TissueSeg:itkEMS Reliability

Results for the itkEMS Reliability study

The Lilly "10" QC dataset of the same subject was segmented with itkEMS. Three dual-channel segmentation trials were performed to compare the segmentation reliability using different versions of itkEMS:

• itkEMS 1.5.4 performing an atlas-to-case affine registration then bias-field correction
• itkEMS 1.7: performing an atlas-to-case deformable registration then bias-field correction
• itkEMS 1.8: performing a bias-field correction then atlas-to-case deformable registration

Parameters are not equal but have been set the same way they are applied in our studies:

• itkEMS 1.5:

 - Filter iterations:	10
 - Filter Time step:	0.01
 - Filter method:	Grad aniso diffusion
 - Max bias degree:	3
 - Prior 1:		1.3
 - Prior 2:		1
 - Prior 3:		0.7
 - Prior 4:		1

• itkEMS 1.7 & itkEMS 1.8:

 - Filter iterations:	6
 - Filter Time step:	0.01
 - Filter method:	Curvature flow
 - Max bias degree:	4
 - Prior 1:		1.3
 - Prior 2:		1
 - Prior 3:		0.7
 - Prior 4:		1
 - Atlas warp:		1
 - Atlas warp grid X:	5
 - Atlas warp grid Y:	5
 - Atlas warp grid Z:	5

The covariance was computed in each trial, not only for each tissue (White matter, grey matter and CSF), but also for the Intra Cranial Volume, sum of the three tissues.

 - Covariance/Tissue	WM		GM		CSF		ICV        
 - itkEMS1.5		1.20%		1.03%		6.26%		0.96%
 - itkEMS1.7		2.89%		1.32%		3.68%		0.46%
 - itkEMS1.8		1.82%		0.89%		4.30%		0.81%

In this study, itkEMS1.7 seems to be the most reliable version concerning the ICV, even if the results differ for each tissue. As the input images don't need a strong bias field correction, we do not seee itkEMS1.8 method's improvement compared to itkEMS1.7.

See data detailed on spreadsheet:

For each itkEMS version and each tissue, an average T1-weighted label image has been computed. Then, using the 'EvaluateSegmentationResult' tool, volume differences, average differences and tanimoto errors have been computed between each case and the previously computed average .

See data detailed on spreadsheet:

cvachet@email.unc.edu