Imaging Biomarker Standardisation (IBSI)

• IBSI is an international initiative that standardizes nomenclature, computational workflows, and reporting guidelines for radiomics feature extraction.
• It defines rigorous mathematical formulations and validation datasets to minimize inter-software and inter-center variability.
• IBSI promotes data pooling and robust clinical modeling through comprehensive, consensus-based validation protocols and benchmark standards.

The Image Biomarker Standardisation Initiative (IBSI) is an international effort to harmonize the mathematical definitions, computational workflows, and reporting standards of imaging biomarkers (“radiomics features”) extracted from medical images. Its primary objective is to enable reproducible, interoperable, and clinically robust quantitative imaging by providing consensus-based nomenclature, reference implementations, comprehensive validation datasets, and exhaustive reporting guidelines (Zwanenburg et al., 2016, Lei et al., 2020).

1. Motivation and Founding Principles

Radiomics transforms medical images into large sets of quantitative features, such as first-order statistics, morphology, and texture descriptors, supporting clinical prediction and biomarker discovery. Early radiomics research exhibited profound inter-software and inter-center variability, with inconsistent definitions and implementations for nominally identical features resulting in irreproducible or conflicting results (Zwanenburg et al., 2016, Lei et al., 2020). IBSI was established as an independent, volunteer-driven collaboration spanning over 30 international institutions, with the explicit goal of providing:

• Consensus mathematical formulations and nomenclature for features
• Standardized pre-processing and analysis workflows
• Reference digital phantoms and feature value benchmarks
• Rigorous validation and reporting procedures

This standardization is essential for pooling data across studies, validating multivariate models, and ensuring that clinical translation of radiomics is scientifically reliable (Zwanenburg et al., 2016, Lei et al., 2020).

2. Consensus Feature Nomenclature and Definitions

IBSI formalizes the taxonomy of image biomarkers using well-defined, unique abbreviations and assigns each feature and parameter a permanent identifier. Feature classes include:

• MORPH: Morphological (shape) features (e.g., volume, surface area, compactness, sphericity)
• STAT/IS: First-order intensity statistics (mean, variance, skewness, kurtosis)
• IH: Intensity histogram features
• IVH: Intensity-volume histogram descriptors
• CM/GLCM: Gray-Level Co-occurrence Matrix features (e.g., joint entropy, contrast)
• RLM/GLRLM: Gray-Level Run-Length Matrix features (e.g., short-run emphasis)
• SZM/GLSZM: Gray-Level Size Zone Matrix
• DZM/GLDZM: Gray-Level Distance Zone Matrix
• NGTDM: Neighborhood Gray-Tone Difference Matrix
• NGLDM: Neighboring Gray-Level Dependence Matrix

Key mathematical expressions are standardized (e.g., for GLCM contrast: $$F_{\mathrm{cm.contrast}} = \sum_{i=1}^{N_g} \sum_{j=1}^{N_g} (i-j)^2 p_{ij}$$). Each feature’s computation is unambiguously defined in the IBSI reference manual, and configuration details (e.g., spatial discretization, texture matrix parameters) are encoded in the feature name for reporting (Zwanenburg et al., 2016, Salmanpour et al., 20 Nov 2025).

3. Reference Datasets, Validation Protocols, and Benchmarking

To substantiate software compliance and eliminate implementation ambiguities, IBSI provides multiple reference objects and specifies detailed validation protocols:

Artifact	Purpose	Methodology
Digital phantom	Feature implementation check	5×4×4 synthetic volume; known gray-levels, hand-calculable ground truth; no interpolation/discretization needed
Radiomics CT phantom	End-to-end validation	Clinical CT (NSCLC); multiple configs (A–E); benchmark values for each feature and processing condition
Convolutional filter phantoms	Filter feature benchmarking	Standardized application of mean, Gaussian, LoG, Gabor, wavelet filters; reference feature response maps

For each feature and configuration, the benchmark table includes the value (consensus median), tolerance, and degree of consensus (weak to very strong). Software is evaluated by relative difference from the benchmark, with <1% typical for well-implemented intensity/texture features. Morphology features are less reproducible due to meshing/numerical discrepancies (Lei et al., 2020).

Compliance procedures entail:

1. Computing required features on IBSI phantoms.
2. Comparing output to benchmark values within set tolerance.
3. Documenting all workflow parameters (interpolation, discretization, gray-level quantization, ROI specification, etc.).

