Hunting for signal in the noise: identifying robust CT radiomic features across variable scanning settings

Objectives: Radiomics offers the potential to derive quantitative biomarkers from medical images, but its clinical implementation remains limited by poor reproducibility and uncertainty about whether features capture true biological signals or merely reflect noise. This study systematically evaluated the repeatability of CT radiomic features from a biological phantom to identify features that are both reproducible and informative. Materials and methods: A bovine phantom was analyzed to assess the reproducibility of radiomic features across different acquisition and reconstruction settings. Radiomic features were extracted and classified as “Reliable,” “Variable,” or “Unreliable” based on their repeatability and comparison with air. The results of this analysis were compared with those from a similar experiment performed at another institution to evaluate inter-site consistency. Finally, cross-referenced results from the phantom analyses were validated in a clinical scenario. Results: Across different slice thicknesses, the number of “Reliable” features was higher in thinner-slice datasets and further improved after voxel resampling, while thicker slices showed inconsistent effects. The comparison between the two phantom experiments demonstrated moderate agreement, highlighting variability across acquisition settings but also identifying a consistent subset of robust features. In the clinical test, 83.3% of the “Reliable” features overlapped with those identified in the phantom experiments, confirming the reproducibility of the selected features in a real-world scenario. Conclusion: Combining repeatability analysis with the noise sensitivity test offers a practical approach to identify robust radiomic features, improving feature selection and model reliability. Further validation is needed to refine thresholds and confirm clinical applicability. Key Points: Question Radiomic features may capture structured noise rather than true biological signal. Identifying which features are genuinely informative remains a major unmet need in radiomics research. Findings Combining repeatability analysis with a noise sensitivity test allowed systematic identification of robust and informative CT radiomic features across multiple acquisition and reconstruction settings. Clinical relevance By distinguishing reproducible and biologically meaningful radiomic features from noise-driven ones, this approach can improve the reliability and generalizability of radiomics-based models, supporting more accurate and trustworthy imaging biomarkers for clinical decision-making and personalized patient care.