RADR: A Robust Domain-Adversarial-based Framework for Automated Diabetic Retinopathy Severity Classification

Sara Mínguez Monedero, Fabian Westhaeusser, Ehsan Yaghoubi, Simone Frintrop, Marina Zimmermann

Keywords: Robustness, domain generalization, adversarial training, diabetic retinopathy

Abstract: Diabetic retinopathy (DR), a potentially vision-threatening condition, necessitates accurate diagnosis and staging, which deep-learning models can facilitate. However, these models often struggle with robustness in clinical practice due to distribution shifts caused by variations in data acquisition protocols and hardware. We propose RADR, a novel deep-learning framework for DR severity classification, aimed at generalization across diverse datasets and clinic cameras. Our work builds upon existing research: we combine several ideas to perform extensive dataset curation, preprocessing, and enrichment with camera information. We then use a domain adversarial training regime, which encourages our model to extract features that are both task-relevant and invariant to domain shifts. We explore our framework in its various levels of complexity in combination with multiple data augmentations policies in an ablative fashion. Experimental results demonstrate the effectiveness of our proposed method, achieving competitive performance to multiple state-of-the-art models on three unseen external datasets.

Submission Number: 79