Ambient Diagnostics - Ambient Intelligence for Scientific Discovery - page 240

Table 3. Test samples (not included in the training dataset).

data set target Db CI a* b* energy entropy

1 10 (P) 1.68 0.62 18.76 9.97 0.25 4.76

2 10 (P) 1.73 1.92 7.75 4.73 0.33 4.59

3 5(HP) 1.64 0.55 9.98 5.35 0.33 4.66

4 5(HP) 1.64 1.23 13.07 5.84 0.19 5.12

5 1 (NP/NHP) 1.63 0.1 11.55 8.77 0.33 4.91

6 1 (NP/NHP) 1.61 0.61 24.56 6.08 0.16 4.64

Table 4. Results for three types (spread factor = 0.4).

data set target GRNN PNN

1 10 (P) 5(HP) 5(HP)

2 10 (P) 1 (NP/NHP) 1 (NP/NHP)

3 5(HP) 4.56 (HP) 5(HP)

4 5(HP) 2.46 (NP/NHP) 1 (NP/NHP)

5 1 (NP/NHP) 1 (NP/NHP) 1 (NP/NHP)

6 1 (NP/NHP) 1 (NP/NHP) 1 (NP/NHP)

We test two variations of radial basis function: probabilistic neural network
(PNN) and generalized regression neural networks (GRNN) provided by Math-
works [62]. Probabilistic neural networks (PNN) are suitable for classification
problems. The PNN model f ( P, T, spread ) takes three arguments: P is an R x
Q matrix of Q input vectors; T is an S x Q matrix of Q target class vectors; and
spread is the width of the radial basis function. To fit data very closely, we use
a spread smaller than the typical distance between vectors. To fit the data more
smoothly, we have to use a larger spread value. Generalized regression neural
networks (GRNN) [61] are a radial basis network that is often used for function
approximation. The GRNN model f ( P, T, spread ) takes three inputs: P is an R
x Q matrix of Q input vectors; T is an S x Q matrix of Q target class vectors;
spread again is the width of the bottom of the radial basis function.
We used 28 samples to train the neural networks and another 6 samples to
test the models. In our first case, we considered three types of targets: Polyps
(P=10), History of Polyps (HP = 5), and the rest of the cases (NP/NHP = C
= HC = H = 1). The test data set is listed in Table 3. We have the results in
Table 4.
In the second test case, we only considered two types of targets: either Polyps
(P = 10), or Non-Polyps (NP = 1). The input data set is listed in Table 5 and
the results are in Table 6.
As the two test scenarios show, PNN performs the same as GRNN. Also we
learned that neural networks work better when the target classes are fewer, e.g.
in our cases, two targets are better than three targets. Although both GRNN
and PNN can correctly identify 2 out of 3 Polyps cases, it just may be a result
of the over-simplified process. It is not necessary to imply that these methods