how to run best skullstripping from command line

[mrneont]

Hi-

I would like to do skullstripping/brain-extraction (i.e., get a detailed mask of the brain within a whole head FOV) from the command line using brainchop. What is the best way to do so?

When I run:

brainchop                                                             \
    "${dset_in}"                                                    \
    --skull-strip                                                     \
    --mask "${mask_out}"                                        \

... I get prompted to answer a question, but instead I want to script this so I can swarm it over many datasets.

Therefore, I added an option --no-optimize to not get prompted:

brainchop                                                             \
    "${dset_in}"                                                    \
    --skull-strip                                                     \
    --no-optimize                                               \
    --mask "${mask_out}"                                        \

... but I would like to not sacrifice masking quality by doing this (the name of the option kind of suggests that).

Is the latter command the best way to script running brainchop for skullstripping on the command line?

[sergeyplis]

Dear @mrneont this issue belongs in https://github.com/neuroneural/brainchop-cli ;)

if you need speed, I recommend running skull stripping once with optimize option. Possibly even with
BEAM=3 brianchop # it will take long
You will get about 2x speedup if you do this in all of your subsequent runs. You will not be asked to optimize again.

Does this work?

[mrneont]

Hi, @sergeyplis -

Ah, sorry to put this on the wrong page... Should I open there and refer to this existing thread?

At the moment, speed is not an issue---it runs quickly as is. I am focused on getting the best, most detailed skull-strip/brain extraction. I wanted to make sure that by adding --no-optimize, I wasn't harming the quality of output. The only issue with optimization running was that I was asked a question to answer on the commandline (is there a -y flag I can provide, to answer yes?), and I wanted everything scripted.

Would you expect adding --beam 3 to provide better skullstripping? (I am testing that to see how much extra time it adds.)

Or are there any other options I should use for best results?

Thanks.

[sergeyplis]

@mrneont if you run the beam search once - not once per input, but just once per model after pip install brianchop - you will not be asked the question again as long as you are running as the same user with the same home directory. The result is cached.
You can pass --beam 0 to avoid optimization or prepend the command line with BEAM=0 like BEAM=0 brainchop ...