IBSI-compliant feature extraction standard for radiomics

Answer

The standard is the Image Biomarker Standardisation Initiative (IBSI) consensus standard: its Reference Manual defines how radiomic texture features are computed, and the Radiology 2020 article by Zwanenburg and colleagues reports the multi-team standardisation behind it 868aa2f9 8a6a63cb. For texture, the manual standardises 25 GLCM features and 16 GLRLM features, alongside GLSZM, GLDZM, NGTDM and NGLDM families, and fixes the preprocessing that precedes them 8a6a63cb. The article standardised 174 features, 169 of them in the first two phases, and validated them on CT and T1-weighted MRI from 51 patients with soft-tissue sarcoma 868aa2f9.

Summary

Use IBSI-defined feature formulas and reference values for GLCM and GLRLM, and treat the preprocessing as part of the standard: interpolation, re-segmentation, intensity discretisation (fixed bin number or bin width), grid distance norms and feature aggregation 8a6a63cb. The standard does not fix one universal bin width or interpolation method for every CT or MRI study; those are configuration settings that must be matched to the relevant IBSI reference configuration or reported 868aa2f9. Features without enough consensus (fewer than 3 matching teams, or not reproduced by an absolute majority) were not standardised and should not be used 8a6a63cb. The two sources agree; there is no disagreement to resolve.

Evidence

• Zwanenburg et al. 2020 (Radiology) - IBSI consensus study; 174 features, 169 standardised in the first two phases; families include GLCM and GLRLM; pipeline of ROI interpolation, masks, optional re-segmentation and discretisation; validated on CT and T1-weighted MRI of 51 soft-tissue sarcoma patients. Does not support a universal bin width or interpolation method. 868aa2f9
• IBSI Reference Manual (arXiv 1612.07003, v11, 2019) - GLCM 25 features, GLRLM 16, plus GLSZM, GLDZM, NGTDM and NGLDM; general image-processing workflow, reporting guidelines, permanent four-character feature identifiers, and reference data sets (a digital phantom and a lung cancer CT image) with reference values. 8a6a63cb

Methodology

Both verified facts name the same standard, one through the consensus article and one through the manual, so I treated them as two views of one standard and did not count them as independent replications. I did not sum or compare feature counts across the two sources because they classify families differently.

Limitations

• Only two records, both from the IBSI authors; no independent audit of software compliance is in the corpus.
• Neither record gives a reference configuration for a specific CT or MRI protocol, so bin width and interpolation choices are left to the user 868aa2f9.
• Clinical performance of the standardised features for any tumour type is not addressed.