Diabetic retinopathy (DR) is a retinal disease caused by diabetes. It is a disorder that develops due to the blood vessel damage that supplies the retina. It advances through the five stages: healthy, mild, moderate, severe, and proliferative, and if at the last stage the patient is left untreated, the vision will be lost. Early diagnosis and treatment are the main keys to preventing vision loss, and computer-aided diagnostic systems can be a big help in the detection process. Conventional screening methods are based on fundus cameras and manual grading, which are both slow and susceptible to variability.

To create and thoroughly evaluate a modified ResNet18 model with Swish activation for five-class DR grading funded by public eye fundus images.

A Kaggle DR dataset of 5000 images was split patient-wise (80/20) before augmentation, with an additional validation through fivefold cross-validation. Models (VGG19, ResNet50, Xception, refined ResNet18, and the proposed architecture) were trained under the same preprocessing (256 × 256 resolution, normalisation), optimizer (Adam), cyclical learning rate schedule, and early stopping. The main endpoints were accuracy, macro-F1, macro-AUC, and macro-AUPRC, with 95% confidence intervals (bootstrap, 1000×). The statistical tests used for the statistical comparisons were McNemar’s and DeLong’s tests, and calibration was carried out using Brier score and reliability curves.

The proposed ResNet18 with Swish got a score of 96.91% (95% CI 95.5–98.0%) for accuracy and 96.92% (95% CI 95.6–98.1%) for macro-F1, which was statistically significantly better than the baseline models. The external validation on APTOS confirmed the generalizability, while the calibration analysis showed that probability estimates were reliable for each class.

The easily portable, well-calibrated deep learning (DL) model is capable of precise and reproducible five-class DR grading and thus facilitates the integration into clinical screening workflows for early detection and intervention.