Re: Super resolution reconstruction for medical images?

On Apr 16, 5:43 pm, "pixel.to.life" <pixel.to.l...@xxxxxxxxx> wrote:
On Apr 16, 2:17 am, "Martin Brown" <|||newspam...@xxxxxxxxxxxxxxxxxx>
wrote:

You are between a rock and a hard place when indirect imaging
techniques like MRI and tomography are used. The measurement data are
in effect the Fourier transform of the object being investigated. The
vital image that the clinician needs is already the result of a very
sophisticated regridding and FFT computation with inherent weaknesses.
Tomography and PET is much worse but can do things not otherwise
available.

The classical linear pseudo inverse is well understood and has failry
well known properties. But in order to compute it certain
approximations have to be made to compensate for the abrupt cut offf
of data in the frequency domain.

Non-linear methods like maximum entropy (and other regularised
methods) can infer the values of unmeasured frequency components by
imposing a positivity constraint on the reconstructed image. The
disadvantage is that you end up with a resolution that is strongly
dependent on the local signal to noise ratio.

Thanks for a very informative explanation. This gives me more doors
(unopened as yet) to explore.

Are you aware of any recent (published) work experimenting with super-
resolution reconstruction not during/after imaging; but at a later
stage (on a viewing workstation when exploring the data, for
visualization e.g.)?

I would be extremely wary of any attempt to interpolate with super-
resolution at the final viewing stage. It is one thing to build a
model with a higher resolution than a conventional pseudo inverse that
is consistent with all the observed measurements, but quite another to
take wild guesses about intervening points based on an existing pseudo
inverse reconstruction. It won't stop people from bicubic
interpolating and unsharp masking though :(

You might be able to find matched filters that do what you want for
simple structures but you run the risk of seeing false positives and/
or missing real features that are just slightly different to your
ideal template.

I am looking to study limitations when doing so (e.g. is it possible
to resolve a sublte ridge in a medical scan that was otherwise lost
during reconstruction due to imaging system limitations? Can it be
done solely looking at the local neighborhood?). If it can be done,
how to validate the results?

I will look for issues of the journal you mentioned.

They are proceedings of the IEEE conference expensive don't buy them -
find in a university library. A lot of it is too specific to radio
astronomy, but some is more general. There are also several books
about the application of Bayesian methods and Maximum Entropy to image
reconstruction some of which include medical diagnostic MRI & X-ray
images.

Regards,
Martin Brown

.

Relevant Pages

  • Re: Super resolution reconstruction for medical images?
    ... The measurement data are ... disadvantage is that you end up with a resolution that is strongly ... resolution reconstruction not during/after imaging; ... Martin Brown- Hide quoted text - ...
    (sci.image.processing)
  • Re: Super resolution reconstruction for medical images?
    ... already with single pixel resolution. ... You are between a rock and a hard place when indirect imaging ... techniques like MRI and tomography are used. ... Noiseless Maxent reconstruction: delta_function ...
    (sci.image.processing)
  • Re: Super resolution reconstruction for medical images?
    ... Example based superresolution is for increasing the resolution of an image ... techniques like MRI and tomography are used. ... Noiseless Maxent reconstruction: delta_function ...
    (sci.image.processing)
Quantcast