Local Image Features Extraction LIFEx

1 Texture User GuideLocalImageFeaturesExtraction LIFEx C. Nioche, F. Orlhac, I. BuvatLIFEx version ,Last update of document: 2022/07/06 Protocol descriptionpage 5 Part I7 Prerequisite / Current knowledge / IBSIStep by stepGUI of Texture protocolIntensity quantization and current limitsSpatial resampling and current limitsCalculation and ResultsScripting seriesLocal Texture (Map)Texture Features - First order featurespage 25 Part IIISBI features27 ISBI featuresnon-ISBI featuresIntensitybased29P25, P75 CalciumAgatstonScoreTLGRIMI ntensityhistogram33 EnergyIntensity-histogram Q1, Q3 Intensity_histogram_CalciumAgatstonScore TLGRIMT exture Features - Second Order featurespage 37 Part IIIGLCM39 GLCM definitionGLCM featuresNGLDM43 NGLDM definitionNGLDM_ContrastGLRLM45 GLRLM definitionGLZLM47 GLZLM definitionReferencespage 49 Part IVUsefulreferences51 BibliographyPart IProtocoldescriptionChapter 1 How to perform textural analysisand estimate conventionalparameters?

2 Step by step Prerequisite / Current knowledge / Prerequisite / Current knowledge / PrerequisiteTextural analysis and measurement of conventional parameters from images are basedon a number of rules related to the Image series, ROI and pixel values, as described be-low: Image series Textural analysis is currently supported for PET, MR, CT, US and NM. Calcu-lations for Computed radiography (CR), Digital Radiography (DX), Radio Fluo-roscopy (RF), Slide Microscopy (SM) and Optical Images are available only if theimages are in TIFF or JPG formats; Texture indices are calculated in a ROI for one or several layers (1 layer = 1 imageseries).

3 To properly interpret the index values, it is important to know that theindex are calculated for the selected Image series (active layer) and, by default, forthe voxel sampling of the selected Image : The ROI should include at least 64 voxels for 3D ROI and 16 voxels for 2D ROI; The ROI should be a single piece (contiguous voxels). It is possible to have a regionincluding several clusters of voxels, in which case only the largest cluster is usedfor calculating textural indices; The ROI definition is an important step, as edges can largely affect textural indexvalues as well as conventional index values (Orlhac et al 2017).

4 In a script, you may addLI This line determines the ROI boundsto be used for intensity scaling when applying the Texture Protocol on a wholecohort. In a script, the lineLI for ROI and check if they are numer-ous index for ROI in a file. That is to say if ROI are parameters: Number of grey levels: this is the number of grey levels (must be >0) used toresample the ROI content. Can typically be 16, 32, 64, 128. The initial pixel valueswill be resampled into that number of grey levels; Size of bin: the width of bin; Resampling: can be absolute (between fixed min and max values regardless of theROI content) or relative (between the min and max values in the ROI); See Orlhacet al Plos One 2015 for the impact of resampling approach onto tissue classification.

5 Min, max bound: to be set for absolute resampling only: min and max boundsdefine the range of resampled Prerequisite / Current knowledge / Current knowledgeIn PET default, texture matrices are computed after a resampling step with64 discrete values, between 0 and 20 SUV units, without spatial resampling [Orlhac2015]. Six texture indices have been shown to be robust to segmentation and are rela-tively independent one from another: GLCM_Homogeneity (from GLCM), GLCM_Entropy (from GLCM), GLRLM_Short-Run Emphasis (GLRLM_SRE, from GLRLM), GLRLM_Long-Run Emphasis (GLRLM_LRE, from GLRLM), GLZLM_Low Gray-level Zone Emphasis (GLZLM_LGZE, from GLZLM), GLZLM_High Gray-level Zone Emphasis (GLZLM_HGZE, from GLZLM).

6 Other set of 30 texture indices, 35 histogram indices, 23 morphological in-dices and 7 conventional (or intensity-based) PET Features are computed. For the resam-pling step, the user can choose the number of discrete values (default: 64) and the range(default: 0-20 SUV units) and the spatial resampling (default: dimension of voxel in theimages).Figure : Texture feature extract (Advanced settings of First Order Features - Morpho-logical - IntensityBased)Figure : Texture feature extract (Advanced settings of First Order Features - LocalIn-tensityBased - IntensityHistogram) Prerequisite / Current knowledge / IBSIF igure : Texture feature extract (Advanced settings of First Order Features - LocalIn-tensityHistrogram)Figure : Texture feature extract (Advanced settings of Second Order Features - Inten-sityBasedRIM)Figure.

7 Texture feature extract (Advanced settings of Second Order Features - GLCM- GLRM) The Image biomarker standardisation initiative (IBSI) LIFEx participates in the Image biomarker standardisation initiative. The primaryaim of this initiative is to improve reproducibility of quantitative Image analyses. LIFE xhas been validated on the digital phantom provided by IBSI in nifti format for directbenchmarking of feature values, without Image processing. This phantom has beendesigned specifically for the purpose of benchmarking software for radiomic featurecalculation.(see ) Step by stepFigure : Texture feature extract (Advanced settings of Second Order Features -NGTDM - GLSZM) Step by stepThis section describes the steps needed to get textural and conventional index Reading of the dicom images for which index should be calculated (link);2.

8 Drawing or loading the ROI (link); of an approach: global(ROI) or Local (Map). If you want to computetexture indices for your ROI, choose the global of the input parameters: default values can be changed. Parameterswithout default values should be set; of the output parameters of interest: click on theSettingsicon at thebottom left of the textural protocol menu. All indexes to be calculated should beselected; the process: click on theRUNicon at the bottom right of the texturalprotocol menu;7. The folder containing thecsvfile listing the results is Graphic User Interface of Texture protocolThis section describes how to get textural and conventional index user interface (see figure ) includes 3 sub-parts:1.

9 The "Spatial Resampling" and "Intensity Resampling" and "Intensity Discretiza-tion"2. A panel to set advanced parameters (see figure?? )3. The "Run" button to start the Intensity quantization and current limitsFigure : Display of textural Intensity quantization and current limitsIn PET, several studies have already investigated the impact of resampling values for number of grey levels, bin width, resampling approach, and bounds,are set based on these MR, CT, US, NM, CR, DX, RF, there is a lack of studies regarding the resamplingsettings. Studies are on-going to define robust parameter values.

10 Meanwhile, usersshould interpret the results with types of intensity resampling are available:Absolute resampling bounds: min maxROI values are rescaled between theminandmaxvalues as set in the parameter box. These values should be set in the appropri-ate units depending on the Image type (Hounsfield units, SUV, #).All voxels are accounted for in the texture index resampling of ROI values: min maxROI values are automatically rescaledbetween theminandmaxvalues of the ROI content. All voxels are accounted for in thetexture index calculation. Voxels with an initial value less thanminare set tomin. Voxelswith an initial value greater thanmaxare set resampling of ROI values: mean-3*Sd mean+3*SdROI values are auto-matically rescaled between mean-3*Sd mean+3*Sd of the ROI content, where Spatial resampling and current limitsand Sd are the mean and standard deviation of the voxels included in the ROI.