Diabetic retinopathy diagnosis using deep neural networks

Abstract

Relevance: is driven by the high prevalence of diabetic retinopathy as one of the leading causes of vision loss worldwide and the limited availability of ophthalmological diagnostic resources. In medical applications, predictive uncertainty assessment is critically important, since erroneous yet confident decisions may lead to serious clinical consequences. Despite significant advances in deep learning-based DR diagnostic systems, insufficient attention has been paid to the systematic interaction between data preprocessing strategies, class imbalance handling, and predictive uncertainty estimation, all of which directly affect model reliability. Aim and objectives: the aim of this study is to develop and experimentally validate an automated DR stage classification framework that achieves high diagnostic performance while systematically evaluating predictive reliability and uncertainty. Methods: the proposed framework integrates computer vision and deep learning methods based on convolutional neural networks with transfer learning and incorporates an adaptive attention mechanism that dynamically regulate channel and spatial attention strength according to feature variability. The approach also incorporates two preprocessing pipelines, class-balancing techniques, and uncertainty estimation through stochastic inference without modifying the training procedure. A systematic experimental study was conducted to evaluate different combinations of preprocessing and balancing strategies under fixed architectural conditions. Model performance was evaluated using standard classification metrics and an original integral reliability indicator, which jointly accounts for classification quality, prediction confidence, and predictive uncertainty. Results: combining advanced preprocessing, targeted balancing strategies, adaptive attention mechanisms, and uncertainty estimation significantly improves both classification effectiveness and predictive reliability. Proposed configuration achieved the highest scores, while uncertainty-aware models exhibited lower predictive variance and higher confidence, particularly in challenging or ambiguous cases. Conclusions: the proposed framework provides a structured methodology for reliability-aware DR classification and contributes an integrated evaluation approach that enhances the practical applicability of deep learning systems in clinical decision support.