Rethinking model of EfficientNet-B9 for brain tumor classification: A high-precision deep learning approach

Automated brain tumor classification using magnetic resonance imaging (MRI) plays a crucial role in timely and accurate diagnosis, reducing reliance on manual interpretation. This study introduces a high-performing deep learning framework built upon a customized EfficientNet-B9 model, drawing inspiration from the architectural principles of EfficientNet-Elite-B9-V2. The proposed model integrates compound scaling strategies across depth, width, and resolution and processes high-resolution input images of size 800 × 800 pixels. The model exhibits strong feature extraction and generalization capabilities with approximately 1413 convolutional layers, optimized hyperparameters, and a high dropout rate of 0.7 for regularization. The training uses the performance of the Adam optimizer and binary cross-entropy loss, leading to stable convergence and minimal overfitting. When evaluated on the Br35H MRI dataset, the model achieves a classification accuracy of 98.33%, along with high precision, recall, and F1-score—demonstrating its effectiveness even without data augmentation. Despite being computationally demanding, this model have been used on conventional GPU computers, which makes clinical integration possible. These results highlight the advantages of high-resolution input processing and advanced architectural design for improving deep learning-based medical image analysis. • A new deep learning methodology is introduced for binary classification of MRI brain images into tumor and non-tumor categories, extending the EfficientNet model with increased depth, width, and resolution scaling specifically optimized for high-resolution MRI images. • A customized architecture incorporating enhanced MBConv layers, dropout regularization, and fully connected layers is designed to improve feature extraction and classification accuracy. • The proposed EfficientNet-B9 model is validated on multiple benchmark brain tumor datasets, demonstrating superior accuracy, sensitivity, and generalization compared to earlier EfficientNet variants and other techniques.