Partial Fourier Methods 2010 Pfi
Partial Fourier Methods 2010 Pfi
Partiial Fourier reconstruction algorithms exploit the redun- of the partial Fourier reconstructions with the full Fou-
dancy in magnetic resonance data sets so that half of the data rier images is obtained using both artificial and real clin-
is c,alculated during image reconstruction rather than ac- ical data.
quired. The conjugate synthesis, Margosian, homodyne de-
tection, Cuppen and POCS algorithms are evaluated using
spatial frequency domain analysis to show their characteris-
tics and where limitations may occur. The phase correction THEORETICAL ANALYSIS OF EXISTING PARTIAL
used in partial Fourier reconstruction is equivalent to a con- FOURIER ALGORITHMS
volution in the frequency domain and the importance of accu-
rately implementing this convolution is demonstrated. New
Partial Fourier algorithms perform considerable data ma-
reconstruction approaches, based on passing the partial data nipulation in either (or both) the image and spatial fre-
through a phase correcting, finite impulse response (FIR), quency (original data) domains. A number of limitations
digital filter are suggested. These FIR and MoFIR algorithms can be exposed by providing a complete analysis of the
have) a speed near that of the Margosian and homodyne de- algorithms in a single domain. This new information can
tectilon reconstructions, but with a lower error; close to that of be used to suggest new algorithms or at least new direc-
the CuppenlPOCS iterative approaches. Quantitativeanalysis tions to take to overcome deficiencies.
of the partial Fourier algorithms, tested with three phase es- For easy ID and 2D image comparison, all 2D images,
timation techniques, are provided by comparing artificial and p(x, y),will be displayed with the partial data dimension
clinical data reconstructed using full and partial Fourier tech-
placed horizontally. It is therefore convenient to assume
niqules.
that the original data, s(u, v); -N/2 5 v < N/2 is inverse
Key words: partial Fourier; phase correction; FIR filter.
discrete Fourier transformed (DFT) in the v dimension to
give a new data matrix, sy( u ) . Each row of sy(u ) , non-zero
INTRODUCTION for -rn 5 u < Nl2, must be reconstructed to form the
corresponding image row of p(x, y),and can be consid-
Part la1 Fourier reconstruction algorithms exploit the re- ered as an independent data set.
dundancy in the magnetic resonance (MR) data set. Such
algorithms are useful when asymmetric data sets arise in
spin and gradient echo MR imaging. The algorithms dis- Conjugate Synthesis, Margosian, and Homodyne
cussed here are the conjugate synthesis (I),Margosian et Detection Imaging
al. (:?, 3), homodyne detection (4), Cuppen et al. (5) and Conjugate synthesis imaging (1) attempts to reconstruct a
Projection onto Convex Sets (POCS) methods (6, 7). The full data matrix based on the assumption that the final
recent review paper by Liang et al. (8) has provided an image p(x, y) has only real components (i.e., no phase
explanation and qualitative comparison of these algo- terms). The procedure uses the complex conjugate of the
rithms. The alternative spatial frequency domain view- known data [O I u < N/2] to fill in the unknown data. The
point used in this paper indicates both the limitations of image is reconstructed using an inverse DFT on the com-
the existing methods and a faster approach to implement- pleted s y ( u ) data set.
ing POCS. This analysis suggests new reconstruction ap- The Margosian partial Fourier algorithm (2) multiplies
proaches, based on passing the partial data through a the data set, s,(u), by a merging filter, with a frequency
phase correcting, finite impulse response (FIR), digital response H( u ) , to produce a new data set s(,’ u ) with its
filter. All the partial Fourier reconstruction algorithms lower and negative frequencies filtered. This filtering
will be evaluated in conjunction with three phase esti- smooths the transition between calculated and known
mation techniques: Margosian’s low frequency filtered data to reduce Gibbs’ ringing during the DFT operation
estimate (Z), the generalized series (9) and a 2D polyno- which follows. The data are then inverse Fourier trans-
mial model estimate (10, 11).A quantitative comparison formed to the image space and each pixel multiplied by
the phase estimator function, 0,(x) = e-j+y(x), where
$y (x)is the phase estimate. In the frequency domain, this
MRM 3 0 5 1 5 9 (1993)
From the Department of Electrical and Computer Engineering, and the is equivalent to convolving the filtered data with the
Department of Radiology (A.C.), University of Calgary, Calgary, Alberta, Fourier transform (FT) a,( u ) of the phase estimate 0,(x)
Canada. giving
Address correspondence to: M. R. Smith, Ph.D., Department of Electrical
and Computer Engineering, University of Calgary, 2500 University Drive, Si(u) = {s,(u)H(u)} 8 Oy(u)where O,(u) = FT(B,(x)).
