[Insight-users] Any advice on which adv. segmentation filter to use here?

[Dan Mueller]

Hi Roman,

2009/11/4 Roman Grothausmann <roman.grothausmann at helmholtz-berlin.de>:
> Dan Mueller wrote:
>>
>> Hi Roman,
>>
>> Not sure if you're still searching for ideas...
>>
>> If you are, then you might consider investigating a simple binary
>> threshold followed by some basic cleaning operations.
>>
>> From your description of the problem I'm not 100% sure what you are
>> trying to achieve (I don't know what a tomogram is, or carbon
>> support), but if I've guessed correctly, the attached image may be
>> fairly close to your desired output. It was generated as follows:
>>    1. binary threshold (lower = 67, upper = 255)
>>    2. morphological close (structuring element = circle, radius = 30)
>>    3. voting binary iterative hole filling (radius = 12, majority
>> threshold = 25)
>>
>> Of course you can tweak the cleaning operations as required. Some of
>> the new label filters in the Review folder may also be useful for
>> cleaning the segmentation.
>
> Many thanks for Your ideas and Your help Dan.
> It seems I was too careful with my filtering. I didn't go above a radius of
> 10. Also I didn't expect this binary voting filter in connection with a hole
> filling.
> Would You recommend the iterative filter or the one that is only applied
> once

I used the iterative version (I think I had max 50 iterations, but
only 10 or so were used).

> Also I wonder if it makes sense to do a morph. closing then a label filter
> another morph. closing and/or binary voting.

You might consider trying some systematic experiments to compare
different parameters/pipelines.

> In the end I would like to close the "pores" with the least modification of
> the already existing "outside surface" (some kind of local convex hull
> surface) but it seems to be a long way to get there.

I guess by "pores" you mean the remaining hole in the structure on the
left? If so, you may consider performing a connect component analysis,
then for any components/labels which touch only one other
(non-background) label, join it with the other label. Not sure if
there is an existing filter to achieve this...

> Thanks again for Your help.
> Roman