A study on the effect of CT imaging acquisition parameters on lung nodule image interpretation

Abstract

Most Computer-Aided Diagnosis (CAD) research studies are performed using a single type of Computer Tomography (CT) scanner and therefore, do not take into account the effect of differences in the imaging acquisition scanner parameters. In this paper, we present a study on the effect of the CT parameters on the low-level image features automatically extracted from CT images for lung nodule interpretation. The study is an extension of our previous study where we showed that image features can be used to predict semantic characteristics of lung nodules such as margin, lobulation, spiculation, and texture. Using the Lung Image Data Consortium (LIDC) dataset, we propose to integrate the imaging acquisition parameters with the low-level image features to generate classification models for the nodules' semantic characteristics. Our preliminary results identify seven CT parameters (convolution kernel, reconstruction diameter, exposure, nodule location along the z-axis, distance source to patient, slice thickness, and kVp) as influential in producing classification rules for the LIDC semantic characteristics. Further post-processing analysis, which included running box plots and binning of values, identified four CT parameters: distance source to patient, kVp, nodule location, and rescale intercept. The identification of these parameters will create the premises to normalize the image features across different scanners and, in the long run, generate automatic rules for lung nodules interpretation independently of the CT scanner types.

A statistical analysis of the effects of CT acquisition parameters on low-level features extracted from CT images of the lung

“…When the kernels from these two groups were compared, there was no significant change in the features. This is additional evidence for what was already concluded in our previous research [12], mainly that these groups are too broad in scope to adequately distinguish the first group from the second group. The broadness stems from the fact that manufacturers each have proprietary convolution kernel algorithms.…”

“…It is the only CT parameter to do so, and this makes sense because the convolution kernel affects how the image is reconstructed (Figure 1). This shows the importance of this particular CT acquisition parameter, and this importance is reinforced by our decision tree classifier results [12] in which this CT acquisition parameter appeared the most (6 times), including multiple appearances within one decision tree. When the convolution kernel FC01 was used on images, all 6 of the Intensity features under the 0.0001 p-value threshold were significantly different from when the other convolution kernels were used to reconstruct images ( Figure 2).…”

“…All 7 of the Gabor features are affected by a kVp value of 140 in a different way than other kVp values. Finally, elongation is shown be unaffected by kVp (although affected by kVp in our previous [12]). …”

“…Subsequent research used the same decision tree classifiers (DT's) but this time added the CT acquisition parameters to the previously calculated image features [12]. The results have shown that CT acquisition parameters do appear within those DT's and thus affect the classifier's results.…”

“…However the decision trees do not allow us to understand how to normalize those features in order to preprocess the images so that the classifiers are independent of the CT acquisition parameters. The decision trees approach did show that there are 7 CT acquisition parameters that affect semantic mappings: slice thickness, exposure, kVp, reconstruction diameter, convolution kernel, Z nodule location, and distance source to patient [12]. This research looks at specifically how the image features are affected by these 7 CT parameters through the use of ANOVA.…”