Constructing Approximation Kernels for Non-Harmonic Fourier Data

Transcription

1 Constructing Approximation Kernels for Non-Harmonic Fourier Data Aditya Viswanathan California Institute of Technology SIAM Annual Meeting 2013 July / 19

2 Joint work with Anne Gelb Sidi Kaber Research supported in part by National Science Foundation grants CNS , DMS and DMS (FRG). 0 / 19

3 Motivating Application Magnetic Resonance Imaging Physics of MRI dictates that the MR scanner collect samples of the Fourier transform of the specimen being imaged. 1 / 19

4 Motivating Application Magnetic Resonance Imaging Collecting non-uniform measurements has certain advantages; for example, they are easier and faster to collect, and, aliased images retain diagnostic qualities. Reconstructing images from such measurements accurately and efficiently is, however, challenging. 1 / 19

5 Model Problem Let f be defined in R with support in [ π, π). Given ˆf(ω k ) = f, e iω kx, k = N,, N, ω k not necessarily Z, compute an approximation to the underlying (possibly piecewise-smooth) function f, an approximation to the locations and values of jumps in the underlying function; i.e., [f](x) := f(x + ) f(x ). Issues Sparse sampling of the high frequencies, i.e., ω k k > 1 for k large. The DFT is not defined for ω k k; the FFT is not directly applicable. 2 / 19

6 Outline Introduction Motivating Application Simplified Model Problem Non-Harmonic Fourier Reconstruction Uniform Re-sampling Convolutional Gridding Harmonic and Non-Harmonic Kernels Designing Convolutional Gridding Kernels Edge Detection Concentration Method Design of Non-Harmonic Edge Detection Kernels 2 / 19

7 Uniform Re-sampling (Rosenfeld) We consider direct methods of recovering f and [f] from ˆf(ω k ). Due to our familiarity with harmonic Fourier reconstructions and the applicability of FFTs, we will consider a two step process: 1. Approximate the Fourier coefficients at equispaced modes 2. Compute a standard (filtered) Fourier partial sum Basic Premise f is compactly supported in physical space. Hence, the Shannon sampling theorem is applicable in Fourier space; i.e., ˆf(ω) = k= sinc(ω k) ˆf k, ω R. 3 / 19

8 Uniform Re-sampling Implementation We truncate the problem as follows ˆf(ω k ) sinc(ω k l) ˆf }{{} l, l M A R 2N+1 2M+1 k = N,, N The equispaced coefficients are approximated using f l = A ˆf(ω k ), where A is the Moore-Penrose pseudo-inverse of A. A and its properties characterize the resulting approximation. Regularization may be used (truncated SVD, Tikhonov regularization) in the presence of noise. A is (unfortunately) a dense matrix in general, with the computation of f requiring O ( N 2) operations. 4 / 19

9 Uniform Re-sampling Sampling Patterns Consider the sampling pattern ω k = k ± U[0, µ], k = N,, N where U[a, b] denotes an iid uniform distribution in [a, b] with equiprobable positive/negative jitter. Jitter µ κ(a) / 19

10 Uniform Re-sampling An Example Reconstruction using jittered samples (µ = 0.5). Error Fourier Modes Recontruction 6 / 19

11 Uniform Re-sampling An Example Reconstruction using jittered samples (µ = 0.5). Error Fourier Modes Reconstruction 6 / 19

12 From Uniform Re-sampling to Convolutional Gridding Recall that for uniform re-sampling, we use the relation ˆf(ω) = k sinc(ω k) ˆf k = ( ˆf sinc)(ω) Since the Fourier transform pair of the sinc function is the box/rect function (of width 2π and centered at zero), we have f Π ˆf sinc Now consider replacing the sinc function by a bandlimited function ˆφ such that ˆφ( ω ) = 0 for ω > q (typically a few modes wide). We now have f φ ˆf ˆφ 7 / 19

