Weierstrass Institute for Applied Analysis and Stochastics Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging Jörg Polzehl (joint work with Karsten Tabelow and Saskia Becker) Mohrenstrasse 39 · 10117 Berlin · Germany · Tel. +49 30 20372 0 · www.wias-berlin.de Neuroimaging Data Analysis, SAMSI, June 9, 2013 Outline MR Physics Data properties and random effects Modeling Adaptive smoothing Examples Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 2 (36) Magnetic resonance imaging (MRI) Figure: Kasuga Huang (Wikimedia) Figure: Franz Wilhelmstötter (Wikimedia) From O. Friman “Adaptive Analysis of Functional MRI Data”, PhD Thesis, 2003 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 3 (36) Magnetic resonance imaging (MRI) Many different imaging contrasts “There is nothing that nuclear spins will not do for you, as Figure: Kasuga Huang (Wikimedia) long as you treat them as human beings” (Erwin L. Hahn 1949) Figure: Franz Wilhelmstötter (Wikimedia) From O. Friman “Adaptive Analysis of Functional MRI Data”, PhD Thesis, 2003 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 3 (36) Magnetic resonance imaging (MRI) T2 -weighted T1 -weighted Diffusion-weighted thanks to: F. Godtliebsen (University Tromsœ), H.U. Voss (Weill Cornell Medical College, NY) and M. Deppe (Uniklinikum Münster) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 3 (36) White matter anatomy White matter characterized by fiber bundles (axon bundles) diameter 30 − 50µm length up to 20 − 25cm highly anisotropic structures Source: John A Beal, PhD Dep’t. of Cellular Biology & Anatomy, Louisiana State University Health Sciences Center Shreveport; Wikipedia Commons MRI can be used to measure water diffusion water diffusion is restricted by anatomic structure Beaulieu, C. The basis of anisotropic water diffusion in the nervous system - a technical review NMR in Biomedicine, 2002, 15, 435-455 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 4 (36) Diffusion MRI Figure: Thomas Schultz (Wikimedia) Notation (Diffusion Propagator) P (~r, ~r 0 , τ ) - probability density for a particle (spin) to “travel” from position ~r 0 to ~r in time τ p(~r 0 ) is the initial probability density of particle location Random walk, diffusion process V (Ensemble Averaged Propagator, EAP): Z ~ τ) = P (R, P (~r, ~r 0 , τ ) p(~r 0 ) d~r 0 , Aggregate over a voxel ~ r −~ ~ r 0 ∈V, R=~ r0 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 5 (36) Diffusion MRI signal and diffusion propagator Diffusion gradients lead to signal attenuation due to diffusion process - loss of phase coherence between precessing spins: S(~ q , τ ) = S0 hexp(iϕ)i Fourier relation Z S(~ q , τ ) = S0 ~ ~ τ ) ei~q R dR ~ P (R, IR3 Measure S(~ q , τ ) at N voxel locations in 3D for 3, . . . , 200 vectors ~ q 3D + S2 Spatial resolution: 0.6-2 mm 1 - ... different b-values Magnetic field strength: 3-7 T Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 6 (36) MR acquisition and noise complex signal in K-space (one coil): 2 sc (k) ∼ N (xc (k), σK ) FFT provides complex image Sc (x) ∼ N (ξc (x), σI2 ) MR image: S(x) usually obtained as norm of linear combinations of Sc from L receiver coils Notation: Si = |S(xi )| |S(x)|2 /σI2 distributed as a linear combination of noncentral χ22 RV, with spatially varying coefficients. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 7 (36) Properties of dMRI images Signal of interest in voxel location ζi = sX xi : ac (xi )ξc (xi )ξ¯c (xi ) c ac (x) depend on the reconstruction algorithm For ac (xi ) = 1 Si /σ ∼ χ2L;ζi /σ Signal density Chi_16^2 − Gaussian Approx. 