Agenda

Overview

Technical Details

Our Team

Overview

Simple task

Nuclei segmentation

Overview

Train

Test

Input images sample variability

Overview

Differing Saliency

  • some nucleus can be easily seen
  • some are barely distinguishable
  • some have homogenous structure
  • some have nucleous in the center 

Overview

Varying shape and size

  • some are elliptical
  • some are oblong
  • some are large
  • some are very small

Overview

Varying level of nuclei attachment

  • some images are very dense
  • some are sparse
  • some have large groups of glued nuclei

Pipeline

Input image

Preprocessing

Morphological pooling

Watershed ensemble generation

Thresholding

Morphological ensemble generation

Technical Details

Watershed pooling

Segmented mask

Preprocessing

Hematoxylin 

Eosin

Background

Technical Details

Color deconvolution

Pipeline

Preprocessing

Hematoxylin 

Eosin

Background

Hematoxylin Channel Extraction

Technical Details

Color deconvolution

Pipeline

Preprocessing

Reinhard's Color Normalization

References:

  • Reinhard, E., Adhikhmin, M., Gooch, B., Shirley, P.: Color transfer between images. IEEE Computer Graphics and Applications 21(5) (2001) 34–41
  • Wang, Y., nd L. Wu, S.C., Tsai, S., Sun, Y.: A color-based approach for automated segmentation in tumor tissue classification. In: Proc. Conf. of the IEEE Engineering in Medicine and Biology Society. (2007) 6577–6580
\begin{matrix} \\ l_{mapped} = \frac{l_{original}-\bar{l}_{original}}{\hat{l}_{original}}\hat{l}_{target}+\bar{l}_{target} \\ \\ \alpha_{mapped} = \frac{\alpha_{original}-\bar{\alpha}_{original}}{\hat{\alpha}_{original}}\hat{\alpha}_{target}+\bar{\alpha}_{target} \\ \\\beta_{mapped} = \frac{\beta_{original}-\bar{\beta}_{original}}{\hat{\beta}_{original}}\hat{\beta}_{target}+\bar{\beta}_{target} \end{matrix}
lmapped=loriginall¯originall^originall^target+l¯targetαmapped=αoriginalα¯originalα^originalα^target+α¯targetβmapped=βoriginalβ¯originalβ^originalβ^target+β¯target \begin{matrix} \\ l_{mapped} = \frac{l_{original}-\bar{l}_{original}}{\hat{l}_{original}}\hat{l}_{target}+\bar{l}_{target} \\ \\ \alpha_{mapped} = \frac{\alpha_{original}-\bar{\alpha}_{original}}{\hat{\alpha}_{original}}\hat{\alpha}_{target}+\bar{\alpha}_{target} \\ \\\beta_{mapped} = \frac{\beta_{original}-\bar{\beta}_{original}}{\hat{\beta}_{original}}\hat{\beta}_{target}+\bar{\beta}_{target} \end{matrix}

Target image

Technical Details

Pipeline

Preprocessing

Hematoxylin Channel Extraction

Reinhard's Color Normalization

Target image

Technical Details

Pipeline

Thresholding

Possible algorithms:

  • K-means clustering
  • Fuzzy c-means clustering

  • Gaussian mixture models

2.    With spatial information

 

1.    No spatial information

 

  • Spatial information Fuzzy c-means clustering
  • Fast spatial distance weighted Fuzzy c-means clustering

  • Dictionary model

References:

  • Zeng, Huang, Kang and Sang. Image segmentation using spectral clustering of Gaussian mixture models. Neurocomputing 144:346–356, 2014. 
  • Guo, Liu, Wu, Hong and Zhang. A New Spatial Fuzzy C-Means for Spatial Clustering. WSEAS Transactions on Computers 14:369-381, 2015.
  • Dahl and Larsen. Learning Dictionaries of Discriminative Image Patches. In: Proc. British Machine Vision Conference. p.77, 2011.
  • Hamed Shamsi and Hadi Seyedarabi, Member, IACSITA Modified Fuzzy C-Means Clustering with Spatial Information for Image Segmentation,International Journal of Computer Theory and Engineering, Vol. 4, No. 5, 2012.

