Imaging Biomarker Standardization Initiative (IBSI)

Updated 30 November 2025

Imaging Biomarker Standardization Initiative (IBSI) is an international consortium that defines precise mathematical radiomic features and protocols to resolve reproducibility issues.
It provides consensus-driven feature definitions across eleven radiomic families, standardizing preprocessing and extraction methods for consistent results.
IBSI offers publicly available benchmark datasets and detailed reporting guidelines to facilitate multi-center validation and clinical adoption of radiomics.

The Imaging Biomarker Standardization Initiative (IBSI) is an international consensus-driven consortium whose mission is to define, standardize, and benchmark the computational extraction of quantitative image features—commonly termed radiomics biomarkers—from medical imaging. IBSI’s framework encompasses precise mathematical feature definitions, standardized preprocessing protocols, comprehensive reporting guidelines, and publicly available benchmark datasets for validation. Its goal is to resolve widespread reproducibility deficits that historically undermined multi-center studies, meta-analyses, and clinical adoption of radiomics in biomedical research and imaging-based precision medicine (Depeursinge et al., 2020, Salmanpour et al., 20 Nov 2025, Lei et al., 2020, Zwanenburg et al., 2016).

1. Origins, Structure, and Motivation

IBSI was initiated in 2016 by imaging scientists, clinicians, and software developers including Zwanenburg, Leger, and Vallières, responding to a recognized reproducibility crisis in radiomics research arising from inconsistent image processing steps, terminology, and feature implementations. The IBSI consortium presently includes over 100 contributors and operates via open collaboration, with results published as reference manuals, publicly accessible benchmark data, and documented best practices (Zwanenburg et al., 2016). All features and processing steps are assigned immutable four-character codes for unambiguous reference.

The reproducibility crisis that motivated IBSI stems from heterogeneity in radiomics software and practices. As observed in multi-tool benchmarking, feature values for the same region of interest (ROI) can diverge by large margins across different pipelines, particularly for mesh-based morphology features and features calculated with inconsistent gray-level discretization schemes (Lei et al., 2020). The absence of common standards impeded validation, generalization, and clinical translation.

2. Standardized Feature Definitions and Taxonomy

IBSI provides consensus mathematical definitions for all commonly used radiomic features, organized into eleven families: morphology, local intensity, first-order statistics, intensity histogram, intensity–volume histogram, and six texture families (GLCM, GLRLM, GLSZM, GLDZM, NGTDM, NGLDM) (Zwanenburg et al., 2016, Salmanpour et al., 20 Nov 2025, Lei et al., 2020).

Each feature is defined by a closed-form formula and classified with its appropriate family and four-character identifier (e.g., "Q4LE" for mean intensity). Taxonomic abbreviations used in IBSI publications include:

Family Abbreviation	Description
MORPH	Morphological features
STAT	Intensity-based statistics
IH	Intensity histogram
IVH	Intensity-volume histogram
GLCM	Grey-level co-occurrence matrix
GLRLM	Grey-level run-length matrix
GLSZM	Grey-level size-zone matrix
GLDZM	Grey-level distance-zone matrix
NGTDM	Neighbourhood grey-tone diff.
NGLDM	Neighbouring grey-level depend.

This unambiguous nomenclature allows inter-paper and inter-tool comparability by disambiguating aggregation, modality, and preprocessing parameters (ex: entropy₍IH,CT,FBS:25HU₎ denotes intensity-histogram entropy from CT imaging, discretized with a 25 HU fixed bin size) (Zwanenburg et al., 2016).

3. Image Preprocessing and Feature Extraction Workflow

IBSI enforces a standardized processing workflow to minimize pipeline-induced feature variability. The canonical steps are (Depeursinge et al., 2020, Salmanpour et al., 20 Nov 2025, Zwanenburg et al., 2016):

Image loading and optional intensity conversion (e.g., PET SUV scaling)
Pre-processing: denoising, bias-field correction (MRI), and motion correction
Segmentation: generating or providing an ROI mask
Interpolation: resampling both the image and mask to isotropic voxels (e.g., 1 × 1 × 1 mm³), with specified interpolation algorithms (tricubic/trilinear)
Re-segmentation: restricting intensity range (e.g., [−1000,400] HU)
Padding: extending boundaries with explicit conditions (constant, nearest, periodic, mirror)
Discretization: quantizing intensities using fixed bin size (FBS) or fixed bin number (FBN)
Feature Calculation: extracting features using the IBSI mathematical definitions, including explicit aggregation strategies for directional invariance

For texture matrices (GLCM, GLRLM, GLSZM, etc.), IBSI prescribes directionality strategies (e.g., 13 directions in 3D, averaging or merging across directions for rotational invariance), neighbor connectivity, and distance norms.

