Poor results in brainExtraction

[ntustison]

Sometimes brainExtraction gives poor results---specifically, nonzero probabilities in the face area. This is due to the training data, in many cases, being defaced. In the short term before adding more training data, in order to improve results for such cases, one can modify the input t1 as follows:

t1 <- antsImageRead(t1_file)

kirby <- antsImageRead(getANTsXNetData("kirby"))
kirby_mask <- thresholdImage(brainExtraction(kirby, modality = "t1", verbose=TRUE), 0.5, 1, 1, 0)
kirby_dilated_mask <- iMath(kirby_mask, "MD", 25)

reg <- antsRegistration(fixed=t1, moving=kirby, typeofTransform="antsRegistrationSyNQuick[a]") 
t1_dilated_mask <- antsApplyTransforms(fixed=t1, moving=kirby_dilated_mask, 
    transformlist=reg$fwdtransforms, interpolator="nearestNeighbor")

t1_defaced <- t1 * t1_dilated_mask   
mask <- brainExtraction(t1_defaced, modality="t1")

[stnava]

@ntustison do you have example ADNI images for which this occurs?

this option might also work

b1=brainExtraction(img, 't1combined' )
b2=brainExtraction( img * iMath( thresholdImage(b1,2,3), "MD", 25 ), "t1combined" )
finabxt = thresholdImage(b2,2,3) %>% iMath("MD",1)

[ntustison]

Yeah, a handful of subjects from James' recent data set have that problem. Let me see if it works.

[ntustison]

It looks like it works for the one case I tried:

[ntustison]

https://github.com/ANTsX/ANTsRNet/commit/bdc79ce324a7fe568b323ce30d8adf2602bee5f3