Diabetic retinopathy (DR) is the largest cause of permanent vision loss in the working-age population, making automated grading critical for timely therapeutic intervention. While recent deep learning algorithms have improved feature discrimination, modern state-of-the-art systems have two fundamental drawbacks. First, most models rely on standard Convolutional Neural Networks, which struggle to capture long-range relationships and lack semantic reasoning, resulting in visual findings that do not correlate with clinical knowledge. Second, present approaches often consider grading as a nominal classification or a pure ordinal regression task, failing to strike a compromise between high classification accuracy and severity-consistent predictions (Quadratic Weighted Kappa). To address these challenges, we propose Dual-SwinOrd, a novel framework that integrates a hierarchical Vision Transformer with a semantically guided dual-head mechanism. Specifically, we use a Swin Transformer backbone to extract hierarchical features, effectively capturing global retinal structures. To handle diverse lesion scales, we incorporate a Progressive Lesion-aware Kernel Attention (PLKA) module and a Semantic Prior Modulation (SPM) module guided by PubMedCLIP, bridging the gap between visual features and medical linguistic priors. In addition, we propose a Dual-Head learning strategy that decouples the optimization objective into two parallel streams: a Classification Head to maximize diagnostic accuracy and an Ordinal Regression Head (DPE) to enforce rank-consistency. This design effectively mitigates the trade-off between precision and ordinality. Extensive experiments on the APTOS 2019 and DDR datasets demonstrate that Dual-SwinOrd achieves state-of-the-art performance, yielding an Accuracy of 87.98% and a Quadratic Weighted Kappa (QWK) of 0.9370 on the APTOS 2019 dataset, as well as an Accuracy of 86.54% and a QWK of 0.9040 on the DDR dataset.
Yu et al. (Tue,) studied this question.