HOME
PDF

Mathematical formulation of the HYPR algorithms

Nasser M. Abbasi

California State University, Fullerton. Summer 2008 page compiled on July 1, 2015 at 10:07pm

1 Introduction

This report is a summary of the HIghly constrained Back PRojection (HYPR) team work performed so far relating to the HYPR research project. We will describe the work done and results found.

The goals set for the HYPR project included formulating the HYPR algorithm and some of its variations (such as Wright-Huang HYPR (WH-HYPR), I-HYPR and HYPR-LR) in a mathematical framework which would allow the study and analyze of these algorithms in relation to other well known non-linear methods such as maximum a-posteriori (MAP) estimation and Maximum Likelihood Expectation Maximization (MLEM). These algorithms, like HYPR, use prior information on the object being reconstructed and they are extensively used in nuclear medicine where the data is intrinsically under sampled.

The initial period of this project, which this report reflects on, was spent becoming familiar with the HYPR algorithm and its connection to MLEM.  Towards this goal, the HYPR algorithm was formulated mathematically and schematic diagrams created which helped in its implementation. MATLAB simulation software was developed to enable more understanding of the algorithm and its behavior by running it on a number of test cases. Initial comparison between the original HYPR and the WH-HYPR made on a number of different test configuration which are described in detail in the simulation section below. The MLEM algorithm was implemented and compared the HYPR algorithm.

In addition, A mathematical connection between HYPR and Expectation Maximization (EM) is described and formulated.

2 Mathematical formulation of the HYPR algorithms

2.1 Original HYPR

2.1.1 Mathematical formulation

Please see the appendix for a complete description of the notation used in this section and throrught the rest of the report.

The mathematics of this algorithm will be presented by using the radon transform R  notation and not the matrix projection matrix H  notation.

The projection st  is obtained by applying radon transform R  on the image It  at some angle ϕ
  t

s = R   [I]
 t    ϕt  t

When the original object image does not change with time one can drop the subscript t  from It  and just write st = Rϕt [I]

Next, the composite image C  is found from the filtered back projection applied to all the st  as follows

     ∑N
C  =    Rfϕt [sti]
     i=1    i

Notice that the sum above is taken over N  and not over Np  . Next, a projection sc  is taken from C  at angle ϕ  as follows

s  =  R  [C ]
 ct    ϕt

Then the unfiltered back projection 2-D image Pt  is generated

       u
Pt = R ϕt [st]

And the unfiltered back projection 2-D image Pct   is generated

       u
Pct = R ϕt [sct]

Then the ratio of PPtc-
  t   is summed and averaged over the time frame and the result multiplied by C  to generate a HYPR frame J  for the time frame.  Hence for the   th
k  time frame we obtain

pict

2.1.2 Schematic diagram

PIC

2.2 Wright-Huang variation of HYPR

2.2.1 Mathematical formulation

This mathematics of this algorithm will be presented by using the radon transform R  notation and not the matrix projection matrix H  notation.

The projection st  is obtained by applying radon transform R  on the image It  at some angle ϕt

st = R ϕt [It]

The composite image C  is found from the filtered back projection applied to all the st

     ∑N   f
C  =    R ϕti [sti]
     i=1

Notice that the sum above is taken over N  and not over Np  . Next a projection sc  is taken from C  at angle ϕ  as follows

sc =  Rϕ [C ]
  t     t

Then the unfiltered back projection 2-D image P
  t  is generated

P  = Ru  [s ]
  t    ϕt  t

And the unfiltered back projection 2-D image Pct   is generated

P  = Ru  [s ]
 ct    ϕt  ct

Now the set of Pt  and Pct   over one time frame are summed the their ratio multiplied by C  to obtain the  th
k  HYPR frame

pict
2.2.2 Schematic diagram

PIC

3 HYPR connection to Expectation Maximization

The following is a discussion of the Mathematics that connects the MLEM algorithm to HYPR.

