Dual-stream disentangled model for microvascular extraction in five datasets from multiple OCTA instruments

Xiaoyang Hu, Jinkui Hao, Quanyong Yi, Yitian Zhao, Jiong ZhangFrontiers in Medicine 2025 · DOI

TL;DR

Researchers have developed D2Net, a dual-stream deep learning model that significantly improves the accuracy of mapping tiny retinal blood vessels in Optical Coherence Tomography Angiography (OCTA) scans. By "disentangling" actual vascular structures from instrument-specific noise and introducing a novel structural prior, the model achieves state-of-the-art performance across five different datasets and multiple imaging devices.

What Problem Does This Paper Address?

Optical Coherence Tomography Angiography (OCTA) is a non-invasive marvel that lets clinicians see retinal capillaries at micrometer resolution. However, it has a "noise" problem. Different imaging instruments produce varying types of artifacts—like white noise or linear streaks caused by involuntary eye movements.

For deep learning models, these artifacts often look like blood vessels, leading to "overshooting" (segmenting noise as vessels) or missing tiny capillaries that are only 1–2 pixels wide. Most existing models are trained on specific datasets and fail when moved to a different machine. The authors sought to create a "universal" extractor that could handle the messy reality of clinical data from various hardware manufacturers.

Methods

The core innovation is the Dual-stream Disentangled Network (D2Net). Instead of a single path for processing images, D2Net uses two parallel encoders:

  1. Vascular Stream: Focuses on learning the latent features of the blood vessels.
  2. Noise/Artifact Stream: Specifically learns to identify and isolate the "style" of the noise and artifacts.

By using a shared decoder and a Kullback-Leibler (KL) regularization term, the network learns to separate (disentangle) the clean vascular structure from the background static.

Additionally, the authors introduced the Distance Correlation Energy (DCE) module. This module treats the image like a graph where pixels are nodes. It calculates "neighborhood energy"—essentially a prior that tells the network: "If these two pixels have similar intensities and are close together, they are likely part of the same continuous vessel." This helps the model maintain connectivity in tiny capillaries that might otherwise look like disconnected dots.

Datasets

The study evaluated D2Net on five distinct datasets to prove its generalization capabilities:

  • FOCA (New): A meticulously annotated dataset of 88 OCTA images specifically focused on microvascular regions, collected from Angiovue and Zeiss Cirrus instruments.
  • OCTA-500 & ROSE-O: Widely used public benchmarks.
  • ROSE-Z & ROSE-H: Private datasets used to test how well the model handles data from instruments it hasn't seen during training (cross-domain generalization).

Key Insights and Ideas

  • Disentanglement is the Key to Generalization: Most models fail across different instruments because they "bake" the noise of the training machine into their understanding of what a vessel looks like. By explicitly modeling noise as a separate stream, D2Net can ignore instrument-specific "style" and focus on the universal "content" of human anatomy.
  • Low-Dimensional Priors Matter: While deep learning is great at high-level features, the DCE module proves that "old school" geometric priors (like pixel correlation) are still incredibly useful for preserving the topology of 1-pixel-wide structures.
  • Correcting the "Gold Standard": The authors found that even professional manual annotations often miss tiny vessels. In their results, D2Net actually identified continuous vascular structures that human experts had overlooked, suggesting the model might be reaching a level of "super-human" consistency in microvascular detection.

Key Results

  • Microvascular Superiority: On the FOCA dataset, D2Net achieved a Dice score of 63.48%, outperforming the next best method (CSNet) by over 2 percentage points.
  • Robustness: When tested on the ROSE-H dataset (which was not used during training), D2Net maintained high performance (83.83% AUC), proving it isn't just memorizing one specific machine's output.
  • Clinical Utility: The researchers applied D2Net to scans of patients with Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI). They found statistically significant drops in Vascular Density (VAD) and Vascular Length Density (VLD) in AD patients compared to healthy controls, reinforcing the potential of OCTA as a biomarker for neurodegenerative diseases.

Limitations and Future Work

The authors acknowledge that while D2Net excels at small Field of View (FOV) scans (3mm x 3mm), it faces challenges when scaling to larger FOVs. There are significant discrepancies in how vascular structures appear in small vs. large scans, and noise/artifacts are often misidentified as vessels in larger, lower-resolution images.

Future work will focus on cross-field segmentation, aiming to bridge the gap between high-detail micro-scans and wide-field retinal imaging, which is crucial for full-scale clinical diagnostics.

Summary updated

Abstract

To precisely evaluate our method on small vessels, we delicately establish OCTA microvascular labels by performing comprehensive and detailed annotations on the FOCA dataset, which includes data collected from different instruments, and evaluated the proposed D2Net effectively mitigates the challenges of microvasculature region recognition caused by noise and artifacts. The method achieves more refined segmentation performance. In addition, we validated the performance of D2Net on four OCTA datasets (OCTA-500, ROSE-O, ROSE-Z, and ROSE-H) acquired using different instruments, demonstrating its robustness and generalization capabilities in retinal vessel segmentation compared to other state-of-the-art methods.
BibTeX
@inproceedings{hu2025dual,
  author = {Xiaoyang Hu and Jinkui Hao and Quanyong Yi and Yitian Zhao and Jiong Zhang},
  title = {{Dual-stream disentangled model for microvascular extraction in five datasets from multiple OCTA instruments}},
  booktitle = {Frontiers in Medicine},
  year = {2025},
  doi = {10.3389/fmed.2025.1542737},
}
Figure