Technical Details

Pipeline

Thresholding

K-means clustering

Parameters:

  • number of color clusters
  • number of most most intensive clusters to be classified as nuclei

Models

  • k-means(3,1)
  • k-means(4,2)
  • k-means(6,3)

input

color clusters

1

6

5

4

3

2

Technical Details

Pipeline

Thresholding

input

color clusters

top 3

most intensive clusters sum

1

6

5

4

3

2

4+5+6

K-means clustering

Parameters:

  • number of color clusters
  • number of most most intensive clusters to be classified as nuclei

Models

  • k-means(3,1)
  • k-means(4,2)
  • k-means(6,3)

Technical Details

Pipeline

Thresholding

Gaussian Mixture Model

Parameters:

  • number of color clusters
  • number of most most intensive clusters to be classified as nuclei
  • which of the distribution parameters (means, covariances,weights) should be updated
  • which covariance option (diagonal,full,tied,spherical) should be used

Models

  • GMM(3,1,{means,covariances},diagonal)
  • GMM(6,3,{covariances},diagonal)
  • GMM(6,3,{covariances},tied)

References:

  • Zeng, Huang, Kang and Sang. Image segmentation using spectral clustering of Gaussian mixture models. Neurocomputing 144:346–356, 2014. 

Technical Details

Pipeline

Thresholding

Spatial distance weighted fuzzy c-means clustering

Parameters:

  • number of color clusters
  • number of most most intensive clusters to be classified as nuclei
  • fuzziness parameter
  • spatial information importance
  • number of neighbours

Models

  • SDWFCM(3,1,2,0.9,8)
  • SDWFCM(3,1,2,0.9,24)
  • SDWFCM(3,2,2,0.9,24)
\begin{matrix} \\v^{(t)}_{i} = \frac{\sum\limits_{k=1}^{n}(u^{(t-1)}_{ik})^{m}x_k }{\sum\limits_{k=1}^{n}(u^{(t-1)}_{ik})^{m} },1\leq i\leq c \\ \\f_{ij}=\frac{\sum\limits_{s_{k}\in NB(s_j)}d_{ik}}{Min\left \{\sum\limits_{s_{k}\in NB(s_j)}d_{lk},1\leq l \leq c \right \}} \\ \\D_{ij}=(1-\lambda )d_{ij}f_{ij}+\lambda d_{ij} \\ \\u^{(t)}_{ik}=\frac{1}{\sum\limits_{j=1}^{c}(D_{ij}/D_{jk})^\frac{1}{(m-1)}} \\ \\J(X;U;V)=\sum\limits_{i=1}^c\sum\limits_{k=1}^n(u_{ik})^m\left \|x_k-v_i \right \|^2 \end{matrix}
vi(t)=k=1n(uik(t1))mxkk=1n(uik(t1))m,1icfij=skNB(sj)dikMin{skNB(sj)dlk,1lc}Dij=(1λ)dijfij+λdijuik(t)=1j=1c(Dij/Djk)1(m1)J(X;U;V)=i=1ck=1n(uik)mxkvi2\begin{matrix} \\v^{(t)}_{i} = \frac{\sum\limits_{k=1}^{n}(u^{(t-1)}_{ik})^{m}x_k }{\sum\limits_{k=1}^{n}(u^{(t-1)}_{ik})^{m} },1\leq i\leq c \\ \\f_{ij}=\frac{\sum\limits_{s_{k}\in NB(s_j)}d_{ik}}{Min\left \{\sum\limits_{s_{k}\in NB(s_j)}d_{lk},1\leq l \leq c \right \}} \\ \\D_{ij}=(1-\lambda )d_{ij}f_{ij}+\lambda d_{ij} \\ \\u^{(t)}_{ik}=\frac{1}{\sum\limits_{j=1}^{c}(D_{ij}/D_{jk})^\frac{1}{(m-1)}} \\ \\J(X;U;V)=\sum\limits_{i=1}^c\sum\limits_{k=1}^n(u_{ik})^m\left \|x_k-v_i \right \|^2 \end{matrix}

cluster center

spatial distance weighted funcion

weighted distance

membership probabilty

objective function

References:

  • Hamed Shamsi and Hadi Seyedarabi, Member, IACSITA Modified Fuzzy C-Means Clustering with Spatial Information for Image Segmentation,International Journal of Computer Theory and Engineering, Vol. 4, No. 5, 2012.

