-
Notifications
You must be signed in to change notification settings - Fork 3
3. Program Interface
This page will explain the options you see in the SVR-LSM GUI interface window.
This choice will open an analysis configuration using default choices and the example dataset that is included with the software.
This choice will save the changes you have made to your analysis.
This choice will save the changes you have made to your analysis, and allow you to save it under a new name.
This option will prompt you to select a directory containing one or more .mat SVR-LSM configuration files that have been saved with the Save option. You can then select one or more from a list of the discovered .mat configuration files to be executed in a batch job, where the selected analyses are executed in a queue without human intervention.
This choice will quit the program.
This option will hopefully bring you to this github website to learn more about the program.
This menu lists the requirements for running SVRLSMGUI. If you get an error when the program launches, check this menu to make sure all of the required components are detected. These requirements include SPM12 on the path, LibSVM (optional), MATLAB's Statistics Toolbox, MATLAB's Parallel Computing Toolbox (optional), and a recent MATLAB version.
This does nothing at the moment.
This will open a typical popup menu giving some information about the software.
This check box will determine whether an output summary file is created for an analysis.
Enabling this option will cause the program to try to utilize MATLAB's parallel pools for computationally costly permutation testing, specifically for the generation of permuted beta maps and voxelwise beta sorting for determining beta value cutoffs. Note that due to the nature of parallelization, estimated time to completion is not displayed. The benefit of parallelization will largely depend on the processors available to your operating system, MATLAB's parallel pools configuration, and your analysis parameters. See DeMarco & Turkeltaub (2018) for some benchmarks relating to the speed improvements one might expect from parallelization. This option can enabled by checking the Parallelize item in the Options file menu at the top of the SVRLGMgui toolbox interface.
This menu item contains options which were previously contained in a Preferences panel (now removed) that configure the support vector regression method, including whether you intend to use libSVM or MATLAB's SVR functionality, and what parameter values should be used for the analysis (see Parameters heading below).
Click on these choices to change the SVR parameters for your analysis.
Gamma defines how far the influence of a single training example reaches. If gamma is low, every point has a far reach. High values mean every point only has a close reach. In practice, this means when we’re fitting our support vector machine that the final boundary will reflect the points that are very closest to it, so it ignores points far away contribute to the decision on where to draw the boundary. WIth a high gamma, we have a very wiggly decision boundary, because a single observation can really pull the boundary to fit it. It’s essentially the degree to which the boundary is more or less jagged (smoothes the boundary). The default value for gamma is 5, which was determined based on simulations in Zhang et al. (2014).
Defines the cost (amount of sacrifice) we are willing to accept in the pursuit of add wiggles to our otherwise smooth the decision boundary in the pursuit of correctly classifying more of the training points. The default value for cost is 30, which was determined based on simulations in Zhang et al. (2014).
In SVR-LSM, to recover the relevant anatomical locations from the SVR analysis, the trained model's predictive hyperplane is back-projected into the input data space. One step in this process is scaling the values in the back-project image. During computation of the permutations, to make SVR maps comparable, they are each scaled by the same value. In the original SVR-LSM (Zhang et al., 2014), and by default, the scaling factor is selected as the maximum absolute value observed in the real, original data ("Maximum"). However, it is not known whether this maximum is robust or stable. Thus we provide the option to scale by the 99th or 95th percentile of the real, original data. This choice can be modified by checking an option in the Beta Scaling submenu in the Options file menu at the top of the SVRLGMgui toolbox interface.
This is the name of the analysis, and will be the name of the subdirectory in which the results of the analysis are written. A subdirectory with this name is created within the folder specified by the Output Folder option.
This is the location on your hard disk where results will be written. Results will be contained in a subdirectory named as the Analysis name field option.
This is the folder containing lesion files for patients to be used in the analysis. These files must be in NiFTI (.nii) format, readable by SPM, and match the RegistryCode column in your score/design (.csv) file.
