GM Mask creation

[armenbod]

In the code, it takes the pv gm mask from the structural data if none is supplied:

oxasl_enable/oxasl_enable/enable.py

Lines 76 to 77 in ca8733e

	wsp.log.write(" - Taking GM ROI from segmentation of structural image\n")
	wsp.enable_gm_roi = Image((wsp.structural.gm_pv.data > 0).astype(np.int32), header=wsp.structural.struc.header)

In Shirzadi2018, they say they are using a 70% probability threshold for GM from which they define the mask, but in this code no threshold is used. How does this affect the output quality score? Do we need to consider a threshold?

[mcraig-ibme]

I agree this looks very dubious. It is using the GM partial volume which will be > 0 in many voxels with very little GM.

The original shell code this was based on used the FAST segmentation directly which is more sensible but doesn't correspond to the 70% threshold either (it will count a voxel as GM if GM is the largest PV which will typically mean > 50%).

I suspect the intention here was to use gm_seg rather than gm_pv which would give a 50% threshold. The python code was based on the shell scripts not the original paper!

I will run a test case using current code and then thresholds of 0.5 and 0.7 to see the effect on the amount of GM detected and the quality measures.