9 Choroidal OCT Analytics
221
Mz =
1
N
N
k=1
z k ,
SDz =
1
N
N
k=1
(z k − Mz) 2 ,
CVz =
SDz
Mz
,
(9.3)
which respectively measure the central tendency, the dispersion, and the standardized
dispersion. In particular, the estimation accuracy of choroidal thickness distribution
is quantified in terms of four quantities—namely, difference (D), absolute difference
(AD), correlation coefficient (CC) and Dice coefficient (DC). Accordingly, we define
specific measures, such as MAD, MCC, SDDC, CVCC, CVDC (replacing z by the
suitable specific quantity). Here note that the standardized dispersion measure CVz
is meaningful only if z is nonnegative. Accordingly, since difference (D) is signed,
CVD is not meaningful and not reported. Further, statistical measures are computed
for results obtained not only by the proposed algorithm (superscripted ‘auto’) but
by manual methods (superscripted ‘ref ’) as well. In particular, the average of the
two manual segmentations is taken as reference in such computations. To ensure
fair comparison, the results are reported vis-à-vis observer repeatability, i.e., the
consistency of performing manual segmentations multiple times by same observer.
A. Difference and Absolute difference: Suppose x i and y i denote the thickness values
at the i-th (i = 1, . . . , N ) column (A-scan index) in two measurements. Then the
difference (D) and the absolute difference (AD) between those measurements at
the i-th column are respectively given by
D i = (x i − y i ),
AD i = |x i − y i |.
(9.4)
For each scan, corresponding mean difference (MD) and mean absolute difference (MAD) are obtained based on (9.3). For the 291 B-scans from the
three datasets, the proposed algorithm achieves MD between −52.98 µm and
13.75 µm with an average of −16.63 µm and standard deviation (SDD) of
19.79 µm, while the MD between the two manual segmentations varies between
−30.67 µm and 20.18 µm with an average of −5.65 µm and SDD of 14.53 µm.
Figure 9.7 provides further dataset-wise details. Difference plots for the three
datasets are furnished in Fig. 9.6. Inspecting those values, there appears to be
a slight negative bias in the proposed method vis-à-vis the reference manual
method, indicating room for further improvement. However, the standard deviations appear to be desirably close.
Proceeding further, the sign of the error is ignored and turned to obtain the absolute difference (AD) measure. The MAD between the estimated thickness and
the reference thickness for our 291 B-scans, are plotted in Fig. 9.8a. To facilitate comparison, MAD between two manual segmentations, measuring observer
repeatability, are also presented. The proposed automated algorithm achieves
MAD between 5.63 and 52.98 µm with an average of 21.88 µm and standard
deviation (SDAD) of 17.43 µm, while the MAD between manual segmentations
varies between 5.55 and 38.71 µm with an average of 13.74 µm and SDAD of
11.67 µm.
221
Mz =
1
N
N
k=1
z k ,
SDz =
1
N
N
k=1
(z k − Mz) 2 ,
CVz =
SDz
Mz
,
(9.3)
which respectively measure the central tendency, the dispersion, and the standardized
dispersion. In particular, the estimation accuracy of choroidal thickness distribution
is quantified in terms of four quantities—namely, difference (D), absolute difference
(AD), correlation coefficient (CC) and Dice coefficient (DC). Accordingly, we define
specific measures, such as MAD, MCC, SDDC, CVCC, CVDC (replacing z by the
suitable specific quantity). Here note that the standardized dispersion measure CVz
is meaningful only if z is nonnegative. Accordingly, since difference (D) is signed,
CVD is not meaningful and not reported. Further, statistical measures are computed
for results obtained not only by the proposed algorithm (superscripted ‘auto’) but
by manual methods (superscripted ‘ref ’) as well. In particular, the average of the
two manual segmentations is taken as reference in such computations. To ensure
fair comparison, the results are reported vis-à-vis observer repeatability, i.e., the
consistency of performing manual segmentations multiple times by same observer.
A. Difference and Absolute difference: Suppose x i and y i denote the thickness values
at the i-th (i = 1, . . . , N ) column (A-scan index) in two measurements. Then the
difference (D) and the absolute difference (AD) between those measurements at
the i-th column are respectively given by
D i = (x i − y i ),
AD i = |x i − y i |.
(9.4)
For each scan, corresponding mean difference (MD) and mean absolute difference (MAD) are obtained based on (9.3). For the 291 B-scans from the
three datasets, the proposed algorithm achieves MD between −52.98 µm and
13.75 µm with an average of −16.63 µm and standard deviation (SDD) of
19.79 µm, while the MD between the two manual segmentations varies between
−30.67 µm and 20.18 µm with an average of −5.65 µm and SDD of 14.53 µm.
Figure 9.7 provides further dataset-wise details. Difference plots for the three
datasets are furnished in Fig. 9.6. Inspecting those values, there appears to be
a slight negative bias in the proposed method vis-à-vis the reference manual
method, indicating room for further improvement. However, the standard deviations appear to be desirably close.
Proceeding further, the sign of the error is ignored and turned to obtain the absolute difference (AD) measure. The MAD between the estimated thickness and
the reference thickness for our 291 B-scans, are plotted in Fig. 9.8a. To facilitate comparison, MAD between two manual segmentations, measuring observer
repeatability, are also presented. The proposed automated algorithm achieves
MAD between 5.63 and 52.98 µm with an average of 21.88 µm and standard
deviation (SDAD) of 17.43 µm, while the MAD between manual segmentations
varies between 5.55 and 38.71 µm with an average of 13.74 µm and SDAD of
11.67 µm.