Enhancing Diabetic Retinopathy Classification Through Advanced Deep Convolutional Neural Network Architectures

Diabetic retinopathy is one of the major causes of vision loss in the world; thus, the development of early diagnosis systems based on artificial intelligence is a medical and scientific necessity. If detected late, it can lead to permanent vision loss and is considered one of the most serious vascular complications of diabetes. Diagnosis is traditionally made by visual inspection of fundus images by an expert medical professional and is time-consuming. Deep learning techniques have shown promise to help detect disease early and improve the accuracy of medical decision- making. This study provides an experimental comparison of four convolutional neural network architectures (Xception, Inception-v3, MobileNet-v3 and ResNet-50), using a balanced dataset of 25,000 colour fundus images across five classes. Of the total images, 80% were used for training, 10% for validation, and 10% for testing. A standardised image size of 1024×768×3 was used to preserve the fine details of diabetic retinopathy. The models were evaluated for a number of metrics such as accuracy, loss, recall, precision, F1-score, AUC, Cohen’s Kappa coefficient and MCC in addition to computational cost and training time. The results indicate that the Xception model outperformed the other models, with a test accuracy of 98.32%, a validation accuracy of 98.52%, the lowest loss rate, and the highest AUC value. Inception-v3 achieved the second-best performance with 95.04% accuracy, and MobileNet-v3 achieved the lowest computational cost, making it suitable for resource- constrained applications, with 90% accuracy. Conversely, ResNet-50 had lower accuracy with high computational cost, which shows that the increased architectural complexity does not necessarily ensure improved performance. The results show that the best model selection for diabetic retinopathy classification depends on a trade-off between diagnostic accuracy, computational efficiency, and practical usability, with Xception being the most suitable for tasks requiring the highest level of accuracy, and MobileNet-v3 as a viable alternative for resource-constrained environments.

@article{Elkardoudi2026,
  title     = {Enhancing Diabetic Retinopathy Classification Through Advanced Deep Convolutional Neural Network Architectures},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Sifeddine Elkardoudi and Khalil Ladrham and Ahmed Eddaoui and Mohamed Talea},
  doi       = {10.14569/IJACSA.2026.0170732},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170732}
}