目的 观察基于Z轴相关性Zero-Shot Noise2Noise(ZS-N2N)方法降低低剂量CT(LDCT)图像噪声的价值.方法 选取癌症成像档案CT数据集,包括正常剂量CT(NDCT)图像和LDCT图像,胸、腹部图像各3组.采用ZS-N2N方法基于Z轴相关性降低LDCT图像噪声,与Self2Self、单纯ZS-N2N及传统Block-matching and 3D filtering(BM3D)方法进行对比,观察各算法峰值信噪比(PSNR)、结构相似度(SSIM)及降噪耗时.结果 降噪后,Self2Self降噪图像噪声明显;BM3D降噪图像结构边缘较模糊,存在部分细节丢失;单纯ZS-N2N和基于Z轴相关性的ZS-N2N降噪图像留有更多细节,质量较好.以Self2Self降低LDCT图像噪声的PSNR和SSIM较差、耗时较长,其余3种方法的PSNR、SSIM和耗时均相近;其中,基于Z轴相关性ZS-N2N的PSNR略高于BM3D和单纯ZS-N2N,但耗时仍略长.结论 基于Z轴相关性ZS-N2N对降低LDCT图像噪声具有较高价值.
Reducing noise of low dose CT images with Zero-Shot Noise2Noise based on Z-axis correlation
Objective To observe the value of Zero-Shot Noise2Noise(ZS-N2N)based on Z-axis correlation for reducing noise of low dose CT(LDCT)images.Methods CT data of the cancer imaging archive were enrolled,including normal dose CT(NDCT)images and LDCT images,with 3 sets of chest and 3 sets of abdominal images.Noise on LDCT images were reduced with ZS-N2N method based on Z-axis correlation,and the peak signal-to-noise ratio(PSNR),structural similarity(SSIM)and time-consuming of reducing noise were compared with those of Self2Self,simple ZS-N2N and traditional Block-matching and 3D filtering(BM3D).Results After reducing noise,noise on Self2Self denoised images remained significant,the structure edges on BM3D denoised images were blurry with some details lost,while simple ZS-N2N and ZS-N2N based on Z-axis correlation denoised images preserved more details and had better quality.PSNR and SSIM of Self2Self denoised images were poor and the time-consuming were longer.PSNR,SSIM and time-consuming of the other 3 methods were similar,among which PSNR of ZS-N2N based on Z-axis correlation were slightly higher than BM3D and simple ZS-N2N,but the time-consuming were also slightly longer.Conclusion ZS-N2N based on Z-axis correlation had high value for reducing noise of LDCT images.
NETL
NSTL
万方数据
