The document discusses applying machine learning algorithms like perceptrons and support vector machines to classify stem cell images. It describes imaging stem cell nuclei, the problem of stem cell classification, and online machine learning methods. The perceptron achieved 70.38% accuracy on test data while support vector machines with radial basis function kernels optimized by CMA-ES achieved 97.02% accuracy, showing machine learning is feasible for stem cell image classification.
Applying Support Vector Learning to Stem Cells Classification
1. Introduction
Online Machine Learning
The Application
Discussion
Applying Support Vector Learning
to Stem Cells Classification
Ofer M. Shir
oshir@liacs.nl
Natural Computing Group
Leiden University
LUMC, MCB Seminar, 25-09-2006
Ofer M. Shir SVM to Stem-Cells Classification
2. Introduction
Online Machine Learning
The Application
Discussion
Outline
1 Introduction
The Problem: Stem Cells Classification
Nucleus Imaging
2 Online Machine Learning
The Teacher-Learner Model
Simple Perceptron
The SVM Algorithm
Images as Instances
3 The Application
Applying Perceptron
Applying SVM
4 Discussion
Conclusions
Prospects
Take-Home Message
Ofer M. Shir SVM to Stem-Cells Classification
3. Introduction
Online Machine Learning The Problem: Stem Cells Classification
The Application Nucleus Imaging
Discussion
Outline
1 Introduction
The Problem: Stem Cells Classification
Nucleus Imaging
2 Online Machine Learning
The Teacher-Learner Model
Simple Perceptron
The SVM Algorithm
Images as Instances
3 The Application
Applying Perceptron
Applying SVM
4 Discussion
Conclusions
Prospects
Take-Home Message
Ofer M. Shir SVM to Stem-Cells Classification
4. Introduction
Online Machine Learning The Problem: Stem Cells Classification
The Application Nucleus Imaging
Discussion
Biological Motivation
The nuclear lamina envelops the nucleus. Intact lamina is
vital for cell survival, knowckdown of lamin B results in
lethal embryos in mice, and mutations in Lamin A cause
premature aging syndromes in human.
In human mesenchemyal stem cells (hMSCs) the lamina
shows a round and flat shape after 3D reconstruction. In
hMSCs undergoing cell death the lamina shape
dramatically changed and precedes the wholemarks of
apoptosis, such as nuclear breakdown and chromatin
fragmentation.
Soon after caspase-8 activation, which ultimately leads to
cell death, intranuclear organization of the lamina are
formed and the depth of the nucleus increased. Similar
changes in lamina organization are found in hMSCs
undergoing replicative senescence.
Ofer M. Shir SVM to Stem-Cells Classification
5. Introduction
Online Machine Learning The Problem: Stem Cells Classification
The Application Nucleus Imaging
Discussion
Biological Motivation
Thus, it is possible that changes in the spatial organization
of the lamina are correlated with the functional state of the
cell. The spatial organization of the lamina can be used as
an early marker to sort between healthy and not-healthy
cells, as changes in lamina organization are visible before
changes in cell morphology are detected.
Here we tested this hypothesis using a machine learning
approach.
Ofer M. Shir SVM to Stem-Cells Classification
6. Introduction
Online Machine Learning The Problem: Stem Cells Classification
The Application Nucleus Imaging
Discussion
Nucleus Imaging
The lamina of hMSCs was detected after transduction of
the Lamin B-GFP lentivirus vector.
Image stacks of the lamin B-GFP were aquired with a
confocal microscope, and 3D reconstruction was obtained
with TeloView.
In control cells the XY and the XZ orientations revealed a
round and flat shape of the lamina.
After activation of caspase-8, the shape of the lamina is
significantly changed.
Ofer M. Shir SVM to Stem-Cells Classification
7. Introduction
Online Machine Learning The Problem: Stem Cells Classification
The Application Nucleus Imaging
Discussion
Control vs. Apoptotic
Ofer M. Shir SVM to Stem-Cells Classification
8. Introduction
Online Machine Learning The Problem: Stem Cells Classification
The Application Nucleus Imaging
Discussion
Nucleus Imaging
Serial slicing along the XZ axis taken from an individual
nucleus with DIPimage toolbox revealed little changes in
the spatial organization of the lamina in a control cell.
