Assessing the multi-software robustness of radiomic biomarkers: a three-tool evaluation

Area of Science:

• Radiomics and Quantitative Imaging
• Medical Physics
• Oncology Imaging

Background:

• Radiomics enables quantitative analysis of medical images.
• Software variability can impact radiomic feature reproducibility.
• Standardization is crucial for multi-center radiomics studies.

Purpose of the Study:

• Assess cross-software reproducibility of radiomic features from Computed Tomography (CT).
• Evaluate three common radiomic platforms: Siemens syngo.via Frontier, 3D Slicer/PyRadiomics, and mint Lesion.
• Identify a robust feature subset for multi-platform and multi-center applications.

Main Methods:

• Retrospective analysis of 97 lesions from contrast-enhanced CT scans.
• Semi-automatic segmentation and radiomic feature extraction across three platforms.
• Harmonization, normalization, and cross-platform similarity assessment using distribution metrics, clustering, and Adjusted Rand Index (ARI).
• Development of a Composite Robustness Index (CI) for feature reproducibility quantification.

Main Results:

• First-order intensity and key Gray-Level Co-occurrence Matrix (GLCM) features showed high stability.
• Siemens syngo.via Frontier and 3D Slicer/PyRadiomics demonstrated strong agreement (mean ARI > 0.85).
• Mint Lesion showed moderate agreement (mean ARI ≈ 0.70-0.75), lacking higher-order textures.
• High-order features (GLDM, GLRLM) exhibited significant variability.
• A reproducible feature set was identified using the CI, including glcm_Correlation, firstorder_Mean, and shape descriptors.

Conclusions:

• A consistent subset of radiomic features is reproducible across different software tools.
• Distribution analysis, clustering, and the CI provide a framework for evaluating cross-platform reliability.
• Findings support multi-tool radiomics and offer a validated feature set for harmonized quantitative imaging.