PaperlayerView exact revision on arXiv

Source-linked AI summary

Image biomarker standardisation initiative

Alex Zwanenburg, Stefan Leger, Martin Vallières, Steffen Löck

arXiv:1612.07003v11cs.CVeess.IV

TL;DR

Radiomics lacks consensus guidelines and definitions for translating acquired images into reproducible image biomarkers. IBSI addresses this gap with a standardised processing scheme, biomarker definitions and nomenclature, reporting guidance, and reference datasets and values; its recommendations show that processing choices materially affect reproducibility. The work is limited to biomarker computation from acquired imaging, not biomarker validation or acquisition and reconstruction harmonisation.

  • Problem

    Radiomics lacks consensus guidelines and definitions for translating acquired imaging into high-throughput image biomarkers, limiting standardisation of the extraction process.

  • Method

    IBSI develops consensus-based recommendations, processing guidelines, biomarker definitions, nomenclature, reporting guidance, and reference datasets and values for image biomarker computation.

  • Results

    Discretisation choice substantially affects intensity distributions, feature values, and reproducibility, with optimal methods depending on imaging modality.

  • Takeaways & Limitations

    The manual provides a common framework for processing, naming, reporting, and checking high-throughput image biomarkers to support reproducible radiomic studies.

  • Takeaways & Limitations

    The work focuses on computing image biomarkers from acquired imaging and does not validate biomarkers or harmonise image acquisition and reconstruction.

Abstract

from arXiv · show

The image biomarker standardisation initiative (IBSI) is an independent international collaboration which works towards standardising the extraction of image biomarkers from acquired imaging for the purpose of high-throughput quantitative image analysis (radiomics). Lack of reproducibility and validation of high-throughput quantitative image analysis studies is considered to be a major challenge for the field. Part of this challenge lies in the scantiness of consensus-based guidelines and definitions for the process of translating acquired imaging into high-throughput image biomarkers. The IBSI therefore seeks to provide image biomarker nomenclature and definitions, benchmark data sets, and benchmark values to verify image processing and image biomarker calculations, as well as reporting guidelines, for high-throughput image analysis.

The image biomarker standardisation initiative

IBSI provides consensus-based standards for extracting, defining, reporting, and validating image biomarkers in high-throughput image analysis. The manual covers image processing, biomarker definitions, reporting nomenclature, reference datasets, and reference values, while excluding biomarker validation and image acquisition standardisation.

  • Nomenclature and references: IBSI provides permanent four-character identifiers for image biomarker definitions and related concepts, rather than page or section numbers.These identifiers are intended to remain stable when document pagination or section numbering changes.
  • Purpose and scope: The manual presents consensus-based recommendations, guidelines, definitions, and reference values to improve reproducibility of radiomic studies.It defines a general radiomics image processing scheme alongside these standardisation resources.
  • Scope boundaries: The manual focuses on computing biomarkers from acquired imaging and excludes validation of imaging biomarkers and broader acquisition and reconstruction harmonisation.It refers to existing recommendations where scope overlaps with other standardisation efforts.
  • Manual structure: The manual covers image processing, image biomarker definitions, radiomic reporting and nomenclature, and datasets and configurations for reference values.These topics are organised across Chapters 2–5.
  • Processing standardisation: Interpolation algorithm choice affects feature reproducibility, with different algorithms yielding more reproducible values for different features in a phantom study.The cited comparison included nearest neighbour, trilinear, and tricubic interpolation.
  • Processing standardisation: Align grid centers is recommended because it is implementation-independent and achieves the desired voxel spacing.This recommendation addresses variation introduced by interpolation-grid alignment.

2.6 ROI extraction

ROI extraction isolates the region used for feature calculation, after which intensities are discretised to make texture-feature calculation tractable. Fixed bin number and fixed bin size preserve different aspects of intensity information, so the choice affects feature values and reproducibility.

