Abstract
Various deep learning models have been developed to segment anatomical structures from medical images, but they typically have poor performance when tested on another target domain with different data distribution. Recently, unsupervised domain adaptation methods have been proposed to alleviate this so-called domain shift issue, but most of them are designed for scenarios with relatively small domain shifts and are likely to fail when encountering a large domain gap. In this paper, we propose DCDA, a novel cross-modality unsupervised domain adaptation framework for tasks with large domain shifts, e.g., segmenting retinal vessels from OCTA and OCT images. DCDA mainly consists of a disentangling representation style transfer (DRST) module and a collaborative consistency learning (CCL) module. Through extensive experiments on retinal vessel segmentation, our framework achieves Dice scores close to target-trained oracle both from OCTA to OCT and from OCT to OCTA, significantly outperforming other state-of-the-art methods.
Introduction
In ophthalmology, Optical Coherence Tomography (OCT) and OCT Angiography (OCTA) are essential imaging modalities. While they observe the same anatomical structures, their visual characteristics differ significantly. OCTA provides high-contrast views of blood vessels, making it ideal for manual annotation. Conversely, standard OCT is more common in clinical settings but is characterized by high noise levels, making vessel segmentation difficult for both human experts and automated systems.
The primary challenge in applying models across these modalities is domain shift. A model trained on clear OCTA images often fails when applied to noisy OCT images due to the vast difference in data distribution. While Unsupervised Domain Adaptation (UDA) aims to bridge these gaps, traditional methods often struggle when the modality gap is as large as that between OCT and OCTA.
The DCDA Framework
The proposed DCDA (Disentangling representation style transfer and Collaborative consistency learning) framework addresses large-gap domain adaptation through two synergistic modules:
1. Disentangling Representation Style Transfer (DRST)
This module operates on the premise that an image can be decomposed into two distinct components: content (the underlying vessel structure) and style (the modality-specific rendering). By utilizing encoders and adversarial training, DRST learns to swap styles between modalities. For instance, it can re-style an OCTA image to appear like an OCT image while preserving the exact vessel geometry, effectively generating synthetic labeled data for the target domain.
2. Collaborative Consistency Learning (CCL)
Rather than relying on a single universal model, DCDA employs two specialized experts: a Source-style expert ($F^S$) and a Target-style expert ($F^T$).
- The Source expert is trained on original labeled source data.
- The Target expert learns from the synthetic images generated by the DRST module.
- Both models engage in collaborative consistency learning on unlabeled data, cross-validating their predictions to ensure agreement on vessel locations regardless of the image's visual style.
Experimental Results
The framework was evaluated using the OCTA-500 dataset, focusing on en-face projections. The experiments compared DCDA against several state-of-the-art (SOTA) methods, including CyCADA and ADVENT.
Quantitative Performance
DCDA demonstrated a significant performance leap over existing methods. In the "OCTA to OCT" adaptation task, DCDA achieved a Dice score of 79.28%, vastly outperforming CyCADA's 18.38%. Remarkably, DCDA's performance closely approached the "Oracle" baseline (81.06%), which represents a model trained directly on labeled target data.
Qualitative Analysis
Visual results highlight DCDA's ability to maintain structural integrity. While competing methods often produce fragmented or noisy vessel maps, DCDA generates clean, continuous vessel structures that align closely with the ground truth.
Conclusion
The DCDA framework effectively bridges the significant domain gap between OCT and OCTA imaging. By disentangling content from style and leveraging collaborative expert models, it achieves near-supervised performance without requiring target-domain labels. Future work may involve extending this 2D en-face approach to full 3D volumes to capture more complex diagnostic details, as well as validating the model's robustness against severe retinal pathologies.
Summary updated