Segmentation of Ocular Pathologies Using Deep Convolutional Neural Network

Diabetes Mellitus (DM) is a chronic, progressive and life-threatening disease. The ocular manifestations of DM, Diabetic Retinopathy (DR) and Diabetic Macular Edema (DME), are the leading causes of blindness in the adult population throughout the world. Early diagnosis of DR and DM through screening tests and successive treatments can reduce the threat to visual acuity. In this context, we propose an encoder decoder based semantic segmentation network SOP-Net (Segmentation of Ocular Pathologies Using Deep Convolutional Neural Network) for simultaneous delineation of retinal pathologies (hard exudates, soft exudates, hemorrhages, microaneurysms). The proposed semantic segmentation framework is capable of providing segmentation results at pixel-level with good localization of object boundaries. SOP-Net has been trained and tested on IDRiD dataset which is publicly available with pixel level annotations of retinal pathologies. The network achieved average accuracies of 98.98%, 90.46%, 96.79%, and 96.70^% for segmentation of hard exudates, soft exudates, hemorrhages, and microaneurysms. The proposed methodology has the capability to be used in developing a diagnostic system for organizing large scale ophthalmic screening programs.