  • ROI masking: ROI masking retains intensities inside the ROI and excludes surrounding voxels, commonly replacing excluded voxels with NaN.The mask isolates the image volume used for feature calculation.
  • Discretisation: Discretisation converts ROI intensities into bins to make texture-feature calculation tractable and suppress noise.The IBSI describes fixed bin number and fixed bin size as the two common approaches.
  • Fixed bin number: Fixed bin number discretisation normalises the observed ROI intensity range, with the number of bins controlling image coarseness and enabling comparison across ROIs.It can be useful when intensity units are arbitrary and contrast is important, but it breaks the relationship with physiological meaning.
  • Fixed bin size: Fixed bin size discretisation assigns bins of width wb from a specified minimum intensity, preserving the relationship with the original intensity scale.The IBSI recommends a common lower bound across samples when a re-segmentation range is available and does not recommend FBS for arbitrary intensity units.
  • Recommendations: The optimal discretisation method is modality-dependent, and discretisation substantially affects intensity distributions, feature values, and reproducibility.Recommendations depend on imaging intensity definitions, re-segmentation ranges, and discretisation algorithms.

3.2 Local intensity features

Local intensity features use 3D neighbourhoods around ROI voxels, including surrounding voxels, and are calculated volumetrically rather than per slice. Peak features summarize mean intensity within 1 cm^3 spherical regions.

  • Local intensity features use ROI voxels as centers but draw neighbourhood intensities from all voxels, including those outside the ROI.
  • Local intensity features are calculated in 3D and not independently for each image slice.
  • The local intensity peak is the mean intensity within a 1 cm^3 sphere centered on an ROI voxel with maximum intensity.
  • The local peak uses a radius of approximately 0.62 cm, includes the center voxel, and selects the highest value when multiple maxima exist.
  • Reference-value tables are provided separately for local and global intensity peak features.

3.3 Intensity-based statistical features

Intensity-based statistical features describe continuous intensity distributions within a 3D ROI without discretisation, but they are not meaningful when the intensity scale is arbitrary.

  • Intensity-based statistical features describe how continuous voxel intensities are distributed within the ROI and do not require discretisation.
  • Features should be calculated from the 3D volume rather than computed per slice and averaged.
  • The feature set includes central tendency, dispersion, distribution-shape, percentile, and range measures such as mean, variance, skewness, kurtosis, median, and IQR.
  • Kurtosis applies a Fisher correction of -3 to center normal distributions on 0, and is set to 0 when intensity variance is zero.
  • The 10th and 90th percentiles are described as more robust alternatives to minimum and maximum intensity, respectively.

3.5 Intensity-volume histogram features

Intensity-volume histogram features relate discretised intensity levels to the volume fraction containing at least each level. Their calculation depends on modality-specific discretisation and should use the full 3D volume.

  • An IVH describes the relationship between discretised intensity i and the volume fraction containing at least intensity i.
  • IVH calculation may require discretising intensities and specifying the total range G and discretisation interval wd.
  • Calibrated discrete modalities such as CT require no discretisation, whereas calibrated continuous modalities such as PET require fixed bin size discretisation.
  • Fixed bin size discretisation replaces bin numbers with bin-center intensities to preserve their relationship with original intensities.
  • Fixed bin number discretisation with Ng = 1000 bins is recommended for arbitrary-intensity modalities such as many MRI sequences.
  • IVH features should be calculated from the 3D volume, and slice-wise calculation followed by averaging is not recommended.
  • The feature set includes volume and intensity thresholds, their differences, and area under the IVH curve; one feature has no reference values and should not be used.

3.6 Grey level co-occurrence based features

GLCM-based features quantify neighbouring grey-level co-occurrences across image directions, with multiple aggregation schemes available to improve rotational invariance. Feature values can depend strongly on the aggregation method.

  • GLCMs describe the distribution of discretised grey-level pairs among neighbouring pixels or voxels along image directions.
  • For each direction, opposite-direction co-occurrence matrices are summed, producing a symmetric GLCM.
  • GLCM features are derived by normalising each matrix into a probability distribution and computing marginal, diagonal, and cross-diagonal probabilities.
  • Merged matrices are formed by summing corresponding co-occurrence counts before probability normalisation, and aggregation can strongly affect feature values.
  • The default neighbourhood uses Chebyshev distance 1; alternative distance-norm constructions are rare and cautioned against because of reproducibility concerns.
  • Distance weighting is not recommended because several neighbourhood, norm, aggregation, and distance conditions can make it ineffective.

