Domain generalization for diabetic retinopathy grading with phase augmentation framework

Diabetic retinopathy (DR), a common diabetic complication, stands as a primary cause of retinal blindness. Rapid automatic DR grading is a crucial approach for prevention. Although deep learning has shown promising results in automated DR grading, its deployment in clinical settings remains challenging due to variations in imaging devices, lighting conditions and other conditions in different hospitals. This study explores the problem of decreased generalization performance in DR grading, stemming from variations in the distributions of the source and target domains, namely addressing the domain generalization (DG) problem in DR grading. To tackle this challenge, we propose a novel Fourier-based domain generalization framework for fundus images, which consists of three key innovation elements: (1) Fourier spectrum enhancement: a Fourier-based image enhancement technique preserves critical high-frequency features by leveraging phase information, significantly improving cross-domain robustness; (2) Collaborative teacher-student knowledge distillation: a dual-network learning mechanism enhances generalization by distilling high-level semantic features from a teacher model to a student model; (3) Feature fusion: a feature fusion module effectively differentiates intra-class and inter-class features, thereby further enhancing classification performance. Extensive evaluation of our framework on six clinically realistic DR datasets demonstrates superior generalization performance compared to existing methods. Furthermore, this study reveals the critical role of Fourier phase information and high-level semantic features in improving generalization, bridging an important research gap in DR grading.