Introduction to FreeSurfer Presented by Sarah Whittle & Dominic Dwyer 19th March 2009

Talk Outline Introduction Individual Cortical (surface-based) and Volumetric Analysis Group Analyses (brief) Structure-Function Integration (brief) Working with Freesurfer

Intro: What can we do with Freesurfer? Surface inflation and manipulation: Visualize structural and functional data; reveal data in sulcal depths Intersubject registration: Alternate spatial normalization Morphometric analysis: Cortical thickness; analysis of folding patterns Cortical Parcellation: Analysis of cortical subregions; fMRI ROI analysis Subcortical segmentation: Volumetric analysis; fMRI ROI analysis White matter parcellation: Volumetric analysis; DTI region of interest analysis Integrate with FSL tools: Spatial normalization of fMRI data; ROI analyses

Intro: FreeSurfer Resources FreeSurfer Wiki can be VERY useful: http://surfer.nmr.mgh.harvard.edu/fswiki “can”: too much info & not 100% logically organised! 2008 Brisbane FSL/FreeSurfer workshop booklet available to loan from MNC lab

Cortical (surface-based) Analysis Surface Reconstruction Theory Input: T1-weighted (MPRAGE,SPGR) Segment white matter from rest of brain. Find white/gray surface Find pial surface “Find” = create mesh Vertices, neighbors, triangles, coordinates Accurately follows boundaries between tissue types “Topologically Correct” closed surface, no donut holes no self-intersections Subcortical Segmentation along the way

Surface Model Mesh (“Finite Element”) Vertex = point of 6 triangles Neighborhood XYZ at each vertex Triangles/Faces ~ 150,000 Area, Distance Curvature, Thickness Moveable The cortical surface is represented by a finite element model using triangles to cover the cortical surface. The reconstruction is the (mostly automated) process of assigning an xyz to each corner of each triangle. Once the xyz of each corner of each triangle is known, then it is possible to characterize the entire cortical surface in terms of area, distance, curvature, and thickness.

White Matter Surface Nudge orig surface Follow T1 intensity gradients Smoothness constraint Vertex Identity Stays

Pial Surface Nudge white surface Follow T1 intensity gradients Vertex Identity Stays

Cortical Thickness pial surface Distance between white and pial surfaces One value per vertex Surface-based more accurate than volume-based white/gray surface lh.thickness, rh.thickness

Curvature (Radial) Circle tangent to surface at each vertex Curvature measure is 1/radius of circle One value per vertex Signed (sulcus/gyrus) Actually use gaussian curvature lh.curv, rh.curv

Rosas et al., 2002 Sailer et al., 2003 Kuperberg et al., 2003 Fischl et al., 2000 Gold et al., 2005 Applications of thickness measurement Salat et al., 2004 Rauch et al., 2004

Surface “Inflation” Inflated Sphere Vertex Identity (index) Preserved All surfaces have the same number of vertices. Different surfaces are made by moving the triangles around to achieve some goal. There is a one-to-one correspondence between surfaces which allows the user to cross-reference vertices from one surface to another (even across subjects) and to the volume. This feature is critical to get properly oriented. In the flat surface, there’s still a one-to-one correspondence, but not all vertices in the orig surface may be represented. Vertex Identity (index) Preserved White Pial

Non-Cortical Areas of Surface Amygdala Amygdala, Putamen, Hippocampus, Caudate, Ventricles, CC

“Spherical” Registration Sulcal Map Spherical Inflation High-Dimensional Registration to Spherical Template

Inter-Subject Registration of Cortical Folding Patterns More in group analysis section

Volume Analysis: Automatic Individualized Segmentation

ROI Atlas Creation Hand label N data sets Volumetric (subcortical): CMA Surface Based: Desikan/Killiany Destrieux Map labels to common coordinate system (using spherical registration). Probabilistic Atlas Probability of a label at a vertex/voxel (global spatial info) Sulcal & gyral geometry Neighborhood relationships You can create your own atlases Incorporates the probable location of a ROI + potential inter-subject variance of the location of the region, derived from the training set employed. Curvature info used to

Automatic Labeling Transform ML labels to individual subject* Adjust boundaries based on Curvature/Intensity statistics Neighborhood relationships Result: labels are customized to each individual. * Formally, we compute maximum a posteriori estimate of the labels given the input data

Why not just register to an ROI Atlas? 12 DOF (Affine) ICBM Atlas

Subject 2 aligned with Subject 1 Problems with Affine (12 DOF) Registration ROIs need to be individualized. Subject 2 aligned with Subject 1 (Subject 1’s Surface) Subject 1

Can’t segment on intensity alone Fischl – 2002 paper in Neurotechnique

Volumetric Segmentation (aseg) Caudate Pallidum Putamen Amygdala Hippocampus Lateral Ventricle Thalamus White Matter Cortex Not Shown: Nucleus Accumbens Cerebellum Whole Brain Segmentation: Automated Labeling of Neuroanatomical Structures in the Human Brain, Fischl, B., D.H. Salat, E. Busa, M. Albert, M. Dieterich, C. Haselgrove, A. van der Kouwe, R. Killiany, D. Kennedy, S. Klaveness, A. Montillo, N. Makris, B. Rosen, and A.M. Dale, (2002). Neuron, 33:341-355.

