<html>
  <head>
    <meta content="text/html; charset=UTF-8" http-equiv="Content-Type">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    Hi Howard,<br>
    <br>
    I've taken a look at your data.<br>
    You can apply tv denoising on the out.mha volume and obtain a
    significantly lower level of noise without blurring structures by
    using the following command : <br>
    rtktotalvariationdenoising -i out.mha -g 0.001 -o
    tvdenoised/gamma0.001.mha -n 100<br>
    <br>
    I was unable to obtain good results with iterative reconstruction
    from the projection data you sent, though. I think the main reason
    for this is that your projections have much-higher-than-zero
    attenuation in air. Your calculation of i0 when converting from
    intensity to attenuation is probably not good enough. Try to correct
    for this effect first. Then you can start performing SART and
    Conjugate Gradient reconstructions on your data, and once you get
    these right, play with ADMM. <br>
    <br>
    You might need to remove the table from the projections to be able
    to restrict the reconstruction volume strictly to the patient, and
    speed up the computations. We can provide help for that too. <br>
    <br>
    Best regards,<br>
    Cyril<br>
    <br>
    <div class="moz-cite-prefix">On 12/17/2014 05:02 PM, Howard wrote:<br>
    </div>
    <blockquote
cite="mid:CAKr9h5XWf4fdf8TpOVf_1TTeauAmY6qA8Mz_OsUeRpEUivsRVQ@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div>Hi Cyril,</div>
        <div> </div>
        <div>I've sent you two files via <a moz-do-not-send="true"
            href="http://wetransfer.com">wetransfer.com</a>: one is the
          sparse projection set with geometry file and the other is the
          fdk reconstructed image based on full projection set. Please
          let me know if you have trouble receiving them.</div>
        <div> </div>
        <div>Thanks very much for looking into this.</div>
        <div> </div>
        <div>-Howard</div>
      </div>
      <div class="gmail_extra"><br>
        <div class="gmail_quote">On Wed, Dec 17, 2014 at 10:19 AM, Cyril
          Mory <span dir="ltr"><<a moz-do-not-send="true"
              href="mailto:cyril.mory@creatis.insa-lyon.fr"
              target="_blank">cyril.mory@creatis.insa-lyon.fr</a>></span>
          wrote:
          <blockquote class="gmail_quote" style="margin:0 0 0
            .8ex;border-left:1px #ccc solid;padding-left:1ex">
            <div bgcolor="#FFFFFF" text="#000000"> Hi Howard,<br>
              <br>
              Thanks for the detailed feedback.<br>
              The image getting blurry is typically due to a too high
              gamma. Depending on you data, gamma can have to be set to
              a very small value (I use 0.007 in some reconstructions on
              clinical data). Can you send over your volume
              reconstructed from full projection data, and I'll have a
              quick look ? <br>
              <br>
              There is a lot of instinct in the setting of the
              parameters. With time, one gets used to finding a correct
              set of parameters without really knowing how. I can also
              try to reconstruct from your cbct data if you send me the
              projections and the geometry. <br>
              <br>
              Best regards,<br>
              Cyril
              <div>
                <div class="h5"><br>
                  <br>
                  <div>On 12/17/2014 03:49 PM, Howard wrote:<br>
                  </div>
                  <blockquote type="cite">
                    <div dir="ltr">
                      <div>Hi Cyril,</div>
                      <div> </div>
                      <div>Thanks very much for your detailed and nice
                        description on how to use the admmtv
                        reconstruction. I followed your suggestions and
                        re-ran reconstructions using admmtotalvariation
                        and admmwavelets with cbct projection data from
                        a thoracic patient.</div>
                      <div> </div>
                      <div>I am reporting what I found and hope these
                        will give you information for further
                        improvement.</div>
                      <div> </div>
                      <div>1. I repeated admmtotalvariation with 30
                        iterations. No improvement was observed. As a
                        matter of fact, the reconstructed image is
                        getting a lot noiser compared to that using 3
                        iterations. The contrast is getting worse as
                        well. I tried to play around with window &
                        level in case I was fooled but apparently more
                        iterations gave worse results.</div>
                      <div> </div>
                      <div>2. Similarly I ran 30 iterations using
                        admmwavelets. Slightly better reconstruction
                        compared with total variation.</div>
                      <div> </div>
                      <div>3. Then I went ahead to test if TV benefits
                        us anything using the tvdenoising application on
                        the fdk-reconstructed image reconstructed
                        from full projection set. I found that the more
                        iterations, the more blurry the image became.
