Automatic pipeline for quantitative brain tissue segmentation and parcellation: Experience with a large longitudinal schizophrenia MRI study

1
Automatic pipeline for quantitative brain tissue
segmentation and parcellation Experience with a
large longitudinal schizophrenia MRI study 1,2G
Gerig, 3,2S Joshi, 1H Gu, 1D Perkins, 1RG Steen,
1R Hamer, 1M Jomier, 1JA Lieberman
Dept. of 1Psychiatry, 2Computer Science,
3Radiation Onc. University of North Carolina,
Chapel Hill, NC 27614, USA gerig_at_cs.unc.edu /
http//www.cs.unc.edu/gerig
Subjects
METHODS
MRI Brain Tissue Segmentation

• Expectation Maximization Segmentation algorithm
  (EMS)
• Single or multiple MRI contrasts
• Multi-modal MRI and Atlas Registration
• Built in bias correction
• Initialization classification governed by
  statistical atlas
• Built in brain stripping
• Efficient 15 minutes
• Van Leemput et al.., Automated model-based tissue
  classification of MR images of the brain , IEEE
  TMI, 18(10), 1999

Summary
Introduction Quantitative large-scale imaging
studies require processing pipelines that are
mostly automatic, reliable, generic in regard to
differences in imaging protocols, and rigorously
tested for reproducibility. Such software, if
distributed to other sites, could help with
cross-validation of results obtained at other
various sites and thus meet concerns in regard to
processing-specific results. We have developed
software within the open-source Insight Toolkit
(ITK, National Library of Medicine), that
includes registration to standard coordinates,
brain tissue segmentation, inhomogeneity
correction, brain stripping, and brain lobe
parcellation. Methods An ITK implementation of
automatic brain segmentation from multi-modal MRI
has been combined with a brain parcellation based
on template deformation. This results in an
automatic pipeline for efficient processing of
large image databases to provide reliable
estimates of gray matter, white matter and
cerebrospinal fluid. A parcellation template is
obtained by an experts parcellation of an
average MRI image. This template is deformed to
each subjects MRI by diffeomorphic large scale
registration. We applied this new method to a
large first-episode longitudinal schizophrenia
study with 91 FE patients and 37 controls at
baseline and 48 patients and 26 controls at 6
month follow-up. All MRI was performed on a 1.5T
GE Signa, using T1-w gradient echo and dual-echo
T2w/PDw spin-echo protocols. Results
Reliability of tissue segmentation was tested on
data from a multi-site reliability study with the
same subject imaged at 5 sites two times.
Reproducibility of full brain tissue volumes were
below 1, demonstrating the excellent stability
of segmentation. Reliability of the parcellation
was tested with two cross-validation studies
using two manually parcellated templates and
qualitative assessment by three experts.
Multi-site, Multi-scanner Validation of
Segmentation

• Dataset Same subject scanned 2-times (24 hour
  window) at 5 different sites (4 GE, 1 Philips)
  within 60 days
• Automatic brain tissue segmentation using
  three-channel (T1, T2w, PDw) MRI
• Results show excellent reliability and stability
  of multi-site scanning and brain tissue
  segmentation
• M. Styner, C. Charles, J. Park, G. Gerig,
  Multisite validation of image analysis methods -
  Assessing intra and inter site variability, Proc.
  SPIE MedIm 02, 09/2002

March 2005 2
March 2005 1
2
RESULTS
Brain Lobe Parcellation
Parcellation Template
Longitudinal Assessment

• Construction of MRI template via unbiased
  building of average template
• S. Joshi, B. Davis, M. Jomier, and G. Gerig,
  Unbiased Diffeomorphic Atlas Construction for
  Computational Anatomy, NeuroImage 23 (1).
• Expert subdivision into major lobar and
  subcortical structures

Application to large study

• Nonlinear deformation (fluid high-dimensional
  deformation) of average template to each subject
• Deformation of parcellation map ? automatic
  parcellation
• Combination of individual parcellation with
  tissue segmentation (gm,wm,csf)

Validation of Segmentation Reliability via Longitudinal Study Validation of Segmentation Reliability via Longitudinal Study Validation of Segmentation Reliability via Longitudinal Study Validation of Segmentation Reliability via Longitudinal Study Validation of Segmentation Reliability via Longitudinal Study Validation of Segmentation Reliability via Longitudinal Study  
Controls (N25, baseline and 6mt follow-up) Controls (N25, baseline and 6mt follow-up) Controls (N25, baseline and 6mt follow-up) Controls (N25, baseline and 6mt follow-up)      
             
  Mean Time1 Mean Time2 Mean Abs Diff Mean Abs Diff Mean Diff Mean Diff
ICV 1320676.0 1318887.6 7386.3 0.56 -1788.4 -0.14
White 453187.8 452786.3 4603.7 1.02 -401.5 -0.09
Gray 657694.3 651534.0 6835.2 1.04 -6160.3 -0.94
Csf 209793.8 214567.3 5719.6 2.70 4773.4 2.25
?
?
Validation of automatic parcellation

• Cross-validation of 2 labeled brains
• Cross-validation of one labeled subject versus
  warped atlas labeling
• Visual qualitative check of major sulci (3
  raters) based on deviation score
• Parcellation by template warping Precision
  depends on anatomical structure

Results Clinical Study ICOS 2005 116401
Longitudinal Changes in Brain Volume in Patients
with First-Episode Schizophrenia An Exploratory
Analysis of 91 Patients R. Steen G. Gerig H.
Gu D. Perkins R. Hamer J. Lieberman There
were no significant differences between 91
patients and 37 controls at baseline. However, in
both patients and controls, there was a
significant (p lt 0.02) decrease in frontal gray
matter (GM) volume over 6 months, which may be
age-related. There were also significant
volumetric changes over 6 months that were unique
to patients, including a 5 increase in
ventricular volume (p 0.0009) a 1 increase in
whole brain white matter (WM) volume (p 0.02)
a 2 increase in parietal WM volume (p 0.006)
and a 5 increase in occipital WM volume (p
0.01).
Software Environment
UNC NeuroLib Insight Toolkit (ITK) developments
(National Library of Medicine Initiative). Tool
freely available to research community
(www.ia.unc.edu/dev, Matthieu Jomier)
CONCLUSIONS

• Automatic MRI analysis
• Excellent reproducibility, no rater variability
• Suitable for large multi-site studies
• Multi-contrast MRI Analysis extended to
  wm/gm/csf inter-relationships
• Lobe parcellation for regional assessment
• Standardized Software Avoids site-specific
  results which are difficult to confirm. Allows
  cross-center validation
• Longitudinal Study
• Implicit validation via controls (assuming no
  longitudinal change)
• Calibration of variability of whole system
  (MRI, biological variability, segmentation
  variability)
• Precision of full-brain tissue volumetry is in
  the range of 1 ? Effect size needs to be larger.
  Lobe volumetry presents larger error.

Imagine Image Processing Pipeline GUI Modules
are connected to produce a pipeline with several
filters. This example shows how to use Imagine
for tissue segmentation and quantification of a
large set of brain MRI.
March 2005 4
March 2005 3