srinath-wp/my-medsam2-endpoint

MedSAM2 - Custom Inference Endpoint

This repository contains a custom handler to deploy the MedSAM2 model using Hugging Face Inference Endpoints.

MedSAM2 is a state-of-the-art medical image segmentation model designed for zero-shot and fine-tuned applications across ultrasound, CT, and MRI modalities.

🔧 Files

• handler.py: Defines a custom inference handler for processing inputs and generating predictions.
• model.py: Loads the pretrained MedSAM2_latest.pt weights directly from the Hugging Face Hub.
• requirements.txt: Python dependencies required for running the model.
• MedSAM2_latest.pt: (Optional) If you are using local weights instead of downloading from the Hub.

📥 Input Format

The model expects a JSON payload with a base64-encoded image:

Example input: { "image": "" }

Encode an image like this in Python: import base64

with open("your_image.png", "rb") as f:
    encoded = base64.b64encode(f.read()).decode("utf-8")

📤 Output Format

The model returns predictions in a JSON object:

Example output: { "output": [[...]] }

Note: The current handler.py uses placeholder preprocessing. You should update it with the actual image preprocessing and model inference logic according to the MedSAM2 architecture.

🚀 Deployment Instructions

1. Push this repo to the Hugging Face Hub
2. Go to: https://huggingface.co/inference-endpoints
3. Create a new endpoint
4. Select:
   • Your model repo (e.g. username/my-medsam2-endpoint)
   • "Custom handler"
   • Hardware: Choose GPU if required
5. Deploy

🧠 Resources

• MedSAM2 Model: https://huggingface.co/wanglab/MedSAM2
• MedSAM2 Paper: https://arxiv.org/abs/2403.03662
• WangLab GitHub: https://github.com/WANG-lab/MedSAM

🤝 License

Please review the license of the original model here: https://huggingface.co/wanglab/MedSAM2

This repository is for research and educational purposes.