Behavioral scores should be organized in a comma separated value (CSV) file where columns represent scores and rows represent individual subjects. The first row should contain headers that contain the name of each column's score. Importantly, one column header must be RegistryCode. This header should contain the NiFTI image names containing the lesion tracings for each subject in the analysis.
This is the behavioral score to be evaluated in the analysis, and is selected from among the headers in the file specified by Score file.
Lesion volume can be a confound to analyses where behavioral scores correlate with lesion volume. See DeMarco & Turkeltaub (2018) for more information.
Lesion volume will not be controlled in your analysis.
This option will implement Direct total lesion volume control (dTLVC; Zhang et al., 2014). This transformation operates on the raw voxel data, dividing each voxel in each subject's lesion by the square root of the size of that subject's lesion. Note that this transformation on affects lesioned voxels (since zero divided by any number is also zero).
This option will include lesion volume as a covariate in the behavioral nuisance model.
This option will include lesion volume as a covariate in the voxelwise lesion data nuisance model.
This option will include lesion volume as a covariate in the behavioral nuisance model and the voxelwise lesion data nuisance model.
This option will constrain analyses to voxels where only more than the specified number of lesions are present in the patient sample. This is an important option because of concerns relating to poor power in the periphery of lesion overlaps.
This choice will cause output maps containing p values to show as 1 minus the p-value. For instance, a p value of .01 will appear as 1 - .01 = .99. This allows easier thresholding in some MRI display software such as MRIcron, which utilizes a minimum value cutoff to showing overlay data. Thus, an intensity cutoff in MRIcron set to .99 will hide all voxels with a p value of greater than .01.
The covariates pane allows you to add an arbitrary number of covariates to your analysis. Covariates are included in the nuisance models that are applied to the data prior to the SVR analysis. To add a covariate to an analysis, choose the name from the dropdown list at the bottom of the Covariates pane and click the + ("plus") button beside the dropdown box. This will add the selected covariate to the list of covariates that can be included in the analysis. To remove a covariate from the analysis, select it in the Covariates listbox and click the - ("minus") button below the listbox.
After adding covariates to the list, it is also necessary to indicate what nuisance model(s) to include the covariates in. There is a nuisance model for the behavioral score and a nuisance model for the voxelwise lesion data. To include the added covariates to the behavioral score nuisance model, check the "Behavioral Score" checkbox. To include the added covariates to the voxelwise lesion data nuisance model, check the "Lesion Data" checkbox. If neither is checked, then the covariates will not be included in your analysis. You can see what covariates were included in the Summary Output file.
When this checkbox is checked, the software will perform permutation testing on the map of SVR-beta values to convert them into p-values, and also perform permutation testing at the cluster level.
This should be a positive integer indicating how many permutations to perform. A good value is about 10,000. The number of permutations indicated is used for both voxel level and cluster level permutations.
This is the voxelwise p value cutoff used to threshold the SVR-beta map derived via permutation testing. A good value is about .005.
This is the clusterwise p value cutoff used to threshold the beta map derived via permutation testing. A good value is about .05.
This checkbox enables continuous permutation-based family-wise error rate (FWER) correction (Mirman et al., 2018). As a correction method, FWER attempts to quantify the number of possible false positive voxels.
With this option selected, the number of permutations to compute is still specified. However, two options replace the Voxelwise P and Clusterwise P editboxes: v, specified in cubic millimeters, and the FWER.
With CFWER enabled, this option is used to specify the rate of false positive voxels. The typical approach is to use the most extreme test statistics computed for each permutation (e.g., v = 1), but Mirman et al. (2018) generalize the approach to control for other values of v. Importantly, in SVRLSMGUI, v is specified in cubic millimeters, as opposed to voxels. Number of voxels depends on the voxel dimensions in the data.
With CFWER enabled, this option specifies the desired family-wise error rate percentile. For instance, for 95th percentile, specify the value 0.05 (default).
The feedback pane shows each step as an analysis proceeds. If there is an error it should show up in the feedback pane.
When you run an analysis you will see a bar showing the progress for each step. This is not supported for some parallelized steps.