An Overview of Medical Image Segmentation

Question 1

What are many different types of objectives for segmentation?

Answer 1

1. Identify grey matter, white matter and CSF
2. Identify tumours, strokes etc
3. Identify different brain structures, hippocampus etc

Question 2

What are the many different types of iamges?

Answer 2

Segmentation quality depends on what is visible in the images

Question 3

What is region growing?

Answer 3

Approach available within many software tools

Define an approximate region within a structure
Scribbles, points, etc
Maybe define some regions outside the structure

Question 4

What does algorithm try to find?

Answer 4

A suitable boundary

Lots of different strategies involved

Question 5

What is classification?

Answer 5

Rather than classify individuals (e.g. patient vs. control), the aim is to classify voxels (e.g. hippocampus vs. not hippocampus).

Question 6

What are the information that can be combined probabilistically for classificatiion?

Answer 6

Voxel intensity
Position within the image
Classes of neighbouring voxels
Texture

Question 7

What is the probability for classifitis (Fatal disease afflicting one in every thousand people)

Answer 7

Out of 1000 people, 999 will not have classifitis.
Of these, 5% will have a positive result – about 50.
One person will have classifitis, with a 99% chance of having a positive result.

About one positive result from someone with classifitis.
About 50 positive results from people without classifitis.
A positive result means about a 2% chance of having classifitis.

Question 8

What does the Bayes rule say?

Answer 8

P(C+ | T+) = P(T+ | C+) P(C+) / P(T+)
= 0.99×0.001 / (0.99×0.001 + 0.05×0.999)
= 0.0194  2%

Question 9

What is tissue classification?

Answer 9

Whereby voxels are assigned to a tissue class according to their intensities the intensity distribution ofeach tissue class needs to be characterised, often from voxelschosen to represent each class.

Question 10

What is another approach?

Answer 10

involves some kind of registration, where atemplate brain is warped to match the brain volume to besegmented (Collins et al., 1995).

Question 11

Why is automated and reliable intensity-based tissue classification complicated?

Answer 11

Spectral overlap of MR intensities if different tissue classes and by the presence of a spatially smoothly varying intensity inhomogeneity or bias field

Question 12

What is iterative expectation-maximisation (EM) procedures?

Answer 12

Interleaved tissue classification with estimation of tissue-class-specific intensity models and bias field correction

Question 13

What does initialisation of the iterative process use?

Answer 13

Digital brain atlas with a priori probability maps for the different tissue classes

Avoids all manual intervention

Question 14

What is the Markov random field (MRF) designed to facilitate?

Answer 14

Discrimination between brain and non brain tissues, while preserving the detailed interfaces between the various tissue classes within the brain

Question 15

What does the algorithm iteratively interleave?

Answer 15

1. Voxel classification
2. Intensity distribution parameter estimation
3. MR bias field correction
4. MRF parameter estimation

Question 16

What are MR images?

Answer 16

Theoretically piecewise constant with a small number of classes

Have a relatively high contrast between different tissues

Question 17

What does the contrast in an MR image dependent upon?

Answer 17

The way the image is acquired

Question 18

How is it possible to highlight different components in the object being images and produce high contrast images?

Answer 18

By altering radio frequency and gradient pulses, and by carefully choosing relaxation timings

Question 19

What two features facilitate segmentation?

Answer 19

1. Altering radio frequency

2. Gradient pulses

Question 20

What is the piecewise constant property degraded considerably by?

Answer 20

1. Electronic noise
2. Bias field
3. Partial volume effect

Question 21

What does the piecewise constant property cause?

Answer 21

Classes to overlap in the image intensity histogram

Question 22

What do many T2-weighted and proton density images have?

Answer 22

Low contrast between GM and WM

Question 23

What is widely employed for brain MR image segmentation?

Answer 23

Statistical approaches

Labels pixels according to probability values which are determined based on the intensity distribution of the image

Question 24

What does statistical approaches attempt to solve?

Answer 24

Problem of estimating the associated class label, given only the intensity for each pixel

Question 25

What is the most widely used models in segmentation?

Answer 25

1. Finite mixture (FM) model

2. Finite Gaussian mixture (FGM) model

Question 26

What is the limitation of FM model?

Answer 26

Spatial information is not taken into account because all the data points are considered to be independent samples drawn from a population

Produce unreliable results

Question 27

What is the typical use of convolutional network?

Answer 27

Classification takes, where the output to an image is a single class label

Question 28

What does the desired output include?

Answer 28

Localisation I.e. a class label is supposed to be assigned to each piece

Question 29

what are the disadvantages of U-net?

Answer 29

1. It is quite slow because the network must be run separately for each patch, and there is a lot of redundancy due to overlapping patches
2. There is a trade-off between localisation accuracy and the use of context

Question 30

What is the main idea of fully convolutional network?

Answer 30

Supplement a usual contracting network by successive layers, where pooling operators are replaced by unsampling operators

These layers increase the resolution of the output

Question 31

What are the features of SPM12 method?

Answer 31

1. Less accurate
2. Intermediate speed (Mins)
3. Generalises fairly well to new scan contrasts
4. Involves image registration
5. requires relatively little labelled training data
6. Partially handles pathologies (lesion etc)
7. A bit interpretable

Question 32

What are features of label-propagation?

Answer 32

1. More accurate
2. Slowest (hours)
3. Generalise poorly to new scan contrasts
4. Involves image registration
5. Requires relatively little labelled training data
6. Does not handle pathologies
7. A bit interpretable

Question 33

What are features of CNN?

Answer 33

1. Very accurate
2. Very fast on GPU(secs)
3. Generalised poorly to new scan contrasts
4. Does not involve image registration
5. Requires lots of labelled training data
6. Handles pathologies well