13 Convolutional Gridding (Jackson/Meyer/Nishimura and others) Gridding is an inexpensive direct approximation to the uniform re-sampling procedure. Given measurements ˆf(ω k ), we compute an approximation to ˆf ˆφ at the equispaced modes using ( ˆf ˆφ)(l) l ω k q α k ˆf(ωk ) ˆφ(l ω k ), l = M,..., M. Now that we are on equispaced modes, use a (F)DFT to reconstruct an approximation to f φ in physical space. Recover f by dividing out φ. α k are density compensation factors (DCFs) and determine the accuracy of the reconstruction. 8 / 19

14 Analysis of the Convolution Gridding Sum The gridding approximation can be written as α k ˆf(ωk ) ˆφ(l ω k ) e ilx f cg (x) = = l M l ω k q φ(x) k l α ( k f(ξ)e iω k ξ dξ ) ( φ(η)e i(l ωk)η dη ) e ilx φ(x) ( ) ( ) f(ξ)φ(η) α k e iω k(η ξ) e il(x η) dξdη k }{{}}{{} A = α N (η ξ) D N (x η) φ(x) (f A α = N )(η) φ(η) D N (x η)dη φ(x) = ([{f Aα N } φ] D N) (x) φ(x) l 9 / 19

15 The Dirichlet Kernel A Review Given ˆf k := f, e ikx, k = N,, N, a periodic repetition of f may be reconstructed using the Fourier partial sum P N f(x) = ˆf k e ikx = (f D N )(x), where k N D N (x) = k N e ikx is the Dirichlet kernel. D N is the bandlimited (2N + 1 mode) approximation of the Dirac delta distribution. 10 / 19

16 The Dirichlet Kernel A Review D N completely characterizes the Fourier approximation P N f. Filtered and jump approximations are similarly characterized by equivalent filtered and (filtered) conjugate Dirichlet kernels. 10 / 19

17 The Non-Harmonic Kernel Consider the non-harmonic kernel A N (x) = k N e iω kx A N is non-periodic. The non-harmonic kernel is a poor approximation to the Dirac delta distribution. Depending on the mode distribution, A N may be non-decaying. Filtering is of no help under these circumstances. 11 / 19

18 The Non-Harmonic Kernel Jittered Modes ω k = k ± U[0, µ], µ = / 19

19 The Non-Harmonic Kernel Log Modes ω k logarithmically spaced 11 / 19

20 Outline Introduction Motivating Application Simplified Model Problem Non-Harmonic Fourier Reconstruction Uniform Re-sampling Convolutional Gridding Harmonic and Non-Harmonic Kernels Designing Convolutional Gridding Kernels Edge Detection Concentration Method Design of Non-Harmonic Edge Detection Kernels 11 / 19

21 Designing Gridding Kernels Recall that the convolutional gridding sum can be written in the form f cg (x) = ([{f Aα N } φ] D N) (x) φ(x) where the (weighted) non-harmonic kernel A α N(x) = k N α k e iω kx. α k are free design parameters which we choose such that A α N is compactly supported and a good reconstruction kernel (such as the Dirichlet kernel) in the interval of interest. 12 / 19

22 Design Problem Formulation Choose α = {α k } N N such that α k e iωkx k N l M e ilx x π 0 else Discretizing on an equispaced grid, we obtain the linear system of equations Dα = b, where D l,j = e iω lx j denotes the (non-harmonic) DFT matrix, and b p = sin(m+1/2)xp sin(x p/2 ) Π are the values of the Dirichlet kernel on the equispaced grid. 13 / 19

23 Design Problem Formulation Choose α = {α k } N N such that α k e iωkx k N l M σ l e ilx x π 0 else Discretizing on an equispaced grid, we obtain the linear system of equations Dα = b, where D l,j = e iω lx j denotes the (non-harmonic) DFT matrix, and b p are the values of the (filtered) Dirichlet kernel on the equispaced grid. 13 / 19

24 Numerical Results Reconstruction ω k logarithmically spaced N = 256 measurements Iterative weights solved using LSQR 14 / 19

