Improving Diabetic Retinopathy Detection Using Fine-Tuned ResNet50 and Data Augmentation Techniques

Abstract

This study introduces a novel deep convolutional neural network (CNN) framework inspired by the fine-tuned ResNet50 model. DR is categorized into four unique stages: mild DR, moderate DR, proliferative DR, and severe DR. This innovative model was crafted to promote feature reuse and guarantee an adequate gradient flow by integrating dense connections among layers. These dense connections are strategically designed to make the model harness data from all earlier layers, thus fostering more in-depth feature representation and spurring learning efficiency. It was performed and evaluated on a sizeable dataset filled with specifically numbered retinal images. Each image was meticulously labeled to indicate the severity of diabetic retinopathy. The findings revealed that this unique model significantly trumped previous methods, yielding an impressive classification accuracy of 99%. Additionally, the model exhibited meaningful progress in recall and F1-score across all classes, reinforcing its reliability in discerning different stages of diabetic retinopathy. The insights from this study suggest that this custom ResNet50 model frame has excellent potential as a vital tool for the precise and early detection of diabetic retinopathy. This could be instrumental in safeguarding individuals with diabetes against vision loss, thus providing a reliable solution for the future.