February 2020
This year, I had the opportunity to work on a personal project alongside two friends whom I share an interest in medicine and healthcare with. We wanted to develop a screening tool that could be used by physicians and patients to diagnose diabetic retinopathy, a leading cause of blindness.
It is generally known that diabetes is caused by the malfunction or insensitivity of insulin and the disease affects millions globally. If diabetes is not managed properly, cells in different parts of the body will begin to lose function and cause severe symptoms. Although diabetes is commonly understood, its symptoms and co-morbidities are less discussed.
Our goal was to develop a web-based prescreening tool that would utilize deep learning to diagnose stages of diabetic retinopathy. The idea was that clinicans would be able to load retinal images into the website taken from patient visits and recieve a diagnosis from 1 - 4 of the severity of the retinopathy. Another possible use case was that diabetes patients would be able to take and load these images into the webapp as well. There’s a few tutorials on the internet that involve using a camera phone to photograph your retina, but it probably requires a friend and a steady hand.
1 - No Diabetic Retinopathy
2 - Mild
3 - Moderate
4 - Severe
Our deep learning algorithm was based on convolutional neural networks, which we implemented using ResNet34 and pytorch. The ResNet models are based on ImageNet, which provides a vast array of images for classification and training. With Resnet, we were able to take off the top few layers of the pre-trained model and train according to our retinal image set.
The result? We were able to accurately diagnose the stage of diabetic retinopathy given retinal scans with an error rate of 18%. Pretty cool!
