A lightweight, novel feature extraction and classification system for the early detection of diabetic retinopathy

One of the main causes of irreversible blindness among working-age adults globally is the eye condition known as Diabetic Retinopathy. Its early detection is imperative for both the treatment and management of the disease. Current methods of detection require specialised equipment and an image processing system that either requires large local resources (memory, processing power) or utilises server-side resources requiring a good internet infrastructure, neither of which may be available in many developing countries. The current research presents three new feature extraction procedures to detect Hard Exudates, Retinal Blood vessels and Microaneurysms respectively, these all being early diagnostic features of Diabetic Retinopathy. Each algorithmic procedure is a processing chain of image processing steps that is fast and robust and but computationally less dependent on resources compared to existing systems. Following feature extraction from eye fundus images a Binary Decision method is used to classify the images into healthy or unhealthy, and validated with known image classes and also by using manual classification conducted by two qualified optometrists. The performance and robustness of the proposed system is evaluated from confusion matrix data, AUC-ROC curves and statistical tests such as the t-test and Chi-square test. The results present a novel decision-based system for the early detection of Diabetic Retinopathy whose light footprint and real-time processing capability has the potential for deployment onto a portable screening system (e.g., smartphone with camera attachment) that can be utilised by people living in remote parts of the world.