Hard Exudate Extraction from Fundus Images using Watershed Transform

Abstract

Diabetic Retinopathy is a medical condition which affects the eyes due to increased blood sugar levels. This is characterized by presence of exudates - deposits of lipids in the posterior pole of the retina. If this ailment is not treated in earlier stages these deposits can cause blurred vision or even permanent blindness. This paper concentrates on extraction of hard exudates and optic disc from the retinal images of eyes using Marker based Watershed approach, which uses the minima imposition method to create mask and marker. The varying contrast across all the images has been taken care by a non-linear equation. Once these bright objects have been extracted from fundus images, area estimation is performed to eliminate the optic disk, thus retaining only exudates. These images have been procured from publicly available databases. Though software systems are easy to install, they prove to be expensive in terms of time and cost; thus this method has also been implemented on FPGA for an on-chip solution. The precision and sensitivity for exudate extraction sans optic disk are found to be 92.4% and 83.78% respectively.  Though other techniques exist which provide better accuracy, the method described in this paper is found to be hardware friendly in comparison with other proven methods. Few steps of the algorithm developed are implemented on FPGA to provide an embedded system approach to this work, considering the advantages of a hardware-software combination.