224
K. K. Vupparaboina et al.
B. Correlation coefficient: For two measurements x i and y i , i = 1, . . . , N , seen
earlier, the correlation coefficient (CC) is defined by [37]
CC =
N
i=1 x i y i
N
i=1 x
2
i
N
i=1 y
2
i
.
(9.5)
The scan wise CC between the estimated choroid thickness and the reference
thickness are now plotted in Fig. 9.8b. In the same figure, the corresponding CC
between the thickness values obtained by two manual segmentations, measuring
observer repeatability, are also plotted for comparison. The proposed automated
algorithm achieves CC between 97.90% and 99.95% with an average (MCC) of
99.54% and standard deviation (SDCC) of 0.31%, across three datasets, while the
CC between manual segmentations varies between 98.18% and 99.97% with an
MCC of 99.77% and SDCC of 0.16%, thus demonstrating the general reliability
of our method.
C. Dice coefficient: Denote the respective sets of pixel indices in the i-th (i =
1, . . . , N ) column of two segmentations by C
x
i (|C
x
i | = x i ) and C
y
i (|C
y
i | = y i ).
Then the Dice coefficient (DC) is defined by [38]
DC =
2
N
i=1 |C
x
i ∩ C
y
i |
N
i=1 |C
x
i | +
N
i=1 |C
y
i |
.
(9.6)
Now scanwise DC between the estimated choroid thickness and the reference
thickness are plotted in Fig. 9.8c, alongside DC between two manual segmentations as a measure of observer repeatability. The proposed algorithm achieves
DC between 88.81% and 98.73% with an average (MDC) of 94.65% and SDDC
of 1.67%, while DC between two manual segmentations varies between 89.44%
and 98.79% with an average (MDC) of 96.73% and SDDC of 1.24%.
9.2.3.1.3 Performance Comparison
A. General comparison: With the above observations, now the performance of
SSIM-based [26] algorithm is compared against that of other reported algorithms. However, it remains problematic to clearly establish the advantages of
recent methods over earlier work. First, experimental datasets used by various
researchers differ in terms of mean wavelengths used in SD-OCT acquisition.
Further, some scans have been obtained from healthy subjects, some from diseased subjects [22–24], some from adult, and some from pediatric subjects [19].
Moreover, some scans have been taken only near the foveal cross section [23,
38], rather than at various vertical locations over a wide range. In addition, a
variety of evaluation criteria, including correlation coefficient (CC) [22], Dice
coefficient (DC) [19, 23, 38], mean border position difference (MBPD) [22, 25],
and mean absolute difference (MAD) [19], have been used. The image quality
and subjective complexity also appear to vary among various datasets rendering
K. K. Vupparaboina et al.
B. Correlation coefficient: For two measurements x i and y i , i = 1, . . . , N , seen
earlier, the correlation coefficient (CC) is defined by [37]
CC =
N
i=1 x i y i
N
i=1 x
2
i
N
i=1 y
2
i
.
(9.5)
The scan wise CC between the estimated choroid thickness and the reference
thickness are now plotted in Fig. 9.8b. In the same figure, the corresponding CC
between the thickness values obtained by two manual segmentations, measuring
observer repeatability, are also plotted for comparison. The proposed automated
algorithm achieves CC between 97.90% and 99.95% with an average (MCC) of
99.54% and standard deviation (SDCC) of 0.31%, across three datasets, while the
CC between manual segmentations varies between 98.18% and 99.97% with an
MCC of 99.77% and SDCC of 0.16%, thus demonstrating the general reliability
of our method.
C. Dice coefficient: Denote the respective sets of pixel indices in the i-th (i =
1, . . . , N ) column of two segmentations by C
x
i (|C
x
i | = x i ) and C
y
i (|C
y
i | = y i ).
Then the Dice coefficient (DC) is defined by [38]
DC =
2
N
i=1 |C
x
i ∩ C
y
i |
N
i=1 |C
x
i | +
N
i=1 |C
y
i |
.
(9.6)
Now scanwise DC between the estimated choroid thickness and the reference
thickness are plotted in Fig. 9.8c, alongside DC between two manual segmentations as a measure of observer repeatability. The proposed algorithm achieves
DC between 88.81% and 98.73% with an average (MDC) of 94.65% and SDDC
of 1.67%, while DC between two manual segmentations varies between 89.44%
and 98.79% with an average (MDC) of 96.73% and SDDC of 1.24%.
9.2.3.1.3 Performance Comparison
A. General comparison: With the above observations, now the performance of
SSIM-based [26] algorithm is compared against that of other reported algorithms. However, it remains problematic to clearly establish the advantages of
recent methods over earlier work. First, experimental datasets used by various
researchers differ in terms of mean wavelengths used in SD-OCT acquisition.
Further, some scans have been obtained from healthy subjects, some from diseased subjects [22–24], some from adult, and some from pediatric subjects [19].
Moreover, some scans have been taken only near the foveal cross section [23,
38], rather than at various vertical locations over a wide range. In addition, a
variety of evaluation criteria, including correlation coefficient (CC) [22], Dice
coefficient (DC) [19, 23, 38], mean border position difference (MBPD) [22, 25],
and mean absolute difference (MAD) [19], have been used. The image quality
and subjective complexity also appear to vary among various datasets rendering