Samseg Aseg Test | March 19 2017

Andrew ran samseg on the Buckner39 test set. mri_compute_seg_overlap was used to compare the subject's segSubSpace.mgz with its seg_edited.mgz.

results

The mean overlap of 18 subjects was above 0.8, but 21 subjects had a mean of ~0.05

subjs_dice.png
fig 1: mean dice scores for all Buckner39 subjects

Every unsuccessful subject's low dice was due to inaccurate registration. Specifically, each segSubSpace volume is placed too high relative to the brain:

981102_vc604_regerror.png
fig 2: subject 981102_vc604 with poor segSubSpace registration

Here are the mean dice coefficients for each label with the 21 outliers removed

labels_dice.no_outliers.png
fig 3: mean dice scores for each label over all non-outlier Buckner39 subjects

data location

main test directory: /autofs/cluster/fsm/users/samseg/tests/2017-03-19_Buckner39

charts and snapshots: /autofs/cluster/fsm/users/samseg/tests/2017-03-19_Buckner39/analysis

Buckner39 test set data: /autofs/cluster/fsm/users/samseg/subjects/Buckner39

SamsegAsegTesting/2017-03-19_Buckner39 (last edited 2017-03-20 14:00:52 by AndrewHoopes)