[Insight-users] Estimation of abdominal volume

[Frederic Perez]

Hello Luis and Insight-users,

> 1) Why did you decided to use CT for this measurement ?
> 
>     There are certain MRI protocols that will highlight
>     fat and therefore help to mark the peritoneum...
> 
>     This may simplify your image processing task, not to
>     mention avoid a significant amount of ionizing radiation
>     to your patient.
> 
>     To quote Dr. Michael Vannier
> 
>         "Good imaging beats good image processing"
> 
>      When segmentation becomes too difficult, you should
>      take that as an indication that the image modality,
>      and/or the imaging protocol are not well suited for
>      this task.

The reason to use the CT studies is simply because they are
already available---their primary use was not to measure 
the abdominal volume but to quantify gas in the abdominal 
track. I will ask my employers for the chances of using MRI
in this facility. (By the way, I like Dr. Michael Vannier's
quote---maybe I will paste it to my web page.)

> 2) It seems to be reasonable to expect that the volume
>     of the abdominal cavity will be extremely variable during
>     the day.  For a longitudinal analysis to make sense you
>     have to solve the very challenging task of establishing
>     conditions where the images acquired at day #1 can be
>     considered similar to those of day #N.  Those daily
>     variations are probably larger than the actual accuracy
>     of a segmentation method.

Another item to discuss with the doctors.

> 3) This problem seems to be a good candidate for an atlas
>     based registration approach.  In this context you could
>     do the following:
> 
>     3.1) Invest a large amount of effort in obtaining a
>          segmentation of the abdominal cavity from an
>          initial image of the patient. This may involve
>          a mixture of supervised segmentation algorithms
>          and a retouching/evaluation stage.
> 
>     3.2) Once you have segmented the initial dataset, you
>          could perform deformable registration between this
>          first original dataset and the new ones, and use
>          the resulting deformation field for mapping the
>          initial segmentation. In this way, the deformation
>          field itself will tell you the changes on the volume
>          in the cavity. In fact, by computing the Jacobian
>          of the deformation field you can estimate local
>          expansions and contractions. (at least in theory)   :)

Thank you very much for the advice (and from other advices 
you gave in the mailing list to other users and were 
particularly helpful too)!
Time for me to (actually) learn about registrations and 
atlases (so far I have been mostly focused on segmentations)...

Best regards,

Frederic Perez


> 
> 
> Regards,
> 
> 
>      Luis
> 
> 
> 
> --------------------------
> Frederic Perez wrote:
> > Dear Insight-users,
> > 
> > being stalled at the moment, I thought I better ask you for hints
> or 
> > advice. Any help will be greatly appreciated. Our problem follows:
> > 
> > [Input]
> > We have a set of 3D CT images, obtained using the same scanner,
> from
> > a set of patients ($i = 1..N$). For each patient $i$ two images $I$
> 
> > are derived, taken at different days, to assess the response to 
> > therapy: $I(i, t_0)$ and $I(i, t_1)$.
> > 
> > [Target]
> > The objective is to measure the difference, per patient, of the 
> > abdominal volume ($V_a(I(i, t_1)) - V_a(I(i, t_0))$).
> > 
> > [How-to?]
> > I guess that the segmentation of the true abdominal cavity is a
> very
> > difficult problem (small narrowness of the peritoneum, i.e. of the 
> > membrane lining the abdominal cavity, the hole due to the
> esophagus,
> > etc.). Hopefully, however, an approximation will suffice.
> > So far, for each image we have segmented most of the trunk by means
> 
> > of a  pipeline of filters (ExtractImageFilter, MaskImageFilter,
> etc.):
> > $T(I(i, t_0))$ and $T(I(i, t_1))$.
> > We can compute volumes with a ImageMomentsCalculator object, but
> only
> > when provided with consistent input we will obtain meaningful
> results.
> > This means that we must crop each pair $T(I(i, t_0))$ and
> > $T(I(i, t_1))$ in such a way that the limiting planes are tightly 
> > related. The question is, how do we define those limiting planes?
> > We thought about using anatomical landmarks, automatically defined,
> > but do not how to do this (maybe with landmarks on the
> bones?---which
> > ones?). Is this the right direction to pursue?
> > 
> > May thanks in advance,
> > 
> > Frederic Perez



		
______________________________________________
Renovamos el Correo Yahoo!: ¡100 MB GRATIS!
Nuevos servicios, más seguridad
http://correo.yahoo.es