A Lightweight Deep Neural Network for Multi-label Eye Diseases Classification

Cataracts, Diabetic Retinopathy, and Glaucoma are the leading causes of vision loss worldwide. Traditionally, these diseases are diagnosed by experienced ophthalmologists who manually inspect the retinal images. Although effective, this approach faces some challenges. In fact, it can be time-consuming and, in some countries with low doctor-patient ratios, patients cannot access the necessary care. To solve this problem, researchers around the world are working on computer-aided diagnosis systems to assist healthcare professionals. In this study, the NASNetMobile architecture is leveraged to classify four eye conditions: Cataract, Diabetic Retinopathy, Glaucoma, and Normal. The proposed model was trained on the EDC dataset, containing 4 classes and 4217 images. Performance metrics used include accuracy, recall, f1-score and area under the curve (AUC). The experimental results were compared against published existing work and showed superior results.