Technical Details

Pipeline

Thresholding

Dictionary model

References:

  • Dahl and Larsen. Learning Dictionaries of Discriminative Image Patches. In: Proc. British Machine Vision Conference. p.77, 2011.

Technical Details

Learning:

  • seeding dictionaries of texture and label patches
  • iterative clustering of patches according to label similarity

Segmentation:

  • assigning the idealized label of the most similar cluster centroid

Pipeline

Thresholding

Parameters:

  • patch size
  • label similarity threshold
  • percentage of seeding and for training
  • number of training iterations
  • adjustment coefficient for centroids of texture clusters

Models

  • DM(1,0.4,0.01,0.1,5,0.05,0.5)
  • DM(1,0.5,0.01,0.1,20,0.05,0.0)
  • DM(2,0.6,0.01,0.1,5,0.05,0.0)
  • DM(3,0.6,0.01,0.1,5,0.15,0.5)

References:

  • Dahl and Larsen. Learning Dictionaries of Discriminative Image Patches. In: Proc. British Machine Vision Conference. p.77, 2011.

Technical Details

Dictionary model

Pipeline

Thresholding

Reality is far from perfect:

  • leftover artifacts
  • hollow shapes
  • glued nuclei groups

 

Technical Details

Pipeline

Thresholding

Operations:

  • dilation and erosion
  • reconstruction by erosion
  • closing and opening
  • fill binary holes

 

Morphological Transformations

Technical Details

Pipeline

Thresholding

Operations:

  • dilation and erosion
  • reconstruction by erosion
  • closing and opening
  • fill binary holes

 

Morphological Transformations

Problems:

  • Sample diversity
  • Diversity within one image
  • "one size fits all" structuring element does not exist

 

Technical Details

Pipeline

Morphological ensemble generation

k-means(3,1)

k-means(4,2)

SDWFCM(3,2,2,0.9,24)

 1

 2

 r

Morphological ensemble

Technical Details

Pipeline

k-means(3,1)

k-means(4,2)

SDWFCM(3,2,2,0.9,24)

 1

 2

 r

Morphological ensemble

Morphological pooling

probabilty map

Technical Details

Pipeline

k-means(3,1)

k-means(4,2)

SDWFCM(3,2,2,0.9,24)

 1

 2

 r

Morphological ensemble

k-means clustering

Morphological pooling

probabilty map

Technical Details

Pipeline

nuclei detachment

Watershed ensemble generation

Watershed algorithm

  • local maxima search space
  • minimal distance between markers
  • other

Technical Details

Pipeline

nuclei detachment

Watershed ensemble generation

Watershed algorithm

  • local maxima search space
  • minimal distance between markers
  • other

Problems:

  • Sample diversity
  • Diversity within one image
  • "one size fits all" parameter set does not exist

 

Technical Details

Pipeline

Watershed ensemble generation

watershed realizations

thresholding output

Technical Details

Pipeline

Watershed ensemble generation

watershed realizations

thresholding output

edge realizations

Technical Details

Pipeline

Watershed pooling

probability map

thresholding output

edge realizations

Technical Details

Pipeline

Watershed pooling

probability map

thresholding output

modified thresholding output

overimpose

edge realizations

Technical Details

Pipeline

Watershed pooling

probability map

thresholding output

modified thresholding output

final segmentation mask

watershed

overimpose

edge realizations

Technical Details

Pipeline

Our Team

Grzegorz Żurek

R&D Stermedia

Wroclaw University of Technology

Jakub Czakon

R&D Stermedia

Piotr Giedziun

R&D Stermedia

Ph.D Witold Dyrka

R&D Stermedia

Wroclaw University of Technology

Piotr Krajewski

CIO Stermedia

Michał Błach

R&D Stermedia

Wroclaw University of Technology

MD Łukasz Fuławka

Patomorphology resident 

Lower Silesian Oncology Center

Contact us 

info@stermedia.pl

jakub.czakon@stermedia.pl

grzegorz.zurek@stermedia.pl

We are looking forward to collaborating with you

Thank you for attention

Segmentation Algorithm

By Stermedia Sp. z o.o.

Segmentation Algorithm

presentation for MICCAI 2015

  • 576
Loading comments...

More from Stermedia Sp. z o.o.