[Paper Review] Bridge the Domain Gap Between Ultra-wide-field and Traditional Fundus Images via Adversarial Domain Adaptation

Lie Ju, Xin Wang+6 more|arXiv (Cornell University)|Mar 23, 2020

Retinal Imaging and Analysis26 references4 citations

TL;DR

This paper proposes a novel adversarial domain adaptation framework to bridge the domain gap between traditional fundus images and ultra-wide-field (UWF) fundus images, enabling knowledge transfer from abundant traditional fundus data to scarce UWF data. By combining cycle-consistent generative adversarial networks, pseudo-labeling, consistency regularization, and MixUp augmentation, the method significantly improves UWF retinal disease classification performance, achieving a 66.71% F1-score—8.28% higher than training solely on UWF data.

ABSTRACT

For decades, advances in retinal imaging technology have enabled effective diagnosis and management of retinal disease using fundus cameras. Recently, ultra-wide-field (UWF) fundus imaging by Optos camera is gradually put into use because of its broader insights on fundus for some lesions that are not typically seen in traditional fundus images. Research on traditional fundus images is an active topic but studies on UWF fundus images are few. One of the most important reasons is that UWF fundus images are hard to obtain. In this paper, for the first time, we explore domain adaptation from the traditional fundus to UWF fundus images. We propose a flexible framework to bridge the domain gap between two domains and co-train a UWF fundus diagnosis model by pseudo-labelling and adversarial learning. We design a regularisation technique to regulate the domain adaptation. Also, we apply MixUp to overcome the over-fitting issue from incorrect generated pseudo-labels. Our experimental results on either single or both domains demonstrate that the proposed method can well adapt and transfer the knowledge from traditional fundus images to UWF fundus images and improve the performance of retinal disease recognition.

Motivation & Objective

• Address the challenge of limited availability of UWF fundus images for training deep learning models in retinal disease diagnosis.
• Bridge the domain gap between traditional fundus images (plentiful, well-annotated) and UWF fundus images (scarce, noisy, clinically realistic) to enable effective knowledge transfer.
• Develop a co-training framework that leverages both source (traditional) and target (UWF) domains using adversarial learning and pseudo-labeling to improve generalization.
• Overcome overfitting and confirmation bias from incorrect pseudo-labels using MixUp augmentation and consistency regularization.
• Demonstrate that domain adaptation from traditional to UWF fundus images can significantly enhance classification performance on real-world clinical UWF data.

Proposed method

• Employ cycle-consistent generative adversarial networks (CycleGAN) to translate traditional fundus images into UWF-like images, reducing domain shift.
• Apply pseudo-labeling to generate synthetic labels for the translated UWF-like images using a pre-trained source-domain classifier.
• Introduce a consistency regularization term to enforce that the model produces similar predictions for both real and translated UWF images, improving robustness.
• Use MixUp augmentation during co-training to reduce overfitting and confirmation bias from potentially incorrect pseudo-labels.
• Co-train the classification model on both real UWF images and pseudo-labeled translated images, with adversarial training to align feature distributions across domains.
• Optimize the framework end-to-end using a combined loss function comprising classification loss, adversarial loss, consistency regularization, and MixUp loss.

Experimental results

Research questions

• RQ1 Can adversarial domain adaptation effectively bridge the domain gap between traditional fundus images and UWF fundus images?
• RQ2 How does pseudo-labeling combined with consistency regularization improve model generalization when training on limited UWF data?
• RQ3 To what extent does MixUp augmentation mitigate overfitting caused by noisy or incorrect pseudo-labels in domain adaptation?
• RQ4 Does knowledge transfer from traditional fundus images improve classification performance on real-world UWF fundus images compared to training from UWF data alone?
• RQ5 How does the number of generated pseudo-labeled images affect model performance, and is there an optimal quantity that maximizes performance?

Key findings

• The proposed method achieves a 66.71% F1-score on UWF fundus image classification, representing an 8.28% improvement over training solely on real UWF data (55.47% F1-score).
• The method outperforms fine-tuning, knowledge distillation, pseudo-labeling, and MixUp baselines across all metrics, with the highest accuracy (81.37%) and F1-score (66.71%) when using 500 pseudo-labeled images.
• Consistency regularization significantly improves model robustness by aligning predictions between real and translated UWF images, reducing domain shift effects.
• Increasing the number of pseudo-labeled images beyond an optimal point (e.g., 500) leads to performance degradation, indicating that image quality matters more than quantity.
• MixUp augmentation effectively reduces overfitting and confirmation bias from incorrect pseudo-labels, contributing to improved generalization and higher F1-score (61.48%) compared to pseudo-labeling alone (60.92%).
• The ablation study confirms that high-quality pseudo-labeled images are critical—excessive low-quality images degrade model performance, highlighting the importance of regularization and data quality control.