GTCG-Net：一种基于门控Transformer的CycleGAN视网膜图像增强方法-《集美大学学报（自然科学版）》

文章信息/Info

Title:: GTCG-Net:an Improved CycleGAN Retinal Image Enhancement Method Based on Gated Transformer

Author(s):: XIONG Bingqing; YU Yuanhu; College of Computer Engineering,Jimei University,Xiamen 361021,China

Keywords:: CycleGAN; Transformer; Gated MLP; skip connections; retinal image enhancement

摘要:: 提出一个基于门控Transformer的CycleGAN视网膜图像增强方法（GTCG-Net）。通过使用生成对抗网络和循环一致性损失来实现视网膜图像的自动增强，从而改善图像质量和病变部分的可视化效果。并利用门控MLP提取更有用的特征，将生成器设计为类UNet结构，使用跳跃连接进行特征融合，以得到丰富的局部特征和细节信息。实验结果显示与其他视网膜图像增强方法GFE-Net、SCRNET、I-SECRET等相比，GTCG-Net不仅可以有效的去除伪影、恢复眼底结构和病理特征，同时在多个数据集上表现出了较好的泛化能力。

Abstract:: A retinal image enhancement method based on CycleGAN with Gated Transformer (GTCG-Net) is proposed.Automatic retinal image enhancement is achieved by using Generative Adversarial Networks and Cyclic Consistency Loss to improve the image quality and visualization of the lesion section.And more useful features are extracted using gated MLP,and the generator is designed as a U-Net-like structure using skip connections for feature fusion to obtain rich local features and detail information.The experimental results show that compared with other retinal image enhancement algorithms GFE-Net,SCRNET,I-SECRET,etc.,GTCG-Net not only effectively removes artifacts and restores fundus structure and pathological features,but also shows better generalization ability on multiple datasets.