Multiple neural network response variability as a predictor of neural network accuracy for chromosome recognition

Human chromosome classification requires all chromosome appearing in a microphotograph of a dividing human cell to be classified within the known normal or abnormal 24 chromosome types. In recent years, research has focused on the use of neural networks for classification of normal chromosomes. Experimental work in this area led us to question whether learning variability, resulting when multiple neural networks are trained to solve the same problem, could be used as a predictor of classification performance. The Copenhagen chromosome data bank, consisting of 30-component feature vectors from 8106 chromosomes isolated from 180 cells, was divided into a training and a test subsets. Back propagation neural networks with 30 input nodes, 1 to 100 nodes in the hidden layer, and 24 output nodes were trained with the same learning parameters. After training, each neural network was tested. The neural network yielding the best classification was labeled as the optimal neural network. An error variability score was calculated for each test chromosome. This score was a function of all (100) neural network outputs obtained for that chromosome. The error variability scores ranged from 0.16 to 1.31 with a mean value of 0.41 and a SD of 0.12. There was significant difference (p < 0.0001) between the variability scores from chromosomes classified correctly (mean = 0.4, SD = 0.1, n = 3804) and incorrectly (mean = 0.62, SD = 0.19, n = 241) by the optimal neural network. When the variability score was used as a threshold to decide whether or not to accept the output of the optimal neural network, a peak classification rate of 98.93% was observed for chromosomes with an error variability score < 0.35. Results indicate that the error variability of multiple neural network responses can be used as a confidence indicator for a optimal neural network.