Information Technology Reference
In-Depth Information
was linked with every region in CC view, thus obtaining 51088 links in total.
We added the link class variable with binary values
true
and
false
following
the definition in equation (2). Finally to every region, view, breast and case,
we assigned class values of
true
(presence of cancer) and
false
according to the
ground-truth information provided by pathology reports.
6.2 Evaluation
To train and evaluate the proposed multi-view CAD system, we used a stratified
two-fold cross validation: the dataset is randomly split into two subsets with
approximately equal number of observations and proportion of cancerous cases.
The data for a whole case belonged to only one of the folds. Each fold is used
as a training set and as a test set. At every level (region, view, breast and case)
the same data folds were used. Although we use the results from the single-view
CAD system, we note that the random split for the multi-view CAD system is
done independently. This follows from the fact that the current framework uses
only a subset of the data (cases without CC views have been excluded).
The Bayesian networks at all stages of our
MV-CAD-Causal
model have been
built, trained and tested by using the Bayesian Network Toolbox in Matlab
([33]). The learning has been done using the EM algorithm, which is typically
used to approximate a probability function given incomplete samples (in our
networks the OR-nodes are not observed).
We compare the performance of our
MV-CAD-Causal
model with the per-
formance of the single-view CAD system (
SV-CAD
). As additional benchmark
methods for comparison at the breast and case level, we use the naive Bayes
classifier (
MV-CAD-NB
) and the logistic regression (
MV-CAD-LR
). Both classifiers
use the same inputs as
MV-CAD-Causal
at both levels. The model comparison
analysis is done using Receiver Operating Characteristic (ROC) curve ([34]) and
the Area Under the Curve (AUC) as a performance measure. The significance
of the differences obtained in the AUC measures is tested using the ROCKIT
software ([35]) for fully paired data: for each patient we have a pair of test results
corresponding to
MV-CAD-Causal
and the benchmark systems.
6.3 Results
Classification accuracy.
Based on the results from
ViewNet
, Figure 7 presents
the classification outcome with the respective AUC measures per MLO and CC
view, respectively.
The results clearly indicate an overall improvement in the discrimination be-
tween suspicious and normal views for both MLO and CC projections. Such an
improvement is expected as the classification of each view in our multi-view sys-
tem takes into account region information not only from the view itself but also
from the regions in the other view. To check the significance of the difference
between the AUC measures we test the hypothesis that the AUC measures for
MV-CAD-Causal
and
SV-CAD
are equal against the one-sided alternative hypoth-
esis that the multi-view system yields higher AUCs for MLO and CC views. The
Search WWH ::
Custom Search