Introduction
Retinal artery/vein (A/V) classification is a fundamental task in medical imaging, providing critical biomarkers for systemic conditions such as hypertension and diabetes. Despite the success of deep learning in medical segmentation, two primary challenges persist: the high cost of pixel-level annotations and the performance degradation caused by domain shift between different imaging devices. This research proposes a semi-supervised, cross-domain framework designed to address these hurdles by leveraging a small amount of labeled data from the target domain alongside a fully labeled source dataset.
Methodology
The core of the proposed framework is a teacher-student model based on the Mean-Teacher architecture. The student model learns from both labeled source/target data and the teacher model’s consistency-based predictions on unlabeled target data. Two specific innovations enhance this process for retinal vasculature:
- Spatial Transformation Module: To maintain consistency across different views, the authors implement a constraint that ensures the model's predictions remain stable even when inputs undergo spatial warping (e.g., rotation or scaling).
- Topology-Aware Regional Mixing: Traditional data augmentation techniques like CutMix often disrupt the anatomical connectivity of blood vessels. The authors introduce a "fan-shaped" mask centered on the optic disc. This strategy respects the natural radial growth of retinal vessels, preserving their topology and creating more realistic synthetic training samples.
Experimental Setup
The framework was evaluated on two datasets:
- DRIVE-AV: A public dataset of 40 images with manual A/V annotations.
- Private Dataset: A large-scale clinical dataset containing 5,336 color fundus images.
To simulate a real-world "low-data" scenario typical of new clinical environments, the researchers used only 10 labeled images from the target domain for training and 88 for testing.
Results and Discussion
The proposed method achieved state-of-the-art results in cross-domain scenarios. In the transition from DRIVE-AV to the Private Dataset, the model reached an accuracy of 88.98% and an F1-score of 87.57%, significantly outperforming standard semi-supervised methods like CPS and UA-MT. The False Discovery Rate (FDR) also dropped to 11.37%, indicating higher precision in identifying vessel types.
The results indicate that incorporating even a small number of target-domain labels (semi-supervision) provides a substantial performance boost over purely unsupervised domain adaptation. Furthermore, ablation studies highlighted that the topology-aware regional mixing was the most critical component for improving sensitivity and specificity, underscoring the importance of biological priors in medical AI.
Conclusion
This work demonstrates that respecting the anatomical topology of retinal vessels during training significantly improves the robustness of A/V classification models across different imaging domains. Future research will focus on validating this approach in cases with severe pathology where the optic disc may be obscured and extending the topology-aware mixing strategy to 3D modalities like Optical Coherence Tomography Angiography (OCTA).
Summary updated