Reliability and prognostic value of radiomic features are highly dependent on choice of feature extraction platform

• DOI: 10.1007/s00330-020-06957-9
• Publication Date: 2020-06-01T08:02:26Z
• Authors: Isabella Fornacon...; Hitesh Mistry; Christoph J. Acke...; Fiona Blackhall; Andrew McPartlin; Corinne Faivre-Finn; Gareth J. Price; James P. B. O’Connor
• Keywords: Feature (linguistics)

Abstract Objective To investigate the effects of Image Biomarker Standardisation Initiative (IBSI) compliance, harmonisation calculation settings and platform version on statistical reliability radiomic features their corresponding ability to predict clinical outcome. Methods The was assessed retrospectively in three datasets (patient numbers: 108 head neck cancer, 37 small-cell lung 47 non-small-cell cancer). Features were calculated using four platforms (PyRadiomics, LIFEx, CERR IBEX). PyRadiomics, LIFEx are IBSI-compliant, whereas IBEX is not. IBSI user-defined by calculating intraclass correlation coefficients confidence intervals. influence choice relationship between biomarkers survival evaluated univariable cox regression largest dataset. Results different software only excellent (ICC > 0.9) for 4/17 when comparing all platforms. Reliability improved ICC 0.9 15/17 analysis restricted IBSI-compliant Failure harmonise resulted poor reliability, even across Software also had a marked effect feature LIFEx. identified as having significant varied platforms, did direction hazard ratios. Conclusion performance prognostic models radiomics. Key Points • varies with version. compliance improves but harmonised. collectively affect value features.