You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Reliable and complete reporting is necessary to ensure reproducibility and validation of results. To help provide a complete report on image processing and image biomarker extraction, we present the guidelines below, as well as a nomenclature system to uniquely features.
Reporting guidelines
These guidelines are partially based on the work of Sollini2017N,Lambin2017,Sanduleanu2018iu,Traverso2018yr. Additionally, guidelines are derived from the image processing and feature calculation steps described within this document. An earlier version was reported elsewhere vallieres2017responsible.
Describe the methods and settings used to filter images, e.g. Laplacian-of-Ga ussian.
Image biomarker computation
Topic
Modality
Item
Description
Biomarker set
58
Describe which set of image biomarkers is computed and refer to their definitions or provide these.
IBSI compliance
59
State if the software used to extract the set of image biomarkers is compliant with the IBSI benchmarks. 11
Robustness
60
Describe how robustness of the image biomarkers was assessed, e.g. test-retest analysis.
Software availability
61
Describe which software and version was used to compute image biomarkers.
Image biomarker computation - texture parameters
Topic
Modality
Item
Description
Texture matrix aggregation
62
Define how texture-matrix based biomarkers were computed from underlying texture matrices.
Distance weighting
63
Define how CM, RLM, NGTDM and NGLDM weight distances, e.g. no weighting.
CM symmetry
64
Define whether symmetric or asymmetric co-occurrence matrices were computed.
CM distance
65
Define the (Chebyshev) distance at which co-occurrence of intensities is determined, e.g. 1.
SZM linkage distance
66
Define the distance and distance norm for which voxels with the same intensity are considered to belong to the same zone for the purpose of constructing an SZM, e.g. Chebyshev distance of 1.
DZM linkage distance
67
Define the distance and distance norm for which voxels with the same intensity are considered to belong to the same zone for the purpose of constructing a DZM, e.g. Chebyshev distance of 1.
DZM zone distance norm
68
Define the distance norm for determining the distance of zones to the border of the ROI, e.g. Manhattan distance.
NGTDM distance
69
Define the neighbourhood distance and distance norm for the NGTDM, e.g. Chebyshev distance of 1.
NGLDM distance
70
Define the neighbourhood distance and distance norm for the NGLDM, e.g. Chebyshev distance of 1.
NGLDM coarseness
71
Define the coarseness parameter for the NGLDM, e.g. 0.
Machine learning and radiomics analysis
Topic
Modality
Item
Description
Diagnostic and prognostic modelling
72
See the TRIPOD guidelines for reporting on diagnostic and prognostic modelling.
Comparison with known factors
73
Describe where performance of radiomics models is compared with known (clinical) factors.
Multicollineari ty
74
Describe where the multicollineari ty between image biomarkers in the signature is assessed.
Model availability
75
Describe where radiomics models with the necessary pre-processing information may be found.
Data availability
76
Describe where imaging data and relevant meta-data used in the study may be found.
Feature nomenclature
Image features may be extracted using a variety of different settings, and may even share the same name. A feature nomenclature is thus required. Let us take the example of differentiating the following features: i) intensity histogram-based entropy, discretised using a fixed bin size algorithm with 25 HU bins, extracted from a CT image; and ii) grey level run length matrix entropy, discretised using a fixed bin number algorithm with 32 bins, extracted from a PET image. To refer to both as entropy would be ambiguous, whereas to add a full textual description would be cumbersome. In the nomenclature proposed below, the features would be called entropyIH, CT, FBS:25HU and entropyRLM, PET, FBN:32, respectively.
Features are thus indicated by a feature name and a subscript. As the nomenclature is designed to both concise and complete, only details for which ambiguity may exist are to be explicitly incorporated in the subscript. The subscript of a feature name may contain the following items to address ambiguous naming:
An abbreviation of the feature family (required).
The aggregation method of a feature (optional).
A descriptor describing the modality the feature is based on, the specific channel (for microscopy images), the specific imaging data (in the case of repeat imaging or delta-features) sets, conversions (such as SUV and SUL), and/or the specific ROI. For example, one could write PET:SUV to separate it from CT and PET:SUL features (optional).
Spatial filters and settings (optional).
The interpolation algorithm and uniform interpolation grid spacing (optional).
The re-segmentation range and outlier filtering (optional).
The discretisation method and relevant discretisation parameters, i.e. number of bins or bin size (optional).
Feature specific parameters, such as distance for some texture features (optional).
Optional descriptors are only added to the subscript if there are multiple possibilities. For example, if only CT data is used, adding the modality to the subscript is not required. Nonetheless, such details must be reported as well (see section 4.1).
The sections below have tables with permanent IBSI identifiers for concepts that were defined within this document.
Abbreviating feature families
The following is a list of the feature families in this document and their suggested abbreviations:
feature family
abbreviation
morphology
MORPH
HCUG
local intensity
LI
9ST6
intensity-based statistics
IS, STAT
UHIW
intensity histogram
IH
ZVCW
intensity-volume histogram
IVH
P88C
grey level co-occurrence matrix
GLCM, CM
LFYI
grey level run length matrix
GLRLM, RLM
TP0I
grey level size zone matrix
GLSZM, SZM
9SAK
grey level distance zone matrix
GLDZM, DZM
VMDZ
neighbourhood grey tone difference matrix
NGTDM
IPET
neighbouring grey level dependence matrix
NGLDM
REK0
Abbreviating feature aggregation
The following is a list of feature families and the possible aggregation methods:
morphology, LI
-,
features are 3D by definition
DHQ4
IS, IH, IVH
2D
averaged over slices (rare)
3IDG
-,3D
calculated over the volume (default)
DHQ4
GLCM, GLRLM
2D:avg
averaged over slices and directions
BTW3
2D:mrg, 2Dsmrg
merged directions per slice and averaged
SUJT
2.5D:avg, 2.5D:dmrg
merged per direction and averaged
JJUI
2.5D:mrg, 2.5D:vmrg
merged over all slices
ZW7Z
3D:avg
averaged over 3D directions
ITBB
3D:mrg
merged 3D directions
IAZD
GLSZM, GLDZM, NGTDM, NGLDM
2D
averaged over slices
8QNN
2.5D
merged over all slices
62GR
3D
calculated from single 3D matrix
KOBO
In the list above, ’–’ signifies an empty entry which does not need to be added to the subscript. The following examples highlight the nomenclature used above:
joint maximumCM, 2D:avg: GLCM-based joint maximum feature, calculated by averaging the feature for every in-slice GLCM.
