Diabetic Retinopathy is the leading cause of blindness in adults as the eye is the first organ to be affected, with 40% to 45% of Americans being affected. This disease as well as others can be detected with a small digital camera attached to any smartphone called a digital ophthalmoscope.
Existing diagnosis requires a hospital visit with an average cost of 250 USD for a consultation with a specialist.
Our solution can be used en masse provided the diagnoses has the mobile opthalmascope and a smartphone with a camera.
What it does
No special medical knowledge is required to capture a photo of a patients retina and the app we built will then collect the image and then give a diagnosis together with an accuracy of the given prediction.
We have an app that collects an image from a mobile ophthalmoscope and uses a model trained on retinal images labeled with varying levels of severity of diabetic retinopathy, from non-existant to a proliferative infection.
How we built it
We trained a model to about 72% categorical accuracy that can recognise retina images taken under a variety of imaging conditions. We then built an app that can collect an image taken by the ophthalmoscope from the smartphones gallery and pass it to the model to provide a diagnosis.
Challenges we ran into
We started with a low accuracy of the model but through lots of iteration on the model architecture and hyperparameter tuning we were able to achieve a reasonable accuracy. Getting the converted tflite model to work also took time.
Accomplishments that I'm proud of
We have been able to create a diabetic retinopathy diagnoser that can detect the retinopathy in its early stages. Early diagnosis and early treatment will slow down its prognosis to blindness.
What I learned
Using deep learning, we can make a huge impact on other people's lives.
What's next for Visual Diagnoser
We plan to be the Uber for preventative healthcare where we connect doctors who can offer more affordable treatment to patients with a positive diagnosis of diabetic retinopathy.