users
users > RE: CMTK does not generate proper reformatted files
Jun 20, 2018 04:06 PM | Torsten Rohlfing
RE: CMTK does not generate proper reformatted files
Hi Mahmood -
Typically, when you see a stack "completely white" that means there was no overlap between the reference and the floating images after registration. As a result, all pixels in the reformatted image will be set to a "padding" value. To explain why that appears white - for integer data, the padding value defaults to "-1", but for unsigned data that is converted to the maximum allowed value, thus white.
Anyway, this suggests registration failure, and here are a few possible reasons why that might happen:
a) bad initialization (e.g., initial transformation is too far off from the correct alignment),
b) excessive optimization step size (i.e., the initial step size is large relative to the size of the images, allowing the algorithm to place one image outside the other altogether),
c) low content overlap between the images (i.e., the images represent different areas of the anatomy and there's only a small portion that's visible in both),
d) unsuitable image similarity metric (e.g., using correlation for label data).
Case a) could be addressed by providing a better initial transformation, for example using the "make_initial_affine_xform" tool. If all else fails, try providing a good initial transformation "by hand".
Case b) could be addressed by reducing the initial step size. Not sure how to do that through the GUI (?), though.
Case c) could be addressed by providing a good initial transformation and, in addition, reducing initial step size.
Case d) is actually quite unlikely, but obviously trying a different metric may help.
Best,
Torsten
Typically, when you see a stack "completely white" that means there was no overlap between the reference and the floating images after registration. As a result, all pixels in the reformatted image will be set to a "padding" value. To explain why that appears white - for integer data, the padding value defaults to "-1", but for unsigned data that is converted to the maximum allowed value, thus white.
Anyway, this suggests registration failure, and here are a few possible reasons why that might happen:
a) bad initialization (e.g., initial transformation is too far off from the correct alignment),
b) excessive optimization step size (i.e., the initial step size is large relative to the size of the images, allowing the algorithm to place one image outside the other altogether),
c) low content overlap between the images (i.e., the images represent different areas of the anatomy and there's only a small portion that's visible in both),
d) unsuitable image similarity metric (e.g., using correlation for label data).
Case a) could be addressed by providing a better initial transformation, for example using the "make_initial_affine_xform" tool. If all else fails, try providing a good initial transformation "by hand".
Case b) could be addressed by reducing the initial step size. Not sure how to do that through the GUI (?), though.
Case c) could be addressed by providing a good initial transformation and, in addition, reducing initial step size.
Case d) is actually quite unlikely, but obviously trying a different metric may help.
Best,
Torsten
Threaded View
| Title | Author | Date |
|---|---|---|
| Mahmood Hoseini | Jun 19, 2018 | |
| Mahmood Hoseini | Jun 25, 2018 | |
| Mahmood Hoseini | Jun 21, 2018 | |
| Torsten Rohlfing | Jun 22, 2018 | |
| Greg Jefferis | Jun 22, 2018 | |
| Torsten Rohlfing | Jun 20, 2018 | |
| Torsten Rohlfing | Jun 20, 2018 | |
| Greg Jefferis | Jun 19, 2018 | |
| Mahmood Hoseini | Jun 19, 2018 | |
| Greg Jefferis | Jun 19, 2018 | |
| Mahmood Hoseini | Jun 19, 2018 | |
| Greg Jefferis | Jun 20, 2018 | |
| Torsten Rohlfing | Jun 20, 2018 | |