Abstract
Diabetic retinopathy is a disease that damages the eyes of people suffering from diabetes (generally referred to as diabetics). The current diagnostic process is a tiresome operation that starts by making a doctor’s appointment, getting eyes scanned and then evaluated by an ophthalmologist. The doctor looks at the images and classifies the images personally and due to the doctor-patient confidentiality, these results cannot be disclosed over the phone. Now, patients must schedule a follow-up appointment and review the results to proceed with the treatment. This may take up to two weeks or more. In today’s day and age, this process is long and tedious and may cost significant time that could instead be used towards treatment. This project proposes a machine learning model that is built using a combination of sequential neural networks that has a stack of layers like Convolution 2D, Dense, Dropout, GaussianNoise, GaussianDropout, Flatten, BatchNormalization and EfficientNets. This model helps in detecting issues at an early stage and determine the possibility of Diabetic Retinopathy. When a person goes to a doctor and gets his/her eyes checked, the scanned images can be run through a classifier, and on that same day an ophthalmologist can then conclude if that individual has retinopathy or not and advise accordingly.
This model is evaluated using a data set from a Kaggle competition APTOS 2019 Blindness Detection that has retinal images taken using fundus photography (photographing the error of an eye) under an assortment of imaging conditions. Images have been rated by a clinician for intensity of diabetic retinopathy on a scale of 0 - 4. Numerous experiments on real world data validate that this model out shined the previous approaches in terms of various metrics for effective diagnosis of the disease.