结合感知损失与双重对抗网络的低剂量CT图像去噪

低剂量计算机断层扫描（LDCT）成像技术在医学诊断中得到广泛应用,但其斑纹噪声和非平稳条纹伪影复杂,目前多数算法仅依靠推断条件后验概率来实现图像去噪,无法应对LDCT图像噪声复杂、数据量少、先验知识缺乏的问题。提出一种结合感知损失的双重对抗网络去噪算法,以实现LDCT图像复原。该算法包含一个去噪器和一个生成器,分别从图像去噪和噪声生成2个角度来建模干净-噪声图像对的联合分布,通过联合学习使得去噪器和生成器相互指导,从而充分学习数据中的噪声信息和清晰图像信息,且学习到的去噪器可以直接用于LDCT图像修复。考虑到通过感知损失学习语义特征差异可以使去噪结果保留更多的细节和边缘信息,提出一种掩膜自监督方法,针对CT图像域训练一个语义特征提取网络用于计算感知损失。实验结果表明,与BM3D、RED-CNN、WGAN-VGG等主流去噪算法相比,该算法可以有效抑制噪声并去除伪影,最大程度地保留边缘轮廓和纹理细节,产生更符合人眼视觉特性的去噪效果,与当下LDCT图像去噪性能较好的SACNN算法相比,所提算法的PSNR和SSIM指标分别提升1.26 dB和1.8%。

Low-Dose CT Image Denoising Combining Perceptual Loss and Dual Adversarial Network

Low-Dose Computer Tomography(LDCT) imaging technology has been widely used for medical diagnosis, but its speckle noise and nonstationary fringe artifacts are complex.Currently, most algorithms only rely on inferential conditional posterior probability to achieve image denoising, which does not deal with the complex noise, small amount of data, and lack of prior knowledge of LDCT images.A dual adversarial network denoising algorithm combined with perceptual loss is proposed to achieve LDCT image restoration.The algorithm includes a denoiser and a generator, which jointly models the joint distribution of clean-noise image pairs from the perspectives of image denoising and noise generation, respectively. By joint learning, the denoiser and generator can guide each other to fully learn the noise information and clear image information in the data, and the learned denoiser can be directly used for LDCT image repairs.Considering that more details and edge information can be retained in denoising results by learning semantic feature differences through perceptual loss, a mask self-supervised method is proposed to train a semantic feature extraction network for the CT image domain to calculate the perceptual loss.The experimental results show that compared with mainstream denoising algorithms such as BM3D, RED-CNN, WGAN-VGG, this algorithm can effectively suppress noise and remove artifacts, maximally retain edge contour and texture details, and induce a denoising effect more consistent with human visual characteristics. Compared with the current SACNN algorithm with improved denoising performance for LDCT images, the Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity(SSIM) indicators of the proposed algorithm are improved by 1.26 dB and 1.8%, respectively.