Complex conjugate artifact removal in FD-OCT using generative adversarial network

In frequency-domain optical coherence tomography (OCT) only half of the available depth range is used. This is due to the occurrence of complex conjugate (CC) ambiguity, which is an artifact resulting from the symmetry properties of the Fourier transform on real-valued spectrum that undermines the optimal sensitivity window. Current approaches require additional active or passive components, and increase systems complexity and cost. We present a novel deep-learning method for CC removal (CCR) based on a generative adversarial network (GAN). The model was trained to learn how to translate OCT scans with CC artifacts into full range images without the requirement of additional equipment or measurement. The data was collected from a phantom sample and human skin in vivo, using a swept source-OCT prototype. The GAN architecture adopted is based on the Pix2Pix model, where the discriminator is a PatchGAN and the generator is a U-net with skipped connections, and has been adapted for high resolution images of 864 x 1024 pixels. CCR-GAN receives as input the complete OCT signal, which consists of intensity and phase images. The findings and the evaluation metrics show that our model is able to effectively suppress CC artifact in OCT scans thereby providing a doubled imaging range. We demonstrated that our model is superior to prior approaches with respect to design complexity, imaging speed, and cost. CCR-GAN can be effectively used to suppress the CC mirror terms and generate full depth range in clinical imaging, that requires large ranging depth and high sensitivity.