Case VI: Partial coverage
Segmentator can be used to segment partial coverage (not whole brain) images. The resulting segmentation could reduce the amount of manual slice-by-slice tissue label editing.
| Package | Used version |
|---|---|
| Segmentator | 1.5.2 |
| FSL-fslmaths | 5.0.9 |
| 3D Slicer | 3.8.1 |
| ITK-SNAP | 3.6.0 |
- Partial coverage MP2RAGE (7 Tesla) image positioned around motor cortex in one hemisphere for a layer-fMRI study (acquired through personal communication with Laurentius Huber).
Researcher's aim is to delineate gray matter (GM) tissue on both white matter (WM) boundary and cerebrospinal fluid (CSF) boundary for further layer-fMRI analyses. The researcher tries several automatic tissue segmentation algorithms with no satisfactory results. The researcher considers falling back to fully manual slice-by-slice labelling. This is the point where the current tutorial comes into play with the promise of more controlled generation of initial tissue labels before considering the laborious manual slice-by-slice editing.
Disclaimer: Here I would like to lay out a possible WM & GM segmentation strategy for partial coverage MR images. This is merely a demonstration case. I do not claim that solution presented here is the best one in every case. Several steps demonstrated in this tutorial can be combined with other methods (eg. fsl-fast) anyway.
I would like to start by mentioning a few preprocessing steps to improve local and global image statistics. These would also help other tissue segmentation methods.
- Coarse masking of irrelevant tissues: I have used ITK-SNAP with a drawing tablet to quickly create a brain extraction mask. I have selected the largest size spherical brush to paint the brain, saved the segmentation image and used it to mask the initial image. This step took around 15 minutes of manual work.
- Denoising to improve local contrast: I have used a non-linear anisotropic diffusion filter to accomplish this. This filter reduces the noise in images while preserving edges.
segmentator_filters /path/to/image.nii.gz --nr_iterations 5
You can see that the definition arc-like structures in 2D histogram improves after filtering.
- Bias field correction to improve global contrast: I have used N4ITK bias correction from 3D Slicer to have a more homogenous image.
You can see that 2D histogram looks much less smeared. This will make it easier to select right clusters.
After the preprocessing I have decided to label tissues of interest by using default Segmentator interface.
- Select and export cerebrum (WM and GM) region:
segmentator /path/to/image.nii.gz --valmin 0 --valmax 4000
- Select and export WM region:
Steps 4 and 5 took around 5 minutes of manual work. At this point most of the WM and GM is grabbed however there are some false positives in form of flimsy structures (eg. vessels). These false labels can be easily removed by basic morphological operations.
-
Erode 1 step to isolate thin/flimsy structures.
-
Apply connected clusters thresholding to remove isolated structures.
-
Dilate 1 step to regain previous thickness.
Cerebrum label is looking much better now. Next step is to improve WM labels.
-
Connected clusters thresholding on WM label.
-
Apply closing (1 step dilation followed by 1 step erosion)
- After reducing the false positives on cerebrum and WM labels, it is time to merge them.
fslmaths /path/to/cerebrum.nii.gz -add /path/to/wm.nii.gz
Steps 6 to 12 does not cost any manual time. After having a decent WM-GM tissue label file the researcher can decide to perform further manual corrections. In my experience, some amount of manual slice-by-slice correction is always needed to have accurate and precise gray matter definition for layer-fMRI research.
