Improved Automatic Grading of Diabetic Retinopathy Using Deep Learning and Principal Component Analysis

Diabetic retinopathy (DR) is one of the most common chronic diseases around the world. Early screening and diagnosis of DR patients through retinal fundus is always preferred. However, image screening and diagnosis is a highly time-consuming task for clinicians. So, there is a high need for automatic diagnosis. The objective of our study is to develop and validate a new automated deep learning-based approach for diabetic retinopathy multi-class detection and classification. In this study we evaluate the contribution of the DR features in each color channel then we pick the most significant channels and calculate their principal components (PCA) which are then fed to the deep learning model, and the grading decision is decided based on a majority voting scheme applied to the out of the deep learning model. The developed models were trained on a publicly available dataset with around 80K color fundus images and were tested on our local dataset with around 100 images. Our results show a significant improvement in DR multi-class classification with 85% accuracy, 89% sensitivity, and 96% specificity.