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

Consensus guidelines and definitions for translating acquired imaging into image biomarkers are scarce, and distinct features can share ambiguous names. The IBSI reference manual presents consensus recommendations and reports that discretisation choices substantially affect feature values and reproducibility, with optimal choices depending on modality.

  • Problem

    Consensus guidelines and definitions for translating acquired imaging into image biomarkers are scarce, and distinct features can share ambiguous names.

    Source: Introduction · p. 11, §4.2 · p. 148

  • Method

    The IBSI reference manual presents consensus recommendations, guidelines, definitions, reference values, and a general radiomics image-processing scheme.

    Source: Introduction · pp. 11–12

  • Results

    Discretisation choices substantially affect intensity distributions, feature values, and reproducibility, while the optimal choice for feature reproducibility depends on modality.

    Source: Recommendations · p. 25

  • Takeaways & Limitations

    The proposed feature nomenclature distinguishes features that would otherwise share an ambiguous name such as entropy.

    Source: §4.2 · p. 148

    The work focuses on computing image biomarkers from acquired imaging; validation of imaging biomarkers falls beyond its scope.

    Source: Introduction · p. 11

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 … 2.7 Intensity discretisation

This section presents IBSI guidance for extracting image biomarkers from acquired imaging, with the aim of improving radiomics reproducibility. It covers processing choices—including interpolation, ROI handling, and intensity discretisation—and their effects on image representation and feature reproducibility.

Permanent identifiers

  • IBSI assigns four-character permanent identifiers to image biomarker definitions and related concepts; references should use these codes rather than page or section numbers.Page and section numbers may change.Source: Permanent identifiers · p. 2

Introduction

  • The manual proposes consensus recommendations, guidelines, definitions and reference values, alongside a general radiomics processing scheme, to improve reproducibility.Its chapters cover image processing, image biomarkers, reporting and nomenclature, and reference data sets and processing configurations.Source: Introduction · pp. 11–12

Image processing

  • Image processing derives features from 3D image stacks on a regular voxel grid, and IBSI presents a general processing scheme whose steps may vary by modality and purpose.Although images may have multiple channels, the manual considers them single-channel; channels may instead be processed separately, selected, or combined.Source: Image processing · p. 13, Figure 2.1 · p. 14

2.1 Data conversion

  • Converting raw data into a more meaningful representation, such as SUV, may be required for some modalities, but assessing conversion methods is outside this work’s scope.Source: §2.1 · p. 13

2.2 Image post-acquisition processing

  • Post-acquisition processing can address modality-specific issues such as noise, intensity non-uniformity and artifacts, but evaluating or standardising these methods is outside scope.Vendor reconstruction software may already perform noise reduction or other post-processing, making additional processing unnecessary.Source: §2.2 · p. 15

2.3 Segmentation

  • ROIs define where features are calculated and may be delineated manually or algorithmically; polygon contours can be rasterised using ray intersections and the even-odd rule.The example implementation may inconsistently classify voxel centers lying on polygon boundaries.Source: §2.3 · pp. 15–16

2.4 Interpolation

  • Texture features require isotropic interpolation for rotational invariance and comparability, but upsampling can infer artificial information while downsampling loses information and may cause aliasing.The manual notes no clear preference between upsampling and downsampling; 2D interpolation may suit large slice spacing but restricts texture calculations to the plane because spacing remains anisotropic.Source: §2.4 · pp. 16–17

Interpolation algorithms

  • Nearest-neighbour, trilinear and tricubic interpolation are common; tricubic convolution is faster than tricubic spline, but both yield similar results and speed differences are practically negligible.Trilinear interpolation uses eight nearby voxels, while tricubic methods use a larger neighbourhood to estimate a smooth third-order polynomial.Source: Interpolation algorithms · p. 17
  • Feature reproducibility depends on interpolation algorithm; trilinear is a conservative choice, while tricubic may retain contrast better and trilinear may suppress high spatial frequencies.Nearest-neighbour interpolation can introduce blockiness and local-texture bias; higher-order interpolation can overshoot at acute intensity transitions. CT intensities are rounded to integers after interpolation.Source: Interpolation algorithms · p. 17, Rounding image intensities after… · p. 17

Partial volume effects in the ROI mask E8H9

  • When a segmented ROI is interpolated, its mask should also be interpolated; nearest-neighbour avoids partial volumes, while trilinear requires thresholding fractions, commonly at δ = 0.5.At least full-precision 32-bit floating-point representation is recommended to avoid rounding errors in image and mask interpolation.Source: Partial volume effects in the ROI mask… · pp. 17–18

Interpolation grid UMPJ

  • Aligning grid centers is recommended because it is implementation-independent and achieves the desired voxel spacing.For the 4 × 4 example, fitting the grid to the original changes the realized spacing to 1.80 mm, whereas origin and center alignment retain 2.00 mm spacing.Source: Interpolation grid UMPJ · p. 18

Interpolation grid dimensions 026Q

Range re-segmentation USB3

Fixed bin number K15C

  • Fixed-bin-number discretisation normalizes the ROI to a chosen bin count, supporting comparisons across ROIs but breaking the link to physiological intensity meaning.Discretisation makes texture-feature calculation tractable and can suppress noise; fixed-bin-number normalization may suit arbitrary intensity units when contrast matters.Source: §2.7 · p. 23, Fixed bin number K15C · p. 24

Fixed bin size Q3RU

  • Fixed-bin-size discretisation preserves the original intensity scale; using a shared minimum from the re-segmentation lower bound is recommended, but this method is not recommended for arbitrary units.Histogram equalisation redistributes intensities to give bins similar voxel counts, while Lloyd-Max iteratively minimizes mean squared discretisation error.Source: Fixed bin size Q3RU · p. 24, Other methods · p. 25

Recommendations

  • Discretisation choice is modality-dependent and substantially affects intensity distributions, feature values, and reproducibility; Table 2.1 gives recommendations by intensity definition and method.The table distinguishes calibrated units such as CT and PET from arbitrary-unit data, and describes recommended combinations of re-segmentation range and fixed-bin methods.Source: Table 2.1 · p. 25, Recommendations · p. 25

2.8 Feature calculation · Image features · 3.1 Morphological features

This section presents IBSI image-feature reference values and standardization criteria, and defines morphological features that quantify ROI geometry and related properties. It describes their construction and the image-processing choices, conventions, units, and computational considerations that affect their calculation.

