SPM Lab II Monday, August 27 2001 1-3PM BME 499.098/Biostat 642 ============================================================================ Name: ___________________________________ Group Name: ___________________________________ Dataset Used: fmriclass ________ (Use *your* group's data) Goals of this Lab ============================================================================ After this lab you will... 1. Be able to perform coregistration between the low resolution and high resolution structural images. You will understand the implications for a image to be the "Object" vs the "Target" in terms of the "world space" of each image. 2. Be able to perform spatial normalization, check its success, and apply the transformation to a contrast image. 3. Be able to include a temporal derivative in a model and assess the necessity of that effect with an F test. Coregistration ============================================================================ SPM's "Coregister" facility is used for registering different types of images from the same subject (it is a "intermodality, intrasubject" registration). To use it you must specify a "Target" or reference image, which does not move, and an "Object" image, which is moved to match the target. Recall that the rigid body transformation is recorded in the .mat file of the object image. Hence, it is not necessary to write out the transformed image (the Object transformed into the space of the Target), though it is sometimes convenient to do so. Before coregistering the low- and high- resolution structural images we need to (I) make sure the two images are in the same orientation and not too far from one another, and (II) correct for intensity inhomogeneities in the high-resolution SPGR image. 0. First see how close the t1_gre and t1_spgr images are to start with. What button do you use to view more than one image? _______________________________________________________________ Multiple images are displayed in the space of the first image. If they are not previously registered this simply means that their origins are lined up. Neither image has had their origin set. What is the default origin? (Not a specific voxel x,y,z, but ...) _______________________________________________________________ 1. If the images were very out of register we could try to get the images closer by manually setting the origin to the Anterior Commissure (AC). A. On the board will be a diagram of how to find the AC on the midsagittal plane. Sketch it here B. Now use the Display button view the t1_gre image and locate the location, as best possible of the AC. You should be able to see it on the axial slice. Hint: Once you are close, it helps to "zoom in". In the display facility, click on the "Full Volume" pop-up menu; select "80x80x80 mm" What is the location of t1_gre's AC in voxels? _________ __________ ___________ Voxels C. Use the Display button view the t1_spgr image and locate the location, as best possible of the AC. What is the location of t1_spgr's AC in voxels? _________ __________ ___________ Voxels D. At this point we *could* set each image's origin using "HDR edit", but we won't. We don't need to because the images are already quite close, but there's a more important reason. By setting the origin on the Target image (t1_gre), we change its "world space"; specifically, the origin of its world space moves from the center of the volume to the AC. Hence it will no longer have the same world space as the functional data *unless* we identically change the origin on all the functional data and any results (beta*, con*, spm_T* etc) we already created. To summarize, if the images are way off, we can set the origin manually to help the Coregistration. But if we change the origin of the image representing the functional space, we have to similarly change the origin on all functional images. 2. The t1_spgr images must be corrected for inhomogeneity before coregistration. As is typical, the SPGR images show a tendency to be brighter in the center of the volume than at the edge. A. Use the display facility to find what the typical intensity of white matter in the corpus collosum? (If you don't know where the corpus collosum is, ask your neighbor.) _______________ B. What are typical white matter values in the frontal cortex? _______________ Perform homogeneity correction on t1_spgr, creating a ht1_spgr image. (See instructions on the board to get the homogeneity correction script file.) C. Now examine the ht1_spgr image. What are typical intensity values of white matter in the corpus collosum? _______________ D. What are typical white matter values in the frontal cortex? _______________ E. Did the homogeneity correction work? Can you notice the difference visually? (e.g. use Check Reg). ________________________________________________________________ ________________________________________________________________ 3. Coregister! You will now "Coregister" the high resolution, ht1_spgr image to the t1_gre image. The t1_gre image has the same space as the functionals, and hence this coregistration will set ht1_spgr's world space to correspond to the functionals. A. What is your "Target" image? _________________________________________________________ B. What is your "Object" image? _________________________________________________________ (Stop! Ask if this is right before you continue.) C. Click on the Coregister button. You only have one subject and you want to "Coregister & Reslice". You will be asked for the modality of "Target" and "Object" image. What is modality of each? (Circle one) Target Modality: PET T1 MRI T2 MRI PD MRI EPI Transm SPECT Object Modality: PET T1 MRI T2 MRI PD MRI EPI Transm SPECT Select the Target and Object images you identified above. It asks for "Other images". This would be useful if we had other images that were acquired in the same space as the SPGR. We don't, so just click 'Done'. Now it will perform the coregistration. The transformation parameters are written in the .mat file of the object image ht1_spgr.img; this sets the world space of ht1_spgr to match that of t1_gre. Since we asked it to "Reslice", it will also create a rht1_spgr.img, an image with the same dimensions and world space as t1_gre. D. Note changes in world spaces. "Display" t1_gre.img and use the "World Space"/"Voxel Space" pop-up button to change between the two spaces. Does the image move as you flip between the two spaces? _____________________________________________________________ "Display" ht1_spgr.img and do the same again. Does the image move as you flip between the two spaces? _____________________________________________________________ Why does one move and the other doesn't? _____________________________________________________________ _____________________________________________________________ D. Check the success of the Coregistration. Use check reg to check the registration. Check the points mentioned in the last lab: i. Frontal pole ii. Occipital (posterior) pole iii. Left & Right sides (e.g. superior temporal gyrus) iv. Corpus collosum: (1) Most anterior, (2) most superior and (3) most posterior extent. Also try tracing sulci. Has the coregistration