From Micro to Macro: Multi‐Scale Lesion Perception for Interpretable Weakly‐Supervised Diabetic Retinopathy Grading

This is the authors' abstract. We don't add key points for this paper.

Diabetic retinopathy (DR) grading from fundus images remains a challenging task, particularly under weak supervision where pixel‐level annotations are unavailable. Existing methods often struggle to capture the diverse, multi‐scale lesions and their complex global inter‐dependencies simultaneously. To address this, we propose a novel micro‐to‐macro weakly‐supervised framework. Our approach features a multi‐level lesion‐aware module that integrates features from shallow to deep layers for comprehensive lesion localization. A lesion relation modelling module then captures long‐range dependencies among the identified lesions. Furthermore, we introduce a class‐aware label smoothing loss that incorporates the ordinal relationship between DR grades into the optimization. Extensive experiments on public datasets demonstrate that our framework achieves state‐of‐the‐art performance. More importantly, the generated lesion activation maps show high clinical concordance with expert annotations, validating the model's superior interpretability and its potential for reliable automated DR screening.