Transcript Mining Structure-Function Associations in a Brain Image Database

Mining Structure-Function Associations
in a Brain Image Database
Vasileios Megalooikonomou
Department of Computer Science
Dartmouth College
BRAID: Brain-Image Database
Nick Bryan
Christos Davatzikos
Joan Gerring
Edward Herskovits
Vasileios Megalooikonomou
What is data mining?
• Now that we have gathered so much data,
what do we do with it?
• Extract interesting patterns (automatically)
• Associations (e.g., butter + bread --> milk)
• Sequences (e.g., temporal data related to stock market)
• Rules that partition the data (e.g., store location problem)
• What patterns are “interesting”?
information content, confidence and support,
unexpectedness, actionability (utility in decision making)
Overview
•
•
•
•
•
Goals
Background
Methods
Results
Discussion - Future Work
Goals
•
•
•
•
Structure-function correlation
Decoupling of signal and morphology
Scalability (large longitudinal studies)
Transparent management of diverse data sources
Background
•
•
•
•
•
Illustra Object-Relational DBMS
Image datablade
Web interface
Lesions identified manually
Images registered to a common spatial standard
(Talairach atlas)
• Clinical information and images are integrated
• Clinical studies (CHS, FLIC, BLSA)
Background: Spatial Normalization of Brain Images
Before
Spatial
Normalization
After
Spatial
Normalization
Background: Spatial Normalization: Example
• 3D elastically deformable model (Davatzikos, 1997)
original
Deform MRI
to Talairach atlas
target
deformed
Background: Talairach Atlas
Background: Gyri Atlas
Background: Sample SQL queries
•
•
•
COMPUTE VOLUME OF A GIVEN STRUCTURE
return volume((select unique image from structures
where side='Left' and atlas='Brodmann' and name='17')) ;
•
•
•
•
•
DISPLAY GIF OF ALL LESIONS SUMMED UP
insert into temp_image_1 values(permanent(map_image(sum_images((
select image from patient_images where image.description='All Lesions')), 'redgreenscale'))) ;
select TS.SliceNo, slice(TS.SliceNo,overlay.image)::GIF as LesionDensity
from TalairachSlices TS, temp_image_1 overlay order by SliceNo ;
Methods
•
•
•
•
•
Segmentation
Registration
Integration into BRAID
Visualization
Statistical analysis
BRAID: Flow of Information
MRI
Registered
Lesions
Atlas
Clinical
Data
Image
Segmentation
Lesions
Image
Registration
Structure-Function
Association Analysis
Methods: Visualization: FLIC study
Sum of lesions for the ADHD- and ADHD+ groups
ADHD(n=61)
ADHD+
(n=15)
Tal-107
Tal-113
Tal-116
Tal-119
Tal-124
SQL query: Sum of lesions for ADHD subjects
•
•
•
•
•
•
•
•
•
•
•
•
•
•
insert into temp_image_1 values(permanent(
map_image(sum_images((select image from patient_images where
image.description='All Lesions' and patient in
(select patient from attributes where varname='ADHD_GRP' and
real_value=2 and patient like 'FLIC%'))),
'redgreenscale')
+
map_image((select unique image from structures where side='Left' and
atlas='Talairach' and name='cortex') + (select unique image from structures
where side='Right' and atlas='Talairach' and name='cortex'), 'bluescale')
+
map_image((select unique image from structures where side='Right' and
atlas='Talairach' and name='putamen'),
'redscale')
+
map_image((select unique image from structures where side='Left' and
atlas='CHS' and name='thalamus'),
'greenscale')));
select TS.SliceNo, slice(TS.SliceNo,overlay.image)::GIF as LesionDensity
from TalairachSlices TS, temp_image_1 overlay order by SliceNo ;
Methods: Statistical Analysis
•Atlas based
•Map each lesion onto at least one atlas structure
•Prior knowledge increases the sensitivity of spatial analysis
•Marked data reduction: 107 voxels
102 structures
•Structural variables: categorical or continuous
•Atlas free (voxel-based)
•No model on the image data
•Cluster voxels by functional association
Methods: Statistical: Atlas Based
• F functional variables, S anatomical structures
• Analysis
• Categorical structural variables
• Exploratory
• F x S contingency tables, Chi-square/Fisher exact test
• multiple comparison problem
• log-linear analysis, multivariate Bayesian
• Directed using visualization, prior knowledge
• small number of hypotheses to test
• no multiple comparison problem
• Continuous structural variables
• Logistic regression, Mann-Whitney
Methods: Statistical: Chi-square
• 2 x 2 contingency tables for categorical variables
• Pearson chi-square
Methods: Statistical: Voxel-based: Logistic Regression
• logit d   log oddsd     1 x1     k xk
where oddsd 
pd .
