Diabetic Retinopathy Lesions Detection using Faster-RCNN from retinal images

Diabetic Retinopathy is an eye disease that damages the retina which can cause vision loss. Early detection of DR is needed because the disease shows little signs in its initial stage due to the slow progression of the disease. The screening process of the eye is a time-consuming, costly, and tedious task due to the examination of every single patient. In this work, we deal with the localization of lesions of DR from retinal images. We have presented a novel method based on the Faster Region-based Convolutional Neural Network (RCNN) to overcome the challenges of DR lesions detection methods and precisely detect the early signs as well. Our method constitutes two steps: first is preprocessing and the other is the localization of abnormalities of DR i.e. hard exudates, soft exudates, microaneurysms, and hemorrhages. For performance evaluation, we have used the publicly available datasets i.e. Diaretdbl and Messidor and achieved average values of accuracy as 0.95 and Intersection over union (IOU) as 0.94. The proposed method achieved remarkable results as compared to state-of-the-art techniques.