Assessing the multi-software robustness of radiomic biomarkers: a three-tool evaluation
Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine
Purpose To assess the cross-software reproducibility of Computed Tomography (CT) radiomic features extracted using three widely adopted platforms (Siemens syngo.via Frontier, 3D Slicer/PyRadiomics, and mint Lesion) and to identify a subset of highly robust features suitable for multi-platform and multi-radiomics applications. Methods A retrospective cohort of 97 lesions (primaryectal cancerectal liver metastases, and hepatocellular carcinoma) who underwent contrast-enhanced Computed Tomography (CT) in the portal venous phase wasd. Semi-automatic 3D lesion segmentations were exported for radiomic extraction across the three platforms. Shared radiomic featurestools were harmonized and z-score normalized. Cross-platform similarity was assessed using distribution distance metrics, hierarchical clustering, and the Adjusted Rand Index (ARI). A novel Composite Robustness Index (CI) integrating Pearson correlation, Kolmogorov–Smirnov statistics, and mean fold-difference was developed to quantify feature-level reproducibility. Results First-order intensity features and key GLCM descriptors (e.g., Correlation, Joint Average, Sum Entropy) demonstrated the highest cross-software stability, with nearly superimposable distributions and strong concordance in clustering structure. Siemens syngo.via Frontier and 3D Slicer/PyRadiomics showed the highest agreement (mean ARI0.85)mint Lesion™—which lacks higher-order texture families—showed moderate deviations (mean ARI ≈ 0.70–0.75). High-order f
Abstract
Purpose To assess the cross-software reproducibility of Computed Tomography (CT) radiomic features extracted using three widely adopted platforms (Siemens syngo.via Frontier, 3D Slicer/PyRadiomics, and mint Lesion) and to identify a subset of highly robust features suitable for multi-platform and multi-radiomics applications. Methods A retrospective cohort of 97 lesions (primaryectal cancerectal liver metastases, and hepatocellular carcinoma) who underwent contrast-enhanced Computed Tomography (CT) in the portal venous phase wasd. Semi-automatic 3D lesion segmentations were exported for radiomic extraction across the three platforms. Shared radiomic featurestools were harmonized and z-score normalized. Cross-platform similarity was assessed using distribution distance metrics, hierarchical clustering, and the Adjusted Rand Index (ARI). A novel Composite Robustness Index (CI) integrating Pearson correlation, Kolmogorov–Smirnov statistics, and mean fold-difference was developed to quantify feature-level reproducibility. Results First-order intensity features and key GLCM descriptors (e.g., Correlation, Joint Average, Sum Entropy) demonstrated the highest cross-software stability, with nearly superimposable distributions and strong concordance in clustering structure. Siemens syngo.via Frontier and 3D Slicer/PyRadiomics showed the highest agreement (mean ARI0.85)mint Lesion™—which lacks higher-order texture families—showed moderate deviations (mean ARI ≈ 0.70–0.75). High-order features, particularly GLDM and GLRLM metrics, exhibited substantial variability across platforms. The CI ranking enabled identification of a reproducible set ofhighly reproducible featuresincluding glcm_Correlation, firstorder_Mean, firstorder_RMS, firstorder_90Percentile, and shape axis-length descriptors. Conclusion Despite intrinsic software differences, a consistent subset of radiomic features remains reproducible across heterogeneous extraction tools. The combined use of distribution analysis, hierarchical clustering, and the Composite Robustness Index offers a rigorous framework for evaluating cross-platform reliability. These findings support the feasibility of multi-tool radiomics and provide a validated feature set for harmonized quantitative imaging pipelines.