Volumetric Segmentation Atlas Description 39 Subjects 14 Male, 39 Female Ages 18-87 Young (1-22): 10 Mid (40-60): 10 Old Healthy (69+): 8 Old Alzheimer's (68+): 11 Siemens 1.5T Vision (Wash U) Whole Brain Segmentation: Automated Labeling of Neuroanatomical Structures in the Human Brain, Fischl, B., D.H. Salat, E. Busa, M. Albert, M. Dieterich, C. Haselgrove, A. van der Kouwe, R. Killiany, D. Kennedy, S. Klaveness, A. Montillo, N. Makris, B. Rosen, and A.M. Dale, (2002). Neuron, 33:341-355.

Automatic Surface Parcellation: Desikan/Killiany Atlas Precentral Gyrus Postcentral Gyrus Superior Temporal Gyrus An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest, Desikan, R.S., F. Segonne, B. Fischl, B.T. Quinn, B.C. Dickerson, D. Blacker, R.L. Buckner, A.M. Dale, R.P. Maguire, B.T. Hyman, M.S. Albert, and R.J. Killiany, (2006). NeuroImage 31(3):968-80.

Desikan/Killiany Atlas 40 Subjects 14 Male, 26 Female Ages 18-87 34 cortical regions Siemens 1.5T Vision (Wash U) An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest, Desikan, R.S., F. Segonne, B. Fischl, B.T. Quinn, B.C. Dickerson, D. Blacker, R.L. Buckner, A.M. Dale, R.P. Maguire, B.T. Hyman, M.S. Albert, and R.J. Killiany, (2006). NeuroImage 31(3):968-80.

Automatic Surface Parcellation: Destrieux Atlas Automatically Parcellating the Human Cerebral Cortex, Fischl, B., A. van der Kouwe, C. Destrieux, E. Halgren, F. Segonne, D. Salat, E. Busa, L. Seidman, J. Goldstein, D. Kennedy, V. Caviness, N. Makris, B. Rosen, and A.M. Dale, (2004). Cerebral Cortex, 14:11-22.

Automatic Surface Parcellation: Destrieux Atlas 58 Parcellation Units 12 Subjects Automatically Parcellating the Human Cerebral Cortex, Fischl, B., A. van der Kouwe, C. Destrieux, E. Halgren, F. Segonne, D. Salat, E. Busa, L. Seidman, J. Goldstein, D. Kennedy, V. Caviness, N. Makris, B. Rosen, and A.M. Dale, (2004). Cerebral Cortex, 14:11-22.

Gyral White Matter Segmentation + + aparc+aseg wmparc Nearest Cortical Label to point in White Matter

Surface-based group analysis Processing Stages Specify Subjects and Surface measures Assemble Data (mris_preproc): Resample into Common Space (fsaverage) Smooth Concatenate into one file Model and Contrasts (GLM) Fit Model (Estimate) (mri_glmfit) Correct for multiple comparisons Visualize (tksurfer)

Surface-based Measures Morphometric (eg, thickness) Functional PET MEG/EEG Diffusion (?) sampled just under the surface

Surface-based Group Analysis in FreeSurfer GLM Command-line based Create a FreeSurfer Group Descriptor File (FSGD) FreeSurfer creates design matrix You still have to specify contrasts Fit model (mri_glmfit) QDEC (GUI) Limited to 2 discrete variables, 2 levels max Limited to 2 continuous variables

Visualisation with tksurfer Saturation: -log10(p), Eg, 5=.00001 Threshold: -log10(p), Eg, 2=.01 False Discovery Rate, Eg, .01 View->Configure->Overlay http://surfer.nmr.mgh.harvard.edu/docs/ftp/pub/docs/freesurfer.groupanalysis.pdf File->LoadOverlay

