The number of processing elements in hidden-layer of back-propagation neural network for karyotyping ~genetic diagnosis\

Abstract

Back-propagation neural networks with various number of processing elements (PEs) in hidden-layer were examined to karyotype of Giemsa-stained human chromosomes. Two learning sets for the experiments were prepared from randomly selected 460 chromosomes. Learning set A consisted of 27 vectors, which included a relative length, a centromeric index, and 25 density vectors extracted from normalized density profile. Learning set B was the same as the learning set A but it had 50 density vectors. For the two learning sets the classification errors in output layer were examined with various number of PEs in hidden layer. Results of the experiment showed that the minimum classification error was obtained in a model trained with 27 input vectors and 48 PEs in its hidden layer.