Publication

Abstract : Diabetic Retinopathy (DR), Cataract, and Glaucoma, collectively called multiple eye diseases (MED), actually affect 2.2 billion people globally, as stated by WHO in 2023. Therefore, a robust system driven by AI to manage multiple vision-threatening diseases is needed. However, there is a challenge with limited annotated data for classifying MED. As a solution, this study has chosen transfer learning architectures such as EfficientNet-B0, ResNet-50, and VGG-16 to understand the comprehensive behavior of each model with optimization techniques. This study has used the ODIR dataset to implement these models, achieving an accuracy of 98% for EfficientNet-B0, 97% for ResNet-50, and 97% for VGG-16. The model first accomplishes the task of feature extraction by employing frozen ImageNet backbones combined with custom trainable classifiers, incorporating dense layers, batch normalization, and dropout. The model addresses overfitting by using early stopping along with hyperparameter tuning to ensure optimization. In addition, the key novelty of the model lies in combining optimized pre-trained models with a customized adaptive data pipeline that integrates prefetch, cache, and shuffle operations, enhancing training efficiency by reducing I/O latency and lowering training time from 1–1.3 minutes per epoch to just 7 seconds. Standard metrics such as F1-score, recall, precision, and accuracy are used for evaluating these models. Thus, this research helps in identifying the optimal model along with optimization parameters that ensure both accuracy and computational efficiency for detecting multiple eye diseases