A Structure-Consistency GAN for Unpaired AS-OCT Image Inpainting

Abstract

Anterior segment optical coherence tomography (AS-OCT) is a crucial imaging modality in ophthalmology, providing valuable insights into corneal pathologies. However, during AS-OCT imaging, intense signals in highly reflective regions can easily lead to saturation effects, resulting in pronounced stripes across the cornea. It compromises the image visual quality and impacts automated ophthalmic analysis. To address this issue, we propose an unsupervised Structure-Consistency Generative Adversarial Network (SC-GAN) that captures the underlying semantic structural knowledge in both the spatial domain and frequency space within the generative model. This strategy aims to mitigate the influence of bright stripes and restore corneal structural details in AS-OCT images. Specifically, SC-GAN introduces a stripe perceptual loss to extract visual representations by utilizing the perceptual similarity between striped and stripe-free images. Moreover, Fourier feature mapping is adopted to learn high-frequency information, thereby achieving crucial structure consistency. The experimental results demonstrate that the proposed SC-GAN can removes stripes while preserving crucial corneal structures, surpassing the competing algorithms. Furthermore, we validate the benefits of SC-GAN in the corneal segmentation task.