A Discrete Tomography Review

Fast Mojette Transform for Discrete Tomography

Abstract

A new algorithm for reconstructing a two dimensional object from a set of one dimensional projected views is presented that is both computationally exact and experimentally practical. The algorithm has a computational complexity of with for an image, is robust in the presence of noise and produces no artefacts in the reconstruction process, as is the case with conventional tomographic methods. The reconstruction process is approximation free because the object is assumed to be discrete and utilises fully discrete Radon transforms. Noise in the projection data can be suppressed further by introducing redundancy in the reconstruction. The number of projections required for exact reconstruction and the response to noise can be controlled without comprising the digital nature of the algorithm. The digital projections are those of the Mojette Transform, a form of discrete linogram. A simple analytical mapping is developed that compacts these projections exactly into symmetric periodic slices within the Discrete Fourier Transform. A new digital angle set is constructed that allows the periodic slices to completely tile all of the objects Discrete Fourier space. Techniques are proposed to acquire these digital projections experimentally to enable fast and robust two dimensional reconstructions.

keywords:
Discrete Radon Transform, Mojette Transform, Discrete Tomography, Image Reconstruction, Projection Mapping, Discrete Fourier Slice Theorem
1

1 Introduction

A common problem in science is to determine the internal structure of an object from indirect measurements. Tomography is concerned with the recovery or reconstruction of the internal structure of an object from its projected “views” or projections. In practical terms, projections provide density profiles of an object at different angles , either from absorption or transmission of, for example, x-rays. The profiles are directly proportional to the attenuation of the x-ray intensity within the object. This attenuation is proportional to the magnitude of the line integral

(1)

at each translate and where are two dimensional (2D) spatial coordinates. The curves (usually) form a set of parallel lines (with compact support ) within the projection having the measure on  ((1)).

An exact reconstruction of the object can then be made when the set of projections is given as

(2)

where the functional notation for has been dropped for convenience and it is assumed that for simplicity2. Equations (1) and (2) define the 2D Radon Transform (RT) of the object first developed for ray lines by Radon ((2)) and Funk ((3)) when using “great circles”. Note that the set (2) is uncountably infinite, therefore requiring an infinite number of projections for an exact or unambiguous reconstruction of the continuous function  ((4)). Thus, the reconstruction is always ill-posed when using the RT, since only a finite number of measurements can be made practically ((5)).

The conventional solution is to acquire a large number of projections (usually) at evenly spaced angles. This has practical implications on the quality of the reconstructed image ((6)). The uniqueness of the reconstruction directly relates to the artefacts known as Ghosts or “invisible distributions” ((7)). Ghosts are always present in reconstructions when using the methods based on this classical RT ((8)).

Discrete RTs attempt to resolve this issue by assuming that the object is partitioned into discrete elements, so that an exact reconstruction may be computed in the discrete domain with a finite number of projections. The inverse problem of RT is then no longer ill-posed and therefore free from reconstruction artefacts. Notable discrete RTs include the arbitrary curve block circulant discrete RT ((9); (10)), the -lines discrete RT ((11); (12)) and the Fast Slant Stack ((13)). A short review of these methods, which are not required for this work, is provided in A. However, most discrete methods tend to suffer from either experimental limitations or high computational complexity. This paper presents a discrete approach that is both experimentally and computationally viable that utilises the Mojette Transform (MT), a discrete RT first constructed by Guédon et al. ((14)).

1.1 Mojette Transform

Let the continuous object be approximated digitally within a point-lattice or digital array (see Fig. 1).

Figure 1: An example of a continuous object (I) and its digital representation (II). The digital version is deliberately represented in low resolution for illustrative purposes. Note that the object and its underlying array need not be rectangular.

This is a reasonable approximation since the reconstruction will eventually be represented and processed digitally. Also, let the rays form part of a parallel beam, in order to simplify what follows. Then, analogous to Eqs (1) and (2), the projections within the MT are defined as

(3)

where (or the vectors ), with respect to the array so that and the (i.e. and are coprime). The fractions are known as Farey fractions and are also related to the distribution of prime numbers ((15); (16)). These rational angle projections have a form equivalent to that of the discrete linogram, i.e. the number of bins is dependent on the angle as

(4)

for each projection of a rectangular image. This has implications on the choice of detector resolution and geometry, which are also discussed later in this work.

The definition of the lines fall into one of two classes, depending on the desired sampling being fully discrete or as intensities distributed over the area between lattice points. In the Dirac Pixel Model (DPM), the pixel is summed to its corresponding bin if and only if the line passes through the centre of the pixel. The lines then form a set of non-periodic and parallel discrete lines

(5)

with of an object having convex support. Convex support simply means that space does not have any local singularities. In interpolated or fractional MTs, the pixels are summed according to the length of the pixel intersected by the lines and can be used to approximate continuous objects ((17)). Arbitrarily complicated pixel models are also possible using splines of high order. The DPM is assumed for the duration of this paper. Fig. 2 shows a simple example of a MT for a image using three projections.

Figure 2: An example of a Mojette Transform for a discrete image of size using the three projections , and . The bold lines within the right-hand grid shows a possible reconstruction path using a corner-based inversion method ((18)).

