An Automatic Early Screening System of Eye Diseases Using UWF Fundus Images Based on Deep Neural Networks

Visual impairment has become a major health problem worldwide. Most eye diseases can be effectively controlled by early screening and timely treatment, preventing visual impairment. At present, the main screening method is manual screening, which can't meet the large-scale clinical requirements. To address the limitations of manual screening, we developed an automatic early screening system using ultra-wide field(UWF) fundus images to identify multiple eye diseases (including myopia, retinal detachment(RD), diabetic retinopathy(DR), and cataract) in this study. The system uses a convolutional neural network-based architecture consisting of two components: feature extractor and classifier. The feature extractor extracts local features of four sub-images of an original UWF fundus image, which are fused and then fed into the classifier to predict its class. Two feature fusion methods, the max feature aggregate operators and concatenation methods, are then explored. The model was trained and evaluated using 7,209 UWF fundus images. The max feature aggregate operators method exhibited 0.9615% accuracy, 0.9695% precision, 0.9697% sensitivity, 0.9744% specificity, and 0.9695% F1_score for task 1, and 0.9330% accuracy, 0.9737% precision, 0.9749% sensitivity, 0.9812% specificity, and 0.9739% F1_score for task 2. Its performance is reliable and was approved by ophthalmologists.