Hybrid 3B Net and EfficientNetB2 Model for Multi-Class Brain Tumor Classification

Brain tumor classification and diagnosis are critical for timely and effective treatment, as brain tumors can severely impact patient health and survival. This paper introduces an enhanced method for the multiclass classification of brain tumors using a novel three-branch convolutional neural network integrated with EfficientNetB2 feature fusion. We systematically compared multi-branch architectures, ranging from single to six branches, to identify the optimal configuration for accurate tumor classification. The three-branch network (3B Net) consistently yielded superior performance across multiple classification tasks. Our experiments, conducted across three distinct datasets, involved binary classification (tumor vs. normal), three-class classification (meningioma, glioma, and pituitary), and four-class classification (non-tumor, meningioma, glioma, and pituitary). The proposed method achieved a binary classification accuracy of 99.50%, a three-class classification accuracy of 94.95% without augmentation, 98.72% with augmentation, and a four-class classification accuracy of 97.80%. Applying data augmentation techniques to the three-class classification dataset led to a marked improvement in classification accuracy. Blind test evaluations confirmed the robustness of the model for both binary and four-class classification, while five-fold cross-validation for the augmented three-class classification dataset further validated its performance. The 3B Net demonstrated the highest accuracy and robustness, significantly outperforming other network configurations. The integration of EfficientNetB2 for feature extraction further enhanced the model’s classification capabilities. These results underline the effectiveness of our 3B Net architecture, suggesting its potential for deployment in automated brain tumor diagnosis and improvement in clinical decision-making processes.