Wbc master thesisdefense

77

Transcript of Wbc master thesisdefense

Page 1: Wbc master thesisdefense
Page 2: Wbc master thesisdefense

Protect us against infection

Lookout for signs of disease

Microscopic scale: 7,000 to 25,000 in a drop of blood

2

“The little warriors of your body”

Page 3: Wbc master thesisdefense

1. Flowing

2. Rolling

3. Adhering

3Microscopy Video of white blood cells

Page 4: Wbc master thesisdefense

Biologists learn cell behaviors by tracking a few cells

4

Page 5: Wbc master thesisdefense

Tedious

Expensive

Subjective

Little Effort

Inexpensive

Objective

5

Page 6: Wbc master thesisdefense

Ray et al. [2002]

Active Contour

Cui et al. [2005]

Monte Carlo

Mukherjee et al. [2004]

Level Set Analysis

6

Page 7: Wbc master thesisdefense

Eden et al. [2005]SmoothnessConstraints

Li et al. [2005]Lineage

Construction

Smith et al. [2008]Probabilistic

Formalization

7

Page 8: Wbc master thesisdefense

8

Ray Cui Mukherjee Eden Smith Li Ours

Tracker Type S S M M M M M

Presentation C C L P P P P

Entries/ Exits

Occlusion

Various Motions

Collision

Optimal Solution

S : single cell M : multiple cells C: contour L: level set P: point

Page 9: Wbc master thesisdefense

As many cells move at a wide range of speed

9

Page 10: Wbc master thesisdefense

In a collision, cell motion and appearance

1. could be different

2. changeabruptly

10

Page 11: Wbc master thesisdefense

Eden et al. [2005]

Our Method11

broken tracks

robust tracks

Page 12: Wbc master thesisdefense

having separate collision statesto describe cells inside and outside of collisions

testing multiple hypothesesof cell motion and appearance as transitions between abrupt motion patterns.

To improve tracking accuracy of colliding cells by:

12

Page 13: Wbc master thesisdefense

Classify each pixel in the image as a Cell or Background13

Page 14: Wbc master thesisdefense

14

Variation in cell appearances within an image

Varied appearance of a cell over time

time

Page 15: Wbc master thesisdefense

15

Idea: combine many “rules of thumb” to a highly accurate prediction rule.

Input: visual features from training samples.

Schema: maintain a strategy to determine “rules of thumb” using weight distribution.

Output: a single strong classifier which is a linear combination of the set of weak classifiers.

Page 16: Wbc master thesisdefense

1. Mean Intensity2. Standard Dev. of Intensity3. Radial Mean

Decision Rules on feature scores16

Page 17: Wbc master thesisdefense

1. Scan each pixel p in the image

2. Compute image feature vector V from a window centered around p

3. Classify p as a Cell pixel if the feature score in Vsatisfies the learned decision rule; otherwise classify p as a Background pixel.

4. Cluster groups of Cell pixels into cell observation.

17

Page 18: Wbc master thesisdefense

1. Predictusing multiple hypotheses

2. Correspondpredictions and measurements

3. Updatebased on the current state

18

Page 19: Wbc master thesisdefense

19

1. Extensively used for tracking.

2. Estimator of the state that is optimal.

3. Consist of two steps:

• Predict: projects next time step state.

• Correct: incorporates a new measurement into a priori estimate to obtain an improved posteriori estimate.

Page 20: Wbc master thesisdefense

Kalman ?

smooth abrupt

H00: No collision – No collisionH11: Collision – Collision

H01: No collision – CollisionH10: Collision – No collision

Collision States 20

Page 21: Wbc master thesisdefense

21

State Vector*:

Observation Vector:

Collision States:

Hypotheses:

State Transition:

Motion and Appearance Models:

state transitionmatrix

control input matrix

process noise vector

~N(0,Qs)

(for )

