Algorithm Questions

[ubersexualShupeng]

Dear all,

I working on a project need to evaluate the PVE on the FDG tumor imaging during the treatment process. Is there are any PVC algorithm in the PETPVC toolbox that can be done without input (tumor volume) mask? Thanks.

[bathomas]

Hi @ubersexualShupengmailto:notifications@github.com,

You have two deconvolution-based options in the toolkit: Van-cittert (pvc_vc) and Richardson-Lucy (pvc_rl). Keep in mind that these techniques tend to amplify noise.

Kind regards,

Ben

Get Outlook for Androidhttps://aka.ms/ghei36

---

From: ubersexualShupeng notifications@github.com Sent: Wednesday, June 28, 2017 7:57:04 PM To: UCL/PETPVC Cc: Subscribed Subject: [UCL/PETPVC] Algorithm Questions (#21)

Dear all,

I working on a project need to evaluate the PVE on the FDG tumor imaging during the treatment process. Is there are any PVC algorithm in the PETPVC toolbox that can be done without input (tumor volume) mask? Thanks.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/UCL/PETPVC/issues/21, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AGX4aSR32RzgD_IQkNxRsPN3Y9uXvKKvks5sIqIAgaJpZM4OIYiJ.

[ubersexualShupeng]

It works. Thank you very much!

Shupeng

[ubersexualShupeng]

Hi @bathomas I still some questions regarding the PVC problem: 1.Is there are any algorithm (in the toolkit) can use paired CT as additional information to perform PVC? (We have CT & PET images obtained almost simultaneously) 2.What is the parameter -x, -y & -z means? what size should be most reasonable? (Our PET images resolution is 444 mm^3 obtained using the PET/CT Philips GEMINI TF Big Bore Scanner )

[bathomas]

Dear @ubersexualShupengmailto:notifications@github.com,

1) At present, you need to create a mask based on the CT data and pass this to the applications in the toolkit. Currently, the toolkit cannot automatically use the CT data directly. It probably never will do this as it would extremely difficult to know what a user wanted to PV-correct based on the data alone.

2) The –x, -y and –z parameters are the full-width at half maximum (FWHM) of the scanner point-spread function (PSF) in each direction. Note that this is not the voxel size. The PSF is dependent on several factors: scanner geometry, isotope and the reconstruction… . Therefore, I cannot recommend a value for you to use. Ideally, you should measure this on the scanner itself, using the same reconstruction as the PET images that you want to PV-correct.

Kind regards,

Ben

From: ubersexualShupeng notifications@github.com Reply-To: UCL/PETPVC reply@reply.github.com Date: Thursday, 29 June 2017 at 15:16 To: UCL/PETPVC PETPVC@noreply.github.com Cc: "Thomas, Ben" b.a.thomas@ucl.ac.uk, Mention mention@noreply.github.com Subject: Re: [UCL/PETPVC] Algorithm Questions (#21)

Hi @bathomashttps://github.com/bathomas I still some questions regarding the PVC problem: 1.Is there are any algorithm (in the toolkit) can use paired CT as additional information to perform PVC? (We have CT & PET images obtained almost simultaneously) 2.What is the parameter -x, -y & -z means? what size should be most reasonable? (Our PET images resolution is 444 mm^3 obtained using the PET/CT Philips GEMINI TF Big Bore Scanner )

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/UCL/PETPVC/issues/21#issuecomment-311980226, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AGX4ackwRUpuclO0YW9daV3bSyXmlESTks5sI7G5gaJpZM4OIYiJ.

[ubersexualShupeng]

Thank you! @bathomas . I do appreciate it.

[ubersexualShupeng]

Hi @bathomas , Any suggestions on selection of the parameter 'iteration number'? Thanks.

Shupeng

[bathomas]

Hi,

Not really. It's a bias vs. noise problem. You need to determine what you feel is appropriate for your data. The deconvolution approaches amplify noise and it tends to be necessary to terminate them prematurely. As for Van-Cittert, you have to look at both the step size (alpha) and iteration number. For example, you might pick a small alpha and large number of iterations or conversely, a large alpha and fewer interations. All I can really suggest is that you test different parameter values for a set of representative images and decide what is best for your data and application.

[ubersexualShupeng]

Thanks @bathomas,

Understand the optimal parameters could be task specific. Are there any quantitative matrix can be used to describe the 'bias' and 'noise' here? Thanks.

Shupeng

[bathomas]

Probably best to start with realistic phantom (well as close to reality as you can get with a phantom) data where you have known activity concentrations. Then perform a parameter sweep and plot the measured activity as a proportion of the true activity on one axis, and variance within the ROI (or maybe contrast) on the other axis. Look at the curves and try the chosen parameters on the real data.

[ubersexualShupeng]

Thanks @bathomas,

Understand that using phantom would be useful to deal with the trade-off between bias and noise. Why the iterative de-convolution based correction method increase image noise? Could you provide some theoretical explanation? or some references? Thanks.

Shupeng