2.8 Feature calculation

  • Image-processing choices can change feature values, and texture optimization tunes interpolation and discretisation to an objective; lower grey-level counts may suit GLSZM, while higher counts may suit GLCM.The text notes that texture optimization may affect subsequent features’ prognostic capability.Source: §2.8 · p. 26

Image features

  • The IBSI feature set builds on earlier proposed sets, and intensity histogram and texture features require prior intensity discretisation, unlike some other feature families.Feature families include intensity-based statistical, histogram-based, intensity-volume histogram-based, morphological, local-intensity, and texture matrix-based features.Source: Image features · p. 27
  • Reference values were established for each feature, but only values reproducible within tolerance and supported by at least moderate consensus were standardized; weak-consensus or non-majority values should not be used.Consensus levels range from weak (<3 matches) to very strong (≥10 matches).Source: Image features · p. 27

Feature aggregation 5QB6

3.1 Morphological features

  • Morphological features describe ROI geometry; voxel-center points suit inner structure and meshes suit outer surfaces, while voxel representations can overestimate surface area and Marching Cubes is suggested for mesh generation.Mesh-based values may vary slightly between implementations because they depend on the meshing algorithm.Source: §3.1 · p. 30

ROI morphological and intensity masks

3.1.1 Volume (mesh)

  • For small ROIs of tens to hundreds of voxels, voxel counting overestimates volume relative to mesh-based volume; mesh-based volume and area are calculated from triangle faces, and surface-to-volume ratio is not dimensionless.Source: §3.1.1 · p. 31, §3.1.3 · p. 32, §3.1.4 · p. 33

3.1.5 Compactness 1

  • Compactness, spherical disproportion, sphericity, and asphericity describe deviation from a sphere and are highly correlated; other descriptors capture centre-of-mass shift, maximum 3D diameter, and PCA-derived axis lengths.Centre-of-mass shift reflects the spatial distribution of low- and high-intensity regions; the maximum diameter can be computed efficiently using the ROI convex hull.Source: §3.1.5 · p. 33, §3.1.9 · p. 35, §3.1.10 · p. 35

3.1.13–3.1.16 Axis features

  • PCA-derived axis lengths describe ROI extent, while inverse axis ratios encode elongation and flatness: 1 denotes non-elongated, non-flat shapes, and lower values indicate greater elongation or flatness.Minor and least axis lengths are twice the semi-axis lengths determined from the second-largest and smallest PCA eigenvalues, respectively.Source: §3.1.13 · p. 37, §3.1.14 · p. 37, §3.1.15 · p. 38

3.1.17–3.1.18 AABB densities

  • Volume density (AABB) compares ROI volume with its axis-aligned bounding-box volume, while area density (AABB) compares ROI surface area with the same box’s surface area.The box is the smallest axis-aligned box enclosing the ROI mesh vertex set; the mesh and convex-hull vertex sets produce an identical box.Source: §3.1.17 · p. 39, §3.1.18 · p. 39

3.1.19–3.1.20 OMBB densities

  • Oriented minimum bounding-box volume and area densities use the smallest-volume box enclosing the ROI, but both features currently lack reference values and should not be used.The oriented box may be approximated at lower computational complexity than the rotating-callipers method.Source: §3.1.19 · p. 40

3.1.21–3.1.22 AEE densities

  • Approximate-enclosing-ellipsoid (AEE) densities use an ellipsoid fitted from PCA of ROI voxel centers; its area estimate uses a Legendre series stopped by default at ν = 20.The ellipsoid volume is defined by its semi-principal axes as Vaee = 4πabc/3.Source: §3.1.21 · p. 40, §3.1.22 · p. 41

3.1.23–3.1.24 MVEE densities

  • Minimum-volume-enclosing-ellipsoid (MVEE) volume and area densities currently lack reference values and should not be used; the smallest enclosing ellipsoid is approximated because direct computation is usually unfeasible.The convex-hull mesh vertex set is recommended for the MVEE because it is sparser than the full vertex set.Source: §3.1.23 · p. 41, §3.1.24 · p. 42

3.1.25–3.1.26 Convex-hull densities

  • Convex-hull volume density compares ROI volume with convex-hull volume and is also called solidity 27,65; convex-hull area density uses the surface area of that same hull.The convex-hull volume and area are computed using the corresponding mesh-feature procedures.Source: §3.1.25 · p. 42, §3.1.26 · p. 42

3.1.27 Integrated intensity

  • Integrated intensity is the average ROI intensity multiplied by volume and, in 18F-FDG-PET, is often called total lesion glycolysis 75.The ROI intensity mask supplies the voxel count used in the feature definition.Source: §3.1.27 · p. 43

3.1.28–3.1.29 Spatial autocorrelation

  • Moran’s I and Geary’s C measure spatial autocorrelation; Geary’s C focuses on voxel intensity differences and is more sensitive to local autocorrelation, while both calculations scale as O(n^2).Moran’s I may require repeated ROI subsampling for tractability at a cost to accuracy; Geary’s C may likewise require approximation for large ROIs.Source: §3.1.28 · p. 43, §3.1.29 · p. 44

3.2 Local intensity features · 3.3 Intensity-based statistical features

Local intensity features summarize 3D neighborhoods around ROI voxels, while intensity-based statistical features describe the ROI’s intensity distribution without discretization. The section defines local and global intensity peaks alongside distribution, dispersion, and magnitude statistics, with 3D volume computation recommended.

3.2 Local intensity features; Aggregating features

  • Local features use ROI voxels only as centers, but their neighborhoods include voxels inside or outside the ROI; local features are calculated in 3D, not per slice.The same 3D-volume recommendation applies to intensity-based statistical features; computing them slice by slice and averaging is not recommended.Source: §3.2 · p. 45, Aggregating features · pp. 45, 47

3.2.1 Local intensity peak; 3.2.2 Global intensity peak

  • Local intensity peaks average a 1 cm^3 sphere centered on each maximum-intensity ROI voxel, while the global peak selects the highest such neighborhood mean across all ROI voxels.If the maximum occurs at multiple voxels, the local peak is computed for each and the highest retained; a spherical mean convolution filter can accelerate the global calculation.Source: §3.2.1 · p. 45, §3.2.2 · p. 45

3.3 Intensity-based statistical features

  • Intensity-based statistical features describe the distribution of ROI intensities without discretization, but are not meaningful when the intensity scale is arbitrary.The ROI voxel intensities are denoted Xgl.Source: §3.3 · p. 47

3.3.1 Mean intensity; 3.3.2 Intensity variance; 3.3.3 Intensity skewness

  • Mean, variance, skewness, and excess kurtosis characterize the ROI intensity distribution; variance is uncorrected, and skewness and kurtosis are set to zero when variance is zero.Excess kurtosis is Fisher-corrected by −3 to center normal distributions on zero.Source: §3.3.1 · p. 47, §3.3.2 · p. 47, §3.3.3 · p. 48

