These are some new features in the MIDAS package:
IMCALC. A new “Image Calculation” program has been added to the MIDAS toolbar. It is started using the calculator icon, , which has been added to the main toolbar. This has options for image manipulation, image analysis using calculations, histograms, or regressions, and displaying combined images. For example, images can be calculated using simple programming expressions (e.g. Cre/(NAA+Cre+Cho) and results displayed and exported. Any IDL command can be used and you can even call your own IDL programs for more complex processing operations. It supports any image type and there are multiple options. Here are some examples:
The first image show the Cho/NAA image superimposed on the T1 and the second is a thresholded overlay of the SUV map for a FET PET study showing the region of significant uptake. The plot shows a regression of choline against NAA with outlying values highlighted. These highlighted voxels are then displayed on the T1 image shown to the right.
To use the image calculation features it is necessary to have the full IDL license. It is recommended that you look at the help file for this application to understand the features.
MINT. This is a new application, developed in collaboration with Dr. Poptani (under NIH grant R21CA170284), that integrates spectra from a predefined region and optionally applies spectral analysis to the integrated spectrum. Fitting using the FITT program can be done automatically or data can be saved in LCModel format. Example applications include integrating spectra from operator-defined regions, for example from a lesion, or from atlas-defined brain regions. Options in the program allow filtering the voxel selection by fitted linewidth and tissue content. By integrating over a region the SNR is improved to the point where fitting of compounds such as glutamate and glutamine can be done reliably. The example spectrum shown here is obtained following integration over a ~10 cc volume and the results from FITT are shown in green. The CRBs for Glu and Gln were both under 5%. Because signal integration is done after correction for B0 inhomogeneities, plus the data is acquired with a small nominal spatial resolution (0.313 cc), an excellent linewidth can be maintained. In an ISMRM abstract Dr. Ding (X. Ding, p2901, 2014) has shown that linewidths can be better than those obtained from a typical single-voxel acquisition.
A test of the MINT program obtained multiple integrated volumes throughout the whole head, which resulted in 139 regions of approximately 6cc volume, of which 133 had Glutamate CRB values of <10%.
The MINT program can be run from the Tools section of the MIDAS toolbar. It is supported by the MAPPER program that imports masks generated in Analyze format and a tool that organizes multiple FITT results into a text file that can be easily imported into a spreadsheet for analysis.
Denoising. A new program, DNPCA, has been developed that uses PCA for denoising of the MRSI data and uses a novel method for automatically choosing the eigenvectors of interest. The program is applied after an initial fit that corrects for phase and B0 shifts, and so is conveniently implemented within the METAFIT processing. To the right is shown an example, with the raw data on top, the standard 2 Hz Gaussian filtering (center), and the result after applying PCA denoising algorithm (bottom).
The program has been shown to improve SNR without impact on the linewidth, resulting in decreased CRB values of spectral fitting. Further evaluation of the program is ongoing and processing pipelines that include this step will be available shortly.
IDL Segmentation. The MRI segmentation used in MIDAS uses the BET and FAST programs from the FSL library. These have previously been called through a Java application. To provide more flexibility this has now been replaced with an IDL application, IDLSEG. This has additional options, including iterative application of BET that handles sagittal acquisitions better, and using the FLAIR image for the brain extraction. One difference with the previous version of the segmentation is that it doesn’t apply the bias field correction to the MRI_T1 and as a result its operation is faster.
PACS. While MIDAS includes a function to output images in DICOM format (in the Viewer) it does not include a method to send data to a PACS system. A separate utility has now been developed for this purpose and is made available on request. It requires purchase of the IDL DICOM library.
Multitasking Using BATCH. The BATCH program has been modified to spawn processes and then continue without waiting for them to finish. This is done using the qualifier “/NOWAIT” after the application name, e.g. a line in the BATCH processing file would read as:
volumizer/nowait 0 1 none 1 MRI_FLAIR none 0,0,0
The multitasking can be applied to IDL functions as well as command-line calls to external processing functions. There are several considerations when using this function, for example error messages from a spawned application will not be known to the BATCH program (although the error log files will still get written) and care is needed to avoid multiple programs competing for access to the same files. Additional discussion on use and limitations is provided in the help file. Initial tests indicate that approximately 30 minutes can be saved by running some steps in parallel.
