Diabetic retinopathy (DR) is estimated to affect over 93 million people worldwide and is the leading cause of blindness among working-age people in developed countries. To make a DR diagnosis, an experienced medical professional must manually examine and analyze digital color photographs of the retinal fundus. This process might take several hours. This paper aims to construct a Residual Neural Network (RNN)-based Deep Learning model to aid in early DR detection and to compare it to existing models. Recently, ResNet architecture has been shown to be the most cutting-edge strategy for image analysis. ResNet152V2, a model pre-trained on Kaggle’s “Diabetic Retinopathy Detection” dataset, APTOS 2019 was used to identify the varying degrees of DR. Preprocessing images for this study involved applying a Gaussian filter and resizing them. To achieve the best possible result in classifying DR levels, pre-trained ResNet models are fine-tuned using various hyperparameters, which is done using the partially frozen layer. The effectiveness of the pre-trained model has been evaluated using a number of measures, including training and test loss, training and validation accuracy graphs, and the confusion matrix. The suggested model outperformed the existing deep learning approaches, as evidenced by its training accuracy of 94.03%, validation accuracy of 90.17%, AUC score of 95.79%, and Cohen Kappa score of 0.901. The F1 score values for each group were as follows: 98% for stage 1, 79% for stage 2, 83% for stage 3, 64% for stage 4, and 58% for stage 5.