3.7 Grey level run length based features

GLRLM features quantify runs of identical discretised grey levels along image directions, with directional and dimensional aggregation choices affecting the resulting feature value.

  • Definition: GLRLMs assess consecutive runs of identical discretised grey levels, unlike GLCMs, which assess neighbouring grey-level co-occurrences.A run length is the length of a consecutive sequence of pixels or voxels with the same grey level along a direction.
  • Aggregation: Six aggregation methods combine directional matrices by averaging features or merging matrices across 2D slices, directions, or 3D directions.The methods include BTW3, SUJT, JJUI, ZW7Z, ITBB, and IAZD.
  • Aggregation: Feature values may depend strongly on aggregation method because merged matrices sum run counts and voxel totals across the contributing matrices.For methods 2, 3, 4, and 6, Nv must also be summed to retain consistency.
  • Distance weighting: Distance weighting changes GLRLM features only under specific conditions, including matrix merging and weighting schemes other than the Chebyshev norm.The default weighting is w = 1; inverse-distance and exponential weighting are alternatives.
  • Reference values: Digital-phantom agreement was strongest for 3D aggregation, with values of 0.705 for averaged and 0.729 for merged configurations.The corresponding consensus labels were very strong.
  • Reference values: Reference tables report feature-specific values, tolerances, and consensus levels across configurations and digital-phantom aggregation methods.The section includes reference values for run-length and grey-level emphasis features, non-uniformity, and related measures.

3.8 Grey level size zone based features

GLSZM features count connected voxel zones by discretised grey level and zone size, with 2D, merged 2D, and 3D aggregation options that can materially affect feature values.

  • Definition: GLSZMs count groups of linked voxels sharing an identical discretised grey level, using 8-connectedness in 2D and 26-connectedness in 3D.The matrix element s_ij records the number of zones with grey level i and size j.
  • Aggregation: Three aggregation methods calculate features from slice-wise matrices averaged over slices, one merged 2D matrix, or one 3D matrix.The methods are identified as 8QNN, 62GR, and KOBO.
  • Aggregation: Merged GLSZMs sum zone counts across slices and require corresponding voxel-count sums for consistency; feature values may depend strongly on aggregation method.The merged approach is illustrated as 2.5D aggregation in Figure 3.4.
  • Neighbourhood: The default GLSZM neighbourhood uses Chebyshev distance δ = 1, whereas alternative norms produce 6-connected 3D and 4-connected 2D neighbourhoods.Using different neighbourhoods is non-standard and cautioned against.
  • Reference values: Digital-phantom small-zone emphasis values were 0.363 for 2D, 0.368 for 2.5D, and 0.255 for 3D.Consensus was strong for 2D and 2.5D and very strong for 3D.
  • Reference values: Reference tables cover zone-size, grey-level, non-uniformity, percentage, variance, entropy, and combined emphasis features.Each table provides reference values for a named GLSZM feature.

3.9 Grey level distance zone based features

GLDZM features combine connected grey-level zones with their distances to the ROI edge, using distance maps and aggregation choices defined for reproducible image analysis.

