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

Optical coherence tomography angiography (OCTA) usually suffers from the inherent random fluctuations of noise and speckles in the imaging system. Previous deep learning methods have mainly focused on improving the quality of B-scan blood flow images or en face projection images. We propose a deep learning method to reconstruct high-quality 3D vasculature, which fully utilizes the volumetric OCTA data and the topological features of the vascular network.

We propose a deep learning method called the three-dimensional C-scan-based generation adversarial network (3DCS-GAN) to improve vascular visualization for volumetric OCTA data.

To train the network, we superimposed the single-shot en face OCTA images on avascular noisy C-scan images to synthesize the input data and used the multiple averaged en face OCTA images as the reference labels. The deep learning algorithm is based on Pix2Pix architecture and consists of a generator model and a discriminator model. A perceptual loss function was utilized by combining content loss and adversarial loss. The proposed algorithm is applied to the C-scan images depth-by-depth to suppress the background noise and enhance vascular visualization in the 3D OCTA data.

The proposed method has improved the contrast-to-noise ratio of cross-sectional OCTA images by $\sim 2$ times. It greatly enhances the visualization of blood vessels in the deep layer and offers much clearer blood vessel topology in the 3D volume-rendering of OCTA. 3DCS-GAN has exhibited superior image enhancement compared with alternative methods. It has been used to enhance the OCTA images of port wine stain disease for clinical investigation.

It demonstrates that the proposed 3DCS-GAN can greatly improve vascular visualization in the deep layer, provide better image quality than the multiple averaged OCTA images, and achieve superior image enhancement for volumetric OCTA data.