X-RAI: Bone Fracture Detection from X-ray Images

Overview

X-RAI is a deep learning project that automatically detects bone fractures from musculoskeletal X-ray images using the MURA v1.1 dataset. The project employs a lightweight convolutional neural network (MobileNetV2) to classify each image as "Fractured" or "Not Fractured."

Google colabnis located at: Google Colab

---

Features

• Fracture Classification: Achieves up to ~80% accuracy on the validation subset.
• Fast Training: Utilizes MobileNetV2 with resized images for efficient training.
• Model Interpretability: Includes Grad-CAM visualizations to highlight regions influencing the model's decisions.
• Label Overlays: Displays predictions with probabilities for better interpretation.

---

Stack

• Dataset: MURA v1.1 (Stanford ML Group)
• Model: MobileNetV2 (pretrained on ImageNet)
• Frameworks: PyTorch, Torchvision
• Visualization: Grad-CAM, Matplotlib
• Environment: Google Colab (T4 GPU) / Kaggle (GPU)

---

Setup for Google Colab

To ensure optimal performance, use a T4 GPU on Google Colab:

1. Open the notebook in Google Colab.
2. Navigate to Runtime → Change runtime type.
3. Select T4 GPU under Hardware Accelerator.
4. Click Save.

Verify GPU availability by running:

import torch
print("GPU Enabled:", torch.cuda.is_available())

---

Installation

Install dependencies in Colab:

!pip install kagglehub torch torchvision matplotlib scikit-learn tqdm grad-cam

---

Usage

1. Dataset Download: The notebook automatically downloads the MURA v1.1 dataset using KaggleHub.
2. Body Part Selection: Modify the bodypart variable (e.g., "humerus", "wrist") to analyze specific body parts.
3. Training: Execute the training loop (5 epochs by default). Adjust num_epochs for extended training.
4. Evaluation: Validation metrics (precision, recall, F1-score) are printed post-training.
5. Visualization: Grad-CAM highlights model focus areas on a random validation sample.

---

Example Output

The notebook displays:

• A classification report with precision, recall, and F1-score.
• Side-by-side images of the original X-ray and Grad-CAM heatmap, annotated with actual and predicted labels.

---

Notes

• The T4 GPU accelerates training significantly compared to CPU-only mode.
• Free Colab tiers may occasionally assign other GPUs, but T4 is prioritized in most sessions.

---

License

This project is open-source and available for educational and research purposes. Cite the MURA dataset and relevant libraries if used in your work.

---

Acknowledgments

• MURA Dataset: Stanford ML Group
• Libraries: PyTorch, Torchvision, Grad-CAM, Matplotlib

For questions or contributions, open an issue or contact the author.