DED Detector

Inspiration

While growing up, my grandmother could recognise people only by their voices. She was one of the many millions that had lost their vision to diabetic eye disease. Early diagnosis is essential to the management and prevention of vision loss. It is my desire to use AI learning tools to create solutions that can aid reduce this burden.

What it does

It detects the different stages of Diabetic eye disease from mild to severe currently with 74% accuracy.

How we built it

We used transfer learning, taking advantage of one of the recent convolutional neural networks EfficientNet. Due to the constraint of time and computation, the algorithm was trained on a small sample. We used data from link the images were preprocessed by applying gausian filters and also transformations like resizing, zooming in, width and height shifting. After training for a few times and doing random hyperparameter tuning, The best accuracy was 74%.

Challenges we ran into

Due to time constraint, we could not train on larger dataset and for longer period(epochs).
We had to use limited GPU resources. With more computational power, we could have trained more samples faster.
Technical challenges - we could not deploy our web app on heroku due to limitation of size. We decided to deploy on google cloud platform.

Accomplishments that we're proud of

We were able to train a model and deploy it using streamlit. We got a good accuracy that can serve as a baseline model. Applying our knowledge of AI to offer practical solutions.