Disentangled Representation Learning for OCTA Vessel Segmentation With Limited Training Data.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography angiography (OCTA) is crucial for retinal vasculature analysis.
• Accurate capillary segmentation in OCTA images is essential for quantitative assessment.
• Deep learning for OCTA segmentation faces challenges with extensive manual labeling and contrast-related artifacts.

Purpose of the Study:

• To develop a deep learning method for accurate OCTA vessel segmentation.
• To address the limitations of manual data acquisition and inherent image artifacts in OCTA.
• To improve the efficiency and reliability of quantitative OCTA analysis.

Main Methods:

• A novel deep learning approach was proposed to disentangle anatomy and local contrast components from paired OCTA scans.
• The method utilizes a contrast-removed anatomy component as input for a deep learning model.
• This allows for segmentation learning with a reduced need for manually labeled training images.

Main Results:

• The proposed method effectively addresses the challenges of limited manual delineations and OCTA image artifacts.
• Segmentation models trained on the disentangled anatomy component require significantly less manual labeling.
• The approach demonstrates state-of-the-art performance in OCTA vessel segmentation.

Conclusions:

• The novel disentanglement method significantly improves deep learning-based OCTA vessel segmentation.
• This approach reduces the labor-intensive requirement for manual annotations.
• The findings offer a more efficient and robust solution for quantitative OCTA analysis in clinical settings.