低剂量CT图像的噪声提取研究

Study on Noise Extraction from Low-Dose CT Images

为了解决CT辐射剂量降低时重建图像质量低的问题,研究了基于StyleGAN2 GAN学习对给定噪声图像的噪声提取,采用从训练的GAN模型中采样大量噪声块的方法,提取噪声特征而不是噪声图像。通过引入轻量级Squeeze-and-Excitation(SE)模块可以更好地为图像不同部分分配权重,使得浅层网络的局部信息在图像分割时能够很好地分割边缘细节图像,深层网络输出的特征图可以捕捉同一幅图像的不同尺度信息。实验结果表明,采用本文方法处理低剂量CT图像的细节还原度真实、局部器官光滑性较好。

To solve the problem of low quality reconstructed images when the radiation dose of CT is degraded,noise extraction based on StyleGAN2 GAN learning for a given noisy image is investigated.This paper employs an approach that samples a large number of noise extraction from the trained GAN model to extract noise features instead of noisy images.By introducing a light-weight Squeeze-and-Excitation(SE)module to better assign weights to different sections of the image,the local information from the shallow network can segment the edge-detail image well during image segmentation,and the feature maps output from the deep network can capture different scale information of the same image.The experimental results show that the low-dose CT images processed using the approach described in this paper have realistic detail reproduction and better local organ smoothness.