The MT may be inverted or reconstructed exactly according to one of two, albeit practically unsatisfactory, definitions of the Mojette projection set. Firstly, one acquires the projections at all possible angles for a given image size, i.e. is the set of all rational vectors of order in all octants of the half-plane for an image. This definition of the MT can be inverted exactly by smearing (or back-projecting) all the projections within this Mojette set ((19)). Although the number of projections are finite in number, in contrast to the RT, it still utilises a large number projections since is large for modest . For example, the inversion of a image requires a total of 20088 projections and is not computationally viable.

An alternative and more flexible definition of the Mojette projection set was proposed by Katz ((6)). Since one needs only as many equations (or translates in this case) as pixels in order to reconstruct exactly, Katz pointed out that a set satisfying

(6)

where is the number of projections of a image, is theoretically sufficient to obtain an exact reconstruction. Normand et al. ((20)) extended the concept to discrete objects within arbitrary convex regions. Although this new definition substantially reduces the number of projections required for an exact reconstruction, it requires new and more sophisticated inversion methods, as the original back-projection approach is no longer applicable. A number of schemes have been proposed, including a Conjugate Gradient method ((21)) and a Geometric Graph approach ((18)). The former is robust in the presence of noise but is not suitably convergent, i.e. an appropriate pre-conditioner is yet to be found, and the latter is very sensitive to noise.

The MT was initially constructed to aid work that modelled the human vision system ((14)). Since then, applications of the MT include image encoding and network transmission ((20); (17)), as well as work on Asynchronous Transfer Mode (ATM), data integrity ((22); (23)), packet networks via an -dimensional MT ((24)), lossless networking ((25)) and in scalable multimedia distribution ((26)). Image watermarking using a Fourier-based method and the MT has also been developed ((27); (28)) as well as another scheme for crypto-marking and transmission over the Internet of medical images ((29); (30)). Finally, the MT has been applied to Computerised Tomography (CT((31); (19); (32); (33)). See Guédon et al. ((34)) for a thorough review of the MT.

This paper presents a new fast digital algorithm for the MT that is both of low computational complexity and robust to noise present in projection data. The algorithm, which is denoted as the Fast Mojette Transform (FMT) for convenience, symmetrises the MT and its projection set to simplify its inversion. A novel analytical and exact mapping of Mojette projections to the Discrete Fourier Transform (DFT) is also constructed. This mapping, together with a new Mojette set definition, allows the direct and exact “reformatting” of Mojette projections so as to completely tile all of its corresponding Discrete Fourier space. The set, which ranges from to number of projections for an image, completely fills DFT space without interpolation and filtering, thus allowing an exact reconstruction. If any filtering is required, it is exact and does not introduce any artefacts or approximations in the reconstruction. The issues associated with detector geometry and resolution of the Mojette projections (as given by Eq. (4)) are also addressed. The result is that the projection angles are not evenly spaced, the acquisition of projections takes advantage of the digital nature of the reconstruction, the inversion is fast and there are no reconstruction artefacts. Section 2.4 proposes an experimental set-up for utilising this FMT, while preserving its simple inversion process. The FMT is made possible by utilising another discrete RT known as the Discrete Radon Transform (DRT).

1.2 Discrete Radon Transform

The Discrete Radon Transform (DRT), which was the first discrete RT constructed ((35); (36); (37); (38); (39); (40); (41); (42); (43)), is the RT constructed within the same finite geometry as the DFT, i.e. the image is assumed to possess periodic boundaries. The DRT provides an exact partition of DFT space in the form of periodic slices, where a periodic slice is the one dimensional (1D) DFT of a DRT projection. The partitioning is known as the Discrete Fourier Slice Theorem (FST((39); (43)). The fast version of the DRT, that utilises the Fast Fourier Transform (FFT((44); (45)), is denoted as the Fast Radon Transform (FRT) in what follows.

The 2D Cartesian DFT is tiled using 1D slices (see Fig. 4) along the congruences (periodic lines)

(7)
(8)

for an FFT space with and p is a prime number (hence, including powers of two) ((46)). The lines (7) and (8) are utilised for the slopes

(9)
(10)

and the set of translates as

(11)

so that the projections are effectively acquired along the vectors and modulo , while the slices are placed along the vectors and in FFT space (see Fig. 4). Note that the FRT projections/slices are all of the same size or length, unlike the MT, which makes the FRT symmetric in comparison and easy to invert (see Fig. 3(a)).

(a)
(b)
Figure 3: The Discrete Radon Transform (DRT). (a) shows the exact or information preserving nature of the DRT. (b) shows the projections within the DRT (dotted lines) and how they are numerically mapped to linogram (Mojette Transform) projections (solid lines) using the nearest neighbour distance between pixels sampled.

