Metaphase recognition by the backpropagation neural network

Abstract

The error-driven backpropagation algorithm for training a neural network is applied to the recognition of metaphase spreads. Images of human blood cells in the metaphase stage together with non metaphase images are used as training samples for the network. The patterns consist of unipolar gray scaled pixels in a 10x10 mesh. Several modifications to the basic architecture and algorithm are implemented such as locally connected models, combination of sigmoid and gaussian activation functions, and varying learning rate.