Publication

Abstract : Diabetic Retinopathy (DR) is a progressive chronic vision-threatening disease of retinal microvasculature associated with prolonged hyperglycaemia, hypertension and other conditions associated with diabetes. Indicators of DR include different kinds of lesions appearing on the retinal surface that are visible in a Digital Fundus Photograph (DFP). Localization of lesions and visual perception is essential to aid physicians in understanding the severity of the condition and to plan an appropriate treatment procedure for the patient. Segmentation inaccuracies due to factors like subtle nature of abnormalities and interference of blood vessels reflect in reduced classification accuracy in case of Feature Based Machine Learning (FML). While Pixel Based Machine Learning (PML) can overcome these issues, they require high computational capabilities and are redundant. Non-segmentation approaches like deep learning have been employed as an alternative for DR diagnosis. However, these techniques directly grade the image through classification and do not allow for visual perception. Thus we have used an intermediate approach called Super-pixel based segmentation that can overcome the problems in FML and PML while retaining the advantages of both. They are consistent with human visual perception and also overcome the data insufficiency problem. In this paper, we have compared the results of multilesion detection associated with DR using super-pixels segmented from three different algorithms namely, Compacted Watershed (CWS), Simple Linear Iterative Clustering (SLIC) & Linear Spectral Clustering (LSC) under a single unified framework. Experimental results show that LSC over-performs both SLIC and CWS quantitatively and qualitatively.