Yes, clearly a very strong bias field is present. Not surprising then that you’d have that gap in your masks; as @jdtournier mentioned, you’ll need to do something about that bias field first.
Don’t use the -fsl option; it introduces more problems than it fixes, especially with better masking in mind. Use the -ants option instead.
Apart from the “don’t use -fsl option” advise, nonetheless: shouldn’t this have been biascorr.mif instead of geomcorr.mif? I.e., the output of dwibiascorrect? But in any case, try to use the -ants option instead for sure.
That result also looks quite like what I would expect from ANTS.
Well, the output you show in that screenshot looks “decent”. It’s hard to truly verify, since we don’t have the ground truth bias field to compare with of course. However, in your original dataset, the problem was (visibly) that the centre of the brain had much lower intensities than the edges. Just look at e.g. the b=0 signal in the ventricles compared to fluid on the outside of the brain to see that. The bias field output (which you showed in hot colour scale in the screenshot) should reflect that, since that’s what it’s trying to estimate… and indeed it does show that pattern: darker in the middle of the brain, brighter on the outside. Beyond the brain area, you can ignore that bias field output (or rather: don’t go by what it does all the way beyond the brain to check its quality), just check it inside of the brain area.