Super-resolution of point-scanning ultra-widefield fundus images based on inter-channel heterogeneity and spatial non-uniformity

Point-scanning laser ophthalmoscopes offer high contrast and a large field of view for fundus imaging; however, the point-scanning mechanism requires longer acquisitions for high resolution, increasing patient discomfort. We propose a deep learning super-resolution method that reconstructs high-resolution images from low-resolution acquisitions, shortening examination time while preserving image quality. To capture the spatial and channel non-uniformity of information in fundus images, we integrate a convolutional block attention module (CBAM) into the Real-ESRGAN baseline, using its channel and spatial attention to improve sensitivity to fine structures such as capillaries. Considering detector noise in point-scanning imaging, we replace the conventional L 1 loss with a vessel-aware weighted Charbonnier loss to balance feature enhancement and denoising. On a self-collected point-scanning fundus dataset, our method achieves improved PSNR and SSIM together with reduced LPIPS and fine-vessel weighted edge width (FV-WEW) for both × 2 and × 4 super-resolution, indicating improvements in reconstruction fidelity, perceptual quality, and fine-vessel sharpness. Tests on clinical images acquired at resolutions of 512 × 512 and 1024 × 1024 further demonstrate stable reconstruction performance across different sampling modes. Our method substantially improves image clarity while preserving topological consistency and suppressing artifacts, offering a reliable technical solution for efficient, high-fidelity ultra-widefield fundus imaging in clinical settings.