摘要
为有效降低低剂量CT图像的噪声,提出一种结合特征金字塔(FPN)的生成对抗网络模型(FGAN)。该模型利用多尺度并行方法从低剂量图像中提取图像细节,在不改变主干模型的前提下,融合不同尺度结构特征,采用批量归一化以及带泄露修正线性单元提高网络性能。通过与非局部均值滤波(NLM)、三维块匹配滤波(BM3D)、多层感知机(MLP)以及生成对抗网络(GAN)4种算法相比,该算法不仅降低了低剂量CT图像的噪声和伪影,而且保留了更多的细节信息,提高了低剂量CT图像质量,验证了该算法的有效性。
In order to effectively reduce the noise of low-dose CT images, a novel generation adversarial network model(FGAN) combining feature pyramid(FPN) is proposed. The model uses multi-scale parallel method to extract image details from low dose images. Without changing the trunk model, the features of different scale structures are fused, and the network performance is improved by batch normalization and band leakage correction linear units. the proposed algorithm is compared by four algorithms: nonlocal mean filtering(NLM), three-dimensional block matching filtering(BM3 D), multilayer perceptron(MLP), and generation adversarial network(GAN). the proposed algorithm not only reduces the noise and artifacts of low-dose CT images, but also retains more detail information and improves the image quality CT low-dose. the effectiveness of the proposed algorithm is verified.
作者
高文波
孔慧华
连祥媛
Gao Wenbo;Kong Huihua;Lian Xiangyuan(College of Science,North University of China,Taiyuan 030051,China;Shanxi Key Laboratory of Signal Capturing Processing,North University of China,Taiyuan 030051,China)
出处
《国外电子测量技术》
北大核心
2021年第8期1-6,共6页
Foreign Electronic Measurement Technology
关键词
低剂量CT
生成对抗网络
多尺度
特征提取
图像去噪
low dose CT
generative adversarial networks
multi-scale
feature extraction
image denoising
