SLIDE 1
Robert Bensch
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
2
www.bioss.uni-freiburg.de
Outline
- Introduction
- Method
– Segmentation – Tracking
- Experiments & Results
- Conclusion
Cell Tracking Challenge (3rd Edition) Cell Segmentation and Tracking - - PowerPoint PPT Presentation
Cell Tracking Challenge (3rd Edition) Cell Segmentation and Tracking - - PowerPoint PPT Presentation
Cell Tracking Challenge (3rd Edition) Cell Segmentation and Tracking in Phase Contrast Images using Graph Cut with Asymmetric Boundary Costs Robert Bensch and Olaf Ronneberger Computer Science Department and BIOSS Centre for Biological
SLIDE 2
SLIDE 3
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
3
www.bioss.uni-freiburg.de
3
Phase contrast microscopy
- Visualize transparent objects with high contrast at
cell borders
Bright-field Phase-contrast
Figure: B. Alberts et al., Molecular Biology of the Cell, 4th Edition, 2002.
Phase-contrast
SLIDE 4
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
4
www.bioss.uni-freiburg.de
4
Phase contrast microscopy
Shade-off Halo pattern Strong edges inside and outside the cell
- Drawback: Artifacts
SLIDE 5
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
5
www.bioss.uni-freiburg.de
5
Standard segmentation algorithms
- Standard edge-based segmentation algorithms fail
- Traditional graph cut with symmetric boundary
costs.
Cyan: Graph cut segmentation result Yellow: Our manual ground truth
SLIDE 6
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
6
www.bioss.uni-freiburg.de
6
Our approach
- True cell borders appear as dark-to-bright transition*
- Search for segmentation mask that favors dark-to-
bright transitions at its boundary
- Graph cut with asymmetric boundary costs
(*positive phase contrast microscopy)
Yellow: Cell outwards direction Green: True cell border Red: Wrong cell border
SLIDE 7
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
8
www.bioss.uni-freiburg.de
- Introduction
- Method
– Segmentation – Tracking
- Experiments & Results
- Conclusion
8
Outline
SLIDE 8
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
9
www.bioss.uni-freiburg.de
9
Segmentation energy functional
- Cost function (Region & boundary term)
- Boundary term
- Asymmetric boundary penalties (dark-to-bright)
→ directed graph with asymmetric edge weights
SLIDE 9
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
10
www.bioss.uni-freiburg.de
Symmetric boundary penalties
3x3 pixel neighborhood, Edges and weights (only
- utwards edges shown)
NW NE N S SE SW W E
SLIDE 10
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
11
www.bioss.uni-freiburg.de
Symmetric boundary penalties
low high 3x3 pixel neighborhood, Edges and weights (only
- utwards edges shown)
NW NW NE N S SE SW W E N SW SE NE W S E
- Low costs at wrong cell borders
(bright-to-dark transitions)
SLIDE 11
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
12
www.bioss.uni-freiburg.de
Asymmetric boundary penalties
low high 3x3 pixel neighborhood, Edges and weights (only
- utwards edges shown)
NW NW NE N S SE SW W E N SW SE NE W S E
- Low costs at correct cell borders
(dark-to-bright transitions)
SLIDE 12
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
13
www.bioss.uni-freiburg.de
Asymmetric boundary penalties
Cyan mask: Segmentation result
- f graph cut with symmetric costs
Yellow: Our manual ground truth Red mask: Segmentation result of proposed method Yellow: Our manual ground truth
SLIDE 13
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
14
www.bioss.uni-freiburg.de
14
Regional penalties
- Standard graph cut
- In our approach
→ hard constraint → soft constraint
SLIDE 14
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
15
www.bioss.uni-freiburg.de
15
Optimization
- Enery minimization problem
- Discretize edge term into 8 directions
→ combinatorial optimization problem
- Solve efficiently by a min-cut approach
SLIDE 15
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
16
www.bioss.uni-freiburg.de
- Introduction
- Method
– Segmentation – Tracking
- Experiments & Results
- Conclusion
16
Outline
SLIDE 16
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
17
www.bioss.uni-freiburg.de
- Propagate Segmentation
Information
- Foreground information using
eroded mask
→ foreground constraint
- Partitioning information using
borders of „support regions“
→ background constraint
17
Tracking: Segmentation propagation
SLIDE 17
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
18
www.bioss.uni-freiburg.de
- Propagate Labels to
- verlapping Segments
- Resolve one-to-many
correspondences
– Propagate label to max. IOU – Invent new labels
- Resolve many-to-one
correspondences
– Take label from max. IOU – Kill other labels
18
Tracking: Label propagation
SLIDE 18
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
19
www.bioss.uni-freiburg.de
- Introduction
- Method
– Segmentation – Tracking
- Experiments & Results
- Conclusion
19
Outline
SLIDE 19
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
20
www.bioss.uni-freiburg.de
20
Datasets: ISBI cell tracking challenge1,2
Glioblastoma-astrocytoma U373 cells on a polyacrylimide substrate*
(1) ISBI Cell Tracking Challenge, Available at: http://www.codesolorzano.com/celltrackingchallenge. (2) M. Maška, V. Ulman, D. Svoboda, P. Matula, and P. Matula, et al., “A benchmark for comparison of cell tracking algorithms,” Bioinformatics, vol. 30, no. 11, pp. 1609–1617, 2014.
*Data provided by Dr. Sanjay Kumar. Department of Bioengineering University of California at Berkeley. Berkeley CA (USA).
†Data provided by Dr. Tim Becker. Fraunhofer Institution for Marine Biotechnology. Lübeck. Germany
Pancreatic Stem Cells on a Poly- styrene substrate (2D)†
- Strong shape variations
- Weak outer borders, strong irrelevant inner borders
- Cytoplasm has same structure as background
SLIDE 20
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
23
www.bioss.uni-freiburg.de
Raw data Graph cut (symm.) Ours (asymm.) Cyan masks: Graph cut with symmetric costs, Red masks: Our approach with asymmetric costs, Yellow borders: Our manual ground truth
23
Experiments: Symmetric vs. asymmetric costs
- Improved detection of very weak boundaries
- Halo boundaries are handled well
SLIDE 21
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
25
www.bioss.uni-freiburg.de
25
Submitted results: PhC-C2DH-U373
SLIDE 22
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
26
www.bioss.uni-freiburg.de
26
Submitted results: PhC-C2DL-PSC
SLIDE 23
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
27
www.bioss.uni-freiburg.de
27
Conclusion
- Direction dependent boundary costs improve
segmentation in phase contrast microscopy
- Our approach outperforms standard min-cut
segmentation with symmetric costs
→ Profit for cell segmentation in other modalities → Open-source MATLAB code (and ImageJ plugin)*:
http://lmb.informatik.uni-freiburg.de/resources/opensource/CellTracking/ *(coming soon ;)
SLIDE 24
Robert Bensch, University of Freiburg, Germany, April 16, ISBI 2015
28
www.bioss.uni-freiburg.de