3.3.5 Median intensity; 3.3.6 Minimum intensity; 3.3.9 Maximum intensity

  • Median, minimum, maximum, and the 10th and 90th percentiles summarize intensity order, with the percentiles offering more robust alternatives to the extremes.The digital-phantom 90th percentile may vary by implementation, with some producing 4.2 rather than 4.Source: §3.3.5 · p. 49, §3.3.6 · pp. 49–50, §3.3.9 · p. 50

3.3.10 Intensity interquartile range; 3.3.11 Intensity range; 3.3.16 Intensity-based quartile coefficient of dispersion

  • The interquartile range measures the span from the 25th to 75th percentiles, while intensity range spans the minimum to maximum.The quartile coefficient of dispersion is presented as a more robust alternative to the intensity-based coefficient of variation.Source: §3.3.10 · pp. 50–51, §3.3.11 · p. 51, §3.3.16 · p. 53

3.3.12 Intensity-based mean absolute deviation; 3.3.13 Intensity-based robust mean absolute deviation; 3.3.14 Intensity-based median absolute deviation; 3.3.15 Intensity-based coefficient of variation

  • Mean and median absolute deviations measure dispersion around the mean and median, respectively; robust mean absolute deviation restricts intensities to the 10th–90th percentile interval to reduce outlier influence.The coefficient of variation measures dispersion using the intensity standard deviation and mean.Source: §3.3.12 · p. 51, §3.3.13 · p. 52, §3.3.14 · p. 52

3.3.17 Intensity-based energy; 3.3.18 Root mean square intensity

3.4 Intensity histogram features

The section defines intensity histogram features from discretised ROI intensities and their bin frequencies, with 3D-volume aggregation recommended over slice-wise averaging. It specifies statistical, distributional, dispersion, and histogram-gradient features, including implementation conditions that affect their interpretation.

Aggregating features

  • Histogram features are computed from discretised ROI intensities and bin frequencies, with occurrence probabilities approximated by each bin count divided by the voxel count; 3D aggregation is recommended over slice-wise averaging.The histogram is formed by discretising the original intensity distribution. The recommended approach computes features using the 3D volume rather than computing per slice and averaging.Source: §3.4 · p. 55, Aggregating features · p. 55

3.4.2 Discretised intensity variance

  • Mean, variance, skewness, and excess kurtosis summarize the discretised intensities; variance is uncorrected, while kurtosis uses a Fisher −3 correction and is set to 0 when variance is 0.Kurtosis measures distribution peakedness and is centered on 0 for normal distributions.Source: §3.4.1 · p. 55, §3.4.2 · p. 55, §3.4.3 · p. 56

3.4.10 Intensity histogram mode

  • The median, 10th and 90th percentiles, interquartile range, and mode characterize distribution location and quantiles; tied modes select the bin nearest the mean, then the left bin if equidistant.The 90th-percentile phantom reference may vary by implementation; some implementations yield 4.2 instead of 4.Source: §3.4.5 · p. 57, §3.4.6 · pp. 57–58, §3.4.10 · p. 59

3.4.6 Minimum discretised intensity

  • Minimum, maximum, and range describe the discretised intensity limits; for fixed-bin-number discretisation the minimum is 1 and the range equals Ng, while the maximum is Ng by definition.With fixed-bin-size discretisation, the minimum may exceed 1.Source: §3.4.6 · p. 57, §3.4.9 · p. 58, §3.4.12 · p. 59

3.4.14 Intensity histogram robust mean absolute deviation

  • Dispersion features include mean absolute deviation, robust mean absolute deviation restricted to intensities between the 10th and 90th percentiles, median absolute deviation, coefficient of variation, and the more robust quartile coefficient of dispersion.The robust mean absolute deviation addresses the potential effect of outliers; median absolute deviation measures dispersion from the median rather than the mean.Source: §3.4.13 · p. 60, §3.4.14 · p. 60, §3.4.15 · p. 61

3.4.18 Discretised intensity entropy

  • Shannon entropy is meaningful for discretised intensities, while uniformity approaches 1 when most intensities occupy one bin and has a lower bound of 1/Ng.Uniformity is sometimes called energy.Source: §3.4.18 · p. 62, §3.4.19 · pp. 62–63

3.4.20 Maximum histogram gradient

  • Histogram-gradient features include the maximum and minimum gradients and their corresponding discretised intensities; the histogram should retain zero-count bins to remain non-sparse.The described gradient method yields a gradient with the same size as the histogram, preserving correspondence between histogram bins and discretised intensities.Source: §3.4.20 · p. 63, §3.4.21 · p. 63, §3.4.22 · p. 64

3.5 Intensity-volume histogram features

The IVH describes the relationship between discretised intensity and the fraction of the ROI volume containing at least that intensity. Its calculation requires modality-appropriate discretisation, an intensity range, and an interval.

3.5 Intensity-volume histogram features

  • IVH features describe the relationship between discretised intensity and the volume fraction at or above that intensity, using a defined intensity range and discretisation interval.Depending on the modality, intensities may require discretisation before calculating the histogram.Source: §3.5 · p. 65

Discrete calibrated image intensities

  • For calibrated discrete intensities such as CT, no discretisation is needed; the range is the re-segmentation range or observed minimum-to-maximum, with interval 1.For a half-open re-segmentation range, its upper limit is max(Xgl).Source: Discrete calibrated image intensities · p. 65

Continuous calibrated image intensities

  • For calibrated continuous intensities such as PET, fixed bin size discretisation is recommended, using the re-segmentation bounds or observed extrema and a modality-dependent bin width.Replace bin numbers with bin-center intensities; then G = [Xgl,min + 0.5wb, Xgl,max − 0.5wb] and wd = wb. The bin width should be small relative to the intensity range; 0.10 SUV is given as an 18F-FDG-PET example.Source: Continuous calibrated image intensities · p. 65

Arbitrary intensity units

  • For arbitrary intensities such as many MRI sequences, fixed bin number discretisation with Ng = 1000 is recommended, yielding Xd,gl = Xd, G = [1, Ng], and wd = 1.The bin width is (Xgl,max − Xgl,min) /Ng; re-segmentation ranges generally cannot be used for non-calibrated intensities.Source: Arbitrary intensity units · p. 65

Calculating the IV histogram

  • The IVH uses fractional volumes and intensities across the full range G, including discretised intensity levels absent from the ROI.Fractional volume is the fraction of voxels with intensity at least i; the recommended features use the 3D volume, not per-slice calculations subsequently averaged.Source: Calculating the IV histogram · p. 66, Aggregating features · p. 66

