1. Dynamic Time Warping for Automated Cell Cycle Labelling
A. El-Labban, A. Zisserman
University of Oxford
Y. Toyoda, A. Bird, A. Hyman
Max Planck Institute of Molecular Cell Biology and Genetics

2. Objectives Segment and track mitotic cells Label mitotic phases
Fully automated system
Interphase
Prometaphase
Anaphase
Prophase
Metaphase
Telophase

3. Data 3D time lapse image stacks Use max intensity z-projections
1-5 minute temporal resolution
0.2 micron xy-resolution

4. Approach
Existing approaches (e.g. Harder et al. 2009, Held et al [CellCognition]):
Track cells
Label cell cycle phase frame-by-frame
Smooth result with HMM (CellCognition)
Our Approach:
Label all frames by using temporal signals of features

5. Temporal signals of features

6. Temporal signals of features
Interphase
Prometaphase
Anaphase
Prophase
Metaphase
Telophase

7. Overview
Part I
Track cells in videos
Part II
Label mitotic phases

8. Part I – Tracking

9. Tracking Tracking by detection Detect first, then associate objects
Here we use detection by classification.

10. Segmentation: Our approach
Logistic regression classifier
Graph Cuts
Logistic regression
classifier
Graph Cut
Input image
Probability map
Binary map

11. Logistic Regression Classifier
Feature:
10 bin intensity histogram in 5x5 window around pixel
Non-uniform bins
Get local neighbourhood information as opposed to single pixel
Histogram gives rotational invariance

12. Logistic Regression
Gives a probability map:

13. Uses local neighbourhood information to make decisions
Graph Cuts
Probability from Logistic Regression Classifier
Gradient dependent pairwise term
Uses local neighbourhood information to make decisions
Pairwise term penalises different labels for adjacent pixels

14. Graph Cuts

15. Tracking
Associate objects based on distance between centroids in consecutive frames.
Easy given segmentation and slow movement of cells.

16. Tracking
Associate objects based on distance between centroids in consecutive frames.
Easy given segmentation and slow movement of cells.

17. Tracking
Associate objects based on distance between centroids in consecutive frames.
Easy given segmentation and slow movement of cells.

18. Tracking

19. Part II – Phase Labelling

20. Simple features
Maximum Intensity:
Interphase

21. Simple features
Maximum Intensity:
Interphase
Prophase

22. Simple features
Maximum Intensity:
Interphase
Prometaphase
Prophase

23. Simple features Maximum Intensity: Interphase Prometaphase Prophase

24. Simple features Maximum Intensity: Interphase Prometaphase Anaphase
Prophase
Metaphase

25. Simple features Maximum Intensity: Interphase Prometaphase Anaphase
Prophase
Metaphase

26. Simple features Maximum Intensity: Interphase Prometaphase Anaphase
Prophase
Metaphase
Telophase

27. Reference signal
Average over training set (±1 standard deviation shaded):

28. Dynamic time warping
Stretch signal onto labelled reference:

29. Dynamic time warping
Stretch signal onto labelled reference:

30. Dynamic time warping Interphase Prometaphase Anaphase Interphase
Prophase
Metaphase
Telophase

31. Dynamic time warping
Find a cost matrix of pairwise distances between points on the two signals
Find minimum cost path through matrix
Test Signal
Reference Signal

32. Features Use 3 features and their gradients at two different scales:
Maximum intensity
Area
Compactness ( 𝑎𝑟𝑒𝑎 𝑝𝑒𝑟𝑖𝑚𝑖𝑡𝑒𝑟 2 )

33. Hidden Markov Model Hidden states, x Observations, y
Mitotic phases
Observations, y
Features
Transition probabilities, a
From one phase to the next
Emission probabilities, b
Of features having a given value in a given phase
Image:

34. Hidden Markov Model DTW essentially a special case of HMM
Easy to extend approach
Can add other classes e.g. cell death
Split phases into sub-phases to account for variation

35. Experiments and Data 54 movies 119 mitotic tracks
27 movies (61 tracks) training, 27 movies (58 tracks) testing

36. Results
Interphase
Prophase
Prometaphase
Metaphase
Anaphase
Telophase

37. Results

38. Outputs

39. Outputs Synopsis video1 of mitotic cells Aligned to start of anaphase
1Rav-Acha et al., 2006

40. Conclusions Novel approach to cell cycle phase labelling
Utilises temporal context
Extendable with HMM

41. Questions?