According to O'Halloran's paper[1] the for Maximum-Likelihood Expectation-Maximization (MLEM) algorithm is mathematically equivalent to HYPR. The MLEM algorithm can be used in image reconstruction for medical purposes. Positron Emission Tomography (PET) and Single-Photon Emission Computed Tomography (SPECT) are two types of image reconstruction processes where the MLEM algorithm is used. The purpose here is to show that the MLEM algorithm will work for HYPR reconstructions.

The MLEM algorithm is a process that approximates the solution to

g = H 𝜃
(1)

In connection to HYPR, we can view H  as a forward projection matrix, 𝜃  as the original image being projected and g  the projection produced. The goal is to relate the above matrix based formulation to the radon transform based formulation seen above in the HYPR mathematical section, which is

st = R ϕt [It]
(2)

We formulate the first iteration of the MLEM algorithm based on equation (1) and see how it can be translated into the HYPR process of image reconstruction. The first step of MLEM is

               M
ˆ𝜃(1)= 𝜃ˆ(0)∗ -1-∑  H    (--gm-)---
 n     n    zn      mn    ˆ(0)
              m=0       H 𝜃    m
(3)

This can be rewritten in matrix form

 ˆ(1)   ˆ(0)-1-⌊⌈  T ⌊⌈ ---g----⌋⌉ ⌋⌉
𝜃n  =  𝜃n z   H     (  ˆ(0))
           n         H 𝜃       n

If we replace 1-
zn   by ---1---
[HT[1]]n   in (4), we obtain

                   ⌊unfiltered back projection⌋
                      ◜ --⌊--◞◟-----⌋◝
                   ||                    ||
ˆ𝜃(1) = ˆ𝜃(0)---1-----||  HT  ⌈ (--g---)⌉   ||
 n     n  [HT  [1]]n |         H ˆ𝜃(0)     |
                   ⌈                    ⌉
                                         n
(4)

The marked portion of the above equation can be viewed as the vector that is produced from unfiltered back projection on the image produced by the ratio

--g---
H ˆ𝜃(0)

Here the division is done element by element to produce the vector whose elements are the ratios of the respective elements of g  and H ˆ𝜃(0).

In HYPR the equation we want to tie to equation (3) above is as follows

Jt = -1-C  ⋅ Ruϕt [-st---]
     Np           Rϕt (C )
(5)

Where the ⋅ represents an element by element multiplication, and the terms in (5) as defined in the section of the HYPR mathematical derivation shown earlier. Hence for (5) and (4) to be equivalent,  We must have

       [      ]
 ˆ(0)   -HT--[1]-n-
𝜃 n =     N     C
            p

Which represents the initial guess for the user image.  Therefore, by using for ˆ(0)
𝜃   as an initial guess for the MLEM algorithm the above weighted term of the composite image C  , the MLEM algorithm will produce  𝜃ˆ(1)   which is a better approximation to the original image from that of the composite C  . And this is what the HYPR algorithm does. It uses the composite image C  to produce the HYPR image J  to approximate the original user image I  . Hence a one step of MLEM is equivalent to running HYPR for one time. Therefore, iterative HYPR algorithms can be seen as a multi step application of MLEM.

4 Software simulation and results

4.1 HYPR simulation

A software simulation written in MATLAB was designed and developed to enable more extensive HYPR testing of different test configurations. The software is GUI based and all test results are saved in a tab-delimited plain text file to allow one further statistical analysis of the data generated by other software. The appendix contains a screen shot of the current version of the simulator (version 1.0).

4.1.1 Description of HYPR simulation and test results

This is a description of the different tests performed. Both the original HYPR and the Wright-Huang HYPR (WH-HYPR) were run and results compared. In the following discussion, we use N
 p  to mean number of projections in one time frame, and Nw  to mean the number of time frames. Hence the total number of projections is NpNw

The table below describes each test. In this table, a test with the letter 'a' represents the test being run using the original HYPR and a test with the letter 'b' represents the test being run using WH-HYPR. Each test was run under both the original HYPR and WH-HYPR.