3.5.1 Volume at intensity fraction; 3.5.2 Intensity at volume fraction

  • Fivh.V10 and Fivh.V90 are the largest volume fractions with intensity fractions of at least 10% and 90%, while Fivh.I10 and Fivh.I90 are minimum intensities present in at most 10% and 90% of volume.Source: §3.5.1 · p. 66, §3.5.2 · p. 67

3.5.3 Volume fraction difference between intensity fractions; 3.5.4 Intensity fraction difference between volume fractions

  • Fivh.V10minusV90 is the difference between volume fractions at two intensity fractions, and Fivh.I10minusI90 is the difference between discretised intensities at two volume fractions.Source: §3.5.3 · p. 68, §3.5.4 · p. 68

3.5.5 Area under the IVH curve

  • Fivh.auc is approximated by the trapezoidal-rule Riemann sum, is defined as 0 when the ROI has only one discretised intensity, and currently has no reference values and should not be used.Source: §3.5.5 · p. 68

3.6 Grey level co-occurrence based features

This section defines grey level co-occurrence matrix (GLCM) texture features from neighbouring discretised grey levels, including directional aggregation and distance handling. It reports reference values for a range of GLCM features across digital phantoms and configuration-based aggregation schemes, with consensus generally strong or very strong.

3.6 Grey level co-occurrence based features

  • A GLCM records neighbouring grey-level pair frequencies in an Ng × Ng matrix; summing opposite-direction counts makes it symmetric, and normalising yields joint probabilities.Source: §3.6 · p. 69

Aggregating features

  • Six aggregation schemes compute features from directional matrices by averaging across directions or slices, or merging matrices before feature calculation.The schemes include 2D and 3D approaches; aggregation is intended to improve rotational invariance, and feature values may depend strongly on the method.Source: Aggregating features · p. 70

Distances and distance weighting

  • Distance weighting can have no effect under several neighbourhood and norm conditions, so the section recommends against using it for GLCM.The default neighbourhood includes voxels within Chebyshev distance 1; alternative direction-vector norms are rare and cautioned against for reproducibility reasons.Source: Distances and distance weighting · p. 71

3.6.1 Joint maximum

  • 0.519 (digital phantom 2D, averaged; very strong) is a joint maximum reference value.Joint maximum is the probability of the most common grey-level co-occurrence.Source: §3.6.1 · pp. 71–72, Table 3.80 · p. 73

3.6.3 Joint variance

  • 73.8 (config. C 3D, merged; tol. 2; very strong) is a joint variance reference value.Joint variance is also called sum of squares and is defined using the joint average.Source: §3.6.3 · pp. 73, 75, Table 3.82 · p. 74

3.6.6 Difference variance

  • 3.06 (digital phantom 3D, averaged; very strong) is a difference variance reference value, computed from diagonal probabilities around the difference average.Source: §3.6.6 · p. 76
  • Difference variance benchmarks include 3.21 for digital-phantom 3D merged and 14.4 (0.5) for config. C 3D averaged and merged; all are reported with very strong or strong consensus.Table 3.85 is identified as the reference-value table for this feature.Source: §3.6.6 · p. 76, Table 3.85 · p. 76

3.6.7 Difference entropy

  • Difference entropy benchmarks span digital-phantom 2D averaged at 1.4 (very strong) and config. C 3D merged at 2.64 (0.03, very strong).Table 3.86 is identified as the reference-value table for this feature.Source: §3.6.7 · pp. 76–77, Table 3.86 · p. 77

3.6.10 Sum entropy

  • Sum entropy benchmarks include 1.79 for digital-phantom 2D slice-merged and 4.56 (0.04) for config. C 3D averaged and merged, with strong or very strong consensus.Table 3.89 is identified as the reference-value table for this feature.Source: §3.6.10 · p. 79, Table 3.89 · p. 79

3.6.11 Angular second moment

  • Angular second moment benchmarks include 0.368 for digital-phantom 2D averaged (very strong) and 0.0447 (0.001) for config. C 3D merged (very strong).This feature is also called energy and uniformity; Table 3.90 is identified as its reference-value table.Source: §3.6.11 · pp. 79–80, Table 3.90 · p. 80

3.6.13 Dissimilarity

  • Dissimilarity benchmarks include 1.42 for digital-phantom 2D averaged and 2.17 (0.05) for config. C 3D averaged, both with strong consensus.The text identifies dissimilarity as conceptually similar to contrast and equivalent by definition to difference average.Source: §3.6.13 · p. 81
  • 0.628 (tolerance 0.001; strong) was reported for configuration B 2.5D direction-merged, alongside 0.966 (0.001; very strong) for configuration C 3D merged and 0.996 (0.001; strong) for configuration A 2.5D direction-merged.These are configuration-specific reference results; the passages do not identify which individual feature each value belongs to.Source: §3.6.13 · p. 81, §3.6.17 · p. 84

3.6.19 Correlation

  • Correlation can be simplified using the symmetry of P∆, with its calculation expressed through row and column marginal probabilities.The passage also gives an equivalent formulation of correlation and notes that symmetry again enables simplification.Source: §3.6.19 · pp. 85–86

3.6.20 Autocorrelation

  • Soh and Tsatsoulis [63] defined autocorrelation, and Table 3.99 provides reference values for the feature.The supplied passages do not include the table's numerical values.Source: §3.6.20 · p. 87, Table 3.99 · p. 87

3.6.21 Cluster tendency

  • Cluster tendency is mathematically equal to the sum variance feature [81].The section also provides cluster tendency reference values, including 276 (tolerance 8; very strong) for configuration C 3D merged.Source: §3.6.21 · p. 87, Table 3.100 · p. 88
  • 79.1 (very strong) was reported for digital phantom 2D averaged cluster shade, versus 147 (strong) for 2.5D direction-merged.Table 3.101 is identified as the reference-values table for cluster shade.Source: §3.6.23 · p. 89, Table 3.101 · p. 89
  • 5.7 × 10^5 (tolerance 1.1 × 10^4; very strong) was reported for configuration C 3D merged cluster prominence.Table 3.102 is identified as the reference-values table for cluster prominence.Source: §3.6.23 · p. 89, Table 3.102 · p. 90
  • Table 3.103 is identified as reporting reference values for the information correlation 1 feature.Source: Table 3.103 · p. 90

