Automated Method and Apparatus for Classification of Cells With Application to the Diagnosis of Anemia

This invention relates to an apparatus for automatically analyzing blood, and through the accumulation of measured properties from individual cells, of thereby classifying the blood specimen according to its close resemblance to the normal or to various pathological conditions. More particularly it is concerned with automatically classifying red blood cells and by accumulating measurements relative to each cell generating characteristic values, which by automated means identify the given blood specimen as typical of either normal or of the pathological condition of a specific type of red cell disorder or anemic condition.

According to present medical practice, the diagnosis of a particular one of more than a dozen major types of anemia uses three broad categories of information; (1) mean descriptors of cell number, size and hemoglobin content, (2) subjective microscopic visual evaluation of the stained blood cells by a trained hemotologist, and (3) specific biochemical or other tests to pinpoint the precise cause of the anemia.

With conventional equipment, the most common of the first category of red cell sample descriptors are; (1) the red cell count, or the number of red cells per unit volume of blood, (2) the hemoglobin content, or amount of hemoglobin per unit volume of blood, (3) the packed cell volume, or the percentage of blood occupied by red cells, (4) a mean cell size parameter, usually taken as the mean cell volume, which is derived by dividing the packed cell volume by the red cell count, (5) a mean cell hemoglobin parameter, which is derived by dividing the total blood hemoglobin content by the red cell count, and (6) the mean cell hemoglobin concentration, which is derived by dividing the total cell hemoglobin by the packed cell volume.

In the second category, a subjective visual evaluation of the stained blood specimen relates to the tedious and time consuming process by a hemotologist of examining a blood film under the microscope and identifying characteristic abnormal cells, such as large cells, or macrocytes, small cells such as microcytes, target cells, elongated cells such as sickle cells, and giving estimates of size variation e.g. anisocytosis +1, +2, or +3, or a subjective evaluation of population shape changes, such as poikilocytosis +1, +2, or +3.

Today, in addition to the overall mean red cell parameters, and the visual descriptions, certain biochemical and other sophisticated tests are often performed to further clarify the pathology of the anemia. These include; iron kinetics tests, serum iron level tests, hemoglobin electrophoresis, folic acid level tests, vitamin B12 tests, and the extraction of a bone marrow sample for evaluation of maturation changes and stainable iron levels. These are very time consuming and expensive, and as to the extraction of a bone marrow sample, very painful.

In accordance with the above, one aspect of the present invention replaces the tedious visual examination with an automatic classification of the normal and abnormal red blood cells into sub-populations and extracts meaningful red blood cells parameters for the separate sub-populations. These and other parameters are used for the automatic classification of the blood specimen with respect to the categories of anemia. An example of automatic classification of red blood cells into sub-populations is fully disclosed in my co-pending continuation-in-part application, Ser. No. 825,673, Filed Aug. 18, 1977. As noted there, one difficulty encountered in separating the normal and abnormal cells into meaningful and widely recognized sub-populations on an automated basis is that of accurately segregating the cells by their morphology and color, particularly where their respective areas or sizes and shapes overlap and their respective principal distinguishing feature is the configuration of their respective central pallors (or a lack of central pallor). Central pallor is the thin, disc-shaped central area of red blood cells which may be circular and particularly pronounced for some cells. For instance, target cells and normocytes may have substantially the same size area and shape, but differ in a central pallor configuration. Thus, to distinguish between these cells, and subsequently to distinguish between anemias, the automated analysis should be able to examine and classify cells on the basis of their interior configuration, as well as their exterior configurations.

Other cells, such as spiculed red blood cells, may have the same general size, area and interior configurations of normocytes or the like, but are distinguished principally by their indented spiculed perimeters. Likewise, adding to the difficulty of classifying abnormal cells such as sickle cells from other elongated cells is that they may have similar peripheral measurements, sizes, and areas, but differ principally from one another in the presence of pointed projections, or spicules. Still other abnormal cells may be separately categorized from other morphologically similar cells only by their hemoglobin content, measured in terms of color or density. Therefore, it may be desirable to distinguish the hypochromic cells from those that are normochromic.

It has been further discovered that the classification of the red cells into subpopulations is not always discerning enough with regard to automatic anemia classification, but that further statistical descriptions of these subpopulations are often required. For example, different anemia bloods may have the same percentage of round biconcave cells but show considerable variation in the dispersion of the cells with regard to size, hemoglobin content and central pallor. Other anemias may result in blood with the same percentage of elongated or spiculed cells, but differ significantly with regard to measures of mean size, hemoglobin, or with regard to the total population skewness relative to a shape measure.

As will be explained in greater detail hereinafter, the present invention is described in connection with a microscopic slide, digital image and pattern recognition system. However, the invention is not to be construed as limited to such a system, as the feature analysis of the sample to classify the blood and to test the blood for abnormalities may be performed using other techniques, such as a coherent optical analysis technique disclosed in U.S. Pat. No. 3,947,123; or a liquid flow process technique such as disclosed in U.S. Pat. Nos. 3,819,270 and 3,822,095. To be commercially feasible, the digital image and pattern recognition process for the blood cells should operate on a real time basis and with sufficient speed and accuracy that it will perform as well as the now commercially accepted leukocyte differential counting systems, such as for example, the LARC manufactured by Corning Glass Works of Corning, N.Y., and generally disclosed in U.S. Pat. No. 3,883,852.

The Coulter Counter, manufactured by Coulter Electronics, Hialiah, Fla., provides results which are helpful in diagnosing anemia in that it provides a red blood cell count and mean red blood cells parameters characterizing the entire population of cells; more specifically, the Wintrobe indices of mean cells volume, mean cell hemoglobin and mean cell hemoglobin concentration along with the number of red cells per cubic millimeter. However, no differentiation between abnormal or normal red blood cells is achieved with the Coulter Counter. Furthermore, the hemoglobin content for individual cells is not determined and measures of dispersion and skewness are not performed. Finally, no automatic categorization or quantitative direct association with known anemias or other pathologies is automatically included as part of the analysis.

Heretofore, some offline experimental work has been performed on image processing of erythrocytes. One of these works, "Bentley, S. A. and S. M. Lewis, The Use of an Image Analyzing Computer for the Quantification of Red Cell Morphological Characteristics, Brit. J. Haemat. 29:81, 1975", describes an offline analysis of dried and stained red blood cells of a total cell population measuring three red blood cells parameters by an image analysis technique. This analysis is similar to the Coulter Counter analysis in that the parameters measured were from the total population of cells being analyzed, and were analogous to the Wintrobe Indices. The drying of the red cells introduced artifacts, and there was a lack of central pallor, or internal red cell analysis to provide a highly refined classification. Moreover, there was not disclosed the capability for differentiating between and classifying normal red blood cells from abnormal red blood cells. Likewise, there was no capability disclosed for the automatic classification of the red blood cells with respect to categories of anemias.

Measurements of normal erythrocytes without differentiation of any abnormal erythrocytes by image processing has been disclosed by J. W. Green and reported in a paper entitled "Green, J. E., `Computer Methods For Erythrocytes Analysis`, Proceedings of Symposium of Feature Extraction and Selection and pattern recognition, IEE Catalog No. 70C 51C pp. 100, Argonne, Ill., 1970". A similar type of paper reporting measurements on the red cells and how to measure their features without any classification thereof was disclosed in a paper entitled "Eden, N., `Image Processing Techniques in Relation to Studies of Red Cell Shape`, edited by M. Bessis, R. Weed and Leblond, Springer-Verlag. New York, pp. 141, 1973".

