Brain Tumor detection and Classification Using Deep Learning

Brain tumor (meningioma, glioma, pituitary) detection is crucial for improving survival rates and prognosis. MRI-based image analysis plays a key role in early detection—even in nascent stages. We propose an ensemble leveraging transfer learning with five fine-tuned pre-trained CNNs: GoogleNet, AlexNet, SqueezeNet, ShuffleNet, and NASNet-Mobile. Each model is adapted by replacing final layers with a 3-neuron output and aggressive fine-tuning (lr = 10, dropout = 0.5). To address data scarcity, we use geometric and intensity-based augmentations and five-fold cross- Validation. All models perform well independently—NASNet-Mobile achieves 94.1%, and AlexNet 92.3%. However, a majority voting system combining the top 3 performers (GoogleNet, SqueezeNet, and ShuffleNet) boosts accuracy to 98.6% and precision to 96.2% on the Figshare dataset, outperforming individual models. Model interpretability is ensured through Grad-CAM visualizations, with tumor localization matching radiologist annotations 95% of the time and reducing false positives by 15%. Lightweight architectures (e.g., SqueezeNet: 1.2M parameters) also favor computational Efficiency. This framework is consistent across varied MRI scans and can be integrated into diagnostic workflows, demonstrating clinical value. The main contributions are: (1) A modular transfer learning pipeline for medical imaging tasks, (2) Enhanced interpretability via tumor- segmented GradCAM heatmaps, and (3) A voting ensemble for improved classification stability (F1 score: 0.93 vs. 0.89 for single models). Applied to 3,064 images, this method delivers top-tier results without complex hybrid setups. Future work will explore federated learning for multi-institutional collaboration under privacy-preserving settings.