Divergent Domains, Convergent Grading: Enhancing Generalization in Diabetic Retinopathy Grading

TL;DR

A novel deep learning method for achieving domain generalization (DG) in DR grading and proposes a new way of generating image-to-image diagnostically relevant fundus augmentations conditioned on the grade of the original fundus image to increase the model's robustness.

Abstract

Diabetic Retinopathy (DR) constitutes 5% of global blindness cases. While numerous deep learning approaches have sought to enhance traditional DR grading methods, they often falter when confronted with new out-of-distribution data thereby impeding their widespread application. In this study, we introduce a novel deep learning method for achieving domain generalization (DG) in DR grading and make the following contributions. First, we propose a new way of generating image-to-image diagnostically relevant fundus augmentations conditioned on the grade of the original fundus image. These augmentations are tailored to emulate the types of shifts in DR datasets thus increase the model's robustness. Second, we address the limitations of the standard classification loss in DG for DR fundus datasets by proposing a new DG-specific loss, domain alignment loss; which ensures that the feature vectors from all domains corresponding to the same class converge onto the same manifold for better domain generalization. Third, we tackle the coupled problem of data imbalance across DR domains and classes by proposing to employ Focal loss which seamlessly integrates with our new alignment loss. Fourth, due to inevitable observer variability in DR diagnosis that induces label noise, we propose leveraging self-supervised pretraining. This approach ensures that our DG model remains robust against early susceptibility to label noise, even when only a limited dataset of non-DR fundus images is available for pretraining. Our method demonstrates significant improvements over the strong Empirical Risk Minimization baseline and other recently proposed state-of-the-art DG methods for DR grading. Code is available at https://github.com/sharonchokuwa/dg-adr.

Figures (9)

• Figure 1: Illustration of DR Progression: From an initial healthy state to advanced stages of DR youtubeDeepLearning. NPDR stands for non-proliferative Diabetic Retinopathy.
• Figure 2: Samples of DR-Aug augmented fundus images, when the prompts used correspond to the original image's grade. First row corresponds to mild non-proliferative DR, the second row shows proliferative DR and third row no DR. The generated augmentations are consistent with the symptoms for the given text prompt and also exhibit some variations of the images present within the dataset (for that particular prompt), even when these variations are not within the original image itself. Temporal continuity for a particular DR grade is induced for a given original image e.g. second row.
• Figure 3: Representation of domain shift using t-SNE embeddings. Different colors correspond to distinct domains. Each plot represents a specific grade, ranging from 0 to 4. The final plot (right bottom) encapsulates the collective features of all grades within each respective domain.
• Figure 4: DG-ADR effectively scatters the feature space weakening the domain-specific information in comparison to ERM and GDRNet. The first two plots in each row depict t-SNE visualizations of two-dimensional features for each domain within a given grade. The final plot encapsulates the collective features of each domain regardless of class label (Grade 0 to 4).
• Figure 5: Samples of poor quality images in the DeepDR dataset.