README.md

metadata

title: Coronary Artery Segmentation
emoji: 🫀
colorFrom: red
colorTo: pink
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: mit

🫀 Coronary Artery Segmentation with Fine-tuned SAM2

An interactive tool for segmenting coronary arteries in X-ray angiography images using fine-tuned Segment Anything Model 2 (SAM2). This Space provides an easy-to-use interface for medical image segmentation with point-based prompting.

---

🎯 How to Use This Space

Step 1: Upload Your Image

• Click on the image upload area or drag and drop your coronary X-ray angiography image
• Supported formats: JPG, PNG, and other common image formats
• The image will be displayed in the interface for annotation

Step 2: Select a Model

Choose from four fine-tuned SAM2 variants optimized for coronary artery segmentation:

• SAM2 Hiera Tiny - Fastest inference, good for quick results
• SAM2 Hiera Small - Balanced speed and accuracy
• SAM2 Hiera Base Plus - Higher accuracy, moderate speed
• SAM2 Hiera Large - Best accuracy, slower inference

Step 3: Add Point Prompts

🟢 Positive Points (Green)

• Click on areas you want to include in the segmentation
• Use positive points to mark the coronary arteries you want to segment
• Multiple positive points help the model understand the full extent of the artery
• Tip: Add points along the length of the artery for better coverage

🔴 Negative Points (Red)

• Click on areas you want to exclude from the segmentation
• Use negative points to refine the segmentation and remove unwanted regions
• Helpful for separating overlapping vessels or removing background
• Tip: Add negative points near the artery boundaries to improve precision

🎨 How to Add Points

1. The first click adds a positive point (green)
2. Hold Shift while clicking to add a negative point (red)
3. You can add multiple points of each type
4. Points are displayed as colored dots on your image

Step 4: Generate Segmentation

• Click the "Segment" button to process your annotations
• The model will generate a segmentation mask based on your point prompts
• The result shows the segmented coronary artery highlighted on the original image

Step 5: Refine (Optional)

• If the segmentation isn't perfect, add more points:
  • Add positive points to include missing artery sections
  • Add negative points to exclude over-segmented regions
• Click "Segment" again to update the result
• Repeat until you're satisfied with the segmentation

Step 6: Clear and Start Over

• Use the "Clear" button to remove all points and start fresh
• Upload a new image to segment a different case

---

💡 Tips for Best Results

1. Start Simple: Begin with 1-2 positive points on the main artery structure
2. Be Strategic: Place positive points at key locations (branches, endpoints, curves)
3. Refine Gradually: Add negative points only where the model over-segments
4. Model Selection: Start with Tiny or Small models for faster iteration, then use larger models for final results
5. Multiple Vessels: For multiple arteries, focus on one at a time with clear positive points
6. Contrast Matters: Images with good contrast between vessels and background work best

---

🔬 Model Architecture

The workflow consists of several key stages:

1. Image Preprocessing - Normalization and enhancement for optimal model input
2. Point Prompt Encoding - Your positive/negative clicks are encoded as spatial prompts
3. SAM2 Processing - Fine-tuned encoder-decoder processes the image with prompts
4. Mask Generation - High-quality segmentation mask output
5. Visualization - Segmented artery overlaid on the original image

---

📊 Model Variants

Model	Parameters	Speed	Accuracy	Best For
Hiera Tiny	~38M	⚡⚡⚡	⭐⭐⭐	Quick experiments, real-time feedback
Hiera Small	~46M	⚡⚡	⭐⭐⭐⭐	Balanced performance, general use
Hiera Base Plus	~80M	⚡	⭐⭐⭐⭐⭐	High-quality results, clinical evaluation
Hiera Large	~224M	⚡	⭐⭐⭐⭐⭐	Best possible accuracy, research

All models have been fine-tuned specifically for coronary artery segmentation on X-ray angiography images.

---

⚠️ Disclaimer

This tool is intended for research and educational purposes only. It is not approved for clinical diagnosis or treatment decisions. Always consult qualified healthcare professionals for medical image interpretation.

---

🚀 Local Installation

If you want to run this application locally, follow these setup instructions:

1. Create Conda Environment

# Create a new conda environment named sam2_FT_env with Python 3.10
conda create -n sam2_FT_env python=3.10.0 -y

# Activate the environment
conda activate sam2_FT_env

2. Install SAM2 Library

# Clone the official repository
git clone https://github.com/facebookresearch/segment-anything-2.git
cd segment-anything-2

# Install the package in editable mode
pip install -e .
cd ..

3. Install Dependencies

pip install uv
uv pip install gradio opencv-python-headless torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

4. Run the Application

# Ensure model checkpoints are in the correct directories
python app.py

Open http://127.0.0.1:7860 in your browser.

---

📚 Citation

If you use this tool in your research, please cite:

@article{ravi2024sam2,
  title={SAM 2: Segment Anything in Images and Videos},
  author={Ravi, Nikhila and others},
  journal={arXiv preprint arXiv:2408.00714},
  year={2024}
}

---

📧 Contact & Support

For questions, issues, or feedback, please open an issue on the GitHub repository or contact the developers.

Built with ❤️ using SAM2 and Gradio