Removing Stray-Light for Wild-Field Fundus Image Fusion Based on Large Generative Models

In low-cost wide-field fundus cameras, the built-in lighting sources are prone to generate stray-light nearby, leading to low-quality image regions. To visualize retinal structures clearer, when fusing two images with complementary patterns of different lighting sources, the fused image might still have stray-light phenomenon near the hard fusing boundaries, i.e., typically in the diagonal directions. In this paper, an image enhancement algorithm based on generative adversarial network is proposed to eliminate the stray-light in wide-field fundus fusing images. First, a haze density estimation module is introduced to guide the model to pay attention to more serious stray-light regions. Second, a detail recovery module is introduced to reduce the loss of details caused by stray-light. Finally, a domain discriminator with unsupervised domain adaptation is employed to achieve better performance generalization on clinical data. Experiments show that our method obtains the best results on both public synthesized traditional fundus image dataset EyePACS-K and private wide-field fundus images dataset Retivue. Compared to the SOTA, the PSNR and Structural Similarity on average upon two above datasets are increased by 1.789 dB and 0.021 respectively.