Diabetic Reinopathy Classification using Deep Learning

Abstract

With diabetes growing at an alarming rate, changes in the retina of diabetic patients causes a condition called diabetic retinopathy which eventually leads to blindness. Early detection of diabetic retinopathy is the best way to provide good timely treatment and thus prevent blindness. Many developed countries have put forward well-structured screening programs which screens every person diagnosed with diabetes at regular intervals. However, the cost of running these programs is increasing with ever increasing disease burden. These screening programs require well trained opticians or ophthalmologist which are expensive especially in developing countries. A global shortage of health care professionals is putting a pressing need to develop fast and efficient screening methods. Using artificial intelligent screening tools will help process and generate a plan for the patients thus skipping the health care provider needed to just classify the disease and will lower the burden on health care professional’s shortage significantly. A plethora of research exists to classify severity of diabetic retinopathy using traditional and end to end methods. In this thesis, we first trained and compared the performance of lightweight architecture MobileNetV2 with other classifiers like DenseNet121 and VGG16 using the Retinal fundus APTOS 2019 Kaggle dataset. We experimented with different image reprocessing techniques and employed various hyperparameter tuning techniques, and found the lightweight architecture MobileNetV2 to give better results in terms of AUC score which defines the ability of the classifier to separate between the classes. We then trained MobileNetV2 using handpicked custom dataset which was an amalgamation of 3 different publicly available datasets viz. the EyePacs Kaggle dataset, the APTOS 2019 Blindness detection dataset and the Messidor2 dataset. We enhanced the retinal features using bio-inspired retinal filters and tuned the hyper-parameters to achieve an accuracy of 91.68% and AUC score of 0.9 when tested on unseen data. The macro precision, recall, and f1-scores are 77.6%, 83.1%, and 80.1% respectively. Our results demonstrate that our computational efficient light weight model achieves promising results and can be deployed as a mobile application for clinical testing.