High variations were found in serial slicing taken from an
apoptotic cell.
Ofer M. Shir SVM to Stem-Cells Classification
9. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Outline
1 Introduction
The Problem: Stem Cells Classification
Nucleus Imaging
2 Online Machine Learning
The Teacher-Learner Model
Simple Perceptron
The SVM Algorithm
Images as Instances
3 The Application
Applying Perceptron
Applying SVM
4 Discussion
Conclusions
Prospects
Take-Home Message
Ofer M. Shir SVM to Stem-Cells Classification
10. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Machine Learning: TRAINING
Online learning considers a situation in which instances
are presented one at a time, where the learner’s task is to
learn a hypothesis which classifies the data correctly.
Training phase: instances {xi }l in Rn , and their labels
i=1
set Y = {−1, +1} are presented to the machine. The
algorithm aims to update its hypothesis h : Rn → {±1} in
order to minimize the prediction error.
Various algorithms have different update rules.
Analogy: teacher, learner, corrections.
Ofer M. Shir SVM to Stem-Cells Classification
11. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Machine Learning: TESTING
This training phase is followed by the testing phase, where
more data is given to the learned hypothesis.
Ideally unseen data. (Why...?)
The correct labels are not presented to the machine!
The accuracy rate is considered - how did the machine
perform?
Ofer M. Shir SVM to Stem-Cells Classification
12. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Simple Perceptron
The Perceptron algorithm (Rosenblatt, 1957) is an online
learning algorithm for finding a consistent hypothesis within the
class of hyperplanes:
C = h(x) = sign wT · x + b w t ∈ Rn , b ∈ R
The optimal hyperplane is defined as the one with the maximal
margin of separation between the two instances classes.
Ofer M. Shir SVM to Stem-Cells Classification
13. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Perceptron: Optimal Hyperplane
Ofer M. Shir SVM to Stem-Cells Classification
14. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Non-Realizable for Hyperplanes Separation
But what if the data is not linearly-separable...?
There is no hyperplane separator hypothesis for the problem!
Ofer M. Shir SVM to Stem-Cells Classification
15. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Mapping...
We would like then to map the instances to a higher
dimensional space, where linear separation is feasible:
Ofer M. Shir SVM to Stem-Cells Classification
16. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Desirable Mapping
Ofer M. Shir SVM to Stem-Cells Classification
17. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
The Algorithm
The Support Vector Machines (SVM) algorithm (Boser, Guyon
and Vapnik, 1992) is a linear method in a high-dimensional
feature space, which is non-linearly interlinked to the instance
space. It allows learning a hypothesis for data which is not
linearly-separable.
Ofer M. Shir SVM to Stem-Cells Classification
18. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
The Kernel Function
The function φ : Rn → F maps the instance vectors onto a
higher dimensional space F, and then the SVM aims to find a
hyperplane separator with the maximal margin in this space.
k (xi , xj ) ≡ φ(xi )T φ(xj )
Ofer M. Shir SVM to Stem-Cells Classification
19. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Kernels
In particular, we consider the following kernel functions:
The polynomial kernel:
d
k (xi , xj ) = γ xT · xj + r
i (1)
Radial basis function (RBF) kernel:
1 2
k (xi , xj ) = exp − xi − xj (2)
2σ 2
The sigmoid kernel:
k (xi , xj ) = tanh κ xT · xj + Θ
i (3)
Ofer M. Shir SVM to Stem-Cells Classification
20. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Images as Instances
Ofer M. Shir SVM to Stem-Cells Classification
21. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Grayscale Images
Ofer M. Shir SVM to Stem-Cells Classification
22. Introduction The Teacher-Learner Model
Online Machine Learning Simple Perceptron
The Application The SVM Algorithm
Discussion Images as Instances
Intermediate Conclusions
Grayscale images are simply matrices with normalized
elements in [0, 1].
In particular, as instance vectors in Rn !
Essentially, an image could be introduced directly to
the learning algorithm, without further processing.
