Avoiding Bias in Longitudinal Image Processing © Martin Reuter mreuter@nmr.mgh.harvard.edu http://reuter.mit.edu Quantifying Imaging Biomarkers Here we focus on www.FreeSurfer.net • measure volume of cortical or subcortical structures • compute thickness (locally) of the cortical sheet • study differences of populations (diseased, control) Neurodegenerative disorder: 14 time points, 6 years, Huntington’s Disease We'd like to: • exploit longitudinal information (same subject, different time points)) Why longitudinal processing? • to reduce variability on intra-individual morph. estimates • to detect small changes, or use less subjects (power) • for marker of disease progression (atrophy) • to better estimate time to onset of symptoms • to study effects of drug treatment ... [Reuter et al, NeuroImage 2012] Example (cross vs. long) Example (subject as own control) Challenges in Longitudinal Designs 1. Over-Regularization: • Temporal smoothing • Non-linear warps (regularization) Ø Potentially underestimate change 2. Bias [Reuter and Fischl, NeuroImage 2011] , [Reuter et al. NeuroImage 2012] • Interpolation Asymmetries [Yushkevich et al., NeuroImage 2010], [Thompson, Holland, NeuroImage 2011], [Fox, Ridgway, Schott, NeuroIm. 2011] • Asymmetric Information Transfer Ø Often overestimate change Interpolation Asymmetries (Bias) Mapping follow-up to baseline: • Keeps baseline image fixed (crisp) • Causes interpolation artifacts in follow-up (smoothing) • Often leads to overestimating change Interpolation Artifacts • Replicate by mapping an image somewhere and then back (will interpolate twice). • For example register A->B to get some transform T, then map A via T and use the inverse transform to map it back. mri_robust_register --mov A.nii --dst B.nii --lta T.lta --satit mri_convert -at T.lta A.nii AatB.nii mri_convert -ait T.lta AatB.nii AatA.nii freeview -v A.nii AatA.nii • Now analyze AatA as second time point with respect to A Interpolation Asymmetries (Bias) Left Subcortical Structures Right Subcortical Structures 10 Sym. Pct. Volume Change Sym. Pct. Volume Change 10 5 0 −5 −10 5 0 −5 −10 Thal Caud Put Pal Hippo Amyg CortGM WM Thal Caud Put Pal Hippo Amyg CortGM WM MIRIAD dataset: 65 subjects First session first scan compared to twice interpolated image. http://miriad.drc.ion.ucl.ac.uk Interpolation Asymmetries (Bias) Left Cortical ROIs Left White Matter ROIs 15 −5 Sym. Pct. Volume Change Sym. Pct. Volume Change 0 −10 −15 10 5 −20 −25 Cuneus PreCun CaMidFr SupFron PreCen SupTmp ParaHip InfPar 0 Cuneus PreCun CaMidFr SupFron PreCen SupTmp ParaHip InfPar MIRIAD dataset: 65 subjects First session first scan compared to twice interpolated image. http://miriad.drc.ion.ucl.ac.uk Tri-Linear vs. Cubic B-Spline Interpolation (Bias) Left Cortical ROIs Left Subcortical Structures [Tri−Linear] (65) [Cubic] (65) 0 Sym. Pct. Volume Change Sym. Pct. Volume Change 10 5 0 −5 −5 −10 −15 −20 −10 [Tri−Linear] (65) [Cubic] (65) Thal Caud Put Pal Hippo Amyg CortGM WM −25 Cuneus PreCun CaMidFr SupFron PreCen SupTmp ParaHip InfPar Regional! Not finding it does not mean it is not there! We need to treat all time points the same! Robust Registration [Reuter et al., NeuroImage, 2010] Inverse