Function-Structure Integration in FreeSurfer Why Is a Model of the Cortical Surface Useful? Local functional organization of cortex is largely 2-dimensional! Eg, functional mapping of primary visual areas: Analysis of fMRI data using cortical surface models offers at least three advantages over more conventional 3-dimensional analysis methods. First, cortical surface models allow for better visualization of activations, providing a more global view than single slices and a better view of the spatial extent of activation foci and their locations relative to each other and to sulcal/gyral landmarks (Dale and Sereno, 1993). Second, statistical methods for the analysis of single subject data can benefit from the exclusion of non-gray matter signals, and smoothing signals along the cortical surface, rather than in 3D results in superior resolution and sensitivity (Kiebel et al., 2000; Andrade et al., 2001; Formisano et al., 2004). Finally, group analysis with cortical surface models employs inter-subject alignment based on the patterns of sulci and gyri, as opposed to Talairach registration, which often ignores sulcal/gyral landmarks and tends to blur activity across neighboring banks of a sulcus (Fischl et al., 1999a,b). The highly folded nature of the cortical surface also makes it difficult to view functional activity in a meaningful way. The typical means of visualization of this type of data is the projection of functional activation onto a set of orthogonal slices. This procedure is problematic as regions of activity which are close together in the volume may be relatively far apart in terms of the distance measured along the cortical surface. In addition, the naturally two-dimensional organization of cortical maps is largely obscured by the imposition of an external coordinate system in the form of orthogonal slices. These problems have led an increasing number of studies to make use of surface-based techniques for visualization. Yellow = mirror image representation of visual field (eg, V1) Blue = non-mirror image (eg, V2) Convenient way to draw borders between areas because adjoining areas often have the opposite visual field sign. Responses to phase-encoded retinal stimulation were recorded with fMRI – analysed with visual field sign method to distinguish mirror- from non-mirror image representations. Allowed accurate tracing of borders between V1, V2, Vp, V3, & V4 in human brain. Also, smooth along surface From (Sereno et al, 1995, Science).

Function-Structure Integration in FreeSurfer Basic Overview of process: First: analyze your data with FEAT (No Smoothing) Register FEAT to FreeSurfer Anatomical Automatic (FLIRT) Manual (tkregister2) Sample FEAT output on the surface Individual Common Surface Space (Atlas/fsaverage) ** Can display any functional data, eg, zstat, fzstat, cope, pe, etc

Function-Structure Integration in FreeSurfer Basic Overview of process (cont’d): Mapping FreeSurfer Segmentations to FEAT ie, displaying functional data on subcortical/cortical segmentation ROI analysis based on segmentation Group Analysis Using GFEAT data Mri_glmfit See Comprehensive Instructions at: http://surfer.nmr.mgh.harvard.edu/docs/ftp/pub/docs/freesurfer.feat.pdf

Working with Freesurfer Unix command-line (Linux, MacOSX) GUI’s for viewing/editing Tkmedit, tksurfer, tkregister Directory structure, naming conventions Pipeline Recon-all: automated surface/volume analysis

File Formats Can store 4D (like NIFTI) FreeSurfer uses a unique file format (mgz = compressed MGH file) Can store 4D (like NIFTI) cols, rows, slices, frames Generic: volumes and surfaces Surface: lh.white Curv: lh.curv, lh.sulc, lh.thickness Annotation: lh.aparc.annot Label: lh.pericalcarine.label Unique to FreeSurfer FreeSurfer can read/write: NIFTI, Analyze, MINC FreeSurfer can read: DICOM, Siemens IMA, GE, AFNI

FreeSurfer Directory Tree Subject ID $SUBJECTS_DIR bert fred jenny margaret …

FreeSurfer Directory Tree Each data set has its own unique SubjectId (eg, bert) Subject ID Subject Name bert bem label morph mri scripts surf tiff label orig T1 brain wm aseg Set SUBJECTS_DIR, issue a command (mksubjdirs) to create directory structure. Convert (using mri_convert) data to COR format and store in mri/RRR, run recon-all. Typically, we acquire 3 high-quality T1 volumes to use as input to reconstruction (128-slice sagital, 1 mm in-plane resolution).

Add Your Data bert bem label morph mri scripts surf tiff label cd $SUBJECTS_DIR mkdir –p bert/orig mri_convert yourdicom.dcm bert/mri/orig/001.mgz mri_convert yourdicom.dcm bert/mri/orig/002.mgz bert bem label morph mri scripts surf tiff label orig 001.mgz 002.mgz

Fully Automated Reconstruction 1. Create directory for data: mkdir –p $SUBJECTS_DIR/bert/orig 2. Copy/Convert data into directory: mri_convert file.dcm $SUBJECTS_DIR/bert/orig/001.mgz 3. Launch reconstruction: recon-all –s bert –autorecon-all Come back in 48 hours … Check your results – do the white and pial surfaces follow the boundaries? -- Can be broken up

