Underexposed fundus image enhancement via illumination guidance and retinal structural modeling

Fundus images are used to diagnose ocular diseases, and a reasonable exposure of fundus images with rich details makes clinical diagnostic results more reliable. However, the exposure of fundus images is often unsatisfactory due to the limited clinical environments and the doctor-patient cooperation. To effectively enhance the exposure of fundus images, a transformer-based model called Illumination Guidance and Retinal Structural Modeling Network (IGRSM-Net) is proposed, where the supervised network predicts the illumination information of the image and simultaneously models the retinal structures to obtain enhanced fundus images. Specifically, an Illumination Guidance Enhancement Block (IGEB) that utilizes illumination representations is proposed to direct the modeling of non-local interactions of regions under different exposure conditions. Unlike simply increasing the image's overall exposure, the illumination prior obtained by Retinex is used to adjust the image exposure further flexibly. To improve the clarity of the retina, a structural modeling branch is developed to compensate for retinal structures, and the retinal structure features are adaptively merged with the enhanced image through the proposed multi-scale fusion strategy. Extensive experiments show that the proposed method achieves competitive performance compared to the state-of-the-art (SOTA) methods in exposure level and vascular texture clarity.