Abstract: Abstract: In recent years, significant progress has been made in using Generative Adversarial Network (GAN) for low-dose CT (LDCT) image denoising. However, existing methods still face challenges such as insufficient modeling capability for complex noise distributions and limited ability to preserve structural details. To address these issues, a Trident Generative Adversarial Network (Trident GAN) for LDCT image denoising was proposed. Firstly, a Trident Uformer was designed. In this generator, a Feature Aggregation Attention (FPA) module was added at the bottleneck layer of the U-Net structure, solving the problem of low spatial resolution in a U-shaped structure. Secondly, a multi-path feature extraction submodule (Trident Block) was designed, in each of the three branches a Local Detail Enhancement Block (LDEB) was introduced to extract detailed features, a Lightweight Channel Attention Block (LCAB) was incorporated to enhance channel-wise features, and a Spatial Interaction Attention Block (SIAB) was utilized to capture important spatial features. Within the SIAB, a multi-level interactive attention function and evaluation mechanism was employed to design the SCAM, addressing the limitations of single-attention mechanisms. Finally, a multi-feature fusion (MFF) module was designed to realize the feature aggregation at the end of the three branches. In the MFF, both local detail information and global semantic information are modelled, which deals with the problem of discontinuous details across different levels. Furthermore, the Multi-Scale Pyramid Discriminator (MSPD) was used to check the quality of the generated results at different dimensions, guiding the generation of globally consistent images. Experimental results show that the proposed method achieves average PSNR and SSIM metrics of 31.5193dB/0.8830 and 33.6331dB/0.9478 on both Mayo and Piglet datasets. Compared with HFSGAN, the number of parameters was reduced by 75% and the testing time was increased by 36%. Compared with existing advanced denoising methods, Trident GAN improves image quality with less computational load.

Key words: Keywords: Low-Dose Computed Tomography (LDCT), image denoising, attention mechanism, Transformer, Generative Adversarial Network (GAN)

摘要： 摘 要: 近些年，生成对抗网络(GAN)用于低剂量CT(LDCT)图像降噪取得了显著进展。然而现有方法仍存在对复杂噪声分布建模能力不足、结构细节保留能力有限等问题。为此，提出一种用于LDCT图像降噪的多路特征生成对抗网络(Trident GAN)。首先，设计特征引导生成器，通过在U-Net结构的瓶颈层增加特征聚合注意(FPA)模块，解决U型结构空间分辨率较低的问题。其次，设计多路特征提取子模块(Trident Block)，在三个分支中分别引入局部细节增强模块(LDEB)提取细节特征，轻量通道注意力模块(LCAB)增强通道特征，空间交互注意力模块(SIAB)获得重要空间特征。在SIAB中采用多级交互式注意力函数和评估机制设计注意力算子集成方法(SCAM)，解决单一注意力受限的问题。最后，设计多特征融合(MFF)模块，在三分支末端进行特征聚合，对局部细节信息和全局语义信息进行建模，解决不同层次之间细节不连续的问题。利用多尺度金字塔判别器(MSPD)在不同维度下检查生成结果的质量，指导生成具有全局一致性的图像。实验结果表明，在Mayo和Piglet数据集上，所提方法的平均峰值信噪比和结构相似性指标分别达到31.5193dB/0.8830、33.6331dB/0.9478，与HFSGAN相比，参数量降低75%，测试时间提高36%。与现有先进降噪方法相比，Trident GAN在较少的计算负荷下提高了图像质量

关键词: 关键词: 低剂量计算机断层扫描, 图像降噪, 注意力机制, Transformer, 生成对抗网络