The first set of tests are designed to detect the effect of Poisson noise on the accuracy of the HYPR algorithm as compared to the WH-HYPR. This was done for different geometry of objects while keeping the number of projections per time frame and the Poisson noise parameter λ  fixed.

The second set of tests are designed to detect the effect of Gaussian noise on the accuracy of the HYPR algorithm as compared to the WH-HYPR. This was done for different geometry of objects while keeping the number of projections per time frame and the Gaussian distribution parameters (mean and variance) fixed. The set of tests used here is smaller than the first set due time limitation.

The third set of tests are designed to detect the effect of increasing the number of projections on the accuracy of the HYPR and WH-HYPR. This was done under one fixed configuration and with the absence of noise.

The main measure of accuracy used was relative RMSE. This was calculated as follows: For each HYPR frame image generated, the set of user images which make up the time frame from which the HYPR frame was generated are averaged to obtain an average time frame image. Then the RMSE was obtained between these 2 images as follows: Assuming these are N  total pixels in each image, the error between each corresponding pixels is found as Hi − Ii  where Hi  is a pixel in the HYPR frame image and Ii  is the corresponding pixel in the averaged time frame image. This error is then squared. This was done for each pixel. The average of these square values is found, and the square root of the result is found. Hence

          ┌│  ------∘-----------
          │∘   1∑N             2
RM  SE  =    N--     (Hi − Ii)
                 i

This quantity is normalized by dividing it by the mean intensity of the averaged time frame image found earlier. This gives a normalized RMSE value for each time frame generated. When there are more than one time frame generated then the average all these RMSE values are used to obtain one representative value of the RMSE for the test, and that is the value showed in the tables below for the purpose of comparing different tests.

Other statistics are calculated to determine the algorithms accuracy. The relative error between the HYPR image and the averaged time frame is found using the standard formula for relative error. This measure however did not appear to be a good indicator for determining the accuracy of the HYPR image. Another statistical measure used is the histogram difference, which is found as follows. The histogram for each HYPR image frame and the corresponding histogram for the averaged time frame image are calculated and the difference between these histograms found. This measure appear to give a good indication of the performance of each test and correlated well with the RMSE measure used. These results are all written to the log file for further analysis, but are not currently taken into account in the following tests due to time limitation. Only the RMSE measure is currently used to determine the accuracy of the algorithm. The following sections describe each set of tests in more details.

4.1.2 The first set of tests
|--------|------------------------------------------------------------------------------------------|
|Test----|Test-description--------------------------------------------------------------------------|
|        |  HYPR  [4] algorithm validation. Using the same parameters as the Wright-Huang           |
|        |                                                                                          |
|        |  paper[2] and validate output as it was described and  shown  in the paper.                 |
|1a      |  This is a fixed disk in the center of the image  whose density changes linearly with  time.   |
|        |                                                                                          |
|        |  Np = 8,Nw  =  16                                                                        |
|--------|------------------------------------------------------------------------------------------|
|1b------|As--above,-but-use-the-WH--HYPR----algorithm.----------------------------------------------|
|        |  Repeat test 1 but with the addition of Poisson noise with λ = 500 to the                 |
|2(a,b)  |                                                                                          |
|--------|--projection-s-generated-from-the-user-images----------------------------------------------|
|        |  A non -time varying two  small white disks close to each others in black background.       |
|3(a,b)  |                                                                                          |
------------Np-=-8,Nw--=--16------------------------------------------------------------------------|
|        |                                                                                          |
|4(a,b)--|As--above,-but-with-Poisson-noise-with-λ-=-500-added--to projection-s.--------------------|
|5(a,b)  |Small  disk that moves  in vertical motion off the center of image. Np = 8,Nw  = 16          |
|--------|------------------------------------------------------------------------------------------|
|6(a,b)--|As--above,-but-with-Poisson-noise-with-λ-=-500-added--to projection-s.--------------------|
|7(a,b)  |2 small disks close to each others that move  in vertical motion. N  =  8,N   = 16           |
|--------|-----------------------------------------------------------------p-------w----------------|
|8(a,b)  |As  above, but with Poisson noise with λ = 500 added  to projection s.                    |
|--------|------------------------------------------------------------------------------------------|
|9(a,b)--|2-small-disks-further-apart-from-each-others-that-move--in-vertical motion.-----------------|
|10(a,b) |As  above, but with Poisson noise with λ = 500 added  to projection s.                    |
|--------|------------------------------------------------------------------------------------------|
|11(a,b)-|one--small-disk-that-move--across the-image-in-the-diagonal-direction.-Np-=--8,Nw--=-16------|
|12(a,b) |As  above, but with Poisson noise with λ = 500 added  to projection s.                    |
----------------------------------------------------------------------------------------------------

