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

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of retinal vasculature is crucial for diagnosing eye diseases.
• Existing deep learning models struggle with noise, artifacts, and segmenting small vessels in OCTA images.

Purpose of the Study:

• To propose a novel Dual-stream Disentangled Network (D2Net) for robust retinal OCTA microvascular segmentation.
• To improve the precision of segmentation by mitigating noise and artifact interference.

Main Methods:

• Developed a dual-stream encoder to separately learn image artifacts and vascular features.
• Incorporated vascular structure prior, including Distance Correlation Energy (DCE) module, for disentangled representation learning.
• Created detailed OCTA microvascular labels on the FOCA dataset for precise small vessel evaluation.

Main Results:

• D2Net effectively mitigates challenges from noise and artifacts in microvasculature recognition.
• Achieved refined segmentation performance, particularly for small vessels.
• Demonstrated robust and generalized performance across four diverse OCTA datasets (OCTA-500, ROSE-O, ROSE-Z, ROSE-H).

Conclusions:

• The proposed D2Net offers superior retinal OCTA microvascular segmentation compared to state-of-the-art methods.
• D2Net shows significant potential for clinical applications in eye disease diagnosis and management.
• The method's robustness across different instruments highlights its practical utility.