부산대학교 도서관

Diabetic retinopathy (DR) is a leading cause of blindness and affects 425 million people worldwide. The ocular disease is particularly pervasive in developing countries, where there exists a severe deficiency of ophthalmologists. Here we propose Eye-Net, a machine learning system to offer an efficient and accurate alternative that may automate DR screening. Exploratory data analysis via dimensionality reduction was first leveraged to interpret similarities between color fundus photos. Subsequently, 16 image enhancement filters were systematically investigated for feature engineering purposes. 24 convolutional neural network architectures, including the previously unconsidered EfficientNet and MobileNet families, were designed, trained, and compared by their efficacy in the binary image classification of DR. Ensemble stacking was leveraged to fuse the top four models through a meta-learner whose weights were optimized by a Dirichlet distribution-based randomized search. Finally, this project developed three small-object detectors, serving as the first to localize microaneurysms, hemorrhages, cotton-wool spots, and hard exudates. Upon evaluation, the ensemble model surpassed all individual models, attaining an area under the curve of 0.995. Coupled with contrast-limited adaptive histogram equalization, the YCbCr color space augmentation was found to be the optimal feature engineering filter. Further, Shapley value analysis and four novel gradient-based visualizations successfully identified imaging biomarkers characteristic of DR. The YOLOv5x algorithm achieved a mean average precision of 0.504, a significant improvement over state-of-the-art lesion localization studies. Eye-Net is a highly performant and transparent DR screening tool. Its lightweight frameworks postulate deployment onto mobile applications as promising avenues.