Introduction
Retinal artery/vein (A/V) classification is a critical component in the automated screening of systemic diseases, including hypertension and diabetic retinopathy. While deep learning models have achieved high accuracy on specific datasets, their performance often degrades significantly when applied to images from different clinical settings. This phenomenon, known as domain shift, arises from variations in camera hardware, illumination, and image resolution across institutions.
Standard Unsupervised Domain Adaptation (UDA) techniques often struggle with pixel-wise tasks like vessel segmentation. Methods that align high-level features can inadvertently destroy fine structural details, leading to "negative transfer." To address this, researchers have introduced Vessel-Mixing Consistency Regularization (VM-CR), a novel approach that leverages unlabeled target data to build robust, domain-invariant models.
Methodology: Vessel-Mixing Consistency Regularization
The proposed VM-CR framework utilizes a "Mean Teacher" architecture, consisting of a Student model and a Teacher model. The Teacher's weights are updated as an exponential moving average (EMA) of the Student's weights, providing stable targets for regularization.
The core innovation is the Vessel-Mixing perturbation. Unlike global mixing strategies, this method performs regional blending:
- Mask Generation: A smooth binary mask ($M$) is generated using a blurred Gaussian random matrix.
- Regional Mixing: Two unlabeled images from the target domain are blended according to the mask $M$. This creates composite images with diverse vascular configurations, forcing the model to handle realistic structural variations.
- Consistency Learning: The Student model processes the mixed image, while the Teacher processes the original images. The system minimizes the discrepancy (Mean Squared Error) between the Student's prediction and the mixed predictions of the Teacher.
This approach ensures that the model's posterior distributions remain consistent under structural perturbations, effectively bridging the gap between the source and target domains without requiring additional manual labels.
Datasets and Experimental Setup
The efficacy of VM-CR was evaluated across four diverse retinal datasets, representing significant domain shifts in resolution and field-of-view.
- AV-DRIVE: Contains 40 images at $584 \times 565$ resolution.
- HRF (High Resolution Fundus): Consists of 45 images at $1200 \times 800$ resolution.
- LES-AV: Includes 22 images at $1620 \times 1444$ resolution.
- INSPIRE-AVR: A challenging dataset of 40 images at $2048 \times 2392$ resolution, providing only centerline labels.
Results and Discussion
VM-CR demonstrated superior performance compared to existing state-of-the-art UDA methods like AdvNet and PnP-AdaNet.
Performance Gains: In the cross-domain task from HRF to DRIVE, the baseline "Source-only" model achieved an F1 score of 79.5%. VM-CR improved this to 87.6%, nearing the supervised upper bound of 89.1%. Similarly, on the INSPIRE-AVR dataset, VM-CR reached an F1 score of 92.5%, a substantial increase from the 80.5% baseline.
Qualitative Improvements: Visual analysis indicates that VM-CR is particularly effective at classifying vessels in complex regions, such as near crossings and terminal branches. By avoiding explicit feature alignment, the method preserves the delicate graphical structure of the vascular tree, which is often lost in other adaptation frameworks.
Conclusion
Vessel-Mixing Consistency Regularization provides a powerful, label-efficient solution for cross-domain retinal A/V classification. By enforcing consistency across regionally perturbed images, the method achieves high robustness to domain shift. Future work will explore the integration of transformer-based backbones and the optimization of computational efficiency to further enhance the clinical utility of this framework.
Summary updated