WM-UNet: A Wavelet-Mamba Hybrid Architecture for Low-Dose CT Denoising

Low-dose computed tomography technology consistently faces a critical contradiction: while achieving reduced ionizing radiation dose, it inevitably suffers from image quality degradation caused by quantum noise and artifacts, which directly constrains the improvement of diagnostic efficacy. Although current denoising methods provide some enhancement in imaging quality, they still exhibit limitations in noise suppression, detail preservation, and computational efficiency. This paper proposes a novel Wavelet Mamba UNet framework (WM-UNet). Firstly, we introduce a Mamba-based 2D-Selective-Scan (SS2D) module that adaptively captures cross-dimensional spatial correlations through direction-aware state propagation and enhances spatial perception via multi-directional scanning trajectories, while preserving fine-grained details through residual feature fusion across encoder-decoder stages. Secondly, a wavelet transform layer is incorporated to decompose images into multi-scale components, enabling hierarchical processing of both high-frequency and low-frequency subbands. This mechanism effectively captures subtle lesion margins and vascular textures in CT scans while maintaining the global anatomical structure of organs. Experimental results demonstrate that WM-UNet achieves a PSNR of 44.73 dB with merely 0.87 million parameters, outperforming competing algorithms.