Visualising deep grading of diabetic retinopathy

Abstract

Diabetic retinopathy is a medical condition that occurs when fluctuating levels of blood sugar damage the blood vessels of the retina. Presence of vascular abnormalities in the retinal images indicate the presence of diabetic retinopathy. This thesis studies an automated method using convolutional neural network to diagnose and grade diabetic retinopathy in color red-green-blue retinal images. The images have a diabetic retinopathy grade as the ground truth. The experiment then applies the convolutional neural network trained on the data-set with the each image having a diabetic retinopathy grade, to a data-set having annotations by experts of the lesions for diabetic retinopathy. The thesis presents a method for visualizing the activations within the convolutional neural network, after training it for diabetic retinopathy diagnosis and grading. The experiments provide evidence that the trained network gives acceptable performance in grading diabetic retinopathy. Also, visualizing the activations provides the basis behind the grading decision made by the trained network.