20 Years of FreeSurfer PETsurfer: MRI-PET Integration using FreeSurfer Nov 17, 2017 Douglas N. Greve, Ph.D. Martinos Center for Biomedical Imaging Massachusetts General Hospital Harvard Medical School

PET Issues Uptake – FDG, Amyloid, Tau; 1 meas, ~20min “Quantitative” – binding potential, metabolism Multiple frames, ~1min, 60-120min Kinetic Models Noisy – requires spatial averaging (ROI, smooth) Partial Volume Effects (PVEs) Like spatial smoothing, 4-15mm FWHM Interaction with anatomy – confound Partial volume correction (PVC)

MRI: FreeSurfer Can Help … Whole-Brain (soon whole-head) Segmentation Subcortical (3D ROIs) Cortical Surface parcellations (2D ROIs) Thickness (3D ROIs, subvoxel accuracy) Surface smoothing (instead of volume) Subject-specific

Two studies done with PETsurfer ROI: FDG vs Aging Non-quantitative ROI-based PVC Voxel-wise: Serotonin Receptor Density Quantitative Voxel-wise PVC

ROI FDG Study of Aging What is the effect of aging on metabolism? With and without Partial Volume Cor (PVC)?

PET Suffers from Partial Volume Effects Anatomy Ideal FDG Real FDG ~Smoothed by 6mm Point Spread Function (PSF) Partial Volume Effect

No PVC: Just average FDG within an ROI

FDG vs Age (No PVC) Widespread decreases in metabolism with age. Is it real?

ROI-based Partial Volume Cor (PVC) Linear model of whole head FDG Each ROI is a regressor Model PVE as linear smoothing Solve to get uptake in each ROI without PVE “Geometric Transfer Matrix” GTM

NoPVC vs PVC GTM-PVC No PVC

What does this happen? A B CSF 20 Year Old A B CSF 80 Year Old

Voxelwise Serotonin Receptor Study Serotonin receptor type 4 (5HT4) Binding potential ~ Receptor Density

ROI Analysis Time Activity Curves Cr(t) TAC(t)

Single Voxel Analysis Single voxels are very noisy!

Voxel-wise Analysis Binding Potential

Smoothing (Volume-based)

Volume-Smoothing Group (N=16) 0.1 1.3 BP Crowns and Fundi have lower BP Banks have higher BP Insula has very large BP, contamination from putamen Volume Smoothing by 10mm FWHM 17

Example 5HT4 BP Map for Group (N=16) Volume Smoothed by 10mm FWHM 0.1 1.3 BP Surface Smoothed by 10mm FWHM Gyral pattern (mostly) gone Insula (mostly) artifact gone Surface has higher BP values 18

Voxel-wise PVC Method (Muller-Gartner) Source MRI WM Uptake Voxel-wise Estimate of WM Signal WM WM Smth X GM GM Smth - Div S + Kinetic Modeling Source PET

Muller-Gartner PVC Method Only performs PVC in GM Some areas where value is 0 (no GM) Division by smoothed GM

MG PVC Results (N=16) Volume Smth Surface Smth MGPVC NoPVC In general BP increases after PVC by 30-60% PVEs might actually be worse with volume smoothing PVEs get a little better with surface smoothing

Summary FreeSurfer helps … PET requires accurate segmentation to analyze and correct for PVEs Voxel-wise analysis requires surface-based analysis, especially with PVC Interaction between PET and anatomy can cause a confound without PVC FreeSurfer helps …

Thanks! Use FreeSurfer Be Happy Bruce Fischl Gitte Knudsen Bruce Rosen David Salat Keith Johnson Aaron Schultz Gitte Knudsen Claus Svarer Vincent Beliveau Melanie Ganz FreeSurfer Team

End of Presentation

Voxelwise Summary Voxel-wise time courses very noisy Smoothing needed Volume smoothing exacerbates PVEs Surface smoothing does not Voxel-wise PVC Uses FreeSurfer ROIs and tissue types Valid in GM only Volume smoothing a disaster in cortex Surface smoothing the only way to smooth PVC

Partial Volume Correction (PVC) Use MRI to segment tissue types and structures Geometric Transfer Matrix (GTM) Rousset, et al, (1998) Purely ROI-based Map-based, voxel-wise Mueller-Gartner Method (1992) Meltzer Method (1996) Often used to perform ROI analysis but less efficient and less accurate than GTM Region-wise Voxel-based (RBV, Thomas 2011) Built on GTM Need to know the point spread function of the scanner (FWHM)

Summary: ROI FDG Study of Aging FreeSurfer generated subject-specific ROIs Subcortical Cortical (subvoxel) Accurate segmentation needed for PVC PVC needed to prevent interaction between FDG uptake and anatomy Non-sparsity makes interaction worse

Kinetic Modeling Analysis (Pseudo-Linear) TAC(t) is on both sides of the equation Not really a linear model Noise in X and y Noise-dependent bias 28

Not Just Artifact Reduction, Noise too Volume Smoothed by 10mm FWHM Surface Smoothed by 10mm FWHM 0.1 1.3 BP 0.01 0.5 CoV Factors of 20-200% reduction in intersubject stddev 29

Voxel-wise Analysis (Individual) No Smoothing Volume Smoothing (5mm) Surface Smoothing (5mm) Surface Smoothing (10mm)

Volume Smoothing near Gyral Crowns Volume Smoothing by 10mm FWHM 31

Volume Smoothing near Gyral Fundi Volume Smoothing by 10mm FWHM 32

Volume Smoothing near Gyral Banks Volume Smoothing by 10mm FWHM 33

Surface (2D) Smoothing 34

Smoothing (Volume-based)