Implementation of Transfer Learning Techniques for retinal image classification with superior Accuracy

In the diagnosis of the majority of retinal diseases, the continuous surveillance of blood vessels is contingent upon retinal image categorization. Deep learning (DL) methodologies are acknowledged as promising strategies for the classification of biological images. Over the last ten years, the classification of retinal diseases has been an active field of study. Various methods exist for classifying retinal diseases, starting with convolution neural networks (CNN) and other deep models. This research suggests using fine-tuned MobileNetV2 and EfficientNetB0 to classify retinal pictures. The Retinal Fundus Multi-disease picture Dataset (RFMiD) was gathered from a public source, and the picture was enhanced to ensure balance. Data augmentation practices were used in the study to improve the training dataset and lessen overfitting. The framework incorporates several techniques for improving images, including contrast-limited adaptive histogram equalization (CLAHE) and the median filter (MF). Two approaches are examined for CNN model training on balanced datasets. The first utilized EfficientNetB0, while the second employed MobileNetV2, resulting in EfficientNetB0 achieving 86.10% validation accuracy and MobileNetV2 attaining 93.77% validation accuracy. The findings of MobileNetV2 demonstrate superior efficacy compared to EfficientNetB0 in the precise classification of retinal pictures.