Jun Motoike’s research while affiliated with Toyohashi University of Technology and other places

The conventional automatic blood cell classifier uses a set of statistical discriminant functions composed of multistage, 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) arid 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 rising repeated BP alone.

This paper discusses the classification of blood cell images. The image input method using the single CCD color TV camera and the special color compensation filter, as well as the method of region segmentation for the blood cell images, are described. The following elaborations are made in the image input stage to separate the red blood cell and the white blood cell images in a stable way using optical means. (1) The bluish green light near 450 ∼ 500 nm is cut off from the white light. (2) The accompanying light imbalance between the blue light and the green/red light is compensated by the special color compensation filter. In the (region) segmentation of the blood cell images, a logical operation is developed in which the region of the red blood cell is extracted based on the binary images obtained by the threshold processing of the subtracted image. Using those methods, 59 blood cell images are processed and a satisfactory segmentation result is obtained.

This paper describes a classification logic to classify a neutrophil either as the band neutrophil or the segmented neutrophil. In a typical case, a band neutrophil has a horseshoe-shaped nucleus while a segmented neutrophil has separated nuclei. So it is very easy to distinguish one from the other. But in the case where nuclei are overlapped or touching, it is difficult to distinguish one from the other. On the other hand, the ratio of the number of band neutrophils to that of the segmented ones is low in normal samples. In the case that the above mentioned ratio is extremely high, the correlation between the counting of the technologist and that of the system might become worse if the fixed classification logic is applied. We developed an adaptive feedback logic for the neutrophil classification, which has three logics. The first is the logic to classify the typical band-form or the typical segmented-form. The second is the logic to presume the ratio of the number of band neutrophils to that of segmented ones in the slide from the number of the typical band class, that of typical segmented class and the total number of neutrophil class, all of which are obtained by automated classification logic. The third is the logic to classify the nontypical form neutrophil into two classes in the discriminant space where discriminant threshold is selected by the ratio presumed in the second logic. This adaptive feedback logic improved the correlation between the counting of the technologist and that of the system in various kinds of samples.