To completely tile all elements of the entire space at least once, i.e. cover all possible coefficients in the FFT space, a total of slices (and hence projections) are required ((46)). For the simplest prime case so that projections are required, the always, so the slices of the same translate may only intersect once at the point . The space is then tiled exactly and exact inversion is possible ((43)). For the case so that projections are required, those lines having slopes coprime to will intersect only once at the DC coefficient for a fixed . However, the slopes not coprime to will intersect p times, leading to under-sampling. The extra projections perpendicular to , i.e. at slopes , are needed in order to sample all of the image. This results in a certain amount of oversampling, which is easily and exactly corrected by dividing each coefficient for all slices by  ((46)).

Thus, for a object, one needs a total of slices for an exact reconstruction via the power of two FRT. The same image would require slices when padded to the nearest prime number using a fast prime-sized FFT via Rader’s algorithm ((45)), which is approximately two to three times slower than the power of two FFT (see Fig. 4 for an example of a prime-sized tiling). The reconstruction is then simply obtained exactly by using the inverse FFT. Reconstruction can be started as soon as slices become available, similar to conventional tomographic methods.

(a)
(b)
(c)
(d)
(e)
(f)
Figure 4: The exact but periodic slices within the Fast Fourier Transform (FFT) (or equivalently the Fast Radon Transform) for the prime case . Each colour represents a slice of a different slope with the DC coefficient/origin centred (black). Note that each vector shown is computed modulo p. (a)-(e) shows the slices with slopes in FFT space. (f) shows the row sum (perpendicular) slice in FFT space. The set of slices tile all of the space exactly once.

Chandra ((47)) also showed that the projections can be mapped to the Number Theoretic Transform (NTT), an integer-only transform analogous to the FFT but with better performance for digital projections. The main issue with the FRT is that the projections are taken along the periodic lines (7) and (8), such as the slices shown in Fig. 4. This makes the experimental geometry not physically meaningful, limiting its practical application to tomography. Despite this, the FRT has been successfully applied to tomography, but no algorithm has resulted that is suitably experimentally and computationally practical ((48); (49); (32); (33)). The next section describes the first major result of the paper, the fusion of the MT and the FRT to produce a discrete RT that is both experimentally and computationally viable for tomography.

2 Fast Mojette Transform

In this section, the algorithm for fast and practical digital tomography is described. Given an image, the objective is to exactly tile an FFT space, where and , so as to efficiently reconstruct the image with the inverse FFT (iFFT). For the FFT space where , redundancy provides the mechanism for suppressing the effects of noise within the projections (see Fig. 5).

Figure 5: The redundancy within the Fast Mojette Transform as an image. The redundancy parameter is chosen so that is a power of two for a image. The original image is assumed to be positioned at the top left for convenience.

Assuming that the size of the image is , in order to utilise the Cooley-Tukey ((44)) algorithm for the FFT, let the total number of projections . If each Mojette projection maps to a unique FRT projection or slice, then they will tile all of FFT space. Once the mapping is determined to be one-to-one, the Mojette projections must be converted to FRT projections. A schematic of these processes are given in Fig. 6.

Figure 6: A schematic of the main processes within the Fast Mojette Transform for . The Mojette projections (I), acquired using a new Mojette set definition, are converted to Fast Radon Transform projections (II) via a new discrete projection mapping, and placed into Fast Fourier Transform space (III) ready for inversion. The structure within the dashed box (II) is a logical step, i.e. the projections are converted in-place.

It will be shown that both these key aspects are possible analytically and are exact in nature. A method is also given to generate optimal projection angle sets. It is then shown how one could experimentally acquire these projections and apply methods to control the noise response within the reconstructions.

2.1 Finite Projection Mapping

The fundamental operation for taking Mojette projections is the same as the FRT because whole pixel intensities are summed along digital rays and placed into discrete bins, i.e. the DPM. The main difference is periodicity, i.e. the MT consist of line segments forming parts of the lines formed from the congruences (7) and (8) of the FRT. The segments can be linked using the shortest distance vectors between pixels sampled by these segments (see Fig. 3(b)). However, these vectors need to be computed using numerical methods which can be computationally inefficient ((50)).

The analytical mapping is as follows. For the case in , the projection of the MT maps to the projection of the FRT as

(12)

where is the multiplicative inverse of that can be computed easily via the Extended Euclidean algorithm. For the case , the projection maps to the and projections as Eq. (12) when the , and

(13)

when the , respectively. These mappings cover all possible vectors/angles because the by definition of the angles. The mapping results from solving and respectively. Examples of how each of these congruences are determined is shown in Fig. 7.

(a)
(b)
Figure 7: A visual presentation of the mapping of Mojette Transform projections to the Fast Radon Transform. Solving for in (a) and in (b) results in the mappings (12) and (13).

It is crucial to ensure that the Mojette set has a one-to-one correspondence to the FRT set, so that the FRT (and equivalently FFT) space is filled completely. This new Mojette set definition, which is finite in number and spans the all octants in the half plane, while providing a discrete coverage the interval , is constructed in the next section.

2.2 Angle Set

In this section, an FMT set is constructed that optimally fills FRT space completely and avoids Mojette projections that are “long”, i.e. projections with a large number of translates , by minimising Eq. (4). This is equivalent to minimising the -norm of the Farey fractions/vectors as

(14)

for each projection in the Mojette set. The simplest way to achieve this is to generate angles through the vectors with in all octants of the half plane. The or value is then checked with the FRT set in question for each vector using the mappings (12) and (13). An example of this set for an image is given Fig. 8(a). Note that this set is not the same as the FRT vectors and , but shares their simplicity. This angle set is a subset of the possible minimal vectors and is computationally inexpensive to generate. Hence, the set is well suited when adapting the detector or image geometry.

