A Text-Guided Lesion Mining Vision-Language Ordinal Classification Framework for Diabetic Retinopathy Grading

Diabetic retinopathy (DR) is a major retinal disease that can cause visual impairment and irreversible blindness. Accurate automated DR grading is essential for large-scale screening and timely clinical intervention. However, most existing methods rely primarily on visual features for classification. Moreover, they often overlook the ordinal structure of DR severity and the intra-class phenotypic heterogeneity arising from diverse lesion combinations. To address these issues, based on the semantic prior information provided by RetiZero, we propose a text-guided lesion mining vision-language ordinal classification framework for DR grading. The proposed framework introduces a text-guided cross-layer lesion mining module that exploits semantic response differences between normal-tissue and lesion-related textual prompts, thereby guiding multi-level visual patch features toward lesion regions relevant to DR grading. To explicitly model the ordered progression of DR severity, we design a conditional ordinal regression branch and an ordinal distribution alignment strategy that jointly encourage the predictions to follow the inherent order of DR grades. Moreover, we introduce a multi-center feature constraint to capture diverse intra-grade phenotypic patterns and enhance feature discriminability. Experiments on APTOS 2019 show that the proposed method achieves 86.3% accuracy, 90.6% AUC, and 70.9% Macro-F1, which improved by 2.4, 0.7, and 5.5 percentage points compared to RetiZero. Furthermore, under the standardized leave-one-domain-out protocol of GDRNet, the proposed method achieves the highest reported average accuracy of 58.6% across six public DR datasets, exceeding the reported result of PAF (54.6%) by 4.0 percentage points. Nevertheless, our approach is limited in F1 and AUC metrics, for which GDRNet delivers superior performance. These results suggest that the proposed framework can improve DR grading performance and the cross-dataset generalization ability of the model to a certain extent.