1 UNC, Stat & OR SAMSI OODA Workshop SAMSI OODA Workshop Dyck path correspondence and the statistical analysis of Brain vascular networks Shankar Bhamidi,

1

UNC, Stat & OR

SAMSI OODA Workshop

Dyck path correspondence and the statistical analysis of Brain

vascular networksShankar Bhamidi, J.S.Marron, Dan Shen,

Haipeng Shen

UNC Chapel Hill

September 14, 2009

2

UNC, Stat & OR

Overview of today’s talk

(Very) Brief introduction to the data Dyck path or Harris correspondence between trees

and functions Modern theory of random trees Exploratory Data Analysis and implications Open problems: some incoherent thoughts

Modeling aspects: Natural probability models of spatial trees? (ISE)

Other datasets of trees?

3

UNC, Stat & OR

Basic take home messages

Last decade has witnessed an explosion in the study of Random tree models in the probability community Many different techniques, universality results Many interesting spatial models

Probability

Large amount of data from many fields Biology (brain networks, lung pathways);

Phylogenetics; “Actual trees” (root pathways) Amazing challenges at all levels (modeling,

probabilistic analysis, statistical methodology, data analysis)

Statistics

4

UNC, Stat & OR

Data Background

Motivating Example:

• From Dr. Elizabeth Bullitt• Dept. of Neurosurgery, UNC

• Blood Vessel Trees in Brains

• Segmented from MRAs

• Study population of trees

Forest of Trees

5

UNC, Stat & OR

Strongly Non-Euclidean Spaces

Trees as Data Objects

From Graph Theory:

• Graph is set of nodes and edges• Tree has root and direction and

leaves

6

UNC, Stat & OR

Blood vessel tree data

From MRA

Segment tree

vessel segments

Using tube tracking

Bullitt and Aylward (2002)

7

UNC, Stat & OR

Goal understand population properties:

PCA: Main sources of variation in the data?

Interpretation? (e.g. age, gender, occupation?)

Discrimination / Classification Prediction Models of spatial trees?

8

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

9

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

10

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

11

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

12

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

13

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

14

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

15

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

16

UNC, Stat & OR

Dyck path Correspondence for one tree

Tree 1

17

UNC, Stat & OR

Dyck Path correspondence continued

One of the foremost methods in probability for analysis of random trees.

Tremendous array of random tree models arising from many different fields e.g. CS, phylogenetics, mathematics, statistical physics

Consider a “random tree” on n vertices Rescale each edge by some factor (turns out 1/√n

is the right factor)What happens?

18

UNC, Stat & OR

Central Result

Theorem [Aldous 90’s]: For many (most?) of the known models of random trees the Dyck path converges to standard Brownian Excursion. This also implies that the trees themselves converge to a random metric space (random fractal) called the Continuum random tree.

Shall come back to this when we look at the spatial aspect.

19

UNC, Stat & OR

Basic intuition

Where does one get such results? Harris: Consider a branching process with geometric

(1/2) offspring This model is “critical” (mean # of offspring=1) Condition on size of the tree when the branching

process dies out to be n. Consider the Dyck path of this tree Has same distribution as a simple random walk

started at 0, coming back to 0 at time 2(n-1) and always above the orign otherwise

20

UNC, Stat & OR

In pictures

21

UNC, Stat & OR

In pictures

22

UNC, Stat & OR

In pictures

23

UNC, Stat & OR

In pictures

24

UNC, Stat & OR

Our data

Have data on a number of trees

Dyck path transformation for all of themExploratory Data Analysis

25

UNC, Stat & OR

Example 1, Assume that we have three following tree data

Tree 1 Tree 2 Tree 3

26

UNC, Stat & OR

Support tree: union of three tree

Tree 1 Tree 2 Tree 3

Tree 1

27

UNC, Stat & OR

Support tree: union of three tree

Tree 1 Tree 2 Tree 3

Tree 1,2

28

UNC, Stat & OR

Support tree: union of three tree

Tree 1 Tree 2 Tree 3

Tree 1,2,3

29

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

30

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

31

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

32

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

33

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

34

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

35

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

36

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

37

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

38

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

39

UNC, Stat & OR

Now, we show how to transform the first tree as curve.

Tree 1/ Support Tree

40

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

41

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

42

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

43

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

44

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

45

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

46

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

47

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

48

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

49

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

50

UNC, Stat & OR

Now, we show how to transform the second tree as curve.

Tree 2/ Support Tree

51

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

52

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

53

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

54

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

55

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

56

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

57

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

58

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

59

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

60

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

61

UNC, Stat & OR

Now, we show how to transform the third tree as curve.

