VMDiff: A novel hybrid CNN-Mamba generative framework for retinal phenotype image enhancement.

Retinal phenotype is crucial for the diagnosis of ophthalmic diseases. Although deep learning models exhibit strong generative capability and detailed reconstruction, they struggle to simultaneously capture fine local details and long-range global dependencies, resulting in poor phenotypic enhancement. To overcome these limitations, we propose VMDiff, a hybrid generative framework based on diffusion models that synergistically integrates CNN and Mamba architectures to enhance retinal images while preserving both local and global features. Specifically, we introduce a high-frequency extraction module (HFEM) with a dedicated high-frequency consistency loss to retain fine structural details more effectively. Moreover, we propose a progressive loss mechanism to decouple the optimization objectives, effectively balancing pixel-level accuracy and perceptual quality. Extensive experiments on synthetically degraded versions of multiple public datasets demonstrate that VMDiff significantly improves image quality. Notably, it achieves highly competitive performance on EyePACS (with a comparable PSNR of 27.8448 and a significantly superior SSIM of 0.9428), while setting state-of-the-art records on DRIVE and IDRiD with PSNR values of 26.7520 and 27.7434, and SSIM scores of 0.9609 and 0.9266, respectively. Further validation on the DRIVE dataset via a vessel segmentation task confirms VMDiff’s efficacy in downstream biomedical analysis, yielding the highest AUC (0.9715), ACC (0.9556), and IoU (0.9516), thereby highlighting its clinical potential.