一种用于低剂量CT的微小细节保护CNN与Transformer融合去噪方法

Low-dose CT denoising method with CNN and Transformer to preserve tiny details

为解决低剂量CT图像因辐射剂量降低而引入大量噪声,导致图像质量下降,从而影响临床诊断准确性问题,构建一种结合卷积神经网络(CNN)与Transformer的网络模型,并在此模型中引入一种内部块特征提取模块,以更好地保护图像中的微小细节。此外,为了解决应用Swin Transformer去噪时出现恢复错误纹理细节的问题,在自注意力部分并入一个多输入通道注意力模块,进而构建一种双重注意力Transformer。本研究在AAPM数据集上进行测试,实验结果表明,与现有的去噪算法相比,本文提出的算法在去噪方面表现出色,可以更好地保护图像的微小细节。

Given that low-dose computed tomography significantly amplifies image noise due to the mitigation of radiation exposure,which degrades image quality and lowers the precision of clinical diagnoses,a novel model incorporating convolutional neural network and Transformer is established,in which an intra-patch feature extraction module is used to effectively preserve tiny details in the image.A double attention Transformer is constructed by incorporating a multiple-input channel attention module into the self-attention for tackling the problem of incorrect restoration of texture details during denoising using Swin Transformer.AAPM dataset is used for testing,and the results demonstrate that the proposed algorithm not only surpasses the existing algorithms in denoising performance,but also excels in preserving tiny details in the image.