Medical image segmentation for clinical application

Jose Bernal, BEng MSc PhD (e)

Email: [email protected]
Preprocessing
Jose Bernal, BEng MSc PhD (e)
Email: [email protected]
Imaging artefacts
Imaging artefacts

• Effects which appear in the reconstructed images and are not present in the true anatomy
• They can lead to misdiagnosis if not adequately identified

Real lesion coalesce with artefact

Real lesion Artefact
4
Categories of imaging artefacts

• Motion artefacts
• Partial volume
• Reconstruction artefacts
• Magnetic field artefacts
• RF artefacts
• Noise

5
Motion artefacts

• Inconsistencies in image acquisition

Time

Measurements
6
Motion artefacts

• Inconsistencies in image acquisition

Time

Measurements
7
Motion artefacts

• Inconsistencies in image acquisition

Time

Measurements
8
Motion artefacts

• Inconsistencies in image acquisition

Time

Measurements
9
Motion artefacts

• Inconsistencies in image acquisition, leading to blurring and ghosting

10
Partial volume effect

• Caused by slice thickness

T1-weighted T2-weighted FLAIR

(2 mm thick) (2 mm thick) (4 mm thick)
11
Partial volume effect

• Caused by slice thickness

20 mm thickness 2 mm thickness

Image credit: http://www.mrishark.com/partial-volume.html 12

Truncation artefacts

• Too little measurements of the object of interest

Image credits: http://mriquestions.com/gibbs-artifact.html 13

Magnetic field inhomogeneity

• Field inhomogeneities are cause by environmental factors, improper shimming and the use
of short bore magnets

Image credits: https://www.sciencedirect.com/science/article/pii/S0730725X17301005 14

Noise

15
Image preprocessing
Image

17
Image preprocessing

• Image enhancement

18
Image preprocessing

• Image enhancement

19
Image preprocessing

• Image enhancement

“Clean” T2-w scan Segmented PVS

20
Image preprocessing

• Image enhancement

“Noisy” T2-w scan Segmented PVS

21
Image preprocessing

• Correction of intensity inhomogeneity

Original T1-w Skull-stripped Bias field Pre-processed

22
Image preprocessing

• Reduction of uninteresting inter-subject inter-scan variability

23
Image preprocessing

• Reduction of uninteresting inter-subject inter-scan variability

24
Image preprocessing

• Reduction of uninteresting inter-subject inter-scan variability

25
Image preprocessing

• Intensity standardisation
• Denoising

26
Histogram of intensities

Frequency

Intensity
27
Histogram of intensities

Frequency

Intensity
28
Intensity standardisation

• Removed unwanted intensity variations

Image credits: Emily Carvajal 29

Intensity standardisation

• Transform intensities to make tissue modes coincide

Image credits: Emily Carvajal 30

Intensity standardisation

• z-score transformation
• Compute mean intensity and standard deviation
• Subtract from all intensities and divide by standard deviation

31
Denoising

32
Mean filter

• For each voxel, replace its value by mean intensity in its neighbourhood

Original values Filtered values

20 30 20 23 31
20 25 30 25 20
30 20 90 20 22 37
25 28 24 26 20
50 12 26 21 23

mean(30+20+90+20+24)

33
Mean filter

• For each voxel, replace its value by mean intensity in its neighbourhood

Original values Filtered values

20 30 20 23 31
20 25 30 25 20
30 20 90 20 22 37 36
25 28 24 26 20
50 12 26 21 23

mean(25+90+22+20+26)

34
Mean filter

• For each voxel, replace its value by mean intensity in its neighbourhood

Original values Filtered values

20 30 20 23 31 23
20 25 30 25 20
30 20 90 20 22
25 28 24 26 20
50 12 26 21 23

mean(20+20+30)

35
Mean filter

• For each voxel, replace its value by mean intensity in its neighbourhood

Original values Filtered values

20 30 20 23 31 23 24 26 25 25
20 25 30 25 20 24 25 38 24 25
30 20 90 20 22 24 39 37 36 21
25 28 24 26 20 32 22 39 22 23
50 12 26 21 23 29 29 21 24 21

36
Mean filter

37
Mean filter

• What if the neighbourhood is larger?

Original values Filtered values

20 30 20 23 31
20 25 30 25 20
30 20 90 20 22 37
25 28 24 26 20
50 12 26 21 23

mean(30+20+90+20+24)

38
Mean filter

• What if the neighbourhood is larger?

Original values Filtered values

20 30 20 23 31
20 25 30 25 20
30 20 90 20 22 32
25 28 24 26 20
50 12 26 21 23

mean(25+30+25+20+90+20+28+24+26)

39
Mean filter

• What if the neighbourhood is larger?

Original values Filtered values

20 30 20 23 31
20 25 30 25 20
30 20 90 20 22 27
25 28 24 26 20
50 12 26 21 23

mean(20+30+20+23+31+20+25+30+25+
20+30+20+90+20+22+25+28+24+26+20
+50+12+26+21+23)
40
Mean filter

Original 3x3

5x5 10x10

41
Median filter

• For each voxel, replace its value by median intensity in its neighbourhood

Original values Filtered values

20 30 20 23 31
20 25 30 25 20
30 20 90 20 22 25
25 28 24 26 20
50 12 26 21 23

median (25, 30, 25, 20, 90, 20, 28, 24, 26)

42
Median filter

Original 3x3

5x5 10x10

43
Mean vs median filter

Original 3x3 Original 3x3

5x5 10x10 5x5 10x10

44
Protecting details from being filtered

• Idea: if edge is detected, do not filter it

45
Protecting details from being filtered

• Idea: if edge is detected, do not filter it; blur otherwise

Original values Filtered values

0 0 20 23 31 0 0 20
0 0 30 25 20 0 0 30
0 0 90 20 22 0 0 90
0 0 24 26 20 0 0 24
0 0 26 21 23 0 0 26

Drastic change of
values!
46
Protecting details from being filtered

• Idea: if edge is detected, do not filter it; blur otherwise

Original values Filtered values

0 0 20 23 31 0 0 20
0 0 30 25 20 0 0 30
0 0 90 20 22 0 0 90
0 0 24 26 20 0 0 24
0 0 26 21 23 0 0 26

Edge

47
Protecting details from being filtered

• Idea: if edge is detected, do not filter it; blur otherwise

Original values Filtered values

0 0 20 23 31 0 0 20 23 25
0 0 30 25 20 0 0 30 25 24
0 0 90 20 22 0 0 90 20 22
0 0 24 26 20 0 0 24 26 22
0 0 26 21 23 0 0 26 21 23

48
Preprocessing
Jose Bernal, BEng MSc PhD (e)
Email: [email protected]