Enhancing Eye Health Diagnosis through Deep Transfer Learning: Unveiling Insights from Low Quality Fundus Images

Abstract

Due to the frequency of eye illnesses, effective and precise diagnostic instruments are required. This work suggests an approach that uses low quality fundus images with deep transfer learning more precisely, the EfficientNetB0 architecture to improve eye health diagnosis. We tackle the problem caused by the quality of fundus photographs that are commonly found in clinical settings, which frequently display noise and abnormalities. Our methodology consists of pretraining the EfficientNetB0 model on a sizable dataset of excellent fundus photos, followed by fine-tuning it on a dataset of poor fundus photos. By employing this transfer learning technique, the model enhances its diagnostic capabilities by learning to identify significant features from the low-quality images. We ran tests on a variety of datasets that included fundus photos of varying degrees of deterioration in order to assess our approach. As compared to conventional techniques, the results reveal a significant improvement in diagnostic accuracy, demonstrating the effectiveness of deep transfer learning for improving eye health diagnosis from difficult fundus images. With fused features from MobileNet and DenseNet-121 models, the ANN specifically achieved accuracies of 98.5% for cataracts, 99.1% for diabetic retinopathy, 99% for glaucoma, and 99.5% for normal conditions.