8 Segmentation of Optic Disc and Cup-to-Disc Ratio Quantification …
205
Table 8.3 Unsigned border error for different segmentation algorithms (Mean ± SD)
A-scan Classification Coarse margin
location
Patch searching
Disc (pixel)
3.245 ± 1.531
3.971 ± 1.870
2.216 ± 1.406
Disc (mm)
0.098 ± 0.046
0.120 ± 0.056
0.067 ± 0.042
Cup (pixel)
2.338 ± 0.993
1.140 ± 0.763
1.164 ± 0.869
Cup (mm)
0.070 ± 0.030
0.034 ± 0.023
0.035 ± 0.026
Table 8.4 Dice similarity coefficient for different segmentation algorithms (Mean ± SD)
A-scan Classification Coarse margin
location
Patch searching
Disc
0.905 ± 0.035
0.868 ± 0.057
0.919 ± 0.034
Cup
0.917 ± 0.113
0.925 ± 0.125
0.928 ± 0.116
Table 8.5 Computational time comparison
LBP
HOG
Fusion
A-scan
classification
Time (s)
86.68
12.57
103.04
109.62
Table 8.6 Error of C/D evaluation by different algorithms (Mean ± SD)
A-scan Classification Coarse margin
location
Patch searching
Evaluation error
0.060 ± 0.044
0.073 ± 0.036
0.045 ± 0.033
where S disc is the area of optic disc and S cup is the area of cup, both defined in
a projection image. Area C/D ratios were calculated by the proposed algorithm,
A-scan classification-based segmentation and the reference standard are shown in
Fig. 8.6a, and the error of C/D ratio evaluation (defined as the difference between
C/D ratios calculated by algorithms and the collected manual segmentation) is shown
in Fig. 8.6b. The mean and standard deviation of the error are also summarized
in Table 8.6. The results indicate that the evaluation error of the patch searchingbased segmentation is lower than those of the other two segmentation algorithms
for automated C/D ratio calculation. The coarse margin location causes the highest
error because it fails to obtain the precise disc margin. Our algorithm proposed here
achieved the best performance in C/D ratio quantification.
8.4 Conclusion
In this chapter, a patch search-based optic disc segmentation algorithm has been introduced for quantifying the C/D ratio in SD-OCT volumes. Compared with traditional
205
Table 8.3 Unsigned border error for different segmentation algorithms (Mean ± SD)
A-scan Classification Coarse margin
location
Patch searching
Disc (pixel)
3.245 ± 1.531
3.971 ± 1.870
2.216 ± 1.406
Disc (mm)
0.098 ± 0.046
0.120 ± 0.056
0.067 ± 0.042
Cup (pixel)
2.338 ± 0.993
1.140 ± 0.763
1.164 ± 0.869
Cup (mm)
0.070 ± 0.030
0.034 ± 0.023
0.035 ± 0.026
Table 8.4 Dice similarity coefficient for different segmentation algorithms (Mean ± SD)
A-scan Classification Coarse margin
location
Patch searching
Disc
0.905 ± 0.035
0.868 ± 0.057
0.919 ± 0.034
Cup
0.917 ± 0.113
0.925 ± 0.125
0.928 ± 0.116
Table 8.5 Computational time comparison
LBP
HOG
Fusion
A-scan
classification
Time (s)
86.68
12.57
103.04
109.62
Table 8.6 Error of C/D evaluation by different algorithms (Mean ± SD)
A-scan Classification Coarse margin
location
Patch searching
Evaluation error
0.060 ± 0.044
0.073 ± 0.036
0.045 ± 0.033
where S disc is the area of optic disc and S cup is the area of cup, both defined in
a projection image. Area C/D ratios were calculated by the proposed algorithm,
A-scan classification-based segmentation and the reference standard are shown in
Fig. 8.6a, and the error of C/D ratio evaluation (defined as the difference between
C/D ratios calculated by algorithms and the collected manual segmentation) is shown
in Fig. 8.6b. The mean and standard deviation of the error are also summarized
in Table 8.6. The results indicate that the evaluation error of the patch searchingbased segmentation is lower than those of the other two segmentation algorithms
for automated C/D ratio calculation. The coarse margin location causes the highest
error because it fails to obtain the precise disc margin. Our algorithm proposed here
achieved the best performance in C/D ratio quantification.
8.4 Conclusion
In this chapter, a patch search-based optic disc segmentation algorithm has been introduced for quantifying the C/D ratio in SD-OCT volumes. Compared with traditional