Individual Stages Green = Manual Intervention? recon-all -help Volumetric Processing Stages (subjid/mri): 1. Motion Cor, Avg, Conform (orig.mgz) 2. Talairach transform computation 3. Non-uniform inorm (nu.mgz) 4. Intensity Normalization 1 (T1.mgz) 5. Skull Strip (brain.mgz) 6. EM Register (linear volumetric registration) 7. CA Intensity Normalization 8. CA Non-linear Volumetric Registration 9. CA Label (Volumetric Labeling) (aseg.mgz) 10. Intensity Normalization 2 (T1.mgz) 11. White matter segmentation (wm.mgz) 12. Edit WM With ASeg 13. Fill and cut (filled.mgz) Surface Processing Stages (subjid/surf): 14. Tessellate (?h.orig) 15. Smooth1 (?h.smoothwm) 16. Inflate1 (?h.inflated) 17. QSphere (?h.qsqhere) 18. Automatic Topology Fixer (?h.orig) 19. Euler Number 20. Smooth2 21. Inflate2 22. Final Surfs (?h.white,?h.pial) 23. Cortical Ribbon Mask 24. Spherical Morph 25. Spherical Registration 26. Spherical Registration 27. Map average curvature to subject 28. Cortical Parcellation (Labeling) 29. Cortical Parcellation Statistics 30. Cortical Parcellation mapped to ASeg Green = Manual Intervention? recon-all -help Note: ?h.orig means lh.orig or rh.orig

Workflow in Stages recon-all –autorecon1 (Stages 1-5) Check talairach transform, skull strip, normalization (?) recon-all –autorecon2 (Stages 6-23) Check surfaces Add control points: recon-all –autorecon2-cp (Stages 10-23) Edit wm.mgz: recon-all –autorecon2-wm (Stages 13-23) Edit brain.mgz: recon-all –autorecon2-pial (Stage 23) recon-all –autorecon3 (Stages 24-30) Note: all stages can be run individually Roughly 15% need alterations/edits

Results Volumes Surfaces Surface Overlays ROI Summaries

Volumes Volume Viewer: tkmedit orig.mgz T1.mgz brainmask.mgz wm.mgz filled.mgz Subcortical Mass $SUBJECTS_DIR/bert/mri All “Conformed” 2563, 1mm3 Many more … aseg.mgz aparc+aseg.mgz Volume Viewer: tkmedit

Surfaces orig white pial inflated sphere,sphere.reg patch (flattened) All surfaces have the same number of vertices. Different surfaces are made by moving the triangles around to achieve some goal. There is a one-to-one correspondence between surfaces which allows the user to cross-reference vertices from one surface to another (even across subjects) and to the volume. This feature is critical to get properly oriented. In the flat surface, there’s still a one-to-one correspondence, but not all vertices in the orig surface may be represented. inflated sphere,sphere.reg patch (flattened) $SUBJECTS_DIR/bert/surf Number/Identity of vertices stays the same (except patches) XYZ Location Changes Flattening not done as part of standard reconstruction Surface Viewer: tksurfer

Surface Overlays lh.sulc on inflated lh.curv on inflated lh.thickness on inflated lh.sulc on pial lh.curv on inflated fMRI on flat All surfaces have the same number of vertices. Different surfaces are made by moving the triangles around to achieve some goal. There is a one-to-one correspondence between surfaces which allows the user to cross-reference vertices from one surface to another (even across subjects) and to the volume. This feature is critical to get properly oriented. In the flat surface, there’s still a one-to-one correspondence, but not all vertices in the orig surface may be represented. Value for each vertex Color indicates value Color: gray,red/green, heat, color table Rendered on any surface fMRI/Stat Maps too lh.aparc.annot on inflated

ROI Summaries: $SUBJECTS_DIR/bert/stats aseg.stats – volume summaries ?h.aparc.stats – desikan/killiany parcellation summaries ?h.aparc.2005.stats – destrieux parcellation summaries wmparc.stats – white matter parcellation Index SegId NVoxels Volume_mm3 StructName normMean normStdDev normMin normMax normRange 1 1 0 0.0 Left-Cerebral-Exterior 0.0000 0.0000 0.0000 0.0000 0.0000 2 2 265295 265295.0 Left-Cerebral-White-Matter 106.6763 8.3842 35.0000 169.0000 134.0000 3 3 251540 251540.0 Left-Cerebral-Cortex 81.8395 10.2448 29.0000 170.0000 141.0000 4 4 7347 7347.0 Left-Lateral-Ventricle 42.5800 12.7435 21.0000 90.0000 69.0000 5 5 431 431.0 Left-Inf-Lat-Vent 66.2805 11.4191 30.0000 95.0000 65.0000 6 6 0 0.0 Left-Cerebellum-Exterior 0.0000 0.0000 0.0000 0.0000 0.0000 …. Routines to generate spread sheets of group data asegstats2table --help aparcstats2table --help