Scanner-Induced Variability in Multicenter PET Radiomics: Comparative Evaluation of Interpolation Methods and ComBat Harmonization

This study aimed to systematically evaluate scanner-induced variability in radiomic features extracted from multicenter [18F]-FDG PET scanners (Siemens, Philips, and GE) and to determine which interpolation method, when combined with ComBat harmonization, most effectively reduces feature variability across scanners. Pre-treated [¹⁸F]-FDG PET scans from 167 stage IIB/III NSCLC patients were obtained from The Cancer Imaging Archive (ACRIN 6668/RTOG 0235 trial). Primary tumors were delineated semi-automatically. The images and masks were resampled into isotropic voxel sizes of 0.5 × 0.5 × 0.5 mm³ using three interpolation methods, namely B-spline, Gaussian, and Nearest Neighbor. A total of 105 radiomic features were extracted. ComBat harmonization was applied to correct for batch effects between scanners. Statistical analysis included the Kruskal–Wallis test, effect size ε², and coefficient of variation (CV) to evaluate variability between scanners before and after ComBat harmonization. ComBat harmonization consistently reduced the variability of radiomic features that emerge from scanner differences. After ComBat harmonization, the Nearest Neighbor interpolation method demonstrated the best performance compared with the B-spline and Gaussian methods. Only 1 out of 105 radiomic features (~0.95%) remained a p-value < 0.05, while approximately 95 of 105 features (90.5%) had CV < 10%. The Nearest Neighbor method also produced the lowest average CV value compared to the B-spline and Gaussian. Radiomic features extracted from different types of scanners can increase radiomic feature variability. The use of Nearest Neighbor interpolation with ComBat harmonization is more effective in reducing radiomic feature variability between different scanners.