Purpose: Radiomic features have shown the potential to assist in clinical decision making and correlate with lung cancer histology. However, the validity of radiomic based predictive models is impeded by concerns of reproducibility. In this study, we explored the variability of lung nodule radiomic features using a dataset with heterogenous CT acquisition and reconstruction conditions.
Methods: Raw projection data were collected from 25 patients with a nodule detected as part of a clinical low dose CT lung cancer screening program. To create a dataset that represented a wide range of acquisition and reconstruction conditions (and noise levels), the raw data for each patientâ€™s CT exam were used to generate simulated images at different dose levels of 50%, 25%, and 10% of original dose level. Iterative reconstructions were obtained at the scanner with reconstruction settings of (SAFIRE I26F, I44F, I50F) and slice thicknesses of (0.6, 1, 2mm), resulting in total of 20 image datasets for each scan. An in-house CAD tool segmented a volume of interest (VOI) for one nodule for each patient at a reference condition. The VOIs were then mapped to images at other acquisition/reconstruction conditions. For each VOI under each condition, 12 GLCM texture features and 5 intensity based radiomic features were calculated. The reproducibility of radiomic features was evaluated by making comparisons of feature values across 20 different acquisition conditions and by testing the significance of differences due to variation of dose level, and reconstruction settings through one-way and two-way ANOVA analysis.
Results: Significant difference of most radiomic features where found when acquisition/reconstruction conditions changed (all p<0.05). The interaction between dose, reconstruction setting, and slice thickness played a pivotal role in reproducibility of radiomic features.
Conclusion: The reproducibility of radiomic features is affected by image acquisition and reconstruction conditions, which should be addressed to ensure their generalizability.