short runs emphasisRLM, 3D:mrg: RLM-based short runs emphasis feature, calculated from an RLM that was aggregated by merging the RLM of each 3D direction.
meanIS: intensity statistical mean feature, calculated over the 3D ROI volume.
grey level varianceSZM, 2D: SZM-based grey level variance feature, calculated by averaging the feature value from the SZM in each slice over all the slices.
Abbreviating interpolation
The following is a list of interpolation methods and the suggested notation. Note that # is the interpolation spacing, including units, and dim is 2D for interpolation with the slice plane and 3D for volumetric interpolation.
interpolation method
notation
none
INT:–
nearest neighbour interpolation
NNB:dim:#
linear interpolation
LIN:dim:#
cubic convolution interpolation
CCI:dim:#
cubic spline interpolation
CSI:dim:#, SI3:dim:#
The dimension attribute and interpolation spacing may be omitted if this is clear from the context. The following examples highlight the nomenclature introduced above:
meanIS, LIN:2D:2mm: intensity statistical mean feature, calculated after bilinear interpolation with the slice planes to uniform voxel sizes of 2mm.
meanIH, NNB:3D:1mm: intensity histogram mean feature, calculated after trilinear interpolation to uniform voxel sizes of 1mm.
joint maximumCM, 2D:mrg, CSI:2D:2mm: GLCM-based joint maximum feature, calculated by first merging all GLCM within a slice to single GLCM, calculating the feature and then averaging the feature values over the slices. GLCMs were determined in the image interpolated within the slice plane to 2 × 2mm voxels using cubic spline interpolation.
Describing re-segmentation
Re-segmentation can be noted as follows:
re-segmentation method
notation
none
RS:–
range
RS:[#,#]
USB3
outlier filtering
RS:#σ
7ACA
In the table above # signify numbers. A re-segmentation range can be half-open, i.e. RS:[#,∞). Re-segmentation methods may be combined, i.e. both range and outlier filtering methods may be used. This is noted as RS:[#,#]+#σ or RS:#σ+[#,#]. The following are examples of the application of the above notation:
meanIS, CT, RS:[-200,150]: intensity statistical mean feature, based on an ROI in a CT image that was re-segmented within a [-200,150] HU range.
meanIS, PET:SUV, RS:[3,∞): intensity statistical mean feature, based on an ROI in a PET image with SUV values, that was re-segmented to contain only SUV of 3 and above.
meanIS, MRI:T1, RS:3σ: intensity statistical mean feature, based on an ROI in a T1-weighted MR image where the ROI was re-segmented by removing voxels with an intensity outside a μ ± 3σ range.
Abbreviating discretisation
The following is a list of discretisation methods and the suggested notation. Note that # is the value of the relevant discretisation parameter, e.g. number of bins or bin size, including units.
discretisation method
notation
none
DIS:–
fixed bin size
FBS:#
Q3RU
fixed bin number
FBN:#
K15C
histogram equalisation
EQ:#
Lloyd-Max, minimum mean squared
LM:#, MMS:#
In the table above, # signify numbers such as the number of bins or their width. Histogram equalisation of the ROI intensities can be performed before the "none", "fixed bin size", "fixed bin number" or "Lloyd-Max, minimum mean squared" algorithms defined above, with # specifying the number of bins in the histogram to be equalised. The following are examples of the application of the above notation:
meanIH,PET:SUV,RS[0,∞],FBS:0.2: intensity histogram mean feature, based on an ROI in a SUV-PET image, with bin-width of 0.2 SUV, and binning from 0.0 SUV.
grey level varianceSZM,MR:T1,RS:3σ,FBN:64: size zone matrix-based grey level variance feature, based on an ROI in a T1-weighted MR image, with 3σ re-segmentation and subsequent binning into 64 bins.
Abbreviating feature-specific parameters
Some features and feature families require additional parameters, which may be varied. These are the following:
An example is glucose present in the blood which competes with the uptake of 18F-FDG tracer in tumour tissue. To reduce competition with the tracer, patients are usually asked to fast for several hours and a blood glucose measurement may be conducted prior to tracer administration.↩
An example of a comorbidity that may affect image quality in 18F-FDG PET scans are type I and type II diabetes melitus, as well as kidney failure.↩
Many acquisition parameters may be extracted from DICOM header meta-data, or calculated from them.↩
Many reconstruction parameters may be extracted from DICOM header meta-data.↩
Spacing between image slicing is commonly, but not necessarily, the same as the slice thickness.↩
The IBSI has not introduced image transformation into the standardised image processing scheme, and is in the process of benchmarking various common filters. This section may therefore be expanded in the future.↩
A software is compliant if and only if it is able to reproduce the image biomarker benchmarks for the digital phantom and for one or more image processing configurations using the radiomics CT phantom. Reviewers may demand that you provide the IBSI compliance spreadsheet for your software.↩