succeeded? _____________________________________________________________ _____________________________________________________________ Spatial Normalization ============================================================================ Spatial Normalization is SPMese for intersubject registration. It is essential for performing intersubject analyses or, for intrasubject analyses, determining Talairach/MNI coordinates of activation foci. While the "Normalize" button can accept any kind of image (T1, T2, etc), to register your subject into the standard atlas space, it is important to use the highest resolution structural data. 1. For us, that is which image: _________________________________________________ (Check your answer before continuing on!) Spatial normalization (or, with a British accent, normalisation) takes three types of images (i) "Images to determine parameters from" These are the high resolution anatomical images from which the spatial transformation is determined. Typically you only specify *one* such image per subject. (ii) "Images to write normalised" These are other images *with* *the* *same* *world* *space* as image (i) above. Typical examples would be statistic or contrast images, or whole set of ra* functional images. (iii) "Template image(s)" Images that define the standard atlas space. You get to choose from images that match the type (or "modality") of the image in (i). It produces a "*_sn3d.mat" file, which records the nonlinear transformation *from* the *world* *space* of image (i) *to* the atlas space. 2. Now that we have coregistered, what does the world space of ht1_spgr correspond to? __________________________________________________________ 3. Thus if we spatially normalize ht1_spgr the resulting _sn3d.mat file will not just be good for ht1_spgr, but for... __________________________________________________________ Click 'Normalize' to start the spatial normalization process. Select 'Determine Parameters Only'. For "Images to determine parameters from", select ht1_spgr.img For "Template images", select T1.img. This image is an average of 152 subject's T1 images, from the MNI/ICBM, smoothed with an 8mm filter. (ICBM=International Consortium for Brain Mapping). After a short while it will finish. This will create a _sn3d.mat and reslice the "image to determine parameter from", creating a nht1_spgr.img. Check the success of the registration, comparing the normalized image (nht1_spgr.img) to the unsmoothed version of the template image, that is avg152T1.img in the spm99/canonical directory. 4. Using the landmarks suggested above, has the spatial normalization succeeded? _______________________________________________________________ _______________________________________________________________ Using temporal derivative to improve model fit. ============================================================================ In class we learned that including a temporal derivative of an experimental predictor can approximate a (nonlinear) temporal shift within the General Linear Model framework. In this section you will fit the block design data again. First! Outside of Matlab, create an analysis directory in c:\data with your group's data number. For example, if your group's data is fmriclass4, create c:\data\analysis4 Since we will be creating more than one analysis, create a subdirectory for this temporal derivative analysis, c:\data\analysis\TempDeriv Using last Friday's lab, repeat the fitting of the block design data (run_01) but this time answer 'Yes' to "add temporal derivatives?" Now select Results. As before, specify the "Faces-Places" contrast and the "Places-Faces" contrast. But note!!! There are now 4 columns in the design matrix. The first is for "Faces", the second is for "Faces" temporal derivative; the 3rd and 4th are for "Places" and it's temporal derivative. 1. For now, ignore the derivatives. Hence the contrasts you want are: Faces-Places: ______ ______ ______ ______ Places-Faces: ______ ______ ______ ______ Results questions: "mask with other contrasts": No. You'd only ever possibly say 'Yes' if you have more than two conditions. Say you have conditions [A B C D], and you are interested in contrasts for A-B and C-D. You might want to view contrast A-B but *only* where C-D is suprathreshold. In that case you'd answer "Yes" here, specify the C-D contrast, give a threshold and the specify "Inclusive". "Title for comparisons": Accept the default or modify it. "Corrected height threshold" Selecting "Yes" will *only* show you voxels that are significant corrected for the multiple comparisons problem of searching over the whole brain. If you say "Yes", you can usually accept the default, 0.05, for the next question. The first time you look at a dataset you should use the corrected threshold to see what is really significant at the voxel level. However, almost immediately, you will probably want to repeat the "Results" with an uncorrected threshold to see what is significant by spatial extent. Recall that an arbitrary primary threshold can be used define clusters, which are assigned p-values. "Fishing" with lower thresholds is post hoc... it doesn't control for the multiple comparisons problem. However, it can be an important tool to generate new research hypotheses. "& extent threshold" Usually set to 0. It is useful to increase it to make the results "prettier" and to make adjust "set level" inference results (set level inference depends both on the primary intensity threshold *and* the extent threshold). Results Interactive Window First, press the p-values "Voxel" button to give you a table of p-values. Use "overlays..." to view the blobs on an anatomical image. To be very certain about anatomical localization, use the ravol*_0001.img image. This will show susceptibility artifacts and signal voids not visible on the SPGR or GRE image. 2. Did the temporal derivatives improve the fit? We can test if the temporal derivatives account for a significant amount of variation with an F-test. The null hypothesis we want to test, at each voxel, is Ho: The true beta associated with the Faces temporal derivative is zero, AND The true beta associated with the Places temporal derivative is zero A single contrast cannot test this hypothesis. For example, a contrast of [0 1 0 1] will only check that their average is equal to zero, or equivalently, that they are equal and opposite in sign. No, the pair of contrasts we want is ______ ______ ______ ______ ______ ______ ______ ______ Run results again and define this F-contrast. Use a corrected threshold. Are there any significant voxels? If so, are they in all previously significant areas, or just some? ______________________________________________________________ ______________________________________________________________ ______________________________________________________________ To conclude, do you think this data justifies the use of a temporal derivative? Justify your answer. ______________________________________________________________ ______________________________________________________________ ______________________________________________________________ ______________________________________________________________ @(#)SPMlab2.txt 1.5 01/08/30