The conventional automatic blood cell classifier uses a set of statistical discriminant functions composed of multi stage, tree structure, linear and second order functions. In this paper, we tried to replace these discriminant functions with three layer neural networks (NN). In this case, chosing the number of hidden layer units becomes a problem. So, we focus our discussion mainly on the optimization of the number of hidden layer units, using a gradual reduction method (GRM), a method we previously proposed. In this paper, we use GRM and examined its effectiveness. We used and compared two types of back propagation (BP) algorithms, standard BP (Std-BP) and BP with forgetting (BP-F), with GRM. From the results, GRM+BP-F is found to suit the above purpose better, from the stability of the recognition curve and from its independency to initial weight conditions. Further the recognition rate is stable, i. e., we don't need to learn NN again. Also we found that the computation time can be shortened with GRM+BP-F as compared to the case of learning using repeated BP alone.