Retinal image enhancement with artifact reduction and structure retention

Enhancement of low-quality retinal fundus images is beneficial to clinical diagnosis of ophthalmic diseases and computer-aided analysis. Enhancement accuracy is a challenge for image generation models, especially when there is no supervision by paired images. To reduce artifacts and retain structural consistency for accuracy improvement, we develop an unpaired image generation method for fundus image enhancement with the proposed high-frequency extractor and feature descriptor. Specifically, we summarize three causes of tiny vessel-like artifacts which always appear in other image generation methods. A high frequency prior is incorporated into our model to reduce artifacts by the proposed high-frequency extractor. In addition, the feature descriptor is trained alternately with the generator using segmentation datasets and generated image pairs to ensure the fidelity of the image structure. Pseudo-label loss is proposed to improve the performance of the feature descriptor. Experimental results show that the proposed method performs better than other methods both qualitatively and quantitatively. The enhancement can improve the performance of segmentation and classification in retinal images.