consistency: • a symmetric displacement model: 1 ⎞ S⎛ 1 ⎞ T⎛ r ( p) = I ⎜ x − d ( p) ⎟ − I ⎜ x + d ( p) ⎟ 2 2 ⎝ ⎠ ⎝ ⎠ • resample both source and target to an unbiased half-way space in intermediate steps (matrix square root) M Source Half-Way Target M −1 Robust Registration [Reuter et al., NeuroImage, 2010] Tumor data courtesy of Dr. Greg Sorensen Tumor data with significant intensity differences in the brain, registered to first time point (left). Robust Registration [Reuter et al 2010] Target Target Robust Registration [Reuter et al 2010] Registered Src FSL FLIRT Registered Src Robust mri_robust_register • mri_robust_register is part of FreeSurfer • can be used for pair-wise registration (optimally within subject, within modality) • can output results in half-way space • can output ‘outlier-weights’ • see also Reuter et al. “Highly Accurate Inverse Consistent Registration: A Robust Approach”, NeuroImage 2010. http://reuter.mit.edu/publications/ • for more than 2 images: mri_robust_template Asymmetric Information Transfer Example: 1. Process baseline 2. Transfer results from baseline to follow-up 3. Let procedures evolve in follow-up (or construct skullstrip in baseline, or Talairach transform …) Can introduce bias! [Reuter 2011, 2012] Robust Unbiased Subject Template 1. Create subject template (iterative registration to median) 2. Process template 3. Transfer to time points 4. Let it evolve there - All time points are treated the same - Minimize overregularization by letting tps evolve freely [Reuter et al., NeuroImage, 2012] Robust Template Estimation • Minimization problem for N images: ˆ 'ˆi } := argmin {I, I,'i N X E(Ii 'i , I) + D('i )2 i=1 • Image Dissimilarity: Z E(I1 , I2 ) = |I1 (x) I2 (x)| dx ⌦ • Metric of Transformations: D(~t, r)2 =k ~t k2 + k R 1 k2F Biased Information Transfer MIRIAD Data (within session at baseline) Biased Information Transfer MIRIAD Data (within session at baseline) Bias in Subcortical Volumes (MIRIAD Session) 6 6 [BASE1] (65) [BASE2] (65) [FS−LONG] (65) [FS−LONG−rev] (65) 2 0 −2 2 0 −2 −4 −4 −6 −6 LPallidum LHippoc LAmygdala RPallidum RHippoc RAmygdala PreCen 0 −2 −4 LatOcc Cuneus PreCun [BASE1] (65) [BASE2] (65) [FS−LONG] (65) [FS−LONG−rev] (65) 4 Sym. Pct. Volume Change Sym. Pct. Volume Change 2 ParaHip 6 [BASE1] (65) [BASE2] (65) [FS−LONG] (65) [FS−LONG−rev] (65) 4 CACing Bias in Left Cortical GM Volumes (MIRIAD Session) Bias in Left Cortical GM Volumes (MIRIAD Session) 6 [BASE1] (65) [BASE2] (65) [FS−LONG] (65) [FS−LONG−rev] (65) 4 Sym. Pct. Volume Change 4 Sym. Pct. Volume Change Bias in Left Cortical GM Volumes (MIRIAD Session) 2 0 −2 −4 −6 −6 CaMidFr SupTmp InfPar Lingual MedOrbFr MidTemp ParaCen Perical PostCen SupPari SupMarg TransTemp How and Why: How to minimize over regularization: ü Only initialize processing, evolve freely How to avoid processing bias: ü Treat all time points the same Why not simply do independent processing then? Ø Increase reliability, statistical power Ø See for yourself … Test-Retest Reliability [Reuter et al., NeuroImage, 2012] Subcortical Cortical Left Subcortical Structures TT−115 Left Cortical Gray Matter Parcellation TT−115 8 [CROSS] (115) [LONG] (115) 7 Abs. Sym. Pct. Volume