Mapping Brain Changes Over Time during Development: Challenges, Limits and Potential

1
Mapping Brain Changes Over Time during
Development Challenges, Limits and Potential

• Guido Gerig
• University of Utah
• Scientific Computing and Imaging (SCI) Institute

2
Time Course of Critical Events in the
Determination of Human Brain Morphometry
Neurodevelopmental processes, cortical synapse
density, and their relationship to gray and white
matter volumes on MRI. Giedd et al. 1999, Sowell
et al. 1999.
3
Understanding early Development

• Brain Development in High Risk Children
• Understanding rate and variability of normal
  development
• Detect differences from typical development
  (autism, at risk, drug addiction,
• Early diagnosis ? early therapy ? better future
  for infants and families
• NIH funding Increased support for discovery
  science

4
Motivation Pediatric Neuroimaging

• The ability to document brain development at a
  period when it undergoes a rapid and critical
  modification is absolutely essential to shed
  light on our understanding of brain development
  and change from normal
• Offers profound scientific implications regarding
  our understanding on how the brain orchestrates
  the complex functional and cognitive
  developments.
• Modeling of normal brain development based on
  neuroimage data.
• Provides great insights into possible mechanisms
  and etiology of developmental disorders.
• Age range 0 to 5 so far poorly understood, lack
  of data.

5
Developmental Trajectory Krabbes
Expected
Asympt.
Sympt.
Notion of normative model/atlas Describe
patients relative to population statistics of
healthy development.
6
Outline

• Imaging Technology for Pediatric Imaging
• Analysis of structural MRI
• Population Studies of DTI
• Towards Longitudinal Analysis

7
I MR Imaging of Children

• Non-sedated neonates and children
• Subject cooperation difficult
• Motion problems
• Safety issues
• Solutions
• MTRAs with special training
• High-speed high-field imaging, high spatial
  resolution
• Parallel Acquisition
• Mock Scanner (Training)
• Motion correction

Courtesy LeBihan 2005
8
Early Human MR Images (Damadian 1972)
9
MRI Scanning today Still rapid progress
Siemens Trio 3T (Allegra 3T) T1w T2w, isotropic
1mm3 Rapid DTI Etc.
10
Autism Center of Excellence DTI(definition of
new pediatric protocol)
TimTrio original DTI data 4 times 25 directions
(mosaic)
25 directions
MD
DTI estimation (25 dir variable b)
3D volume reconstruction
tractography
FA
Robert McKinstry, WU Carlo Pierpaoli, NIMH
11
High-Speed Imaging Neonatal MRI at 3T
T1 3D MPRage or FLASH 1x0.9x0.9 mm3
FSE T2w 1.25x1.25x1.95mm3
FSE PDw 1.25x1.25x1.95 mm3
UNC Weili Lin 3T Siemens Allegra Scan Time
Structural MRI (T1 SpinEcho) 8min, DTI 4min
-gt 12 Min tot
12
Training of infants Mock Scanner

• Old MRI scanner used for practice sessions
  (pictures Yale Univ.)
• Subjects learn to remain still for up to 30
  minutes
• Head tracking coupled with video presentation
  (Duke)

13
High-Speed Imaging Infant MRI at 3T
T1
T2
2 weeks
1 year
2 years
UNC Weili Lin 3T Siemens Allegra
14
Rapid Change First year of Development
2 weeks
Movie
1 year
15
Example UNC Scan Success Rate
Siemens 3T Allegra, Weili Lin team
16
New Technology MRI Motion Correction

• GE PROPELLER - Motion Correction Imaging
  (unique pattern of k-space filling that acquires
  data in radial "blades" rotating in sequence)
• Periodically Rotated Overlapping Parallel Lines
  with Enhanced Reconstruction

• Siemens Navigator pulse for online correction
  (Flash T1, courtesy of MGH ).

17
Prenatal MRI / Premies
Petra Hueppi, Geneva and Harvard (26 weeks premie)
Julia Fielding, UNC (intra-utero)
18
Most Recent Parallel Acquisition and matrix coils
23 channel prototype array at 1.5T Gram
Wiggins and Larry Wald, MGH
19
23 channel array for 1.5T

• SNR Maps
• Grad. Echo
• Normalized to volume coil average (1.0)
• SNR gain
• 4 fold in cortex
• 1.75x in corpus
• callosum

Volume coil
Siemens volume coil
23 channel array
23 Channel Bucky
Courtesy Bruce Rosen, MGH
20
UNC MGH Project Modeling Head Shapes for
Infant Coil Design
Example 95 head and brain size for 2yr group.
Statistical modeling of head shapes for infant
matrix coils Collaboration with Larry Wald, MGH
and W. Lin, UNC
21
Images used for Measurements Calibration
Commercial MRI 3D Phantom
22
Example Duke BIAC BIRN scanner calibration
23
Recent Autism Network Study Design