1  pd .
• Identify “causal brain region” that best discriminates affected/unaffected
subjects
• logit d   af  b
where
• f = volume(intersect(Lesion, Sphere)) / volume(Sphere)
• d = deficit (e.g., hemiparesis)
• a = log odds / lesioned fraction of sphere volume
• b = prior log odds of d
• Optimize sphere parameters x, y, z, r
Results: Atlas based: FLIC study- ADHD
Structural
Variable
Fisher’s Exact
p-value
Mann-Whitney
p-value
Right Putamen
0.065
0.033
Left Thalamus
0.095
0.093
Right Caudate
0.168
0.115
Left Putamen
0.670
0.824
Results: Atlas based: CHS study
Structure
Function
R globus pallidus
L hippocampus
R gyri angular
R gyri orbital
R gyri cuneus
R optic tract
R hemiparesis
R visual defect
L pronator drift
L visual defect
L visual defect
L pronator drift
Chi-square S-Bonf. Correct.
p-value
p-value
0.00001
0.00001
0.00002
0.00003
0.00003
0.00003
0.0039
0.0095
0.0195
0.0224
0.0224
0.0224
Results: Voxel-based: FLIC study
Results: Voxel based: 3D reconstruction: FLIC study
Results: Voxel-based Regression Analysis
ADHD+
ADHD-
Optimal_Regression_Sphere
Methods: Validation
•Objective: to evaluate BRAID’s analytical capabilities
•Problems: not enough subjects, true assocs unknown,
registration error
•Approach:
•Lesion-Deficit Simulator (LDS) + Monte Carlo analysis
•measure effect of strength of assocs, model complexity,
registration error, statistical power of tests
•Application: a test-bed for development and evaluation
of S-F correlation methods
Validation: Background
• Bayesian Network Model for S-F associations
• Consider 3 cases for cond. prob. table, noisy-OR model
case description deficit cond. probs. struct1 struct2 p(func=normal)
1
2
3
strong
moderate
weak
0/1
0.25 / 0.75
0.49 / 0.51
N
N
A
A
N
A
N
A
0.75
0.25
0.25
0.06
Validation: Lesion-Deficit Simulator (LDS)
• For each subject p
• produce lesions:
• obtain params for lesion size, number, spatial distr.
• construct pdfs
• produce simulated lesions given the pdfs
• model registration error
• estimate 3D Gaussian using landmarks
• produce displacements of lesion centroids
• find lesioned structures and priors of abnormality
• use fraction of lesioned volume and threshold
S
• Sample priors for abnormality of structures and produce S p
• Generate BN model of assocs among S-F
• For each subject p instantiate S-nodes to produce F Fp
Results: Simulator
Results: Simulator
Results: Simulator
Results: Simulator
• N is inversely proportional to the smallest
prior/conditional probability
• The degree of assocs affects more the performance than
the number of assocs
• On average 87% of assocs were found in registered
images compared with perfect registration
Discussion - Future Work
•
•
•
•
•
•
neural-network and other non-statistical models
bayesian multivariate analysis
more complex spatial models
increase number of subjects in BRAID
automate methods for image segmentation
statistical analysis of morphological variability
Analysis, Classification and Visualization of Probabilistic 3D Objects
For more information...
• www.cs.dartmouth.edu/~vasilis, braid.rad.jhu.edu
• V. Megalooikonomou, C. Davatzikos, E. Herskovits, “Mining Lesion-Deficit Associations
in a Brain Image Database”, ACM SIGKDD, Aug. 1999, San Diego, CA, pp. 347-351.
• V. Megalooikonomou, C. Davatzikos, E. Herskovits, “A Simulator for Evaluation of
Methods for the Detection of Lesion-Deficit Associations”, Human Brain Mapping, in
press.
• V. Megalooikonomou and E. Herskovits, “Mining Structure-Function Associations in a
Brain Image Database”, chapter in Medical Data Mining and Knowledge Discovery, K. J.
Cios (ed.), Springer-Verlag, to appear in 2000.
• V. Megalooikonomou, J. Ford, L. Shen, F. Makedon, “Data Mining in Brain Imaging”,
Statistical Methods in Medical Research, to appear (invited paper).
• E. H. Herskovits, V. Megalooikonomou, C. Davatzikos, A. Chen, R. N. Bryan, J. Gerring,
“Is the spatial distribution of brain lesions associated with closed-head injury predictive of
subsequent development of attention-deficit hyperactivity disorder? Analysis with brain
image database”, Radiology, Vol. 213, No. 2, pp. 389-394, 1999.