Purpose: The uniqueness of radiomic features, combined with their reproducibility, determines the success of radiomic studies. In general, radiomic features extracted from a defined region of interest (ROI) is hypothesized unique to the ROI. This study is to investigate if the hypothesis is valid.
Methods: Two cohorts of NSCLC patients were retrospectively retrieved from GE and Siemens CT scanners. The lung nodule (defined region) was delineated manually and radiomic features were extracted using IBEX. The same ROI was then translated randomly to four other tissue regions of the same image set: adipose, heart, lung beyond nodule, and muscle. The same features were extracted from the four tissue regions for comparison. Coefficient of Variation (CV) was calculated to test variation of features within different ROIs. The Concordance Correlation Coefficient (CCC) between radiomic features of the defined region (lung nodule) and a given tissue region were calculated to determine feature uniqueness. Identifying features as non-unique follows this process: 1) A feature is flagged within a tissue region when the CCC>0.85 for ≥50% of patients; then 2) a feature is identified as non-unique when ≥2 tissue regions are flagged for that feature.
Results: We analyze 14 patients from GE and 18 patients from Siemens. Of 123 features, 18 radiomic features in GE and 16 features in Siemens are identified as non-unique, with 11 features overlapping. The overlap features are features included in the GradientOrientHistogram, and InverseDiffMomentNorm and InverseDiffNorm within the 3D GrayLevelCooccurrence (GLC). The remaining non-unique features for GE are SumAverage from the 3DGLC and features in the Intensity Histogram Gauss Fit. While for Siemens the features are Energy and Entropy for both 2D/3DGLC and MaxProbability for 3DGLC.
Conclusion: Approximately 10% or more radiomic features are not unique to a defined region. Radiomic feature uniqueness needs to be considered in future radiomics studies.