control input vector

to predict the state in the next frame

Page 22: Wbc master thesisdefense

Multiple Hypotheses

No Collision-No Collision

H00

No Collision

– Collision

H01

Collision –Collision

H11

Collision –No Collision

H10

Motion

Position KalmanCollision location

From collision

speed

From restoredvelocity

Velocity KalmanApproaches

zeroApproaches

zeroRestore to

pre-collision

Appearance Area ConstantSum of

colliding cells

Sum of colliding

cells

Restore to pre-collision

22

Page 23: Wbc master thesisdefense

23

0. Initial state

Page 24: Wbc master thesisdefense

24

1. Predict motion

Page 25: Wbc master thesisdefense

25

2. Predict collision

Page 26: Wbc master thesisdefense

26

3. Get measurements

Page 27: Wbc master thesisdefense

27

4. Get errors in position & area

Page 28: Wbc master thesisdefense

28

5. Correspond predictions to measurements

no measurement

no predictioncorrespond to H00

correspond to H01

Page 29: Wbc master thesisdefense

29

6. Update

lost cell

new cell

non-colliding cell

colliding cells

Page 30: Wbc master thesisdefense

30

7. Repeat for the next time step

stay colliding

split away

keep moving

start moving

keep moving

Page 31: Wbc master thesisdefense

31

Page 32: Wbc master thesisdefense

32

Page 33: Wbc master thesisdefense

33

Page 34: Wbc master thesisdefense

34

1. Smoothness Constraints (SC)

2. Single Hypothesis (SH)

3. Multiple Hypotheses (MH)

Same detection results

Same blood flow directions

Same tracking parameters (SH & MH)

Page 35: Wbc master thesisdefense

35

Recall : TP / (TP + FN)

Precision: TP / (TP + FP)

Page 36: Wbc master thesisdefense

36

Dataset SC SH MH

Colliding positions 7.11 7.08 6.91

Non-colliding positions 5.63 5.99 5.78

All positions 5.82 6.15 5.95

RMSE : Root mean squared errors of position (pixel)

1. SH introduces additional error in positions.

2. MH does not introduce any additional error.

3. Estimating colliding cells’ positions is more difficult.

-0.03 -0.17

+0.36 -0.21

+0.33 -0.20

Page 37: Wbc master thesisdefense

37

Dataset SC SH MH

Colliding positions 36 63 72

Non-colliding positions 53 76 79

All positions 49 73 77

PTP: Percentage of Tracked Positions (%)

1. Large improvement in colliding positions.

2. Improvement overall.

3. Tracking colliding cells is more difficult.

+27 +9

+23 +3

+24 +4

+28

Page 38: Wbc master thesisdefense

MH SC SH

38

Page 39: Wbc master thesisdefense

39

The effect of collision duration on tracking

1. Exclude SC from being considered for collision.

2. Classify colliding positions into bins based on the number of frames of the collision.

Page 40: Wbc master thesisdefense

40

The effect of collision duration on RMSE

Page 41: Wbc master thesisdefense

41

The effect of collision duration on PTP

Page 42: Wbc master thesisdefense

42

The impact of detection on tracking

Data with good detection results before and after collision (+/- 2 frames)

Page 43: Wbc master thesisdefense

43

The impact of detection on RMSE

Dataset SH MH

Colliding positions 7.08 6.91

Colliding positions with good detection

6.95 5.86

-0.13 -1.05

-0.17

-1.09

1. Different improvement between dataset.

2. Different improvement between methods.

Page 44: Wbc master thesisdefense

44

The impact of detection on PTP

Dataset SH MH

Colliding positions 63 72

Colliding positions with good detection

70 88

+7 +16

+9

+18

1. Large improvement between dataset.

2. Large different between methods.

3. MH achieves high performance in tracking.

Page 45: Wbc master thesisdefense

Our Method

45

Page 46: Wbc master thesisdefense

1. Add more features to improve detection.

46

Page 47: Wbc master thesisdefense

2. Incorporate a probabilistic approach to transition between collision states.

47

Page 48: Wbc master thesisdefense

3. Expand to track cells with more complex motions and behaviors.

48

Page 49: Wbc master thesisdefense

49

1. A method for tracking colliding cells.

2. Incorporate Kalman filter and multiple hypotheses for each collision state.

3. Improve 28% in tracked position coverage

compared to a previous work .

4. Achieve 88% in tracked position coverage in

tracking colliding cells.

Page 50: Wbc master thesisdefense

50

Page 51: Wbc master thesisdefense

S. J. Schmugge, S. Keller, N. Nguyen, R. Souvenir, T. H. Huynh, M. Clemens, M. C. Shin. "Segmentation of Vessels Cluttered with Cells using a Physics based Model". 11th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), New York, September 6-10 2008.

N. Nguyen, S. Keller, T. Huynh, M. Shin. “Tracking Colliding Cells”. IEEE Workshop on Applications of Computer Vision (WACV), Snowbird, UT December 07-09 2009.

to be submitted to IEEE Transactions on Medical Imaging

N. Nguyen, S. Keller, Eric Norris, T. Huynh, M. Shin. “Tracking Colliding Cells in Intravital Microscopy Images”.

51

Page 52: Wbc master thesisdefense

52

Eden et al. [2005]

Our Method

Page 53: Wbc master thesisdefense

53

1. Predict motion

2. Predict collision

3. Get measurements

4. Get errors in position & area

5. Match with minimal error

Page 54: Wbc master thesisdefense

State Vector*:

Observation Vector:

Collision States:

Hypotheses:

