A Comparative Study of EfficientNetB0 and Vision Transformer (ViT-B16) Architectures for Brain Tumor Classification Using MRI Scans

Abstract

Accurate detection and classification of brain tumors from MRI scans is critical for effective clinical diagnosis and treatment planning. While Convolutional Neural Networks (CNNs) like EfficientNetB0 have been widely used for medical image analysis due to their strong feature extraction capabilities, their performance is often hindered by limited spatial context awareness. Vision Transformers (ViTs), by contrast, leverage self-attention mechanisms to capture global contextual relationships, potentially overcoming these limitations. This study presents a rigorous comparative analysis of EfficientNetB0 and ViT-B16 architectures on the Brain Tumor MRI Dataset, which includes four classes: Glioma, Meningioma, Pituitary Tumor, and No Tumor. Both models were trained under identical preprocessing, augmentation, and hyperparameter settings using Kaggle’s cloud infrastructure. Evaluation based on accuracy, precision, recall, F1-score, AUC-ROC, and interpretability (via Grad-CAM and Attention Maps) revealed stark differences. EfficientNetB0 exhibited severe overfitting, achieving high training but poor test performance (30.89% accuracy), misclassifying most tumor types. In contrast, ViT-B16 achieved superior generalization with a test accuracy of 71.62% and balanced performance across tumor categories. Interpretability analyses confirmed ViT-B16's ability to localize tumors more effectively. These results highlight the promise of transformer-based architectures for robust and clinically viable brain tumor classification.