(a)
(b)
Figure 8: (a) shows an example of a simple angle set, formed from the vectors with in all octants in the half plane, used within the Fast Mojette Transform for array. Note that this set is not necessarily symmetric in all octants. (b) shows an example of a true minimised angle set for the same array. Both mappings tile or cover FRT (and hence FFT) space for the image exactly. The origin is marked as a red point.

To generate the minimal angle set, one computes

(15)

where is the floor (round-down) operators, beginning the computation with and until . Eq. (15) is a modified form of the standard way to generate Farey fractions. An example of this set for an image is given in Fig. 8(b). Once generated, the set must be sorted by ascending and the first mappings to the FRT angle set chosen. However, the set requires the generation of a large number of fractions and sorting of these fractions, making the computation expensive. Thus, this set is optimal for an unchanging detector and image geometry, since it need only be computed once.

2.3 Finite Conversion

Once the projection set is known for a given image and FFT space, the Mojette projections need to be converted to the finite FRT projections. This can be done by equating Eqs in (5), as well as Eqs (7) and (8), with the mappings (12) and (13). For a Mojette translate and an FRT translate , Eq. (5) becomes

when multiplying by . This expression is equivalent to Eq. (7) via Eq. (12) when taken modulo . Likewise, for the finite projections

which is equivalent to Eq. (8) via Eq. (13) when taken modulo . Similar equations result for the angles in the other octants. Hence,

(16)

The conversion can be done prior to the inversion of the Mojette projections, after having been acquired and stored, or as the projections are being acquired. The latter is more desirable as the Mojette projections are effectively “compacted”, since the FRT projections are generally shorter than the Mojette projections and all have the same size. The next section discusses experimental methods to acquire Mojette projections.

2.4 Mojette Acquisition

Computationally, the Mojette projections can be acquired trivially using the Eqs in (5), since each pixel can be summed to its respective translate  ((14)). The major result of this paper is the experimental data acquisition geometry, which is the topic of the remainder of this section.

The MT can be acquired experimentally in one of three ways. The simplest method is to use a 1D detector array of size , which is typically at most in size for rectangular array when using a minimal angle set (see Fig. 9(a)).

(a)
(b)
Figure 9: Illustrations of the proposed experimental set-ups for the Fast Mojette Transform. Experiment (a) utilises a large detector with a resolution equal to that of the Mojette projection requiring the highest resolution. Experiment (b) utilises a small fixed resolution detector that is scanned along at an angle (see inset and Eq. (18)) with respect to the view angle of the Mojette projection. The angle is used to ensure that the (fixed width) pixels of the detector encompass the Mojette ray sums.

This detector has to have the maximum resolution of

(17)

for unit pixel area, which is effectively the separation between the translates of the angle corresponding to (see ((51), pg. 20-21) for proofs). Since the remaining projections require a lower detector resolution, these projections become a sub-region within this detector and can be recovered by sub-sampling measurements on the detector. The other redundant parts of the detector in these cases may be used to estimate noise on those projections. The projections can then be converted exactly to FRT projections using the Eqs in (16) and reconstruction process may be started immediately as slices become available.

For detectors smaller than in size, the detector can be translated or scanned across the object in the direction perpendicular to the projection angle, similar to the linogram approach ((52)) (see Fig. 9(b)). To overcome limitations in fixed detector resolution and fixed pixel width or size , the detector is scanned at angle with respect to . The angle is given as

(18)

for each vector , where is the separation between the translates (see also inset of Fig. 9(b)). This ensures that the entire width of the pixel in the detector encompasses its corresponding ray sum without the need for a multi-resolution detector. However, the detector sensitivity may be adversely affected by the angle .

The final method is to use a multi-resolution detector, either of size (being in size) or smaller to scan the detector at the angle in relation to the object. When utilising the DPM for a continuous object, the projection needs to be convolved with a beam-spread function to account for fractional sampling of the array  ((53)). For the simple angle set, this function is always a triangle with a base of and a height of one for the vector . In the next section, the issues regarding the suppression and the effects of noise are discussed.

2.5 Noise

The response of the FMT algorithm to inconsistencies in projection data, such as noise from detectors, can be minimised in two non-exclusive ways. Firstly, the redundancy parameter can be used to “spread out” these inconsistencies over a larger image area. For noise that is statistical in nature, and therefore the same on average for different image sizes, the same noise is present over a larger area for larger image sizes. Thus, this noise is not as prominent over the reconstruction sub-region. Equivalently, the noise is averaged out due to having more translates in the projection data. However, since the larger image area requires more projections, it exposes the object to potentially more radiation. This is a non-issue for data encoding and/or transmission applications.

