eye-sight-insight

• 

  ai powered diabetic retinopathy detection

Inspiration

The idea for Eye-Sight-Insight stemmed from conversations with my elder brother few years back, who is pursuing an MS in Ophthalmology. During his academic journey, he worked on a research paper around retinal diseases and mentioned the alarming rise in diabetic retinopathy cases. He also pointed out the lack of accessible, automated tools for early detection in real-world settings especially in rural and low-resource regions. That conversation sparked my desire to build a solution that uses AI for social good.

What it does

Eye-Sight-Insight is a web-based AI tool that allows users (doctors or patients) to upload retinal fundus images and instantly receive predictions about the presence of diabetic retinopathy. It returns the predicted stage of the disease and confidence score to reflect model certainty with a clean, fast UI to make the tool accessible for healthcare use.

How we built it

Used EfficientNetB0, a pretrained CNN model, for image classification. Deployed using FastAPI with DockerFile pushed into HuggingFace space using GitBash to build a scalable and efficient backend. Integrated with Lovable to create an interactive frontend that uploads images and displays results. The entire system is built with extensibility in mind, allowing future integration with EMR or PACS systems.

Challenges we ran into

Preprocessing and formatting retinal images correctly for EfficientNet input. Handling cases where the model returned prediction without confidence, causing NaN% UI errors. Troubleshooting backend deployment and matching frontend expectations. Working with a pretrained model without custom fine-tuning due to dataset limitations

Accomplishments that we're proud of

Successfully created a working prototype that bridges AI and healthcare. Built an end-to-end system with real-time prediction, usable UI, and backend integration. Tackled and solved deployment issues and confidence calculation errors. Created something that could genuinely help doctors and patients.

What we learned

How to use transfer learning for medical image classification. Effective use of FastAPI + Lovable to connect ML models to real-world apps. Importance of robust error handling and user feedback in healthcare tools. Insights into how clinical workflows (EMRs) can be enhanced using ML.

What's next for eye-sight-insight

Integrating with an EMR app for hospital deployment. Improving the model with custom training using larger datasets like EyePACS or APTOS. Adding visual explanations like Grad-CAM to make predictions explainable. Partnering with clinics to pilot the tool in real-world screening environments.

Built With

• docker
• fastapi
• gitbash
• huggingface
• kaggle
• lovable