Before asking questions or reporting problems to the contact address, please make sure your question has not been answered in the past.
How to cite LIFEx?
Please see the How to cite LIFEx? page.
How convert my csv result file to Excel format:
Please follow the instructions given at convert csv to columns
What is the ideal solution between the absolute or relative resampling (= intensity discretization) in PET?
In PET images, we do not recommend to use the relative resampling. As explained in https://www.ncbi.nlm.nih.gov/pubmed/26669541, relative resampling induces a high correlation between texture indices and metabolic volume, and we proposed to use the absolute resampling instead. In PET, we resample with 64 number of grey levels (size bin equals to 0.3) between 0 and 20 SUV units (because this is the standard range of SUVs in oncology with FDG). Generally speaking, a good approach is to determine the range of variations of SUV across the regions of interest of in all patients and set the resampling values as a function of that. For instance, if you are studying regions in which SUV is always between 0 and 9, then you can use absolute resampling with 64 grey level between 0 and 10 (bin size of 0.15).
What is the ideal solution between the absolute or relative resampling (= intensity discretization) in CT?
In CT images, we recommend using absolute resampling, as in PET. Again, you have to set the range of HU and the number of bins (or bin size) for your data, as a function of the data content. Basically, the bin size has to be set as a function of the meaning of HU difference. If a HU difference of 10 does not mean much, then the bin size could be set to 10 for instance (using a smaller value is useless) and a range of HU between -1000 HU and 3000 HU for example (including all HU possible variations). If you are only interest in one organ, for instance willing to investigate radiomic features in the liver only, then you should use a much narrower range of values, for instance between 0 HU and 50 HU for non-enhanced CT using 20 bins (bin size of 2.5).
What is the ideal solution between the absolute or relative resampling (= intensity discretization) in MR?
In MR, the impact of different resampling schemes has been studied in http://iopscience.iop.org/article/10.1088/1361-6560/aabd21. Based on that paper, we recommend using a fixed bin width, which corresponds to absolute resampling. Intensity discretization setting does depend on the MR sequence that is used, but usually, we would recommend using 128 bins. Min and max setting completely depends on the MR sequence and the type of lesion of interest. A way to identify the min and max value setting is to go over all patients, look at the min and max in each patient ROI and then take the min of min and max of max over all patients. And use these values for each patient. What is important is that you use the same min and max values for all patients. Yet, in MR, you first have to check that the MR intensity values in a reference healthy tissue are about the same for all patients. Otherwise, you have to correct for the inter-scan variability first, before performing radiomic analysis. What we could recommend is that you first focus on a healthy reference region and check that in that region, you observe consistent radiomic feature values across patients (ie that they follow a distribution close to a Gaussian distribution). Only after this first step is validated, you can go on and analyze lesions.
What is the ideal solution between absolute or relative resampling (= intensity discretization) in ultrasound imaging (US)?
In US images, we recommend using absolute resampling, ie fixed bin size, as in PET. You have to set the range of Intensity values and the number of bins (or bin size) for your data, as a function of the data content. Basically, the bin size has to be set as a function of the meaning of intensity difference. If an intensity difference of 10 does not mean much, then the bin size could be set to at least 10 for instance (using a smaller value is useless). If you are only interested in one organ, for instance willing to investigate radiomic features in the liver only, then you should use a range of values corresponding to values observed in that organ, for instance between 0 and 128 using 10 bins (bin size of 12.8).
Does LIFEx have any functionalities for harmonizing radiomic features calculated from images coming from different scanners or protocols?
Why is texture processing time so long?
Calculations of some radiomic features are very time consuming, such as SUVpeak1mL, SUVpeak05mL and RIM. To speed up the process, first start by un-selecting them in the advanced texture parameters form. And then restart the calculation.
In PET, the DICOM fields that have to be properly filled for correct SUV correction are:
DICOM fields to SUV conversion:
(0008,0032) TM #6  Acquisition Time
(0010,1020) DS #6  Patient’s Size
(0018,1074) DS #10  Radionuclide Total Dose
(0018,1075) DS #12  Half Life
(0018,1072) TM #6  Radiopharmaceutical Start Time
(0054,1001) CS #4  Units
(0028,1052) DS #2  Rescale Intercept
(0028,1053) DS #8  Rescale Slope
About the Spatial resampling XYZ spacing.
Which values should be used?
The main point of the spatial resampling XYZ spacing is to make sure that all calculations are performed with the same voxel size whatever the voxel size of the input image. This is important because textural feature values depend on voxel size. If all input images have the same voxel size, the user does not have to worry about this spacing, as it will be set by default to the voxel size so will always be the same. If the voxel size is not always the same for the different images processed in a study, then a fixed voxel size should be chosen (eg 2 x 2 x 2 mm in PET), and then used that for all images. This will remove the variability in texture values due to different voxel size.