Eye Disease Detection Using Deep Learning Models with Transfer Learning Techniques

Abstract

INTRODUCTION: Diabetic Retinopathy, Cataract and Glaucoma are the major eye diseases posing significant diagnostic challenges due to their asymptotic nature at their early stages. These diseases if not detected and diagnosed at their early stages may lead to severe visual impairment and even can cause blindness in human beings. Early detection of eye diseases showed an exceptional recovery rate. Traditional diagnostic methods primarily relying on expertise in the field of ophthalmology involve a time-consuming process. With technological advancements in the field of imaging techniques, a large volume of medical images have been created which can be utilized for developing more accurate diagnostic tools in the field. Deep learning (DL) models are playing a significant role in analyzing medical images. DL algorithms can automatically learn the features which indicate eye diseases from eye image datasets. Training DL models, however, requires a significant amount of data and computational resources. To overcome this, we use advanced deep learning algorithms combined with transfer-learning techniques. Leveraging the power of deep learning, we aim to develop sophisticated models that can distinguish different eye diseases in medical image data. OBJECTIVES: To improve the accuracy and efficiency of early detection methods, improve diagnostic precision, and intervene in these challenging ocular conditions in a timely manner. METHODS: The well-known Deep Learning architectures VGG19, InceptionV3 and ResNet50 architectures with transfer learning were evaluated and the results are compared. RESULTS: VGG19, InceptionV3 and ResNet50 architectures with transfer learning achieved 90.33%, 89.8% and 99.94% accuracies, respectively. The precision, recall, and F1 scores for VGG19 were recorded as 79.17%, 79.17%, and 78.21%, while InceptionV3 showed 82.56%, 82.38%, and 82.11% and ResNet50 has 96.28%, 96.2%, and 96.24%. CONCLUSION: The Convolutional Neural Network models VGG19, Inception v3, ResNet50 combined with transfer learning achieve better results than the original Convolutional Neural Network models.