The appendix shows the output obtained from the above set of tests. We now present a summary of the results

|-----------|-(a)-Original--|--(b-) Wright--|-----------------|
|Test       | HYPR          |  HYPR        |  Selected        |
|           |               |              |  algorithm       |
|-----------|-RMSE----------|--RMSE--------|-----------------|
|1          |0.639          |0.636         | WH  -HYPR        |
|-----------|---------------|--------------|-----------------|
|2 (noise)  |1.7298         |1.2079        | WH  -HYPR        |
|-----------|---------------|--------------|-----------------|
|3----------|1.0329---------|1.0411--------|-Original-HYPR----|
|4 (noise)  |1.9879         |1.4917        | WH  -HYPR        |
|-----------|---------------|--------------|-----------------|
|5----------|2.6349---------|3.095---------|-Original-HYPR----|
|           |               |              |                 |
|6-(noise)--|4.9216---------|4.3288--------|-WH---HYPR--------|
|7          |2.1157         |2.3496        | Original HYPR    |
|-----------|---------------|--------------|-----------------|
|8-(noise)--|2.99-----------|2.7793--------|-WH---HYPR--------|
|9          |2.151          |2.3524        | Original HYPR    |
|-----------|---------------|--------------|-----------------|
|10-(noise)-|2.9983---------|2.818---------|-WH---HYPR--------|
|           |               |              |                 |
|11---------|2.558----------|3.083---------|-Original-HYPR----|
|12 (noise) |4.881          |4.3884        | WH  -HYPR        |
-------------------------------------------------------------

Observation from running the first set of tests  The original HYPR algorithm performed better in each test when noise is absent from projection. This occurs in either time varying or non-time varying configuration. On the other hand, WH-HYPR performed better in each case when noise was present. This occurs in either time varying or non-time varying configuration.

4.1.3 The second set of tests

These tests are a repeat of the first set of tests, but with noise generated from normal distribution. Due to time limitation only test 2,6 and 10 described above are repeated since these 3 tests are good representative of the overall tests. The letter N is added to the test name to indicate the use of Normal distribution.

|----------|--------------------------------------------------------------------------------|
|-Test-----|-Test description---------------------------------------------------------------|
|          |  Repeat test 1 but with the addition of Normal  noise with μ = 0 and σ2 = 500   |
| 2N (a,b) |                                                                                |
|          |  to the projection s generated from the user images                              |
|----------|----------------------------------------------------------------------2---------|
|-6N(a,b)--|-Repeat-test-5, but-with-the-addition-of Normal-noise-with-μ-=-0-and-σ--=-500---|
| 10N(a,b) | Repeat test 9, but with the addition of Normal noise with μ = 0 and σ2 = 500   |
--------------------------------------------------------------------------------------------

The appendix shows the output obtained from the above set of tests. Summary of the results is shown below. To clarify the nature of tests below a short description is given again below.

