Multicenter study of quantitative computed tomography analysis using a computer-aided three-dimensional system in patients with idiopathic pulmonary fibrosis

Abstract

Purpose

To evaluate the feasibility of automated quantitative analysis with a three-dimensional (3D) computer-aided system (i.e., Gaussian histogram normalized correlation, GHNC) of computed tomography (CT) images from different scanners.

Materials and methods

Each institution’s review board approved the research protocol. Informed patient consent was not required. The participants in this multicenter prospective study were 80 patients (65 men, 15 women) with idiopathic pulmonary fibrosis. Their mean age was 70.6 years. Computed tomography (CT) images were obtained by four different scanners set at different exposures. We measured the extent of fibrosis using GHNC, and used Pearson’s correlation analysis, Bland–Altman plots, and kappa analysis to directly compare the GHNC results with manual scoring by radiologists. Multiple linear regression analysis was performed to determine the association between the CT data and forced vital capacity (FVC).

Results

For each scanner, the extent of fibrosis as determined by GHNC was significantly correlated with the radiologists’ score. In multivariate analysis, the extent of fibrosis as determined by GHNC was significantly correlated with FVC (p < 0.001). There was no significant difference between the results obtained using different CT scanners.

Conclusion

Gaussian histogram normalized correlation was feasible, irrespective of the type of CT scanner used.

Appendix

The Gaussian histogram normalized correlation (GHNC) method classifies pixels in a target image into different categories. In the GHNC method, a set of Gaussian histograms of CT attenuation values of the original image and the differential image are extracted from samples of each category in advance. The normalized correlations between the Gaussian histograms in the surrounding local 50-pixel area of a pixel and those of the categories are then compared for classification into the predefined categories. To avoid the effects of random noise, pixel attenuation is assumed to have random noise with a Gaussian distribution. Gaussian convolution filtering is applied to the extracted histograms from the local area and the samples in the categories.

The Gaussian histogram of the predesigned sample is obtained from the following formula:

$$\alpha_{\text{g}} (x)\; = \;\frac{1}{{N_{\alpha } }}\sum\limits_{\alpha \in D\alpha } {\frac{1}{{\sqrt { 2\pi \sigma } }}\;\exp } \;\left\{ {\frac{{(x\; - \;\alpha )^{2} }}{{2\sigma^{2} }}} \right\},$$

(1)

in which α is the Hounsfield units of each pixel and N _α is the number of pixels in the predesigned sample area D _α. The variable σ is the standard deviation of the Gaussian random noise.

The Gaussian histogram of the target area D _β is given by

$$\beta_{\text{g}} (x)\; = \;\frac{1}{{N_{\beta } }}\sum\limits_{{\beta \in D_{\beta } }} {\frac{1}{{\sqrt { 2\pi \sigma } }}\;\exp } \;\left\{ {\frac{{(x\; - \;\beta )^{2} }}{{2\sigma^{2} }}} \right\},$$

(2)

in which β and N _β also denote the Hounsfield units and the number of pixels in the target area D _β, respectively.

The normalized correlation between α _g(x) and β _g(x) is given by the formula

$$r(\alpha_{g} ,\beta_{g} ) = \frac{{\int {\alpha_{g} (x)\beta_{g} (x)\;dx} }}{{\sqrt {\int {\alpha_{g} (x)\alpha_{g} (x)\;dx} } \sqrt {\int {\beta_{g} (x)\beta_{g} (x)\;dx} } }}.$$

(3)

In our method, the normalized correlations are calculated in the local area of each target pixel for the original and differential images. The product of both correlations is then used as the similarity of the target and the predesigned sample pattern in each category to classify the target pixels into the predesigned categories.