Motion and Appearance Models:

state transitionmatrix

control input matrix

process noise vector

~N(0,Qs)

(for )

control input vector

to predict the state in the next frame

Measurement Model:

measurement transition matrix

measurement noise vector

~N(0,R)54

Page 55: Wbc master thesisdefense

Predicted State Vector:

Predicted Covariance:

H00 H01 H11 H10

A

B

*

State Vector of cell i :

Zero Matrix Zero Matrix

Zero Vector Zero Vector

55

Page 56: Wbc master thesisdefense

X1

X2

X3

Xk

Cell Detector

z1

z3

zn

z2

x’1a

x’1b

x'2a

x’2b

x’3a

x’3b

x’ka

x’kb

z1x’1a

x’2b z3

znx’kb

x’3a

z2

Cell 3 is missing

New cell has entered

X1

X2

X3

Ck

X4

Calculate error between all possible pairs

Sm & On

Criteria: Position & Area.

Assign pairs via Greedy Search

2. Correspondence

56

Page 57: Wbc master thesisdefense

Predicted State Vector:

Hypothesized Measurement Vector:

measurement transition matrix

Error of hypothesis :

observation from the detector

weight vector

Rule 1:

Rule 2:

error threshold of Unlikely (i, k) pair

Stop corresponding condition:

57

Page 58: Wbc master thesisdefense

Remaining Observation : new cell

leukocyte typicaldiameter

area

Remaining Cell :

Not corresponded for 3 frames:

Updated State Vector :

Kalman gain

Updated Covariance:

depends on the cell current state s Eliminate the affect cause by abrupt change in collision

58

Page 59: Wbc master thesisdefense

59

Robert Schapire algorithm (1996).

Idea: Combine many “rules of thumb” a highly accurate

prediction rule.

Maintain a strategy to determine “rules of thumb” weight.

Terms: Learner = Hypothesis = Classifier.

Weak Learner: <50% error rate.

Strong Learner: linear combination of weak learner.

Page 60: Wbc master thesisdefense

60

Plastic AppleReal Apple

Page 61: Wbc master thesisdefense

61

Weight

Color

Page 62: Wbc master thesisdefense

62

Weak Learner

Optimal Guessing line

Wrong guess

Wrong

guess

Page 63: Wbc master thesisdefense

63

Right guess

Gets smaller

Get smaller

Wrong guess

Gets larger

Page 64: Wbc master thesisdefense

64

More weighted ones

get classify correctly

Wrong guess

Wrong guess

Page 65: Wbc master thesisdefense

65

Wrong guess

Gets larger

Right guess

Gets smaller

smaller

smaller

Page 66: Wbc master thesisdefense

66

Page 67: Wbc master thesisdefense

67

Page 68: Wbc master thesisdefense

68

1st

2nd

3rd

4th

Page 69: Wbc master thesisdefense

69

Decision

Boundary

Real Apple

Region

Plastic Apple

Region

Page 70: Wbc master thesisdefense

70

4. Get errors in position & area

Page 71: Wbc master thesisdefense

71

7. Repeat for the next time step

Page 72: Wbc master thesisdefense

NoCollision-

CollisionDuring

CollisionPost

Collision

Position KalmanCollision location

From collision

speed

From restoredvelocity

Velocity KalmanApproaches

zeroApproaches

zeroRestore to

pre-collision

Area ConstantSum of

colliding cells

Sum of colliding

cells

Restore to pre-collision

72

Page 73: Wbc master thesisdefense

73

Page 74: Wbc master thesisdefense

74

Dataset SC SH MH

All cells .49 .73 .77

Non-colliding cells .53 .76 .79

Colliding cells .40 .63 .72

Treated colliding cells .51 .70 .88

+.24

+.23

+.23

+.21

+.04

+.03

+.09

+.18

Page 75: Wbc master thesisdefense

75

The impact of detection on PTP

Dataset SC SH MH

Colliding positions 40 63 72

Colliding positions with good detection

51 70 88

+7 +16

1. Large improvement between dataset.

2. Large different between methods.

3. MH achieves high performance in tracking.

+11

Page 76: Wbc master thesisdefense

1. Mean Intensity2. Standard Dev. of Intensity3. Radial Mean

Decision Rules on feature scores76

Page 77: Wbc master thesisdefense

77

Ray[2004]

Cui[2006]

Eden[2005]

Smith[2008]

Li[2008]

Dzyubachyk

[2010]

Ours[2010]

Tracker Type Single Single Multiple Multiple Multiple Multiple Multiple

Imaging Modality in-vivo in-vivo in-vivo ex-vivo in-vitro in-vitro in-vivo

Entries/ Exits

Occlusion

Various Motions

Collision