25 Numerical Results Reconstruction Error ω k logarithmically spaced N = 256 measurements Iterative weights solved using LSQR 14 / 19

26 Numerical Results DCF weights α ω k logarithmically spaced N = 256 measurements Iterative weights solved using LSQR 14 / 19

27 Outline Introduction Motivating Application Simplified Model Problem Non-Harmonic Fourier Reconstruction Uniform Re-sampling Convolutional Gridding Harmonic and Non-Harmonic Kernels Designing Convolutional Gridding Kernels Edge Detection Concentration Method Design of Non-Harmonic Edge Detection Kernels 14 / 19

28 Concentration Method (Gelb, Tadmor) Approximate the singular support of f using the generalized conjugate partial Fourier sum S σ N[f](x) = i N k= N ˆf(k) sgn(k) σ ( ) k e ikx N σ k,n (η) = σ( k N ) are known as concentration factors which are required to satisfy certain admissibility conditions. Under these conditions, S σ N[f](x) = [f](x) + O(ɛ), ɛ = ɛ(n) > 0 being small i.e., SN σ [f] concentrates at the singular support of f. 15 / 19

29 Concentration Factors Factor Expression π sin(α η) Trigonometric σ T (η) = Si(α) α sin(x) Si(α) = dx 0 x Polynomial σ P (η) = p π η p p is the order of [ the factor ] Exponential σ E (η) = Cη exp 1 α η (η 1) C - normalizing constant α - order π C = 1 1 N 1 N [ exp 1 α τ (τ 1) ] dτ Table: Examples of concentration factors Figure: Envelopes of Factors in k-space 16 / 19

30 Some Examples Jump Approximation Trigonometric Factor Jump Approximation Exponential Factor f S N [f] f(x) 0 f(x) f S N [f] x x (a) Trigonometric Factor (b) Exponential Factor Figure: Jump Function Approximation, N = / 19

31 Designing Non-Harmonic Edge Detection Kernels Choose α = {α k } N N such that i sgn(l)σ( l /N)e ilx x π α k e iωkx l M k N 0 else Discretizing on an equispaced grid, we obtain the linear system of equations Dα = b, where D l,j = e iω lx j denotes the (non-harmonic) DFT matrix, and bp are the values of the generalized conjugate Dirichlet kernel on the equispaced grid. 18 / 19

32 Numerical Results Jump Approximation and Corresponding Weights ω k logarithmically spaced N = 256 measurements Iterative weights solved using LSQR 19 / 19

33 Summary and Future Directions 1. Applications such as MR imaging require reconstruction from non-harmonic Fourier measurements. 2. Assuming the function of interest is compactly supported, the underlying relation between non-harmonic and harmonic Fourier data is the Shannon sampling theorem (sinc interpolation). 3. Convolutional gridding is an efficient approximation to sinc-based resampling. 4. A set of free parameters known as the density compensation factors (DCFs) allow us to design gridding kernels with favorable characteristics. 5. To do compare results with frame theoretic approaches, use banded DCFs to obtain better gridding approximations. 19 / 19

34 Selected References 1. D. Rosenfeld, An Optimal and Efficient New Gridding Algorithm Using Singular Value Decomposition, in Magn. Reson. Med., Vol. 40, No. 1 (1998), pp J. Jackson, C. Meyer and D. Nishimura, Selection of a Convolution Function for Fourier Inversion Using Gridding, in IEEE Trans. Med. Img., Vol. 10, No. 3 (1991), pp J. Fessler and B. Sutton, Nonuniform Fast Fourier Transforms Using Min-Max Interpolation, in IEEE Trans. Sig. Proc., Vol. 51, No. 2 (2003), pp J. Pipe and P. Menon, Sampling Density Compensation in MRI: Rationale and an Iterative Numerical Solution, in Magn. Reson. Med., Vol. 41, No. 1 (1999), pp A. Gelb and E. Tadmor, Detection of Edges in Spectral Data, in Appl. Comp. Harmonic Anal., 7 (1999), pp / 19