Cell classification by a learning principal component analyzer and a backpropagation neural network
Maricor Soriano, Caesar A. Saloma · Bioimaging · 1995
Efficient image classification of metaphase spreads, unburst white blood cells, and noise is demonstrated using a learning principal component analyzer and a backpropagation network classifier. For an N-pixel image input, the analyzer produces a compact representation consisting of M (M < N) projection coefficients that are computed with respect to a set of principal component images. Sanger's learning algorithm is employed to determine the M principal components of an ensemble of metaphase cells, ordinary cells and noise. The M outputs of the analyzer are utilized as inputs to a three-output feedforward neural network. The use of a principal component analyzer reduces the complexity of the pattern recognition network, making it easy to implement in personal computers. The trained recognition network achieves a classification success rate of 97.6% for image inputs that are not part of the training set, and therefore generalizes excellently.