1.
Test 2N is a small fixed disk, changes intensity linearly with time. Np = 8,Nw  =  16
2.
Test 6N is one disk which moves vertically, off center. N  =  8,N   = 16
  p       w
3.
Test 10N is two disks separated from each others that move vertically across the image. Np =  8,Nw =  16
|-----|----------------|--------------|---------------|-------------|
|     |  (a) Original  | (b) Wright   |               |             |
|Test |  HYPR          | HYPR         |Abs difference  | Selected    |
|     |                |              |               | algorithm    |
|-----|--RMSE----------|-RMSE---------|---------------|-------------|
|     |                |              |               |             |
|2N---|-1.7583----------|1.7179---------|0.0404---------|WH---HYPR----|
|6N   | 4.0069          |3.9797         |0.0272         |WH  -HYPR    |
|-----|----------------|--------------|---------------|-------------|
-10N----2.7754-----------2.7737----------0.0017----------WH---HYPR-----

Observations from running the second set of tests  WH-HYPR performed better in all 3 cases. This correlated well with results found from the first set of tests where it was observed that WH-HYPR performed better each time Poisson noise was added. In the above 3 tests, normal noise was added and it is observed that WH-HYPR performed better.

4.1.4 Third set of tests

As was mentioned earlier, the goal of these tests is to measure the relative accuracy of the algorithms on the same configuration but with increasing number of projections per time frame. It is expected that the accuracy of each algorithm will improve, and we wish to obtain the measure of this improvement as a function of the number of projections per time frame.

For this purpose, the following test configuration was used: small white disk moving vertically and off center, no noise added. One time frame was used and the following number of projections {8,16,32,64, 128,256,512, 1024} . These tests as named 8r,16r,32r,128r,256r, 512r  and 1024r  respectively. The table below show the result of the tests.

|-------|-----------------------|------------------|----------------|----------------|
| Test  |  (a) Original HYPR    | (b) WH  -HYPR    | Abs difference  | Selected       |
|       |  RMSE                 | RMSE             |                | algorithm       |
|-------|-----------------------|------------------|----------------|----------------|
|-8r----|-1.6879-----------------|2.0836------------|-0.3957----------|Original-HYPR---|
|       |                       |                  |                |                |
|-16r---|-1.3772-----------------|1.59--------------|-0.2128----------|Original-HYPR---|
| 32r   | 1.0994                 |1.18845           | 0.0891          |Original HYPR   |
|-------|-----------------------|------------------|----------------|----------------|
|-64r---|-0.774------------------|0.8315------------|-0.0575----------|Original-HYPR---|
| 128r  | 0.5095                 |0.5355            | 0.0260          |Original HYPR   |
|-------|-----------------------|------------------|----------------|----------------|
|-256r--|-0.3722-----------------|0.3765------------|-0.0043----------|Original-HYPR---|
|       |                       |                  |                |                |
|-512r--|-0.2847-----------------|0.2825------------|-0.0022----------|WH---HYPR-------|
| 1024r | 0.2469                 |0.2459            | 0.0010          |WH   HYPR       |
-------------------------------------------------------------------------------------

Observations from running the third set of tests  As the number of projections per time frame increased, the accuracy of WH-HYPR improved. At high number of projections (over 512 per time frame) WH-HYPR bypassed original HYPR and became more accurate. It is not clear at this time if such high number of projections per time frame will conflict with other MRI requirements (sampling rate limitation or other issues), but the above shows that, even with the absence of noise, the WH-HYPR can become more accurate than the Original HYPR but at a cost of having large number of projections per time frame.

4.1.5 Conclusions drawn from HYPR test results

Original HYPR performed better than WH-HYPR when the number of projections is relatively low (below 256 per time frame) and when there was no noise present (noise added to projections taken from the original images). This occurred in all configurations (both objects moving in time or fixed).

WH-HYPR performed better when noise is present (both Poisson and Normal noise) and for all number of projections and for all configurations.