                        For example, with 50 iterations the contrast on
                        the denoised image is very low so that the
                        vertebrae and surrounding soft tissue are hardly
                        distinguishable. Changing gamma's at 0.2, 0.5,
                        1.0, 10  did not seem to make a difference on
                        the image. With 5 iterations the denoising seems
                        to work fairly well. Again, changing gamma's
                        didn't make a difference.</div>
                      <div>I hope I didn't misused the
                        totalvariationdenoising application. The command
                        I executed was: rtktotalvariationdenoising -i
                        out.mha -o out_denoising_n50_gamma05 --gamma 0.5
                        -n 50</div>
                      <div> </div>
                      <div>In summary, tdmmwavelets seems perform better
                        than tdmmtotalvariation but neither gave
                        satisfactory results. No sure what we can infer
                        from the TV denoising study. I could send my
                        study to you if there is a need. Please let me
                        know what tests I could run. Further help on
                        improvement is definitely welcome and
                        appreciated.</div>
                      <div> </div>
                      <div>-Howard</div>
                    </div>
                    <div class="gmail_extra"><br>
                      <div class="gmail_quote">On Mon, Dec 15, 2014 at
                        4:07 AM, Cyril Mory <span dir="ltr"><<a
                            moz-do-not-send="true"
                            href="mailto:cyril.mory@creatis.insa-lyon.fr"
                            target="_blank">cyril.mory@creatis.insa-lyon.fr</a>></span>
                        wrote:
                        <blockquote class="gmail_quote"
                          style="margin:0px 0px 0px
0.8ex;padding-left:1ex;border-left-color:rgb(204,204,204);border-left-width:1px;border-left-style:solid">
                          <div bgcolor="#FFFFFF" text="#000000"> Hello
                            Howard,<br>
                            <br>
                            Good to hear that you're using RTK :)<br>
                            I'll try to answer all your questions, and
                            give you some advice:<br>
                            - In general, you can expect some
                            improvement over rtkfdk, but not a huge one<br>
                            - You can find the calculations in my PhD
                            thesis <a moz-do-not-send="true"
                              href="https://tel.archives-ouvertes.fr/tel-00985728"
                              target="_blank">https://tel.archives-ouvertes.fr/tel-00985728</a>
                            (in English. Only the introduction is in
                            French)<br>
                            - Adjusting the parameters is, in itself, a
                            research topic (sorry !). Alpha controls the
                            amount of regularization and only that (the
                            higher, the more regularization). Beta,
                            theoretically, should only change the
                            convergence speed, provided you do an
                            infinite number of iterations (I know it
                            doesn't help, sorry again !). In practice,
                            beta is ubiquitous and appears everywhere in
                            the calculations, therefore it is hard to
                            predict what effect an increase/decrease of
                            beta will give on the images. I would keep
                            it as is, and play on alpha<br>
                            - 3 iterations is way too little. I
                            typically used 30 iterations. Using the CUDA
                            forward and back projectors helped a lot
                            maintain the computation time manageable<br>
                            - The quality of the results depends a lot
                            on the nature of the image you are trying to
                            reconstruct. In a nutshell, the algorithm
                            assumes that the image you are
                            reconstructing has a certain form of
                            regularity, and discards the potential
                            solutions that do not have it. This
                            assumption partly compensates for the lack
                            of data. ADMM TV assumes that the image you
                            are reconstructing is piecewise constant,
                            i.e. has large uniform areas separated by
                            sharp borders. If your image is a phantom,
                            it should give good results. If it is a real
                            patient, you should probably change to
                            another algorithm that assumes another form
                            of regularity in the images (try
                            rtkadmmwavelets)<br>
                            - You can find out whether you typical
                            images can benefit from TV regularization by
                            reconstructing from all projections with
                            rtkfdk, then applying