Tree 3/ Support Tree

62

UNC, Stat & OR

Advantages of this encoding

If we are only interested in topological aspects then mathematically this is reasonable

Main reason: Suppose f, g are encodings of two trees, s and t, then the sup norm between the two functions bounds the Gromov-Haussdorf distance

However a number of issues as well

63

UNC, Stat & OR

Actual Data

64

UNC, Stat & OR

Raw Brain Data - Zoomed

65

UNC, Stat & OR

Raw Brain Data - Zoomed

66

UNC, Stat & OR

Some Brain Data Points(as corresponding trees)

67

UNC, Stat & OR

Some Brain Data Points(as corresponding trees)

68

UNC, Stat & OR

Some Brain Data Points(as corresponding trees)

69

UNC, Stat & OR

Some Brain Data Points(as corresponding trees)

70

UNC, Stat & OR

Some Brain Data Points(as corresponding trees)

71

UNC, Stat & OR

Some Brain Data Points(as corresponding trees)

72

UNC, Stat & OR

Data Representation- Youngest

73

UNC, Stat & OR

Data Representation- oldest

74

UNC, Stat & OR

Average Tree-Curve and picture of the average tree

75

UNC, Stat & OR

Illust’n of PCA View: PC1 Projections

76

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

77

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

78

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

79

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

80

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

81

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

82

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

83

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

84

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC1 direction

85

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC2 direction

86

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC2 direction

87

UNC, Stat & OR

PCAPictures of trees that we get when we move in PC2 direction

88

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC2 direction

89

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC2 direction

90

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC2 direction

91

UNC, Stat & OR

PCA Pictures of trees that we get when we move in PC2 direction

92

UNC, Stat & OR

PCAPictures of trees that we get when we move in PC2 direction

93

UNC, Stat & OR

PCAPictures of trees that we get when we move in PC2 direction

94

UNC, Stat & OR

DWD

95

UNC, Stat & OR

DWD

96

UNC, Stat & OR

DWD

97

UNC, Stat & OR

DWD

98

UNC, Stat & OR

DWD

99

UNC, Stat & OR

DWD

100

UNC, Stat & OR

DWD

101

UNC, Stat & OR

DWD

102

UNC, Stat & OR

DWD

103

UNC, Stat & OR

DWD/relabeling

random relabeling: Suppose we randomly relabel each tree as male or female.

How does the DWD direction behave?

104

UNC, Stat & OR

DWD/relabelling

105

UNC, Stat & OR

DWD/relabelling

106

UNC, Stat & OR

DWD/relabelling

107

UNC, Stat & OR

DWD/relabelling

108

UNC, Stat & OR

DWD/relabelling

109

UNC, Stat & OR

DWD/relabelling

110

UNC, Stat & OR

DWD/relabelling

111

UNC, Stat & OR

DWD/relabelling

112

UNC, Stat & OR

DWD/relabelling

113

UNC, Stat & OR

Implications

“Eyeballing” the data, the PC1 directions (and PC2) do not seem to be capturing variation in the data

Because of encoding all the trees to form a support tree?

Perhaps because inherently PCA works well in the Euclidean regime?

Path of Dyck paths a weird subset of function space? Any math theory that can be developed about

families of large trees? Modeling of these trees?

114

UNC, Stat & OR

Blood vessel tree data

Marron’s brain:

From MRA

Segment tree

of vessel segments

Using tube tracking

Bullitt and Aylward (2002)

115

UNC, Stat & OR

Blood vessel tree data

Marron’s brain:

From MRA

Reconstruct trees

in 3d

Rotate to view

116

UNC, Stat & OR

Blood vessel tree data

Marron’s brain:

From MRA

Reconstruct trees

in 3d

Rotate to view

117

UNC, Stat & OR

Blood vessel tree data

Marron’s brain:

From MRA

Reconstruct trees

in 3d

Rotate to view

118

UNC, Stat & OR

Blood vessel tree data

Marron’s brain:

From MRA

Reconstruct trees

in 3d

Rotate to view

119

UNC, Stat & OR

Blood vessel tree data

Marron’s brain:

From MRA

Reconstruct trees

in 3d

Rotate to view

120

UNC, Stat & OR

Blood vessel tree data

Marron’s brain:

From MRA

Reconstruct trees

in 3d

Rotate to view

121

UNC, Stat & OR

Thoughts I: Probabilistic models of spatial trees?

What are natural models of spatial trees such as those in this talk?

At least two natural directions to proceed in ISE (Integrated Superbrownian Excursion): Arising

from modelling of critical random systems in euclidean space

Engineering and biological principles of flow distribution: (Constructal theory)

122

UNC, Stat & OR

ISE: Integrated Superbrownian excursion

Formulated in the late 90s by Aldous Has now come to be one of the standard models of

spatial trees Arises as the scaling limit of many different systems Example: Random trees on the integer lattice Critical contact process in high dimensions etc Thought to be the scaling limit of many systems at

criticality

Use Standard Brownian excursion and Brownian motion to construct a random tree in 3 (or higher dimensions)

123

UNC, Stat & OR

ISE: in pictures

124

UNC, Stat & OR

ISE in pictures

125

UNC, Stat & OR

ISE in pictures

126

UNC, Stat & OR

ISE in pictures

127

UNC, Stat & OR

ISE

Any notion of data driven ISE?

128

UNC, Stat & OR

Blood vessel tree data

Notion of ISE on the

sphere?

Notion of ISE where the

Brownian motion has

some sort of drift?

How does one estimate

drift from data?

Model of thickness on

edges to the data?

129

UNC, Stat & OR

Other examples of tree data?

130

UNC, Stat & OR

Data on actual root systems

131

UNC, Stat & OR

PCA and Random Walks on Tree space?

In this study we tried usual notion of PCA

Ok when data are “Gaussian in nature”

Tree space intuitively very non-linear

Can one use random walks to explore this space?

132

UNC, Stat & OR

Intuition

133

UNC, Stat & OR

Random walk on data points

134

UNC, Stat & OR

Folded Euclidean Approach