README.md

---
pipeline_tag: image-to-text
license: mit
library_name: transformers
---

<div align="center">

<img src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg" width="120"/>

# 🩺 MedFusion-AI (Unified Radiology Pipeline)

**Pro + Lite unified radiology model — powered by MedSigLIP & MedGemma**

[![Use this model](https://img.shields.io/badge/💻_Use-Model-2C3E50?style=for-the-badge&logo=huggingface&logoColor=yellow)](https://huggingface.co/spaces/fokan/MedFusion-AI)
[![Deploy](https://img.shields.io/badge/🚀_Deploy-HF_Space-2C3E50?style=for-the-badge)](https://huggingface.co/spaces/fokan/MedFusion-AI)
[![Train](https://img.shields.io/badge/🧠_Train-Colab-2C3E50?style=for-the-badge&logo=googlecolab)](https://colab.research.google.com/github/fokan/MedFusion-AI)
[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg?style=for-the-badge)](LICENSE)

</div>

---

## 🧠 About the Model
**MedFusion-AI** is a unified multimodal medical-AI pipeline integrating vision encoders and text decoders to produce full radiology reports from X-ray or DICOM inputs.

| Mode | Encoder | Decoder | Precision |
|------|----------|----------|------------|
| **Pro** | `fokan/medsiglip-448-fp16-pruned20` | `fokan/medgemma-4b-it-fp16-pruned20` | FP16 + Pruned (High accuracy) |
| **Lite** | `fokan/medsiglip-448-int8` | `fokan/medgemma-4b-it-int8` | INT8 (Compact & fast) |

---

## 🩻 Usage (Python)

```python
from medfusion_pipeline import MedFusionPipeline

pipe = MedFusionPipeline.from_pretrained(".", mode="pro")  # or 'lite'
report = pipe.analyze("sample_xray.jpg")
print(report)
```

---

## ⚙️ Modes
- **pro** → FP16 + Pruned (High accuracy)
- **lite** → INT8 (Compact speed-optimized)

---

## 💡 Features
- Handles **X-ray / DICOM inputs** automatically  
- Generates **structured radiology reports**  
- Plug-and-play **dual pipeline (Pro & Lite)**  
- Optimized for **medical education + research**  

---

## 🧩 Deployment Options
| Platform | Description |
|-----------|--------------|
| **🤗 Hugging Face Spaces** | One-click Gradio demo or inference API |
| **🚀 HF Inference Endpoint** | GPU-backed endpoint for production |
| **💻 Local Deployment** | Python + Torch runtime (CPU/GPU friendly) |

---

## 📊 Model Specs
- Architecture: MedSigLIP encoder + MedGemma decoder  
- Params: ~4 B (Teacher) → ~0.4 B (Student Distilled)  
- Input Resolution: 224 / 448 px  
- Optimized for: Chest X-rays & general radiographs  

---

## 📜 Citation
If you use **MedFusion-AI** in research, please cite:

```
@software{fokan_medfusion_ai_2025,
  title={MedFusion-AI: Unified Radiology Encoder-Decoder Pipeline},
  author={Karrar Alhdrawi},
  year={2025},
  url={https://huggingface.co/fokan/MedFusion-AI}
}
```

---

<div align="center">
Built with ❤️ by <a href="https://huggingface.co/fokan">fokan</a>
</div>