In U.S. Pat. No. 3,851,156 Green provides a technique for scene segmentation of stained red and white blood cells through the use of a color algebra technique. In so doing features of perimeter, size and color are generated for red blood cells. These are measured on the total population of cells and no classification into subpopulations is performed. Further, a precise central pallor analysis is not considered, and means are not provided to acquire sub-population statistical measures, such as the bivariate dispersion of the hemoglobin and size, nor is it indicated that they are important to achieve an anemia categorization, or a profile of similarity measures to prototype anemias.

In short, none of the aforementioned systems have the ability to analyze cells by their features, particularly the inner features of cell pallor, to quickly classify the blood sample or report it as similar.

With the method and apparatus disclosed in my above-identified co-pending patent application, a quantitative analysis of abnormal subpopulations of cells has been performed on a scale heretofore not possible. Because this equipment and techniques enable analysis of cells more quickly and accurately than can the human eye or other existing equipment, it has been possible to gain a better quantitative understanding of abnormal red blood cell subpopulations for different anemias and the relationship of the abnormal red blood cell subpopulations to normal cell subpopulations. It has now been found that a red blood cell sample of blood contains significantly more information concerning the type of anemia present than heretofore known or thought.

Thus, with the present invention, it has been possible to quantify and identify blood from patients with anemia by variables or measures of characteristic values of subpopulations, such as size, hemoglobin content, percentages of cells, and other parameters such as measures of dispersion and skewness for certain single and combined parameters over different subpopulations as well as the population of cells as a whole. These measured properties of the subpopulation of cells and the population as a whole provide a robust description of a patient's blood sample.

Also, with the present invention, it is now possible to compare these red blood cells descriptors relative to characteristic values for the red blood cells in a standard normal blood sample and to those typically found in each of a plurality of recognized kinds of anemia. It is also possible to generate an indice of the relative closeness of the blood sample to one or more standard anemias so that the clinician is given a powerful quantitative relationship to aid in his diagnosis. Additionally, by testing the patient's blood at different times, particularly after successive treatments, one can generate a series of indices of the patient's blood relative to each type of anemia. Thus, it is possible to determine a patient's progress to see if his blood is deteriorating or is progressing towards a more normal blood. With expanded and accepted usage of the present invention, it is thought that some of the other time consuming, painful and/or expensive tests, above discussed, and now commonly used in the diagnosis of anemia, may be eliminated. Therefore, and as will be explained in greater detail hereinafter, a real time analysis of the red blood cell sample and to determine its make-up and a comparison to standard types of anemia is feasible.

To this end, the present invention uses multiple logic systems operating simultaneously and in a controlled relationship one with another to divide and perform tasks therebetween. Thus, the ability to analyze hundreds of red blood cells and to extract their various features and then to define the parameters for the subpopulations for comparison with anemia standards can be accomplished by the division of the functions and tasks between these simultaneously cooperating logic systems, all as will be explained in greater detail hereinafter.

A general object of the invention is to provide a new and improved, as contrasted to the prior art, system for the automatic analysis and classification of cells.

An object of the invention is to provide a method and apparatus for automatically classifying the blood and its relationship to recognized categories of anemias.

Another object of the invention is to provide an improved method and apparatus for automatically testing blood for abnormalities.

These and other objects of the invention will become apparent from the following detailed description and the accompanying drawings in which:

FIG. 1 is a perspective view of an apparatus for practicing the method of blood analysis and embodying normal features of the invention;

FIG. 2 is a block diagram showing the operation of the apparatus illustrated in FIG. 1;

FIG. 3 is a block diagram of the preferred process for analyzing and classifying blood cells;

FIG. 4 illustrates a scanning technique for locating cells and determining the boundary point of cells in an image;

FIGS. 5a, 5b, and 5c are flow charts of the preferred classification technique for classifying the blood cells into mutually exclusive subpopulations.

FIG. 6 is a diagrammatic view of a model for red blood cell central pallor measurement.

FIG. 7 illustrates a chain code description and analysis method for three diagramatic red blood cell outlines;

FIG. 8 is a block diagram of the preferred process for determining whether a cell is round;

FIG. 9a, 9b and 9c are graphs illustrating thickness/density profile measurements for three different, typically appearing cell types, measured in two orthogonal directions. These profiles are used to measure the cell central pallor features and target cell features. FIG. 9a illustrating a "flat" cell having little or no central pallor development;

FIGS. 10a, 10b, and 10c are graphs illustrating the profiles of the cells of FIGS. 9a, 9b and 9c with the peaks and valleys of each profile labelled;

FIG. 11 is a schematic of the preferred process for accomulating red blood cell subpopulation parameters;

FIG. 12a, 12b, 12c, 12d and 12e are schematics illustrating the preferred process of computing the sub-population characteristics from the accumulated values from a plurality of cells;

FIGS. 13a, 13b, 13c and 13d are graphs of bivariate distributions of red blood cells subpopulations exhibiting normal and anemic characteristics;

FIG. 14 is a schematic block diagram of the preferred process for generating a similarity, or n-space distance, measure between a given sset of measurement values describing a blood and a stored set of characteristic values, prototypic of various anemic conditions, or the normal blood.

FIG. 15a is a graph of population distributions of the cell circularity shape measure, illustrating the differences in skewness of these distributions, over all subpopulations, for normal blood compared to that of sickle cell anemia.

FIG. 16 is a graph of population distributions of individual cell central pallor measurements, for the 2 subpopulating, biconcave and spherocyte, illustrating the differences in mean values and dispersion for bloods of spherocytosis, normal and iron deficiency anemia.

As shown in the drawings for purposes of illustration, the invention is embodied in a method and apparatus for automatically classifying red blood cells and for analyzing the relationship of the patient's blood sample to at least one recognized category of anemia or to a normal blood, or to a red blood cell disorder other than an anemia. More specifically, individual red blood cells are automatically examined and classified into different cell subpopulations such as, for example, a spherocytic cell subpopulation, an elongated cell subpopulation, an irregular shape cell subpopulation, a target cell subpopulation and a generally round and biconcave cell subpopulation, and then a plurality of characteristic values are generated for the patient's subpopulations and population of cells as a whole for comparison with reference characteristic values which define a recognized anemia. By way of example, the selected characteristic values, which identify a given anemia, have been found and are given hereinafter for the following anemias: iron deficiency, chronic disease, B-thalassemia, megaloblastic, hemoglobin SS, hemoglobin SC and spherocytic; and likewise, reference characteristic values defining a normal blood, i.e, substantially all normocytic cells or the like, has been developed and will be given hereinafter.

Also, as will be explained in greater detail hereinafter, a population or subpopulation dispersion measure of the red blood cells in a patient's blood relative to characteristics such as, for example, hemoglobin, mean cell size (or area) shape, and central pallor, may be reported to the clinician. For example, broadly speaking, the bivariate red cell distribution of size and hemoglobin content for each cell is generally in the form of an elliptically shaped profile, as best seen in FIGS. 13a-13d and having axii at 45.degree. and 135.degree.. The length and width as measures of the bivariate dispersion, and the location of the profile by measures of the means values, may be reported to provide the clinician with an impression of the patient's total cell make up. Similarly, measures of central tendency dispersion and skewness of pallor and shape are provided to further quantify the total cell make up on the total cell population, or on subpopulations, as illustrated in FIGS. 15 and 16.

The patient, who has anemia generally is experiencing difficulty in either manufacturing new normocytic red blood cells or his existing red blood cells are being destroyed at an abnormal rate or by an abnormal process. Red blood cells typically have a life of about 120 days and their generation, growth and death is a continuous process. An anemic disorder generally manifests itself in blood cells having unusual sizes or shapes relative to a normal red blood sample, which predominantly contains round normocytic red blood cells, or in blood cells having hemoglobin characteristics differing from the hemoglobin characteristics for normal blood cells. Thus, since the currently existing cells, for example in a normal patient who has just developed an underlying disease process leading to anemia, have a life span of 120 days, new cells with differing characteristics will tend to produce a wider dispersion of population measurements, as in FIG. 13c compared to FIG. 13a, when sampled and examined by the precise measurement techniques described herein. Such information provides the clinician with a knowledge of the presence of any previously visually estimated anisocytosis, i.e. a large measure of cell variations in size, and also as well the variations in hemoglobin content. The mean cell hemoglobin and mean cell size information locate the central tendency of the cell distributions in FIGS. 13a-13d.

