Automated Brain Tumor Analysis from MRI Using Deep Learning

Accurate brain tumor classification from MRI remains essential for computer-assisted diagnosis, yet manual interpretation is time-consuming and variable. This study presents an EfficientNet-B0-based convolutional neural network for multi-class classification of glioma, meningioma, pituitary tumors, and no-tumor cases. The model was trained and evaluated on a public MRI dataset of 7023 images using a strict patient-level split to ensure unbiased assessment. A fixed EfficientNet-B0 backbone with a lightweight classification head reduces overfitting while maintaining stable learning. Performance was assessed via accuracy, precision, recall, F1-score, and specificity. The model achieved class-wise (one-vs-rest) accuracies of 96.5% for glioma, 99.1% for no tumor, 95.6% for meningioma, and 97.9% for pituitary tumors, with high specificity across all classes. A controlled comparison against ResNet50 and MobileNetV2 under identical conditions shows that EfficientNet-B0 offers a balanced trade-off between predictive performance and computational cost, achieving competitive accuracy with significantly fewer parameters and faster inference than deeper architectures. This study provides a reproducible evaluation framework, patient-level validation protocol, and systematic backbone comparison to support efficient and reliable deep learning models for multi-class brain tumor classification.