retfound-visionai

<div align="center">

👁️ RETFound VisionAI — Retinal Disease Classifier

Fine-Tuned Medical Vision Transformer for Multi-Condition Retinal Screening

</div>

---

👥 Credits & Roles

---

👨‍💻 Fine-Tuning Engineer Profile

<table align="center"> <tr> <td align="center"> <h3><b>Karthikeya Chavala</b></h3> <p><i>AI Engineer & Full-Stack Developer</i></p> <p> <a href="https://github.com/Karthikeya-Chavala7893"><b>GitHub Profile</b></a> • <a href="https://github.com/Karthikeya-Chavala7893/Eye_Diseases_Classification"><b>VisionAI Repository</b></a> • <a href="mailto:[email protected]"><b>Contact Email</b></a> </p> </td> </tr> </table>

🌟 Project Contributions:

• 🎯 Transfer Learning & Training: Fine-tuned the 307M-parameter RETFound ViT-Large architecture, reaching 91.62% classification accuracy.
• ⚡ Full-Stack AI Architecture: Built the VisionAI screening system (Next.js 14, Python Flask REST API, Firebase Auth/Firestore).
• 🛡️ Zero-Disk In-Memory Processing: Implemented volatile RAM inference (io.BytesIO) ensuring raw retinal photos are never stored on disk.

---

📊 Model Performance & Specifications

• Base Architecture: RETFound ViT-Large/16 (24 Transformer Layers, 16 Attention Heads)
• Foundation Pre-training: 1.6 Million unlabelled retinal fundus images (Nature, 2023)
• Input Resolution: $224 \times 224$ pixels (RGB Retinal Fundus Photograph)
• Test Set Accuracy: 91.62%
• Target Classes (4 Pathologies):
  • Normal (Healthy Retina)
  • Diabetic Retinopathy
  • Glaucoma
  • Cataract

---

⚠️ Medical & Research Disclaimer

IMPORTANT NOTICE FOR USERS & CLINICIANS:

1. Educational & Research Use Only: This artificial intelligence model is fine-tuned and published strictly for academic research, educational demonstration, and technological evaluation purposes.
2. Not an FDA / CE Cleared Medical Device: This model is NOT a certified medical diagnostic device. It has not been approved by the FDA, CE, CDSCO, or any regulatory healthcare authority for autonomous clinical diagnosis or treatment planning.
3. No Substitute for Professional Medical Advice: Predictions and confidence scores generated by this model must never replace clinical examinations, expert ophthalmic interpretations, or diagnostic consultations by licensed ophthalmologists or healthcare professionals.
4. No Liability: The fine-tuner and base model authors assume no legal responsibility or clinical liability for decisions or outcomes resulting from the use or interpretation of this model.

---

🔬 Scientific Citation

@article{zhou2023foundation,
  title={A foundation model for generalizable disease detection from retinal images},
  author={Zhou, Yukun and Chia, Mark A and Wagner, Siegfried K and others},
  journal={Nature},
  volume={622},
  number={7981},
  pages={156--163},
  year={2023},
  publisher={Nature Publishing Group}
}