Validation of Voxel-Size Normalization of CT Radiomic Features in Lung Cancer Patients

Abstract

Radiomic features are potential candidates for therapy response assessment in radiation oncology. However, the robustness of these features with imaging parameters is not well established. For example, some radiomic features related to intensity, GLRLM and NGTDM were suggested as potential biomarkers in recent studies. However, all these features were found to be intrinsically dependent on voxel size as indicated by a recent texture phantom study. Normalization by voxel size is a potential method to remove these voxel-size dependencies. The purpose of this study was to validate the voxel size normalization in a patient cohort. Eight patients with non-small lung cancer (NSCLC) of varying tumor volumes were analyzed. The segmented volume of interest (VOI) along with original CT scan of each patient was down and up-sampled to various voxel sizes using linear interpolation. For each patient, VOI was resampled to 5 different pixel sizes from 0.4 to 1.0 mm and 5 different slice thicknesses from 1 to 4 mm. There was a total of 88 CT data sets used in this study. Ten features including entropy, energy, TGV, contrast from intensity histogram; mean and inverse variance from GLCM; gray level non-uniformity (GLNU) and run length non-uniformity (RLNU) from GLRLM; and coarseness and texture strength from NGTDM were extracted from 3D VOIs using an in-house program. To investigate the usefulness of voxel-size normalization, each feature was modified to include voxel size or number of voxels in its definition. The absolute value of the spearman rank correlation coefficient (rs) was calculated for each feature to determine which features were correlated with voxel size or number of voxels in VOI before and after normalization. The features having values rs > 0.9, 0.5 < rs < 0.9 and rs< 0.5 were respectively categorized as high, moderate and no correlations with voxel size. Without voxel size normalization, eight out of 10 features: Intensity-energy, Intensity-entropy, Intensity-TGV, GLCM-mean, GLCM-inverse variance, GLRLM-GLNU, GLRLM-RLNU and NGTDM-coarseness were highly correlated with voxel size (rs > 0.9). Intensity based contrast and NGTDM based texture strength were moderately correlated with voxel size (0.5 < rs < 0.9) before normalization. After voxel size normalization, all ten features showed low values of spearman rank correlation coefficient (rs< 0.5), indicating intrinsic dependencies of these features on voxel size were reduced or eliminated. We conclude that voxel size normalization results from a texture phantom study also apply to lung patients cohort. This study highlights the importance of investigating the robustness of radiomic features with CT imaging parameters and exploring methods to fix this variability before using them for prognostic and predictive studies.