These results are simply horrible. I have similar problems with one database currently which have large hyper signal in the fat around the brain. I am planning on releasing an update of the procedure which will be robust to these type of effects. I am not sure what tool you used though. If if is NIAK_BRICK_MASK_BRAIN_T1, I would suggest to try NIAK_BRICK_T1_PREPROCESS as well (the linear coregistration is sometimes surprisingly robust to disastrous initial masks, and the subsequent brain extractions use a quite different approach). If you tried both and both failed, then I would need to see an image to figure out what may cause the algo failure. But if it is similar to the problem I ran into, then it is related to a highly skewed histogram with outliers (high) values in the tissues surrounding the brain. Manually clipping the volume may fix your issue :

>> [hdr,vol] = niak_read_vol('file_anat.mnc');
>> hist(vol(:))
>> vol(vol>1000) = 1000;
>> hdr.file_name = 'file_anat_clipped.mnc';
>> niak_write_vol(hdr,vol);

where the value 1000 is an example and needs to be determined based on the histogram. The changes I mentioned above will involve some automated clipping. If you would be willing to share one problematic brain, please email me and I'll have a look. I am very interested in collecting those "hard" brains to test further updates of the brain segmentation/coregistration tools.

Best regards,

Pierre