N.W., Calgary, Alberta, Canada T2N 1N4.
Received May 4, 1992; revised October 27, 1992; accepted January 26,
111
1993. The reconstructed image is taken to be the twice the real
0740-3194/93 $3.00
Copyright 0 1993 by Williams & Wilkins part of the image. This is equivalent to adding the fre-
All rights of reproduction in any form reserved. quency domain data to the reflected conjugate of itself.
51
52 McGibney et al.
The frequency domain analysis of the Margosian method The frequency domain analysis of this Cuppen imple-
is shown in Fig. 1. mentation is shown in Fig. 2. Conjugating the data in the
'The Margosian image is normally displayed using a image domain is equivalent to conjugating and reflecting
magnitude display. One of the problems with the magni- the spatial frequency domain data around the zero fre-
tude display of images is the noise bias which leads to a quency point. This also reverses the sign of the phase
decreased detectability for low intensity objects. This can error in the image domain, which must be corrected by
be avoided by displaying a phase corrected real image multiplying with the function O;(x) = e'2'r(x). The image
(12-14) provided that the phase estimator is not corre- phase correction operation corresponds to a convolution
lated to the noise (10).Since the images obtained from a in the frequency domain with the Fourier transform
partial Fourier reconstruction are commonly displayed of O'Jx). Finally the new data set is placed in the original
using a magnitude display, it is not often realized that data set locations where no data had been collected, and
these images are phase corrected if properly recon- the process iterated.
structed. If these images were displayed as a real data set, POCS (6, 7) is based on the principle that the correct
then the improved low intensity detectability suggested image is the intersection of all images whose Fourier
by Berstein et al. (13) would become available. This is the transform agrees with the measured partial data and all
essence of the homodyne detection approach (4) which is images whose phase is the same as the phase estimate. To
equivalent to a Margosian reconstruction with a real find the intersection of these two images sets, the data set
rather than a magnitude display of the reconstructed is first inverse Fourier transformed into the image do-
image. main, where the complex values of the image are pro-
jected onto a line that is at an angle equal to the phase
Cuppen and POCS Methods estimate,
-
These iterative approaches attempt to overcome the
shortcomings of the Margosian method. The Lindskog The new image is then Fourier transformed back to the
Cuppen implementation (15) inverse Fourier transforms spatial frequency domain where it is used to replace the
the original data to the image domain where it is conju- unknown data. The process is repeated and on the last
gated and multiplied by eJZm'(x).This new image is then iteration, instead of substituting the new data into the
Fourier transformed back into the spatial frequency do- data set from the previous iteration, a merging filter is
main to be used to fill in the missing data points. The used.
process is then iterated until a suitable convergence is The relationship between the POCS and Cuppen's al-
found. gorithms is more obvious if the projection operation is
ORIGINAL
DATA SET
MULTIPLY
PHASECORRECI'ION
(FOOURIUI'IRANFOFMED)
11 DATA SET
"("'REFLECTED
CONVOLVE
O W
DATA SET
DATA SET
FIG. 1. Frequency domain analysis of the Margosian/homodyne FIG. 2. Frequency domain analysis of the Cuppen's partial Fou-
detection partial Fourier reconstruction method. Removal of the rier reconstruction algorithm. The POCS algorithm is equivalent to
phase correction stage provides the analysis for the conjugate averaging the Cuppen's corrected data with the original data until
synthesis approach. the final iteration, when a merging filter is applied.