• ACE grant Autism center of excellence.
• Longitudinal study of babies at risk for Autism
  scanned at 6mo 1y and 2yr
• 4 scanning sites
• Seattle
• St Louis
• Philadelphia
• Chapel Hill (2 scanners)
• Data Coordination Center MNI
• Image Processing SCI Utah

Montreal
Seattle
Salt Lake City
St Louis
Philadelphia
Chapel Hill
24
Dataset

• 2 human phantoms scanned twice at each site (t1
  and t2)
• 4 scanners
• 3 Siemens 3T Tim Trio (Type A1,2,3)
• 1 Siemens 3T Allegra (Type B)
• Total of 16 scans

25
Traveling Phantom Multi-site Comparison
26
Tissue segmentation evaluation

• Tissue segmentations computed using itkEMS
• Comparison of individual tissue classes to the
  atlas ones.
• Measurements
• Tissue volumes
• Probabilistic Overlap Volume (POV) Measurement

Extension of the Dice coefficient for
probabilistic images
- Reliability via Kullback-Leibler divergence
Describes differences between probability
densities p(xc) and q(xc) of locations x spread
across the whole image volume.
27
Siemens 3T Trio Cross-site Variability
Full brain tissue volumes
Horizontal axis wm1, gm2, csf3,
icv4 Standard deviations wm0.48, gm0.35,
csf1.05, icv0.17
28
Z-scores Trio versus Allegra (Z scores are
simply how many standard deviations from the mean
a score is )
Horizontal axis Scans (Trios 1-12, Allegra 13 to
16) Vertical axis z-scores per tissue type
(solid line) and scanner (number) Result
Allegra scans show much higher z-scores than 12
Trio scans, with overestimation of wm, icv and
underestimation of csf
29
Tissue volume reliability
POV measure for the tissue segmentation of each
case compared to the atlas tissue segmentation
Coefficients of variation (STDEV/MEAN) for tissue
volume of scanner type A
Scanner type A stability
Gouttard et al., MICCAI08
30
Sub-cortical structure segmentation

• 10 sub-cortical structures of the brain segmented
  using an atlas based registration segmentation.
• Measurement
• Volume stability (coefficient of variation)
• Probabilistic overlap volume (POV)
• Surface distances

31
Sub-cortical structures
Coefficient of variation (COV) within scanner
type A data
Caudate
Amygdala
Hippocampus
Pallidus
Putamen
Scanner type A stability
32
Sub-cortical structures
Surface distances - Maximum absolute distance
(MAD) - Hausdorff
Probabilistic overlap
Gouttard et al., MICCAI08
33
Analysis of structural MRI
34
Normative Pediatric MRI Database (MNI A. Evans
Consortium)
Movie
35
Rapid Change First year of Development
2 weeks
Movie
1 year
36
Contrast changes in early development
T1
T1
T1
5D
6Mo
14Mo
T2
T2
T2
5D
6Mo
14Mo
Courtesy Keith Smith, UNC Radiology
37
Contrast changes in early development
neonate
1 year
2 years
PD
T1
T2
UNC/Utah infant neuroimaging study
38
Neonatal MRI Segmentation

wm-myel
gm
wm nm
Int
39
4 and 6 month old subjects
Intermediate stage of contrast flip between white
and gray, with no differentiation in T1w at 4-6
mt and in T2 at 6-8 mt. T1 and T2 are not in sync
w.r.t. tissue contrast
40
Challenge in Segmentation of 1years olds

• Difficulties for tissue segmentation
• Strong bias inhomogeneity
• Gradual degree of myelination decreasing from
  central to peripheral regions
• Very low contrast in cortical white/gray
• T2 lags behind T1 in its ability to depict wm
  contrast and therefore even shows less contrast.

T1w axial and zoomed
T2w axial and zoomed
41
White matter is very heterogeneous at early
infant stage

• Notion of gray/white tissue with constant
  brightness questionable
• Both contrasts (T1/T2) characterize development
• Cortical thickness?

