Análisis comparativo del desempeño de un modelo CNN personalizado frente a modelos preentrenados para la clasificación multietapa de la retinopatía diabética

Abstract

This work develops and evaluates a deep learning system aimed at multi-stage classification of diabetic retinopathy from fundus images. The study follows a comparative approach, whose purpose is to analyze the performance of a proposed CNN model and contrast it with different pretrained architectures under homogeneous training and evaluation conditions, in order to identify differences attributable to the structure of each model. To conduct this analysis, a balanced dataset across several diagnostic categories was consolidated by integrating public fundus image repositories and organizing it through a stratified split into training, validation, and test sets. The images underwent a standardization and preprocessing procedure to ensure consistent inputs, and the models were trained under a controlled experimental configuration. Final evaluation was performed on an independent test set using multiclass metrics and analysis tools such as the confusion matrix and ROC curves under a One-vs-Rest scheme. The results showed that the proposed CNN model achieved competitive overall performance compared to the pretrained architectures, highlighting a good balance between discriminative capability and training stability. In general, most confusions occurred in intermediate classes, which is consistent with the visual overlap between adjacent disease stages. Likewise, the generalization analysis indicated the presence of moderate overfitting in the evaluated models, suggesting the need for additional regularization strategies and external validation to strengthen the system’s robustness on unseen data. As an operational validation, an interactive web platform was implemented that integrates the selected model and enables inference by uploading fundus images, applying automatic preprocessing, and presenting the multiclass prediction in an accessible manner. Overall, this work provides a reproducible framework for comparing deep learning architectures for multi-stage diabetic retinopathy classification and lays the groundwork for integration into digital decision-support environments.