Retinal Disease Early Detection using Deep Learning on Ultra-wide-field Fundus Images

Ultra-wide-field Fundus Imaging captures the main components of a patients eyes such as optic dics, fovea and macula, providing doctors with a profound and precise observation, allowing diagnosis of diseases with appropriate treatment. In this study, we exploit and compare deep learning models to detect eye disease using Ultra-wide-field Fundus Images. To fulfil this, a fully-automated system is brought about which pre-process and amplify 4697 images using cutting-edge computer vision techniques with deep neural networks. These neural networks are state-of-the-art methods in modern artificial intelligence system combined with transfer learning to learn the best representation of medical images. Overall, our system is composed of 3 main steps: data augmentation, data pre-processing and classification. Our system demonstrates that ResNet152 achieved the best results amongst the models, with the area under the curve (AUC) score of 96.47% (95% confidence interval (CI), 0.931-0.974). Furthermore, we visualise the prediction of the model with the corresponding confidence score and provide the heatmaps which show the focal point focused by the models, where the lesion exists in the eye because of damage. In order to help the ophthalmologists in their assessment, our system is an essential tool to speed up the process as it can automate diagnosing procedures and giving detailed predictions without human interference. Through this work, we show that Ultra-wide-field Images are feasible and applicable to be used with deep learning.