3D interpolation/hotspot removal

[Thomson Neil (EAST KENT HOSPITALS UNIVERSITY NHS FOUNDATION TRUST)]

Dear list

In brief: I would be interested (and grateful) to hear other's perspectives on good (3D) interpolation methods to "fill in the blanks" left by hotspot removal. Cheers.

The long version:
A colleague and I are coding some software to remove "hotspots" in 3 dimensions from tomographic ventilation scintigrams (3D images of radioactive aerosol deposition within the lungs). In certain cases, the aerosol is trapped in clumps and it is desirable to remove this clumping for analysis.

We have removed hotspots using a published method (no reference to hand), which involves subtracting the median value (med_old) of the data set from each pixel to create a new data set. This is used to calculate a new median value (med_new). A mask is then produced by calculating (pixel_value_old - med_old)/med_new. If this is greater than some number (set by trial and error, but published data says 6), it is in the mask and we remove it from the original data.

The tricky step is to "fill in the blanks." Our initial trial is a basic linear interpolation from just one pixel out from the mask. This gives a noisy fit and also still a bit "hot." The hotness could be reduced by increasing the size of the mask, but this can affect normal areas elsewhere within the lung volume.

Naturally, I have uploaded a composite image<http://www.fishlegs.co.uk/10b5030k-1.tif> onto the web, which depicts a single slice in four views, one of the original data set (top left), interpolated (top-right), the mask used to remove the hotspots (bottom-left) and the original data with hotspots removed (bottom-right).

Once again, I'd be really grateful to hear what you have to say!

Thanks, Neil

Neil Thomson,
Nuclear Medicine Physics Section,
Medical Physics,
Kent and Canterbury Hospital,
UK. CT1 3NG
[phone]

********************************************************************************************************************

This message may contain confidential information. If you are not the intended recipient please inform the
sender that you have received the message in error before deleting it.
Please do not disclose, copy or distribute information in this e-mail or take any action in reliance on its contents:
to do so is strictly prohibited and may be unlawful.

Thank you for your co-operation.

NHSmail is the secure email and directory service available for all NHS staff in England and Scotland
NHSmail is approved for exchanging patient data and other sensitive information with NHSmail and GSi recipients
NHSmail provides an email address for your career in the NHS and can be accessed anywhere
For more information and to find out how you can switch, visit www.connectingforhealth.nhs.uk/nhsmail

********************************************************************************************************************

[Gabriel Landini]

On Tuesday 26 Apr 2011, Thomson Neil wrote:

> The tricky step is to "fill in the blanks." Our initial trial is a basic
> linear interpolation from just one pixel out from the mask. This gives a
> noisy fit and also still a bit "hot." The hotness could be reduced by
> increasing the size of the mask, but this can affect normal areas
> elsewhere within the lung volume.

Is this similar to "inpainting" the saturated part? If so, you might want to
look laplacian diffusion.
If implemented naively (the only way I seem to be able to do it :-) ), it
converges very slowly but there must be some clever speed up (apparently
Matlab has a roifill which does something similar).

I seem to remember that somebody holds a patent (is it MS?) on this :-/

Here is a thread where we discussed this in the past:
http://imagej.588099.n2.nabble.com/Filling-a-roi-using-Laplace-diffusion-
equation-td2650221.html

I hope it is helpful

Gabriel

[dscho]

Hi,

On Tue, 26 Apr 2011, Gabriel Landini wrote:

> On Tuesday 26 Apr 2011, Thomson Neil wrote:
>
> > The tricky step is to "fill in the blanks." Our initial trial is a
> > basic linear interpolation from just one pixel out from the mask. This
> > gives a noisy fit and also still a bit "hot." The hotness could be
> > reduced by increasing the size of the mask, but this can affect normal
> > areas elsewhere within the lung volume.
>
> Is this similar to "inpainting" the saturated part? If so, you might
> want to look laplacian diffusion.

Note: inpainting is pretending that there is more information than you
actually recorded. Yes, it looks nicer, but the temptation to analyze the
data after inpainting must be resisted! Oh, and inpainting can be
detected. So there are even more good reasons to resist the temptations.

Of course, if you want to produce a piece of art, inpainting is quite a
cool technique.

Ciao,
Johannes

[Gabriel Landini]

On Wednesday 27 Apr 2011, Johannes Schindelin wrote:

> Note: inpainting is pretending that there is more information than you
> actually recorded. Yes, it looks nicer, but the temptation to analyze the
> data after inpainting must be resisted! Oh, and inpainting can be
> detected. So there are even more good reasons to resist the temptations.

Johannes is right, analysing the filled image part would be very silly.
I have no idea why this inpainting was needed, but I presume that the ROI -as
it is known- would be avoided in the analysis.

I have used a slow version of iteratively averaging a ROI for retouching an
old photo which had creases and cracks.
This does not produce good results when edges cross the area to be repainted
or the region is very textured. But it does look much better than filling with
a flat colour and yes, it is easily detectable because the result texture is
very smooth.

Albert had one project idea to use a wavelet method to reconstruct lost or
undesirable data:
http://pacific.mpi-cbg.de/wiki/index.php/GSoC_2011_Ideas
This sounds quite interesting, is there anybody working on this?

