A Decision Support System For Retinal Image Defect Detection

Abstract

Deep learning has become the de facto method for image classification. In this work, a common framework for decision support system is presented that can be reused for diagnosing multiple retinal clinical conditions. Retinal fundus images provide a non-invasive way to diagnose eye-related diseases like glaucoma and diabetic retinopathy (DR). State-of-the-art deep learning methods focus on the detection of key regions of the retina including fundus, optic disc and retinal vessels individually. In order to achieve acceptable precision and recall for a clinically deployable system, a decision support system that combines state-of-the-art deep learning system and relevant explainable features are built. The proposed method is tested on two retinal pathology use cases - glaucoma and for the detection of hard exudates that is critical in diagnosing DR. The proposed model is validated using DRIVE dataset with average Jaccard index of more than 96% for fundus, around 98% for OD and around 90% in identifying retinal vessels using a five-fold cross-validation. For disease detection, the above key regions are combined and validated using standard datasets with good outcomes.