HDMamba:基于状态空间模型的低剂量CT去噪方法OA

In low-dose CT denoising tasks,to address the issues of excessive GPU memory consumption caused by the high com-putational complexity of the quadratic calculation in self-attention mechanism-based models and loss of high-frequency details due to lack of adaptability in multi-scale feature fusion,this paper proposes a novel denoising model called HDMamba.The model adopts an encoder-decoder architecture with a Haar wavelet-based feature enhancement(HWFE)module deployed at the front end to provide richer feature representations for subsequent processing.Within each encoder-decoder layer,a multi-scale feature-aware state space module(MFSSM)is introduced to replace traditional self-attention mechanisms,employing spatial SSMs to achieve linear-complexity long-range dependency capture.Additionally,learnable parameters α and β are incorporated after the dilated convolution branch in the multi-scale feed-forward network to dynamically weight features at different scales.Experimental results on the AAPM dataset demonstrate that the proposed model maintains a low parameter count(31.13 M)while significantly outperforming existing denoising methods in all other evaluation metrics.