Machine Learning for
Bio-Image-Analysis

10.05.2021

Volker Bäcker, Jean-Bernard Fiche,
Cedric Hassen Khodja, Francesco Pedaci

Introduction

  • What is bio-image analysis?

  • How is it done without machine
    learning?

  • What is machine learning?

  • How is bio-image analysis done
    with machine learning?

Bio-Image-Analysis

“The extraction of information from digital images in the context of biological research”

The Image-Analysis Workflow

Select a scale

  • Use “Gaussian blur”-filter to select a scale
    • Low-pass filter
    • Removes high frequencies from the image
      • The higher sigma, the lower the remaining frequencies

Convolution

Features at different scales

feature σ=3.5 σ=7.0 σ=10.
variance
sobel

Machine Learning

Machine learning algorithms build a mathematical model of sample data, known as ”training data”, in order to make predictions or decisions without being explicitly programmed to perform the task.

Machine Learning phases

  • training
    • a model is learned from training data

  • validation
    • the trained model is validated
      against test data

  • application
    • use the trained model to
      make predictions on new data

Machine Learning vocabulary

  • supervised
    • a model is learned from pairs of input and output data
  • unsupervised
    • a model is learned from the inherent structure of the input data alone
  • classification
    • the result is a category
  • regression
    • the result is a real number

Machine Learning -
How is that even possible?

  • ML algorithm implements a
    mathematical model with a number
    of model parameters

  • given the training data,
    find parameter values that minimize
    the prediction error

Machine Learning Example 1
Linear Regression

Training Data:

Femur length (cm) Height (cm)
45 153
44 168
44 177
47 180
44 171
50 168

estimate body height f(x) given the femur length x.

model: f(x) = ω1 + ω2 × x

parameter of the model:
 ω1 and ω2

Machine Learning Example 1
Linear Regression

  • find parameters
    ω1, ω2

  • so that error
    between
    training data
    and model
    is minimal

Example 1 - Squared Loss function

Example 1 - Gradient descent

  • find the minimum of
    the loss function

  • by using gradient
    descent

Example 1 - Predictions

f(x) = ω1 + ω2 × x

ω1 = 131.13cm

ω2 = 0.87

f(55cm) = 131.13cm + 0.87 × 55cm

f(55cm) = 179.42cm

Femur example

  • Supervised or Unsupervised?

  • Classification or Regression?

The programs

  • ImageJ/FIJI

  • Weka / Labkit

  • Ilastik

  • Cellprofiler / CP Analyst

  • Orbit

ImageJ/FIJI

  • Demo ImageJ 01
    • Open Image
    • Threshold
    • Binary Watershed
    • Compare to GT

ImageJ/FIJI

  • Demo ImageJ 02
    • Revert Image
    • Laplacian of Gaussian (scale 3)
    • Threshold (Yen)
    • Binary Watershed
    • Compare to GT

Ilastik

  • Demo Ilastik
    • import image(s)
    • select features and scales
    • name classes
    • select training data
    • export result
    • batch

Exercises 01

Clustering

  • A machine learning method

    • Unsupervised

    • Classification

Clustering

  • Clustering

    • Group objects in a way that

      • objects in the same cluster are
        more similar to each other

      • than to objects in other clusters

Clustering algorithms

  • K-means
  • DBScan
  • hierarchical clustering
  • expectation-maximization
  • …

k-means clustering

  • Algorithm:
    • Start with k initial means
    • Repeat until convergence
      • Assign feature-vectors to clusters
      • Recalculate the means of the clusters

  • Partition the feature-
    space into k-clusters

  • Each feature-vector
    belongs to the cluster
    with nearest mean

K-means example

K-means clustering in
machine learning

  • Classification of unknown data:

    • calculate the feature vector

    • assign it to the cluster
      with the nearest mean

  • Training phase:

    • randomly select a number of
      feature vectors

      • for example 5% of the data

    • run the k-means clustering on
      the selected feature vectors

    • the resulting means are the
      classifier

Classify pixels by color

  • RGB

  • input image

CIEL*a*b* color-space

  • CIELab* color-space
    • L = lightness
    • a = green (-) to red (+)
    • b = yellow (-) to blue (+)

  • Designed, so that
    • distances correspond to perceived
      distances between colors.

Software

  • color clustering in FIJI
  • comes with WEKA
  • Plugins>Segmentation>Color Clustering

Exercises 02