Structure Preservation Generative Adversarial Network for Noise Reduction in SD-OCT Images

In order to remove speckle noise in spectral-domain optical coherence tomography(SD-OCT),this paper proposes a structure preservation generative adversarial network to generate high quality enhanced-depth imaging optical coherence tomography(EDI-OCT)from SD-OCT images.The proposed method learns the mapping from the SD-OCT to EDI-OCT domain based on a cycle-consistent adversarial network(Cycle-GAN)architecture.To alleviate the structural difference incurred by the Cycle-GAN,our model forces a global loss function to preserve global structural consistency utilizing the similarity of the continuous frame.Meanwhile,the local loss function utilizes a modality independent neighborhood descriptor to preserve anatomic details.The experimental results of denoising on 50 Cirrus SD-OCT datasets show that the algorithm has a PSNR value of 29.03 dB,a SSIM value of 0.82,and an EPI value of 0.50,which are better than the existing algorithms.