CSGAN: a consistent structural GAN for AS-OCT image despeckling by image translation

Sanqian Li, Muxing Xiong, Risa Higashita, Jiang Liu

Keywords: AS-OCT, image despeckling, structural consistency, GAN

TL;DR: We incorporate some strategies into the style learning network for preserving structural details among feature domain, spatial dimensions, and frequency space.

Abstract: Anterior segment optical coherence tomography (AS-OCT) is a recent imaging technique for visualizing the physiological structure of the anterior segment. The speckle noise inherited in ASOCT images degrades the visual quality and hampers the subsequent medical analysis. Previous work was devoted to removing the speckles and acquiring satisfying images. According to the clinical requirements, it might be desirable to maintain locally higher data fidelity instead of enforcing visually appealing but rather wrong image structural features. Catering to this expectation, we propose a Consistent Structural Generative Adversarial Network (CSGAN) to learn the clean style of low-speckle in repeated AS-OCT images and simultaneously preserve the tiny but vital structural knowledge among the latent feature, spatial and frequency domains. Specifically, we design a latent constraint into the generator to capture the inherent content in the feature domain and adopt the perceptual similarities to directly preserve structural detail in the spatial dimension. Besides, we introduce a focal frequency scheme that adaptively represents and distinguishes hard frequencies to compensate for the spatial loss and refine the generated image to improve image quality. Finally, the experimental results demonstrate that the CSGAN can achieve satisfactory despeckling results with preserving structural details on the AS-Casia dataset.