  • Definition: GLDZMs count connected zones sharing a discretised grey level and the same distance to the ROI edge, thereby capturing the relation between location and grey level.A grey-level zone map and a distance map are required.
  • Connectivity and distance: Voxel linkage uses 26-connectedness in 3D and 8-connectedness in 2D, while distances to the ROI edge use 6-connectedness and 4-connectedness, respectively.A linked zone’s distance is the minimum distance among its voxels.
  • Definition: The definition departs from the original by using iterative 6- or 4-connected distance determination for non-rectangular ROIs.The procedure repeatedly erodes the ROI, increments distances for eroded voxels, and stops when the ROI is empty.
  • Distance convention: Distances start at 1 rather than 0 for voxels on the ROI edge, preventing division by 0 in some features.This is a stated difference from the original definition.
  • Definition: The intensity mask determines zones, while the morphological ROI mask determines the distance map used by the GLDZM.This dual-mask construction distinguishes GLDZM computation from GLSZM computation.
  • Aggregation: Three aggregation methods average slice-wise features, merge all 2D matrices, or compute from a 3D matrix; aggregation can strongly affect feature values.Merged GLDZMs sum zone counts and voxel totals across slices.
  • Limitations: Using non-default distance norms is cautioned against because Euclidean distances are less efficient, may require rounding, and can create reproducibility issues.The default boundary-distance norm is Manhattan.
  • Reference values: Digital-phantom small-distance emphasis values were 0.946 for 2D and 0.917 for 2.5D, with strong and moderate consensus, respectively.The section also provides reference values for large-distance, grey-level, non-uniformity, percentage, and variance features.

3.10 Neighbourhood grey tone difference based featuresIPET

NGTDM features quantify differences between discretised voxel intensities and neighbouring averages, with definitions adapted for irregular ROIs and 2D or 3D aggregation. Distance norms and weighting are non-standard because they may impair reproducibility.

  • Definition: NGTDM sums grey-level differences between voxels and the average intensities of their neighbouring voxels within Chebyshev distance δ.For 3D volumes, the neighbourhood contains W = (2δ + 1)^3 − 1 voxels; 2D neighbourhoods contain W = (2δ + 1)^2 − 1 pixels.
  • ROI handling: A neighbourhood is valid when at least one neighbouring voxel belongs to the ROI, rather than requiring the complete neighbourhood to lie inside the ROI.This modification accommodates irregular ROI masks; the neighbourhood size W_k counts only ROI voxels.
  • Normalisation: NGTDM grey-level probabilities are p_i = n_i/N_v,c, where N_v,c counts ROI voxels having at least one valid neighbour.If every voxel has a valid neighbour, N_v,c equals N_v; the example has N_g = 4 and N_g,p = 3.
  • Aggregation: Three aggregation schemes compute features from slice-wise matrices averaged over slices, merged 2D matrices, or a single 3D matrix.When matrices are merged, s_i and n_i are summed and N_v,c and p_i are recomputed; feature values may depend strongly on aggregation.
  • Reproducibility: Distance norms other than Chebyshev and distance weighting are non-standard and cautioned against because of potential reproducibility issues.Manhattan or Euclidean norms and weights such as ||m||^-1 or exp(-||m||^2) are described as possible alternatives.

3.11 Neighbouring grey level dependence based featuresREK0

NGLDM features count neighbouring voxels whose discretised grey levels differ from a centre voxel by at most a coarseness parameter. IBSI modifies the definition for irregular ROIs and cautions that non-standard neighbourhood choices can compromise reproducibility.

  • Definition: The NGLDM counts neighbouring voxels dependent on a centre voxel when their discretised grey-level difference is at most α within Chebyshev distance δ.The resulting matrix records counts by centre grey level and dependence count.
  • ROI handling: IBSI considers all voxels in the ROI mask even when neighbourhoods extend outside the ROI, with outside voxels contributing nothing to dependence counts.This avoids inconsistent results for non-cuboid ROIs and makes N_v equal to N_s.
  • Parameters: The typical settings are α = 0 and Chebyshev distance δ = 1, although other coarseness parameters and larger neighbourhood radii are possible.These parameters determine which neighbouring grey levels are counted as dependent.
  • Aggregation: NGLDM features can be computed from slice-wise matrices averaged over slices, a merged matrix, or a single 3D matrix.Merged matrices require element-wise summation of dependence counts and corresponding updates to N_v and N_s; feature values may depend strongly on aggregation.
  • Reproducibility: Distance norms other than Chebyshev and distance weighting are non-standard and cautioned against because of potential reproducibility issues.The generalised formulation permits alternative norms and weights, but IBSI does not recommend them as standard practice.
  • Feature relationships: The normalised grey-level non-uniformity from a single 3D NGLDM is equivalent to the intensity histogram uniformity feature.Grey-level non-uniformity is low when dependence counts are evenly distributed across grey levels.
  • Definition correction: The dependence count entropy definition corrects Sun and Wee’s formulation by using count probability p_ij rather than dependence count s_ij inside the logarithm.The probability is defined as p_ij = s_ij/N_s.

