mhj In Vivo Quantitative Assessment of Carotid Plaque component with multi-contrast MRI Jannie Wijnen 22 may 2003 using clustering algorithms, implemented.

/mhj

In Vivo Quantitative Assessment of Carotid Plaque component with

multi-contrast MRI

Jannie Wijnen

22 may 2003

using clustering algorithms, implemented in Mathematica

/mhj

Question from the clinic

•If the highly thrombogenic plaque content is exposed to

the blood, thrombo-embolic incidents as myocardial

infarction, stroke, and peripheral vascular disease might

occur

•Design a program that automatically detects the

various components of the atherosclerotic plaque

• The program should be accurate and easy to use

/mhj

Solution to the question

•5 different MR images of the plaque -> notebook

•combining these images in a 5-dimensional

feature space

•search for clusters in feature space that represent

a tissue (hemorrhage, fibrous tissue, calcium and

lipid)

•show these clusters in the original image

/mhj

Feature space

Feature space:

An n dimensional plot showing the

combination of grey values in each of the n

images for all corresponding points

Example:

/mhj

ik10 3 74 6 71 8 0

y{0

2.5

5

7.5

10

2

4

6

8

2

4

6

8

10

0

2.5

5

7.5

10

2

4

6

8

/mhj

Clustering

Example of 3D feature space of the images

Kmeans cluster algorithm : Place K points into the

space represented by the objects that are being

clustered. These points represent initial group

centroids. Assign each object to the group that has the

closest centroid

Example of clusters in 3D space

/mhj

Problems

For the best classification the clusters should

be small and well- separated from each

other

•background intensity

•image miss-registration

•noise

/mhj

Background intensities

Example of clusters in images with a large intensity

gradient

Background correction by minimisation of

entropy of image with a known gradient.

Conclusion:

no background correction needed in these

images.

/mhj

Image Registration

•Pixels that are compared to each other must

represent exactly the same anatomical positions.

Example of non- registered images

•Minimisation of the mutual information (entropy)

leads to the displacement needed to register the

images.

Example of registered images

/mhj

Noise Reduction

The less noise the smaller the clusters

Euclideanshortening: preserves the

edges

example

Variance of the clusters after

euclideanshortening is smaller

/mhj

Starting Values

Difference between the clustering with randomly chosen starting values and starting values chosen by user

/mhj

Starting values

•The clustering is not reproducible when

starting values are chosen by hand

•It is better to extract starting values form

the information in the 5D space

•try to find starting values from the

maximum intensities in the 5D space

/mhj

Resultsklusters T1w_TSE PDw T1w_TFE T2w_TSE T2w_FSE

1067klusters_eucl T1w_TSE PDw T1w_TFE T2w_TSE T2w_FSE

Calcium is too small in lower serie

information loss by euclideans shortening

/mhj

Resultsklusters1025 T1w_TSE PDw T1w_TFE T2w_TSE T2w_FSE

klusters_eucl1025 T1w_TSE PDw T1w_TFE T2w_TSE T2w_FSE

The algorithm does not find hemorrhage and lipid core, the two dangerous components

/mhj

Discussion /Conclusion

•Cluster techniques can be used for

tissue detection in atherosclerotic plaque

•A lot of progress can be made:

starting values, cluster technique, non-

linear registration, smoothing.