Frequency Domain Breast Lesion Classification Using Ultra-Wideband Lesion Response - Ultra-Wideband and 60 GHz Communications for Biomedical Applications - page 245

Table 2 Network performances in homogeneous environment
Network ACC (%) TPR (%) FNR (%) FPR (%)
K1 77.78 87.04 31.48 12.96
K2 79.07 82.96 24.81 17.04
C1 65.74 59.63 28.15 40.37
ACC accuracy TPR true positive rate, FNR false negative rate, FPR false positive rate, K1
Kolmogorov network 1, K2 Kolmogorov network 2, C1 9-5-5-2 network

Table 3 Network performances in heterogeneous environment
Network ACC (%) TPR (%) FNR (%) FPR (%)
K1 62.59 64.44 39.26 35.56
K2 62.78 57.78 32.22 42.23
C1 60.74 56.67 35.19 43.33
ACC accuracy TPR true positive rate, FNR false negative rate, FPR false positive rate, K1
Kolmogorov network 1, K2 Kolmogorov network 2, C1 9-5-5-2 network

responses. Hence, input vectors can be constructed using these highly variable poles
and fed to the neural network for classification.
Table 4 shows the results after reducing the number of poles by PCA.
Comparing the results in Table 4 to that obtained in Table 3 , the increase in the
accuracy of Kolmogorov network is evident. K1 is more accurate by 2.78 % when
the 18-pole inputs were used instead of the 25-pole inputs before PCA. At its best,
K1 is found to be 6.11 % more accurate with the 12-pole inputs. K2 also follows a
similar trend as K1 with its accuracy peaking at the 12-pole inputs scenario. C1 on
the other hand did not perform better after pole reduction except for the 7-pole inputs
case, where it fared marginally better by only 2.22 %.
To further increase the performance of the classifier, a cascaded form of the feed-
forward back-propagation network is proposed. This structure will guarantee more
updates of the biases and weights while they will also be propagated back as shown
in Fig. 7 .
As K1 has the highest accuracy, only the cascaded form of K1 (K1c) was developed
for discussion. The result of the accuracy of K1c in the various scenarios is presented
below (Table 5 ).
The accuracy of K1c is noticeably higher than the previously attained with any
network in the heterogeneous environment.
It is worth noting that PCA results in some loss of detail in the input signals but
this trade-off has allowed for the development of a significantly accurate network
that coupled with the benefits of a cascaded network, is able to attain a performance
comparable to other studies that have achieved an accuracy of more than 70 % [ 32 ].
From the figure and table above, it is clear that K1c with 12-pole vector inputs is
the most accurate network.
Specificity is mathematically defined as TN/ ( TN

+

FP ). In practice, a positive
result in a high specificity test is used to confirm a disease. Specificity of 73.9 %
obtained through this classifier which indicates that should this test be administered