                            rtktotalvariationdenoising on the
                            reconstructed volume (try 50 iterations and
                            adjust the gamma parameter: high gamma means
                            high regularization). If this denoising
                            implies an unacceptable loss of quality,
                            stay away from TV for these images, and try
                            wavelets<br>
                            <br>
                            I hope this helps<br>
                            <br>
                            Looking forward to reading you again,<br>
                            Cyril
                            <div>
                              <div><br>
                                <br>
                                <div>On 12/12/2014 06:42 PM, Howard
                                  wrote:<br>
                                </div>
                              </div>
                            </div>
                            <blockquote type="cite">
                              <div>
                                <div>
                                  <div dir="ltr">
                                    <div>I am testing the ADMM total
                                      variation reconstruction with
                                      sparse data sample. I could
                                      reconstruct but the results were
                                      not as good as expected. In other
                                      words, it didn't show much
                                      improvement compared to fdk
                                      reconstruction using the same
                                      sparse projection data. </div>
                                    <div> </div>
                                    <div>The parameters I used in ADMMTV
                                      were the following:</div>
                                    <div> </div>
                                    <div>--spacing 2,2,2 --dimension
                                      250,100,250 --alpha 1 --beta 1000
                                      -n 3</div>
                                    <div> </div>
                                    <div>while the fdk reconstruction
                                      parameters are:</div>
                                    <div> </div>
                                    <div>--spacing 2,2,2 --dimension
                                      250,100,250 --pad 0.1 --hann 0.5</div>
                                    <div> </div>
                                    <div>The dimensions were chosen to
                                      include the entire anatomy. 72
                                      projections were selected out of
                                      646 projections for a 360 degree
                                      scan for both calculations.</div>
                                    <div> </div>
                                    <div>What parameters and how can I
                                      adjust (like alpha, beta, or
                                      iterations?) to improve the ADMMTV
                                      reconstruction? There is not much
                                      description of this application
                                      from the wiki page.</div>
                                    <div> </div>
                                    <div>Thanks,</div>
                                    <div> </div>
                                    <div>-howard</div>
                                    <div> </div>
                                  </div>
                                  <br>
                                  <fieldset></fieldset>
                                  <br>
                                </div>
                              </div>
                              <pre>_______________________________________________
Rtk-users mailing list
<a moz-do-not-send="true" href="mailto:Rtk-users@public.kitware.com" target="_blank">Rtk-users@public.kitware.com</a>
<a moz-do-not-send="true" href="http://public.kitware.com/mailman/listinfo/rtk-users" target="_blank">http://public.kitware.com/mailman/listinfo/rtk-users</a><span><font color="#888888">
</font></span></pre>
                              <span><font color="#888888"> </font></span></blockquote>
                            <span><font color="#888888"> <br>
                                <pre cols="72">-- 
--
Cyril Mory, Post-doc
CREATIS
Leon Berard cancer treatment center
28 rue Laënnec
69373 Lyon cedex 08 FRANCE

Mobile: <a moz-do-not-send="true" href="tel:%2B33%206%2069%2046%2073%2079" target="_blank" value="+33669467379">+33 6 69 46 73 79</a></pre>
                              </font></span></div>
                        </blockquote>
                      </div>
                    </div>
                  </blockquote>
                  <br>
                  <pre cols="72">-- 
--
Cyril Mory, Post-doc
CREATIS
Leon Berard cancer treatment center
28 rue Laënnec
69373 Lyon cedex 08 FRANCE

Mobile: <a moz-do-not-send="true" href="tel:%2B33%206%2069%2046%2073%2079" target="_blank" value="+33669467379">+33 6 69 46 73 79</a></pre>
                </div>
              </div>
            </div>
          </blockquote>
        </div>
      </div>
    </blockquote>
    <br>
    <pre class="moz-signature" cols="72">-- 
--
Cyril Mory, Post-doc
CREATIS
Leon Berard cancer treatment center
28 rue Laënnec
69373 Lyon cedex 08 FRANCE

Mobile: +33 6 69 46 73 79</pre>
  </body>
</html>