Change Abs. Sym. Pct. Volume Change 7 8 6 5 4 3 2 1 0 [CROSS] (115) [LONG] (115) 6 5 4 3 2 1 Thalamus Caudate Putamen Pallidum Hippocamp Amygdala 0 Cuneus PreCun CaMidFr SupFron PreCen SupTmp ParaHip InfPar [LONG] significantly improves reliability 115 subjects, ME MPRAGE, 2 scans, same session Test-Retest Reliability [Reuter et al., NeuroImage, 2012] Diff. ([CROSS]-[LONG]) of Abs. Thick. Change: Significance Map [LONG] significantly improves reliability 115 subjects, ME MPRAGE, 2 scans, same session Increased Power [Reuter et al., NeuroImage, 2012] Left Hemisphere: Right Hemisphere 100 90 80 70 60 50 40 30 20 10 0 Sample Size Reduction (Right Hemisphere) Percent Subjects needed (LONG vs. CROSS) Percent Subjects needed (LONG vs. CROSS) Sample Size Reduction (Left Hemisphere) Thalamus Caudate Putamen Pallidum Hippocamp Amygdala 100 90 80 70 60 50 40 30 20 10 0 Thalamus Caudate Putamen Pallidum Hippocamp Amygdala Sample Size Reduction when using [LONG] Huntington’s Disease (3 visits) [Reuter et al., NeuroImage, 2012] Independent Processing Longitudinal Processing Atrophy in Huntington’s Disease [CROSS] Atrophy in Huntington’s Disease [LONG] 2 [CN] (10) [PHDfar] (16) [PHDnear] (19) [HD] (9) 1 0 −1 −2 −3 −4 LThalamus LCaudate LPutamen RThalamus RCaudate RPutamen 3 Pct. Volume Change (per year w.r.t. baseline) Pct. Volume Change (per year w.r.t. baseline) 3 2 [CN] (10) [PHDfar] (16) [PHDnear] (19) [HD] (9) 1 0 −1 −2 −3 −4 LThalamus LCaudate LPutamen RThalamus RCaudate RPutamen [LONG] shows higher precision and better discrimination power between groups (specificity and sensitivity). Huntington’s Disease (3 visits) [Reuter et al., NeuroImage, 2012] Rate of Atrophy Baseline Vol. (normalized) Atrophy in Huntington’s Disease [LONG] −3 2 [CN] (10) [PHDfar] (16) [PHDnear] (19) [HD] (9) 1 0 −1 −2 −3 6 Stucture Volume / Intracranial Volume Pct. Volume Change (per year w.r.t. baseline) 3 5 x 10 Volume at Baseline in Huntington’s Disease CN (10) PHDfar (16) PHDnear (19) HD (9) 4 3 2 1 −4 LThalamus LCaudate LPutamen RThalamus RCaudate RPutamen 0 LThalamus LCaudate LPutamen RThalamus RCaudate RPutamen Putamen atrophy rate is significant between CN and PHD far, but baseline volume is not. Sources of Bias during Acquisition BAD: these influence the images directly and cannot be easily removed! • Different Scanner Hardware (Headcoil, Pillow?) • Different Scanner Software (Shimming Algorithm) • Scanner Drift and Calibration • Different Motion Levels Across Groups • Different Hydration Levels (season, time of day) Hydration Levels 14 subjects, 12h dehydration, rehydration 1L/h [with A. Bartsch et al. – submitted] Links: 1. Software and Wiki: http://freesurfer.net • http://freesurfer.net/fswiki/LongitudinalProcessing 2. Facebook: http://facebook.com/FreeSurferMRI 3. Data (MIRIAD): http://miriad.drc.ion.ucl.ac.uk 4. Publications: • • Reuter, Rosas, Fischl: Highly Accurate Inverse Consistent Registration: A Robust Approach. NeuroImage 53(4): 1181-1196, 2010. http://reuter.mit.edu/papers/reuter-robreg10.pdf Reuter et al.: Within-Subject Template Estimation for Unbiased Longitudinal Image Analysis. NeuroImage 61(4): 1402-1418, 2012. http://reuter.mit.edu/papers/reuter-long12.pdf