Room: AAPM ePoster Library
Purpose: Lack of rotational invariance in quantitative evaluation of image textural features means that the feature value will be sensitive to the angular sampling technique employed in computing the feature. In this study, we investigated the impact of image rotation on the quantitative values of 95 textural features from 6 different radiomic feature categories, using in-house developed software and a 3D digital phantom.
Methods: A 3D computational phantom was developed with the following specifications: 4 axial slices with 32×32 (2×2×2mm3) voxels, 4 corner-to-corner intensity gradients in 2 orthogonal directions with voxel intensities ranging from 1-to-64. The gradient phantom was rotated 90 times about the baseline (1-90 degrees, using nearest-neighbor interpolation technique) using an IBSI-validated, in-house-developed software (ROdiomX), and 95 textural features from the following 6 textural feature categories (3D with 27 directions) were computed: Gray-Level-Co-occurrence (GLCM: 25-features), Gray-Level-Run-Length (GLRLM: 16-features), Gray-Level-Size-Zone (GLSZM: 16-features), Gray-Level-Distance-Zone (GLDZM: 16-features), Neighborhood-Grey-Tone-Difference (NGTDM: 5-features), and Neighboring-Grey-Level-Dependence (NGLDM: 17-features). Coefficient of variance (CV), intraclass correlation coefficient (ICC, two-way random-average-score), and mean absolute deviation were calculated to measure dispersion of each feature around the baseline value.
Results: ICC values for all feature categories were poor (ICC<0.40). Coefficient-of-variation values were: 1.75 (GLDZM), 2.23 (NGTDM), 2.96 (NGLDM), 3.51 (GLRLM), 4.03 (GLSZM), and (6.02) GLCM suggesting that GLDZM was least impacted and GLCM was most impacted by rotation. None of the features were found to be rotationally invariant given that CV were different than 1.0. Results imply that GLSZM, NGTDM, GLSZM, GLDZM, GLCM, NGLDM, and GLRLM features are affected more similarly against rotations respectively.
Conclusion: study investigated the sensitivity of 95 textural features from 6 different feature categories against rotation. Lack of rotational-invariance suggests that specific protocols for angular sampling in computation of these features is necessary to reduce variation among different algorithms used to compute the same features.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by a grant from Varian Medical Systems (Palo Alto, CA).
Texture Analysis, Feature Extraction, Feature Selection