformed individual models and existing approaches, illustrating improved robustness and reliability. These findings establishes a foundation for subsequent research to explore diverse imaging sequences, segmented data analyses, and multi-institutional studies, thereby enhancing the scope and applicability of the findings in advancing automated grading systems and holding significant promise for real-world clinical use.