Ocula

415 million people across the world have diabetes, 33% of which have diabetic retinopathy (DR). Current methods of detecting DR, requires a trained ophthalmologists to review retinal images and it is difficult for them to categorize severity. Ocula is a novel innovation which uses neural networks to detect and group DR severity from 0 to 4 for easier prognosis of Diabetes.

For this project, 35,000 fundus images were used through the California Healthcare Foundation and EYEPACS – these images were verified / classified by ophthalmologists in 5 categories. These images came from different sources and different cameras, resizing and ‘noise removal was done for RGB channel using median filter. Five features (bifurcation of blood vessels, blot hemorrhages, edemas, exudates area, and microammeter) were used for training to focus on. A neural network consisting of 25 layers was built with 4 convolutional, RELU and pooling layers, along with three fully connected layers. Based on qualitative criteria, stochastic gradient method was used with learning rate of 0.01, as it provide 94% of accuracy along with 68 hours of processing on a single CPU. Prediction was done for 50 cycles consisting of 9,249 images. Consistently this resulted in accuracy close to or above 90% with a standard deviation of 1.2 and median of 91.9%.

This project shows that a neural network with multiple layers can be developed for prediction, classification of diabetic retinopathy for early prognosis of diabetes. My next steps are working with ophthalmologists to move into clinical trials of Ocula. The end goal is to completely develop this system so that diabetes detection can be highly precise and cost-efficient. I truly believe that with Ocula, we can achieve this, simply - One EYE at a time!

An example of two eye images with varying levels of Diabetes. To the human eye, these differences are almost indistinguishable. But not to Ocula!