Image data manipulation

Question 1

Why is digital image processing useful?

Answer 1

Allows the acquired image to be manipulated to improve the display

Question 2

What is a digital image?

Answer 2

A function with x and y co-ordinates and an intensity function (greyscale function), each with finite discrete quantities

Question 3

What are the three approaches to image processing?

Answer 3

Point operations
Local operations
Global operations

Question 4

What are point operations?

Answer 4

Calculate a new pixel value for each pixel in an image

Question 5

What are the five common point operations?

Answer 5

Inversion
Contrast enhancement/stretching
Thresholding
Windowing
DICOM display calibration curve

Question 6

What is windowing and leveling?

Answer 6

Applies a look up table to adjust the relationship between pixel values and displayed values

Question 7

What are local operations?

Answer 7

The application of a kernel to the image via a convolution in order to apply a filter or window a portion of the image

Question 8

What are global operations?

Answer 8

Processes that manipulate the image as a whole with a calculation performed in frequency space rather than real space

Question 9

Why are global operations performed in frequency space?

Answer 9

The same processes are performed quickly with fewer steps and calculations

Question 10

How do global operations work?

Answer 10

Apply a fourier transform to both the image and the kernel
Convolve in fourier space by multiplying the transformed image and kernel
Apply the inverse transform

Question 11

What image registration and what is its purpose?

Answer 11

The process of finding the spatial transforms that maps points on an object in one image to points on an object in another image. Usually used for aligning medical images

Question 12

What is the definition of image registration?

Answer 12

The mathematical operation of aligning two or more image datasets so that similar or complementary information can be transformed onto a common reference

Question 13

What is the definition of image fusion?

Answer 13

The process of combining information in two or more image datasets into a more informative display. Accurate image registration is a pre-requisite of image fusion

Question 14

What are the four things that are needed for an image registration algorithm?

Answer 14

A metric
A transform
An optimiser
An interpolator

Question 15

What is the metric?

Answer 15

A measure of how similar the two images are, often the mean square difference. This is what is optimised

Question 16

What is the transform?

Answer 16

A mathematical method for moving the one image to match another, ie a transformation

Question 17

What is the process of optimisation?

Answer 17

Minimise the MSD using transformations - calculate it for all possible transforms
Select the transform with the lowest MSD

Question 18

What are the three types of transformations and how many degrees of freedom do they allow?

Answer 18

Rigid - translations 3DoF, Rotations 6DoF
Affine - Scaling/sheering 12DoF
Deformable - n DoF

Question 19

How is interpolation used in image registration?

Answer 19

If the spatial moves do not correspond to whole numbers of pixels can interpolate between overlapping pixels in spatial coordinates
Must be carried out during optimisation

Question 20

What are the three common interpolation methods?

Answer 20

Nearest neighbour - intensity of voxel nearest in distance is returned
Linear - returned value is a weighted average of the surrounding voxels, with the distance to each voxel taken into account
Polynomial

Question 21

Why is the B-spline value of an interpolation method important?

Answer 21

Higher B-spline = more accurate interpolation but more computing power required

Question 22

What are the five most common kinds of metric used?

Answer 22

Mean squared difference
Sum of squared difference
Correlation coefficient
Ratio image uniformity
Mutual information

Question 23

When are MSD and SSD metrics used?

Answer 23

Geometric only problems
Pixel intensity based problems only
Only mono-modality as very sensitive to a few pixels with large intensity differences

Question 24

How does mutual information work?

Answer 24

Uses entropy and produces image histograms for the separate and combined images. Therefore the on image probability predicts the other image

Question 25

What is the equation for the entropy of a message?

Answer 25

H = nlog(s) where n = length of message and s = number of symbols

Question 26

As the number of messages increases what happens to the amount of information gained and the uncertainty of it being each message?

Answer 26

Information and uncertainty increases

Question 27

How did Shannon entropy include the probability of a message?

Answer 27

H = sum((1/s^n)log(s^n)) - a weighted measure of entropy | The rarer an event, the more information contained

Question 28

How is Shannon entropy used for images?

Answer 28

It is a probability distribution of the occurrence of grey values low entropy - zero info High entropy - equal quantities of lots of grey values - lots of info

Question 29

What is mutual information used for?

Answer 29

Registration of multimodality images. Assumes regions of similar grey levels in one image correspond to the same region in another

Question 30

How are joint histograms constructed?

Answer 30

Intensities/probabilities from one image are plotted against the second image to create feature space Feature space is constructed by counting the number of times a combination of grey values occur x axis is from image 1, y from image 2 Each time the particular combination of grey values are found the entry in feature space increases

Question 31

How are joint histograms used for optimisation?

Answer 31

Histogram shows increasing dispersion as the mis-registration increases Can estimate joint probability distribution using H = -sum(p(i,j).log(p(i,j)))

Question 32

How can difference images be used in opitimsation?

Answer 32

Subtract the two images - perfectly aligned gives a uniform intensity, slightly misaligned gives edge features Therefore there is an increased entropy, so register by iteratively attempting to minimise the entropy of the difference image

Question 33

What is the mathematical equation for mutual information?

Answer 33

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

Question 34

What is the advantage in using mutual information over joint entropy?

Answer 34

It includes the entropies from the separate images so it will not produce low values for complete misregistrations where only the backgrounds overlap - it is also less sensitive to the size of the overlap

Question 35

How is image registration accuracy assessed?

Answer 35

In commissioning - through the use of gold standards | In clinic - visually