In the embodiment of the invention described herein, each of the specified anemias is identified by 16 stored parameters or properties. The patient's body is analyzed on an individual cell basis with each cell being classified into a subpopulation and then parameters such as mean cell size, mean cell hemoglobin and the percentage of cells in the subpopulation of the total cell population are generated to give subpopulation results. Also, a plurality of other measured properties or parameters of the patient's blood are generated from the subpopulation parameters, to total sixteen parameters to define a set of reference characteristic values, i.e. an n-space location, and a calculation is made of the closeness of the patient's blood location relative to the eight reference character values or n-space locations for the seven anemias and the normal blood. A report of the closeness of the patient's blood sample relative to these standard anemias provides the clinician with a statement as to what type of anemia, if any, the patient has, or how similar it is to a known type. Then, after the patient's treatment, the clinician is able to make later analyses and achieve new quantified results showing whether the patient is progressing towards a more normal blood or is deteriorating.

To achieve the analysis of the individual cells and the classification of same into subpopulations and the comparison of the blood subpopulations variables to those defining a specific anemia on a real time basis, the preferred equipment employs first and second logic systems which operate simultaneously and in a controlled manner so as to proportion the work and efforts therebetween. Also, as will be explained in greater detail, the present apparatus and method include a number of powerful and novel techniques and means of and for cell classifying and analyzing which result in an efficient and less expensive method and apparatus for doing the red blood cell analysis. For instance, the present invention recognizes that a normal blood sample generally will have a very high percentage of round cells with identifiable central pallor which can be grouped into a common subpopulation called a "biconcave" cell subpopulation and that seven different anemias can be identified when using only four other subpopulations with this biconcave subpopulation. It is to be understood, however, that the present invention is not limited to any subpopulations described or defined herein, as the particular names and make up of subpopulations may be varied and still fall within the purview of the invention herein claimed.

As shown in FIGS. 1 and 2 of the drawings, for purposes of illustration, the invention is embodied in an apparatus 10 which comprises a microscopic digital image processing and pattern recognition system which analyzes a mono layer of red blood cells on a microscope slide 12 with the cells being spaced from each other to ease the automated classification thereof. Suitable high resolution microscope optics 14 form an optical image for each red blood cell on a vidicon television camera tube or other detector 16 which converts the scanned electronic charged distribution of the optical image point by point into a numerical or digitized image representing the optical transmission of the points in each image. The output of the vidicon camera is applied to digitizer electronics 20 which includes an analog to digital converter which is connected to an image processing logic 22 which controls the digitizer electronics 20 and receives and stores the digitized cell images into a memory store. The image processing logic 22 operates on the digitized cell images in a manner that will be hereinafter described which includes cell feature extraction and cell classification.

A suitable stage motor means 24 is provided and controlled by stage motor electronics 26 which are in turn controlled by a master control logic 28. The stage motor 24 is provided to shift the slide 12 in order to iteratively process different image areas of the blood specimen on the slide. To control the focus of the microscope, a focus control motor means 30 is connected to the microscope and is operated by focus motor electronics 32 which are also controlled by the master control logic 28 by means of the focus parameter electronics 34. Focus control of slides for image analysis is well known in the art, e.g. U.S. Pat. No. 3,967,110.

The apparatus 10 shown in FIG. 1 includes a housing 38 having a cover 40 enclosing the microscope optics 14 and the television vidicon 16. An upper section 42 of the housing 38 houses the control switches of the apparatus the next lower section 44 houses the master control logic 28 with the next two lower portions 46 and 47 of the housing containing the memory store for the image processing logic 22 and master control logic 20 and the motor electronics 26 and 32. A terminal 48 is connected to the master control logic 28 and has a keyboard 50 for input of identifying information about the specimen or for other instructions. A monitoring screen 52 provides a visual display of the final report, and preferably a written printout is also made by a printer means 54 to afford a permanent record. A TV monitor 55 provides desired pictorial displays. The TV camera electronics are housed in a section 49 below the monitor. The next lower section 51 houses the analog to digital converter with the first section 53 housing the image processing logic 22. The results of the red cell analysis may also be transmitted for storage in a medical computer data bank.

In accordance with the present invention, red blood cells may be examined such that normal cells are distinguished from abnormal cells and classified by the apparatus 10 into subpopulations automatically in a detailed fashion heretofore not possible by a manual/visual examination of cells. Also, each of the red blood cells being examined may be classified into mutually exclusive subpopulations and reported out so that the presence of a minor number of abnormal cells is not overlooked or forgotten and so that accurate parameters about a given subpopulation may also be provided. For the first time, the individual red blood cells may be examined individually for the hemoglobin contents. Thus, a report may be made not only of the kind of cells found in the subpopulation but also of their number of their hemoglobin characteristcs. Advantageously, the individual red blood cells may be analyzed and classified with less subjectivity into a large number of mutually exclusive subpopulation (Table I) such as biconcave (round cells with central pallor), elongated cells, targets, and irregular cells (cells not fitting into any of the above classifications).

The preferred hemoglobin characteristic gathered from the analysis of the hemoglobin contents of the individual cells within a given subpopulation and reported out is the mean cell hemoglobin (MCH) for a given subpopulation of cells, such as shown in Table I. In addition to the hemoglobin parameters, the individual cells are counted for each subpopulation to provide their respective percentages of the total population; and likewise mean cell area (MCA) for each subpopulation may also be reported as shown in Table I. It has been found to be helpful in detecting abnormalities in blood samples to determine multivariate distributions of the red blood cells in particular subpopulations of a sample with respect to a plurality of quantifiable features. In this regard a bivariate distribution is shown in FIG. 13a as a distribution of round biconcave cells, with respect to a preferred quantifiable feature cell area on one axis and the cell hemoglobin content on the other axis.

By means of a measurement and analysis procedure to be described, parameters are reported which describe this distribution with regard to its central disposition, or mean values over the plurality of variables, and its variability, spread, or dispersion. The mean cell area and mean cell hemoglobin describe the center of the distribution and are reported as shown in Table I. Two other statistical parameters EV1 and EV2 are reported in Table I and describe the variance of the dispersion of the distribution in the orthogonal directions of its major and minor elliptical spread. EV1 and EV2 stand for eigenvalue 1 and eigenvalue 2, respectively, and describe the dispersion or spread of the distribution. If the points of the distribution are through of as defining an ellipse, then EV1 and EV2 can be thought of as relating to the length and breath of the ellipse. Advantages derived from reporting parameters relating to a distribution of a particular subpopulation will be more fully described hereinafter.

Other parameters reported in Table I include the mean pallor volume (PAL) for the biconcave and spherocyte cells as we-1 as the standard deviation for the distribution of the biconcave and spherocyte cells with respect to central pallor volume. The pallor volume standard deviation (PSD) is a parameter which describes the variance of the distribution of this measure over these subpopulations of cells. Another parameter reported is the skewness (SKW) which measures the skewness of the distribution of all the cells with respect to the quantifiable feature (perimeter of the cell) 2 area of the cell.

This data has been unavailable prior to this invention from any commerical instrument, or in any other fashion, such as from special research instrumentation. The closest analogous instrument is the Coulter counter (Coulter Company, Hialeah, FL) which is unable to classify red blood cells into subpopulations and which reports the mean cell size and mean cell hemoglobin for the entire population of red blood cells. As seen in Table I, the present invention is also capable of reporting the total population, or average mean cell hemoglobin as well as the average mean cell area (which is related to the mean cell size) in addition to the other parameters suggested. In that table these are denoted in the line with AVERAGE parameters.

