Abstract
1752
Introduction: Radiomics is a process of extracting a large number of quantitative imaging features (QIF) capturing textural properties within the volume of interest (VOI) [1, 2]. In order to examine potential effects of interrogated volume on the robustness of QIFs in PET imaging, we have compared radiomic features extracted from concentric spherical VOIs with different diameters that were segmented on a homogenous 18-FDG PET phantom.
Methods: The Homogenous 18-FDG PET Phantom was imaged on an integrated PET/MR system (Biograph mMR; Siemens Healthineers). Six Spherical concentric VOIs with varying diameters: 2,3,5,7,9, and 12 cm were delineated, all centered at the center of the scanner FOV. In total 801 textural and intensity features were extracted using Pyradiomics software (http://pyradiomics.readthedocs.io) for each region of interest from the original image and after decomposition of the image using all combinations of low and high pass filters in 3 dimensions [3] Coefficient of variance (CV) was calculated for each individual feature. The difference in robustness of each 5 following textual families: Gray Level Co-occurrence Matrix (GLCM, n=22), Gray Level Size Zone Matrix (GLSZM, n=16), Gray Level Run Length Matrix (GLRLM, n=16), Neighbouring Gray Tone Difference Matrix (NGTDM, n=5) Gray Level Dependence Matrix (GLDM, n=14) extracted from original image was assessed using Kruskal-Wallis H test. The robustness of features before and after applying filters and amongst different filter combinations were compared using Wilcoxon Rank-Sum test, using a Bonferroni adjusted P value threshold of 0.01 (0.05/5)
Results:
Kruskal-Wallis H-test showed that robustness of textural families and intensity radiomics extracted from the original image are significantly different when VOI size changes (p-value = 6.3 ×10-5). QIFs are more robust when extracted from the original image in comparison to wavelet decomposed images (p-value = 0.0009). High-Pass and Low-Pass filter decomposition were only significantly different in terms of QIFs robustness in the x-dimension ( p-value = 0.0054).
Conclusions: VOI size has substantial impact on the measurement of many radiomic features. QIFs with low CV can potentially be utilized to assess tumor microenvironment over time as we can be more confident the changes in these features are not resulting from the changes in VOI size. Hence, NGTDM and GLDM families and high pass filter decomposition in the x-dimension may have limited utility in longitudinal studies assessing QIFs due to their size-dependent instability