Improved Automatic Classification of Biological Particles from Electron-Microscopy Images Using Genetic Neural Nets

In this paper several neural network classiication algorithms have been applied to a real-world data case of electron microscopy image classiication in which it was known a priori the existence of two diierentiated views of the same specimen. Using several labeled sets as a reference, the parameters and architecture of the classiier (both LVQ trained codebooks and BP trained neural-nets) were optimized using a genetic algorithm. The automatic process of training and optimization is implemented using a new version of the g-lvq (genetic learning vector quantization) and g-prop algorithms, and compared to a non-optimized version of the algorithms, Kohonen's lvq (learning vector quantization) and MLP trained with QP. Dividing the all available samples in three sets, for training, testing and validation, the results presented here show a low average error for unknown samples. Usually G-PROP outperforms G-LVQ, but G-LVQ obtains codebooks with less parameters than the perceptrons obtained by G-PROP. The implication of this kind of automatic classiication algorithms in the determination of three dimensional structure of biological particles is nally discused.