Differences between revisions 66 and 67
Deletions are marked like this. Additions are marked like this.
Line 67: Line 67:
||ADNI_T2TCV||[[attachment:ADNI_Total_Cranial_Vault_Segmentation_Method_20121108.pdf|33 subjects, verified Total Cranial Vault (TCV) segmentations from T2 scans]],[[attachment:TCV_DICT.csv|TCV_DICT.csv]], [[attachment:TCV.csv|TCV.csv]]|| #||ADNI_T2TCV||[[attachment:ADNI_Total_Cranial_Vault_Segmentation_Method_20121108.pdf|33 subjects, verified Total Cranial Vault (TCV) segmentations from T2 scans]],[[attachment:TCV_DICT.csv|TCV_DICT.csv]], [[attachment:TCV.csv|TCV.csv]]||
Line 80: Line 80:
The SASHA and ADNI_T2TCV datasets are the primary validation sources, as is based on expert manual segmentation of T2 data (SASHA) and manual verification of T2-based segmentation (ADNI_T2TCV, where 33 grade 3 subjects of 300 possible were chosen...only subjects with DICOMs selected)). ADNI30 data consists of all intracranial voxels labeled in T1 image. The SASHA dataset is based on expert manual segmentation of T2 data. ADNI30 data consists of all intracranial voxels labeled in T1 image.
Line 98: Line 98:


== Project Notes ==

==== Tasks ====
 * NS - confirm Usage format and Validation plan with Contributors
 * CL/AH - update the Method section with the sbTIV algorithm/approach
 * AH - copy SASHA datasets to samseg/subjects dirs from locations shown in [[eTIV]] page [DONE]
 * AH - run SPM12 eTIV function on SASHA, Buckner40, ADNI60, IXI_79 datasets
 * CL/AH - create validation scripts, whereby given a set of sbTIV measures for a dataset, plot against the reference dataset (either manual values or eTIV) see Methods section of [[eTIV]] for details on SASHA scripts, and the Malone paper for their approach.

This page is readable only by those in the LcnGroup.

sbTIV - segmentation-based (estimated) Total Intracranial Volume

Contributors: Nick Schmansky (NS), Christian Larsen (CL), Bruce Fischl (BF), Doug Greve (DG), Andrew Hoopes (AH), Bram Diamond (BD), Oula Puonti (OP), and Koen Van Leemput (KVL)

Total intracranial volume (TIV/ICV) is an important covariate for volumetric analyses of the brain and brain regions. It is commonly used to correct for head size variation (i.e., 'normalize' hippocampal volume size). The gold-standard method is manual delineation of T2 scans. Freesurfer currently provides the eTIV measure, described here. It has been shown to be a robust covariate:

However, etiv is not obtained via direct segmentation of all structures within the skull boundaries. Rather, its estimation depends on the linear transform between the T1 image and a T1w atlas.

There are other estimation methods. A recent example, where a comparison to freesurfer eTIV is made, is:

Our segmentation-based estimate of total intracranial volume, sbTIV, relies on structure segmentations (which includes CSF and other intra-cranial non-brain structures) obtained by the Samseg processing stream.

Usage

Samseg computes and outputs the file samseg.stats, which contains the volumes of each labeled structure, e.g.:

# Measure Soft_Nonbrain_Tissue, 987441.150308, mm^3

sbTIV takes samseg.stats as an input together with an optional list of structure names that define intracranial volume. Names within the list must correspond to those used within the atlas used by samseg. sbTIV outputs sbtiv.stats, which contains the intra-cranial volume estimate, e.g.:

# Measure Intra_Cranial_Volume, 160000.123456, mm^3

Commandline arguments for sbtiv can be found by running sbtiv with the --help parameter.

Additionally...

mri_segstats --sbTIVin <sbtiv.stats> ...

