Diagnosis of Diabetic Retinopathy by using EfficientNet-B7 CNN Architecture in Deep Learning

This research study intends to assess the effects of diabetes on the thickness of each inner, outer, and peripheral retinal layer. For a very long time, it was believed that diabetic retinopathy (DR) was predominantly a retinal microvascular condition brought on by the immediate consequences of hyperglycemia and the metabolic pathways it activates. However, some current research has shown that retinal neurodegeneration, which is caused by an unfavorable ratio of neurotoxic to neuroprotective components, exists even before the onset of clinically observable microvascular damage. One such retinal illness that affects persons with long-term diabetes is diabetic retinopathy. A sign of diabetic retinopathy is the development of abnormal blood vessels within the retina. Damage to the vascular system in the eyes is a result of diabetes. Although there may not be any symptoms initially, visual issues may gradually become apparent. The pro evaluation of the physiological changes taking place within the human body is one of the most critical issues in bioengineering. Due to the numerous complexities involved in the process, detecting abnormalities in the human eye is very challenging. By 2050, 16.0 million Americans will be diagnosed with diabetic retinopathy, according to estimates. Most conventional methods for identifying diseases from retinal pictures rely on manual work. Due to the significant error rate of human observation, these strategies have a less success rate. The suggested model additionally considered the dataset’s substantial class imbalance. Aptos2019-diabetic-retinopathyoversampled-256 $\times$ 256 dataset was used to train the CNN model. In order to automatically detect diabetic retinopathy, this study developed a deep learning technique called the EfficientNet-B7 model. To determine the ideal parameters to employ, hyperparameter tweaking was done. We evaluated the Efficientnet-B7 model employing an augmentation strategy that can reduce the occurrence of overfitting in comparison to models without augmentation. To increase classification accuracy and analyze the efficiency of the proposed work with the algorithms. The convolutional neural networks layout and levelling the method by using EfficientNetB7 use a compound coefficient to scale all depths, breadth, and clarity parameters equally and it obtained 94% accuracy.