Parti8dFourier Reconstruction Algorithms 53
statements to the contrary by Nolls et al. (4). The differ- in Fig. 5 which represents some simplistic data that is
ence between the two methods becomes greater as the shifted in the frequency domain by the effect of a simple
phase shifts become larger. linear phase term. The phase correction operation is rep-
The merging filter using in the Margosian and homo- resented by O ( u ) = 6 ( u - n). The merging filter is illus-
dyne techniques was designed to smooth the transition trated using a ramp filter although a shifted Hanning
between known and zero values to avoid Gibbs' ringing filter would actually be used. The shaded areas in Fig. 5
during reconstruction. The merging filter in the FIR al- schematically show the edge distortions that would be
gorithm also smooths the transition, but is designed more present if the phase correction operator had a bandwidth
to remove the data distorted by the phase correction edge of 2 p + 1. It can clearly be seen that the amount of data
effects. The differences are schematically demonstrated affected by the convolution edge effects is considerably
lower in the FIR approach. The Cuppen/POCS algorithms
achieve a similar reduction in edge effects by virtue of
their iteration. The reason that the homodyne detection
and Margosian reconstructions are limited to images
with slowly varying phase (small p ) can be seen by the
+ CONVOLVE
PHASE CORRECTED
DATA SET
large shaded area for these algorithms. The use of a merg-
ing filter prior to the phase correction means that the
errors are introduced directly into the large amplitude
terms around 1u1 = 0 , maximizing the distortion.
As will be seen in the reconstructed images, the sim-
plistic approach of the FIR reconstruction is degraded by
rmm I uL
the windowing operation on the phase estimates which
introduces truncation artifacts if the true bandwidth is
LT--l
MULTIPLY high. This is a problem for gradient echo images where a
really narrow bandwidth phase estimator has not yet
been found (10).These artifacts are reduced in the mod-
PHASE CORRECIED
DATA SET
ified FIR (MoFIR) approach which uses all the phase
correction function, e(u ) , without truncating it. This
however reintroduces some of the convolution edge ef-
fects discussed earlier. However if the phase estimator is
properly chosen, its values outside the range +- p are still
small so that the smearing of the noncollected data will
not extend too far into the valid data and will be effec-
tively removed by the use of the narrow transition merg-
ing filter. When successful at reconstruction, this algo-
rithm phase corrects the data so that the image can be
FIG. 4. Frequency domain analysis of the N R and MoNR partial
displayed with a real display to obtain increased low
Fourier reconstruction algorithms.
k 1x_
FIG. 5. Schematic diagram showing
the distortions introduced into the
Margosian/homodyne partial Fourier MARGOSIAN RECONSTRUCTION
MERGING FXTER PHASE CORRECTION R E N C l l O N & ADDITION
reconstructed data set as the merging 4
filter is applied prior to the phase cor-
rection. The FIRalgorithm removes this
distortion and accounts more accu-
t sfuJHfu) /sfuJHfuJJ'o(uJ
intensity detectability mentioned for the homodyne Analysis on artificially generated data can be rightly
technique. criticized as being unrealistic and not representing the
The MoFIR phase correction convolution can be imple- true clinical picture. However, when artificially gener-
men ted directly in the frequency domain. However, the ated data is used in conjunction with clinical data, it can
long length of the phase estimator means that it is more give an insight on the reasons why certain distortions
computationally efficient (Order(Nlog,N) compared to appear on the clinical data. The final test was on a high
Order(2pN) complex operations) to inverse Fourier signal-to-noise clinical image: an axial view of the head
transform the partial data into the image domain, multi- acquired with a spin echo series and a data set size of 256
ply by the phase correction, Fourier transform back to the X 256 (Fig. 6B). The phase of the image was slowly vary-
frequency domain to apply the merging filter before re- ing with sharp local phase changes associated with blood
flecting the data and reconstructing. Thus the MoFIR re- flow near the bottom of the image and in several locations
construction is “2 FFTs” slower than the Margosian ap- around the circumference of the brain.
