This program is about 3D cone-beam CT for flat detector.
This is made for students who start to learn the CT medical imaging.
- I provide the projection, back-projection codes using built-in GPU functions (semi-GPU codes)
- Above version Matlab 2013b with parallel computing toolbox are needed.
- you can switch on/off the usage of GPU.
- The computational time of this code is several times faster than CPU-based code. Speed depends on your GPU.
- FDK ( known as filtered backprojection)
- MLEM (Maximum Likelihood Expectation Maximization),
- SART (Simultaneous Algebraic Reconstruction Technique)
- SQS (Separable Quadratic Surrogates).
You can run:
0. Please check "ParamSetting.m" carefully.
1. Make a measurement: MeasurementGen.m: F5
1.1 Poisson Noise can be imposed.
2. Reconstruction examples
If I have a set of 2D projections, how do I convert these images into the "img128.m" data file?
Super helpful code. Was able to run it and get a reconstruction on the first try with my own data. Saved me from having to implement FDK from scratch! As with all reconstruction code: read the documentation *first* before attempting to do a reconstruction so that you understand how the code is configured. And *understand your geometry* and CT geometry in general.
how can i run this code ?? My CUDA driver or my GPU device is not proper.
Where is the input into the Matlab code for the location of the central ray from the x-ray source onto the detector? This would be a location on the detector in pixels [vertical and horizontal location]
Hi, how can I make the 'img' from my own 2D images. Thank you
Which Body Part it is?
To use this package,
1. Extract the ZIP file to a folder and add that folder to your path.
2. Run MeasurementGen(). This creates the proj.mat file which is missing when you try to run the other files.
Hi all I dont seem to find the main Demo.m code , could someone please advise?
I put in input a matrix 138*157*101. The only way to run MeasurementGen.m is to impose in param setting x=157 and y=138. Is it normal? After that the code back projection gave me this error
Error in backprojection (line 25)
vol(:,:,iz) = (Ratio.*interp2(proj',pu,pv,param.interptype));
Could someone help me?
Hi Aubrey, you should convert your projection image to double or single data format. When you have image input yo can use imread and then stack it into 3D data.
What format should i insert 'img' as?
I have the series of 2D image projections as TIF
Hello, Kyungsang Kim!
Could you refer me to the paper where you studied to implement the MLEM?
Thanks in advance.
Who knows which paper this file refered to？
thanks you for the codes!
it`s helpful for me!
nice program. just wonder the following line in filtering.m
proj(:,:,i) = fproj(end/2-param.nu/2+1:end/2+param.nu/2,:)/2/param.du*(2*pi/param.nProj)/2*(param.DSD/param.DSO);
can you explain why we need the last several terms instead of just fproj(end/2-param.nu/2+1:end/2+param.nu/2,:)?
Thank you for the code.
Is there beam hardening correction algorithm?
The projection x-ray is mono-energitic or poly-energitic spectrum and what with kVp and mAs?
Thank you so much
good job!besides I wonder if you can tell me which paper did your work refer to ?
Note that the projection implementation has an asymptotically increasing error when it reaches to vertical edges at 45 degrees. This is due to the fact that the way to project and back project is to "stretch" the image to transform the cone to a parallel, thus making the closest single voxels (at 45) incredibly wide. You can test this by projecting a single image with a full white "cube image" at 45 degrees. Strange artifacts arise. This is less relevant the closer to the centre of the image one is. The toolbox is still great to learn about medical imaging, but this error makes it (unfortunately) not a good toolbox for research purposes or real image reconstruction.
You need to run MeasurementGen.m first to generate the proj.mat. Could have been a lot less confusing if proj.mat is pre-generated in the folder. Thanks though.
Where is your proj.mat file? cannot find it?
Dear Kyungsang, may I ask two questions?
1- In comments you mentioned page 96 Jiang Hsieh 2nd edition; however, I cannot follow the code in projection.m, i.e., how it is representing the formula 3.59. May I ask for your hints on it?
2- How can we reach from a CT images of a patient to the sample file which you have provided?
Sorry my mistake, I found proj.mat
Where is proj.mat file?
Great contribution, thank you very much for the code.
Great contribution for others!
Thank you Kyungsang Kim!
Really cool! Thank you for this code! Especially I like to see a real phantom and animated reconstructions.
Thank you Kyung :-)
Please do as following:
1) Open "ParamSetting.m"
2) Change to "param.gpu = 0"
Then, please try again.
This option needs "Parallel Computing toolbox" and "GPU (Hardware)",
If you have a GPU, please check memory of GPU and number of cores.
Also, it would be useful above 2013b version.
Thank you for the code. Is there a description file to explain what each file does and how to run the code please?
When I run Recon1_FDK or MeasurmentGen, they both come up with errors saying some file or function is missing or undefined. More specifically, when running Recon1_FDK error suggests proj.mat does not exist and when I run MeasurementGen the error suggest gpuArray (called in Projection.m) is undefined.
Amny thanks for assistance in advance.
Would it be possible to use this program to forward project a 3D image where dx=dy, but dx=dy!=dz? Right now, my parameters are:
%% Parameter setting %%
param.nx = 400;
param.ny = 400;
param.nz = 66;
%The real detector panel pixel density (number of pixels)
param.nu = 400;
param.nv = 400;
% Detector setting, according to Varian Trilogy OBI (real size)
param.su = 350; % mm
param.sv = 350; % mm
% X-ray source and detector setting
param.DSD = 1500; % Distance source to detector
param.DSO = 1000; % X-ray source to object axis distance
% This confusing stuff
param.sx = param.nx*res; % mm
param.sy = param.ny*res; % mm
param.sz = param.nz*res; % mm
but this doesn't create the correct forward projections, so far as I have been able to check. Am I trying to do something that is impossible in image processing?
