Enhanced ConvNeXt-V2 Framework with Squeeze-and-Excitation Attention Module for Grading of Diabetic Retinopathy using Fundus Images

Diabetic retinopathy (DR) is a progressive microvascular complication associated with diabetes and is a leading cause of preventable blindness globally. The automated and accurate grading of DR from retinal fundus images is challenging due to high intra-class similarity, variations in illumination, and class imbalance. This study introduces an improved convolutional architecture for classifying DR severity, utilizing the ConvNeXt-V2-Tiny backbone enhanced with Squeeze-and-Excitation (SE) attention modules. A customized preprocessing pipeline using green-channel CLAHE and Rayleigh-gamma is applied to enhance lesion contrast, while LDAM loss and class-balanced sampling address skewed label distributions. The model is trained on the APTOS-2019 dataset, which includes 3,662 fundus images across five severity levels, and is evaluated using clinically relevant metrics. The proposed framework achieves an accuracy of 87.2%, a macro-F1 score of 0.78, and a Quadratic Weighted Kappa of 0.70, outperforming the ResNet-50 and Xception baselines. These findings demonstrate the effectiveness of ConvNeXt-V2-SE in delivering robust, interpretable, and computationally efficient diagnostic performance for large-scale DR screening.