Biomedical Engineering Reference
In-Depth Information
2.7 Cross-Validation
All the quantifications were implemented in MATLAB 7.12 (The Mathworks Inc.).
The ability of the congruity measure is cross-validated for several tasks of diagnosis
that include cross-sectional separation of KL 0 vs. KL 1, KL 2 vs. KL 3, KL 0 vs.
KL
1. The scans were randomly divided into training
and validation subsets. Using the training subset, the scale and number of iterations
were optimized for the best AUC, and used the optimized scale and number of
iterations on validation subset to compute the validated AUC. The process was
repeated 100 times and the median value was reported. The congruity indices were
computed for 16 combinations of curvature scale (1.6 mm, 2.4 mm, 3.6 mm, and
5.4 mm) and number of iterations (2, 4, 8, and 16) to experiment and optimize
the scale and number iterations that give best AUC for various diagnosis tasks
mentioned above.
>
0 and KL
1 vs. KL
>
3 Results
The precision of the CI was 7.5
%
that was computed at the scale (2.4 mm) and
iteration number (4) corresponding to the maximum trained AUC of all the tasks
(see Table
2
). In general, for all combinations of scale and iterations, the range of
RMS CV was 6.9
%
-8.5
%
.
Congruity values were higher in healthy knees and lower in knees with OA.
The congruity map for a healthy knee (KL 0) and knee with advanced OA (KL 3)
used in the evaluation was shown in Figs.
2
a and
2
b, respectively. The AUCs for CI
to separate healthy knees from different levels of OA were significant and were
shown in Table
2
.
The diagnostic ability of CI quantified as AUC to separate KL 0 KL1 was 0.64
(
p
<
0.001). Further, the AUC was 0.69 (
p
<
0.0001) for separating KL 0 vs.
KL
>
0, and 0.73 (
p
<
0.0001) for separating KL
1 vs. KL
>
1. The cross-
sectional separation of healthy (KL 0) and knees with OA (KL
>
0) was shown in
Fig.
2
c.
Table 2 Statistical scores to show the ability of the CI to separate KL 0 vs. KL 1, KL 2 vs. KL 3,
KL 0 vs. KL
>
0 and KL
1 vs. KL
>
1 cross-sectionally
Iteration
Scale
Train
Validated
Task
(median
SD)
(median
SD)
AUC
AUC
KL 0 vs. KL 1
2
3
5.4
1.0
0.66**
0.64**
KL 2 vs. KL 3
8
5
2.4
1.4
0.73*
0.63
KL 0 vs. KL
>
0
2
2
2.4
0.3
0.70****
0.69****
KL
1 vs. KL
>
1
4
3
2.4
0.4
0.75****
0.73****
*
p
<
0.05, **
p
<
0.01, ****
p
<
0.0001, SD: Standard Deviation