Please include analytical equation for better understanding. Or please explain the algorithm in projection.m. If possible please include analytical equation references
That's a good point.
This factor is used in the weighted backprojection for fan-beam and cone-beam geometries.
(see Jiang Hsieh 2nd edition, Figs 3.35 (reason) and 3.37(geometry))
Result is similar compared to the previous one, In real geometry, this factor is not critical factor for the image, because CT rotates 360 deg (average factor ~ 1) as well as itself is close to 1.
However, I think I omitted it for fast calculation, soon I will fix it for accuracy.
Thanks a lot.
ps) you can directly send e-mail to me, it is sometimes difficult to check comments here.
Hi, I compared you work with the FDK equation in some papers, it seems that your FDK algorithm omited a factor. I added this factor in the "backprojection.m" of your algorithm, the code is shown as follows.
vol(:,:,iz) = Ratio.*interp2(uu,vv ,proj',pu,pv,'linear');
Do you think it is necessary to add this factor? I am waiting for your reply, Thank you.
If you have real projections ("RealProj"), then first calculate:
1. I0 = max(RealProj(:)); % if you don't know I0
2. Proj = -log(RealProj/I0);
Here, "I0" is the blank image (intensity of source) as mentioned in the text book.
If you think that the final image has a problem, check 1) "RealProj" has 0 values or 2) "I0" is too high than background values.
Solution of cases
1) step 3. Proj(isinf(Proj)) = 0;
2) step 1. Define I0 as a background (high intensity) average.
step 2. Proj = -log(min(RealProj/I0,1));
Then, you can execute FDK or MLEM as in Demo.m.
Parameter setting is very important. Image resolution is defined by detector size and distances of DSO, DSO. In your case, DSD is 700 and DSO is 560. so ratio is 560/700 = 0.8. Then detector size of a pixel is 50/1024, so maximum resolution of a voxel is 50/1024*0.8.
Now, voxel resolution "dx=dy=dz" is defined. then define the number of voxels that covers whole images (nx, ny, nz).
Thus, sx = nx*dx, sy=ny*dy, sz=nz*dz;
I modified just for fast calculation. But if you can see Jiang Hsieh's book (second edition,page 96), it is easy to understand.
Thank you for your answers. I have a few questions left.
If I have already got real projections from x-ray imaging system, for example, I got the projections of a mouse, do I need to do the "CTprojection"? Under this circumstance, if I only need to do "CTbackprojection", how to set following params about the reconstructed object:nx, ny, nz, sx, sy, and sz.
The params of the detector are setted as follows:
Last question, would you please give some reference about FDK algorithm not using iso-center(DSO) domain.
1. w = cos(theta), to calculate w, my code considers distances in detectors, as your comment, many books calculate in iso-center (DSO) domain. In this case, uu and vv should be differently calculated by considering distance ratio from source to center and source to detector. Please think about angle between iso-center line and extension line source-voxel-detector.
2. In my code, pi/2 rotate shift is for calculating exactly your comment. "interp2" function results in 90 degree rotated image. So I did "angle" -> "angle-pi/2", and I think DSO+ry is right.
3. If you want to get as real projection data, first set blank scan intensity "b", this is related to "dose", then measurement M = b*exp(-projection(obj)).
Thanks your comment.
I am wondering if I have real projected images, do I need to use the function "CTprojection"?
Good work. I have some questions.
Firstly, in the file of "backprojection.m", you set "w = (param.DSD)./sqrt((param.DSD)^2+uu.^2 + vv.^2);", but in many books, they said that w should be setted as "w = (param.DSO)./sqrt((param.DSO)^2+uu.^2 + vv.^2);". Can you give me some reason?
Secondly, also in the "backprojection.m", I think the description of rx, ry, pu and pv should be:
Would you please give me some suggestion, thank you.
Hi, Masoud Hashemi,
This program is for flat detectors,
I think you should modify for the arc detector geometry,
Thanks for using my software.
I am wondering if the projection function is for Arc detectors (3rd Generation) or Flat detectors?
The codes can run but the result is not very good,can you upload the referrences you refer for the codes,so tha we can understand clearly. Thank you！
This program is similar to other programs but
To accelerate speed,
(1) define the reconstruction object and projection images as a "single" precision,
and then (2) use "matlabpool".
You can use "matlabpool",
for i=1:n (angle)
projection or backprojection
parfor i=1:n (angle)
projection or backprojection
!!please delete figure plot functions
This increases the performance more than twice in quad cores.
Please change the function code "projection_pixel.m":
dist = (SAD+SDD)./sqrt((SAD+SDD)^2 + uu.^2 + vv.^2)*abs(ys(2)-ys(1));
dist = sqrt((SAD+SDD)^2 + uu.^2 + vv.^2)./(SAD+SDD)*abs(ys(2)-ys(1));
Thank you, :)
Examples updated, Matlab built-in function based GPU code updated, Parker weight updated.
Fixed filtering function
Debug and add some descriptions
Update code and debug
I modified codes simply.
I changed title, tag
Fixed a bug
I optimized the source codes
Download apps, toolboxes, and other File Exchange content using Add-On Explorer in MATLAB.