Using MobileNetV2 to Classify the Severity of Diabetic Retinopathy

Sarah Sheikh and Uvais Qidwai (Qatar University, Qatar)

With the increase in the number of diabetic patients, diabetic retinopathy has become a serious cause of concern. Screening programs aimed at screening diabetic patients at regular intervals have been put forward in many developed countries but the increasing cost of operating such programs and the global shortage of health care professionals brings forth the need for artificially intelligent systems to be used for grading purposes. Having POCT (Point-of-Care-Technology) based smartphone screening tools which will eliminate the need of having health care providers needed just to classify the retinal fundus scans. In this paper we trained, validated and then tested our classifier which is based on the MobileNetV2 architecture, a computationally efficient classification system on a retinal fundus custom dataset created by the amalgamation of 3 different publicly available datasets popularly used. We enhanced the retinal features using bio-inspired retinal filters and tuned the hyper-parameters to achieve an accuracy of 91.6% and an accuracy of 0.9. The macro precision, recall, and f1-score values are 77.6%, 83.1%, and 80.1% respectively. Our results demonstrate that our model achieves state-of-the-art results when compared with previous related work and can be deployed in a mobile application for clinical testing.

Conference: UKSim-AMSS 22nd International Conference on Computer Modelling and Simulation, UKSim2020

Published: Mar 25, 2020

DOI: 10.5013/IJSSST.a.21.02.16