Secondly, the redundancy in the entire image, after the inversion of the MT, can be used as part of a iterative algorithm for reducing the noise. For example, an or -norm minimising (iterative) algorithm (applied to the pixel values) can be used to make image values in the redundant region converge to zero ((54); (55)). The result of this minimising algorithm gives an approximation to the noise in the projection data, reducing the effects of these inconsistencies within the image. This approach will be most useful in very noisy or incomplete projection data.

The redundancy within the image area can also be utilised to reduce the number of projections as described by Chandra et al. ((32)) for the noise-free projection data. Any missing FRT projections cause finite ghosts superimposed on the image that can be removed exactly via a non-iterative algorithm applied to these ghosts in the redundant area. The number of projections is then just the number of rows in the initial object, which is generally much less than . Ghost removal in the presence of noise is still an active area of research. The next section describes the performance and noise response of the FMT and the MT.

2.6 Results

Numerical simulations of the FMT were conducted using the DPM of a image of Lena on a Intel™ Core2Duo™ E6600 (2.4GHz) processor. Fig. 10 shows an example of a simulated reconstruction using the FMT of the image with 3% Gaussian noise present in the projections within a FFT space (i.e. the redundancy parameter of ).

(a)
(b)
(c)
Figure 10: A noise response numerical simulation of the inverse Fast Mojette Transform using the simple angle set. (a) shows the original image of Lena. (b) shows the reconstruction with and Gaussian noise. (c) shows the actual errors between (a) and (b) as a surface plot. On average, the error is approximately 5 out of 256 grey-scales per pixel.

The reconstruction (Fig. 10(b)) shows that the result is stable to moderate levels of noise, having a Peak-Signal-to-Noise (PSNR) of approximately 35dB, possibly suitable for lossy image and video encoding. The errors present on the reconstruction are shown in Fig. 10(c), which has a Root Mean Squared Error (RMSE) of 4.8 grey-scales out of a possible 8-bit (256) grey-scales. Note that the noise is uniformly distributed and no image related artefacts are apparent.

(a)
(b)
Figure 11: A comparison of the angle sets for the Fast Mojette Transform (FMT) with a image of Lena. (a) shows the Root Mean Squared Error (RMSE) with increasing redundancy . (b) shows the computation times of the inverse FMT and the forward Mojette Transform (MT) on a log scale with the time in micro-seconds ( secs).

Graph 11(a) shows the noise response of the two different angle sets as a function of the redundancy parameter . The parameter values are defined as powers of two to ensure that the FFT space is also a power of two for the Cooley-Tukey algorithm ((44)). Increasing the parameter spreads out the noise as predicted and exponentially decreases the error on the reconstruction. The simple angle set, which is a subset of the minimum set, has a better noise response because it possesses more translates than the full minimum set.

Graph 11(b) shows the computation times of the MT and the inverse FMT in micro-seconds. The inverse FMT, although an order of magnitude slower than the MT for very large images, is suitably fast as the time for reconstructing a image from 6144 projections is approximately 13 seconds. The time could be greatly reduced using Graphical Processing Units (GPUs). Considering both graphs, the optimal value for the parameter appears to be around or 4, in order to balance noise response with the speed of computation. Further work on the FMT needs to be done in selecting optimal experimental geometries and conducting tomographic experiments with real data.

Conclusion

A noise tolerant algorithm for fast digital-to-digital tomography was constructed (see graphs in Fig. 11). The angle set for the Mojette Transform was redefined to one that was more symmetric and minimal in size, so that inversion could be computed with low complexity (see Eq. (14)). Two choices for the angle sets were constructed that were either computationally or experimentally suitable for variable or fixed image geometry respectively (see Eqs (12) and (13)). A new analytical mapping of these projections was constructed that allowed them to be compacted directly and exactly into a Discrete Fourier space of desired size (see Eq. 16). Once the projections are mapped, the (robust to noise) reconstruction can be obtained with a computational complexity with for an space (see Fig. 10). Redundancy within this space was used to control inconsistencies in projection data, such as detector noise. The redundancy also allows the exploitation of or -norm minimising (iterative) algorithms to values within the space to further reduce noise. Experiments and techniques to apply this Fast Mojette Transform for real measurements were also proposed (see Fig. 9).

Acknowledgements

S. Chandra would like to thank the Faculty of Science, Monash University for a Ph.D scholarship and a publications award. N. Normand would like to thank the Australian Research Council for his International Fellowship.

Appendix A Discrete Tomography Review

Discrete reconstruction methods take advantage of digital image geometry of the reconstructions to allow exact inversion schemes. The methods of Beylkin ((9)) and, Kelley and Madisetti ((10)) utilise projections along arbitrary curves and reconstruction via Block-Circulant matrices. However, the method requires the storage of a large number of matrices and has an unfavourable computational complexity of for a 2D image.

Figure 12: The -lines for a image used to approximate continuous lines ((11); (12)). A line, parametrised by -intercept and rise , is divided into two lines comprising the to and to pixels respectively. The final -intercept of the first becomes and the -intercept for the second is redefined as .

Götz and Druckmüller ((11)), and Brady ((12)) independently introduced the concept of -lines, which approximate lines of arbitrary (including irrational) slopes (see Fig. 12). Sums along -lines approximate the classical linogram and an iterative algorithm has been developed which can recover the image with desired accuracy by Press ((56)).