0.6 Signal density Chi_2 − Gaussian Approx. 0.5 NCP 0 NCP 2 NCP 4 Signal attenuation → decrease in SNR Bias(NCP,coils) 0 1 2 3 4 x 5 6 Bias/sigma 2 3 1 coil 2 coils 4 coils 8 coils 1 0 0.0 0.0 0.1 0.1 0.2 0.2 dchi 0.3 dchi 0.3 4 0.4 0.4 0.5 NCP 0 NCP 2 NCP 4 0 2 4 6 x 8 0 2 4 6 8 NCP Strong Bias E S/σ − ζ/σ for small SN R. Models thermal noise, other sources of noise ? ac (xi ) 6= 1 ? Varying ac (xi ) lead to heteroskedasticity ..., varying effective number of coils. Noise estimation –> Aja Fernandez (2009, 2011) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 8 (36) 10 The diffusion tensor model Gaussian diffusion: (Homogeneity within a voxel, no effect of fiber structure) 1 ~ uT D−1 ~ u ~ τ ) = P (r~ exp −r2 . P (R, u, τ ) = p 4τ det D(4πτ )3 Diffusion Tensor Model: E(~ q , τ ) = E(q~ u, τ ) = e−b~u Fully characterized by the Diffusion Tensor D (or R : Estimation by nonlinear regression: R(Si ; θ, R) θ̂i R̂i T D~ u D = RR> ) Si = {Si (~0), Si (~ q1 ), . . . , Si (~ qm )} = m X (Si (~ qj ) − θ exp(−bj (~ u> uj R)> ))2 j R)(~ 2 σ j=0 = argmin R(Si ; θ, R) ! θ,R Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 9 (36) Characteristics in the diffusion tensor model Mean diffusivity T r(D) = µ1 + µ2 + µ3 = 3µ̄ Fractional anisotropy (FA) FA = 3 (µ1 − µ̄)2 + (µ2 − µ̄)2 + (µ3 − µ̄)2 2 µ21 + µ22 + µ23 1/2 Geodesic anisotropy (GA) (Fletcher (2004), Corouge (2006)) GA = 3 X (log(µi ) − log(µ))2 i=1 3 1X log(µ) = log(µi ) 3 i=1 !1/2 Color coded FA / GA maps FA / GA coded as image intensity Principal eigenvector coded in RGB Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 10 (36) Models addressing shortcomings of DTI Diffusion tensor model (DTI) Assumptions: Assumes anisotropic Gaussian diffusion does not reflect effects of fiber geometry. homogeneous fiber structure within a voxel Reality: high percentage of voxel with fiber crossings or bifurcations Consequences: Uninformative tensor estimates Reduction in FA, Biased or non-existent directional information Need a better description of the data!! Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 11 (36) Models addressing shortcomings of DTI Diffusion tensor model (DTI) Assumptions: Assumes anisotropic Gaussian diffusion does not reflect effects of fiber geometry. homogeneous fiber structure within a voxel Reality: high percentage of voxel with fiber crossings or bifurcations Consequences: Uninformative tensor estimates Reduction in FA, Biased or non-existent directional information More detailed models: Orientation distribution function (Tuch 2004, Wedeen 2005, Aganji 2010, Özarslan 2006) Positive definite EAP and ODF estimation (Cheng 2012) Tensor Mixture Models (Behrens 2003/2007, Hosey 2005, Tabelow 2012, Jian 2007, Leow 2009) Kurtosis Imaging (Özarslan 2003, Liu 2005, Hui 2008, Jensen 2010, Tabesh 2011) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 11 (36) Clinical Applications Mean diffusivity / FA MD/FA based diagnostics: decreased Mean Diffusivity –> Stroke changes of FA over time differences in FA between groups (control/patient) Surgery planning Color coded FA Color coded FA / MD Experimental questions: increased resolution reduction of recording time Both lead to significant loss in Signal to Noise Ratio (SNR). Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 12 (36) Modeling example. Data: H. Voss, Weill Cornell Medical College Tensor ODF ODF (Qball) Tensor Mixture ODF 3.0 Tesla GE Signa Excite MRI Scanner 8 Channel receive only head coil 10S0 images and 140 gradient directions T E = 73.2ms, T R = 14s 66 slices Acquisition matrix size: 128 × 128 zero filled to 256 × 256 CCFA Kurtosis Tensor Voxelsize; 0.9 × 0.9 × 1.8mm3 b-value 1000 s mm2 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 13 (36) Research problems, high resolution interest in higher resolution (leads to loss in SNR) interest in stronger signals (higher field strength 7T) high resolution Magnetom 7T Siemens. Data: R. Heidemann (MPI Leipzig/Siemens) s 1000 mm 2, FOV 143 × 147mm, 0.8mm isotropic. b-value left: 60-gradients 7 S0 images recording time 15 min. right: 240-gradients 28 S0 images recording time 65 min. Need for image enhancement / noise reduction methods preserving the structure adaptive smoothing for functional data in adaptive smoothing for image data in R3 (Tabelow et al (2009)) S 2 n R3 (Becker et al (2012, 2013)) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 14 (36) Research problems, high resolution interest in higher resolution (leads to loss in SNR) interest in stronger signals (higher field strength 7T) really bad DTI images 3T GE, 40-gradients, 3 S0 images, b-value 1000s/mm2 128x128 (1.72x1.72x0.9) 256x256 (0.86x0.86x0.9) dito 4 averages Need for image enhancement / noise reduction methods preserving the structure adaptive smoothing for functional data in adaptive smoothing for image data in R3 (Tabelow et al (2009)) 2 S n R3 (Becker et al (2012, 2013)) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 14 (36) Structural adaptive smoothing: Setup/General idea Design: x1 , . . . , xn ∈ X ⊆ IRp Yi ∼ Pθ(xi ) θ : IRp → Θ (i.i.d.) SM Structural assumption: ∃ Partitioning X = m=1 Xm such that Observations: (Polzehl & Spokoiny 2000,2006) Y1 , . . . , Yn ∈ Y ⊂ IRq θ(x) = θ(xi ) ⇔ ∃m : x ∈ Xm ∧ xi ∈ Xm i.e. θ constant on each Xm – local homogeneity structure Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 15 (36) Structural adaptive smoothing: Setup/General idea Design: x1 , . . . , xn ∈ X ⊆ IRp Yi ∼ Pθ(xi ) θ : IRp → Θ (i.i.d.) SM Structural assumption: ∃ Partitioning X = m=1 Xm such that Observations: (Polzehl & Spokoiny 2000,2006) Y1 , . . . , Yn ∈ Y ⊂ IRq θ(x) = θ(xi ) ⇔ ∃m : x ∈ Xm ∧ xi ∈ Xm i.e. θ constant on each Xm – local homogeneity structure General idea: Determine structure and estimates in an iterative procedure local partitioning described by a weighting scheme “Learn” about partitioning from estimates “Learn” about (k) (k) W (k) (xi ) = wi1 , . . . , win ∀i θ̂(xi ) by local comparisons θ(xi ) using the weighting scheme W (k) (xi ) in weighted log-likelihood estimation. ( ideally for large k: (k) wij ≈ 1 0 if ∃m : xi , xj ∈ Xm else. go from small to large scales with iterations Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 15 (36) Structural adaptive smoothing: Algorithm Propagation-Separation algorithm (0) (0) k = 0, Wi such that wij = δij , θ̂(xi ) as weighted likelihood or least squares estimate, h(0) = 1. Initialization: Step Adaptation: ∀i, j define (k) (k) wij = Kloc (∆(xi , xj )/h(k) )Ks (sij /λ) Estimation: ∀i define θ̂(k) (xi ) = arg max l(Y, Wi ; θ) θ Iterate: Stop if ∗ ( or k ≥ k , else consider next scale h (k+1) arg min R(Y, Wi ; θ)) θ = ch h(k) , set k := k + 1 and continue with adaptation. Statistical penalty: (k−1) sij = Ni KL(Pθ̂(k−1) (xj ) , Pθ̂(k−1) (xi ) ), (k) Ni (k−1) = max(Ni , n X j=1 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 16 (36) (k) wij ) Illustration Univariate local constant regression with additive Gaussian errors Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 17 (36) Illustration Univariate local constant regression with additive Gaussian errors Weighting schemes Location weights Statistical penalty Combined weights Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 