A Deep Learning based Approach for Grading of Diabetic Retinopathy using Large Fundus image Dataset

Abstract

Diabetic Retinopathy DR is one of the commonly existing disease in diabetic patients that may cause vision impairment. DR has five stages of severity, wherein the last stage is one of the major cause of blindness worldwide. In most of the cases, DR does not show any symptoms until it reaches the last stage. If a mean is available to timely detect DR during the early stages, it can be cured properly. An easy and accurate system therefore is required to detect DR at very minute level. Patients seeking help and doctors providing medical support both will be the immediate beneficiary of this detection system. Doctors will greatly augment their diagnostic, supportive or corrective decisions, as the system will support their decision with certainty of DR severity stage. In this paper, we designed and compared multiple automated systems for classification of DR using colored eye fundus images. The main focus was to design a system that detects the presence of DR and classifies the disease (if present) based on severity with best accuracy possible. We proposed three frame works, Frame Work 1: was designed a cascaded classifier using more than one combinations of Convolution Neural Networks (CNN) for feature extraction and comparaed best conventional classifier RFT for classification. It provided improved accuracy of 75% with Sensitivity/Specificity 89%/76.8% and dealt with over-fitting and complexity of knowing minor details.Frame Work 2: we applied multiple pre-processing techniques and reviewed their results we also infused results of multiple CNNs to get better accuracy of 78.6% with Sensitivity/Specificity of 96%/75.14%. Frame Work 3: we added LSTM layer to our CNN for getting better Accuracy of 76.9% with Sensitivity/Specificity 96.08%/73.81%. Using these multiple Frame works we achieved best using ensemble systems. This research concludes that by selecting right kind of pre-processing, data set management and architecture working accuracy can be improved.