Thus, as indicated above, herein, the invention will be described as having the ability to classify red blood cells into the several mutually exclusive subpopulations set forth in Table I. The subpopulations listed are the preferred subpopulations for classifying blood with respect to recognized categories of anemias but there may be other subpopulations defined. The mean cell area (MCA) is reported in microns.sup.2 with the mean cell hemoglobin (MCH) reported in picograms (pg).

The several subpopulations described and their associated parameters hereinafter are:

In accordance with another aspect of the present invention, samples of blood may be analyzed and thereby classified by "similarity" or "distance" measures being reported at to compare said sample to recognized categories of anemic or normal bloods. In the preferred embodiment, 24 parameters are measured for the subpopulations of the sample of blood taken from the patient. Of these, 16 are used for the tested sample of blood define a point in this 16-space. Consequently , the typical parameter values for a particular anemia also define a point in the 16 parameter space. In accordance with the present invention, the distance, from the point representing the values for the sample blood taken from the patient, to each of the points representing the typical parameter values for each of the categories of anemia, may be determined. Thus, a physician would be able to determine which of the categories of anemia the sample of blood taken from the patient most closely resembles and could make a diagnosis from that information. Alternatively, simple decision logic could point out the most probable diagnosis. The normalized distance of the parameter values for a sample of blood is shown for a normal category of blood as well as the recognized categories of anemia in Table I. As seen in Table I, this particular sample of blood is closest to normal since 0.9 is less than any other distance reported and therefore the blood most closely resembles normal blood.

With reference now to another aspect of the present invention, a multiple parallel logic architecture has been found to provide the rapid processing necessary for efficient analyzing of cells on a slide. Thus, in the preferred embodiment, there is provided a first processing means, the master control logic 28, and a second processing means, the image processing logic 22 as shown in FIG. 3. The analysis of the cells on a slide requires a sequence of operations to be performed, and since one operation often requires the results of a previous operation, there are provided synchronizing means for synchronizing the processors so that the results necessary to perform a particular operation are available when that operation is begun.

FIG. 3 illustrates the specific interrelationships between the master control logic 28 and the image processing logic 22. Because of this multiple parallel logic or architecture, the master control logic may proceed with one task or operation while the image processing logic is proceeding with another operation.

As seen in FIG. 3, the operations carried out by the master control logic 28 are listed in the lefthand column with the operations of the image processing logic 22 in the righthand column. The master control logic, after clearing its associated accumulators, proceeds to operation 56 in which a start signal is sent to the image processing logic and thereafter continues to operation 58. The image processing logic meanwhile is waiting for the start signal (operation 60) from the master control logic. Upon receipt of the start signal, the image processing logic 22 proceeds to operation 62 which includes digitizing the image produced by the vidicon camera 16 (FIG. 2). Upon completion of the digitizing, the image processing logic sends a "digitizing done" signal (operation 64) to the master control logic indicating the completion of the digitizing process and proceeds to operation 66. The master control logic operation 58 is currently waiting for the "digitizing done" signal and upon its receipt, proceeds to move the stage (operation 60) on which the slide rests so that a new field of cells may be imaged since the previous field has already been digitized by the image processing logic 22. The optics 14, FIG. 2, are providing an imaging means of the cells on the slide. The stage motor drive 24, and the focus motor drive 30, and their associated electronics, are controlled by the master control logic 28. After moving the stage so that a new field may be imaged, the master control logic proceeds to operation 70 wherein the field is focused and then proceeds to operation 72.

After transmitting the "digitizing done" signal, the image processing logic scans the digitized image for a cell boundary point (operation 66). If a cell boundary point is found (operation 74), the image processing logic extracts the cells boundary and features (operation 76) and classifies the cell as to its proper subpopulation (operation 78).

The image processing logic then returns to operation 66 and continues scanning the image for another cell boundary point. The scanning, feature extraction, and cell classification operations will be describe in more detail below. If the logic section 74 determines that a new boundary point has not been located, then the image processing logic proceeds to operation 80 wherein the features of each cell located as well as each cell's subpopulation classification is transmitted to the master control logic which will be in the process of executing operations 68, 70, or 72. The transmittal of the information is on an interrupt basis, i.e., should the master control logic be in the process of controlling the imaging means (operations 68 or 70), the master control logic will interrupt these operations and store the information received from the image processing logic before proceeding with moving the stage and focusing the microscope. However, if these operations have already been completed then the master control logic proceeds to operation 72 wherein the master control logic waits for the data to be transmitted from the image processing logic. In response to the receipt of the data, the master control logic will transmit an acknowledge signal (operation 82) to the image processing logic and then proceeds to operation 84 wherein the subpopulation data for each subpopulation is updated, as will be more fully explained below.

Upon receipt of the acknowledge signal, the image processing logic proceeds to digitize the image of the new field that has been moved into view by the master control logic. The master control logic, upon completing the update of the subpopulation data, determines at logic section 88 whether N, the total number of cells processed, is equal to one thousand. If 1000 cells have not been processed, the master control logic returns to operation 58 and waits for the "digitizing done" signal from the image processing logic, otherwise the master control logic calculates the subpopulation parameters (operation 90), proceeds with an anemia classification (operation 100), and prints the results (operation 102), as will also be more fully explained below.

Thus, because of the dual processor architecture, the master control logic is free to control the imaging means wherein a new field is brought into view to be imaged while the image processing logic is proceeding with the digitizing and analyzing of the image from the previous field. Similarly, while the master control logic is accumulating the data extracted from the image by the image processing logic, the image processing logic may simultaneously digitize and analyze a new image provided by the new field which had been brought into view by the master control logic. It should be noted that although for purposes of illustration, only one image processing logic is described as associated with the master control logic, it is capable of utilizing information from several image processing logics operating in parallel and independently on different images.

The present invention is directed to the optimization of the time of analysis as well as the number of features used in the classification logic so that the amount of storage and classifying techniques may be reduced substantially along with equipment requirements therefor. With an optimization of analysis time for classification, there is a danger that the reliability and accuracy of the classification are compromised. Despite this, a relatively foolproof feature set and classification logic has been invented for a large number of subpopulations such as those shown in Table 1. The preferred classification features are size, hemoglobin content, spicularity, roundness, elongation, central peak height (if present) from cross-sectional cell scans, and central pallor. By suitable combinations and analyses of such features, it is possible to differentiate from normal blood and to identify biconcave round cells, spherocytes, target cells, irregular shaped cells, and elongated cells.

In the preferred method and apparatus, the cell classifications are achieved by an image processing and pattern recognation with great accuracy and reliability by rendering white blood cells and other artifacts substantially invisible to the optics 14 by using a light having an optical wavelength of about 415 Nanometers. At this optical wavelength, the red blood cells are relatively contrast enhanced to the ultraviolet sensitive Vidicon camera without staining, while the white blood cells and other formed elements are substantially invisible. The staining of the red blood cells prior to being analyzed by a microscopic image processing technique has been found to be a time-consuming process, as well as undesirable in that the staining may introduce a number of stained artifacts which detract from the accuracy of the analysis. Furthermore, many of the stains are not stoichiometric in the representation of hemoglobin concentration according to density, thus distorting the quantization of the hemoglobin content of the cell on a per-cell basis. A particular manner of vapor fixing of cells before they dry without staining thereof to prevent the formation of artifacts by distortion of the central pallor is disclosed in my co-pending application entitled "Method and Apparatus for the Preparation of Blood Samples for Automated Analysis", filed Dec. 21, 1977, and hereby incorporated by reference as fully reproduced herein. Thus, by rapidly preparing the specimens in a monolayer and fixing with a formaldehyde vapor prior to the drying of red blood cells, as disclosed in the afore-mentioned co-pending patent application, and by not employing a time consuming staining to contrast enhance the cells, as in white blood cell analysis, these specimens may be quickly prepared and analyzed accurately.

