GDAFormer: Transformer-Driven Fundus Image Enhancement with Gated Dual-Attention

Fundus image enhancement plays a crucial role in retinal disease diagnosis but remains hindered by challenges such as uneven illumination, low contrast, and domain variability. We present GDAFormer, a novel transformer-based framework tailored for fundus image enhancement, incorporating multi-head depth-wise convolutional self-attention and a gated dual-attention fusion block. This design captures long-range dependencies while adaptively integrating global structures and local pathological features, preserving vascular continuity and mitigating over-smoothing. A segmentation-aware loss, guided by a pretrained retinal segmentation network, further reinforces structural fidelity without requiring paired supervision. Extensive evaluations on the FIQ, RCF, and RF datasets show that GDAFormer consistently outperforms state-of-the-art methods in both PSNR and SSIM metrics. Our approach achieves strong generalization across diverse imaging protocols, making it a robust and clinically meaningful enhancement tool.