Figure 13: The Fourier Slice Theorem mapping within the Fast Slant Stack ((13)). The row-wise pseudo-polar sampling, created by the linogram mapping of slices, is cast to Cartesian Fourier space in order to use the Fast Fourier Transform. The same is done to the columns of the mapping.

The last approach is to utilise the pseudo-polar Fourier Transform (FT((57)). The Fast Slant Stack method bypasses the need for the filtering within the FST via this pseudo-polar FT having a linogram mapping of the slices ((13)). However, the coefficients are still interpolated (via the Chirp- transform) to Cartesian Fourier space in order to use the FFT (see Fig. 13).

Acronyms

1D
one dimensional
2D
two dimensional
RT
Radon Transform
CT
Computerised Tomography
FT
Fourier Transform
FST
Fourier Slice Theorem
DRT
Discrete Radon Transform
FRT
Fast Radon Transform
MT
Mojette Transform
FMT
Fast Mojette Transform
FFT
Fast Fourier Transform
DFT
Discrete Fourier Transform
NTT
Number Theoretic Transform
DPM
Dirac Pixel Model

Footnotes

  1. journal: Signal Processing
  2. A discussion of the effects of scattering and beam hardening in polychromatic x-ray beams is beyond the scope of this paper.

References

  1. A. G. Ramm, A. I. Katsevich, Radon Transform and Local Tomography, CRC Press, 1996.
  2. J. Radon, Über die bestimmung von funktionen durch ihre integralwerte längs gewisser mannigfaltigkeiten, Berichte Sächsische Acadamie der Wissenschaften, Leipzig, Math.-Phys. Kl. 69 (1917) 262–267.
  3. P. Funk, Über eine geometrische Anwendung der Abelschen Integralgleichung, Mathematische Annalen 77 (1915) 129–135. doi:10.1007/BF01456824.
  4. K. T. Smith, D. C. Solmon, S. L. Wagner, Practical and mathematical aspects of the problem of reconstructing objects from radiographs, Bulletin of the American Mathematical Society 83 (6) (1977) 1227–1270.
  5. B. F. Logan, The uncertainty principle in reconstructing functions from projections, Duke Mathematical Journal 42 (4) (1975) 661–706.
  6. M. Katz, Questions of Uniqueness and Resolution in Reconstruction from Projections, Lecture Notes in Biomathematics, Springer-Verlag, 1977.
  7. R. N. Bracewell, J. A. Roberts, Aerial Smoothing in Radio Astronomy, Australian Journal of Physics 7 (1954) 615–640.
  8. A. K. Louis, Ghosts in tomography - the null space of the Radon transform, Mathematical Methods in the Applied Sciences 3 (1981) 1–10. doi:10.1002/mma.1670030102.
  9. G. Beylkin, Discrete Radon transform, Acoustics, Speech, Signal Processing, IEEE Transactions on 35 (2) (1987) 162–172.
  10. B. Kelley, V. Madisetti, The fast discrete Radon transform, ICASSP-92: 1992 IEEE International Conference on Acoustics, Speech and Signal Processing (Cat. No.92CH3103-9) (1992) 409 – 12.
    URL http://dx.doi.org/10.1109/ICASSP.1992.226189
  11. W. Götz, H. Druckmüller, A fast digital Radon transform-an efficient means for evaluating the Hough transform, Pattern Recognition 29 (4) (1996) 711–718.
    URL http://dx.doi.org/10.1016/0031-3203(96)00015-5
  12. M. L. Brady, A fast discrete approximation algorithm for the Radon transform, SIAM Journal on Computing 27 (1) (1998) 107–119. doi:10.1137/S0097539793256673.
    URL http://link.aip.org/link/?SMJ/27/107/1
  13. A. Averbuch, D. Donoho, R. Coifman, M. Israeli, J. Walden, Fast slant stack: A notion of Radon transform for data on a cartesian grid which is rapidly computable, algebraically exact, geometrically faithful and invertible, Tech. Report, Stanford University 11.
  14. J.-P. Guédon, D. Barba, N. Burger, Psychovisual image coding via an exact discrete Radon transform, Proc. of the SPIE - The International Society for Optical Engineering 2501 (1995) 562–572.
  15. J. Franel, Les suites de farey et le probléme des nombres premiers, Gottinger Nachr (1924) 191–201.
  16. E. Landau, Bemerkungen zu der vorstehenden abhandlung von herrn franel, Göttinger Nachr (1924) 202–206.
  17. J.-P. V. Guédon, N. Normand, Mojette transform: applications for image analysis and coding, Visual Communications and Image Processing ’97 3024 (1) (1997) 873–884.
    URL http://link.aip.org/link/?PSI/3024/873/1
  18. N. Normand, A. Kingston, P. Évenou, A geometry driven reconstruction algorithm for the Mojette transform, in: Lecture Notes in Computer Science (LNCS), Vol. 4245, Springer Berlin / Heidelberg, 2006, pp. 122–133.