Ofer M. Shir SVM to Stem-Cells Classification
23. Introduction
Online Machine Learning Applying Perceptron
The Application Applying SVM
Discussion
Outline
1 Introduction
The Problem: Stem Cells Classification
Nucleus Imaging
2 Online Machine Learning
The Teacher-Learner Model
Simple Perceptron
The SVM Algorithm
Images as Instances
3 The Application
Applying Perceptron
Applying SVM
4 Discussion
Conclusions
Prospects
Take-Home Message
Ofer M. Shir SVM to Stem-Cells Classification
24. Introduction
Online Machine Learning Applying Perceptron
The Application Applying SVM
Discussion
Experimental Procedure: Modus Operandi
Training phase: provide the machine with shuffled 2000
slices and their correct labels.
Testing phase: test the machine with shuffled 1040 slices
without their labels - and check its accuracy.
Correct classification means that the output of the machine
per given instance is its correct label as in our database.
Wrong classification (error rate) - vice versa.
Ofer M. Shir SVM to Stem-Cells Classification
25. Introduction
Online Machine Learning Applying Perceptron
The Application Applying SVM
Discussion
Applying Perceptron
Applying the Perceptron was straightforward, with respect
to parameter settings, and did not require any preliminary
tuning.
However, the algorithm obtained, after training, a test
accuracy of 70.38% (732/1040 images were classified
correctly).
This result led us to the conclusion that the data was not
linearly-separable, and a stronger approach was much
needed.
Ofer M. Shir SVM to Stem-Cells Classification
26. Introduction
Online Machine Learning Applying Perceptron
The Application Applying SVM
Discussion
Applying SVM - Preliminary
Applying SVM (libsvm package) to the classification
problem with default settings yielded test accuracy of 55%
on average.
Thus, tuning the kernel parameters was essential - several
parameters as well as the profile of the kernel (Eq. 1, 2, 3)
and its various appropriate parameters ({γ, r, d}, {σ} and
{κ, Θ}).
The Covariance Matrix Adaptation Evolution Strategy
(CMA-ES) [Hansen et al., 2001] was selected as the
optimization tool: the cross-validation accuracy rate was
the objective function to be optimized.
Each objective function evaluation takes 11 minutes on a
single processor: runs were limited.
Ofer M. Shir SVM to Stem-Cells Classification
27. Introduction
Online Machine Learning Applying Perceptron
The Application Applying SVM
Discussion
SVM - Numerical Results
CMA-ES found an RBF kernel with 98.90%
cross-validation.
Testing phase:
Accuracy of 97.02% - 1009/1040 images were classified
correctly!
Highly satisfying! Beyond any expectation!
Ofer M. Shir SVM to Stem-Cells Classification
28. Introduction
Conclusions
Online Machine Learning
Prospects
The Application
Take-Home Message
Discussion
Outline
1 Introduction
The Problem: Stem Cells Classification
Nucleus Imaging
2 Online Machine Learning
The Teacher-Learner Model
Simple Perceptron
The SVM Algorithm
Images as Instances
3 The Application
Applying Perceptron
Applying SVM
4 Discussion
Conclusions
Prospects
Take-Home Message
Ofer M. Shir SVM to Stem-Cells Classification
29. Introduction
Conclusions
Online Machine Learning
Prospects
The Application
Take-Home Message
Discussion
Conclusions
Machine learning as a way of life.
Machine classification of stem cells is feasible!
Numerical results are remarkably excellent.
No further image analysis, after the image acquisition, is
required.
Behind everything in life there is a matrix...
Ofer M. Shir SVM to Stem-Cells Classification
30. Introduction
Conclusions
Online Machine Learning
Prospects
The Application
Take-Home Message
Discussion
Prospects
Classification of other ”colors”.
Classification of 3D images!
Analysis of time-dependent 3D movies.
Ofer M. Shir SVM to Stem-Cells Classification
31. Introduction
Conclusions
Online Machine Learning
Prospects
The Application
Take-Home Message
Discussion
Take-Home Message
Natural computing, machine learning and data mining are
rich fields with a lot to offer!
Find yourself a nice computer-scientist, and invest in your
relationship.
You may prefer to consider those tools as a black-boxes.
BUT then apply and boost medicine...
Ofer M. Shir SVM to Stem-Cells Classification