proach, but computationally more efficient than the Partial data sets were generated by setting the values
Cuppens and POCS approaches which are more than “21 for u < -m equal to 0 in the full data sets. Each partial
FFTj” slower, where I is the number of iterations. Fourier algorithm was tested on the clinical images with
the filtered (low frequency), the generalized series (GS)
and the 2D polynomial estimate. The low frequency
TEST METHODS AND RESULTS phase estimate ( 2 ) uses as a phase estimate the image
Two test images were evaluated. The first was an artificial generated by zero padding the center portion of the data
image, a simple box shape (see Fig. 6A) generated in the -m 2 u < m to -Nl2 2 u < Nl2 and reconstructing using
spatial frequency domain from a truncated sinc function. the standard Fourier technique. By contrast, the general-
This, artificial image was modified with a quadratic glo- ized series (9) attempts to model all the known frequency
bal phase error and two rapid localized phase changes information using a separable function where one com-
( 0 . 1 and
~ 0 . 5 high,
~ 5 pixels wide, respectively). These ponent contains only image amplitude terms (based on
phase errors represent static field and instrumentation an edge enhanced magnitude image) and the other only
effects and a small and large flow induced phase changes. phase information. Further details may be found in the
Both exact and inaccurate (quadratic only) phase esti- review by Liang et al. (8).
mates were supplied to the partial Fourier algorithms. MacFall et al. (16) have attempted to generate a phase
This was intended to correspond to practical situations estimate using a 1D polynomial. This work was extended
where the exact phase is not determined because of the first by Berstein and Perman (12) to low order polynomi-
effect of image noise or the nature of the phase estimation als in both data directions and then by McGibney et al.
algorithm. (10, 11) to a true 2D polynomial fit. A major advantage of
synthesis approach for only a small increase in compu- higher bandwidth filtered and generalized series phase
tation time. estimates. This allows its use with both spin and gradient
Given the exact phase terms, both the Cuppens and echo images, unlike the direct FIR approach. Errors from
POCS iterative approaches quickly move to a line image an inaccurate phase estimate are again localized, avoid-
that is equivalent to that of the full Fourier reconstruc- ing the systematic distortions found in the Margosianl
tion. In practice, the phase estimate will not be exact and homodyne detection reconstructions.
the procedures have a tendency to become unstable in An estimate of the calculation time required for the
this situation if iterated too often. Given an inaccurate algorithms was obtained by implementing the algorithm
phase estimate (quadratic terms only), the Cuppen algo- on a TAAC application's accelerator connected to a SUN
rithm had large amplitude distortions in the regions of 31160 workstation and does not include system overhead
local rapid phase change which spread as low frequency such as image movement. The results are given in Table
oscillations into areas where the phase estimate had been 1. The times for phase estimation plus partial Fourier
corrwt. The rippling occurs because of the mismatch of reconstruction range over several orders of magnitude.
the positive and negative data sets where they meet in the The times for POCS reconstruction is for the faster algo-
frequency domain; producing Gibbs' ringing artifacts in rithm suggested by McGibney (10). The actual times for
the image domain. This ringing is particularly evident in the Cuppen and POCS algorithms depends on the num-
the full image when the phase estimate is generated from ber of iterations required, which is image-dependent.
