Image denoising substantially improves accuracy and precision of intravoxel incoherent motion parameter estimates

Applicability of intravoxel incoherent motion (IVIM) imaging in the clinical setting is hampered by the limited reliability in particular of the perfusion-related parameter estimates. To alleviate this problem, various advanced postprocessing methods have been introduced. However, the underlying algorithms are not readily available and generally suffer from an increased computational burden. Contrary, several computationally fast image denoising methods have recently been proposed which are accessible online and may improve reliability of IVIM parameter estimates. The objective of the present work is to investigate the impact of image denoising on accuracy and precision of IVIM parameter estimates using comprehensive in-silico and in-vivo experiments. Image denoising is performed with four different algorithms that work on magnitude data: two algorithms which are based on nonlocal means (NLM) filtering, one algorithm that relies on local principal component analysis (LPCA) of the diffusion-weighted images, and another algorithms that exploits joint rank and edge constraints (JREC). Accuracy and precision of IVIM parameter estimates is investigated in an in-silico brain phantom and an in-vivo ground truth as a function of the signal-to-noise ratio for spatially homogenous and inhomogenous levels of Rician noise. Moreover, precision is evaluated using bootstrap analysis of in-vivo measurements. In the experiments, IVIM parameters are computed a) by using a segmented fit method and b) by performing a biexponential fit of the entire attenuation curve based on nonlinear least squares estimates. Irrespective of the fit method, the results demonstrate that reliability of IVIM parameter estimates is substantially improved by image denoising. The experiments show that the LPCA and the JREC algorithms perform in a similar manner and outperform the NLM-related methods. Relative to noisy data, accuracy of the IVIM parameters in the in-silico phantom improves after image denoising by 76–79%, 79–81%, 84–99% and precision by 74–80%, 80–83%, 84–95% for the perfusion fraction, the diffusion coefficient, and the pseudodiffusion coefficient, respectively, when the segmented fit method is used. Beyond that, the simulations reveal that denoising performance is not impeded by spatially inhomogeneous levels of Rician noise in the image. Since all investigated algorithms are freely available and work on magnitude data they can be readily applied in the clinical setting which may foster transition of IVIM imaging into clinical practice.体素内不相干运动成像（intravoxel incoherent motion, IVIM）在临床场景中的应用，受制于其参数估计（尤其是灌注相关参数估计）的可靠性不足。为缓解这一难题，学界已提出多种先进后处理方法，但此类算法往往难以直接获取，且普遍存在计算负荷增加的缺陷。与之相反，近年来已有多款计算效率优异的图像去噪方法问世，此类方法可在线获取，且有望提升IVIM参数估计的可靠性。本研究旨在通过全面的仿真（in-silico）与在体（in-vivo）实验，探究图像去噪对IVIM参数估计准确性与精准度的影响。本次实验采用四种针对幅度数据的去噪算法：两种基于非局部均值（nonlocal means, NLM）滤波的算法、一种依赖扩散加权图像局部主成分分析（local principal component analysis, LPCA）的算法，以及一种利用联合秩与边缘约束（joint rank and edge constraints, JREC）的算法。我们针对空间均匀与非均匀莱斯噪声（Rician noise）水平下的信噪比变化，分别在仿真脑体模与在体真值数据中，考察IVIM参数估计的准确性与精准度。此外，我们还通过对在体测量数据进行自助法分析（bootstrap analysis），评估了参数估计的精准度。实验中，IVIM参数的计算分别采用两种方式：a）分段拟合方法（segmented fit method）；b）基于非线性最小二乘估计（nonlinear least squares estimates）的全衰减曲线双指数拟合（biexponential fit）方法。无论采用何种拟合方式，实验结果均表明，图像去噪可显著提升IVIM参数估计的可靠性。实验显示，LPCA与JREC算法的表现相近，且优于两类NLM相关去噪方法。当使用分段拟合方法时，相较于含噪数据，仿真脑体模中各IVIM参数的准确性提升幅度为76%~79%、79%~81%、84%~99%，精准度提升幅度则分别为74%~80%、80%~83%、84%~95%（对应灌注分数、扩散系数与伪扩散系数）。除此之外，仿真实验还表明，图像中的空间非均匀莱斯噪声不会削弱去噪效果。由于本次研究涉及的所有算法均可免费获取且支持幅度数据处理，因此可直接应用于临床场景，有望推动IVIM成像向临床实践转化。