Hisatake YOKOUCHI’s research while affiliated with Hitachi, Ltd. and other places

This paper describes an automated classification algorithm of RBC's morphological abnormalities, which is one of the examinations of the blood cell differentials. As there are many RBCs in the image (which is called multi-object pattern), one of the important problems for this automation is to develope the segmentation of the multi-object pattern. We developed a new method of the segmentation of the multi-object pattern, which was done only by the local operations based on the wave propagation method. Feature extraction is done only by the local operations, too. Moreover, we developed a logic to detect various kinds of morphological abnormalities by the partition of multidimensional space, which is defined by the feature parameters of RBC's. Classification rate of morphological abnormalities by this method is 92.6%.

A prototype of automated blood cell analyzer has been developed. The process is performed by the pattern recognition method. Four types of examinations are automatically operated : 1) differential count of normal and abnormal white blood cells, 2) morphological analysis of red blood cells, 3) numerical platelet estimates and 4) reticulocyte count (a slide preparation differing from other three). The WBC differential count is the essential of the process and applies an adaptive and hierarchical classifier.fw Here, the word “adaptive” means self compensation using standard cells in each sample. The standard cells are automatically picked up from input cells. The WBC differential count is processed at the rate of one sample per minute. Net discriminating rate of normal WBC differential count is better than 98%.

The evaluation of several findings taken from fundus photographs is a technique widely used in mass screening of geriatric disorders and diagnoses of hypertension. The examination of fundus photographs is left to the pattern recognition ability of the individual physician. Along with the increasing occurrence of geriatric disorders, the quantity of work could well come to exceed the capacity of the physicians available. Nevertheless, there has been no research to date on the automatic reading of fundus photographs nor has there been any direct discussion of the technical possibility of automation in this area.

Fundus photographs are widely used in diagnosis of hypertensive diseases to facilitate the early detection of changes in arteriolar vessels. This type of disease is responsible for the great majority of fatal illnesses in Japan. Among the many kinds of abnormal phenomena which can be observed in fundus photographs the arterio-venous crossing phenomenon is one of the most important factors in evaluating the degree of arteriolar change ; moreover, the degree of change considered to be quite difficult to recognize automatically. In this paper, the authors describe a system of image processing and an automatic recognition method of crossing phenomena. The processing flow of this system is as follows : 1) Improvement in the quality of images with the aid of color information (pre-processing). 2) Extraction of the contour lines of blood vessels from black and white photographs (feature extraction). 3) Quantitative classification of the grades of crossing phenomena, which agree with visual classification by physicians (classification). A recognition experiment on nine samples has been carried out under a fully automatic system. In the experiment a successful distinction was made between normal crossing sites and abnormal crossing sites, attaining full agreement with the physician's judgment.

Ophthalmoscopic findings have been considered as an important factor in multiphasic screening systems, or in the diagnosis of the hypertensive disease. However, the study of its automatic and quantitative analyzing system has never been tried. In this paper, the authors have made an trial to classify the grade of changes at the arterio-venous crossing in the retina, with development of such an analyzing system in fundus photography as a final goal. Important technical points for developing such a system are as follows : 1 Analysis of chromatic characteristics in the fundus photographs, and a translation technique of photo-images to electrical-images, which should be improved in signal to noise ratio using above result (pre-processing). 2 Extraction technique of contour lines of blood vessels from above images (feature extraction). 3 Quantitative classification of the grade of the diseases by the analysis of above contour lines (recognition). This report is concerned with the second and third points. Results obtained were as follows : 1 The minimum value of the diameter of the retinal vein normalized by, its mean value was found to be a good measure in classifying the crossing phenomena. 2 The two-stage extraction method, i. e., the coarse extraction and the precise extraction, was found to be useful for extracting contour lines of blood vessels, which included some new techniques for the image processing. Using above results, the normal patterns and abnormal patterns in the crossing phenomena were successfully distinguished at the accuracy of 90 percent.