18 (36) Parameters Parameters k∗ determines the maximum bandwidth ch = 1.251/p provides exponential growth of sum of location weights Kernels Kloc (z) = (1 − z 2 )+ and Ks (z) = min(1, 2(1 − z))+ Propagation condition (Becker & Mathé, 2013) Let the probability law Let θ(xi ) ≡ θ and Pθ and the metric of the design space X be specified (k) N̄i be the sum of non-adaptive weights (= without Ks ()) in step k (k) (k) Zλ (k, p, θ) = inf{z > 0 : P (N̄i K(θ̂i (λ), θ) > z) ≤ p} λ is set to be chosen according to the Propagation Condition at level if Zλ (., p, θ) is non-increasing ∀p > A value λ can be chosen by simulation using functions from our R-packages aws or dti the value of λ does not depend on the data at hand Theoretical results for exponential family models in Becker & Mathé, 2013 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 19 (36) Propagation condition Propagation condition for Poisson variates on a 3D grid. 0 5 1.08 10 1.47 15 step 20 2.24 3.12 bandwidth 25 4.49 30 6.51 35 9.44 Poisson 3 −dim. ladj= 1 Exceed. Prob. 0 5 10 15 20 25 z Level: = 0.001, Red lines correspond to the non-adaptive estimates. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 20 (36) Smoothing of Diffusion Weighted Data Two different views: DWI as functional data: consider observations related to a voxel as sample from a function on one (multiple) spheres S : R3 → Rn describe this function by a model define the statistical penalty in terms of a distance between models or their parameters DWI as data in orientation space consider observations related to a voxel as sample from a function on one (multiple) spheres S : R3 n S2 → RB+1 define the statistical penalty in Kullback-Leibler-divergences between noncentral χ-distributions. first considered in Franken (2008) and Duits and Franken (2011) using nonlinear diffusion methods Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 21 (36) Functional approach based on Tensor Model Tabelow et al 2008 Smoothing algorithm (0) (0) Initialization: k = 1, h(1) = ch . Set ζ̂b,i = Sb,i , D̂i (0) (0) , θ̂0,i Ni = 1. Adaptation: For every pair i, j compute (k−1) Ni (k) sij = λCi (g, h(k−1) ) (k) (k−1) wij = Kloc ∆(i, j, D̃i (k−1) R(ζ̂.,i (k−1) , θ̂0,j (k) (k) = Pn j=1 (k−1) ) − R(ζ̂.,i (k−1) , θ̂0,i θ̂0,i (k) D̂i (k−1) , D̂i (k) )/h(k) Kst sij /σj2 , Estimation of diffusion weighted images: Ni (k−1) , D̂j ! (k) = arg minθ,D R(ζ̂.,i , θ, D), Set (k) wij . Stopping or Iterate: Stop if k = k ∗ , otherwise h(k+1) = ch h(k) , k := k + 1 The algorithm may incorporate a Bias correction for ζ̂ . Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 22 (36) ) Comparison Adaptive smoothing provides more stable estimates without loss of structure enables to reduce recording time sensitive only to contrasts reflected by the model A: unsmoothed B: non-adaptive C: adaptive Data: M. Deppe, Univ. Münster. Color scheme codes FA intensity only. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 23 (36) Position orientation adaptive smoothing (POAS) Becker et al. (2012) Position orientation adaptive smoothing (POAS) Algorithm ( model=“chi” in dwi.smooth) Choose a sequences of bandwidths Initialization (k = 0): Adaptation (Step (0) sg1 g2 q , k)}k=0,··· ,k∗ {h(~ q , k)}k=0,··· ,k∗ and {κ(~ = 0, h(~ q , 0) = 1 k): ∀g1 , g2 ∈ R3 o S2 define adaptive weights ! (k) ∆κ (g1 , g2 )2 sg1 g2 (k) wg1 g2 =Kloc Ks h2k λ X (k) (k) (k−1) (k−1) sg1 g2 = wg1 g2 · KL Ŝg1 /σ̂, Ŝg2 /σ̂ g2 Estimation (Step k): ∀g1 ∈ R3 o S2 define weighted mean as X X Ŝg(k) := wg(k) S / wg(k) 1 1 g2 g2 1 g2 g2 ∈R3 oS2 Iterate: Stop if g2 ∈R3 oS2 k ≥ k∗ , else k := k + 1 and continue. (k−1) The KL-divergence between noncentral χ2L -distributions KL Ŝg1 (k−1) /σ̂, Ŝg2 /σ̂ needs to be approximated. Distributions are parametrized by their expected values. