DBP2:UNC:Tissue Segmentation Evaluation

Back to NA-MIC Internal Collaborations, UNC DBP 2

Objective

We would like to evaluate the brain tissue segmentation using two different algorithms: itkEMS and Slicer3 EMSegment.

• A reliability study was performed on adult dataset.
• A qualitative study was performed on pediatric dataset.

Reliability study on adult dataset

Method

A UNC dataset has been used for this study, one patient being scanned 10 times on the same scanner. Two dual-channel segmentation trials were performed to compare the segmentation reliability using the two algorithms. A single T1-weighted image was registered to the UNC template atlas and the rest of the T1 images were affinely registered and scaled to that original T1 image to minimize scanner-influenced differences.

Results

One case failed at the execution using EMSegment and has been removed.

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

 - Covariance/Tissue	WM		GM		CSF		ICV        
 - itkEMS		1.27%		0.99%		6.27%		0.88%
 - EMSegment		7.66%		3.15%		14.84%		1.51%

See results detailed on spreadsheet:

Media:UNCDBP2_TissueSegmentationReliabilityStudy.xls

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

See data detailed on spreadsheet:

Media:UNCDBP2_TissueSegmentationReliabilityStudy2.xls

Screenshots

EMSegment tissue segmentation on adult case

itkEMS tissue segmentation on adult case

Qualitative study on pediatric dataset

EMSegment tissue segmentation on pediatric case

itkEMS tissue segmentation on pediatric case