Skip to main content

Table 1 Automatic vs. manual measurements of Drusen per image

From: Automated drusen detection in retinal images using analytical modelling algorithms

#

OP1

OP2

OP3

OP4

TE1

TE2

TE3

TE4

AD3RI

CVExperts

CVAD3RI

1

1.8

2.2

2.0

2.2

1.3

1.3

1.6

2.0

3.0

21%

63%

2

3.2

4.4

3.3

4.2

2.9

3.4

2.7

6.2

1.8

30%

52%

3

8.3

4.4

5.8

7.8

6.9

4.3

6.1

7.1

7.4

23%

17%

4

2.7

1.3

0.9

1.2

1.2

1.5

1.5

0.6

1.7

46%

24%

5

4.9

3.7

4.3

5.0

2.0

3.5

5.0

5.7

5.8

27%

36%

6

10.7

5.8

7.6

10.7

10.2

9.1

10.0

18.3

11.9

36%

15%

7

13.0

9.5

9.7

18.1

10.3

14.4

12.3

22.8

14.0

34%

2%

*8

10.1

3.3

0.1

4.5

3.6

3.0

2.2

9.5

5.2

77%

14%

9

13.8

13.8

14.6

20.2

15.1

3.9

20.6

31.7

12.9

48%

23%

*10

7.7

1.8

4.2

2.5

7.6

4.9

10.2

12.0

10.7

57%

68%

*11

2.1

3.9

7.5

17.8

20.3

5.3

20.1

12.1

11.4

67%

2%

12

5.9

5.5

6.7

6.1

6.7

6.6

8.3

7.2

5.6

13%

15%

13

1.2

3.4

2.3

2.3

2.9

2.2

3.7

2.7

1.4

30%

47%

14

3.5

3.6

3.6

5.6

3.8

4.5

6.5

3.5

5.3

26%

23%

15

1.4

5.2

8.4

6.7

5.7

6.6

6.4

7.3

6.1

35%

3%

16

1.9

5.3

4.3

4.8

3.4

4.0

4.5

5.9

4.8

29%

12%

17

1.2

1.6

1.1

0.4

0.8

1.3

1.9

1.4

2.6

38%

113%

*18

1.0

1.1

2.0

0.2

0.8

1.4

1.7

4.1

1.2

77%

19%

19

31.0

25.2

51.0

45.5

44.9

41.7

55.9

63.2

52.2

28%

16%

20

11.7

16.2

18.4

20.5

17.0

15.5

16.7

27.2

19.3

25%

8%

21

9.3

15.8

9.1

7.1

6.7

11.2

14.2

16.1

15.1

34%

35%

*22

12.5

7.3

4.6

0.7

0.1

0.0

0.0

0.8

0.0

141%

100%

  1. This table shows the measurements of drusen areas as a percentage of the total macular area for each expert and for AD3RI. It is also shown the CV among experts and the CV of the automated method when compared to the ground truth for the twenty two test images. As it can be seen in the outlier images signed with an asterisk (*), the CV among experts was significant (> 50%), especially in image #22 where only 3 out of 8 experts graded significant drusen areas. It should also be noticed that no significant differences in the results produced by the group of Ophthalmologists (OP) and by the group of Technicians (TE) were observed. In this table it can be observed that AD3RI obtained a good accuracy with a CV lower than the average CV of the experts in 10 out of 17 images (dataset without outliers). In the remaining 7 images AD3RI obtained a higher CV mainly due to the small drusen areas in which even a small difference in the area value penalizes significantly its CV.