EfficientNet-based Diabetic Retinopathy Classification Using Data Augmentation
K.T Harithalakshmi, Rajeev Rajan, K.M Nadheera
Abstract
Diabetic Retinopathy(DR) is an eye disease that leads to visual defects in people who have diabetes. Diabetes also causes the chance of arising other eye problems, including cataracts and glaucoma. Hence detection of symptoms associated with DR at early stages is essential. The DR grading process is complex due to tiny lesions, data inconsistency, and variation within the class. The solution for fine-grained DR grading is finding different features that affect the visual difference, such as Changes in blood vessel diameter, microaneurysms, soft exudates, hard exudates, and hemorrhages. However, small lesions are difficult to distinguish using convolutional neural networks (CNNs). The uneven distribution of DR data causes the model to focus on DR levels with many data, affecting the grading performance. The main aim of this paper is to explore the impact of various deep neural networks to improve scoring performance and address all the above challenges. Finally for improving the model performance, we introduce an image augmentation technique using the DDR dataset and compare the model’s reaction. The results show that data augmentation can significantly improve classification performance compared to baselines.