A comparison of wavelet and Fourier descriptors for a neural network chromosome classifier

This paper compares the efficacy of wavelet and Fourier descriptors in neural networks used for chromosome classification. The backpropagation (BP) neural network architecture was used. Absolute chromosome length and wavelet or Fourier coefficients derived from the densitometric profile formed a feature vector for each chromosome. Four learning sets for both wavelet- and Fourier-based networks were prepared from 1584 randomly selected chromosomes. When the test sets consisted of intact chromosomes, the best classification accuracy of the Fourier-trained networks was 90.3%; for wavelet-trained networks, it was 87.5%. The wavelet networks took less time to stabilize and the best wavelet classifier required fewer coefficients than the best Fourier classifier for similar results. The strengths of both wavelet-trained and Fourier-trained networks were seriously compromised when truncated chromosomes were included in the test sets, with the wavelet networks yielding a higher percentage of misclassified chromosomes (best classification accuracy of 53.3% correct for Fourier-trained networks, and 38.5% for wavelet-trained networks).