Radiomics reporting guidelines and nomenclature

Reliable and complete reporting is presented as necessary for reproducible and validated image-processing and image-biomarker extraction results.

  • The guidelines aim to support complete reporting of image processing and image biomarker extraction, alongside a nomenclature system for uniquely identifying features.

4.1 Reporting guidelines

The reporting guidelines specify acquisition, reconstruction, processing, segmentation, interpolation, and voxel-dimension information needed to describe radiomics workflows. They also identify the relevant parameters for different imaging modalities and processing stages.

  • Acquisition and patient factors: The guidelines require reporting the region of interest, patient preparation, drug and tracer administration, contrast-agent details, comorbidities, and acquisition protocol.They also request scanner type, imaging modality, scan timing, motion handling, calibration, and scan duration.
  • Modality-specific acquisition: Modality-specific acquisition details include CT voltage, current, exposure, MRI sequence and timing parameters, PET reconstruction settings, and tracer information.The MRI items include echo, repetition, inversion, and flip-angle parameters; PET items include iterations, subsets, and point-spread-function modelling.
  • Image processing: Reporting should cover image corrections, data conversions, interpolation, noise suppression, smoothing, skull stripping, non-uniformity correction, and intensity normalisation.The guidelines also request other post-acquisition processing methods not separately listed.
  • Segmentation: Segmentation reporting should describe the method, expert number and expertise, consensus strategy, automation settings, modality used, and conversion to a voxel-based mask.
  • Interpolation: Interpolation reporting should specify the algorithm, grid positioning and dimensions, extrapolation handling, and final interpolated voxel size.

4.2 Feature nomenclature

The IBSI nomenclature combines a feature name with a subscript that records details needed to distinguish otherwise ambiguous image features. Subscripts can encode modality, processing, aggregation, interpolation, re-segmentation, discretisation, and feature-specific settings using standardized abbreviations and notation.

  • Core structure: Feature names use subscripts to distinguish features that may otherwise share the same name, incorporating only details for which ambiguity may exist.The example contrasts intensity histogram-based entropy with grey level run length matrix entropy using distinct family, modality, and discretisation descriptors.
  • Core structure: Subscripts may specify feature family, aggregation method, modality or channel, spatial filters, interpolation, re-segmentation, discretisation, and feature-specific parameters.The feature family abbreviation is required; the other descriptors are optional and address potentially ambiguous settings.
  • Processing descriptors: Re-segmentation notation records intensity ranges and outlier filtering, which may be combined in either order.Examples include CT ranges, PET SUV thresholds, and MRI intensity filtering using µ ± 3σ.
  • Processing descriptors: Discretisation notation distinguishes fixed bin size, fixed bin number, histogram equalisation, and Lloyd-Max methods, while additional parameters encode weighting or distance choices.Feature-specific settings include weighting functions and Chebyshev distances for several matrix families; co-occurrence symmetry can be symmetrical or asymmetrical, with asymmetry not recommended.
  • Aggregation: Aggregation notation distinguishes 2D, 2.5D, and 3D calculations, including averaging or merging directions across slices or volumes.Three-dimensional calculation over the volume is the default, while alternatives include slice-wise averaging and merged directional matrices.
  • Processing descriptors: Interpolation notation records the method, dimensionality, and uniform spacing, with examples including nearest neighbour, linear, cubic convolution, and cubic spline interpolation.The spacing marker includes units, and dimensionality or spacing may be omitted when clear from context.

Reference data sets

IBSI provides reference values for image features using a digital image phantom and a lung cancer CT image. These data sets can also be used to verify radiomics software implementations.