Brain Atlas in Subject Space. The PRANA program now has an option to apply the inverse spatial transform of an atlas file to the SI-resolution subject space. This requires that the deformable spatial registration to the atlas has already been done. The function is started using the “Atlas to Subject Output” option on the display widget pull-down menu (shown right). A widget that allows selection of specific atlas regions is then displayed, but generally all regions can be taken. The subject-space version of the atlas gets put into a PRANA node under the SI/Maps, as illustrated below in the MIDAS Browser.
The atlas can be viewed in any of the display programs in the selection if SI maps. On the left is shown an example for a slice at the taken at the level of the thalamus for a modified version of the AAL atlas (N. Tzourio-Mazoyer et al., Neuroimage, 15: 273 (2002)). Remember this is now at the SI resolution in subject space.
The regions defined in the subject-space atlas can be used for spectral integration in the MINT program and it can be useful in the IMCALC program to limit voxel selection or look at data values in specific regions.
There is an important limitation of this atlas in that it does not account for partial volume contributions between neighboring regions. To minimize file sizes the atlas is saved in an integer “indexed” format. However, all analyses of SI data should take into account the relatively broad spatial response function of the SI data. As discussed in the previous newsletter the effective voxel volume for our standard EPSI implementation is equivalent to 1.5 cc.
Signal-to-Noise Calculation. MIDAS has included a utility function that creates a map of the spectral SNR measurement for some time, but this could only be run in batch mode. A widget has now been added to this program, which can be run from the Tools section of the MIDAS Toolbar. The program allows selection of frequency ranges for measurement of a peak height and of the noise. It is assumed that the NAA peak will be selected and the program then produces two results, the SNR_NAApkht and SNR_NAAarea. The Area result is obtained using the result of the spectral fitting of NAA divided by the RMS value of the noise, and has the advantage over the peak measurement in that it accounts for varying linewidth.
Zip Study Tool. This utility was introduced last year as a convenient way of packing all data or a selection of files related to a study into a zip file. In addition to expanding some of the file types that get included, this now includes an option to copy all of the processing files that are used in a batch processing pipeline. This utility therefore provides a convenient way to pack up all data files and processing files to send a study to another site or to copy a project to another location. Full operational details are in the help file.
QUID Level & Width Values. The QUID program (introduced in the 2013 Newsletter) provides a mosaic format for displaying images. When making up reports it is frequently desirable to show multiple images with the same color scale, e.g. for comparisons between studies. To do this the level and width values can be entered manually using the text entry boxes in the window Level/Width portion of the widget, shown to the right. The example shown here for a mean kurtosis image for the that is displayed for the Level and Width values shown in the widget above:
Spectral integration in SID. The MINT program, described above, does spectral integration over predefined 3D volumes or brain atlas regions, but for more ad hoc analyses the same type of processing can also be done in the SID program. Averaging spectra from multiple voxels can be done from the Options menu on the Reference Image “Sum Region On/OFF”, or by starting the Processing>Read Image Values function and clicking the button “Turn on ROI Averaging”. Once enabled, each selection of a new voxel will add the new selection to the displayed spectrum.
The Image Values widget also has some additional options, which are shown to the right. Some of the standard maps can be used to automatically guide the signal integration. In the example shown to the right the white-matter map has been selected with a threshold of 0.5 (the map goes from 0.0 to 1.0), and only voxels having a fitted linewidth of between 3 and 12 Hz will be selected. This integration is only done at the selected slice, but can be continued for other slices by changing slice and repeating the operation. The plot shows a result obtained from 311 voxels (remember these are interpolated voxels) and the reference image shows which voxels have been selected:
Other functions on this widget include saving the data values to a text file and saving the individual spectra from the selected ROI. For these options it is frequently convenient to reduce the number of data points by subsampling the voxels selected, in light of the fact that the data is interpolated and neighboring voxels are not independent. This is achieved using the “Skip Factor”, which can be seen on the widget shown above. An additional new feature is that spectral fitting can be done on this spectrum (or any single spectrum) using FITT. This is done from the “Spectral Fitting” pull-down menu option on the spectrum window (shown to the right). This brings up a new widget, shown here:
On this widget you need to enter a filename to save the spectrum and the fit results and select a processing file. These FITT processing files need to be specific for single-voxel fitting and we are using the convention that the filenames start with FITT2_SVS_*. On hitting “Save and Fit” the FITT program will run and if the “Auto Show Result” option was checked the results will be displayed in a new window. For the example shown above the fitted linewidth was 6.5 Hz, not bad for a 97 cc volume!
