Harmonization of chest CT scans for different doses and reconstruction methods

Abstract

Purpose: To develop and validate a computed tomography (CT) harmonization technique by combining noise-stabilization and autocalibration methodologies to provide reliable densitometry measurements in heterogeneous acquisition protocols.

Methods: We propose to reduce the effects of spatially variant noise such as nonuniform patterns of noise and biases. The method combines the statistical characterization of the signal-to-noise relationship in the CT image intensities, which allows us to estimate both the signal and spatially variant variance of noise, with an autocalibration technique that reduces the nonuniform biases caused by noise and reconstruction techniques. The method is firstly validated with anthropomorphic synthetic images that simulate CT acquisitions with variable scanning parameters: different dosage, nonhomogeneous variance of noise, and various reconstruction methods. We finally evaluate these effects and the ability of our method to provide consistent densitometric measurements in a cohort of clinical chest CT scans from two vendors (Siemens, n = 54 subjects; and GE, n = 50 subjects) acquired with several reconstruction algorithms (filtered back-projection and iterative reconstructions) with high-dose and low-dose protocols.

Results: The harmonization reduces the effect of nonhomogeneous noise without compromising the resolution of the images (25% RMSE reduction in both clinical datasets). An analysis through hierarchical linear models showed that the average biases induced by differences in dosage and reconstruction methods are also reduced up to 74.20%, enabling comparable results between high-dose and low-dose reconstructions. We also assessed the statistical similarity between acquisitions obtaining increases of up to 30% points and showing that the low-dose vs high-dose comparisons of harmonized data obtain similar and even higher similarity than the observed for high-dose vs high-dose comparisons of nonharmonized data.

Conclusion: The proposed harmonization technique allows to compare measures of low-dose with high-dose acquisitions without using a specific reconstruction as a reference. Since the harmonization does not require a precalibration with a phantom, it can be applied to retrospective studies. This approach might be suitable for multicenter trials for which a reference reconstruction is not feasible or hard to define due to differences in vendors, models, and reconstruction techniques.

Keywords: COPD; CT scanner; Hounsfield Unit correction; calibration; lung density; quantitative imaging.

Figure 1

Scheme of the proposed harmonization method. (a) The image is statistically characterized by a mixture model that provides the probability of belonging to each tissue class (posterior probabilities). (b) The conditioned local statistical moments are calculated through the local characterization per tissue by using the posterior probabilities. Then, they can be aggregated to estimate the moments. The signal and the spatial variance are the first and second order moments, respectively. (c) The functional relationship between signal and variance, $b(x,{\widehat{\sigma }}^2)$, is estimated and the spatially variant bias removed. Then, the systematic bias is corrected considering two anatomical structures (trachea: −1000 HU, and descending aorta: 50 HU), resulting in the harmonized signal estimate: $\widehat{X}(\mathbf{r})$. (d) The harmonized signal can be combined with the stabilized residual in order to preserve any details in the structure that might be codified within the noise. The detail, D(r), can be added with an average standard deviation, $\overline{D}$, set as a parameter. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 2

Example of correction of functional dependence between density and noise variance in the trachea air for a high‐dose and soft kernel reconstruction. (a) Coronal view of the minimum intensity projection to perceive the trachea (in red). (b) Regression of the computed tomography numbers and the spatial variance of noise in the trachea. (c) The harmonization corrects the functional dependence and reduces the bias in air. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 3

(a) Synthetic image for two different visualization windows (left: lung parenchyma [−1024,−700] HU; right: soft tissue [−1024,200] HU). (b) Normalized nonstationary standard deviation of noise. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 4

Probability density function of a nc‐Γ distribution known signal value set in the mode and increasing noise. Note that the most likely tissue density relies in the mode of the distribution, whereas the mean (represented as dots) depends on the variance of noise. This interrelationship will generate a bias (difference between mean and mode) that depends on the reconstruction kernel and the dose. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 5

Visual representation of harmonization of synthetic images reconstructed for a prototypical configuration for high‐dose acquisitions with soft kernels (κ = 0.4 and $\overline{s}=40$ HU). The image is divided in two regions where different visualization windows are applied to improve the visualization of the effect of noise and harmonization. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 6

Root mean square error for different reconstruction kernels and average standard deviation of noise. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 7

Analysis of resolution for the width required to raise the edge response from 10% to 90% of reference levels for air (−1000 HU) and surrounding fat (−90 HU). [Color figure can be viewed at wileyonlinelibrary.com]

Figure 8

Concordance correlation coefficient (${\rho }_{\mathrm{ccc}}$) for a pair of images with different configurations of dose ($\overline{s}$) and kernels (k). Note that ${\rho }_{\mathrm{ccc}}$ for the noisy data shows important discrepancies as the noise differences of images increase. The harmonized data shows a more homogeneous concordance across doses and kernels. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 9

Examples of harmonization for each dataset and detailed region with emphysema in the lung window [−1024,−700] HU. Note that the effect of noise reduces the visibility of anatomical structures in the sharp HD reconstruction for the reference images, and the reduction of the contrast in the low‐dose reconstructions as a result of the noise and reconstruction method. The harmonization reduces the noise, improves the visibility of anatomical structures and increases the contrast in the low‐dose reconstructions to the same levels observed in the high‐dose reconstructions. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 10

Details and significance of the hierarchical linear model for the Siemens dataset. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 11

Details and significance of the hierarchical linear model for the GE dataset. [Color figure can be viewed at wileyonlinelibrary.com]