  • Reference data sets: Reference feature values were obtained from a digital image phantom and a CT image of a lung cancer patient.The same reference data sets are intended for verification of radiomics software implementations.
  • Reference data sets: The reference data sets are publicly available through the IBSI data repository.The passage identifies the repository as https://github.com/theibsi/data sets.

5.1 Digital phantom

The digital phantom is a small, manually specified voxel volume designed for direct feature calculation and comparison with radiomics software. Its ROI exclusions, intensity range, and matrix settings define the benchmark conditions.

  • Phantom design: The digital phantom contains 5 × 4 × 4 voxels, with slices stacked along the z direction.Each slice comprises the voxels in the x-y plane at a particular z position.
  • Phantom design: The ROI excludes several exterior voxels and one internal voxel, while grey levels 2 and 5 are absent and the present range is 1 through 6.Excluded voxels are identified in blue in the phantom figure.
  • Calculation settings: The phantom requires no image processing before feature calculation, so features are calculated directly from the phantom.Fixed bin size 1 or fixed bin number 6 would preserve the phantom contents but is not required.
  • Calculation settings: GLCMs are symmetrical and use Chebyshev distance δ = 1; NGLDM and NGTDM also use δ = 1, with NGLDM coarseness α = 0.These settings establish the matrix configurations used for the phantom calculations.
  • Calculation settings: Because the phantom is not discretised, most intensity-based statistical features match their intensity histogram-based analogues in value.The phantom uses identical ROI morphological and intensity masks because no re-segmentation is applied.

5.2 Lung cancer CT image

The lung cancer CT reference data set uses a delineated tumour ROI and multiple 2D and 3D processing configurations. Interpolation and diagnostic descriptors are emphasized because implementation details can alter feature values and reveal processing discrepancies.

  • Data and configurations: The CT reference image comes from the first patient in a public four-patient non-small-cell lung carcinoma data set, with the gross tumour volume used as the ROI.The image and segmentation are available in DICOM and converted NIfTI formats.
  • Data and configurations: Five processing configurations test different methods, separating 2D in-slice processing from 3D volumetric interpolation and texture extraction.Configurations A and B are 2D, whereas C-E are 3D.
  • Interpolation: Interpolation has a major influence on feature values, and implementations of the same method may differ because they use different interpolation grids.The recommended checks include grid origin, ceiling-based grid size, aligned grid centers, boundary handling, and full-precision image and ROI representations.
  • Diagnostic features: Diagnostic features describe changes in the image and ROI masks after each processing step and have reference values in the IBSI compliance spreadsheet.They are intended to help identify implementation issues whose effects can propagate across sequential processing steps.
  • Diagnostic features: Diagnostic descriptors include image and ROI dimensions, bounding boxes, voxel counts, and minimum, mean, and maximum intensities before and after interpolation or re-segmentation.The same ROI descriptors can characterize interpolated and re-segmented regions of interest.
  • CT calculation settings: For the CT image, discretisation causes intensity-based statistical features to differ from their intensity histogram analogues, and re-segmentation makes the morphological and intensity masks non-identical.GLCM, NGLDM, and NGTDM use symmetrical or distance-based settings with δ = 1 and α = 0; IVH processing varies by configuration.

Appendix A

Appendix A provides reference texture matrices extracted from the digital phantom, covering 2D and volumetric 3D image-biomarker representations. The tables include directional and merged forms across several matrix families.

  • Reference texture matrices are extracted from the digital phantom for comparison and verification purposes.
  • Grey-level co-occurrence matrices are provided for 2D xy-plane and volumetric 3D extraction, including directional and merged forms.The directional tables use Chebyshev distance 1 and identify directions in (x, y, z) coordinates.
  • Grey-level run length matrices are provided for 2D xy-plane and volumetric 3D extraction, with separate directional and merged tables.
  • Grey-level size zone and distance zone matrices are reported for both 2D xy-plane and volumetric 3D extraction.
  • Neighbourhood grey tone difference and neighbouring grey level dependence matrices are reported in 2D and 3D forms using Chebyshev distance 1.The neighbouring grey level dependence matrices additionally specify coarseness 0.
Loading 1612.07003v11…