The location of the cell image and the identification and feature extraction has been greatly simplified as described below to locate and define the cells by a boundary procedure which defines the cell in the form of an octal chain code. The use of octal chain codes as an image processing technique is described in a paper by Freeman H., Computer Processing of line-drawing images, ACM Computing Surveys 6:57, 1974. As will be explained in greater detail, the octal chain code allows feature extraction as to: (1) cell size, (2) perimeter length and roundness shape measure, (3) irregular shape measure and (4) elongation shape measure. This is followed by extracting the summed density or hemoglobin feature, and then by extracting cross-sectional scans (thickness/density profiles) for central pallor measurement and target cell measurement. Finally, inner central pallor boundaries are determined and features analyzed for more precise target cell identification.

After having extracted these identifying features, the cells are then categorized by a classification means. The preferred classification means (FIGS. 5a, 5b and 5c) comprise either a digital logic system of electrical devices or a programmed microprocessor which uses Boolean logic to classify the red blood cells.

Referring now in greater detail to the specific features of the illustrated embodiment of the invention, the images of the cells are digitized (operation 62 of FIG. 3) in a manner known to the art, e.g., U.S. Pat. No. 3,883,852 as a television digitizing system. Magnified blood cell images are obtained by using microscope optics with ultraviolet illumination, arranged to provide a 0.23 micron pixel resolution in the image plane. A pixel is a picture element having a specific location in the digitized image stored in the memory analyzer.

Referring now to FIG. 4 which illustrates in greater detail the operation 66 (FIG. 3) by the image processing logic. An original microscopic image which had been digitized is stored as represented by the image 108 for the purpose of further analysis. This analysis is carried out by the image processing logic and is represented by the blocks indicated at 115 which comprise the operations 76 and 78 (FIG. 3). In this preferred embodiment of the invention, individual cells 110 and 112 in a digitized image 108 are located by a technique in which a raster scan is made of the digitized image to locate objects above a critical threshold, such as illustrated for cell 110 in block 113. The boundary of the cell is traced by examining the neighboring pixel elements by a counterclockwise search, by techniques which are well known in the art, one such technique is disclosed in U.S. Pat. No. 3,315,229. During this counter-clockwise boundary tracing operation herein, the picture element at the "top" of the cell, pixel 114a, which is usually the pixel located first, and the one at the "bottom" of the cell, here pixel 114f, are stored for reference in the later analysis. The analysis process then proceeds to extract features and to classify the located cell into one of a plurality of subpopulations, as in block 115, and as described in detail later.

The raster scan of the digitized image is then continued from the bottom pixel 114f to hit the next digitized cell 112 by impacting a pixel 112a which is above the threshold as seen in block 116. After the boundary is traced and the features for this cell are extracted and the cell is classified, the raster scan continues from the bottom pixel 112b, and, as seen in block 118, no more cells are located in the image field. At this time, the image processing logic transmits the cell features and subpopulation classifications to the master control logic (operation 80) as shown in FIG. 4.

The initial image processing done by the image processing logic outlined in FIG. 3 is shown in greater detail in FIG. 5a. After the image has been digitized (operation 62), the image is scanned to locate a cell (operation 66) and the boundary is traced as explained above.

During this boundary tracing operation octal chain codes are formed in an operation 119. The outer boundaries, defining a cell, is processed in the following manner. Each pixel element defining the boundary is stored in a list as a series of numbers indicating a line description of the cell. For instance, referring to FIG. 7, a digital image of cells as defined by their boundary pixels 120 are illustrated.

As is well known in the art, e.g., as described in "Bacus, J. W. and J. H. Weens, `An Automated Method of Differential Red Blood Cell Classification with Application to the Diagnosis of Anemia,` Journal of Histochemistry and Cytochemistry, 25:7, 1977," a plurality of features F1-F4 can be computed from this chain code. The details of this computation are fully described in the aforementioned publication, which is hereby incorporated by reference as if fully reproduced herein.

The above features are combined with other features for use in the classification of the cells. In this regard, the following features are used herein:

As indicated above, features F1-F4 are calculated in an operation 124 by the image processing logic as shown in FIG. 5a. Feature F1 relates to the area or size of the cell as determined by the number of picture elements or pixels that are enclosed by the cell boundary. Feature F2 is the (boundary perimeter).sup.2 /area and is of assistance in classifying round and non-round objects. A round object would have a theoretical value of 4.pi. and non-round objects have greater values.

In actual practice the value of the perimeter squared divided by the area for round digitized objects varies as a function of the number of pixels, and in addition always involves quantization error, such that in practice for quantized circles the value approximates 14.0, and is a better approximation to this reference number as the number of pixels, or size, of an object increases. For total cell areas above 500 pixels the quantization error is within .+-.0.2 units.

Features F3 and F4 relate to the spicularity and elongation shapes, respectively, F3 being a count of the number of spicules in a chain code boundary, and F4 measuring the non-roundness due to elongation of the boundary, as shown in FIG. 7. Feature F5 is the integrated optical density of the cell (operation 136). It is the sum of the grey levels within the enclosed boundaries of the cell. Feature F6, which is a measure of the pallor volume, assists in distinguishing cells with large pallors, such as hypochromic cells from normocytes. Feature F7 is equal to the larger of the two central peaks of two cross sectional orthogonal 3 peaked thickness/density profiles, either having a central peak, and is used to detect target cells. Feature F8, is a measure of the depth of the central pallor, as determined from two cross sectional, orthogonal, 2 peaked, thickness/density profiles. Feature F9 is a measure of the degree of roundness of the pallor itself, and is also used in distinguishing target cells.

The logic decisions for determining the various features that have been briefly described are carried out by the image processing logic using the logic flow chart shown in FIGS. 5a, 5b, and 5c. The logic decisions are made using the various features together with threshold values that are identified as T1 through T11. The thresholds T1-T11 are described in Table 5 and specific values are also provided. As shown therein, the thresholds are used by the logic with the various features in making logic decisions leading to the classification of the cell of interest in accordance with the flow chart shown in FIGS. 5a, 5b, and 5c. In this regard, FIGS. 5a, 5b, and 5c illustrate various decisions that are made on the basis of various features either exceeding or being less than certain threshold values as will be specifically described.

Referring to FIG. 5a, an object that is located is examined by logic section 138 to determine if it is sufficiently large to be a cell, rather than a noise or dirt artifact, and thus is to be further analyzed. If feature F1, which is the size or area of the object under consideration, is less than the threshold value T1 which may be a value of about 6 microns.sup.2, then the object is not considered by the decision logic and another object will be located for analysis and classification. However, if the area of the cell is greater than the threshold value T1, feature F5 is computed in operation 136 wherein the hemoglobin content of the cell is determined. This is simply a summing of the grey levels inside the boundary of the chain coded cell and then dividing by a conversion factor 1290 or thereabout to convert the grey level measurements to picograms of hemoglobin per cell.

For this purpose the electronics generating the television signal and digitizing said signal should be adjusted to produce grey levels corresponding to the following optical density at 418 nanometers:

Also, for calculation of hemoglobin and the area the optics and television electronics should be adjusted such that round objects of the following dimensions produce the given number of pixels.

