Re: segmentation of images with poor noise statistics

On 1 Jun., 19:12, "Pixel.to.life" <pixel.to.l...@xxxxxxxxx> wrote:
On May 31, 3:53 pm, goanicks <spam...@xxxxxxxx> wrote:





Hi there,

I am looking for a clever approach to perform an image segmentation in
SPECT images with typically poor noise porperties. According to the
present (Poisson) noise, simple first-order-methods, such as edge
detection filtering, or simply applying a threshold usually fail.
Browsing the WEB I found, that there are so called statistical region
based methods based on a 'snake', which also take the noise properties
of the image into account and which are (from the authors) reported to
be very stable and reliable ...

Now my questions:

- Is this the method of choice for the type of images I
want to process?
- Are there any free code sources or even programs which
can be used to try out this approach?

Thanks a lot in advance!

Regards
Andy

Hi, Andy,

By nature, SPECT images will have low SNR, low spatial resolution, and
sometimes low dynamic range too. That would make traditional intensity
based methods not very useful as you would notice: you cant
effectively estimate gradients and separate regions easily (see
example images here:http://www.alasbimnjournal.cl/revistas/15/images/carreira2.jpg
).

In my experience, when information content is low, or corrupted,
additional information needs to be used in conjunction with the
intensity to achieve a practical analysis result. I have seen
geometric methods being successfully employed in the pursuit (active
contours, level-sets e.g.)

If you wish to work in 2D images, you may start with active contours.
But ultimately I would recommend a level-sets based approach (fast
marching) owing to its robustness to topology changes when evolving an
interface.

Here is anexample work:

http://ipg.zesoi.fer.hr/papers/mi01spect.pdf

You may find on page 4 how the authors deal with noise and fuzziness.
If you imagine a geometric contour evolving towards the boundary of
the desired object, you can employ additional constraints other than
just intensity value, graidient etc. You can adjust the speed function
based on curvature to avoid leaking into areas with fuzzy boundary.

Here is another post with similar discussion, although image modality
being different:

http://groups.google.com/group/medicalimagingscience/browse_thread/th...

Do keep us posted with the results...

Pixel.To.Life
[http://groups.google.com/group/medicalimagingscience?hl=en]- Zitierten Text ausblenden -

- Zitierten Text anzeigen -

Hi,

thanks that you took your time to send me such a comprehensive reply.
Although I am not (yet ;-) an expert in this field, I fully agree with
your statement about the applicability of standard (intensity based)
methods.

I already started with an active contour approach, but I am not yet
happy with the achieved result. According to what I saw so far, it
works only with konvex objects which is not appropriate in my case.
However, I don't think this is a normal behaviour, but rather a result
of the chose contour movement rules and segmentation quality criteria
I have chosen. So, there is clearly room for further improvement.

Maybe I get some more ideas from reading the paper and having a look
at the earlier postings you mentioned.

Thanks again

Andy

.

Relevant Pages