冉瑞生,张思文,李进,等.基于特征融合的低剂量CT图像降噪方法[J]. 微电子学与计算机,2024,41(5):11-21. doi: 10.19304/J.ISSN1000-7180.2023.0279
引用本文: 冉瑞生,张思文,李进,等.基于特征融合的低剂量CT图像降噪方法[J]. 微电子学与计算机,2024,41(5):11-21. doi: 10.19304/J.ISSN1000-7180.2023.0279
RAN R S,ZHANG S W,LI J,et al. A noise reduction method of low-dose CT images based on feature fusion[J]. Microelectronics & Computer,2024,41(5):11-21. doi: 10.19304/J.ISSN1000-7180.2023.0279
Citation: RAN R S,ZHANG S W,LI J,et al. A noise reduction method of low-dose CT images based on feature fusion[J]. Microelectronics & Computer,2024,41(5):11-21. doi: 10.19304/J.ISSN1000-7180.2023.0279

基于特征融合的低剂量CT图像降噪方法

A noise reduction method of low-dose CT images based on feature fusion

  • 摘要: 近年来低剂量CT(Low Dose CT, LDCT)被广泛应用于临床诊断中,但LDCT会产生不规则的噪声。已有的降噪方法往往缺乏对全局特征信息的考虑,以及不注重边缘特征信息和重建图像的视觉效果。为此,提出了一种基于特征融合的低剂量CT图像降噪方法。首先,利用Transformer优异的全局感受野提取图像的全局特征信息,并利用卷积神经网络(Convolutional Neural Network, CNN)良好的局部特征提取能力提取图像的局部特征信息。在Transformer模块中加入维度变换思想,以更好地抑制噪声;在CNN模块中使用稠密连接的方式将浅层网络的特征信息复用于深层网络中,以此保存更多的特征信息。其次,为了获取更加丰富的图像细节特征,使用了改进的索伯边缘增强算子来加强模型对边缘特征信息的提取能力。最后,将Transformer模块和CNN模块获取的特征信息进行融合并输出重建图像。此外,为了使降噪重建后的图像有更好的质量和视觉效果,设计了一个多尺度复合损失函数。实验表明:在AAPM-Mayo数据集的降噪实验中,与当前主流的LDCT图像降噪方法相比,本文方法取得了更好的降噪效果。

     

    Abstract: Low-Dose CT (LDCT) is widely used in clinical diagnosis, but it also generates some irregular noise. The existing noise reduction methods often lack consideration of global feature information, as well as do not focus on edge feature information and the visual effect of the reconstructed image. Then, a feature fusion-based noise reduction method for low-dose CT images is proposed. Firstly, the excellent global receptive field of Transformer is used to extract the global feature information of the image, and the good local feature extraction ability of Convolutional Neural Network (CNN) is used to extract the local feature information of the image. The dimensional transformation idea is added to the Transformer module to better suppress the noise; a dense connection in CNN is used to reuse the feature information from the shallow network to the deep network, so as to achieve pre-fusion of features and preserve more feature information. An improved Sobel edge enhancement operator is used to enhance the model's ability to extract edge feature information. Finally, the feature information acquired by the Transformer module and the CNN module is fused and the reconstructed image is output. In addition, a multiscale composite loss function is designed to make the reconstructed image with better quality and visual effect after noise reduction. The experiments show that in the noise reduction experiments of AAPM-Mayo dataset, compared with the current mainstream LDCT image noise reduction methods, the method in this paper achieves a better noise reduction effect.

     

/

返回文章
返回