TissueSeg:EMS1.7 Reliability

The Lilly "10" QC data set of the same subject was segmented with EMS 1.7. Two separate trials of 10 and 20 filter iterations were performed, each with a 3 channel and a 1 channel (T1 only) segmentation and the same grid parameters. A single T1 image was registered to the template atlas and the rest of the T1 images were registered and scaled to that original T1 image to minimize scanner-influenced differences. Some of the T1 corrected images for the 1 channel segmentations seemed very washed out; including much extra brain tissue that at the very least would need to be cleaned up manually.

The coefficient of variance was calculated for each trial. The coefficient of variance for intracranial volume (GM+WM+CSF) for 10 filter iterations was determined to be 1.18 for the 3 channel and 3.11 for the 1 channel segmentation. For the second trial of 20 iterations the CV was calculated to be 1.18 and 3.16 for the 3 channel and 1 channel segmentations respectively. The spreadsheet includes a break down of constituent brain matter; as well as the new data and the older run data (20 iteration in gray) for comparison.

See data detailed on spreadsheet:

Media:LQC_EMS17_trial_10.xls

jblocher@med.unc.edu

The test with 10 filter iterations was repeated with the warping feature of EMS disabled. Similarly to the first trial, both 1 channel and 3 channel segmentations were tested. In general, the results show that disabling the warping does not significantly improve the 1 channel reliablilty. However, the quality of the segmentations were markedly improved by including less extra-brain tissue.

The 3 channel results do reflect an improvement in reliablity without the warping feature. The spreadsheet below details both non-warped segmentations and older data from the warped trials for comparison.

Again in this case the coefficient of variance was calculated for each trial. With the warping feature of EMS 1.7 disabled the CV for the intracranial volume was calculated to be 0.804 for the 3 channel and 2.94 for the 1 channel segmentation.

See data detailed on spreadsheet:

Media:LQC_EMS17_trial_10 NoWarp.xls

jblocher@med.unc.edu

A further test was conducted to compare EMS 1.7 using EMS 1.5.4 (an earlier production version), the same Lilly QC data set, and identical segmentation parameters. The coefficient of variance for the intracranial volume was again calculated and determined to be 0.866 for the 3 channel and 1.78 for the T1 only segmentation.

See data detailed on spreadsheet:

Media:LQC_EMS154_trial_10.xls

jblocher@med.unc.edu