3.6.25 Information correlation 2

  • The second information-theoretic correlation measure is estimated for symmetric PΔ.Source: §3.6.25 · p. 91
  • Digital phantom values were 0.487 for 2D averaged, 0.263 for 2D slice-merged, 0.487 for 2.5D direction-merged, and 0.291 for 2.5D merged.For 3D, averaged was 0.52 and merged was 0.269; the 2D/2.5D entries are marked strong and the 3D entries very strong.Source: §3.6.25 · p. 91
  • Configuration A 2D averaged 0.863 and slice-merged 0.851; 2.5D direction-merged 0.879 and merged 0.88.The reported tolerances are 0.003, 0.002, 0.001, and 0.001, respectively; each is marked strong.Source: §3.6.25 · p. 91, Table 3.104
  • Configuration B 2D averaged 0.837 and slice-merged 0.792; 2.5D direction-merged 0.821 and merged 0.819.All four are marked strong, with reported tolerances of 0.001.Source: §3.6.25 · p. 91, Table 3.104
  • Configuration C 3D averaged 0.9 and merged 0.899; configuration D 3D averaged 0.845 and merged 0.846.The reported tolerances are 0.001 for C and 0.003 for D; both are marked strong.Source: §3.6.25 · p. 91, Table 3.104
  • Configuration E 3D averaged 0.813 (moderate) and merged 0.813 (strong), both with tolerance 0.004.Source: §3.6.25 · p. 91, Table 3.104

3.7 Grey level run length based features

This section describes grey level run length matrix features, which quantify discretised grey levels across consecutive runs along image directions, with directional values averaged or matrices merged. It reports reference values and consensus ratings for selected normalized non-uniformity, variance, and entropy features using digital phantoms and specified configurations and aggregation methods.

Aggregating features

  • Directional GLRLM features can be averaged across directions or calculated from merged matrices, and their values may depend strongly on the aggregation method.The described approaches include averaging 2D matrices across directions and slices and merging all 3D directional matrices.Source: Aggregating features · p. 92

Distance weighting

  • Distance weighting multiplies run lengths by a weight, but is non-standard and cautioned against because it may cause reproducibility issues.The default is w = 1; inverse-distance and exponential weights are examples, and weighting has no effect with Chebyshev norm or before matrix merging.Source: Distance weighting · p. 92

3.7.2 Long runs emphasis

  • Long runs emphasis is defined to emphasise long run lengths.The supplied passage introduces its definition but does not provide the formula.Source: §3.7.2 · p. 94

3.7.3 Low grey level run emphasis

  • Low grey level run emphasis is the grey-level analogue of short runs emphasis, emphasising low rather than short run lengths.The supplied passage introduces its definition but does not provide the formula.Source: §3.7.3 · p. 95

3.7.4 High grey level run emphasis

  • 0.294 (digital phantom 2D, averaged; very strong) and 0.372 (digital phantom 3D, merged; very strong) are reference values for high grey level run emphasis.This feature is described as the grey-level analogue of long runs emphasis, emphasising high grey levels.Source: §3.7.4 · pp. 95–96, Table 3.109 · p. 96

3.7.7 Long run low grey level emphasis

  • Long run low grey level emphasis targets runs in the upper-right GLRLM quadrant, where long run lengths and low grey levels occur.The feature definition is attributed to marker 22 in the supplied passage.Source: §3.7.7 · p. 97

3.7.8 Long run high grey level emphasis

  • 17.4 (digital phantom 2D, averaged; very strong) and 15.6 (digital phantom 3D, merged; very strong) are reference values for long run high grey level emphasis.This feature targets the lower-right GLRLM quadrant, where long run lengths and high grey levels occur; the supplied passage attributes the definition to marker 22.Source: §3.7.8 · p. 98, Table 3.113 · p. 99

3.7.10 Normalised grey level non-uniformity

  • 0.46 (digital phantom 2D, averaged; very strong) and 0.43 (digital phantom 3D, averaged; very strong) are reference values for normalised grey level non-uniformity.The feature is described as a normalised version of grey level non-uniformity.Source: Table 3.114 · p. 99, §3.7.10 · pp. 100–101
  • Config. C 3D averaged and merged both yielded 0.102 with tolerance 0.003 for normalised grey level non-uniformity, with strong and very strong consensus, respectively.Source: §3.7.10 · p. 100, Table 3.115 · p. 100

3.7.12 Normalised run length non-uniformity

  • The normalised run length non-uniformity reference value was 0.513 for the digital phantom 3D averaged method, with very strong consensus.The feature is described as the normalised version of run length non-uniformity.Source: §3.7.12 · p. 101, Table 3.117 · p. 102

3.7.14 Grey level variance

  • Grey-level variance reference values included 3.35 for the digital phantom 2D averaged method (very strong consensus) and 101 with tolerance 3 for config. C 3D averaged (strong).Source: §3.7.14 · p. 103, Table 3.119 · p. 103
  • Run-length variance reference values included 2.17 for the digital phantom 2D averaged method and 2.62 for the 3D merged method, both with very strong consensus.Source: §3.7.14 · p. 104, Table 3.120 · p. 104
  • Run entropy was 5.35 with tolerance 0.03 for config. C 3D averaged and merged; consensus was strong and very strong, respectively.Source: §3.7.14 · p. 104, Table 3.121 · p. 105

3.8 Grey level size zone based features

GLSZM features quantify zones of linked voxels sharing a discretised grey level, with results depending on matrix aggregation and neighbourhood choices. The section defines zone-emphasis and distribution features and provides reference values for several configurations.

3.8 Grey level size zone based features

  • The GLSZM counts groups of linked voxels with identical discretised grey levels; 3D uses 26-connectedness and 2D uses 8-connectedness, which can split cross-slice zones.Whether this 2D issue affects predictive performance or reproducibility is undetermined.Source: §3.8 · p. 106

Distances

  • The default GLSZM neighbourhood uses Chebyshev distance δ = 1; alternative norms or larger distances are non-standard and cautioned against because of reproducibility risks.Manhattan and Euclidean norms yield 6-connected 3D and 4-connected 2D neighbourhoods; larger distances can join separate same-intensity zones.Source: Distances · p. 106

Aggregating features

  • GLSZM features can be averaged across 2D slices, computed after merging slice matrices, or computed from a 3D matrix; values may depend strongly on aggregation.When merging matrices, voxel counts Nv should also be summed; Figure 3.4 illustrates the approaches.Source: Aggregating features · p. 106, Figure 3.4 · p. 107

3.8.1 Small zone emphasis; 3.8.2 Large zone emphasis; 3.8.4 High grey level zone emphasis

3.8.7 Large zone low grey level emphasis

  • 503 (digital phantom, 3D) and 241 (config. D, 3D) are reference values for large zone low grey level emphasis.The corresponding reported tolerances are — and 14, respectively; both have strong consensus.Source: §3.8.7 · p. 110, Table 3.129 · p. 111

