Two-Step Watershed Segmentation

of Epithelial Cells

Kevin Keraudren, Martin Spitaler, Vania M.M. Braga,Daniel Rueckert, Luis Pizarro

Department of Computing, National Heart & Lung Institute,

Imperial College London, UK

6th International Workshop on Microscopic Image Analysis withApplications in Biology

Heidelberg, September 2nd, 2011

Introduction Segmentation Feature analysis Conclusion

1 IntroductionBiological backgroundChallenges

2 SegmentationProposed methodComparison with Voronoi-based segmentation

3 Feature analysis

4 Conclusion

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 2 / 44

Introduction Segmentation Feature analysis Conclusion

1 IntroductionBiological backgroundChallenges

2 Segmentation

3 Feature analysis

4 Conclusion

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 3 / 44

Introduction Segmentation Feature analysis Conclusion

Biological background

Biological background

Cells junctions essential for life of multi-cellular organisms

E-cadherin:I master cell-cell adhesion receptorI frequently disrupted in different pathologies and cancer

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 4 / 44

Introduction Segmentation Feature analysis Conclusion

Biological background

Biological backgroundDuring malignancy different oncogenesdisassemble junctions to enable cell invasion

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 5 / 44

Introduction Segmentation Feature analysis Conclusion

Biological background

High-Throughput RNAi Screens

Studying the stability of junctions for Cancer Research

E-cadherin Actin Nuclei

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 6 / 44

Introduction Segmentation Feature analysis Conclusion

Biological background

Junctions phenotypes

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 7 / 44

Introduction Segmentation Feature analysis Conclusion

Biological background

ChallengesAccurate cell segmentation for per cell measurements:

E-cadherin & Actin localization/quantification

Shape of junctions: straight, wavy, fragmented

=⇒

NucleiCytoplasmJunctionsCell-cell contact

Background

No current method

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 8 / 44

Introduction Segmentation Feature analysis Conclusion

Biological background

ChallengesAccurate cell segmentation for per cell measurements:

E-cadherin & Actin localization/quantification

Shape of junctions: straight, wavy, fragmented

=⇒

NucleiCytoplasmJunctionsCell-cell contact

Background

No current method

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 8 / 44

Introduction Segmentation Feature analysis Conclusion

Challenges

Challenges

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 9 / 44

Introduction Segmentation Feature analysis Conclusion

Challenges

Challenges

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 10 / 44

Introduction Segmentation Feature analysis Conclusion

1 Introduction

2 SegmentationProposed methodComparison with Voronoi-based segmentation

3 Feature analysis

4 Conclusion

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 11 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Segmentation pipeline overview

1 Preprocessing

2 Nuclei segmentation

3 Background thresholding

4 Edge map

5 Two-step watershed

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 12 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Junctions image

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 13 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Block-matching filtering (BM3D, Dabov et al., 2007)

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 14 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Block-matching filtering (BM3D, Dabov et al., 2007)

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 15 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Coherence enhancing diffusion (CED, Weickert, 1999)

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 16 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Background tresholding on Actin

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 17 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Background tresholding on Actin

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 18 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Adaptive thresholding + filling holes on Nuclei

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 19 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Sobel filter + morphological closing on Junctions

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 20 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Sobel filter + morphological closing on Junctions

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 21 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

1st watershed step on Edge map

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 22 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

1st watershed step on Edge map

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 23 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

1st watershed step on Edge map

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 24 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Eroding cells at junctions

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 25 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

Eroding cells at junctions

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 26 / 44

Introduction Segmentation Feature analysis Conclusion

Proposed method

2nd watershed on raw Junctions

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 27 / 44

Introduction Segmentation Feature analysis Conclusion

Comparison with Voronoi-based segmentation

Comparison with Voronoi-based segmentation

Voronoi-Based Segmentation (T. Jones et al., 2005)I implemented in CellProfiler and Bioconductor (propagate)

Pixel distance =Gradient + λ×Euclidean distance

1 + λ

λ = 0 λ = 1000

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 28 / 44

Introduction Segmentation Feature analysis Conclusion

Comparison with Voronoi-based segmentation

Comparison with Voronoi-based segmentationpropagate λ = 10−5

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 29 / 44

Introduction Segmentation Feature analysis Conclusion

Comparison with Voronoi-based segmentation

Comparison with Voronoi-based segmentationProposed segmentation pipeline

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 30 / 44

Introduction Segmentation Feature analysis Conclusion

1 Introduction

2 Segmentation

3 Feature analysis

4 Conclusion

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 31 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extractionJunctions region with per cell threshold

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 32 / 44

Introduction Segmentation Feature analysis Conclusion

Feature analysis

0.1 0.2 0.3 0.4 0.5

0.05

0.10

0.15

0.20

Mean HECD1 intensity at junctions Vs non−junction

Mean HECD1 intensity at junction (per cell threshold)

Mea

n H

EC

D1

inte

nsity

in c

ytop

lasm

(pe

r ce

ll th

resh

old)

CytoplasmicWavyFragmented

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 33 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extractionCorner-to-corner contours

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 34 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extractionCorner-to-corner contours

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 35 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extraction

●

●

Cytoplasmic Wavy Fragmented

1.32

1.34

1.36

1.38

1.40

1.42

1.44

Cell perimeter / Corner−to−corner perimeter (mean)

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 36 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extractionSmoothed contours

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 37 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extractionSmoothed contours

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 38 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extraction

●

●

●

●

Cytoplasmic Wavy Fragmented

1.07

1.08

1.09

1.10

1.11

Cell perimeter / Smoothed perimeter (mean)

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 39 / 44

Introduction Segmentation Feature analysis Conclusion

1 Introduction

2 Segmentation

3 Feature analysis

4 Conclusion

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 40 / 44

Introduction Segmentation Feature analysis Conclusion

Main contribution

Novel segmentation method for RNAi depleted cells

Accurate localization of cell structures

Allows precise quantification of E-cadherin & Actin

NucleiCytoplasmJunctionsCell-cell contactBackground

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 41 / 44

Introduction Segmentation Feature analysis Conclusion

Limitations of our method

Tailored for specific images: cell-cell junctions

Suffers from watershed leaks or local maxima

Unable to recognise“no-nucleus”cells

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 42 / 44

Introduction Segmentation Feature analysis Conclusion

Future work

Segmentation quality assessment

Cell under-segmentation

More feature extraction

Share code with EBImage authors

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 43 / 44

Introduction Segmentation Feature analysis Conclusion

Thanks!

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 44 / 44

Introduction Segmentation Feature analysis Conclusion

watergrow function

Watershed with seed regions

Region growing with shape prior

Example with a 5x5 box neighbourhood, 6 votes required:

1 1 1 1 1

3 3 2 3 3

3 3 2 3 3

3 1 1 3 3

3 1 1 3 3

2 will not be included

1 1 1 1 1

3 2 2 2 3

3 2 2 2 3

3 1 1 1 3

1 1 1 1 1

2 will be included

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 44 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extractionJunctions region with per cell threshold

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 44 / 44

Introduction Segmentation Feature analysis Conclusion

Feature extractionJunctions region with constant width

Keraudren, Spitaler, Braga, Rueckert, Pizarro () Two-Step Watershed Segmentation September 2nd, 2011 44 / 44