7. Previous Automatic Methods Ray et al. [2002] Active Contour Cui et al. [2005] Monte Carlo Mukherjee et al. [2004] Level Set Analysis 6
8. Previous Automatic Methods Eden et al. [2005] SmoothnessConstraints Li et al. [2005] Lineage Construction Smith et al. [2008] Probabilistic Formalization 7
11. Challenges In a collision, cell motion and appearance 1. could be different 2. change abruptly 10
12. Example Eden et al. [2005] broken tracks robust tracks Our Method 11
13. Approach To improve tracking accuracy of colliding cells by: having separate collision states to describe cells inside and outside of collisions testing multiple hypotheses of cell motion and appearance as transitions between abrupt motion patterns. 12
14. Cell Detection Classify each pixel in the image as a Cell or Background 13
15. Detection Problems 14 Variation in cell appearances within an image time Varied appearance of a cell over time
16. AdaBoost 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.
17. Training 100 Cell Samples 100 Background Samples Features Mean Intensity Standard Dev. of Intensity Radial Mean Decision Rules on feature scores 16
18. Detection Procedure Scan each pixel p in the image Compute image feature vector V from a window centered around p Classify p as a Cell pixel if the feature score in V satisfies the learned decision rule; otherwise classify p as a Background pixel. Cluster groups of Cell pixels into cell observation. 17
19. Cell Tracking 1. Predict using multiple hypotheses 2. Correspond predictions and measurements 3. Update based on the current state 18
20.
21.
22. Motion and Appearance Model 21 Collision States: State Transition: Hypotheses: to predict the state in the next frame control input vector State Vector*: Motion and Appearance Models: (for ) state transition matrix control input matrix process noise vector ~N(0,Qs) Observation Vector:
37. Tracking Performance 36 RMSE : Root mean squared errors of position (pixel) -0.03 -0.17 +0.36 -0.21 +0.33 -0.20 SH introduces additional error in positions. MH does not introduce any additional error. Estimating colliding cells’ positions is more difficult.
38. Tracking Performance 37 PTP: Percentage of Tracked Positions (%) +27 +9 +23 +3 +4 +24 +28 Large improvement in colliding positions. Improvement overall. Tracking colliding cells is more difficult.
40. Discussions 39 The effect of collision duration on tracking 6 112 Exclude SC from being considered for collision. Classify colliding positions into bins based on the number of frames of the collision. colliding cells bins of collision duration
43. Discussions 42 The impact of detection on tracking 38 596 Data with good detection results before and after collision (+/- 2 frames) cell positions treated colliding cells
44. Discussions 43 The impact of detection on RMSE -0.17 -0.13 -1.05 -1.09 Different improvement between dataset. Different improvement between methods.
45. Discussions 44 The impact of detection on PTP +9 +7 +16 +18 Large improvement between dataset. Large different between methods. MH achieves high performance in tracking.
47. Future Work 1. Add more features to improve detection. 46 5 7 6 8
48. Future Work 2. Incorporate a probabilistic approach to transition between collision states. 47 72 73 75 76
49. Future Work 3. Expand to track cells with more complex motions and behaviors. 48 49 50 51 52
50. Conclusion 49 A method for tracking colliding cells. Incorporate Kalman filter and multiple hypotheses for each collision state. Improve 28%in tracked position coverage compared to a previous work . Achieve 88% in tracked position coverage in tracking colliding cells.
52. 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”. Related Publications 51
54. 53 Tracking Steps Predict motion Predict collision Get measurements Get errors in position & area Match with minimal error
55. Collision States: Hypotheses: to predict the state in the next frame control input vector State Vector*: Motion and Appearance Models: (for ) state transition matrix control input matrix process noise vector ~N(0,Qs) Observation Vector: Measurement Model: measurement noise vector ~N(0,R) measurement transition matrix 54
56. State Vector of cell i : Predicted State Vector: Zero Matrix Zero Matrix Zero Vector Zero Vector Predicted Covariance: 55
57. z1 x’1a Calculate error between all possible pairs Sm & On Criteria: Position & Area. Assign pairs via Greedy Search HФ 2. Correspondence X1 x’1b HФ Cell Detector z2 x'2a HФ X2 HФ x’2b z3 x’3a HФ X3 HФ x’3b zn X1 x’ka HФ z1 x’1a Xk HФ x’kb X2 x’2b z3 Cell 3 is missing x’3a X3 New cell has entered z2 X4 Ck zn x’kb 56
58. 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
59.
60. 59 II. AdaBoost1. Adaptive Boosting method 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.
76. Discussions 75 The impact of detection on PTP +7 +16 +11 Large improvement between dataset. Large different between methods. MH achieves high performance in tracking.
77. Training 100 Cell Samples 100 Background Samples Features Mean Intensity Standard Dev. of Intensity Radial Mean Decision Rules on feature scores 76