Executive Summary : | Ophthalmic diseases, such as glaucoma, diabetic retinopathy, age-related macular degeneration, cataract, and myopia, are prevalent worldwide and can lead to partial or complete vision loss if left untreated. Early diagnosis is crucial for preventing blindness, but it is challenging due to the occurrence of only a few symptoms in the early stages. Fundus imaging, a non-invasive and cost-effective technique, is widely used for clinical inspection. However, manual identification of these diseases is laborious and time-consuming, necessitating the development of automated tools for screening from fundus images at an early stage. Automated tools based on machine learning and deep learning methods, particularly convolutional neural networks (CNNs), have made significant strides in recent years. However, these methods rely on identifying single ophthalmic diseases and cannot detect multiple diseases simultaneously. This project aims to design an intelligent multi-label ophthalmic disease screening tool that can identify multiple ophthalmic diseases. However, multi-label classification is challenging due to high inter-class similarity, minute size variations among different lesion types, lack of large datasets, and class imbalance problems. The project aims to develop a lesion-aware attention-based CNN model to capture fine-grained features even with limited data. Additionally, a large and diverse dataset consisting of all possible ophthalmic diseases will be developed. An effective vision transformer (ViT)-based model will be developed to exploit dependencies among visual features and disease labels. The proposed diagnostic tool will be cost-effective and help detect multiple ophthalmic diseases on a larger scale. |