Biomedical Engineering Reference
In-Depth Information
expert's performance. We also used three global distance-based metrics to provide
an overall indicator of the performance of the algorithm. The mean difference
(MD) is a measure of the segmentation bias. The mean absolute difference (MAD)
quantifies the error in segmentation. The maximum difference (MAXD) indicates
the maximum error in segmentation. These global distance-based metrics are
computed for each image j using the following equations:
N
MD j
=
d j ( θ i ) /N,
(7a)
i =1
N
MAD j
=
|
d j ( θ i ) |
/N,
(7b)
i =1
d θ i
j
MAXD j
=
ma i {|
|}
.
(7c)
2.3.2.2. Area-based metrics Prostate volume is often used in prescribing the
appropriate radiation dose and in interpreting the Prostate-Specific Antigen (PSA)
test for diagnostic purposes and for treatment evaluation. This involves outlining
the prostate in 2D images that span the length of the prostate, then adding up the
areas enclosed by the individual outlines and multiplying by the distance between
slices. Since area calculations affect volume estimates, we used three area-based
metrics — (1) fractional area difference, P d , (2) sensitivity, C s , and (3) accuracy,
C a — to compare the algorithm outlines to manual outlines, which were taken to
be the “true” outlines of the prostate. The three metric are defined as
P d
=
A d /A m ,
(8a)
C s
=
TP/A m ,
(8b)
C a
=1 ( FP + FN ) /A m ,
(8c)
where TP is the true positive area that is the common region between the manual
outline and the algorithm outline, FP is the false positive area that is defined as
the region enclosed by the algorithm outline but outside of the manual outline,
FN is the false negative area that is the region enclosed by the manual outline that
is missed by the algorithm, A m = TP + FN is the area enclosed by the manual
outline, and A d =( TP + FP ) ( TP + FN )= FP
FN is the area difference.
2.4. Results and Discussion
2.4.1. Classification and examples
Approximately 36% of the images were classified as being “easy” to segment,
51% “moderate,” and the remaining 13% difficult. The left-hand column in Fig-
ure 5 gives examples of manually outlined boundaries in the (a) easy, (b) moderate,
Search WWH ::




Custom Search