Accelerating the Nonuniform Fast Fourier Transform
Summary (1 min read)
1. Introduction.
- The authors describe an extremely simple and efficient implementation of the nonuniform fast Fourier transform .
- Let us begin, however, with a more precise description of the computational task.
- There is some confusion in the literature about the use of NUFFTs in this context (see Remark 1 below).
- Not all schemes for reconstructing Fourier integrals of the type (3) can be represented formally as a quadrature of the type (1) .
3. Fast Gaussian Gridding.
- Following standard practice, the authors will refer to these processes as gridding and the M r -point mesh as the oversampled mesh.
- This cost (in either storage or CPU time or both) becomes a significant burden in two, three, and higher dimensions.
- It is sometimes called the curse of dimensionality; in the absence of a separable coordinate system, interpolation-type processes have costs that grow exponentially with dimension.
- This expression for f τ looks much more expensive than it actually is.
- An elementary calculation shows that EQUATION ) Careful organization of the loop shows that, for each source point, two exponential evaluations are required, followed by two multiplications at each of 2M sp regular mesh points.
Example 2 (Fast Gridding Compared to Gridding).
- Both algorithms use the standard FFT on the oversampled mesh, and the time for this step is indicated in Figure 3 of the math coprocessor, cache size, etc.
- Note that the type-1 and type-2 transforms are very similar in terms of floating point operations; the differences in CPU time are due mainly to memory caching issues.
- In any case, the speed-up of the fast gridding algorithm is significant in two dimensions and would be even more significant in the three-dimensional case.
- The authors have not carried out such fine-tuning.
Example 3 (Comparison with Direct Method).
- The authors compare the computational performance of their fast gridding algorithm with direct summation and the standard FFT.
- The direct code was implemented and compiled with the same options: the gcc-2.95 compiler with -O2 optimization on a 450MHz Sparc Ultra-60.
- Table 2 shows how the computational cost for gridding grows with precision.
- In summary, the NUFFT is about 4 times more expensive in one dimension than a traditional M -point FFT for single precision accuracy.
- The MRI hardware is able to acquire the Fourier transform of a particular tissue property at selected points in the frequency domain.
Did you find this useful? Give us your feedback
Citations
669 citations
545 citations
376 citations
Cites background from "Accelerating the Nonuniform Fast Fo..."
...Two useful properties of the Gaussian window function (C.1) that can be exploited within the presented framework have recently been reviewed in [ Greengard and Lee 2004 ]....
[...]
326 citations
Cites background from "Accelerating the Nonuniform Fast Fo..."
...kernel can be quite efficient; in one dimension, the Gaussian kernel requires two exponential evaluations per data sample point plus two multiplications for each point on the grid to which this data sample is deposited [12]....
[...]
272 citations
References
1,251 citations
"Accelerating the Nonuniform Fast Fo..." refers background or methods in this paper
...While they concentrated on the one-dimensional case, higher dimensional versions have been considered by a variety of authors [3, 6, 14]....
[...]
...For a variety of technical reasons, however, nonuniform data sampling techniques are much better suited for fast data acquisition, motion correction, and functional MRI [4]....
[...]
...In this example, we create simulated MRI data by using a type-2 transformation in two dimensions: F (skx, s k y) = ∑ j1 ∑ j2 f(j1, j2) e−i(j1,j2)·(s k x,s k y) ,(17) followed by a type-1 transformation to reconstruct the image, f̃(j1, j2) = N−1∑ k=0 Fk e i(j1,j2)·(skx,s k y)....
[...]
...We restrict our attention here to one: function (or image) reconstruction from Fourier data as discussed in [6, 8, 11, 14]....
[...]
...One of the important applications of the nonuniform FFT is to magnetic resonance imaging (MRI) [6, 8, 11, 12, 13, 14]....
[...]
1,187 citations
"Accelerating the Nonuniform Fast Fo..." refers background or methods in this paper
...One of the important applications of the nonuniform FFT is to magnetic resonance imaging (MRI) [6, 8 , 11, 12, 13, 14]....
[...]
...We restrict our attention here to one: function (or image) reconstruction from Fourier data as discussed in [6, 8 , 11, 14]....
[...]
...There are a host of applications of such algorithms, and we refer the reader to the references [2, 6, 8 , 11, 13, 14, 17] for examples....
[...]
848 citations
536 citations
"Accelerating the Nonuniform Fast Fo..." refers background or methods in this paper
...For a variety of technical reasons, however, nonuniform data sampling techniques are much better suited for fast data acquisition, motion correction, and functional MRI [4]....
[...]
...In this example, we create simulated MRI data by using a type-2 transformation in two dimensions: F (skx, s k y) = ∑ j1 ∑ j2 f(j1, j2) e−i(j1,j2)·(s k x,s k y) ,(17) followed by a type-1 transformation to reconstruct the image, f̃(j1, j2) = N−1∑ k=0 Fk e i(j1,j2)·(skx,s k y)....
[...]
...We restrict our attention here to one: function (or image) reconstruction from Fourier data as discussed in [6, 8, 11, 14]....
[...]
...One of the important applications of the nonuniform FFT is to magnetic resonance imaging (MRI) [6, 8, 11, 12, 13, 14]....
[...]
...Example 4 (MRI Image Reconstruction)....
[...]
359 citations
"Accelerating the Nonuniform Fast Fo..." refers background in this paper
...While they concentrated on the one-dimensional case, higher dimensional versions have been considered by a variety of authors [3, 6, 14]....
[...]
...Subsequent papers, such as [1, 3, 9, 10], described variants based on alternative interpolation/approximation approaches....
[...]
Related Papers (5)
Frequently Asked Questions (4)
Q2. What is the purpose of the MRI?
In most clinical systems, the device is designed to acquire data on a uniform Cartesian mesh, from which a standard FFT can be used for image reconstruction.
Q3. What is the simplest way to do this?
2. Precompute E3(l) = e−(πl/Mr)2/τ for 0 ≤ l ≤ Msp and E4(k) = E4(M−k) = eτk2 for |k| ≤ M2 .Step C: Convolution for Each Source Point (xj , yj) 1.
Q4. How many factors affect the computation time of the fast gridding algorithm?
The actual computation time depends on a number of factors, including compiler options, type of CPU, performanceof the math coprocessor, cache size, etc.