the generalized series algorithm as this approach is itself Despite the fact that the FIR reconstruction used a
susceptible to ringing artifacts (10).The POCS conver- small bandwidth filter during the direct convolution, it
gence is slower than for the Cuppen's algorithm because
of the averaging action. However, the local phase error
effects do not spread as far through the image because of Table 1
the s,moothing action of the averaging and the use of a Comparison of the Times of the Partial Fourier Reconstruction
merging filter. Algorithms Both Individually and in Conjunction with the Phase
Because the high frequency bandwidth of the exact Estimation Algorithms
phase estimate invalidates its basic assumptions, the FIR
Partial Partial
method performs poorly on the artificial box data. With Partial Fourier
the lower bandwidth of the smoothed phase estimate, it Partial Fourier Fourier
provides a reconstruction on both the line and clinical Fourier PIUS genera,ized plus
data similar to the Cuppen or POCS algorithms, with only filtered polynomial
(s) estimate estimate
artifacts near the rapid phase changes. Unlike Cuppen
,and IPOCS reconstructions, these distortions are strictly (s) (S)
(S)
localized to the area where the phase has been incorrectly Conjugate-symmetry 2.4 7.1 464 7.1
determined (sharp changes). This localizing effect is also Margosian-hornodyne 2.9 7.6 464 7.6
seen when the FIR is used on the clinical image with Cuppen (4 iterations) 12.5 17.2 474 17.2
filterled and generalized series phase estimates, despite POCS (4 iterations) 14.6 19.3 476 19.3
their larger bandwidth. FIR (direct) 9.6 14.3 471 14.3
FIR (circular) 5.3 10.0 467 10.0
The MoFIR reconstruction approach was designed to MoFIR (circular) 5.3 10.0 467 10.0
work with both the low bandwidth 2D polynomial and
58 McGibney et al.
was still considerably slower than the MoFIR reconstruc- values of the partial data width m are available by using
tion which required a number of additional transforms a low bandwidth phase estimator and a narrow merging
between the frequency and time domains. This is a con- filter. In particular, the merging filter must be applied
sequence of the 2pN and Nlog,N time relationship be- after the phase correction so that the merging filter char-
tween the direct and circular convolution methods. In acteristics are not distorted, and to ensure that the phase
retrospect, the FIR reconstruction, despite its shorter fil- correction edge effects are removed. When only high
ter length, could also have been determined this way. The bandwidth phase estimators are available, as in a gradi-
current implementation of the Cuppen and POCS algo- ent echo image, a larger value of m is required for optimal
rithms via the computationally efficient circular convo- reconstruction. After phase correction, a sharp transition
lution must be balanced against the fact that the circular merging filter is used in the MoFIR algorithm to remove
convolution mixes the high positive and negative data (reduce) the data made invalid by the phase correction
frequencies. Direct convolution can allow better control convolution operation, and then the data is reflected and
of the edge effects. By contrast, there is no difference in reconstructed.
the FIR and MoFIR algorithms implemented by either However full removal of the edge effects can only be
approach. This is because there are not yet any negative achieved by reconstructing using the undistorted data
frequency data components when the phase correction between [-m + p I u < Nl2 - 1 - Zp]. This implies some
convolution is applied. technique, such as modeling (8, 17), to implicitly or ex-
plicitly extrapolate the data beyond u = N12 - 1 - 2p to
avoid the convolution distortion associated with smear-
CONCLUSIONS ing the high positive and negative frequencies of the par-
tial data set. In addition, since the phase estimator band-
Five existing methods of partial Fourier imaging (conju-
width may be greater than the partial data width, it will
gate symmetry, Margosian, homodyne detection, Cup-
also be necessary to generate the data for u 5 -m + 2p.
pen, and POCS) were analyzed using equivalent fre-
After these extrapolations, the partial Fourier reconstruc-
quency domain operations to determine how they
tion can be completed. A number of papers (10, 18, 19)
introduced the missing data and what their weaknesses
have reported early results from the combination of mod-
were. The analysis was used to suggest two new partial
eling and partial Fourier reconstruction, although those
reconstruction techniques (FIR and MoFIR) in an attempt
algorithms were not implemented with the intention of
to obtain the best features of the other algorithms but in
removing the phase correction distortions. The quantita-
a faster implementation.
tive evaluation of the joint technique is to be the subject
Global and local quantitative measures of the algo-
of a future paper. The evaluation is difficult as it is nec-
rithms was made using an artificial I D image and a 2D
essary to distinguish between the effects of modeling on
clinical images. Similar distortion effects and systematic
both the full Fourier and partial Fourier reconstructions
errors were seen for the algorithms whether reconstruct-
and the correction of the phase estimation edge effects.
ing artificial or clinical data with the conjugate symme-
tr,y, Margosian and homodyne approaches giving the
worst results. ACKNOWLEDGMENT
The Cuppen and POCS algorithms converged to the
exact image (given enough iterations) if a true phase es- The authors wish to thank the University of Calgary and
timate was provided. However, if the phase estimate was the Natural Sciences and Engineering Research Council
not exact, both algorithms produced major non-localized (NSERC) of Canada for scholarship and operating funds.
artifacts. However, their global performance was better The clinical image was provided by Dr. Fernando Boada
than the other algorithms. (Case Western Reserve University, Cleveland, Ohio).
