Unsupervised retina image synthesis via disentangled representation learning

Abstract

Fluorescein Fundus Angiography (FFA) is an effective and necessary imaging technology for many retinal diseases including choroiditis, preretinal hemorrhage, and diabetic retinopathy. However, due to the invasive operation, harmful fluorescein dye, and the consequent side effects and complications, it is also an image modality that both doctors and patients are reluctant to use. Therefore, we propose an approach to use Fluorescein Fundus (FF) images, which are non-invasive and safe, to synthesize the invasive and harmful FFA images. Additionally, since paired data are rare and time-consuming to get, the proposed method uses unpaired data to synthesize FFA images in an unsupervised way. Previous unpaired image synthesis methods treat image translation between two domains in two separate ways and thus ignore the implicit feature correlation in the translation process. To solve that, the proposed method first disentangles domain features into domain-shared structure features and domain-independent appearance features. Guided by the adversarial learning, two generators will learn to synthesize FFA-like images and FF-like images correspondingly. Perceptual loss are introduced to preserve the content consistency during translation. Qualitative results show that our model could generate realistic and mimic images without the usage of paired data. We also make quantitative comparisons on Isfahan MISP dataset to demonstrate the superior image quality of the synthetic images.