1 of 36

Comparative Study of �Retinal Vessel Segmentations

SIIT-Chiba University joint project, October 15-22, 2015

2 of 36

Project Team in Retinal Group

  • Project Advisor:

Asst. Prof. Dr. Pakinee Aimmanee

​

​

​

  • Project head:

Ms. Nittaya Muangnak

(Ph.D Candidate from SIIT)

​

SIIT-Chiba, October 15-22, 2015

2

3 of 36

Project Team in Retinal Group

  • Project members:

1) Mr. Faisal Khan from SIIT

​

​

2) Mr. Kohei Satoh from Chiba University

3) Mr. Takayuki Okamoto from Chiba University

SIIT-Chiba, October 15-22, 2015

3

4 of 36

Project Objective

  • To study existing retinal vessel segmentation techniques
  • To compare performance of those retinal vessel segmentations.

SIIT-Chiba, October 15-22, 2015

4

5 of 36

Project Scope

  • To understand retinal vessel segmentations through following approaches:

1) Local entropy thresholding

2) Gradient orientation analysis

3) Mathematical morphology

SIIT-Chiba, October 15-22, 2015

5

6 of 36

Project Scope (con’t)

SIIT-Chiba, October 15-22, 2015

6

segmentation approaches

Input Image

Output Image

7 of 36

Performance Measures

SIIT-Chiba, October 15-22, 2015

7

Measurements

Description

Recall (Sensitivity)

B/(A+B)

Precision (Positive Predictive Value)

B/(B+C)

8 of 36

Study of Vessel Segmentation through Selected Approach

SIIT-Chiba, October 15-22, 2015

8

9 of 36

Vessel Segmentation Functional Diagram

SIIT-Chiba, October 15-22, 2015

9

Preprocessing to enhance quality of image

Vessel centerline detection

Vessel segmentation

10 of 36

Image Preprocessing

SIIT-Chiba, October 15-22, 2015

10

Preprocessing to enhance quality of image

11 of 36

Retinal Image Labeling into Normal & Abnormality

SIIT-Chiba, October 15-22, 2015

11

12 of 36

Retinal Image Labeled as Normal

  • Image ID: 9 Cases

SIIT-Chiba, October 15-22, 2015

12

13 of 36

Retinal Image Labeled as Abnormal

  • Image ID:

11 Cases

SIIT-Chiba, October 15-22, 2015

13

14 of 36

Relevant Features of Vessel Segmentation

  • High gradient variation/Pixel value
  • Line-like shape
  • Tree-like structure

SIIT-Chiba, October 15-22, 2015

14

15 of 36

Relevant Features of Vessel Segmentation

SIIT-Chiba, October 15-22, 2015

15

​

1) Local Entropy Thresholding

2) Gradient Orientation Analysis

3) Mathematical Morphology

Method based

unsupervised method, matched filtering, mathematical morphology

unsupervised method, multiscale approach, mathematical morphology

unsupervised method, mathematical morphology

Required preprocessing

Yes

No, <green component>

Yes, <green component>

Gradient variation

Yes

Yes

​

Pixel intensity value

Yes

Yes

Yes

Line-like shape

Yes

Yes

Yes

Tree-like shape

No

No

No

16 of 36

Selected Vessel Segmentations through demo applications

SIIT-Chiba, October 15-22, 2015

16

Vessel centerline detection

Vessel segmentation

17 of 36

1) Local Entropy Thresholding

SIIT-Chiba, October 15-22, 2015

17

18 of 36

1) Local Entropy Thresholding (con’t)

SIIT-Chiba, October 15-22, 2015

18

step 1: Green band of image

step 2: Mask image

step 4: Local entropy thresholding

step 5: Short component removal

step 3: Match filtered image

19 of 36

1) Local Entropy Thresholding (con’t)

SIIT-Chiba, October 15-22, 2015

19

step 1: Green band of image

step 2: Mask image

step 4: Local entropy thresholding

step 5: Short component removal

step 3: Match filtered image

20 of 36

2) Gradient Orientation Analysis

SIIT-Chiba, October 15-22, 2015

20

21 of 36

2) Gradient Orientation Analysis

SIIT-Chiba, October 15-22, 2015

21

22 of 36

SIIT-Chiba, October 15-22, 2015

22

After 3 Scales combination

and ridge removal

Green band of image

GOA-scale1

GOA-scale3

GOA-scale2

23 of 36

2) Gradient Orientation Analysis (con’t)

SIIT-Chiba, October 15-22, 2015

23

Inverted ridged removal image

Vessel segmented image

24 of 36

3) Mathematical Morphology

SIIT-Chiba, October 15-22, 2015

24

Original image

Green band

Enhanced green

Histogram eq.