In addition, WH-HYPR performed better when there was no noise added, but when the number of projections per time frame was increased.

These results seem to be a direct consequences of the fact that WH-HYPR sums the backprojection images over a time frame period before taking the ratio of these sums in order to obtain the mask image, while in the original HYPR the ratio for each backprojection images is first found and the ratios added and averaged. More analysis will be needed to better understand this difference and to explain mathematically this observed difference between HYPR and WH-HYPR.

Since real MRI data is characterized by low SNR, this leads one to conclude that WH-HYPR should be the preferred choice between these 2 algorithms.

4.2 Expectation Maximization simulation

4.2.1 Description of simulation

The original HYPR algorithm was compared to 1 step of the MLEM algorithm. A time-invariant white disk with radius 25  pixels centered in a 256  by 256  black image. 128  different projection angles were used (ordered using bit-reversed ordering), and the size of the window was set to 8 projections.

4.2.2 Results of simulation

The figures below are the actual images produced. The composite image, the HYPR reconstruction for the first HYPR frame, and the corresponding MLEM image. The HYPR and the MLEM images are indistinguishable, although the mean absolute error is slightly higher for HYPR than for MLEM. More detailed comparisons of MLEM and HYPR are planned.

PIC

5 Future work

1.
Continue research into iterative HYPR and its connection to Expectation Maximization. Both analytically and through simulation.
2.
Work more on understanding the artifacts (modeling errors) and in the extreme, the pathological cases in which the HYPR algorithms will fail (worst-case scenario).
3.
Characterizing the noise amplification and resolution of the HYPR algorithm through. Simulate HYPR algorithm with projection subjected to different noise distributions and determine which variations of the algorithm are most accurate and under which  conditions.
4.
Investigate and implement a new Iterative HYPR variation proposed during work on this project which uses the Wright-Huang as its iterative step and compare to the current standard I-HYPR which uses the original HYPR and compare performance.
5.
Investigate possibility of a better measure to compare the accuracy of a HYPR image to a time frame image than was used in this report (RMSE), and if one is found, use the new measure for future testing.
6.
Obtained a mathematical description of the HYPR method based on the matrix formulation and not based on the radon transform. Apply for a simple geometrical shape which is time varying.

6 Appendix

6.1 nomenclature

1.
MLEM Maximum-Likelihood Expectation-Maximization
2.
PET Positron Emission Tomography
3.
SPECT Single-Photon Emission Computed Tomography
4.
I  A 2-D image. This represent the original user image at which the HYPR algorithm is applied to.
5.
It  When the original image content changes during the process, we add a subscript to indicate the image I  at time instance t  .
6.
R  radon transform.
7.
R ϕ  radon transform invoked at a projection angle ϕ  .
8.
ϕt  When the projection angle ϕ  is not constant but changes with time during the MRI acquisition process, we add a subscript to indicate the angle at time instance t  .
9.
R ϕt   radon transform invoked at a projection angle ϕt  .
10.
s = R ϕ[I]  . radon transform applied to an image I  at angle ϕ  . This results in a projection vector s  .
11.
H  Forward projection matrix. The Matrix equivalent to the radon transform R  .
12.
𝜃  Estimate of an image I  .
13.
H 𝜃  Multiply the forward projection matrix H  with an image estimate 𝜃  .
14.
g = H 𝜃  Multiply the forward projection matrix H  with an image estimate 𝜃  to obtain a projection vector g  . Notice that for the inner dimensions of the matrix multiplication operation H 𝜃  to be equal, this requires that the 2D image 𝜃  be linearized. In other words, the 2D image 𝜃  be written as a column vector.
15.
  u
R ϕ[s]  The inverse radon transform applied in unfiltered mode to a projection s  which was taken at angle ϕ  . This results in a 2D image.
16.
Rfϕ[s]  The inverse radon transform applied in filtered mode to a projection s  which was taken at angle ϕ  . This results in a 2D image.
17.
  T
H  g  The transpose of the forward projection matrix H  multiplied by the projection vector g  . This is the matrix equivalent of applying the inverse radon transform in an unfiltered mode to a projection s  (see item 12 above).
18.
H+g  The pseudo inverse of the forward projection matrix H  being multiplied by the projection vector g  . This is the matrix equivalent of applying the inverse radon transform in filtered mode to a projection s  (see item 13 above).
19.
C  Composite image generated by summing all the filtered back projections from projections st  of the original images It  . Hence      ∑N   f
C =     R ϕti [sti]
     i=1
20.
Pt  The unfiltered backprojection 2D image as a result of applying Ru  [st]
  ϕt  where st  is projection from user image It  taken at angle ϕt  .
21.
Pct   The unfiltered backprojection 2D image as a result of applying  u
Rϕt [st]  where st  is projection from the composite image C  taken at angle ϕt  .
22.
Np  Number of projections used to generate one HYPR frame image. This is the same as the number of projections per one time frame.
23.
N  The total number of projections used. This is the number of time frames multiplied by Np
24.
Jk  The kth  HYPR frame image. A 2-D image generate at the end of the HYPR algorithm. There will be as many HYPR frame images Jk  as there are time frames.

