Vasileios Megalooikonomou

1
Mining Structure-Function Associations in a
Brain Image Database
Vasileios Megalooikonomou
Department of Computer Science
Dartmouth College
2
BRAID Brain-Image Database
Nick Bryan Christos Davatzikos Joan
Gerring Edward Herskovits Vasileios
Megalooikonomou
3
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 --gt 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)
4
Overview

• Goals
• Background
• Methods
• Results
• Discussion - Future Work

5
Goals

• Structure-function correlation
• Decoupling of signal and morphology
• Scalability (large longitudinal studies)
• Transparent management of diverse data sources

6
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)

7
Background Spatial Normalization of Brain Images
Before Spatial Normalization
After Spatial Normalization
8
Background Spatial Normalization Example

• 3D elastically deformable model (Davatzikos, 1997)

deformed
original
target
Deform MRI to Talairach atlas
9
Background Talairach Atlas
10
Background Gyri Atlas
11
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_i
  mage(sum_images((
• select image from patient_images where
  image.description'All Lesions')),
  'redgreenscale')))
• select TS.SliceNo, slice(TS.SliceNo,overlay.ima
  ge)GIF as LesionDensity
• from TalairachSlices TS, temp_image_1
  overlay order by SliceNo

12
Methods

• Segmentation
• Registration
• Integration into BRAID
• Visualization
• Statistical analysis

13
BRAID Flow of Information
MRI
Atlas
Registered Lesions
Clinical Data
Image Segmentation
Structure-Function Association Analysis
Image Registration
Lesions
14
Methods Visualization FLIC study
Sum of lesions for the ADHD- and ADHD groups
ADHD-
(n61)
ADHD
(n15)
Tal-113
Tal-116
Tal-119
Tal-124
Tal-107
15
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'Al
  l Lesions' and patient in
• (select
  patient from attributes where varname'ADHD_GRP'
  and
  
• 
  real_value2 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

16
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

17
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

18
Methods Statistical Chi-square

• 2 x 2 contingency tables for categorical
  variables

• Pearson chi-square

19
Methods Statistical Voxel-based Logistic
Regression
where

Identify causal brain region that best
discriminates affected/unaffected subjects

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

20
Results Atlas based FLIC study- ADHD

• Structural Fishers Exact Mann-Whitney
• Variable p-value 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

21
Results Atlas based CHS study
Chi-square p-value
S-Bonf. Correct. p-value
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
0.00001 0.00001 0.00002 0.00003 0.00003 0.00003
0.0039 0.0095 0.0195 0.0224 0.0224 0.0224
22
Results Voxel-based FLIC study
23
Results Voxel based 3D reconstruction FLIC
study
24
Results Voxel-based Regression Analysis
ADHD
ADHD-
Optimal_Regression_Sphere
25
Methods Validation

• Objective to evaluate BRAIDs 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

26
Validation Background

• Bayesian Network Model for S-F associations

• Consider 3 cases for cond. prob. table, noisy-OR
  model

struct1
struct2
p(funcnormal)
case
description
deficit cond. probs.
N N A A
N A N A
0.75 0.25 0.25 0.06
1 2 3
strong moderate weak
0 / 1 0.25 / 0.75 0.49 / 0.51
27
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

F

• For each subject p instantiate S-nodes to
  produce F

p
28
Results Simulator
29
Results Simulator
30
Results Simulator
31
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

32
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

33
Analysis, Classification and Visualization of
Probabilistic 3D Objects
34
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.