Radiomic Feature Stability in Low-Dose CT and 3D Segmentation for Pulmonary Nodules: Phantom and Multicenter...

Purpose To identify a stable radiomic feature (RF) subset for pulmonary nodules by analyzing test-retest and acquisition parameter variability using phantoms and to evaluate its clinical utility. Materials and Methods CT scanners at multiple centers were used to scan a chest phantom embedded with artificial pulmonary nodules under varied doses and scanning and reconstruction parameters. An artificial intelligence-based, fully automated three-dimensional segmentation method was used. Three sources of variability were examined: test-retest, inter-CT scanner, and intra-CT scanner. RF stability was assessed using the concordance correlation coefficient (CCC), dynamic range, and intraclass correlation coefficient (ICC). Statistical tests were used to analyze the impact of these factors, with stable RFs identified through UpSet plots. Multiple radiomic models were constructed and validated using multicenter clinical ground-glass nodule data. Results In terms of test-retest variability, repeatable RFs represented 1715 of 2264 features (76%, CCC > 0.9). The reproducibility rate for inter-CT variability was 64% (1439 of 2264 features, ICC > 0.9). Variance component analysis identified section thickness/interval (24.9%), low-dose scanning (24.3%), inter-CT scanner variation (17.5%), and field of view (FOV, 14.9%) as the largest contributors to radiomic feature variability, with consistent influence patterns across all acquisition parameters. A stable subset of 466 RFs was used to construct six radiomics models, which, compared with prescreening RF models, demonstrated greater clinical efficacy and predictive performance on the validation sets (AUC range, 0.86 [95% CI: 0.81, 0.91] to 0.90 [95% CI: 0.86, 0.94)]. Conclusion Variations in section thickness, FOV, interscanner differences, and low-dose settings significantly affected RF stability. A stable RF subset for pulmonary nodules was identified and demonstrated clinical utility. Keywords: Phantom Studies, Applications-CT, Clinical Testing, Machine Learning, Model Training, Model Validation, Radiomics, Pulmonary, Lung, Pulmonary Nodule, Computed Tomography, Repeatability, Reproducibility Supplemental material is available for this article. © The Author(s) 2026. Published by the Radiological Society of North America under a CC BY 4.0 license.