42
Hi-res T1 1year-old (Weili Lin, UNC)
Fuzzy gm/wm boundary Cortical thickness or
alternative probabilistic measure?
43
Follow-up Hi-res T1 (Weili Lin, UNC)
1yr
2yrs
T1 MRI of same child at 1yr and 2yrs with wm
probability maps wm/gm boundary more fuzzy at
1yr.
1yr
2yrs
44
Brain Segmentation 1year old

• Advanced version of expectation-maximization
  segmentation (M. Prastawa)
• Prior Age-specific atlas
• Nonlinear registration of atlas to subject
• Robust, nonparametic clustering
• Parametric bias field correction

45
Current Solution Individual Tissue Segmentation
at each time point
Segmentation procedures Prastawa et
al., Warfield et al., Rueckert/Aljabar et al.
GM WM CSF
0.7 months
13.4 months
24.2 months
46
Results Gender Differences at birth
p 0.0030
gray matter
47
Results Early Brain Growth
Early growth mostly attributed to gray matter
development.
Knickmeyer et al., ACNP07
48
Analysis of DTI in Infants
49
Challenge Common Coordinate Frame for Population
Analysis
50
Co-registration of image sets
Not registered
51
Atlas Building
I2
I1
Atlas
Î1
I3
I5
I4
Joshi et al 2004 Goodlett et al 2006
52
Estimation of coordinate transformations
Structural Operator
Transformation (Affine, Fluid)?
Structural Average

Deformation Fields (1N)?
H-1-fields (1N)?
53
Computation of tensor means
DTI Images
Rotate Tensors based on JH-1
Tensor Averaging
DTI Atlas
H-fields (1N)?
Riemannian Symmetric Space
54
Tensor Field Deformations

• Local rotation of Jacobian Alexander et al 2001
• Log-Euclidean Tensor Statistics Arsigny et al
  2005

55
Unbiased atlas building (N150)
56
Neonate Tractography
Fornix and Uncinate
Success in Atlas of 95 neonate cases on
tracts which could not be computer in single cases
Genu, Splenium, Internal Capsule
57
Co-registration From linear to nonlinear
58
Neurodevelopmental atlas
Neonate
1 year
2 year
Adult
Neonate, 1 year, 2 year data courtesy John
Gilmore UNC Psychiatry
Adult Data courtesy Marek Kubicki Brigham and
Women's Hospital. Psychiatry Neuroimaging
Laboratory
59
Neurodevelopmental atlas
Neonate
1 year
2 year
Adult
60
Neurodevelopmental atlas
61
Neurodevelopmental atlas
62
Sample Quantitative Statistics
63
Sample Quantitative Statistics
64
From Diffusion Tensors to Connectivity?
DT-MRI measurements
65
Data Parametrization
66
Quantitative Tractography
FA along tracts
Tract ROIs
Tensors statistics along spines
FA motor tract
MD motor tract
Eigenvalues
- Tractography for ROI definition - Tensor
analysis for statistics along tracts
67
Fiber Tract Property Analysis

• Computation of pointwise mean andstandard
  deviation of diffusion properties
• Averaging of tensors in cross-sections defined by
  common length
• Calculation of tensor-derived parameters (FA, MD,
  Eigenvalues)

68
Example Uncinate Fasciculus
69
FA distributions in cross-sections
70
Combining bothPopulation-based analysis of
white matter fiber bundles
71
Fiber Tractography via Atlases
cingulate
fornix and uncinate
motor tracts
uncinate
genu, splenium, motor
Neonate atlas DTI
1yr atlas DTI
72
Mapping of Fiber Tracts
Image B
Image A
Atlas
73
Group statistics of DTI fiber bundles
74
Atlas Fiber Bundle Statistics Schizophrenia
study, Harvard
Top Curve of mean /- std. dev. of FA Left
Color display of atlas tract
75
Example curve samples
76
Functional Data Analysis

• Underlying biology of measurements is continuous
• Tract analysis samples from the continuous
  biology
• Global vs. point-wise statistics
• Smoothing
• Dimensionality Reduction
• Ramsay and Silverman 2002
• Functional Statistics

77
Functional Statistics
78
B-spline fitting

• B-spline curve fit to tensor measures
• Uniformly spaced knots
• very fine sampling for FiberViewer curves
• coarser sampling for volumetric segmentation
• Enforces smoothness
• Reduces dimensionality

79
Functional PCA

• Assume mean subtracted
• Dimensions N x \inf N x M M X \inf
• N subjects
• M basis functions

80
Functional PCA
81
Hypothesis testing and discrimination

• Permutation test using T2 statistics

• Linear discriminant embedded in T2

82
Pediatric Example

• Working example of 1 year vs. 2 year subjects
• Not intended as clinical experiment
• Significance expected
• Discrimination provides interpretation

