Domain Modelling For A Lightweight Convolutional Network Focused On Automated Exudate Detection in Retinal Fundus Images

Retinal Fundus images contain important signs implying various stages of diabetic retinopathy. One of the early signs of diabetic retinopathy is exudates. The timely detection of this sign from fundus images can prevent or at least suppress the advancement of this condition. To this effect, this research presents a CNN-based automated exudate detection architecture for the timely detection of this early sign. In order to overcome the challenges of a limited dataset several domain specific augmentations are proposed for improving the generalization capacity of the developed architecture. The lightweight architecture consists of only 6.42 million parameters, compared to Resnet-18 (11.69 million) whilst achieving an overall F1 score of 89%