The decision logic then operates to determine whether the cell is round or non-round. This is performed by a logic section indicated generally at 140. The logic section 140 is shown in FIG. 8 to include logic subsections 142, 144, and 146. The subsections 142, 144 and 146 are operable to jointly make the roundness determination with the features F2, F3 and F4 being examined with respect to thresholds T4, T5 and T6. If the cell has a small roundness value, a small spiculated value and a small elongated value, then it is considered to be round and is passed on to the next operation 148 (FIG. 5a) which is the first step in the target cell analysis and central pallor analysis. Similarly, if it is determined that the cell is not round, then logic subsection 150 (FIG. 5a) operates to determine if the size of the cell exceeds an upper boundary threshold T2, and if it does, the cell is not further analyzed and a new cell will be considered. The effect of the subsection 150 is to eliminate double cells such as that shown in the pictorial representation 152. It should be appreciated from the pictorial representation that such a double cell would not pass the roundness test, but it is also not a non-round cell of the type for cells of classes 3 and 4. Thus, it cannot be accurately classified and it is for this reason that the subsection 150 eliminates such cells from further consideration.

As previously mentioned, the roundness of the cell is determined by feature F2 which will have a value of 14.0 for a perfect circle and will increase as the shape of the cell departs from circular. Thus, the threshold value T4 is chosen to reflect reasonably good circularity and if the feature F2 exceeds the threshold T4, that is an indication that the shape is not circular, hence the logical flow to subsection 150 indicating that the object is not round. If feature F2 is not greater than threshold T2, it is one indication that the cell is round and if the decision from the subsections 144 and 146 also indicate adequate roundness, the logic flow then proceeds to logic subsection 148 (FIG. 5a).

In operation 148 thickness/density profiles are extracted from the cell image. These profiles are illustrated in FIGS. 9a-9c and 10a-10c. A thickness density profile is determined by the grey levels of the pixels along a particular direction across the cell image. As noted earlier, the grey level of a pixel is determined by the hemoglobin density at that point. It has been found that the grey level of the cell at a particular point is related to the hemoglobin density and the cell thickness at that point. Two such thickness/density profiles, profile a and profile b are shown in FIG. 9a for a biconcave cell determined in two orthoginal or transverse directions, a and b. Two profiles each are also illustrated in FIGS. 9b and 9c for a target cell and a spherocyte cell. As seen in FIG. 9b, one direction (direction a) practically missed the center area. Since these profiles are used to distinguish target cells (Features F7), two transverse directions are preferably analyzed. Thus for each cell, two cross-sectional profiles are determined wherein the profile relates to the thickness of the cell along the points of the cross-sections.

A profile for each cell of FIG. 9 is discussed more fully in connection with FIGS. 10a-10c. As seen in FIG. 10a, the profile has two "peaks," P1 and P2, and one "valley,"V1. P1 and P2 are relative maxima of the profile of the cell with respect to the cell thickness and thus determine the two relative maximum thickness density points along the profile. V1 determines the relative minimum point of thickness density. Similarly, the target cells have three relative maximums P1 P2 and P3 with two relative minima, V1 and V2 as shown in FIG. 10b. The spherocyte has one peak P1 and no valleys (FIG. 10c). These profiles are utilized in a target cell analysis and a central pallor analysis as will be more fully explained hereinafter.

After the image processing logic extracts the thickness/density profiles for the cell, it proceeds to the target cell analysis performed by the logic section, referred to generally at 156 of FIG. 5b. The first step of the target cell analysis is to smooth the two profiles, profile a and profile b, as shown in operations 156 and 158, which is performed by the image processing logic before proceeding to a logic subsection 160. The logic subsection 160 determines whether a profile has three peaks and if so forwards it to an operation 162 which determines half the average of the two non-center peaks, P1 and P3, or "LEV1." A logic subsection 164 determines whether the two valleys, V1a and V2a are less than LEV1 and if so then the cell located might be a target cell and the image processing logic proceeds to examine profile b. If not, then the valleys are not deep enough in profile a to be a target cell so the center peak, P2a, is set to zero in an operation 166 and profile a is smoothed to two peaks or less in an operation 168.

After profile a is examined, profile b is examined for three peaks in a logic subsection 170. If the logic subsection determines that profile b has three peaks, it is forwarded to an operation 172 and logic subsection 174 wherein the two valleys, V2a and V2b are compared to LEV2 which is half the average of the two non-center peaks P1b and P3b as for profile a. If the two valleys are less than LEV2, then it is forwarded to operation 176 wherein the feature, F7, is determined as to which is the larger of the two center peaks, P2a and P2b, of the profiles a and b. Features F7 is compared to a threshold T7 in a logic subsection 178, and if larger, the cell is classified as a target cell (C5). In other words, if the larger of the two center peaks is larger than a certain threshold, then the cell is determined to be a target cell. If not, then the center peaks of the profiles are probably due to "noise" in the image video and digitizing and not due to a center area of a target cell. In that case, both profiles are smoothed to two peaks or less in operations 180 and 182. However, if the logic subsection 174 determined that the valleys of profile b were not less than LEV2, then the profile b is forward to a logic subsection 184 which checks whether the center peak of profile a had been set to zero. If not, then profile a may have detected a target cell and thus P2b is set to zero and subsection 176 determines the maximum value for F7 as described.

If the center peak, P2a had been set to zero, then neither profile has passed the tests at logic subsection 164 and 174 respectively. Thus the cell is probably not a target cell and profile b is also smoothed to two peaks or less at operation 182. However, some target cells might not be detected in this analysis, therefore, other tests are performed on the cell as will be explained later.

After the center peaks of profiles a and b have been examined as explained above, a logic subsection 186 determines whether profile a has only one peak. If so, the variables P1a, P2a and V1a are set equal to each other in an operation 188. In either case, the image processing logic then examines profile b to determine whether it has only one peak, at the logic subsection 190. If profile b has only one peak then the variables P1b, P2b, and V1b are set equal to each other in an operation 192.

Continuing with FIG. 5c therein, a feature F8, which is the average value of the 2 valleys subtracted from the average value of the four peaks of the two profiles of the cell is determined by subsection 194. Then the cell feature F1 is examined to determine whether the size of the cell is larger than a threshold T8 at a logic subsection 196.

If the cell is large, i.e., F1 is greater than T8, it is possible that the cell is a target cell despite the previous target cell analysis and therefore another target cell analysis will be performed beginning in operation 198. Therein, a variable LEV3, is set equal to one-half the value of feature F8 (operation 198).

Next, a search for the central pallor of the cell is initiated by searching a direction along the line from the top pixel of the cell through the center of the cell looking for a threshold condition, i.e., hitting a pixel which is below the threshold LEV3, before the center is reached. The chain code is then formed for the central pallor boundary (operation 202). The pallor circularity feature F9 is then computed in an operation 204. F9 is calculated as the number of pallor boundary pixels squared divided by the area of the central pallor. F9 is then compared to a threshold value T9 at a logic subsection 206 to determine the circularity of the central pallor. This operation is necessary since the two profiles from the previous target cell analysis may have missed the central area as shown for the cell 208. Thus, if circularity feature F9 is greater than the threshold T9, then the cell is a target cell otherwise, the cell is forwarded to the operation 209 wherein a feature relating to the size of the central pallor of the cell is computed.

The central pallor feature is defined as the percentage volume of a cylinder, with the height and area of the cell under consideration, not occupied by hemoglobin. This is illustrated in FIG. 6, where T represents the cell height or thickness, and 132 indicates the indented central pallor region. The cell area is known from previous analysis on that cell, i.e., F1. Also, feature F5 is the sum of the grey levels for pixels enclosed by the chain code defining the boundary of the cell. As noted above, the hemoglobin density is related to the thickness of the cell and in this manner, the hemoglobin feature F5 defines a volume which is related to the thickness or volume of the cell. The cylinder height, or thickness (T), is derived by using the average value of the peaks of the two thickness/density profiles of the cell, as: ##EQU1##

Thus, the volume of the central pallor may be calculated as: T times the area of the cell (F1) minus the hemoglobin content. Finally, the percentage pallor volume F6 is: ##EQU2##

After this feature has been computed, the image processing logic proceeds to a logic subsection 210 wherein the cell is distinguished between biconcave cells (C1) and spherocyte cells (C2) as it has already been determined that the cell is not an elongated cell (C3), an irregular cell (C4) or a target cell (C5). The logic subsection 210 compares the percentage pallor volume feature F6 to a threshold value T10 and the pallor depth feature F8 to a threshold T11 and if either feature is less than its associated threshold then the cell is deemed a spherocyte cell (C2), otherwise it is a biconcave cell (C1).