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 24 (36) Discrepancy function on R3 × S2 Embedding of R3 × S2 into SE(3) Notations S2 := {~ q ∈ R3 : k~ q k = 1} = 2-sphere SE(3) := R3 o SO(3) = 3-dimensional Euclidean motion group, where SO(3) := {R ∈ R3o3 : RT = R−1 , det(R) = 1} stab(~ez ) := {R ∈ SO(3) : R = rotation around the z -axis} Then it holds S2 ∼ = SO(3)/stab(~ez ) and R3 × S2 ∼ = SE(3)/(0 o stab(~ez )). F : SO(3) → R, which satisfies F (R(α,β,γ) ) = F (R(0,β,γ) ) for all α ∈ [0, 2π), can be identified one-to-one with a function f : S2 → R. Any function Parametrization: ~ e Parametrization of SO(3): R(α,β,γ) := R~eγz Rβy R~eαz ∈ SO(3) for β ∈ / {0, π} Parametrization of S2 : ~ q(β,γ) := (cos γ sin β, sin γ sin β, cos β)T ∈ S2 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 25 (36) From geodesics to a discrepancy function on R3 o S2 Riemannian 2-norm k.kR on SE(3) - Approximation kgkR ≈ inf 6 X !1/2 ki2 : i=1 6 Y i=1 exp(ki Ai ) = Mg ≡ g ∈ SE(3) using g ≡ Mg := R(α,β,γ) 0 0 0 ~v T 1 ! and the left invariant basis matrices {Ai }6i=1 . Discrepancy on R3 o S2 We set for g1 , g2 ∈ R3 × S2 with gj = (~ vj , q~j ) ∆κ (g1 , g2 ) ≈ inf 3 X where ĝ := !1/2 ki2 + κ−2 (k42 + k52 + |k6 |) i=1 : 6 Y exp(ki Ai ) = Mĝ i=1 −1 −1 2 2 Rq~ (~ v1 − ~v2 ), Rq~ Rq~1 , and κ balances between distances on S2 and in R3 . Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 26 (36) Multishell POAS, Becker et al 2013 Multi-shell data More sophisticated models for dMRI require measurements for different b-values: multi-shell data Sb : V × Gb 3 (~v , ~g ) 7→ Sb (~v , ~g ) ∈ R, b ∈ B0 := B ∪ {0} Consider a vector form S : V × G 3 (~v , ~g ) 7→ (S0 (~v ), Sb1 (~v , ~g ), ..., SbB (~v , ~g ))T ∈ RB+1 some Sb (~v , ~g ) may be unobserved Improvements: Incorporates information from different shells including S0 Improved and faster approximation for KL-divergence Simplified metric on R3 o S 2 Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 27 (36) msPOAS Multi-shell POAS algorithm (iteration only) b ∈ B and m := (~vm , ~gm ) ∈ V × Gb fill the missing values (k) (m) and Ñm,b , b0 ∈ B \ {b} by spherical interpolation Iteration: For each (k−1) of S̃b0 Compute the statistical penalty (k−1) (k) smn = X (k−1) Ñm,b KL S̃b σ̂ b∈B0 (k−1) (m) S̃b , (n) ! σ̂ , n ∈ V × Gb , and the adaptive weights (k,b) (k) w̃mn = Kloc δκ (m, n)/h(k) · Kad smn /λ , n ∈ V × Gb . Compute the adaptive estimator X (k) S̃b (m) = (k) (k,b) w̃mn Sb (n)/Ñm,b n∈V ×Gb and the corresponding sum of weights (k) Ñm,b Iterate after increasing = max 0 k ≤k X (k0 ) w̃mn . n∈V ×Gb h and adjusting κ. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 28 (36) Examples: Really bad SNR Color Coded FA maps from DTI really bad DTI images 3T GE, 40-gradients, 3 S0 images, b-value 1000s/mm2 Reconstructions by POAS, 16 Steps 128x128 (1.72x1.72x0.9) 256x256 (0.86x0.86x0.9) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 29 (36) POAS Repeated acquisition Acquisition time: 65 min vs. Acquisition time: 15 min! Acquisition time: 15 min Example: High resolution single-shell data Data: R. Heidemann, MPI for Cognitive Neuroscience Leipzig MAGNETOM 7T (Siemens), b-value of 1000s/mm2 , 60(240) directions. FOV: 143 × 147mm2 , 91 slices, isotropic resolution of800µm. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 30 (36) Example: High resolution multi-shell data Observed images from, same gradients but b-values 800 and 2000 smoothed images L = 4, σ = 30 Data: S. Mohammadi, Wellcome Trust Centre for Neuroimaging, UCL Institute of Neurology, London. 