Noninvasive Bloodglucose Prediction via Fundus Imagery

fundus image
D-eye digital opthalmascope
Dexcom CGM
after 5 Epochs , GAN is able to recognize features such as blood vessels

Inspiration

I wanted to build a solution to check blood sugar levels non-invasively. I have a Type 1 diabetic child and want to build tools that can help him manage his blood sugar levels and improve the quality of his life and of others with diabetes.

What it does

We wanted to investigate if it is possible to predict blood glucose levels in realtime via retinal scans.

Our work is motivated by this paper, where the researchers describe constriction of vascular smooth muscle due to hyperglycemia - https://bpspubs.onlinelibrary.wiley.com/doi/full/10.1111/bph.13399

Since the retina is rich in vascular smooth muscle, we would like to study the effects of constriction and expansion of such tissue and determine if it is correlated to blood glucose levels

How we built it

We captured fundus photographs using a D-eye digital ophthalmoscope. The retinal images were segmented using SIFT detectors . We then built a image super-resolution model based on SR-GAN . code is located on Github

Out plan was to collect realtime blood glucose readings using a Dexcom CGM and use that to build a prediction model to infer glucose readings from retinal imagery based on Attention models, but we ran out of time