Three-dimensional C-scan-based generation adversarial network with synthetic input to improve optical coherence...

Area of Science:

• Medical Imaging
• Deep Learning
• Vascular Biology

Background:

• Optical coherence tomography angiography (OCTA) imaging is often degraded by noise and speckles.
• Existing deep learning methods primarily focus on improving 2D OCTA images (B-scan or en face).
• Reconstructing high-quality 3D vasculature from volumetric OCTA data remains a challenge.

Purpose of the Study:

• To propose a novel deep learning method for reconstructing high-quality 3D vasculature from volumetric OCTA data.
• To fully utilize volumetric OCTA data and vascular network topological features for improved visualization.
• To enhance the visualization of deep-layer blood vessels in OCTA.

Main Methods:

• Developed a three-dimensional C-scan-based generative adversarial network (3DCS-GAN).
• Synthesized training data by superimposing en face OCTA images onto noisy C-scan images.
• Employed a Pix2Pix-based architecture with generator and discriminator models, incorporating perceptual loss (content and adversarial loss).
• Applied the algorithm depth-by-depth to C-scan images to reduce noise and enhance vascular visualization.

Main Results:

• Significantly improved the contrast-to-noise ratio of cross-sectional OCTA images.
• Greatly enhanced visualization of deep-layer blood vessels and clarified 3D vascular topology.
• Demonstrated superior image enhancement compared to alternative methods.
• Successfully applied to enhance OCTA images for port wine stain disease investigation.

Conclusions:

• The proposed 3DCS-GAN effectively improves deep-layer vascular visualization in OCTA.
• Achieved superior image quality and enhancement for volumetric OCTA data compared to averaged images.
• Validates the potential of 3DCS-GAN for clinical investigation and improving OCTA analysis.