83
Pediatric Example Genu (FA)
84
Pediatric Example Genu (Norm)
85
Pediatric Example Left motor tract
86
Pediatric Example Left motor tract
87
Pediatric Example Genu Discrimination
88
Statistical analysis of tracts as 1-D curves
functional data analysis (FDA)
Group Statistics of DTI Fiber Bundles Using
Spatial Functions of Tensor Measures Casey
Goodlett, P. Thomas Fletcher, John Gilmore, and
Guido Gerig, MICCAI 2008
89
Pediatric Example Motor Tract Discimination
90
DWI is part of multimodal MRI protocol
91
DTI is part of multi-modal MRI protocol
T1 MPRage. T2 3D TSE, MD, FA (2yrs old. Weili
Lin, UNC)
92
Quantitative Tractography to study early wm
development
John H. Gilmore et al., Early Postnatal
Development of Corpus Callosum and Corticospinal
White Matter Assessed with Quantitative
Tractography, AJNR Nov. 2007
93
Towards Longitudinal Analysis
94
Manifold Kernel Regression (B. Davis)

• What are we looking for A weighted Fréchet mean
  image as a function of age!
• Weights depend on the age

Age
motivation kernel regression manifold kernel
regression application conclusion
95
Manifold Kernel Regression (B. Davis)

• What are we looking for A weighted Fréchet mean
  image as a function of age!
• Weights depend on the age

Age
motivation kernel regression manifold kernel
regression application conclusion
96
Aging Brain via Population Shape Manifold Kernel
Regression

• B. Davis, E. Bullitt, (UNC)
• S. Joshi, T. Fletcher (Utah)
• D. Marr Prize, ICCV07 best paper award

97
Fallacy of global versus local analysis
Nonlinear growth of human brain
98
Longitudinal Study Design Normative NIH Brain
Study

• Challenges
• Mixed Cross-sectional and longitudinal design
• Missing data (1, 2 or 3 data points per subject)

99
Longitudinal changes of MR images of population
0.7
13.47"
24.2
0.7
12.8
24.4
12.6
24.8
1.3
12.6
100
Properties of data

• Correlation, similarity between repeated MR scans
• Missing Data
• Unbalanced spacing, different time points
• Correlation between tissues, inter-subject
  variance, etc.
• Multivariate features Dimensionality
• Regression not suitable

101
Challenge Multivariate Longitudinal data analysis
Multivariate Longitudinal Statistics for
Neonatal-Pediatric Brain Tissue Development, Sh.
Xu, M. Styner, J.H. Gilmore, G. Gerig, SPIE
2008 Multivariate Nonlinear Mixed Model to
Analyze Longitudinal Image Data MRI Study of
Early Brain Development, Shu Xu et al., MMBIA
2008
102
Challenge Multivariate Longitudinal data analysis
Multivariate Longitudinal Statistics for
Neonatal-Pediatric Brain Tissue Development, Sh.
Xu, M. Styner, J.H. Gilmore, G. Gerig, SPIE
2008 Multivariate Nonlinear Mixed Model to
Analyze Longitudinal Image Data MRI Study of
Early Brain Development, Shu Xu et al., MMBIA
2008
103
Conclusions

• Pediatric Imaging Image Analysis
• Amazing progress of imaging technology
• Image processing tools newly developed
• Wealth of new results
• Fascinating research area Full of discoveries
• Potential impact Better understanding ? early
  detection ? therapy
• Research field needs
• Multidisciplinary research Biology, anatomy,
  medicine, CS, statistics
• Link between MRI findings and underlying
  neurobiology
• Sharing of data and analysis tools
• Fundamental computational and statistical
  problems
• Everything changes Contrast, size, shape,
  appearance
• Statistics of growth of images and structures 4D
  statistical atlases
• (Longitudinal) multivariate statistics of imaging
  features patient parameters

104
Announcement MICCAI 2008 Workshop
105
Acknowlegements

• Clinical Research Partners
• Joseph Piven (UNC)
• John H. Gilmore (UNC)
• Janet Lainhart (Utah)
• Weili Lin (UNC)
• Computer Science Partners
• Sarang Joshi (Utah)
• Tom Fletcher (Utah)
• Marcel Prastawa (Utah)
• Sylvain Gouttard (Utah)
• Casey Goodlett (Utah)
• Shu Xun (UNC)
• Martin Styner (UNC)
• Ron Kikinis (SPL Harvard)

106
ACE Autism Network of Excellence
PI Joseph Piven, UNC Infants at risk 6mo to 2
years, longitudinal study 4 acquisition sites,
DCC, processing Utah
107

• Neonatal Brain Development in High Risk Children
  (J. H. Gilmore, MD)
• Understanding rate and variability of normal
  development
• Detect differences from typical development
• Early diagnosis ? early therapy ? help families

108
Neuroimaging for Phenotyping
Motor Tract
Motor Circuitry
UCBT
Cerebellar Peduncles
Maria Escolar, Martin Styner
109
Structure