Frequency Domain Breast Lesion Classification Using Ultra-Wideband Lesion Response - Ultra-Wideband and 60 GHz Communications for Biomedical Applications - page 244

Fig. 6 9-5-5-2 network (C1)

TP/P . It indicates the accuracy
of the network in classifying malignant lesions. The complement of TPR is the false
positive rate (FPR) which may also be labeled as the false alarm rate. This value is
of interest to doctors as a false positive result may cause the patient undue stress.
False negative rate (FNR) is of higher concern than the aforementioned TPR and
FPR in this project. The FPR is known to be found as FNR

True positive rate ( TPR ) is computed as TPR =

( TN/N ).
This is because a high FNR means too many cases of the cancer will be left
undetected which is highly undesirable.
Table 2 details the networks' four main performance indicators for homogenous
environment (no clutter).
From Table 2 , the FNR values of the networks are unsatisfactory. However, be-
cause these preliminary networks are fairly accurate (except for network C1, which
not only has an unsatisfactory accuracy of 65.74 % but also a high false alarm rate of
40.37 %), the results obtained in the homogeneous environment still prove that the
lesions are classifiable with the technique proposed.
The same networks were trained with the backscatter signal obtained from a
heterogeneous environment (with clutter) to observe their performances in more
realistic heterogeneous environment.
The results of this application for the same networks in the heterogeneous
environment are presented in Table 3 .
The general trend of the performance is similar to that shown in Table 2 .K2is
still the network with the highest accuracy although its performance is degraded in
noisy environment.
On the other hand, performance of C1, unlike Kolmogorov networks, has lost only
5 % of its accuracy. This suggests that the binary classification problem is indeed
more complex in the heterogeneous environment and the Kolmogorov networks may
be underfitting the problem.
To improve the results in heterogeneous environment, dimension reduction of the
input vectors is investigated. From the results found earlier, the current networks are
clearly underfitting the problem. From Table 3 , a significant decrease in accuracies
of the networks can be observed when compared to the results obtained for Table 2 .
Classifiers were not able to perform as well in the heterogeneous environment because
the responses from the other tissue present are also detected and the number of
neurons in the networks may not be sufficient, thus causing underfitting.
Principal component analysis (PCA) was used to reduce the number of poles
extracted from frequency response of each lesion. Based on PCA results, only a
few number of poles significantly contribute to the total variations in different lesion

= 1 − TNR = 1 −