Association between image acquisition parameters and MIP
We took images from the machine type with the largest sample size (GE machine type 1) and fitted a series of linear regression models with MIP as an outcome variable and acquisition parameters as covariates (See Table S1). As a whole, the acquisition parameters were strongly associated with MIP (p <1.0×10-16) after adjusting for VPD. The regression model with all nine acquisition parameters had a residual deviance of 896.32 and an R2 value of 0.299. Including one acquisition parameter at a time, no model had a residual deviance close to 896.32. We then tried all possible combinations of pairs of acquisition parameters. We found that a model including Exposure Time and Body Part Thickness as covariates had almost as good fit as that including all nine acquisition parameters - the difference in goodness of fit was only just statistically significant (p=0.0134; See Table S1). When Body Part Thickness and Exposure Time were included one-at-a-time, the models provided significantly worse explanations of MIP (See Table S1). Residual deviances for the next eight best fitting models with two acquisition parameters are also listed in Table S1. In combination, Body part thickness, with one other acquisition parameter (Exposure Time, Relative X-ray Exposure or Organ Dose) were best able to explain MIP measurements, although R2 values were not high (approximately 0.3). We carried out similar analyses for the other machine types. For all machines with more than 350 images, the nine acquisition parameters were significantly associated as a whole with MIP (data not shown). R2 levels were similar across all machines, based on models with all nine acquisition parameters (i.e., approximately 0.30). The acquisition parameters that best accounted for the associations differed, though, across machines. For example, for Philips machine type 8, Exposure Time and Exposure in uAs together obtained a similar goodness of fit to a model based on all nine acquisition parameters (Body Part Thickness was not important for explaining MIP’s association with the acquisition parameters for this machine type).
Table S1. Residual deviance tests to select the most relevant acquisition parameters combination that can explain MIP. The residual deviances correspond to fitted linear regression models with MIP as an outcome variable and the acquisition parameters as covariates, adjusting for VPD.
Covariates / Residual deviance / Degrees of freedom / P-value (**) / Direction of the association / Correlation between A&B(A) / (B)
All nine acquisition parameters included / 896.32 / N/A / N/A / N/A / N/A / N/A
No acquisition parameters included / 1072.19 / 9 / <1.0×10-16 / N/A / N/A / N/A
The nine best fitting models with two acquisition parameters:
(A)Body Part Thickness, (B)Exposure Time / 914.01 / 7 / 0.0134 / + / - / 0.707
(A)Body Part Thickness , (B)Relative X-ray Exposure / 914.57 / 7 / 0.0109 / + / - / 0.758
(A)Body Part Thickness, (B)Organ Dose / 915.25 / 7 / 0.0084 / + / - / 0.744
(A)Body Part Thickness, (B)Exposure / 924.43 / 7 / 2.1×10-4 / + / - / 0.568
(A)Body Part Thickness, (B)Exposure in uAs / 924.72 / 7 / 1.8×10-4 / + / - / 0.567
(A)Body Part Thickness, (B)Compression Force / 981.81 / 7 / 1.0×10-16 / + / - / 0.146
(A)Body Part Thickness, (B)kVp / 989.53 / 7 / <1.0×10-16 / + / - / 0.790
(A)Body Part Thickness, (B)X-ray Tube Current / 989.61 / 7 / <1.0×10-16 / + / + / - 0.619
(A)Exposure, (B)Relative X-ray Exposure / 999.69 / 7 / <1.0×10-16 / - / + / 0.952
Two models with one acquisition parameter:
(A)Body Part Thickness / 989.69 / 8 / 1.1×10-16 / + / N/A / N/A
(A)Exposure Time / 1071.05 / 8 / <1.0×10-16 / + / N/A / N/A
(*) All models were adjusted for VPD. (**) Likelihood Ratio test p-values (testing difference in fit from the model with all nine acquisition parameters).
1