This is a known problem with the method as it currently stands. Because only
tissue probability maps of GM, WM and CSF are used, there is nothing that
explains the large intensity difference between the scalp and surrounding
air. If the whole image was used during the segmentation, then there would
be a tendancy for the tissue probability maps to be expanded to cover the
whole head (i.e. all non-air voxels in the images). To get around this
problem, the images are thresholded so that the large expanse of fresh air
around the head is not included in the calculations. Your CSF voxels are
probably very low intensity, and therefore below this intensity threshold.
Ideally, an MRF-like procedure within the model should help the situation, as
would including a few morphological operations (e.g. hole filling), but I
haven't got around to this yet.
Best regards,
-John
> I'm interested in getting CSF maps in native space by segmenting
> T1-weighted images with good contrast. The CSF maps show the CSF
> boundaries but the ventricles as well as the larger sulcal spaces have no
> signal inside them. The results were the same even I when I increased the
> bias regularisation. Smoothing the scans a bit (3x3x3 kernel) made the
> boundaries thicker but the ventricles were still empty. It's as though the
> CSF was mistaken for background. Are there any other defaults I should try
> changing?
>
> Ironically, I had no such problem with lesser quality scans in which the
> contrast was not as good.
|