The object of the study is the accuracy of diabetic retinopathy diagnosis based on retinal images. This study investigates convolutional neural network (CNN) models for the automatic detection of diabetic retinopathy (DR) from retinal images. The main problem lies in the insufficient effectiveness of basic CNN models in recognizing DR stages on fundus images. The core problem addressed is the suboptimal performance of baseline CNNs in identifying DR stages from medical imagery. To solve this, two CNN architectures were thoroughly evaluated: a baseline model and an enhanced model integrating advanced preprocessing techniques such as image resizing (256 × 256 and 512 × 512), the image normalization, and data augmentation methods. The enhanced model outperformed the original, achieving a validation accuracy of 91% compared to 88% for the baseline, and demonstrating reduced loss during both training and validation. This improvement is attributed to the optimized input image quality and increased variability in the training set, which enhanced the model’s ability to generalize and avoid overfitting. A distinctive feature of the results lies in the synergy between preprocessing and CNN architecture, which enabled significantly improved classification performance even under hardware constraints. These limitations suggest that further gains are possible with extended computational resources and access to larger datasets. The practical applicability of the findings is evident in the potential deployment of such models in clinical screening systems to support early and accurate DR diagnosis. The models were trained on a proprietary dataset of expert-labeled, high-resolution retinal images, similar in format to EyePACS and APTOS, though not publicly available due to ethical considerations