Cho/NAA Ratio Values. This question has come up a few times: “Why do the Cho/NAA ratios in MIDAS not match the values I’m used to?
The answer is that the spectral fitting used in MIDAS calculates the relative concentration, whereas historically, at least in the in vivo MRS world, people have relied on visual analysis of peak heights or fitting routines that return results for peak areas. The relative concentrations and peak areas are different because the peaks for NAA and Creatine come from 3 protons in one molecule but the peak from Choline comes from 9. To convert relative concentrations to values equivalent to those obtained using peak heights or peak areas then wherever you see Cho multiply by 3. Taking the example spectrum shown above (in the “Spectral integration in SID” section), which is for normal brain, the fit returns the following values:
Choline = 281
Creatine = 930
NAA = 1176
Cho/NAA = 0.24
Cho/Cre = 0.30
NAA/Cre = 1.26
Here the NAA/Cre ratio agrees visually with the plot but the Cho/NAA and Cho/Cre do not. To calculate the relative area we have to multiply Cho by three:
Cho/NAA = 3*281/1176 = 0.71
Cho/Cre = 3*281/930 = 0.91
The values for the ratio of the areas now agree qualitatively with what we can visually estimate from the spectrum.
Note: There is a source for confusion here since the MIDAS naming convention for the fitted metabolite amplitude values returned by the FITT program are labeled with the suffix _Area, e.g. NAA_Area, but keep in mind that this is really the relative concentration. Following signal normalization the _Area label is dropped, i.e. it becomes NAcetylaspartate.
MIDAS FAQ. The previous item discussing ratio values, as well as other questions and a selection of image artifacts that we have encountered are discussed in the MIDAS_FAQ.pdf document. This is located in the “Installation” subdirectory of the MIDAS documentation as well as on the documents page of the web site. It’s worth checking this occasionally.
EPSI Releases. The EPSI sequence has undergone occasional changes, with the last release done in December 2013. If you have not had a release of the sequence since then it may be worthwhile getting the latest version. The most recent sequence is epsi_extWSr_vsTE_smgrappa1_el3_prf1
These are the changes made before Dec. 2013:
· The GRAPPA encoding was changed to have 13 central (fully sampled) k-space lines< which was increased from the previous 7 central encoding lines. This improved SNR and made the sequence less sensitive to motion.
· Elliptical k-space encoding was added. This maintained a similar acquisition time with the increased number of central k-space lines.
· The phase encoding was moved to after the 180 refocusing pulse. This improved a zipper artifact resulting from transverse magnetization coming from the refocusing pulse.
Help! Artifacts! This is an artifact we are bewildered by and we are looking for ideas. This was a study of a 53 y.o. female with chronic migraines. The SI data showed excellent quality with narrow linewidths and no evidence of motion, but had a terrible spectrum in a very localized region in the occipital lobe. The image below shows the creatine map and the plot shows the spectrum in the middle of the white spot. This artifact was centered at slice 17, i.e. the center in z, where we have previously noted the zipper artifact that runs along the center of X and Z, but that could not be seen in this case. It covers about 5 slices in Z, with no evidence of anything propagating from above or below. The water image and line shape at this location (shown below) showed good quality and no artifact, indicating good B0 homogeneity.
One thought was a flow artifact, but there was no evidence of this on the perfusion image done in the same study. Another thought is stimulated echos in some out-of-slab region.
Any ideas would be welcome!
Spectral Simulation for FITT: The basis functions used in FITT are created using the VESPA simulation program. The prior spectral parameter information used for this simulation is based on our 2000 report: V. Govindaraju, et al. NMR Biomed. 13: 129-153 (2000). An update of some parameters and corrections to the original report has just come out, in:
Govind V, Young K, Maudsley AA. Corrigendum: Proton NMR chemical shifts and coupling constants for brain metabolites. Govindaraju V, Young K, Maudsley AA, NMR Biomed. 2000; 13: 129-153. NMR Biomed. 28(7):923-924 (2015).