  19. M. Servières, N. Normand, J.-P. Guédon, Y. Bizais, The Mojette transform: Discrete angles for tomography, in: G. Herman, A. Kuba (Eds.), Proceedings of the Workshop on Discrete Tomography and its Applications, Vol. 20, Electronic Notes in Discrete Mathematics, 2005, pp. 587–606.
  20. N. Normand, J.-P. Guédon, O. Philippe, D. Barba, Controlled redundancy for image coding and high-speed transmission, Proc. of the SPIE - The International Society for Optical Engineering 2727 (1996) 1070–1081.
  21. M. Servières, J. Idier, N. Normand, J.-P. Guédon, Conjugate gradient Mojette reconstruction, Proc. of the SPIE - The International Society for Optical Engineering 5747 (1) (2005) 2067–2074.
    URL http://dx.doi.org/10.1117/12.593399
  22. O. Philippé, J.-P. Guédon, F. Terrien, ATM source-channel image coding, Visual Communications and Image Processing 3024 (1) (1997) 1220–1230. doi:10.1117/12.263202.
    URL http://link.aip.org/link/?PSI/3024/1220/1
  23. F. Terrien, J.-P. V. Guédon, O. Philippe, Secure coding for medical sequence transmission over ATM network, Medical Imaging 1997: PACS Design and Evaluation: Engineering and Clinical Issues 3035 (1) (1997) 200–209. doi:10.1117/12.274572.
    URL http://link.aip.org/link/?PSI/3035/200/1
  24. B. Parrein, P. Verbert, N. Normand, J.-P. V. Guédon, Scalable multiple descriptions on packets networks via the n-dimensional Mojette transform, Quality of Service over Next-Generation Data Networks 4524 (1) (2001) 243–252.
    URL http://link.aip.org/link/?PSI/4524/243/1
  25. P. Verbert, J.-P. V. Guédon, B. Parrein, Distributed and compressed multimedia transmission using a discrete backprojection operator, Internet Multimedia Management Systems III 4862 (1) (2002) 315–325. doi:10.1117/12.473047.
    URL http://link.aip.org/link/?PSI/4862/315/1
  26. J.-P. V. Guédon, N. Normand, P. Verbert, B. Parrein, F. Autrusseau, Load-balancing and scalable multimedia distribution using the Mojette transform, Internet Multimedia Management Systems II 4519 (1) (2001) 226–234.
    URL http://link.aip.org/link/?PSI/4519/226/1
  27. F. Autrusseau, J. V. Guédon, Image watermarking for copyright protection and data hiding via the Mojette transform, Security and Watermarking of Multimedia Contents IV 4675 (1) (2002) 378–386. doi:10.1117/12.465296.
    URL http://link.aip.org/link/?PSI/4675/378/1
  28. F. Autrusseau, J. V. Guédon, Image watermarking in the Fourier domain using the Mojette transform, 2002 14th International Conference on Digital Signal Processing Proceedings. DSP 2002 (Cat. No.02TH8628) 2 (2002) 725–728.
    URL http://dx.doi.org/10.1109/ICDSP.2002.1028193
  29. F. Autrusseau, J. Guédon, Y. Bizais, Mojette cryptomarking scheme for medical images, Proc. of the SPIE - The International Society for Optical Engineering 5032 (2003) 958–965.
    URL http://dx.doi.org/10.1117/12.480296
  30. M. Babel, B. Parrein, O. Deforges, N. Normand, J.-P. Guédon, J. Ronsin, Secured and progressive transmission of compressed images on the internet: application to telemedicine, Internet Imaging VI 5670 (1) (2005) 126–136.
    URL http://link.aip.org/link/?PSI/5670/126/1
  31. M. Servières, J.-P. Guédon, N. Normand, A discrete tomography approach to PET reconstruction, in: Proc. 7th International Conf. on Fully 3D Reconstruction in Radiology and Nuclear Medicine, 2003.
  32. S. S. Chandra, I. Svalbe, J.-P. Guédon, An exact, non-iterative Mojette inversion technique utilising ghosts, in: Lecture Notes in Computer Science (LNCS), Vol. 4992, Springer Berlin / Heidelberg, 2008, pp. 401–412.
  33. H. Fayad, J. P. Guedon, I. Svalbe, Y. Bizais, N. Normand, Applying Mojette discrete Radon transforms to classical tomographic data, Medical Imaging 2008: Physics of Medical Imaging 6913 (1) (2008) 69132S. doi:10.1117/12.770478.
    URL http://link.aip.org/link/?PSI/6913/69132S/1
  34. J.-P. Guédon, N. Normand, A. Kingston, B. Parrein, M. Servières, P. Évenou, I. Svalbe, F. Autrusseau, T. Hamon, Y. Bizais, D. Coeurjolly, F. Boulos, E. Grail, The Mojette Transform: Theory and Applications, ISTE-Wiley, 2009.
  35. C. Jordan, Traité des substitutions et des équations algébriques, Gauthier-Villars, Paris, 1870.
  36. J. P. S. Kung, The Radon transforms of a Combinatorial Geometry, I, Journal of Combinatorial Theory, Series A 26 (2) (1979) 97 – 102. doi:DOI:␣10.1016/0097-3165(79)90059-1.
    URL http://www.sciencedirect.com/science/article/B6WHS-4D8D%˙PY1-5N/2/2919cc9ecf355376df75bf47d0901506
