Feature Pool Exploitation for Disease Detection in Fundus Images

Abstract

Retinal fundus diseases without immediate diagnoses and treatment may lead to serious consequences such as permanent visual impairment. Recently, many machine learning (ML) and deep learning (DL) models have been introduced for fundus image classification. However, those heavy models require high-end graphics processing units for training and testing, thus not suitable for real-case usage in fundus cameras with limited computation power. In this paper, we demonstrate the effectiveness of our proposed two-block model, Feature Exploitation Lightweight network (FEL-net) which consists of a Feature Exploitation Block (FEB) and Lightweight Classification Block (LCB) by comparing it with other DL models. The experiment was carried out on a dataset of 21, 697 fundus images and our model achieves 99% binary classification accuracy. The proposed robust FEB is used to generate a refined feature pool for fundus images to build an efficient ML classifier that can distinguish fundus images with disease from the normal case with high accuracy and low running time.