The mutuality and robustness of radiomics features in [¹⁸F]FDG PET/CT lung cancer studies

Objectives To assess the relationships between and robustness of advanced quantitative tracer uptake imaging features (i.e. radiomics) in PET/CT lung cancer.

Methods Ten lung cancer patients received two baseline whole-body [¹⁸F]FDG scans acquired on a time-of-flight PET/CT scanner (Philips Healthcare, Cleveland, OH). All scans were reconstructed using reconstruction settings that complied with the EANM guidelines for tumor PET imaging. Volumes of interest (VOI, n=18) were delineated on lesions larger than 10 mL using a background-corrected 50% isocontour method. For each VOI, 101 radiomics features were determined (22 first order gray-level statistics, 9 geometric and 70 textural features). Absolute test-retest variability (TRT) and intraclass correlation coefficients (ICC) between test and retest data were determined. A principal component analysis (PCA) was then performed on pooled test and retest data to determine unique radiomics features that do not strongly correlate.

Results About half (46) of the radiomics features demonstrated poor TRT (>15%) and ICC (<0.9). After excluding these features, initial PCA revealed that 36 features showed a strong correlation with one or more other features (r>0.95). In total 28 sets were identified, where nine sets had more than one member. Out of these 9 sets, only those features were retained that provided best TRT and/or ICC. The following PCA identified 5 components (Kaiser-Meyer-Olkin’s measure of sampling adequacy: 0.682, Bartlett’s test of sphericity: p<0.001). The five features that correlated well with these components and had minimal mutual correlation were sphericity and 4 gray-level co-occurrence matrix-based features.

Conclusions Only 5 out of 101 radiomics features have minimal mutual correlation (i.e. they substantially reflect different tumor tracer uptake characteristics) and are recommended for further assessment of their prognostic value in PET/CT lung cancer studies.

Research Support Philips Healthcare