This also includes a supplementary file with the full list of metabolites.
Changing FITT Parameters. The spectral fitting program, developed by Brian Soher, has multiple program options that require some time to get familiar with. However, a relatively simple change that many researchers may want to do is to change the metabolites included in the spectral model. Examples may be to add lactate for a study with a tumor, or to switch between Glx and Glutamate plus Glutamine. To do this, start up FITT2 from the Toolbar and load a dataset and processing file. Note, we use the FITT2 program, , although the older FITT program is still available on the toolbar for backward compatibility.
The data must be reconstructed (FDFT and LITE) but it doesn’t matter if the FITT program has been previously run. Then click the “Editor” and “Viewer” buttons. Starting with the Viewer, shown in the following Figure, this allows selection of a voxel (in the image window to the lower left) and viewing the fitted result, here is an example:
On first use the plots shown will be different from that shown here. The plot scaling will need to be adjusted and the information displayed will need to be set using the “Number of Plots” and the data type selection – see the menus under Plot1, Plot2 etc. See the help files (for FITT and FITT2) for a description of all the options. Clicking “ReFitVox” runs the fitting and the result is displayed. The plots shown here are 1) the data and initial estimate, 2) the data and fit result, 3) the individual components of the fit, and 4) the residual.
The next step will be to change the metabolite selection using the Editor. The available metabolites are listed in the lower section:
If the metabolite you want is not listed then the basis functions need to be created and imported using the “Prior File” section at the top of the widget. The “Fixed T2 value” defines a fixed Lorentzian component to the lineshape and the value depends on the lineshape model used. For LorGauss this should be set to a large number (essentially disabled). For Gaussian the T2 value can be used, note it is defined in seconds. The “Use DB ppm values” determines whether the peaks are allowed to move around a little when getting the initial values or whether the exact database values should be used. Some improvement is obtained for lactate with this on.
When changing the metabolite selection the program will inform you that any previous fit result is now invalid and will be deleted, but you can now go back to the Viewer window and refit the spectrum using the new model. If the result looks good then save the new processing file, File->Save.
Congratulations and many thanks to our collaborators for contributions to these recent reports:
· M. Sabati, S. Sheriff, M. Gu, J. Wei, H. Zhu, P. B. Barker, D. M. Spielman, J. R. Alger, A. A. Maudsley. Multi-vendor implementation and comparison of volumetric whole-brain echo-planar mr spectroscopic imaging. Magn. Reson. Med., Epub ahead of print (2014).
· TV Veenith; M Mada; E Carter; J Grossac; VFJ Newcombe; JG Outtrim; V Lupson; S Nallapareddy; GB Williams; S Sheriff; DK Menon; AA Maudsley; JP Coles. Comparison of inter-subject variability and reproducibility of whole brain metabolite ratio proton spectroscopy. PLOS One, 17;9(12):e115304 (2014).
· AA Maudsley, V Govind, B Levin, G Saigal, S Sheriff, and L Harris. Distributions of altered MR diffusion and spectroscopy measures with traumatic brain injury. J. Neurotrauma, 32(14):1056-1063, (2015).
· V.V. Eylers, A.A. Maudsley, P. Bronzlik, P.R. Dellani, H. Lanfermann, X-Q Ding. Detection of normal aging effects on human brain metabolite concentrations and microstructure with whole brain MR spectroscopic imaging and quantitative MR imaging. AJNR, In Press (2015).
· A. Abdoli and A.A. Maudsley. Phased-array combination for MR spectroscopic imaging using a water-reference. Magn. Reson. Med., In Press (2014).
This report showed that the multichannel combination method implemented in the FDFT program performs as well as or better than others that have been proposed.
Midas Blenney (by Gary McKinney)
Andrew Maudsley, September 2015
This newsletter is aimed at providing information to the developers and users of the MIDAS software package. If you would prefer not to receive notification of these reports just let me know (amaudsley at med.miami.edu).