Hybrid EfficientNet-Transformer Model for Multi-Class Brain Tumor Classification from Imbalanced MRI Images

Abstract

Brain tumor classification from MRI scans is critical for diagnosis but remains challenging due to subtle inter-class variations (e.g., glioma vs. meningioma) and significant class imbalance in clinical datasets. Current deep learning approaches relying solely on Convolutional Neural Networks (CNNs) struggle to capture global contextual relationships, leading to suboptimal performance in tumor types.

Base classifiers like CNN fail to model long-range spatial dependencies in tumor boundaries, class imbalance biases models toward majority classes (e.g., no_tumor), and limited explainability in feature fusion reduces clinical trust. To handle the addressed issues, we introduce a hybrid EfficientNetB4-Transformer architecture that synergizes local feature extraction with global attention mechanisms. Our model addresses class imbalance via Focal Loss and test-time augmentation, while transformer blocks explicitly model tumor morphology across MRI slices.