4. Convolutional Filtering Standardization

Convolutional filters (e.g., Gaussian, Laplacian-of-Gaussian, mean, Laws, Gabor, wavelet, Riesz transforms) are used to emphasize spatial patterns such as edges, blobs, and scale-specific structures prior to feature extraction. IBSI’s convolutional-filter manual version 9 formalizes the definitions, implementation strategies (boundary conditions, separability, rotation-invariance), and compliance testing (Depeursinge et al., 2020):

Gaussian Smoothing: $g_σ[x] = (1/(√{2π} σ)^D) e^{−||x||^2/(2 σ^2)}$ with practical kernel truncation.
LoG (Laplacian-of-Gaussian): $g_σ[x]= −(1/σ^2)... e^{−||x||^2/(2σ^2)}$ .
Mean, Gabor, Laws, Wavelets, Riesz: Each with explicit parameterization and reference frequency responses.
Mandatory reporting parameters: kernel size, σ, λ, boundary condition, kernel family, level, pooling/aggregation specifics.

Compliance requires each step—from padding and rotating to frequency-domain formulation—to match the manual’s prescriptions. IBSI establishes phantoms and reference response maps for each filter configuration, along with tolerance windows for output verification.

5. Benchmarking and Software Compliance Testing

IBSI supplies digital phantoms and clinical datasets (e.g., a 5×4×4 voxel digital phantom, and a lung cancer CT phantom) with ground-truth feature values for each processing configuration (Zwanenburg et al., 2016, Lei et al., 2020, Salmanpour et al., 20 Nov 2025). Benchmark datasets are used in a staged compliance process:

Phase 1: Digital phantom, testing ~50 filter configurations, with a pass criterion of ≤1% per-voxel deviation in filter responses for 99% of voxels.
Phase 2: Clinical CT images, reference values for first-order features post-filtering, tolerances at 1% of feature distribution range.
Phase 3: Multi-modality PET/CT/MRI sarcoma data, assessing robustness across imaging types.

Reproducibility is quantified as the percent of features within predefined error thresholds when compared to IBSI reference values. For example, PySERA achieved >94% IBSI-compliant reproducibility, closely matching MITK and outperforming other toolkits (Salmanpour et al., 20 Nov 2025).

Inter-software comparison studies reveal that while first-order and texture features generally exhibit high inter-tool agreement (median relative differences <2% across most pipelines), morphology features remain more variable (sometimes >10% difference) due to differences in mesh calculation, surface node placement, and interpolation (Lei et al., 2020).

6. Reporting Guidelines and Best Practices

IBSI’s reference manuals include a comprehensive reporting checklist (76 items) covering the full radiomics workflow (Zwanenburg et al., 2016). Mandatory reporting parameters comprise:

Imaging modality, acquisition/reconstruction protocol
Segmentation protocol and inter-operator strategy
Interpolation and grid alignment methods
Intensity range usage and outlier handling decisions
Discretization mode and parameters (FBN/FBS, number/size of bins)
Feature aggregation schemes (direction averaging or merging)
Software and algorithm versioning
All processing and parameter settings required to guarantee reproducibility

Transparent logging of these parameters is required to permit independent audit, facilitate multi-center harmonization, and adhere to the FAIR (Findable, Accessible, Interoperable, Reusable) principles (Salmanpour et al., 20 Nov 2025).

Best practices, as distilled in recent IBSI-compliant software such as PySERA, further include the use of fixed random seeds, the full application of all preprocessing pipelines, validation on supplied phantoms, and unified preprocessing steps for both handcrafted and deep learning features to avoid domain shift (Salmanpour et al., 20 Nov 2025).

7. Limitations, Open Issues, and Future Directions

Despite its consensus-driven rigor, IBSI acknowledges several unresolved challenges (Depeursinge et al., 2020, Lei et al., 2020):

Morphological features derived from mesh representations remain poorly harmonized due to differences in surface generation and interpolation.
Convolutional-filter rotation invariance is only approximate for separable kernels and not fully realized in all cases; some Riesz filter validation remains open.
Interpolation, especially prior to high-pass filtering, can alter frequency content in nontrivial ways.
Physical-to-voxel unit conversions (σ, λ) across images with differing grid spacing require careful, explicit handling.
Non-convolutional image processing (e.g., median or morphological filtering) is outside current IBSI scope.
Software adherence and exact reproducibility are currently verified for a finite set of feature classes; as radiomics expands into deep learning and multimodal contexts, further standardization may be required.

A plausible implication is that ongoing extension of IBSI to cover hybrid (handcrafted and deep) feature representations, as implemented in PySERA, is vital for future-proofing radiomics reproducibility (Salmanpour et al., 20 Nov 2025).

References: