Diabetic Retinopathy (DR) detection is essential for managing Diabetes Mellitus, as it can lead to vision impairment and blindness. This study investigates various deep learning, convolutional neural network architectures (CNNA) utilized as automating DR severity classification. Three thousand trained dataset of fundus images are tested on 2900 images across five DR severity classes, the proposed model integrates Res-Net, Efficient-Net, Inception, Xception, and Dense-Net. Utilizing both binary and multiclass classification techniques, the models are evaluated based on training accuracy, validation accuracy, and loss metrics. Findings reveal distinct performance profiles among the models, highlighting their suitability for DR classification tasks. This analysis underscores the importance of selecting appropriate DL architectures tailored to optimize accuracy and mitigate potential overfitting challenges in DR detection.