Referring back to FIG. 3, the feature extraction operation 76 and the cell subpopulation classification operation 78 have been completed for the cell that had been located in the image scan. The image processing logic will then continue scanning the image for another cell (operation 66) and if no other cells are found then the features for those cells located as well as the cells' subpopulation classifications will be sent to the master control logic in the operation 80.

While the determination of the various features and decisions contained in the logic diagram of FIGS. 5a, 5b and 5c is carried out utilizing the threshold values contained in Table 6, it should be understood that the threshold values are based upon empirical and statistical analysis and can be varied somewhat without appreciably effecting the eventual classification of the cells. It should also be appreciated that the threshold values are believed to be optimum values which have been fixed to maximize the accuracy of the classification.

Upon completion of the feature extraction and cell classification analyses for the cells located in the image, these features are transmitted to the master control logic as illustrated in FIG. 3. After acknowledging the receipt of the data (operation 82), the master control logic proceeds to update subpopulation measurements for each cell class located in the image just analyzed (operation 84). A diagram illustrating the updating operation in greater detail is shown in FIG. 11. A plurality of accumulators are provided to produce a running total of a plurality of measurements for the cell subpopulations or classes. Each accumulation is a function of one or more cell features, such as the cell feature value itself or the value squared, for example. The cell feature values F1, F2, F4, F5 and F6 for a particular cell are provided as inputs to the accumulaters together with the cell classification C.sub.i to which the cell features pertain. After the measurements for the cell have been accumulated, then the other cells in the image are similarly processed to further accumulate the measurements based on all of the cell's features.

Thus, the feature F2 (cell circularity feature) is provided at a line 212 to an accumulator 214. The accumulator 214 produces a running total S1, i.e., accumulates the measurement (F2-14.1).sup.3 for all the cells located by the image processing in logic wherein F2 is the cell circularity feature (Table 4), This measurement is used in a later calculation which provides a parameter describing the skewness of the distribution of all the red blood cells located, with respect to the circularity feature of the cells.

Also accumulated is the elongation feature F4 which is provided at a line 216 to accumulators 218 and 220. The accumulator 218 sums the total (S2) of the feature F4 for all the cells which is used to calculate the average elongation for the cells. The accumulator 220 provides a sum or running total (S3) of the elongation feature F4 squared, i.e., (F4).sup.2, which is used to calculate a parameter describing dispersion, or variation of the distribution of the red blood cells with respect to the mean of the elongation feature F4.

Not all feature measurements are accumulated for each subpopulation. For example, the feature F6 (pallor volume) is only accumulated for the biconcave cells (subpopulation C1) and the spherocyte cells (subpopulation C2). Therefore, in addition to the features for a particular cell, the subpopulation classification for the particular cell to which the features pertain is provided which is shown as C.sub.i at line 222. A plurality of logic utilize the input C.sub.i to discriminate among the cell subpopulations. Thus, the cell classification C.sub.i is provided to the inputs of a logic AND gate 224 and an AND gate 226 with subpopulation C1 constant (i.e. a 1) provided to the other input of the AND gate 224 and subpopulation C2 constant (1.e., a 2) provided to the other input of AND gate 226. The output of these AND gates are provided to an OR gate 228 which may enable the accumulators 230 and 232. The accumulator 230 provides a summation of the feature F6 (central pallor volume) as indicated by input lines 242, but only when enabled by the logic OR gate 228. Similarly, the accumulator 232 accumulates the sum of the feature (F6).sup.2 but only when enabled. Thus, the gates 224, 226, and 228 permit the accumulators 230 and 232 to accumulate the measurements derived from the feature F6 only when the feature had been extracted from a C1 or C2 biconcave or spherocyte class cell. The output of the accumulator 232 is provided at S5 which is used to compute the dispersion parameter of the distribution of spherocyte and biconcave cells with respect to the mean volume of the central pallor of the cells. The output of the accumulator 230 is provided at S4 which is also used to calculate the dispersion parameter and also to calculate the mean or average central pallor volume for the spherocyte and biconcave cells.

Similarly, a logic AND gate 234 enables accumulators 236, 238 and 240 when C.sub.i at line 222 is equal to a 2, i.e., the cell features appearing on the feature lines 244, and 246 were extracted from a class C2 (spherocyte) cell. The accumulator 236 accumulates the feature F1 (cell area) which is provided at S11, which will be used to calculate the mean cell area parameter for the cells in the C2 classification. The accumulator 238 provides at S12 the accumulated total of feature F5 (cell hemoglobin content) which is used to calculate the mean cell hemoglobin content for the class C2. The accumulator 240 provides a total of the number of cells in the C2 class, i.e., N2 equals the number of spherocyte cells located by the image processing logic.

In a similar manner the total cell area for the elongated (C3), the irregular (C4) and target (C5) cells are provided at S13, S15 and S17, respectively. The total of all cells hemoglobin content for the elongated, irregular, and target cells are provided at S14, S16 and S18, respectively. The total number of cells in each of the above subpopulations is provided at N3, N4 and N5.

Likewise, the total of all of the cells areas for the biconcave subpopulation is provided at S6, the total of all the cells hemoglobin contents is provided at S7, and the total number of biconcave cells is provided at N1. For additional accumulated measurements on the biconcave subpopulation, additional logic gates permit accumulators to discriminate among the class cells. Thus, an AND gate 248 enables accumulators 250, 252, and 254 when the features appearing at the lines 244 and 246 have been extracted from a C1, i.e., a biconcave cell. The accumulator 250 provides the accumulated sum of the measurement (F1).sup.2 at S8. The accumulator 252 similarly provides the accumulated total of the measurement (F5).sup.2 at S9. Finally, the accumulator 254 provides the accumulated sum of the product of the feature F1 times the feature F5 (F1 .times. F5). The accumulated S9 and S10 are used to calculated parameters descriptive of the dispersion, or variation of the bivariate distribution which will be further explained hereinafter.

Thus the features for each cell examined by the image processing logic provide the inputs to the logic described in FIG. 11 for up-dating or accumulating measurements based upon the cell features with the particular measurements up-dated for each cell depending upon the subpopulation classification to which that particular cell belongs. The measurements up-dated by the logic of FIG. 11 provide an intermediate step for the calculation of parameters which are descriptive of each subpopulation classification as well as parameters which are descriptive of multivariate distributions of cell subpopulations with respect to different cell features.

Referring back to FIG. 3, it is seen that at logic subsection 88 the determination is made whether a preset total of N cells have been processed. If not, the master control logic returns to operation 58 wherein it waits for the "digitizing done" signal indicating that the image processing logic has completed digitizing the next field. If N cells have been processed, e.g., N=one thousand, then the accumulated measurements which had been updated as illustrated in FIG. 11 for those N cells are used to calculate the parameters descriptive of the subpopulations (operation 90) which is illustrated in greater detail in FIGS. 12a through 12e.

