Deep Learning Framework for Brain Tumor Classification and Segmentation
NeuroSeg Classification is a comprehensive deep learning framework designed for automated brain tumor classification and segmentation from MRI scans. The project implements and compares multiple training strategies including separate, joint, and attention-based U-Net architectures to achieve state-of-the-art performance on the BRISC 2025 dataset.
- Multi-Task Learning: Simultaneous tumor classification and segmentation
- Joint Training Architecture: Shared encoder for improved feature learning
- Attention Mechanisms: Enhanced U-Net with attention gates
- Comprehensive Evaluation: Detailed metrics and visualizations
- Efficient Training: 30% faster than separate training approaches
Dataset Statistics:
| Metric | Value |
|---|---|
| Total Samples | 6,000 |
| Image Size | 224×224 |
| Format | Grayscale MRI |
| Training Set | 4,000 (66.2%) |
| Validation Set | 1,000 (16.6%) |
| Test Set | 1,000 (17.2%) |
Class Distribution:
| Class | Samples | Percentage |
|---|---|---|
| Glioma | 1,401 | 23.4% |
| Meningioma | 1,635 | 27.3% |
| No Tumor | 1,207 | 20.1% |
| Pituitary | 1,757 | 29.3% |
Dataset Source: BRISC 2025 on Kaggle
-
Classification Model (U-Net Encoder + Classifier)
- U-Net encoder for feature extraction
- Fully connected classifier head
- CrossEntropy loss with class weighting
-
Segmentation Model (Standard U-Net)
- Encoder-decoder architecture
- Skip connections for spatial information
- Combined BCE + Dice loss
-
Joint Training Model
- Shared encoder for both tasks
- Separate decoder heads for classification and segmentation
- Multi-task loss optimization
-
Attention U-Net
- Attention gates in skip connections
- Enhanced feature selection
- Improved segmentation accuracy
| Parameter | Value |
|---|---|
| Learning Rate | 1e-4 |
| Batch Size | 32 |
| Optimizer | AdamW |
| Weight Decay | 1e-3 |
| Loss Function | CrossEntropy (Weighted) |
| Early Stopping | Patience=15 |
| Epochs | 100 |
| Parameter | Value |
|---|---|
| Learning Rate | 1e-3 |
| Batch Size | 16 |
| Optimizer | AdamW |
| Weight Decay | 1e-3 |
| Loss Function | BCE + Dice Loss |
| Scheduler | ReduceLROnPlateau |
| Early Stopping | Patience=15 |
| Epochs | 100 |
| Parameter | Value |
|---|---|
| Learning Rate | 1e-3 |
| Batch Size | 12 |
| Optimizer | AdamW |
| Weight Decay | 1e-3 |
| Loss Function | BCE + Dice + CrossEntropy |
| Combined Metric | (mIoU + Accuracy) / 2 |
| Scheduler | ReduceLROnPlateau |
| Early Stopping | Patience=15 |
| Epochs | 100 |
All models were trained on CUDA-enabled GPU with the following characteristics:
- Smooth convergence across all training phases
- Effective early stopping to prevent overfitting
- Consistent improvement in validation metrics
| Model | Test Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|
| Separate Training | 0.9560 | 0.9571 | 0.9560 | 0.9557 |
| Joint Training | 0.9860 | 0.9860 | 0.9860 | 0.9860 |
Improvement: +3.0% accuracy with joint training
| Model | Test mIoU | Dice Score | Pixel Accuracy | Test Loss |
|---|---|---|---|---|
| Separate Training | 0.7563 | 0.8551 | 0.9952 | - |
| Joint Training | 0.7663 | 0.8613 | 0.9955 | - |
| Attention U-Net | 0.7734 | 0.8662 | 0.9956 | 0.1975 |
Best Model: Attention U-Net achieves highest segmentation performance
Joint training improves classification accuracy by ~1%
Joint training improves segmentation mIoU by ~2%
Shared encoder learns better general features
Multi-task learning provides regularization effect
Training time: 30% faster than separate training
Classification Model Performance:
- Strong diagonal dominance indicating high accuracy
- Minimal confusion between tumor types
- Perfect classification for pituitary tumors (300/300)
Joint Model Performance:
- Improved classification across all classes
- Reduced misclassifications compared to separate training
- Near-perfect results for no_tumor and pituitary classes
The visual results demonstrate:
- Accurate tumor localization across different MRI views
- Precise boundary detection
- Robust performance on various tumor types and sizes
- High agreement between ground truth and predictions
- Python 3.8+
- CUDA-capable GPU (recommended)
- 8GB+ RAM
# Clone the repository
git clone https://github.com/yourusername/neuroseg-classification.git
cd neuroseg-classification
# Create virtual environment
python -m venv venv
#Activate virtual environment
.\venv\Scripts\activate
# Install dependencies
pip install -r requirements.txtAll dependencies are listed in requirements.txt. Key packages include:
- PyTorch (>=2.0.0) - Deep learning framework
- Albumentations (>=1.3.1) - Advanced image augmentation
- OpenCV (>=4.8.0) - Image processing
- Scikit-learn (>=1.3.0) - Metrics and evaluation
- Matplotlib & Seaborn - Visualization
- NumPy - Numerical computing
- 98.6% Classification Accuracy on test set
- 77.34% mIoU for tumor segmentation
- 30% Training Time Reduction with joint learning
- Multi-task Learning improves both tasks simultaneously
- Attention Mechanisms boost segmentation performance
- Final Test Accuracy: 95.60%
- Best Validation Accuracy: 93.00% (Epoch 98)
- Training completed in 100 epochs
- Final Test mIoU: 75.63%
- Best Validation mIoU: 70.21% (Epoch 93)
- Dice Score: 85.51%
- Final Test Classification Accuracy: 98.60%
- Final Test Segmentation mIoU: 76.63%
- Combined Score: 84.20%
- Convergence: Epoch 99
- Final Test mIoU: 77.34%
- Dice Score: 86.62%
- Best Performance: Epoch 100
Abdullah Al Fahad
Student of Computer Science and Engineering
BRAC University
- Dataset: BRISC 2025 Dataset from Kaggle
- Framework: PyTorch Deep Learning Framework
- Architecture Inspiration: U-Net and Attention U-Net papers
- Training Infrastructure: CUDA-enabled GPU acceleration
This project is licensed under the MIT License - see the LICENSE file for details.
Abdullah Al Fahad
