AdvBlur: Adversarial Blur for Robust Diabetic Retinopathy Classification and Cross-Domain Generalization

TL;DR

This work tackles the challenge of robust diabetic retinopathy classification under distribution shifts by introducing AdvBlur, a training framework that injects heavily blurred images as adversarial examples and relies on a dual-loss (custom loss) to disentangle domain-specific from task-relevant features. By labeling blurred images as a separate class during loss computation but not in final predictions, the model learns to ignore non-informative high-frequency patterns while preserving five clinically meaningful DR severity classes. Extensive experiments across camera types and external datasets, along with ablations on loss design and blur method, demonstrate competitive or superior generalization to unseen data without requiring camera metadata. The approach holds promise for more robust, scalable DR screening and could be extended to other medical-imaging domains where domain shift is a critical bottleneck.

Abstract

Diabetic retinopathy (DR) is a leading cause of vision loss worldwide, yet early and accurate detection can significantly improve treatment outcomes. While numerous Deep learning (DL) models have been developed to predict DR from fundus images, many face challenges in maintaining robustness due to distributional variations caused by differences in acquisition devices, demographic disparities, and imaging conditions. This paper addresses this critical limitation by proposing a novel DR classification approach, a method called AdvBlur. Our method integrates adversarial blurred images into the dataset and employs a dual-loss function framework to address domain generalization. This approach effectively mitigates the impact of unseen distributional variations, as evidenced by comprehensive evaluations across multiple datasets. Additionally, we conduct extensive experiments to explore the effects of factors such as camera type, low-quality images, and dataset size. Furthermore, we perform ablation studies on blurred images and the loss function to ensure the validity of our choices. The experimental results demonstrate the effectiveness of our proposed method, achieving competitive performance compared to state-of-the-art domain generalization DR models on unseen external datasets.

Figures (6)

• Figure 1: AdvBlur-Proposed methodology. As shown in the diagram, a new dataset is prepared by adding heavily blurred versions of the original images. These blurred images are labeled as Class 6. During training, classification is performed, and the loss function is applied based on the image label. If the image is blurred, the loss function $L_{\text{BI}}$ (blurred image loss) is used. If the image is original, the loss function $L_{\text{OI}}$ (original image loss) is applied.
• Figure 2: Examples of original (left) and blurred (right) fundus images. The blurring is applied using a median blur with a kernel size of 151.
• Figure 3: Examples of fundus images processed with different blur techniques. The median blur method (top center) effectively removes all blood vessels and other retinal features, leaving only the background.
• Figure 4: Heat map and the respective masked images
• Figure 5: t-SNE plot after training- Differentiate class 0 and other classes.