The FIR and MoFIR methods are fast non-iterative al-
gorithms implemented using an approach that allowed REFERENCES
for the correction of edge effects associated with the
1. D. A. Feinberg, J. D. Hale, J. C. Watts, L. Kaufman, A. Mark,
phase correction operation. When the phase estimate was
Radiology 161, 527 (1986).
not exact, artifacts were produced as with the other re- 2. P. Margosian, in “Proceedings, 4th SMRM Conference,”
construction approaches. Unlike the other reconstruc- 1024 (1985).
tions, these inaccuracies were confined to the region of 3. P. Margosian, F. Schmitt, D. E. Purdy, Health Care Instrum.
the error and did not produce significant ringing in the 1, 195 (1986).
correctly phase estimated areas. The MoFIR algorithm 4. D. C. Nolls, D. G . Nishimura, A. Macovski, IEEE Truns. Med.
was the more stable in the presence of high frequency h a g . MI-10(2), 154 (1991).
phase components, making it applicable for both spin 5. J. Cuppen, A. van Est, Magn. Reson. Imaging.5, 526 (1987).
arid gradient echo reconstructions. 6. E. D. Lindskog, E. M. Haacke, J. D. Mitchel, Z. P. Liang, in
All the algorithms performed the best with a low band- “Proceedings, 8th SMRM Conference, 1989,” p. 363.
7. E. M. Haacke, E. D. Lindskog, W. Lin, J.Magn. Reson. 92,126
width phase estimator. Through frequency domain anal-
(1991).
ysis of the partial Fourier reconstruction algorithms we 8. Z . P. Liang, F. E. Boada, R. T. Constable, E. M. Haacke, P. C.
have shown the importance of properly accounting for Lauterbur, M. R. Smith, Rev. Magn. Reson. Med. 4(2), 67
the edge effects associated with the phase correction (1992).
techniques used in existing partial Fourier algorithms. 9. Z. P. Liang, in “Proceedings, 9th SMRM Conference, 1990;’’
These edge effects can be minimized when only small p. 552.
Partial Fourier Reconstruction Algorithms 59
10. G. McGibney, “Phase Sensitive Reconstruction of MR Imag- 15. E. D. Lindskog, “Partial Fourier Imaging,” M.Sc. thesis, De-
es,” M. Sc. thesis. Electrical and Computer Engineering, partment of Electrical Engineering and Applied Physics,
University of Calgary, Canada (1991). Case Western Reserve University, (19891.
11. G. McGibney, M. R. Smith, S. T. Nichols. in “Proceedings, 16. J. R. MacFall, N.J. Pelc, R. M. Vavrek, Magn. Res. Imclging 6,
9th SMRM Conference, 1990,” p. 563. 143 (1988).
17. M. Smith, S. T. Nichols, R. Constable, R. M. Henkelman,
12. M. A. Berstein, W. H. Perman, in “Proceedings, 6th SMRM
Magn. Reson. Med. 19, l ( 1 9 9 1 ) .
Conference, 1987,” p. 801. 18. M. R. Smith, W. Saar, S. T. Nichols, in “Proceedings, 5th
13. M. A. Bernstein, D. M. Thomasson, W. H. Perman, Med. SMRM Conference, 1986,” p. 81.
Phys. 16(5), 813 (1989). 19. J. M. Goldfarb, E. M. Haacke, F. E. Boada, W. Lin, in “Pro-
14. J. Hua, G. C. Hurst, J. Magn. Res. Imaging 2, 347 (1992). ceedings, loth SMRM Conference, 1991,” p. 1 2 2 6 .