DeLashNet: A Deep Network for Eyelash Artifact Removal in Ultra-Wide-Field Fundus Images

Abstract

The interference of eyelash artifacts in ultra-wide-field fundus (UWF) images has always been a serious problem in preventing precise clinical observations of pathology. Currently, the automatic removal of eyelash artifacts in UWF images remains unsolved and thus will eventually affect the diagnosis accuracy. In this paper, we propose a deep learning architecture called DeLashNet to eliminate eyelash artifacts from UWF images. Our DeLashNet consists of two stages: the first stage is the eyelash artifact removal stage based on a conditional generative adversarial network, and the second stage is the background refinement stage using an encoder-decoder structure. To solve the issue of lacking training samples with eyelashes, we design a novel eyelash growing model to generate synthetic eyelashes with labels and finally established a paired synthetic eyelashes (PSE) dataset. Experiments are conducted to verify the effectiveness of our proposed DeLashNet on eyelash artifact removal. The comparative and ablation studies demonstrate that the proposed DeLashNet achieved satisfactory removal performance on eyelash artifacts of UWF images.