Utilizing ocular imaging data and deep learning to diagnose open angle glaucoma and diabetic retinopathy

Preventable eye diseases represent a significant public health and economic challenge, impacting over 2.2 billion individuals. Glaucoma (GC) and Diabetic Retinopathy (DR) are primary causes of vision loss worldwide, but early diagnosis can significantly slow disease progression. Currently, over 80 million people are affected by glaucoma and more than 103 million people have DR. Traditional diagnosis relies on ophthalmologists manually analyzing fundus images which is time-consuming and costly. To address these limitations, this study leverages transfer learning with pre-trained deep learning models, including MobileNet and ResNet, to enhance the efficiency and accuracy of disease diagnosis from fundus images. These architectures and classifiers were trained, validated, and tested using public datasets that combined together, and included 4,377 retinal images for three categories: GC, DR and normal eyes. We compared the two Convolutional Neural Network (CNN) models, ResNet50 and MobileNetV2, and among the tested architectures, ResNet50 demonstrated a testing accuracy of 97.45%. Meanwhile MobileNetV2 had an accuracy of 97.19%. The results show that ResNet50 has superior diagnostics. This study shows that diagnosis of GC and DR can be automated which would make the process streamlined, fast and efficient. It is cheaper than traditional diagnosis and requires less equipment.