Explainable-Swin-Unet-Brain-Tumor Official implementation of Hybrid Swin-Unet for Brain Tumor Detection with Grad-CAM Explainability

Explainable Hybrid Neural Networks for Trustworthy Brain Tumor Detection

This repository contains the implementation of a Hybrid Swin-Unet architecture for multi-class brain tumor classification from MRI scans. The model integrates hierarchical Swin Transformers with a U-Net decoder to capture both global context and local spatial features.

Key Features:

1. Hybrid Swin-Unet Architecture: Combines transformer-based global modeling with CNN-like local feature recovery.
2. Weighted Random Sampling (WRS): Addresses class imbalance across Glioma, Meningioma, and Pituitary tumor types.
3. Explainable AI (XAI): Integrated Grad-CAM heatmaps for clinical validation and transparency.
4. High Performance: Achieves 97.94% accuracy on the benchmark Brain Tumor MRI dataset.

Installation:

1. Clone the repository:
2. git clone git clone https://github.com/masifchishti/Explainable-Swin-Unet-Brain-Tumor.git
3. cd Explainable-Swin-Unet-Brain-Tumor
4. Install dependencies:
5. pip install torch torchvision numpy matplotlib seaborn scikit-learn

How to Run:

1. Training and Evaluation

To train the model and generate performance metrics (Confusion Matrix, ROC, etc.): python main_training.py

2. Generate Grad-CAM Heatmaps

To visualize the model's decision-making on a specific MRI slice: python visualize_gradcam.py --image_path path/to/mri_image.jpg

Results Summary:

 Accuracy: 97.94%  Mean AUC: 0.98  Inference Time: 28ms (NVIDIA Tesla T4)

Visual Results

Grad-CAM Explainability

The model provides explainable predictions using Grad-CAM heatmaps.

Model Performance

Repository Structure:

 model/: Swin-Unet architecture definition.  utils/: Data loaders, preprocessing, and WRS implementation.  results/: Saved high-resolution figures (PNG).  weights/: Trained model checkpoints.

License:

This project is licensed under the MIT License - see the LICENSE file for details.