Token2Token Top-k Vision Transformer for Low-dose CT Denoising

Token2Token Transformer methods have shown outstanding performance in medical CT image denoising, primarily due to their ability to gradually aggregate surrounding tokens and enhance local information. However, existing methods struggle to dynamically distinguish between useful information and noise during feature aggregation, which affects the quality of image reconstruction. To address this issue, we propose an efficient low-dose CT denoising model, Token2Token Top-k Vision Transformer (TKformer). This model selectively retains the most relevant self-attention values to distinguish between image information and noise, thereby enhancing the precision of feature aggregation and facilitating high-quality image reconstruction. Specifically, we introduce Top-k Attention (TKA), which adaptively selects the most relevant features after reshaping, suppressing unnecessary information flow and improving aggregation precision. The synergistic effect between TKA and DGFN enables the model to better focus on key image details while effectively suppressing noise. The experimental results demonstrate that the proposed model outperforms state-of-the-art methods on the AAPM 3mm dataset, showcasing exceptional denoising and image recovery capabilities.