Diabetic Retinopathy (DR) is a prevalent eye complication among individuals with diabetes. Early detection of DR can prevent vision loss and lower associated problems. Manual diagnosis via retinal fundus images requires the expertise of medical professionals to assess several complex characteristics, including visual acuity tests, pupil dilation, and optical consistency tomography. The process is challenging, time-consuming, and prone to errors. This paper proposes a deep learning (DL) technique, DRDnet22, to detect and categorize DR into five distinct stages. The APTOS 2019 dataset is utilized in this study, which contains 3662 images. The study employs dense block and transition layer with global average pooling (GAP) techniques, and the combination of distinct Convolutional Networks (ConvNets) results in an enhanced representation beyond the utilization of features from a single ConvNet alone. To conduct an ablation analysis, the conventional pre-trained model’s performance was compared. The performance evaluation points out that the proposed model achieved 81.6% accuracy, which is an advantage over other DCNN models. Despite facing a few inaccuracies, the model’s competence in spotting vital features underscores its diagnostic potential.