Deep Learning for Diabetic Retinopathy (DR) Classifier

Abstract

High prevalence in diabetes contributed to the rising of the complications such as diabetic retinopathy (DR) which cause the patient to suffer from vision loss and this vision loss may become permanent if it is not well manage and treat. Including other medication non-adherence factor, this rising trend has been made more challenging for physician in keep apace with demand using manual method of retina screening to diagnose DR. Therefore, this project aimed to develop a classifier of deep learning for DR in 5 different level of severity which are normal (no DR), mild, moderate, severe and proliferative DR, to speed up the clinical assessment for early DR detection or DR monitoring progression of disease. The classification was applied using Scottish grading with consider three categories for fundus abnormalities which are hemorrhage, lipid exudates and microaneurysms for a training from scratch model and a pre-trained model using Inception V3. Verified and trustable dataset of retina screening or fundus image available online such as kaggle.com used to train and test of the system. Besides, the execution of this project was completed via on cloud training using Google Cloud Platform (GCP). The obtained results are measuring the classifier performance tested in term of accuracy, sensitivity, and specificity. Transfer learning using Inception V3 has been shown a better performance compare to own training model with accuracy 81.2% using transfer learning Inception V3 and 74.7% for own training model.