In India ,the number of people affected by Diabetes is over 65 million.1.
According to the WHO, 31.7 million people were affected by diabetes mellitus (DM) in India in the year 2000. This figure is estimated to rise to 79.4 million by 2030, the largest number in any nation in the world.
Almost two-third of all Type 2 and almost all Type 1 diabetics are expected to develop diabetic retinopathy (DR) over a period of time.2,3,4At present, diabetic retinopathy (DR) is a slow and strenuous process that requires trained Ophthalmologists to analyze color photographs of retinas. They can then classify the level of deterioration the patient’s eye has experienced into one of four categories. While this process is effective, it is very slow. On an average it takes more than 48 hours for the reports of the patient to come on record. Furthermore, in areas where access to trained clinicians or suitable equipment is limited such as rural areas, individuals are left without any support. As the statistics suggest the increase in the diabetic retinopathy cases, such a system may well become insufficient.
An Automated grading of diabetic retinopathy has potential benefits such as increasing efficiency, reproducibility, and a larger coverage of screening programs; hence increasing the access to the resources for the public in need; and improving patient outcomes by providing early detection and treatment. To maximize the accuracy metrics of the automated grading, an algorithm to detect imputable diabetic retinopathy is needed. We propose a model for classifying retina images as having DR using Convolutional Neural Networks trained with transfer learning. At a later stage a detailed comparison of various other image classifiers such as – SVMs, Random Forests, Random Kitchen Sink (RKS) etc.
, will be tabulated. The input to the model is a pre-processed 256px x 256px retina image, and the output is a class label indicating whether or not the retina has DR.