Reliability of CT radiomic features reflecting tumour heterogeneity according to image quality and image processing parameters

Abstract

The reliability of radiomics features (RFs) is crucial for quantifying tumour heterogeneity. We assessed the influence of imaging, segmentation, and processing conditions (quantization range, bin number, signal-to-noise ratio [SNR], and unintended outliers) on RF measurement. Low SNR and unintended outliers increased the standard deviation and mean values of histograms to calculate the first-order RFs. Variations in imaging processing conditions significantly altered the shape of the probability distribution (centre of distribution, extent of dispersion, and segmentation of probability clusters) in second-order RF matrices (i.e. grey-level co-occurrence and grey-level run length), thereby eventually causing fluctuations in RF estimation. Inconsistent imaging and processing conditions decreased the number of reliably measured RFs in terms of individual RF values (intraclass correlation coefficient ≥0.75) and inter-lesion RF ratios (coefficient of variation <15%). No RF could be reliably estimated under inconsistent SNR and inclusion of outlier conditions. By contrast, with high SNR and no outliers, all first-order RFs, 11 (42%) grey-level co-occurrence RFs and five (42%) grey-level run length RFs showed acceptable reliability. Our study suggests that optimization of SNR, exclusion of outliers, and application of relevant quantization range and bin number should be performed to ensure the robustness of radiomics studies assessing tumor heterogeneity.

Conflict of interest statement

The authors declare no competing interests.

Figures

Figure 1

Study processes for lesion selection, image matrix generation, and reliability tests. A radiologist identified three representative patterns of tumour necrosis on the CT images of patients who treated with metastatic urothelial carcinoma. The original CT values in the left upper-boundary of the image matrices were replaced by 2 × 2 voxel clusters of outlying grey value, to evaluate the effect of unintended outlier in the region-of-interest (black squares indicating the location of outliers). By adding noise levels, three image matrices were generated (high, intermediate, and low signal-to-noise level). Consequently, a total of 27 image matrices were generated. These imaging data were quantized by applying various quantization ranges, and bin numbers. Accordingly, reliability for radiomics feature values and inter-lesion ratios were evaluated.

Figure 2

Alterations in the histogram according to SNR and outlier conditions. Histograms from the original CT image (high SNR, (a), low-SNR CT image (b), outlier-containing high-SNR CT image (C), and outlier-containing low-SNR CT image (d). Low SNR and unintended outliers (red arrows in (c) and (d)) increased the SD and range of the histogram. The mean value was increased by the extremely high CT values of bone-equivalent outliers.

Figure 3

Alterations in the GLCM according to image and image processing conditions. (a) The GLCM from the original CT image (high SNR) shows a diagonal probability distribution. The distribution was most concentrated with a distribution of the mean ± 3 SD, and then min‒max, followed by the mean ± 2 SD. Low SNR dispersed the probability distribution. The addition of outliers concentrated the distribution, shifting it towards the upper left direction in the min‒max quantization range. (b) Increasing the bin number segmented the probability distribution while maintaining the overall shape.

Figure 4

Alterations in the GLRLM according to image and image processing conditions. (a) The probability in the GLRLM was distributed mostly in the upper rows as the run length of the original CT image was 3 or less. The distribution was most concentrated with mean ± 3 SD, followed by min-max, and mean ± 2 SD. Low SNR dispersed the probability distribution. The addition of outliers concentrated the distribution and shifted it to the left with the min‒max quantization range. (b) Increasing the bin number segmented the probability distribution while maintaining the overall shape.

Figure 5

Heatmap showing the ICCs of GLCMs and GLRLMs with image and image processing conditions. The left-most column represents the option being varied, while the second to fourth columns indicates the other options being fixed. The green to yellow colour shows the acceptable range of ICCs (i.e., equal to or greater than 0.75). Abbreviations: IMC = information measure of correlation, IDM = inverse difference moment, IDMN = inverse difference moment normalized, IDN = inverse difference normalized, GLNU = Grey-Level Nonuniformity, HGRE = High Grey-Level Run Emphasis, LRE = Long Run Emphasis, LRLGE = Long Run Low Grey-Level Emphasis, LRHGE = Long Run High Grey-Level Emphasis, LRGE = Low Grey-Level Run Emphasis, RLNU = Run Length Nonuniformity, SRE = Short Run Emphasis, SRHGE = Short Run High Grey-Level Emphasis, SRLGE = Short Run Low Grey-Level Emphasis.

Figure 6

Heatmap presenting the CVs of inter-lesion ratios in the GLCMs and GLRLMs with varying image and image processing conditions. The left-most column represents the option being varied, while the second to fourth columns indicate the other options being fixed. The yellow to green colour shows the acceptable range of CV (i.e. equal to or less than 15%). Abbreviations: IMC = information measure of correlation, IDM = inverse difference moment, IDMN = inverse difference moment normalized, IDN = inverse difference normalized, GLNU = Grey-Level Nonuniformity, HGRE = High Grey-Level Run Emphasis, LRE = Long Run Emphasis, LRLGE = Long Run Low Grey-Level Emphasis, LRHGE = Long Run High Grey-Level Emphasis, LRGE = Low Grey-Level Run Emphasis, RLNU = Run Length Nonuniformity, SRE = Short Run Emphasis, SRHGE = Short Run High Grey-Level Emphasis, SRLGE = Short Run Low Grey-Level Emphasis.