Diabetic Retinopathy (DR) is one of the major causes of blindness. As an early detection mechanism, the type and number of lesions (microaneurysms (MAs), hemorrhages (HEM), and exudates (EX)) that appeared on retinal images were used to classify DR and its severity level via automated diagnosis systems. An automated diagnosis system usually contains two main processing techniques: segmentation and classification. The accuracy of this process is related to the parameters of the input retinal image captured by the camera inside the system. Segmentation plays a crucial role as it is the initial step for detecting types and number of lesions on the retinal image. The accuracy of this system is heavily dependent upon the retinal and image parameters, which are intensity level differences between background (BG)-blood vessels (BV), BV-bright lesions, BV-dark lesions, and noise levels. Even though several researchers have already proposed many automated diagnosis systems with different image segmentation algorithms, their accuracy and consistency are rarely explored. In this study, we analysed the accuracy of an automated diagnosis system to detect DR and its severity levels. The focus is on image segmentation based on fuzzy entropy multi-level thresholding, as it is one of the most popular algorithms and has been shown to work very well in many applications. The analysis aimed to develop conditions to ensure accurate DR detection and its severity level based on obtainable images and retinal parameters. Firstly, a retinal image model was developed representing retinal under variation of all retinal and image parameters. Overall 45000 images were developed using the retinal model. Secondly, feasibility and consistency analysis were performed based on a specific design Monte Carlo statistical method to quantify the successful detection of DR and its severity levels over all retinal and image parameters. The results of the analyses provide optimum conditions to ensure the detection of DR and its severity levels via segmentation of MAs, HEM and EX on the retinal images. The conditions for accurate DR detections are: BG to BV > 30% and BV to the dark lesions (MAs) > 15% for mild DR, BG to BV > 40% and BV to the dark lesions (MAs and HEM) > 20% for moderate DR, and BG to BV > 30% and BV to the dark lesions (MAs and HEM) > 15%, and BV to the bright lesions (EX) > 55% for severe DR. Finally, the validity of these conditions were verified by comparing their accuracy against real retinal images from publicly available datasets. The verification results demonstrated that the condition for the analysis could be used to predict the success of DR detection using fuzzy entropy multi-level thresholding from obtainable image and retinal parameters. The conditions are also very important as a guide for system developers and computer vision practitioners in developing DR detection systems.