3T MAGNETOM Trio scanner (Siemens), reduced FoV-technique (Heidemann2010), FoV: 161 × 58mm around the motor cortex, isotropic in-plane resolution of 1.2mm. 34 slices of 1.3mm slice thickness. b-values: b = 800s/mm2 and b = 2000s/mm2 each with 100 gradient directions. 21 S0 images. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 31 (36) Variability-Reduction by Smoothing Mean absolute deviation of direction from Behrens 1-stick-1-ball model. b) MAD Improvements vrs. FA. d) sampled directions describing the posterior d. e) MAD orig. vs. smoothed data. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 32 (36) Effect of smoothing: Deterministic streamline fiber tracks Original data Smoothed Smoothed (roiy = 91) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 33 (36) Collaborations Cooperation: Citigroup Biomedical Imaging Center, Weill Medical College, Cornell University, NY, U.S.A. University of Münster BNIC, Charité, Berlin Max-Plank Institute for Human Cognitive and Brain Sciences, Leipzig Wellcome Trust Centre for Neuroimaging at UCL, ION UCL, London, UK R-Community: CRAN Task View: Medical Image Analysis Special volume 44 on Magnetic Resonance Imaging in R of Journal of Statistical Software Acknowledgments: We thank the Weill Medical College, Cornell University, the Max Planck Institute for Human Cognitive and Brain Sciences, the Wellcome Trust Centre for Neuroimaging at UCL, the University of Münster and the NIH/NCRR Center for Integrative Biomedical Computing (P41-RR12553) for providing functional and diffusion-weighted MR datasets. Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 34 (36) Software R-Package dti: Modeling, Smoothing, Fiber tracking, Visualization of DWI data relies on Packages oro.DICOM, oro.nifti, gsl, adimpro, rgl other related packages: fmri and aws The examples in this talk have been done using R-Package dti POAS will also be part of the ACID toolbox for SPM (S. Mohammadi) Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 35 (36) Bibliography (Selection) J. Polzehl, V. Spokoiny (2000). Adaptive Weights Smoothing with applications to image restoration, J. R. Stat. Soc. Ser. B Stat. Methodol., 62: 335–354. J. Polzehl, V. Spokoiny (2006). Local Likelihood Modeling by Adaptive Weights Smoothing. Probability Theory and Related Fields, 135: 335–362. S. Becker, P. Mathe (2013). A new perspective on the Propagation-Separation approach: Taking advantage of the propagation condition WIAS Preprint #1766 K. Tabelow, J. Polzehl, V. Spokoiny, H.U. Voss (2008). Diffusion tensor imaging: Structural adaptive smoothing. Neuroimage, 39(4): 1763–1773. K. Tabelow, H.U. Voss, J. Polzehl (2012). Modeling the orientation distribution function by mixtures of angular central Gaussian distributions. Journal of Neuroscience Methods, 203, 200-211. J. Polzehl, K. Tabelow (2011). Beyond the Gaussian Model in Diffusion-Weighted Imaging: The Package dti Journal of Statistical Software, 44(12), 1–26. S. Becker, K. Tabelow, H.U. Voss, A. Anwander, R. Heidemann, J. Polzehl (2012). Position-orientation adaptive smoothing of diffusion weighted magnetic resonance data (POAS) Medical Image Analysis, 16, 1142–1155. S. Becker, K. Tabelow, S. Mohammadi, N. Weiskopf, J. Polzehl (2013). Adaptive smoothing of multi-shell diffusion-weighted magnetic resonance data by msPOAS WIAS Preprint # 17xx Position Orientation Adaptive Smoothing (POAS) in Diffusion Weighted Imaging · Neuroimaging Data Analysis, SAMSI, June 9, 2013 · Page 36 (36)