3.8.9 Grey level non-uniformity; 3.8.10 Normalised grey level non-uniformity

  • Grey level non-uniformity measures zone-count distribution across grey values, while normalised grey level non-uniformity is its normalised version.Config. A 2D has grey level non-uniformity 82.2 (tolerance 0.1; strong consensus).Source: §3.8.9 · p. 111, §3.8.10 · p. 112

3.8.12 Normalised zone size non-uniformity; 3.8.13 Zone percentage; 3.8.14 Grey level variance

3.8.15 Zone size variance

  • Zone size variance estimates variance in zone counts across zone sizes; the digital phantom 3D reference value is 331 (very strong consensus).Source: §3.8.15 · p. 114

3.9 Grey level distance zone based features

This section describes grey level distance zone-based features, including their relationship to discretised grey levels and distance from the ROI edge, distance calculation and aggregation options. It also provides reference values for zone distance variance and zone distance entropy.

3.9 Grey level distance zone based features

  • The GLDZM counts linked voxels sharing a discretised grey level and distance to the ROI edge; the intensity mask defines zones and the morphological mask defines distance.Voxel linkage uses 26-connectedness in 3D or 8-connectedness in 2D.Source: §3.9 · p. 116
  • Distance uses 6-connectedness in 3D and 4-connectedness in 2D, with each zone assigned the minimum voxel distance; the lowest distance is 1 to avoid division by zero.Unlike the original definition, the method iteratively erodes the ROI and increments distances until the mask is empty, avoiding reliance on Chamfer maps for non-rectangular ROIs.Source: §3.9 · p. 116

Aggregating features; Distances

  • Features can be computed per 2D slice and averaged, computed after merging 2D matrices, or computed from a 3D matrix; non-default distance norms may reduce efficiency or reproducibility.Merging requires summing zone counts and voxel counts; feature values may depend strongly on aggregation. Manhattan distance is the efficient default, while Euclidean distance requires nearest-non-ROI searches and rounding.Source: Aggregating features · p. 117, Distances · p. 117

3.9.1 Small distance emphasis

  • Small distance emphasis reference values include dig. phantom 2D 0.946 and config. B 2D 0.36.The digital phantom 2D value has strong consensus; config. B 2D has tolerance 0.005 and strong consensus.Source: §3.9.1 · pp. 117–118, Table 3.140

3.9.2 Large distance emphasis

  • Large distance emphasis reference values include dig. phantom 2D 0.371 and config. D 3D 0.0409.The digital phantom 2D value has strong consensus; config. D 3D has tolerance 0.0005 and strong consensus.Source: Table 3.141 · p. 118, §3.9.3 · pp. 118–119

3.9.4 High grey level zone emphasis; 3.9.7 Large distance low grey level emphasis

  • High grey level zone emphasis has config. D 3D 0.0302, while large distance low grey level emphasis has config. C 3D 0.0154.Both config. D and config. C values have strong consensus; their respective tolerances are 0.0006 and 0.0005.Source: Table 3.143 · p. 119, §3.9.5 · p. 120, Table 3.146 · p. 121

3.9.9 Grey level non-uniformity; 3.9.10 Normalised grey level non-uniformity

3.9.13 Zone percentage; 3.9.15 Zone distance variance

  • Zone percentage measures realised zones against the maximum potential zones, while zone distance variance estimates the variance in zone counts across distances.Highly uniform ROIs produce a low zone percentage; zone distance variance uses p_ij = d_ij/N_s.Source: §3.9.13 · p. 123, §3.9.15 · p. 124

3.9.15 Zone distance variance

  • Table 3.154 presents reference values for the zone distance variance feature.Source: Table 3.154 · p. 125
  • Table 3.155 presents reference values for the zone distance entropy feature.Source: Table 3.155 · p. 125

3.10 Neighbourhood grey tone difference based featuresIPET

NGTDM features summarize differences between discretised voxel intensities and the average intensities of their neighbours. The section defines ROI-aware neighbourhoods and feature calculations, describes aggregation and distance options, and gives reference values for selected features.

3.10 Neighbourhood grey tone difference based featuresIPET

  • NGTDM records each discretised grey level’s summed difference from neighbouring intensities; its 3D neighbourhood uses Chebyshev distance δ and excludes the centre voxel.In 2D, neighbourhoods are calculated within slices, with W = (2δ + 1)^2 − 1; in 3D, W = (2δ + 1)^3 − 1.Source: §3.10 · p. 126

Aggregating features

  • Aggregation can average 2D feature values across slices (8QNN), merge 2D matrices (62GR), or calculate from a 3D matrix (KOBO).For merged matrices, sum s_i and n_i across slices and update N_v,c and p_i; feature values may depend strongly on the aggregation method.Source: Aggregating features · p. 127

Distances and distance weighting

  • Chebyshev distance is the default, while Manhattan or Euclidean norms and distance weighting are possible but considered non-standard because of reproducibility concerns.Weights may depend on neighbour displacement, for example w = ||m||^-1 or w = exp(−||m||^2).Source: Distances and distance weighting · pp. 127–128

3.10.1 Coarseness

  • Coarseness reflects small grey-level differences in coarse textures, and its maximum is set arbitrarily to 10^6 when the denominator is zero.This explicit maximum replaces the original unspecified small ε in the denominator.Source: §3.10.1 · p. 128

3.10.2 Contrast

  • Contrast combines grey-level dynamic range with the degree of intensity change, and is zero when only one grey level has non-zero probability.The dynamic-range and intensity-change terms use copies of the grey-level probabilities with separate iterators.Source: §3.10.2 · p. 129

3.10.3 Busyness

  • Busyness references were 2.99 for digital phantom 2D (strong), 6.8 for 2.5D (strong), and 6.54 for 3D (very strong).These are reference values for the busyness feature.Source: §3.10.3 · p. 129, Table 3.159 · p. 130
  • Busyness references were 1.39 for config. C (tol. 0.01; very strong), 5.14 for D (0.14; strong), and 4.65 for E (0.1; strong).These are 3D reference values for the busyness feature.Source: §3.10.3 · p. 130, Table 3.159 · p. 130

3.10.5 Strength

  • Strength references were 2.88 for digital phantom 2D (strong), 0.741 for 2.5D (strong), and 0.763 for 3D (very strong).These are reference values for the strength feature.Source: §3.10.5 · p. 130, Table 3.161 · p. 131

3.11 Neighbouring grey level dependence based featuresREK0

The section defines neighbouring grey level dependence features from voxel neighbourhoods and describes their aggregation, distance options, feature families, and benchmark values. Its implementation uses every ROI voxel as a centre, and the entropy definition corrects an error in Sun and Wee 69.

