10
$\begingroup$

I'm trying to automatically detect some medical defined anatomical landmarks in a CT reconstructed volume. Medical doctors use these landmarks to measure some patient specific parameters. I have attempted to use the SIFT feature descriptor, since these anatomical landmarks are kind of "keypoints". This did not work very well since the landmarks are points (or tiny regions) that are in general not "interest points" as defined by SIFT. I have been looking many pattern/template matching algorithms but, when I do not have rotation/translation/scale problems, I find that the extracted features do not differentiate each landmark enough (from the rest of the landmarks and from the rest of the non landmark patches) to train a classifier that performs well enough (at least an 80% of detection accuracy).

Please let me know if I'm not stating the problem clearly enough.

I would really appreciate any advise.

Thanks!

Example image:

The small x crosses are the landmarks I want to detect. The lines represent the measures taken. These are some slices of different cases (of course, I cannot post the full 3D volume)

The small x crosses and little squares are over the landmarks I want to detect (I forgot to mention that I have a training set, with the labeled landmarks). The white lines represent the measures taken. These are some slices of different cases (of course, I cannot post the full 3D volume).

$\endgroup$
  • $\begingroup$ Could you post some representative pictures and point out the features that you are trying to detect? $\endgroup$ – Jim Clay Apr 20 '12 at 14:59
  • $\begingroup$ I see the Xs and boxes in the image, but I don't understand what makes them landmarks. Were the ones in the image selected by hand? If you can describe how they are chosen that would help a lot. $\endgroup$ – endolith Apr 20 '12 at 15:53
  • $\begingroup$ Yes, these landmarks are selected by hand by MDs. Actually, mainly their position in the bone and their curvature are what make them detectable by the clinician. Also the cortical bone width is maybe taken into account (this is natural for them, it is really hard to reverse engineer how they find these points) because it is thinner than in other parts of the bone. My difficulty is actually in modelling all this in a feature extractor. $\endgroup$ – Federico Apr 20 '12 at 16:26
5
$\begingroup$

I hesitate to write this as an answer, but given that you're asking only for advice, I will do so.

I suggest investigating techniques based on the Dual-Tree Complex Wavelet Transform (DTCWT). These have shown to be useful for generating descriptors that have good tolerance to shift, scale and rotation of the source images. It's not the classic problem in that you're not allowing the points to be assigned for you, but I suspect with some thought you can adapt the techniques to predefined landmarks.

Clearly, the landmarks have some interest from the perspective of a clinician, so there is something of interest about them - it's simply a case of modelling that in the descriptor. Wavelet techniques (in particular the DTCWT) tend to be good at modelling features that the eye picks up on.

A starting point would probably be this fairly recent paper.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.