Vessels

Noise Removal

Shade corrected

Vessels

Final Result

25 of 36

Experiment Results on Selected Vessel Segmentation Approaches

SIIT-Chiba, October 15-22, 2015

25

26 of 36

SIIT-Chiba, October 15-22, 2015

26

Original Image

Ground truth

Result1

Result2

Result3

27 of 36

SIIT-Chiba, October 15-22, 2015

27

Original Image

Ground truth

Result1

Result2

Result3

28 of 36

SIIT-Chiba, October 15-22, 2015

28

Original Image

Ground truth

Result1

Result2

Result3

29 of 36

SIIT-Chiba, October 15-22, 2015

29

Original Image

Ground truth

Result1

Result2

Result3

30 of 36

Performance Measures

SIIT-Chiba, October 15-22, 2015

30

Methodology

Image Case

Recall (%)

Precision (%)

1) Local Entropy Thresholding

Normal case

76.31

64.55

​

Abnormal case

69.37

55.95

2) Gradient Orientation Analysis

Normal case

61.62

66.60

​

Abnormal case

54.92

54.10

3) Faisal’s approach

Normal case

73.58

56.18

​

Abnormal case

60.36

40.16

Highest accuracy in normal cases is in red text

Highest accuracy in abnormal cases is in blue text

31 of 36

Outstanding Performance & Limitation of the Studied Vessel Segmentations

SIIT-Chiba, October 15-22, 2015

31

​

Outstanding Performance

Limitation

1) Local Entropy Thresholding

  • simplicity method
  • well perform in distinguishing between vessel and background
  • due to the presence of lesion in abnormal cases, lead to return false detection and non-vessels also be detected

2) Gradient Orientation Analysis

  • the segmented vessels can be yielded varying diameter of blood vessel at multiple scale
  • shading problem can lead false detection

3) Faisal’s approach

​

  • segmented vessels are not continuously detected, so we obtain incomplete result.

32 of 36

Discussion

  • We found that false detection in this method is due to low contrast and non-uniform illumination of images.

→ We recommend existing grayscale generating approaches which are following methods.

1) NTSC Coef. Method

Grayscale = 0.298912 * R + 0.586611 * G + 0.114478 * B

2) ITU Coef. Method　etc. (To generate gray-scale image method)

X = 1.0 – 3.0

R = (R^X) * 0.222015

G = (G^X) * 0.706655

B = (B^X) * 0.071330

Grayscale = (R + G + B)^(1 / X)

​

SIIT-Chiba, October 15-22, 2015

32

33 of 36

Discussion (con’t)

​

  • We can combine these three approaches to outperform vessel segmentation results.

SIIT-Chiba, October 15-22, 2015

33

34 of 36

Conclusion

  • We conclude that vessel segmentation approaches require not only using Green component but also Red and Blue band combination.
  • Vessel segmented results with high robustness still require preprocessing scheme and additional anatomical constraints to separate the lesions in the final vascular tree.
  • Application of multi-scale approach can yield more accurate in vessel diameter variation.

SIIT-Chiba, October 15-22, 2015

34

35 of 36

Suggestion from our Presentation

  • Same constraint of finding threshold value should be the same in comparative vessel segmentation approaches.
  • Do more review in Gradient Orientation Analysis approach.
  • Vessel segmentation approaches are not only focused on the accuracy of detection, but structure itself is also important.
  • Three results from three approaches should be combined to yield the outperform segmented results.

SIIT-Chiba, October 15-22, 2015

35

36 of 36

Thank you

SIIT-Chiba, October 15-22, 2015

36