  37. V. Labunets, Superfast multidimensional Fourier Radon transforms and multidimensional convolutions, Statistical Methods of Signal Processing IX (1985) 140–142.
  38. P. Diaconis, R. L. Graham, The Radon transform on , Pacific Journal of Mathematics 118 (1985) 323–345.
  39. A. Grigoryan, New algorithms for calculating the discrete Fourier transforms, J. Vichislit. Matem. i Mat. Fiziki 25 (9) (1986) 1407–1412.
  40. E. D. Bolker, The Finite Radon Transform, Contemporary Mathematics (American Mathematical Society) 63 (1987) 27–49.
  41. I. Gertner, A new efficient algorithm to compute the two-dimensional discrete Fourier transform, Acoustics, Speech and Signal Processing, IEEE Transactions on 36 (7) (1988) 1036–1050. doi:10.1109/29.1627.
  42. J. A. Fill, The Radon Transform on , SIAM Journal on Discrete Mathematics 2 (2) (1989) 262–283. doi:10.1137/0402023.
    URL http://link.aip.org/link/?SJD/2/262/1
  43. F. Matúš, J. Flusser, Image Representation via a Finite Radon Transform, Pattern Analysis and Machine Intelligence, IEEE Transactions on 15 (10) (1993) 996–1006.
    URL http://dx.doi.org/10.1109/34.254058
  44. J. W. Cooley, J. W. Tukey, An Algorithm for the Machine Calculation of Complex Fourier Series, Mathematics of Computation 19 (90) (1965) 297–301. doi:http://dx.doi.org/10.2307/2003354.
    URL http://dx.doi.org/10.2307/2003354
  45. C. M. Rader, Discrete Fourier transforms when the number of data samples is prime, Proceedings of the IEEE 56 (6) (1968) 1107–1108.
  46. A. Kingston, I. Svalbe, Generalised finite Radon transform for NxN images, Image and Vision Computing 25 (10) (2007) 1620 – 1630, Discrete Geometry for Computer Imagery 2005. doi:DOI:␣10.1016/j.imavis.2006.03.002.
    URL http://www.sciencedirect.com/science/article/B6V09-4KF7%˙84N-1/2/8c70f3504e93d23e9b11e87981762893
  47. S. S. Chandra, Exact image representation via a Number-Theoretic Radon Transform, IET Computer Vision In Peer Review, submitted April.
  48. P. M. Salzberg, R. Figueroa, Tomography on the 3D-torus and crystals, in: G. T. Herman, A. Kuba (Eds.), Discrete Tomography: Foundations, Algortihms and Applications, Birkhäuser, 1999, Ch. 19.
  49. I. Svalbe, D. van der Spek, Reconstruction of tomographic images using analog projections and the digital Radon transform, Linear Algebra and Its Applications 339 (2001) 125–145.
    URL http://dx.doi.org/10.1016/S0024-3795(01)00487-6
  50. I. Svalbe, A. Kingston, Farey sequences and discrete Radon transform projection angles, Electronic Notes in Discrete Mathematics 12 (2003) 154 – 165, 9th International Workshop on Combinatorial Image Analysis. doi:DOI:␣10.1016/S1571-0653(04)00482-2.
    URL http://www.sciencedirect.com/science/article/B75GV-4FWS%˙P7W-7S/2/9f36405de4b0cf6719fe4dac61e78f10
  51. C. Olds, A. Lax, G. Davidoff, The Geometry of Numbers, Vol. 41 of The Anneli Lax New Mathematical Library, The Mathematical Association of America, 2000.
  52. P. Edholm, G. Herman, Linograms in image reconstruction from projections: Part 1. back projecting using the Radon transform, Proc. Twenty-First Annual Hawaii International Conference on System Sciences 4 (1988) 48–50.
  53. A. Kingston, I. Svalbe, Mapping between digital and continuous projections via the discrete Radon transform in Fourier space, Proc. 7th International Conference on Digital Image Computing: Techniques and Applications, CSIRO Publishing (2003) 263–272.
  54. K. P. Bube, R. T. Langan, Hybrid minimization with applications to tomography, Geophysics 62 (4) (1997) 1183–1195. doi:10.1190/1.1444219.
    URL http://link.aip.org/link/?GPY/62/1183/1
  55. E. Candès, J. Romberg, T. Tao, Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information, Information Theory, IEEE Transactions on 52 (2) (2006) 489–509. doi:10.1109/TIT.2005.862083.
    URL http://dx.doi.org/10.1109/TIT.2005.862083
  56. W. H. Press, Discrete Radon transform has an exact, fast inverse and generalizes to operations other than sums along lines, Proceedings of the National Academy of Sciences (PNAS) 103 (51) (2006) 19249–19254.
  57. A. Averbuch, R. Coifman, D. Donoho, M. Israeli, J. Walden, The pseudo-polar FFT and its applications, Technical Report YALEU/DCS/RR, Yale University 1178.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
105162
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description