Deep-EFNet: An Optimized EfficientNetB0 Architecture With Dual Regularization for Scalable Multi-Class Brain Tumor Classification in MRI

Brain tumor classification from MRI images plays a crucial role in early diagnosis and treatment planning. While deep learning approaches have shown promise in this domain, achieving high accuracy (ACC) by maintaining model generalization across different datasets remains challenging. Existing approaches in automated brain tumor classification often struggle with feature extraction complexity and model generalization, particularly when dealing with datasets of varying sizes. Additionally, many current methods lack robust performance validation across multiple evaluation metrics. To develop and validate a transfer learning-based deep learning model for accurate multi-class brain tumor classification that demonstrates consistent performance across different dataset sizes while maintaining high classification ACC and reliability. We presented Deep-EFNet, a novel architecture based on EfficientNetB0 with transfer learning. The proposed model leverages the powerful feature extraction capabilities of EfficientNetB0 while incorporating strategic enhancements for improved performance. These enhancements include batch normalization layers for stable training, dropout and L2 regularization mechanisms for preventing overfitting, and a hierarchical dense layer structure for improved feature representation. To validate Deep-EFNet approach comprehensively, we evaluated the model on two distinct datasets containing 3,064 and 7,023 MRI images respectively, each encompassing four classes of brain tumors. The experimental evaluation of Deep-EFNet demonstrated exceptional performance across multiple metrics and datasets. The model achieved an ACC of 95% on 3K-DS (3,064 images) and notably improved to 98% ACC on the larger Dataset 2 (7,023 images). The reliability of these results is further validated by impressive ROC curve scores, achieving perfect 1.00 values across all classes in Dataset 2, while 3K-DS showed equally strong performance with ROC scores of 0.99 for class 0 and 1.00 for the remaining classes. Comprehensive validation through ACC, precision, recall, F1-scores, confusion matrices, and calibration curves further demonstrated the model’s robust and consistent performance across all evaluation metrics. Deep-EFNet demonstrates state-of-the-art performance in brain tumor classification, surpassing existing approaches in both ACC and reliability. The model’s consistent performance across datasets of varying sizes suggests strong generalization capabilities, making it a promising tool for clinical applications in brain tumor diagnosis.