SemiET: Semi-supervised fundus enhancement with efficient Transformers

Retinal fundus imaging is essential for the screening, diagnosis, and management of ocular diseases. High-quality fundus photographs are a prerequisite for reliable clinical interpretation and automated analysis. However, in real-world acquisition, image quality varies widely due to defocus, illumination, motion blur, and sensor noise, making robust enhancement necessary. Traditional enhancement methods often fall short by either over-enhancing images or losing vital details. Similarly, deep learning approaches, especially CycleGAN-based models, struggle with preserving structural details and bridging domain gaps. To address these limitations, we introduce SemiET, a novel semi-supervised fundus enhancement framework that utilizes an efficient Transformer structure. SemiET jointly leverages paired synthetic degradations for pixel-level reconstruction and unpaired real low-quality images for distribution alignment. To make high-resolution Transformer enhancement efficient, SemiET adopts Restormer-inspired channel-wise attention, avoiding a large spatial-affinity map while preserving global-context modeling. To assess utility beyond PSNR and SSIM, we develop a comprehensive evaluation framework that spans quality grading, vessel and optic disc/cup segmentation, fovea localization, and held-out DDR lesion-preservation detection. Across multiple datasets, SemiET achieves leading full-reference image-quality results, competitive structural downstream-task performance, and the highest mAP@0.5 among the evaluated enhancement methods in the DDR probe, while retaining the smallest parameter count among the compared methods.