Xu Wei, Mueller Klaus
Visual Analytics and Imaging Lab, Center of Visual Computing, Computer Science Department, Stony Brook University, Stony Brook, New York 11794-4400, USA.
Med Phys. 2012 Aug;39(8):4748-60. doi: 10.1118/1.4736528.
Low-dose CT has attracted increasing attention due to growing concerns about radiation exposure in medical scans. However, the frugal use of x-ray radiation inevitably reduces the quality of the CT images, introducing artifacts such as noise and streaks which make the reconstructed images difficult to read in clinical routine. For follow-up CT exams a prior scan is often available. It typically contains the same anatomical structures, just somewhat deformed and not aligned. This work describes a two-step technique that utilizes this prior scan to achieve high-quality low-dose CT imaging, overcoming difficulties arising from noise artifacts and misalignment. We specifically focus on reducing the dose by lowering the number of projections. This gives rise to severe streak artifacts which possibly lower the readability of CT images to a larger extent than the fine-grained noise that results from lowering the mA or kV settings.
A common approach is to apply image filtering to reduce the noise artifacts. These techniques typically utilize pixel neighborhoods in the degraded image to estimate the true value of a pixel at the center of this neighborhood. However, this can lead to poor results when the image is severely contaminated under very low low-dose situations. We propose a method that utilizes the nondegraded, clean prior to determine higher quality pixel statistics to form the pixel estimates, supported by the matching scheme of the non-local means filter. To make this matching reliable, a good registration of prior and low-dose image is required. For this, we employ a state-of-the-art registration method, called SIFT-flow, which can tolerate the high amount of streak noise. But even for properly registered images, using an artifact free prior for the matching yields inferior results. We hence describe a scheme that first constructs a tandem-prior with streak artifacts resembling those in the low-dose image, and then employs this image for the matching, but uses the corresponding high-quality prior to determine the pixel estimates.
Two experimental studies are carried out, using a head phantom and a human lung with projections gathered via simulation. We assess the quality of the processed reconstruction with various metrics: mathematical and perceptual. We find that the quality that can be obtained with the artifact-matched prior-based scheme significantly exceeds that of all competing schemes. Even though the general prior-based approach is able to eliminate the streak artifacts, only the artifact-matched scheme can restore small detail and feature sharpness.
The reduced-projection low-dose image reconstruction algorithm we present outperforms traditional image restoration algorithms when a prior scan is available. Our method is quite efficient and as such it is well suited for fast-paced clinical applications such as image-assisted interventions, orthopedic alignment scans, and follow-ups.
由于对医学扫描中辐射暴露的担忧日益增加，低剂量CT受到了越来越多的关注。然而，X射线辐射的节约使用不可避免地降低了CT图像的质量，引入了诸如噪声和条纹等伪影，这使得重建图像在临床常规中难以解读。对于后续的CT检查，通常可以获得先前的扫描图像。它通常包含相同的解剖结构，只是有些变形且未对齐。这项工作描述了一种两步技术，该技术利用先前的扫描来实现高质量的低剂量CT成像，克服了由噪声伪影和未对齐引起的困难。我们特别关注通过减少投影数量来降低剂量。这会产生严重的条纹伪影，与降低毫安或千伏设置所产生的细粒度噪声相比，可能在更大程度上降低CT图像的可读性。
一种常见的方法是应用图像滤波来减少噪声伪影。这些技术通常利用退化图像中的像素邻域来估计该邻域中心像素的真实值。然而，当图像在非常低的低剂量情况下受到严重污染时，这可能会导致效果不佳。我们提出了一种方法，该方法利用未退化的干净先前图像来确定更高质量的像素统计信息，以形成像素估计值，并由非局部均值滤波器的匹配方案提供支持。为了使这种匹配可靠，需要对先前图像和低剂量图像进行良好的配准。为此，我们采用了一种称为SIFT-flow的先进配准方法，该方法可以容忍大量的条纹噪声。但是，即使对于正确配准的图像，使用无伪影的先前图像进行匹配也会产生较差的结果。因此，我们描述了一种方案，该方案首先构建一个带有类似于低剂量图像中条纹伪影的串联先前图像，然后使用该图像进行匹配，但使用相应的高质量先前图像来确定像素估计值。
使用头部模型和通过模拟收集投影的人体肺部进行了两项实验研究。我们使用各种指标评估处理后的重建质量：数学指标和感知指标。我们发现，基于伪影匹配先前图像的方案所获得的质量显著超过所有竞争方案。即使基于一般先前图像的方法能够消除条纹伪影，但只有伪影匹配方案能够恢复小细节并保持特征清晰度。
当有先前扫描图像时，我们提出的减少投影低剂量图像重建算法优于传统的图像恢复算法。我们的方法非常高效，因此非常适合诸如图像辅助干预、骨科对齐扫描和随访等快节奏的临床应用。
