Image Processing Flashcards

1
Q

How does windowing and levelling work?

A

Window is how much of the greyscale you will display and level is the centre
Everything outside of this window will be black or white

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

How does thresholding work and when does it work best?

A

Binary windowing

Works best for robustly quantitative imaging modalities with large intensity differences

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

What are some clinical examples of thresholding?

A

Setting patient external
Bone outlining

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

Why does thresholding not work well for MR?

A

Image intensities are variable, same sequence: different intensity spread and different threshold required

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

What are kernel operators used for?

A

To change image in a desired way: smoothed or sharpened

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

How are kernel operations done?

A

A matrix of pixel weighting factors are applied to each pixel in an image using a convolution:

Centre kernel on a pixel
Multiply the overlying values and add together
Divide by sum of kernel values

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

What is the purpose of registration?

A

To align two images so that the same anatomy represented in each image is overlaid

Optimisation problem: no analytical solution

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

What are the four ingredients of a registration algorithm?

A

Metric: measure of image similarity
Transform: function which maps moving image to fixed image
Optimiser: method used to find the transform which optimises the metric
Interpolator: for resampling during transforms

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

What is the simplest metric for registration?

A

Mean square difference
Minimise mean square difference

Difference between fixed image intensity at a position and the intensity of the moving image at that position

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

What is the simplest transform for registration?

A

A translation

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

What does the optimiser do (msd example)?

A

Calculate all possible transforms and mean square difference values for each
Find the minimum value for the optimum registration

Not practical to calculate all values in more complex methods

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

How can interpolation be done?

A

Nearest neighbour
Linear interpolation
N-th order B-spline

Moving image needs to be resampled to match fixed image frame of reference

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

When can mean square difference be used?

A

On unimodal registrations (CT-CT)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

What is mean square difference sensitive to?

A

Small number of voxels with large intensity differences

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

What is the most common metric used in RT?

A

Mutual information
Intensities don’t have to be the same: multi-modal

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

What is entropy?

A

A measure of the disorder of a system

In the context of messages, tells you the amount of information in a message (0 if can only say one word..)

Maximum entropy would be every pixel is a different colour, H=0 is black image

17
Q

How does joint histogram work?

A

Each pair of points is plotted in joint histogram showing the frequency of the combinations in grey scale

Work out joint probability distribution and then entropy
Greater entropy, poorer registration

18
Q

What is equation definition of mutual information?

A

I(A,B) = H(A) + H(B) - H(A,B)

Maximise mutual information by minimising entropy of joint histogram
Addition of entropy ensures optimum value doesn’t just overlap backgrounds

19
Q

What types of registration can you have?

A

Rigid (translation or translation and rotation)
Affine (rigid plus scaling and shearing)
Deformable (no geometric relationship in moving image preserved)

20
Q

What are advantages and disadvantages of rigid registration?

A

+ patient geometry is not altered

  • requires patients to be set up the same way
  • significant problems trying to use diagnostic images outside of the brain
  • cannot cope with differences in internal anatomy
21
Q

When is rigid registration used?

A

On-treatment verification
Registering images for contouring

22
Q

How many degrees of freedom in deformable registration?

A

Up to 3N where N is the number of voxels in the image

23
Q

How is deformable registration initialised?

A

With prior rigid registration

24
Q

What is deformable registration used for?

A

Dose tracking

25
Q

What are the issues with deformable registration?

A

Ensuring deformation models the actual change in the patient - realistic deformations difficult - bladder shape with bladder filling deforms differently to weight gain
Evaluating algorithms - need realistic gold standards, phantoms are not realistic and patients not standard
Assessing registration accuracy - visual assessment is gold standard

26
Q

How does atlas based image segmentation work?

A

Have an atlas of images with expert contours
Acquire new image of same modality
Deformably register each atlas image to new image
Transfer atlas contours to new image using DIR
Combine contours to produce one final contour set (there are different ways to do this)

27
Q

What are issues with atlas based image segmentation?

A

Number of patients in atlas (more patients, better accuracy but slower)
Unusual patient anatomy is difficult
Requires atlas images to be similar to the current image
May produce nonsense - edits reduce time saving
Depends on accuracy of contours in atlas

28
Q

How does deep learning image segmentation work?

A

Starts with atlas based segmentation
Involves multiple layers with each layer using output from other layer
Each layer has parameters
Most common architecture is U NET which uses multiple CNN layers with different sampling resolutions, can extract features at different resolutions

29
Q

What are issues with deep learning image segmentation?

A

Depends on quality of manual contouring
Requires large amount of training data
Training times are long
Requires image to be similar to images used in training data
Blackbox

30
Q

What are some evaluation methods of image segmentaion?

A

Overlap metrics (no real correlation to whether volume is clinically acceptable)
Distance metrics (max/mean distance to agreement)
Impact on planned DVHs
Clinician rating (most clinically relevant but requires clinician time)

31
Q

What are three forms of image processing and when are they used?

A

Processing (windowing, thresholding, kernel operations), used at planning

Registration, used at planning, verification, dose assessment

Contouring, used at planning and dose assessment