When <sbtiv.stats> is provided to mri_segstats, the sbTIV measure is read and added as an additional measure of brain volume to its output summary file. From its help file:

  5. sbTotalIntracranialVolume (sbTIV) - estimate of the intracranial volume
  based on the Samseg-based segmentation. See surfer.nmr.mgh.harvard.edu/fswiki/sbTIV

and as found in aseg.stats:

# Measure sbTotalIntracranialVolume, sbTIV, samseg-based estimated Total Intracranial Volume, 1667606.252292, mm^3

Method

For each voxel in the image, Samseg computes the probability that it belongs to each of the labels governed by an underlying brain atlas used during segmentation. The full list of segmented labels include skull, non brain tissue and csf. This is in contrast to the older FreeSurfer .gca -based atlas, which was limited to brain structures which did not fully 'fill' the skull. This means that an ICV estimate made from a samseg segmentation is in principle immune to shrinking brain structures (where the .gca -based structures do not contain labels allowing immunity to this effect). This is a critical characteristic of an ICV estimate and is particularly important in measuring structures in an aging population, where ICV does not change with age, but many structures shrink.

Since the samseg processing stream supports specifying a custom atlas for segmentation, sbtiv correspondingly allows you to specify a list of labels to be used for ICV computation. If none is provided, the default list of labels is used and also printed to the console for inspection. NOTE: Since you can specify your own, custom atlas for samseg to use, you could in principle use one without skull and non brain tissue. However, this makes the segmentation less robust and renders sbtiv unusable.

The ICV computation itself is trivial, and is the sum of intracranial label probabilities over all voxels multiplied by the volume of a voxel. This weighing accounts for voxels that is not entirely certain to be intracranial, e.g., due to partial volume effects. This is more accurate than summing over a hard segmentation. Since samseg already outputs the weighed volume for each label, sbtiv computes the sum of the label volumes within the ICV list. As such, you may chose to omit usage of sbtiv altogether and simply use the samseg output directly for ICV computations.

Validation

The sbTIV measure was validated against these reference sources:

dataset

description

SASHA

22 subjects, manual TIV from Buckner/OASIS T2 scans

#||ADNI_T2TCV||33 subjects, verified Total Cranial Vault (TCV) segmentations from T2 scans,TCV_DICT.csv, TCV.csv||

ADNI30

30 subjects, all intracranial voxels labeled manually

Buckner40

40 subjects, eTIV from FS v6.0, and SPM12

ADNI60

60 subjects, eTIV from FS v6.0, and SPM12

IXI79

70 subjects, eTIV from FS v5.3, and SPM12

This link has notes on generating the ICV measure in SPM12. ...and this paper. ...and this matlab file.

The test subjects and scripts are located here: /cluster/fsm/users/samseg/<dataset>.

The SASHA dataset is based on expert manual segmentation of T2 data. ADNI30 data consists of all intracranial voxels labeled in T1 image. The other datasets compare against the freesurfer eTIV and SPM12 eTIV.

Manual segmentations

Pearsons correlation coefficient and the absolute % difference in ICV volume between each method and the 'gold truth' manual segmentations were computed, and the data points correspondingly scatter plotted.

  • sasha.png

    adni30.png

Method agreement

As a secondary validation measure, correlations and absolute % difference in measured ICV were computed for a pairwise rotation of freesurfer 6.0, spm 12 and samseg, to explore to what extent the methods 'agree' with each other.

Dataset

FS 6 vs spm12

samseg vs FS

samseg vs spm12

ADNI30

r=0.922, diff=6.588%

r=0.975, diff=2.697%

r=0.961, diff=4.131%

ADNI60

r=0.913, diff=7.453%

r=0.968, diff=3.148%

r=0.952, diff=4.666%

Buckner40

r=0.922, diff=6.632%

r=0.974, diff=3.467%

r=0.966, diff=3.397%

IXI79

r=0.892, diff=9.236%

r=0.914, diff=4.011%

r=0.969, diff=6.234%

SASHA

r=0.754, diff=6.442%

r=0.949, diff=2.725%

r=0.870, diff=3.898%

sbTIV (last edited 2023-01-05 11:00:54 by NickSchmansky)