Retinal image enhancement with artifact reduction and structure retention

Highlights：

• We develop an unpaired fundus image enhancement method, which can effectively reduce artifacts and ensure structural consistency.

• We summarize three causes of artifacts | severe blurriness, imperfect illumination, and misleading information. High frequency prior is incorporated into our generative networks to reduce the artifacts by the proposed high-frequency extractor.

• A feature descriptor is trained alternately with the generator to ensure the fidelity of image structure. Pseudo-label loss is proposed to extract a better vessel feature in blurry images.

• Both visual comparison and quantitative evaluation prove the superiority of this method. And the enhancement can improve retinal image processing such as vessel segmentation, disease classification.

摘要

•We develop an unpaired fundus image enhancement method, which can effectively reduce artifacts and ensure structural consistency.•We summarize three causes of artifacts | severe blurriness, imperfect illumination, and misleading information. High frequency prior is incorporated into our generative networks to reduce the artifacts by the proposed high-frequency extractor.•A feature descriptor is trained alternately with the generator to ensure the fidelity of image structure. Pseudo-label loss is proposed to extract a better vessel feature in blurry images.•Both visual comparison and quantitative evaluation prove the superiority of this method. And the enhancement can improve retinal image processing such as vessel segmentation, disease classification.

论文评审过程：Received 21 April 2022, Revised 14 June 2022, Accepted 9 August 2022, Available online 10 August 2022, Version of Record 15 September 2022.