6.2 Simulation software screen shots

PIC

6.3 HYPR simulation results

6.3.1 Test 1a

PIC

6.3.2 Test 1b

PIC

6.3.3 Test 2a

PIC

6.3.4 Test 2b

PIC

6.3.5 test 3a

PIC

6.3.6 Test 3b

PIC

6.3.7 Test 4a

PIC

6.3.8 Test 4b

PIC

6.3.9 Test 5a

PIC

6.3.10 Test 5b

PIC

6.3.11 Test 6a

PIC

6.3.12 Test 6b

PIC

6.3.13 Test 7a

PIC

6.3.14 Test 7b

PIC

6.3.15 Test 8a

PIC

6.3.16 Test 8b

PIC

6.3.17 Test 9a

PIC

6.3.18 Test 9b

PIC

6.3.19 Test 10a

PIC

6.3.20 Test 10b

PIC

6.3.21 Test 11a

PIC

6.3.22 Test 11b

PIC

6.3.23 Test 12a

PIC

6.3.24 Test 12b

PIC

6.3.25 Test 2N a

PIC

6.3.26 Test 2N b

PIC

6.3.27 Test 6N a

PIC

6.3.28 Test 6N b

PIC

6.3.29 Test 10N a

PIC

6.3.30 Test 10N b

PIC

6.3.31 Test 8r a

PIC

6.3.32 Test 8r b

PIC

6.3.33 Test 16r a

PIC

6.3.34 Test 16r b

PIC

6.3.35 Test 32r a

PIC

6.3.36 Test 32r b

PIC

6.3.37 Test 64r a

PIC

6.3.38 Test 64r b

PIC

6.3.39 Test 128r a

PIC

6.3.40 Test 128r b

PIC

6.3.41 Test 256r a

PIC

6.3.42 Test 256r b

PIC

6.3.43 Test 512r a

PIC

6.3.44 Test 512r b

PIC

6.3.45 Test 1024r a

PIC

6.3.46 Test 1024r b

PIC

6.4 References

References

[1]   Iterative projection reconstruction of time-resolved images using highly-constrained back-projection (HYPR) by Rafael L. O'Halloran, Zhifei Wen, James H. Holmes, Sean B. Fain

[2]   Highly Constrained Back projection for Time-Resolved MRI by C. A. Mistretta, O. Wieben, J. Velikina, W. Block, J. Perry, Y. Wu, K. Johnson, and Y. Wu

[3]   Principles of computerized Tomographic imaging by Kak and Staney

[4]   Highly Constrained Backprojection for Time-Resolved MRI by C. A. Mistretta, Wieben,z J, Velikina,W. Block,J. Perry,Y. Wu. K. ohnson and Y. Wu.