Efficient Transfer Learning Model for Automated Diabetic Retinopathy Grading in Resource-Constrained Environment

The first among the outcomes of diabetes mellitus was Diabetic Retinopathy (DR), can cause severe damage for the retinal arterial blood vessels and vision. Early detection of DR stage may reduce the damage to the retina and vision loss. There are different stages of DR. Many existing methodologies has been presented to predict the stage of DR which uses conventional CNN and succeeded in achieving better performance. But the drawback of these models is training such huge network from scratch takes lot of time and resources. In this paper we proposed a novel hybrid approach to solve the DR problem by minimizing the resources utilization. This approach is a 2-step process. During the first step, the essential features from the retinal fundus images are extracted by using transfer learning technique called EfficientNet. And these feature representations are given as input to second step. In second step an ensemble machine learning boosting algorithm is used to predict the DR stage. Transfer learning techniques enable to use the existing pretrained weights which make the training process fast and minimize the resource consumption. And also, the hybrid approach which uses deep learning and machine learning together to deliver better accuracy. This study has conducted experiments on 2 datasets APTOS and IDRiD datasets by applying 3 different boosting algorithms like AdaBoost, XGBoost and LightGBM. The combination of EffiecientNet with Xtreme Gradient Boost (XGBoost) has given 99.1% & 99.2% accuracy on given 2 datasets.