3.11 Neighbouring grey level dependence based featuresREK0

  • NGLDM counts neighbours within Chebyshev distance δ whose discretised grey levels differ from the centre by at most α; all ROI voxels are centres, and outside-ROI voxels add no dependence.NGLDM captures overall texture coarseness and is rotationally invariant; α = 0 and δ = 1 are typical choices, but alternatives are possible.Source: §3.11 · p. 132

Aggregating features

  • Features can use slice-wise 2D matrices, merged 2D matrices, or a 3D matrix; merging sums matrix elements and voxel/neighbourhood counts, while aggregation affects feature values.Other distance norms and distance weighting are considered non-standard and cautioned against because of potential reproducibility issues.Source: Aggregating features · pp. 132–133

Note on feature references

3.11.1 Low dependence emphasis; 3.11.2 High dependence emphasis; 3.11.3 Low grey level count emphasis; 3.11.4 High grey level count emphasis

  • The emphasis features cover low or high dependence counts, low or high grey levels, and combinations of low/high dependence with low/high grey levels.Low and high dependence emphasis are also called small and large number emphasis by Sun and Wee 69.Source: §3.11.1 · p. 134, §3.11.2 · p. 134, §3.11.3 · p. 135

3.11.5 Low dependence low grey level emphasis; 3.11.6 Low dependence high grey level emphasis; 3.11.7 High dependence low grey level emphasis; 3.11.8 High dependence high grey level emphasis

3.11.9 Grey level non-uniformity; 3.11.10 Normalised grey level non-uniformity; 3.11.11 Dependence count non-uniformity; 3.11.12 Normalised dependence count non-uniformity; 3.11.13 Dependence count percentage

  • Grey-level and dependence-count non-uniformity assess distribution across grey levels and dependence counts, respectively; normalized grey-level non-uniformity equals intensity histogram uniformity for a single 3D matrix.Dependence count percentage evaluates to 1 under this definition, so it may be omitted.Source: §3.11.9 · p. 138, §3.11.10 · p. 138, §3.11.11 · p. 139

3.11.14 Grey level variance

  • Grey level variance reference values were 2.7 for the digital phantom with 2D aggregation, 3.05 with 2.5D aggregation, and 3.05 with 3D aggregation.The feature estimates variance in dependence counts over grey levels.Source: §3.11.14 · p. 140

3.11.16 Dependence count entropy

  • The dependence-count entropy definition remedies Sun and Wee 69’s error of using dependence count sij rather than count probability pij inside the logarithm.Digital-phantom entropy reference values were 2.71 with 2D, 3.36 with 2.5D, and 4.4 with 3D aggregation.Source: §3.11.16 · p. 141

Radiomics reporting guidelines and nomenclature

Reliable, complete reporting is necessary for reproducibility and validation; the section presents reporting guidelines and a nomenclature system to support complete image-processing and biomarker-extraction reports.

Radiomics reporting guidelines and nomenclature

4.1 Reporting guidelines

The reporting guidelines specify information to document across imaging and image-processing workflows, drawing partly on prior work and the steps described in the document. They cover re-segmentation, discretisation, filtering and biomarker computation, as well as the biomarker set, software, IBSI compliance, robustness assessment and texture-matrix computation.

Reporting guidelines

  • Report the region of interest and patient preparation, including pre-acquisition instructions and drugs administered.The guidelines also call for reporting equipment used for patient comfort during scanning.Source: §4.1 · p. 143
  • For PET and SPECT, report the radioactive tracer, administration method, and injected activity.The guidelines also ask authors to describe uptake time and control of competing substance levels.Source: §4.1 · p. 143
  • Document the acquisition protocol, scanner type and vendor, imaging modality, scan timing, calibration, patient instructions, motion handling, and scan duration.For dynamic scans, report acquisition time per frame and any temporal modelling technique.Source: §4.1 · p. 144
  • Report modality-specific acquisition parameters, including CT tube voltage and current and MRI sequence, repetition time, echo time, and inversion time.The guidelines also cover additional MRI sequence options and other acquisition settings.Source: §4.1 · pp. 144–145
  • Describe data conversions and post-acquisition processing, including SUV normalisation, ADC computation, other conversions, anti-aliasing, and intensity normalisation.The list also includes noise suppression, PET spatial smoothing, MRI skull stripping and non-uniformity correction, and other processing methods.Source: §4.1 · p. 146
  • Specify segmentation methods and settings, including manual-expert consensus, automated approaches, the image used for multimodal segmentation, and interpolation choices.The guidelines also cover converting polygonal or mesh segmentations to voxel masks and handling partially masked voxels after ROI interpolation.Source: §4.1 · pp. 146–147

Re-segmentation methods

  • Report the methods and settings used to re-segment the ROI intensity mask.Source: §4.1 · p. 147

Image processing - discretisation

  • Describe the discretisation method, bin count or size, and lowest intensity in the first bin for fixed-bin-size discretisation.If discretisation is performed separately to create intensity-volume histograms, describe that as well.Source: §4.1 · p. 147

Image processing - image transformation

  • Report image-filtering methods and settings, such as Laplacian-of-Gaussian.IBSI has not added image transformation to its standardised processing scheme and is benchmarking common filters, so this section may expand in the future.Source: §4.1 · p. 147

Image biomarker computation

  • Specify the computed biomarker set and its definitions, software and version, and how biomarker robustness was assessed.State whether the software reproduces IBSI reference values; compliance requires reproducing values for the digital phantom and at least one radiomics CT phantom processing configuration, and reviewers may request the compliance spreadsheet.Source: §4.1 · p. 147

Image biomarker computation - texture parameters

  • Define how texture-matrix-based biomarkers were computed from the underlying texture matrices.Source: §4.1 · p. 147

4.2 Feature nomenclature

IBSI proposes a concise but complete feature nomenclature: a feature name is paired with a subscript that adds only details needed to resolve ambiguity. Suggested abbreviations specify feature families and, where relevant, aggregation, imaging and processing choices, discretisation, and feature-specific parameters.

