Вы находитесь на странице: 1из 25

Medical Image Processing Lectures

Medical Image Segmentation


Prof. Leo Joskowicz
School of Engineering and Computer Science
The Hebrew University of Jerusalem, ISRAEL

MIS techniques: classification (1)


1. Manual methods
2. Thresholding and classification
– Shape-based histogram techniques
– Non-parametric optimal thresholding
– Parametric optimal thresholding
– Minimum distance / maximum likelihood methods
3. Edge-based techniques
– Border tracing
– Graph search
– Hough transforms
4. Morphological operators
– Opening, closing, erosion, dilation
2
1
2
MIS techniques: classification (2)
5. Region-based techniques
– region growing
– region splitting and merging
– connected components labeling
– watershed segmentation
6. Deformable models
– active contours
– snakes
– level-set
7. Graph-cut
8. Hybrid methods

MIS methods: user input


Manual: time consuming, error prone, subjective, not
reproducible, but may be the most accurate if an expert is
doing the work (non-fatigued + good interactive tools).
Automatic: segment and partition an image entirely
automatically, with no human intervention. Computer
intensive; may be error prone due to large variety of
image types and image characteristics.
Semi-automatic: Combination of computer processing and
human intervention. The expert initializes the process by
selecting initial boundaries or starting information. After
processing, the expert can correct erroneous results.
Nearly automatic: minimal interaction by end-user
4
2
4
MIS methods: prior knowledge
• Knowledge about the structures of interest can
greatly help in the segmentation process
• Prior knowledge comes in many forms:
• Intensity ranges, distributions
• Shape and morphology
• Adjacencies, texture, …
• Explicit  precoded
• Implicit  derived (learned) from a training set
• Prior knowledge is task and anatomy-dependent
5

6
3
6
Partial Volume Effect -- PVE

1. Manual MIS: cerebelum segmentation

Used to obtain ground truth since it is unbiased 8


4
8
2. Thresholding and classification
Basic idea:
a structure is uniquely characterized by
its intensity values in the image
valid for highly contrasted structures

Basic thresholding approach:


– thresholding between two grey-levels (windowing)
– representative value + range
– derived from histogram
9

Intensity ranges variations in CT

10
5
10
11

11

Example: simple global thresholding

12
6
12
Segmentation with histograms (1)
Object vs. background

background background
object object

Several objects

13

13

Segmentation with histograms (2)


Region-based adaptive thresholding

14
7
14
MIS : simple thresholding

Copyright L. Joskowicz, 2010


15

15

Example: head CT
Three peaks: background,
brain, skull.
Separating thresholds:
11 and 185.

Copyright L. Joskowicz, 2010


16
8
16
Three-class thresholding

background brain tissue skull


Copyright L. Joskowicz, 2010
17

17

Morphology operations

brain mask largest connected


component
18
9
18
Morphology operations
final result

brain threshold dilation erosion


5×5 5×5
Copyright L. Joskowicz, 2010
19

19

Morphology operations
original image dilation set dilation dilation
operation result

erosion erosion original vs


erosion set operation result result 10
20

20
Morphological operators
• Binarize image
• Dilation:
1 1 0 1 0
I = image; B =
1 1 1 1 1
0 1 0

• Erosion: dual of dilation


complement I; reflect B; dilate; complement result

Copyright L. Joskowicz, 2010


21

21

MIS: thresholding example (1)

Abdominal CT Scan Thresholded Image

22
11
22
MIS: thresholding example (2)

After mathematical Surface model Marching Cubes


morphology operations + surface simplification

23

23

MIS with thresholding: limitations (1)


• Works only for segmentation based on intensities
• Not specific to anatomy
• Robust only for images with global uniformity and
high contrast to noise
• Local variability caused by imaging inhomogeneities
and distortions
• Intrinsic assumption is made that the probability of
features is uniformly distributed
24
12
24
MIS with thresholding: limitations (2)
• Choice of threshold can be computed from grey-
level histogram.
• Does not assume any spatial correlation of voxel
intensity.
• Does not take into account the partial volume
effect (PVE)

motivates classification methods

25

25

Limitations of thresholding

26
13
26
Intensity values of brain structures in MRI

27

MIS challenges: intensities overlap

14
28
MIS challenges: intensities overlap

29

Inhomogeneities in MRI images

original image inhomogeneity field corrected image

Image: Univ. of Utah 15


30
3. Edge-based segmentation
Ideal edge Ramp edge

Gray-level profile of a horizontal line through image

31

Image edge derivatives

16
32
3. Edge-based segmentation techniques
• Edge = boundary One-dimensional edge profiles
between two
regions with distinct
gray-level
properties.
• Most edge detection
algorithms compute
a local derivative
operator.

33

33

Edge detection with derivative operators

34
17
34
Derivatives
Derivative: instantaneous rate of change of a function f

Approximate partial derivatives



x f (i , j )  f (i , j )  f (i  1, j )

y f ( i , j )  f ( i , j )  f ( i , j  1)
• Approximate derivatives by convolution with (1,-1)

35

Gradient operators (1)


• 2D first derivatives 
f (i, j)  f (i, j)  f (i 1, j)
x

• Magnitude of the gradient

• Direction of the gradient

• Numerical approximation: f  G x2  G y2  G x  G y
– Roberts Operator
– Sobel Operator  Gy 
 ( x, y )  tan 1  
– Prewitt Operator  Gx 
36
18
36
Image derivatives

Ix = * -1 1 =

-1
Iy = * 1
=

37

Image gradient

Ix2 + Iy2 =
Gonzalez & Woods: Digital Image Processing, 2001.
38
19
38
Gradient operators (2)
Roberts Operator

G f  x, y  f i, j   f i 1, j 1  f i 1, j   f i, j 1 


 Gx  Gy

Gx  Gy 

39

39

Gradient operators (3)


Prewitt operator

M  S x2  S y2 Sx  Sy 

20
40
Second derivative operators (1)
Laplacian operator: 2D equivalent of 2nd derivative

2 f 2 f
 f  2  2
2

x y

2 f x2 f xx  2 f x f y f xy  f y2 f yy

n 2
f x2  f y2

41

41

Second derivative operators (2)


• Second directional derivative
• Laplacian of Gaussian
 x2  y2 
h  x , y   exp   
 2  2  "Mexican hat" operator
 
 2 h  x , y   f  x , y    2 h  x , y   f  x , y 

 x 2  y 2  2 2   x2  y2 
LoG   h  x , y   
2
 
 exp   2  2  
  4
  
 r 2  2    r2 
   exp  
2 
   2 
4
 21
42

42
Cross section of intensity function

43

43

44
22
44
(a) Original image
(b) Convolving
image (a) with
(c) Binarization
(d) Zero crossing

45

45

Example: lung CT slice


original image Laplacian derivative

Sobel derivative Closing

46
23
46
Edge detection: carpal bones

47

Edge detection: vertebra results

48
24
48
MIS with edge detection: limitations
• Produces disconnected boundaries
• Over- of under-segmentation and spurious edges
• Connecting edges and filling in is challenging
• Problem is harder in 3D!

 boundary-based methods

49

49

25

Вам также может понравиться