4. Workflow Standardization and Reporting Guidelines

The IBSI pipeline prescribes a sequence of processing steps, each with explicit mathematical and implementation requirements:

1. Image loading and conversion: Acquisition metadata must be archived (modality, DICOM tags, normalization procedures).
2. Preprocessing: Denoising, bias correction (especially MRI), and, if applicable, quantitative conversion (SUV normalization in PET).
3. Segmentation: ROI definition, including mask topology, morphological basis, and observer consensus documentation.
4. Interpolation and resampling: To isotropic grids (e.g., to 1 mm for 3D); supported interpolation modes documented (nearest, linear, trilinear, tricubic spline).
5. Intensity normalization: Z-score, min-max, or none, with explicit reporting of mean/σ and normalization domain.
6. Gray-level discretization: Fixed-bin number (FBN) or fixed-bin size (FBS); IBSI prescribes formulae for each (e.g., FBS: $d_i = \lfloor \frac{v_i-v_{min}}{w} \rfloor+1$).
7. Convolutional filtering (if applied): Standardized filter kernels and parameters (support, σ, kernel family). All parameters must be reported with permanent IBSI IDs (Depeursinge et al., 2020).
8. Texture/feature calculation: As per IBSI definitions; includes all relevant directions, distances, and aggregation steps.
9. Output and reporting: Full manifest of processing parameters; feature values linked to explicit feature IDs.

Minimal reporting comprises at least 76 items spanning patient preparation, acquisition, reconstruction, preprocessing, segmentation, feature extraction, and software versioning (Zwanenburg et al., 2016, Salmanpour et al., 20 Nov 2025).

5. Software Ecosystem and Benchmarking Outcomes

Multiple open-source and proprietary radiomics libraries now cite or implement the IBSI standard (Lei et al., 2020, Salmanpour et al., 20 Nov 2025):

• PyRadiomics: Python, partial IBSI compliance; robust for first-order and most texture features.
• MITK: C++, considered IBSI reference for benchmarking.
• LIFEx, SERA, CaPTk: Varying feature coverage and compliance fidelity.
• PySERA: Python-native, modular implementation strictly operationalizing IBSI-compliant preprocessing, discretization, feature extraction, and deep radiomics integration. In benchmarking, achieved >94% feature-level reproducibility against IBSI/MITK, outperforming PyRadiomics and LIFEx for agreement and predictive accuracy (Salmanpour et al., 20 Nov 2025).

A comparative study found that non-morphology features across implementations generally matched IBSI values within <1% relative difference, while shape features were more discordant due to meshing algorithm divergence. Diagnostically critical features such as GLCM, GLRLM, GLSZM, and NGLDM are particularly sensitive to discretization settings, so precise parameter documentation is mandatory (Lei et al., 2020).

6. Impact on Reproducibility and Clinical Modeling

Adherence to IBSI standards demonstrably reduces inter-site and inter-software variation in feature extraction, enabling pooled multi-center model development and robust clinical biomarker deployment (Zwanenburg et al., 2016, Salmanpour et al., 20 Nov 2025). In application, such as osteoarthritis assessment on synthetic X-ray data, strict operationalization of IBSI protocols in resampling, normalization, discretization, and subsequent feature stability filtering (e.g., intraclass correlation coefficient ICC ≥ 0.90) yields machine learning models with high predictive ability and domain-robust performance (Alzubaidi et al., 14 Jan 2026).

Nevertheless, discrepancies in non-IBSI-compliant workflows, omitted features, or incomplete reporting continue to threaten reproducibility. Morphological features are specifically highlighted as an ongoing challenge due to their dependence on mesh parameterizations and ROI geometric uncertainties (Lei et al., 2020).

7. Ongoing Developments and Future Directions

IBSI actively updates its standards to encompass new filter families (e.g., Laws, Gabor, non-separable wavelets, Riesz transform), higher-dimensional preprocessing, and deep radiomics integration. The initiative publishes evolving reference manuals and maintains phantoms, compliance checklists, and benchmark tables via a central GitHub repository (Depeursinge et al., 2020).

The modular design of emerging libraries (notably PySERA) ensures that future features or plug-ins inherit IBSI-compliant preprocessing and reporting via object-oriented architectures, facilitating scalable and traceable radiomics pipelines. A plausible implication is that future advances in quantitative imaging—spanning classical and AI feature spaces—will be subsumed within a unified, auditable, and reproducible IBSI-compliant research infrastructure (Salmanpour et al., 20 Nov 2025).

By embedding rigorous, formally validated standards, the IBSI is foundational to the continued reproducibility, interoperability, and clinical deployment of quantitative imaging biomarkers across the medical imaging research landscape.