Enhanced ConvNeXt Classifier for Diabetic Retinopathy Severity Classification

Abstract

Diabetic retinopathy is a disease where appropriate treatment according to its stage is key to preventing blindness. Yet, manual grading of fundus photographs is time-consuming and subject to inter-observer variability. Consequently, automated severity classification using artificial intelligence has already entered practical use in clinical settings. Deep learning–based image classification has evolved through architectures such as ResNet, Vision Transformer (ViT), and ConvNeXt, achieving increasing levels of performance. However, further improvement in diagnostic accuracy remains an important challenge. To address this, we propose D-ConvNeXt, a novel architecture that integrates the biological information processing mechanisms of dendritic neuron models into ConvNeXt. This design aims to mimic neuronal dendritic computation and enhance feature extraction. We quantitatively evaluated D-ConvNeXt against conventional deep learning models on the international 5-stage DR severity classification task. The results demonstrate that D-ConvNeXt achieved superior performance across all evaluation metrics except Recall, including Accuracy, Precision, AUC, and F1 score. Notably, Precision improved to 71.8％ , compared to 69.6％ for ConvNeXt. In severity-specific analysis, the most significant improvement was observed in proliferative DR（ PDR），where both Precision and Recall increased. These findings highlight the clinical potential of D-ConvNeXt for more reliable and accurate DR screening.