Retino AI

AI That Sees Before It’s Too Late.

Comment

Inspiration

Diabetic Retinopathy (DR) is one of the leading causes of preventable blindness worldwide. Early detection significantly reduces the risk of severe vision loss, yet many patients — especially in rural and under-resourced regions — lack timely access to ophthalmologists. We were inspired to build Retino AI to bridge this gap using artificial intelligence. Our goal was to create a scalable, accessible screening tool that assists in early diagnosis and supports healthcare professionals.

What it does

->Retino AI is an AI-powered web application that: ->Analyzes retinal fundus images ->Predicts the stage of Diabetic Retinopathy ->Provides confidence scores ->Suggests precautionary steps based on severity ->The system classifies images into:

  1. No DR
  2. Mild
  3. Moderate
  4. Severe
  5. Proliferative DR This enables faster preliminary screening and helps prioritize high-risk patients.

How we built it

We built Retino AI using: 🧠 Machine Learning:

->PyTorch for model training and inference ->SwinViT architecture for efficient image classification ->Image preprocessing using torchvision transforms

🌐 Frontend:

->Google AI Studio prototype interface ->Interactive UI with image upload and prediction display ->Fetch API integration for prediction calls

⚙ Backend (Model Serving):

->FastAPI for creating a prediction endpoint ->REST API to connect frontend and model ->We trained the model on retinal datasets and optimized it for multi-class DR classification.

Challenges we ran into

Integrating the trained .pth model with a cloud-based frontend Handling cross-origin (CORS) issues between frontend and backend Ensuring consistent image preprocessing between training and inference Debugging deployment and server accessibility issues Managing limited time during development These challenges pushed us to deeply understand model serving and frontend-backend integration.

Accomplishments that we're proud of

Successfully training a multi-class DR classification model Building an end-to-end pipeline from image upload to prediction Designing an intuitive and user-friendly interface Integrating AI into a real-world healthcare use case Overcoming deployment and integration obstacles Most importantly, we transformed a research idea into a functional prototype.

What we learned

Practical experience in deploying deep learning models Model optimization for real-world applications API integration between frontend and backend systems Debugging deployment and server communication issues The importance of accessibility and usability in healthcare AI We also learned how crucial collaboration and structured problem-solving are in hackathon environments.

What's next for Retino AI

We plan to: Improve model accuracy using larger, more diverse datasets Add Grad-CAM visualization for explainable AI Deploy on a scalable cloud infrastructure Develop a mobile-friendly version Integrate with hospital management systems Seek clinical validation with healthcare professionals Our long-term vision is to make Retino AI a reliable AI-assisted screening tool for global healthcare accessibility.

Built With

Share this project:

Updates