SGF-RWKV: Semantic-Guided and Frequency-Enhanced RWKV for Unified Medical Image Denoising

Medical image denoising is essential for enhancing diagnostic accuracy and clinical applicability. Since diverse imaging modalities exhibit heterogeneous noise characteristics, developing a unified all-in-one frame work is critical to avoid modality-specific solutions and improve generalization. However, existing approaches struggle with three key challenges: difficulty in uniformly modeling multimodal noise, loss of structural details, and limited spatial context modeling. To addressthis issue, wepropose SGF-RWKV, a unified medical image denoising framework built upon the Receptance Weighted Key Value (RWKV) model for consistent noise modeling across modalities. Specifically, a semantic guided module leverages CLIP based language embeddings to generate task-aware attention maps for adaptive noise suppression, a frequency enhancement module refines structural and edge details via Fourier magnitude modeling, and a dilated multi-scale shift mechanism extends spatial context modeling to preserve continuity and fine texture. By jointly integrating semantic, frequency, and spatial cues, SGF-RWKV achieves modality agnostic denoising within a unified framework. We con ducted comprehensive experiments on five publicly available medical imaging datasets spanning multiple modalities, including MRI, CT, ultrasound, OCT, and cryo-EM. Experimental results demonstrate that our unified denoising framework achieves superior performance across diverse medical imaging modalities, excelling in key quantitative metrics such as PSNR, SSIM, and RMSE. The source code is available at https://github.com/zwj666a/SGF_RWKV.