Diabetic retinopathy (DR) is a critical state that could lead to permanent blindness if not detected and treated early. However, it takes a lot of time and is liable for errors while evaluating retinal images manually for the diagnosis. We propose the DenseNet-121 model using an XGBoost classifier for DR classification to overcome these challenges. The features of the classifier are generated from a pre-trained network called DenseNet-121. We have used the APTOS-2019 Blindness Detection dataset from Kaggle for this work and our proposed model achieved a correctness of 92.20%. Comparative analysis with existing methods confirms that DenseNet-121 outperforms various other methods. These results showcase how deep learning (DL) can be effectively utilized for early diagnosis and classification of DR. On a test set of retinal pictures, our model obtained 95% accuracy, 94% precision, 97% recall, and 94.5% F1-score. The experimental results confirm that DenseNet-121 can support clinicians in the timely diagnosis of DR.