4.2 Feature nomenclature

  • The subscript requires a feature-family abbreviation and may add aggregation, modality or other data descriptors, and processing settings when needed to distinguish otherwise ambiguous features.For example, the two entropy features are denoted entropyIH, CT, FBS:25HU and entropyRLM, PET, FBN:32; optional descriptors are omitted when no ambiguity exists, although the details must still be reported.Source: §4.2 · pp. 148–149

4.2.1 Abbreviating feature families

  • The suggested feature-family abbreviations include MORPH for morphology, IS or STAT for intensity statistics, IH for intensity histogram, and GLCM or CM for co-occurrence matrix.The list also assigns abbreviations to intensity-volume histogram and other texture-matrix families.Source: §4.2.1 · p. 149

4.2.3 Abbreviating interpolation

  • Interpolation notation identifies the method and can include dimensionality and uniform-grid spacing, with these attributes omissible when clear from context.Suggested forms include NNB:dim:# for nearest-neighbour, LIN:dim:# for linear, and CSI:dim:# or SI3:dim:# for cubic spline interpolation.Source: §4.2.3 · p. 150

4.2.4 Describing re-segmentation

  • Re-segmentation notation represents ranges and outlier filtering, allows half-open ranges, and can combine range and filtering methods.Examples include RS:[-200,150] for a CT range and RS:3σ for removing T1-weighted MR ROI voxels outside µ ± 3σ.Source: §4.2.4 · p. 151

4.2.5 Abbreviating discretisation

  • Discretisation notation identifies the method and its parameter, including fixed bin size (FBS:#), fixed bin number (FBN:#), or histogram equalisation (EQ:#).Histogram equalisation may precede the listed discretisation algorithms, and its parameter gives the number of bins in the equalised histogram.Source: §4.2.5 · p. 151

4.2.6 Abbreviating feature-specific parameters

  • Feature-specific parameter notation covers co-occurrence symmetry, distance, and weighting, with asymmetrical co-occurrence matrices not recommended.The listed conventions include SYM for symmetrical matrices, ASYM for asymmetrical matrices, δ:# for Chebyshev distance, and w:f for weighting by function f.Source: §4.2.6 · p. 152
  • Matrix-family parameters also specify Chebyshev distance for size-zone and distance-zone features, while neighbourhood tone and dependence features allow distance and weighting notation.No weighting is the default where listed, and a function can be specified using w:f.Source: §4.2.6 · p. 152

Reference data sets

Reference feature values were derived from a digital image phantom and a lung cancer patient’s CT image. These data sets can also verify radiomics software implementations.

Reference data sets

  • Feature reference values came from a digital image phantom and a lung cancer patient’s CT image, and both data sets can verify radiomics software implementations.The data sets are available from the IBSI GitHub repository.Source: Reference data sets · p. 154

5.1 Digital phantom

The digital phantom was designed for manual image-feature calculation and comparison with radiomics software implementations. It also specifies the conditions for calculating features directly from the phantom.

5.1 Digital phantom

  • The phantom contains 5 × 4 × 4 voxels, with (x, y) slices stacked along z.Source: §5.1 · p. 154
  • Its ROI excludes several outer voxels and one internal voxel; grey levels 2 and 5 are absent, while 1 and 6 are the ROI’s lowest and highest levels.Source: §5.1 · p. 154

5.1.1 Computing image features

  • Features are calculated directly without preprocessing or required discretisation; fixed bin size 1 or six bins would leave the phantom unchanged.Source: §5.1.1 · p. 154
  • Grey level co-occurrence matrices are symmetric at Chebyshev distance δ = 1; neighbouring grey level dependence and neighbourhood grey tone difference matrices use the same distance, with coarseness α = 0.Source: §5.1.1 · pp. 154–155
  • Without discretisation, most intensity-based statistical features match their intensity histogram-based analogues, and the ROI morphological and intensity masks are identical because there is no re-segmentation.Source: §5.1.1 · p. 155

5.2 Lung cancer CT image

The IBSI uses CT images from a public four-patient non-small-cell lung carcinoma dataset, selecting patient PAT1 to provide reference feature values across image-processing configurations. The section outlines configuration choices, interpolation considerations, diagnostic descriptors, and feature-calculation conditions.

5.2 Lung cancer CT image

  • Five configurations compare 2D in-slice with 3D volumetric processing, testing fixed-bin-number and fixed-bin-size discretisation.For 2D configurations A and B, interpolation and texture extraction are in-slice; configurations C–E use three-dimensional interpolation and volumetric feature extraction.Source: §5.2 · p. 155

5.2.1 Notes on interpolation

  • Implementation-dependent interpolation grids can change feature values; check DICOM-origin alignment, ceiling-sized centered grids, boundary padding, and full-precision image and ROI masks.The interpolation grid should be at least as large as the original, with boundary voxels padded using nearest-boundary intensity; 32-bit precision helps avoid rounding errors, including unintended ROI voxel exclusion.Source: §5.2.1 · pp. 155, 157

5.2.2 Diagnostic features

  • Diagnostic features track image and ROI changes through processing, with reference values available in the IBSI compliance check spreadsheet.Descriptors cover initial and interpolated images, and initial, interpolated, and re-segmented ROIs; they include dimensions, voxel dimensions, intensity summaries, mask bounding boxes, and voxel counts.Source: §5.2.2 · pp. 157–158

5.2.3 Computing image features

  • Feature calculations use symmetric GLCMs at Chebyshev distance δ = 1 and NGLDM/NGTDM at δ = 1, with NGLDM α = 0; discretisation and re-segmentation affect feature values and masks.Intensity-based statistical features can differ from histogram-based analogues after discretisation. For IVH, configurations A, B, and D use calibrated discrete intensities without separate discretisation, whereas C and E re-discretise ROI intensities for their respective test cases.Source: §5.2.3 · p. 158

Appendix A … A.16 Neighbouring grey level dependence matrix (3D)

Appendix A provides digital-phantom texture matrices for reference. Its tables document 2D and volumetric 3D extractions, including merged variants and stated extraction settings.

Appendix A; Digital phantom texture matrices

A.1 Grey level co-occurrence matrix (2D); A.2 Grey level co-occurrence matrix (2D, merged); A.3 Grey level co-occurrence matrix (3D); A.4 Grey level co-occurrence matrix (3D, merged)

A.5 Grey level run length matrix (2D); A.6 Grey level run length matrix (2D, merged); A.7 Grey level run length matrix (3D); A.8 Grey level run length matrix (3D, merged)

A.9 Grey level size zone matrix (2D); A.10 Grey level size zone matrix (3D); A.11 Grey level distance zone matrix (2D); A.12 Grey level distance zone matrix (3D)

A.13 Neighbourhood grey tone difference matrix (2D); A.14 Neighbourhood grey tone difference matrix (3D); A.15 Neighbouring grey level dependence matrix (2D); A.16 Neighbouring grey level dependence matrix (3D)

Related papers

Loading 1612.07003v11…