Overview This repository contains a comprehensive deep learning pipeline for brain tumor classification using state-of-the-art architectures including Vision Transformers (ViT), Convolutional Neural Networks (CNNs), hybrid models, and ensemble methods. The project demonstrates advanced techniques including knowledge distillation from a large ViT teacher model to an efficient MobileNetV2 student model. Dataset Combined Public MRI Dataset Sources: Merged datasets from BRISC, GTS AI, Mendeley, Figshare, and Zenodo Classes: 4 categories No tumor Glioma Meningioma Pituitary tumor Data Split: Training: 20,000 images (5,000 per class - balanced) Validation: 2,142 images Test: 2,311 images Preprocessing: All images resized to 224×224 pixels with augmentation techniques including elastic deformations, random flips/rotations, and color jittering Model Architectures 1. Vision Transformer (ViT) - Teacher Model Parameters: 235 million Architecture: 12 transformer encoder layers with multi-headed self-attention Patch size: 16×16 Embedding dimension: 768 Test Accuracy: ~94.6% Macro F1-score: ~0.93 2. MobileNetV2 - Student Model (Knowledge Distillation) Parameters: 2.4 million Training Method: Knowledge distillation from ViT teacher model Temperature: 5.0 for soft label generation Performance: Achieves near-teacher accuracy with 98% fewer parameters 3. Hybrid CNN-ViT Model Combines EfficientNet-B0 (CNN backbone) with ViT Features concatenated from both architectures Test Accuracy: 94.72% Total Parameters: 90.5 million 4. Ensemble Models (PyTorch) Ensemble 1: EfficientNet-B2 + DenseNet121 Test Accuracy: 92.43% Ensemble 2: DenseNet121 + VGG16 Test Accuracy: 92.8% Ensemble 3: EfficientNet-B2 + VGG16 Test Accuracy: 92.0% Training Details Optimization Optimizer: AdamW with learning rate 2×10⁻⁵ Loss Function: Cross-Entropy Loss Regularization: Dropout (0.5), weight decay (1×10⁻⁵) Learning Rate Schedule: ReduceLROnPlateau with warmup Early Stopping: Patience of 8-10 epochs Hardware Requirements GPU: Tesla P100-PCIE-16GB or equivalent Framework: PyTorch 2.6.0, TensorFlow 2.18.0 CUDA: 12.4 Key Features Knowledge Distillation Pipeline: Transfer learning from 235M-parameter ViT to lightweight MobileNetV2 Dataset Curation: Comprehensive preprocessing and balancing of multi-source MRI data Multi-Framework Implementation: Both PyTorch and TensorFlow/Keras implementations Extensive Evaluation: Classification reports, confusion matrices, ROC curves, and AUC scores Production-Ready Models: Saved checkpoints and training logs included File Structure ├── Dataset_workings_of_the_Brain_Tumor_Classification_for_Neuro_Triad_ViT.ipynb│ └── Dataset merging, augmentation, and balancing├── Neuro_Triad_ViT-235M-_Teacher_Model_with_MobileNet_V2_student_Model_for_Brain_Tumor_Classification_using_Knowledge_Distillation.ipynb│ └── ViT teacher + MobileNetV2 student with knowledge distillation├── 85M_parameterized_vision_transformer-Pretrained-ViT.ipynb│ └── Standalone ViT architecture and experiments├── hybrid_model_cnn_vit-CNN-ViT.ipynb│ └── Hybrid EfficientNet + ViT model├── effiecienrnetb2_densenet121-Ensemble.ipynb│ └── Ensemble: EfficientNet-B2 + DenseNet121├── densene121_vgg16-Ensemble.ipynb│ └── Ensemble: DenseNet121 + VGG16└── efficientnetb2_vgg16-Ensemble.ipynb └── Ensemble: EfficientNet-B2 + VGG16 Requirements # Core Dependencies torch>=2.6.0 torchvision>=0.21.0 tensorflow>=2.18.0 transformers>=4.52.4 # Data Processing numpy>=1.26.4 pandas opencv-python Pillow # Visualization matplotlib>=3.7.2 seaborn # Utilities scikit-learn tqdm Installation pip install torch torchvision tensorflow transformers numpy pandas opencv-python Pillow matplotlib seaborn scikit-learn tqdm Usage Each notebook is self-contained and can be run independently: Dataset Preparation: Run Dataset_workings...ipynb to prepare and balance the dataset Model Training: Execute any model-specific notebook Evaluation: Each notebook includes comprehensive evaluation metrics and visualizations Results Summary Model Test Accuracy Parameters Notes ViT Teacher 94.6% 235M Highest accuracy Hybrid CNN-ViT 94.72% 90.5M Best hybrid approach DenseNet121 + VGG16 92.8% - Top ensemble EfficientNet-B2 + DenseNet121 92.43% - - MobileNetV2 (Distilled) ~94% 2.4M Most efficient Citation Acknowledgments Brain tumor MRI datasets from BRISC, GTS AI, Mendeley, Figshare, and Zenodo Pre-trained models from Hugging Face and PyTorch/TensorFlow model zoos Hardware support from Kaggle/Colab GPU infrastructure License Apache 2.0 Contact For questions or collaborations, please open an issue in this repository or contact (bin.abdullah@northsouth.edu).