Segmentation

Question 1

What can segmantic segmentation be used for?

Answer 1

1. Quantitative analysis, e.g. volum
2. Location and extent of e.g. tumor
3. Creating 3D models

Question 2

What are partial volume effects?

Answer 2

Due to coarse sampling, the boundaries of the objects get blurred, meaning a boundary pixel might represent two different objects.

Question 3

How does the problem of intensity inhomogenities arise?

Answer 3

Some images, especially MRI and ultrasound might av different intensities for the same tissue type, denpending on where in the image the tissue is.

Question 4

What is the problem of Anisotropic resolution?

Answer 4

Most 3D imaging has different resoloution along the different axis.

Question 5

Name one example how imaging artifacts may appaer?

Answer 5

Medical imaging equipment might be disturbed by e.g. metal.

Question 6

What is the problem with limited contrast?

Answer 6

When segementing different tissue might have similar intensity resulting in segmentation “leak”.

Question 7

What is morphological variability?

Answer 7

Objects might stretch, compress, change shape e.g.

Question 8

What is the difference between gold standard and ground truth?

Answer 8

Ground truth is the actually solution, most of the time only availible for simulations. Gold standard is a solution createt by an “expert”.

Question 9

What are the problems of gold standard, and how can the be solved?

Answer 9

Problems:

1. Intra-observer variability, the same observer might create different solutions on different occasions.
2. Inter-observer variability, diffeferent observers might create different solutions
3. Requires expert knowledge and is time consuming

Solutions:

1. Same observer can make several solutions
2. Different observers can make several solutions
3. Agreement/ disagreement can be quantified.

Question 10

What is the conceptual difference between precission and accuracy.

Answer 10

Accuracy is high when bias is low, precission is high when variance is low.

Question 11

What is robustness?

Answer 11

The performanace with different noise levels, (outliers).

Question 12

In classifiation, what is:

1. Accuracy,
2. Precision
3. Recall
4. Sensitivity

Answer 12

1. (TP+TN) / (P+N)
2. TP / (TP + FP)
3. TP / P
4. TN / N

Question 13

What is the F1 score?

Answer 13

The harmonic mean of precission and recall:

2TP / (2TP + FN + FP)

Question 14

Name some methods to determine overlapp between reference and computed segmentation.

Answer 14

1. DSC 2*|A and B| / (|A| + |B|), equivalent to F1.

2. Jaquard |A and B|/ |A U B|.

Question 15

What are some limitations of DSC?

Answer 15

Quite huge deformations might give the same score as small translations.

Question 16

How can we meassure surface distance?

Answer 16

We can calculate the distance to the boundary in one segmentation and trace the boundary of the second segmentation in the first one. This will give us pixel wise surface distance.

Question 17

For segmentation, what are the advantages/ disadvantages of intensity thresholding?

Answer 17

Advantages:

1. Simple
2. Fast

Disadvantages:
1. Regions must have homogenous and distinct intesities,
2. Diffcult to find consistant thresholds across several images.
3 Leakage.

Question 18

For segmentation, what are the advantages/ disadvantages of region growing?

Answer 18

Advantages:
1. Fast. 
2. Connected regions
s
Disadvanteges:
1. Requires user selected points. 
2. Leakages and rough boundaries are likely 
3. Regions must be homogenous.

Question 19

What are the techniques used for graph cuts?

Answer 19

Markov random fields.

Question 20

For segmentation, what are the advantages/ disadvantages of graph cuts?

Answer 20

Advantages:

1. Accurate
2. Reasonable efficeint

Disadvantages:

1. Semi automatic, requires input.
2. Diffucult to select hyperparameters.

Question 21

In graph cut, what does the pih_ij(x) term do?

Answer 21

Describes “loss” for adjacent pixels having different values.

Question 22

For segmentation, what are the advantages/ disadvantages of active contours?

Answer 22

Advantages:

1. Can incorportate shape constraints
2. Smooth boundaries avoid leakage

Disadvantages:

1. Complex energy regularization terms
2. Computationally expensive
3. Non-trivial hyperparameters

Question 23

What is the idea behind atlas based segementation

Answer 23

Leverage one or more map/ chart of the object during segementation by doing registration.

Question 24

For segmentation, what are the advantages/ disadvantages of Atlases?

Answer 24

Advantages:

1. Robust and accurate
2. Yields plausible segmentations
3. Fully automatic

Disadvantages:

1. Computational expensive
2. Cannot deal well with abnormalities
3. Not suitable for tumor segementation

Question 25

For segementation, what would be the input to a random forrest?

Answer 25

A patch around the pixel

Question 26

For segmentation, what are the advantages/ disadvantages of random forrest?

Answer 26

Advantages:

1. Robust and accurate (ensamble)
2. Computational efficient
3. Fully automatic

Disadvantages:

1. Shallow model, no hierarchical feauters (unlike NN)
2. No guaranty on connectedness

Question 27

How can we efficiently do segmentation via Dense classification?

Answer 27

Use a convo only net to get a classification map. This map can be upsampled by transposed convolution.

Question 28

What is tiling and why is it used?

Answer 28

For a dense classification, processing the whole image might require too much memory. Splitting it up into tiles will decrease memory consumption, but increase computations.

Question 29

What is multi-scale processing?

Answer 29

Feed the image at different resolutions into two or more different pathways.

Question 30

Why will we often use small kernels?

Answer 30

Several layers with small kernels might give the same receptive field, with fewer parameters.

Question 31

What kind of architecture is the U-Net

Answer 31

Encoder/ decoder FCN network with skip connections (diagram looks like a U).

Question 32

What si a transposed convolution?

Answer 32

Upsampling by adding zeros (zero fill) or nearest neighbour fill and and doing a convolution.

Question 33

What will the different results be of using nearest neighbour fill and zero-fill?

Answer 33

Nearest neighbour fill will usually result in a smoother image, while zero-fill might result in artifacts.

Question 34

Why do we usually not want to use zero-padding for upsampling?

Answer 34

It might result in artificats.

Question 35

What is EMMA?

Answer 35

Ensamble of Multiple Models and Architecures, use several different approaches for segmeting and do a majority vote. Will reduce inconsistent errors (made by model choise) but doesn’t help much for consisten errors (from dataset).