Cheers

Gabriel

[Thomson Neil (EAST KENT HOSPITALS UNIVERSITY NHS FOUNDATION TRUST)]

Thanks for the responses.

I'm fully aware of the "gotchas" with in-filling. The entrie stack is going to be analyses whether in-filling is performed or not. Some in-filling has to be performed otherwise there are too many empty pixels and this is distracting during qualitative interpretation - a smoothed segment gives better results for this reason. The images we are working with are typically small (128*128*128) and speed is not an issue as we only have a couple to look at per day.

I'm now racing my colleague; I'm coding for gaussian blur and he's coding for 9-point smooth!

Thanks again, Neil

Neil Thomson,
Nuclear Medicine Physics Section,
Medical Physics,
Kent and Canterbury Hospital,
UK. CT1 3NG
[phone]
________________________________________
From: ImageJ Interest Group [[hidden email]] On Behalf Of Gabriel Landini [[hidden email]]
Sent: 27 April 2011 15:34
To: [hidden email]
Subject: Re: 3D interpolation/hotspot removal

On Wednesday 27 Apr 2011, Johannes Schindelin wrote:

> Note: inpainting is pretending that there is more information than you
> actually recorded. Yes, it looks nicer, but the temptation to analyze the
> data after inpainting must be resisted! Oh, and inpainting can be
> detected. So there are even more good reasons to resist the temptations.

Johannes is right, analysing the filled image part would be very silly.
I have no idea why this inpainting was needed, but I presume that the ROI -as
it is known- would be avoided in the analysis.

I have used a slow version of iteratively averaging a ROI for retouching an
old photo which had creases and cracks.
This does not produce good results when edges cross the area to be repainted
or the region is very textured. But it does look much better than filling with
a flat colour and yes, it is easily detectable because the result texture is
very smooth.

Albert had one project idea to use a wavelet method to reconstruct lost or
undesirable data:
http://pacific.mpi-cbg.de/wiki/index.php/GSoC_2011_Ideas
This sounds quite interesting, is there anybody working on this?

Cheers

Gabriel

********************************************************************************************************************

This message may contain confidential information. If you are not the intended recipient please inform the
sender that you have received the message in error before deleting it.
Please do not disclose, copy or distribute information in this e-mail or take any action in reliance on its contents:
to do so is strictly prohibited and may be unlawful.

Thank you for your co-operation.

NHSmail is the secure email and directory service available for all NHS staff in England and Scotland
NHSmail is approved for exchanging patient data and other sensitive information with NHSmail and GSi recipients
NHSmail provides an email address for your career in the NHS and can be accessed anywhere
For more information and to find out how you can switch, visit www.connectingforhealth.nhs.uk/nhsmail

********************************************************************************************************************

[dscho]

Dear Neil,

On Wed, 27 Apr 2011, Thomson Neil (EAST KENT HOSPITALS UNIVERSITY NHS FOUNDATION              TRUST) wrote:

> The entrie stack is going to be analyses whether in-filling is performed
> or not. Some in-filling has to be performed otherwise there are too many
> empty pixels and this is distracting during qualitative interpretation -
> a smoothed segment gives better results for this reason.

Just make sure that you do not analyze ghosts. And I would suggest to
refrain from publishing papers containing such analyses.

Ciao,
Johannes

http://www.uab.edu/researchintegrityandimages/guidelines/guideline_6.html

[Cammer, Michael]

Could this problem be rectified by using a median filter and clearly stating "median filtering N X N neighborhood used".
-Michael C.

-----Original Message-----
From: ImageJ Interest Group [mailto:[hidden email]] On Behalf Of Johannes Schindelin
Sent: Wednesday, April 27, 2011 11:27 AM
To: [hidden email]
Subject: Re: 3D interpolation/hotspot removal

Dear Neil,

On Wed, 27 Apr 2011, Thomson Neil (EAST KENT HOSPITALS UNIVERSITY NHS FOUNDATION              TRUST) wrote:

> The entrie stack is going to be analyses whether in-filling is performed
> or not. Some in-filling has to be performed otherwise there are too many
> empty pixels and this is distracting during qualitative interpretation -
> a smoothed segment gives better results for this reason.

Just make sure that you do not analyze ghosts. And I would suggest to
refrain from publishing papers containing such analyses.

Ciao,
Johannes

http://www.uab.edu/researchintegrityandimages/guidelines/guideline_6.html

------------------------------------------------------------
This email message, including any attachments, is for the sole use of the intended recipient(s) and may contain information that is proprietary, confidential, and exempt from disclosure under applicable law. Any unauthorized review, use, disclosure, or distribution is prohibited. If you have received this email in error please notify the sender by return email and delete the original message. Please note, the recipient should check this email and any attachments for the presence of viruses. The organization accepts no liability for any damage caused by any virus transmitted by this email.
=================================

[dscho]

Hi,

On Wed, 27 Apr 2011, Cammer, Michael wrote:

> Could this problem be rectified by using a median filter and clearly
> stating "median filtering N X N neighborhood used".

This would be clearly within the appropriate bounds, in my humble opinion.

Although the quantitative analysis should still be made on the original
data, since the Median filter is a non-linear filter. (But expert opinions
differ on the question whether quantitative analyses on median-filtered
images makes sense or not. I can give you a couple of arguments for 'pro'
as well as for 'contra' :-)

Ciao,
Johannes