Difficulties with segmentation

[Jack Tattersall]

I'm having troubles segmentating some my images. Im trying to segmentate an AVG + the artery and vein it's connected to. However, Thresholding, level set or machine learning alldont work for me. Sometimes I get an error "contour not created and added since it's invalid", but I'm not sure what's invalid. Is the MRI resolution too low / too noisy? Also when trying to use threshold no contour is created, when pressing twice. I think this has to do with the same error. When I try the machine learning option, the contour is way too small and I need to zoom in a lot to see the faulty contour.

Thanks in advance,
Jack Tattersall

[David Parker]

Hi Jack,

The error message means that SV could not find an isocontour with value 0.0 (taken from the Preset slider). It would be best to disable Preset I think.

The resolution of the segmentation slice (the red square) looks too small, the left hand image windows look blocky, click on the Size button to make it larger by maybe 2x.

There does not seem to be much signal in the vessel you are trying to segment. That will cause problems with all image based segmentation methods (threshold, level set, machine learning, etc.) You can manually define a segmentation using the SplinePoly method.

Cheers,
Dave

[Jack Tattersall]

Hello David,
Thanks for the quick response, just a couple more follow up questions.
I attached 2 more pictures of a different series of the same patient but with white vessels and zoomed out.
1) Machine Learning gives (as can be seen in the attachments) a very small contour and I'm not sure why. Also Thresholding still does not seem too
work, without the preset does not make a difference, I doesnt seem to initiate, i dont get to see that it starts to make contours when i activate
thresholding (the button is blue) and click and slide on the 2D plane (like in the tutorial). Instead nothing
happens. LevelSet still gives the invalid contours error.
2) Do you think the 2D view still is too blocky/ too low res to segmentate. Maybe this is the problem, if so I can try and request for future patients that the resolution is higher if poss, because I know this was difficult.

Thanks for the help!
Cheers, Jack

[Jack Tattersall]

Hello David,

I have just talked to the MR clinical physician, and we are going to test if 3D isometric images will give better resolution in the axial plane. Hopefully this will resolve some of the issues.
Cheers Jack

[David Parker]

Hi Jack,

1) The ML algorithm is not able to create a good segmentation because the image widow does not contain enough data (pixel values) to distinguish a vessel boundary. None of the image based algorithms will work for this data.

2) The second picture you attached seems like the segmentation should work although the image gradient (the lower left screen) does not show a clear boundary. The image window may still not have enough data for the ML segmentation to work well. You could try threshold on this. Level set is a bit tricky, there are lots of parameters that may need to be set to segment this data well.

Can you send me this data, cleaned of patient information of course!

I've attached a picture of an MRI data that we typically use in your research. You will see how much better the resolution is.

Cheers,
Dave

[Jack Tattersall]

Hi David,

I could send you the data, but yesterday I've discussed a new protocol for the image acquisition with the MR fysician and it allready seemed much better, so I'll wait for the next patient and try it with the new data.

Cheers,
Jack