Influence of Wavelets and Boundary Conditions on the Diagnosis of Multiple Sclerosis Using Artificial Neural Networks

Artificial neural networks with radial basis functions are used to diagnose patients with multiple sclerosis. But the results of the training of this type of network varies greatly with the wavelet selected to compress the data and with the boundary conditions applied to the signal to compensate for the usage of a closed interval. In this paper we compare the results obtained for several wavelets when several boundary conditions are used to extend the signal. The most significant coefficients were extracted following the "Largest Coefficient" criterion. The training process used the left-out method. We did not change the network architecture. The network was trained 20 times for each case, starting each time with a different set of random values. We collected the averages of each of them. Our results showed a remarkable difference between a few of the cases, were the network did a perfect job, and the rest of them, which gave conventional results. The excellent results were obtained by a combination of specific wavelets and particular boundary conditions. The specific wavelets were either of the Biorthogonal Wavelets or Reverse Biorthogonal wavelets family. The particular boundary conditions were the ones obtained by extending the signals using any one of the following approaches: antisymmetric half point padding, smooth padding of order zero or zero padding.