[MINC-users] classify: bg and csf

[Alex Zijdenbos]

Hi Pedro,

It looks like the background tags are not used at all. I think the
immediate solution for you would be to add -mask_classified to the
classify_clean invocation; that will force '0' everywhere outside the mask.

Although I don't recall at the moment, this was probably a design decision
somewhere along the way; if you train a classifier on background, you are
actually likely to introduce 'holes' in the classified image. The current
approach, which is to only assign '0' to voxels outside the mask,
essentially forces all voxels inside the mask to end up in one of the 3
tissue classes.

Clearly classify_clean could use some updating and clarifications. Also, on
your image, the tag cleaning step (if you add -clean_tags) actually
crashes, which it also should not do.

-- A

On Mon, Oct 10, 2016 at 10:29 AM, Pedro <ptcougopinto at gmail.com> wrote:

> Transformed tags seem fine. I will send you the files.
> Thanks
> Pedro
>
> > Em 10 de out de 2016, à(s) 11:18, Alex Zijdenbos <zijdenbos at gmail.com>
> escreveu:
> >
> > Hi Pedro,
> >
> > Hrm - you are right - so that is a bug in the documentation:
> > 'classify_clean -help' clearly states "non-linear transformation to map
> the
> > tags from stereotaxic space to subject", which is wrong in a few ways
> (the
> > xfm is in fact inverted, and it doesn't need to be a non-linear
> > transformation either).
> >
> > I would suggest you double-check (using the log file and -tmpdir and
> > -keeptmp) whether the tags that are provided to classify itself, are
> still
> > good (contain all classes, and actually match the image). If that doesn't
> > clarify it, send the mnc and xfm files and I can take a look.
> >
> > -- A
> >
> > On Mon, Oct 10, 2016 at 9:57 AM, Pedro <ptcougopinto at gmail.com> wrote:
> >
> >> I had understood that the .xfm file was from native to ICBM. I think
> this
> >> transformation is inverted before transforming the tags. Quoting a prior
> >> email thread:
> >>
> >> <image-to-tags.xfm> is a tranform from your native data space
> >> (input.mnc) to the space of the tags. If you are using the ones in the
> >> standard package this will be a transform from native space to the
> >> icbm152 model.
> >>
> >>> Em 10 de out de 2016, à(s) 10:45, Alex Zijdenbos <zijdenbos at gmail.com>
> >> escreveu:
> >>>
> >>> Hi Pedro,
> >>>
> >>> Is it possible that your transform is mapping the tags into oblivion?
> The
> >>> standard set of tags are in ICBM stx space, and the argument to
> >>> -tag_transform should be an xfm that maps those standard tags to the
> >>> subject image. Just from the file name of the xfm you are using, I'd
> say
> >>> you may need to invert it first?
> >>>
> >>> -- A
> >>>
> >>> On Mon, Oct 10, 2016 at 9:06 AM, Pedro <ptcougopinto at gmail.com> wrote:
> >>>
> >>>> Hi Alex
> >>>>
> >>>> $ classify_clean -tag_transform t1_2talnlin.xfm -mask mask.mnc
> >>>> -mask_source t1.mnc t1_class.mnc
> >>>>
> >>>> Shouldn’t classify_clean be using background tags from
> >>>> /opt/minc/share/classify/ntags_1000_bg.tag by default?
> >>>>
> >>>>
> >>>>> Em 7 de out de 2016, à(s) 22:27, Alex Zijdenbos <zijdenbos at gmail.com
> >
> >>>> escreveu:
> >>>>>
> >>>>> Hi Pedro,
> >>>>>
> >>>>> My guess would be that you are not providing classify with (a)
> >> sample(s)
> >>>> of 'bg'. Exactly how are you invoking it?
> >>>>>
> >>>>> -- A
> >>>>>
> >>>>>
> >>>>> Sent from mobile
> >>>>>
> >>>>>> On Oct 7, 2016, at 12:51, Pedro <ptcougopinto at gmail.com> wrote:
> >>>>>>
> >>>>>> Hi,
> >>>>>>
> >>>>>> I´m getting background and CSF to have the same label value with
> >>>> classify. Is it possible to avoid this? I tried using -mask
> -mask_source
> >>>> brain_mask.mnc, still with background labels as ‘1’.
> >>>>>>
> >>>>>> Thanks
> >>>>>> Pedro
> >>>>>> _______________________________________________
> >>>>>> MINC-users at bic.mni.mcgill.ca
> >>>>>> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
> >>>>> _______________________________________________
> >>>>> MINC-users at bic.mni.mcgill.ca
> >>>>> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
> >>>>
> >>>> _______________________________________________
> >>>> MINC-users at bic.mni.mcgill.ca
> >>>> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
> >>>>
> >>> _______________________________________________
> >>> MINC-users at bic.mni.mcgill.ca
> >>> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
> >>
> >> _______________________________________________
> >> MINC-users at bic.mni.mcgill.ca
> >> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
> >>
> > _______________________________________________
> > MINC-users at bic.mni.mcgill.ca
> > http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
>
> _______________________________________________
> MINC-users at bic.mni.mcgill.ca
> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-users
>