The output S1 of the accumulator 214 (FIG. 11) is used in the calculation of a dispersion parameter which describes the skewness of a distribution. Herein, a distribution of all the cells with respect to the elongation feature (F4). Skewness is calculated as: ##EQU3##

Thus a logic subsection 256 having inputs S2 and N produces the skewness parameter: ##EQU4##

The calculation of the skewness parameter is quite helpful in describing a population of cells. For example, a distribution of normal cells is shown in FIG. 15, generally at 255. The distribution is with respect to Feature F2 (circularity). Also shown is a distribution of sickle cell anemia cells, generally at 257. As can be seen there, the distribution of sickle cells is greatly skewed toward the right, indicating a great number of elongated cells. Note, however, that the mode of both distributions is identical. Thus, the skewness parameter is a valuable comparison tool for indicating anemias.

A logic subsection 258 having inputs S2 (the sum of the elongation measurements for the cells) and N (the total number of cells) produces the mean cell elongation parameter (ELN).

The general formula for the dispersion in the form of the standard deviation of a distribution with respect to a variable X is given by: ##EQU5##

A logic subsection 260 produces the standard deviation of the elongation distribution of cells with respect to the elongation features. The logic subsection 260 has an input S2 equal to ##EQU6## (FIG. 11) and an input S3 equal to ##EQU7## and produces the elongation standard deviation (ESD) after the square root of the output has been taken by a logic subsection 262.

A parameter for the mean central pallor volume (PAL) of the bioconcave and spherocyte cells is provided by a logic subsection 264 having inputs N1 (the number of biconcave cells), N2 (the number of spherocyte cells) and S4 (the accumulated sum of the volumes of the central pallors of those subclassifications). A parameter of the distribution of the biconcave and spherocyte cells with respect to the central pallor volume, herein, the central pallor volume standard deviation (PSD) is provided by a logic subsection 266 having inputs S4 and S5 and a logic subsection 266 to finally produce the parameter PSD in a manner similar to that of the parameter ESD.

A distribution of three different populations of cells, normal, spherocytic, and iron deficient, with respect to the features F6, the percentage volume of central pallor is shown in FIG. 16. It is important to note that the distribution of normal cells at 267 has the same mean value (PAL) as the distribution of iron deficient cells at 269, yet they have a different variation or standard deviation (PSD) in central pallor volume. On the other hand the distribution of normal cells has the same standard deviation as the distribution of spherocyte cells at 271 but a different mean value. Thus both parameters have been found advantageous in the classification of blood with respect to anemias.

Two other parameters, EV1 and EV2, are computed utilizing the accumulated sums S6-S10 and N1 and which are descriptive of the amount of dispersions of a bivariate distribution of the biconcave cells. The two variables of the bivariate distribution are the cell size and the cell hemoglobin content. Four such distributions are illustrated in FIGS. 13a-13d wherein the cell area defines the absicissa axis and the cell hemoglobin content defines the ordinate axis. Each "X" represents a biconcave cell with its location within the graph defining the cell's area and hemoglobin content. Thus, as can be seen in the four Figures, the cells are distributed mainly on a 45.degree. line passing through the origin. The mean cell area (MCA) and the mean cell hemoglobin (MCH) define the center of each distribution. The values, EV1 and EV2, define the dispersion or the amount of spread of the distribution in 2 principal independent axes. In particular, EV1 describes the amount of spread of the cluster or distribution along the direction at essentially 45.degree., or along the line of major dispersion of the ellipse with EV2 describing the dispersion in a direction which is orthogonal or transverse, that is 90.degree., relative to the dispersion of EV1.

Referring to FIG. 12a, a logic diagram is shown for the computation of the parameters EV1 and EV2. The general formula for computing the variance of a distribution with respect to a variable is similar to that given for the standard deviation. The variance of the distribution with respect to cell area is provided by a logic section 270 which has inputs N (the number of biconcave cells) S8 (the summation of (F1).sup.2 for each biconcave cell), and S6 (the summation of F1 for each biconcave cell). The variance of the distribution with respect to hemoglobin content is provided by a logic section 272 which has inputs N1, S9 (the summation of (F5).sup.2), and S7 (the summation of (F5)). A logic section 274 provides the sum, K, of the output of the logic sections 270 and 272 and a logic section 276 provides the product, A, of the output of the logic sections 270 and 272.

The co-variance of the distribution with respect to both the cell area and the cell hemoglobin content is provided by a logic section 278 having inputs N1, S7, S6 and S10 (the summation of the product F1 times F5 for each biconcave cell). A logic section 280 squares the output of the logic section 278 to produce an output, B. A logic section 282 subtracts the output A of the logic section 276 from the output B of the logic section 280 to provide an output D. K and D are coefficients of a quadratic equation wherein a logic section 282 produces the first solution, EV1, to the quadratic equation, and the logic section 284 produces the second solution, EV2, to the equation.

A logic section 286 produces the mean cell hemoglobin parameter for the biconcave cells by dividing the total hemoglobin content S7 for all the biconcave cells by the number (N1) of the biconcave cells. The mean cell area (MCA) of the biconcave cells is produced by a logic section 288 which divides the total cell area (S6) of the biconcave cells by the total number (N1) of the biconcave cells.

In a similar manner, as shown in FIG. 12b, the mean cell area and mean cell hemoglobin parameters are computed for the remaining four classes or subpopulations, i.e., the spherocytes, elongated, irregular, and target cells by eight logic sections 290-297. The number of cells in each subpopulation, N1-N5 are each transformed into a percentage of the total number of cells by five logic subsections 300-304, in FIG. 12b. For example, the percentage of biconcave cells (NC1) is provided by logic subsection 300 which divides the number of biconcave cells (N1) by a total number of cells located by the image processing means (N) and multiplies by 100.

Finally, in the preferred embodiment, two other parameters are calculated which describe the entire population of cells analyzed as illustrated in FIGS. 12d and 12e. First, a mean cell area parameter (MCA) is calculated as a weighted average by multiplying the percentage of a subpopulation (i.e. Ni-NC5 being first divided by 100) by the mean cell area for that subpopulation for each subpopulation and adding the products to produce the weighted average. For example, the percentage of biconcave cells (NC1) is multiplied by the mean cell area (MCA1) for the biconcave subpopulation by means of a logic section 306 and the percentage of the spherocyte cells (NC2) is multiplied by the mean cell area of the spherocyte cells (MCA2) by means of a logic section 308 and so on for the other subpopulations and adding these five products by means of a summation logic section 310 to produce the mean cell area (MCA) for the entire population. A weighted average of the hemoglobin content for the entire population (MCH) is produced in a similar manner by a plurality of "multiply" logic sections 312-316 and a summation logic section 318.

In the above manner, 24 parameters descriptive of the various subpopulations of red blood cells and the entire population of red cells as a whole may be calculated, 22 of which are listed in Table I. They are the percentage of the entire population for each subpopulation, the mean cell area (MCA) and the mean cell hemoglobin (MCH) for each subpopulation, the MCA and MCH for the entire population, the mean central pallor volume (PAL) of the distribution of biconcave and spherocyte cells, the standard deviation (PSD) of the central pallor volume distribution, and the skewness (SKW) of the circularity distribution of the entire population. Two parameters, the mean of the elongation distribution (ELN) and the standard deviation of the elongation distribution (ESD) are calculated but in the preferred embodiment are not reported, as in Table I. The parameters in Table I show values calculated for a sample of blood taken from a patient. Similarly, a sample of blood may be taken from another that is known to exhibit one of the known categories of anemia such as iron deficiency for example and the sixteen parameters may be calculated for the known anemic sample. Subsequently, the parameters calculated from the analysis of the sample of blood taken from the patient may be compared to the parameters of an iron deficient anemic sample to determine if the sample from the patient resembles iron deficient blood. Likewise parameters may be calculated for a plurality of known anemic samples of blood wherein the parameters of the patient's sample may be compared and in this manner the patient's blood may be classified with respect to those recognized categories of anemic blood. Referring to Table I, it is seen that the sample, from which the parameters of Table I were calculated, has been compared to eight types of blood which are normal, iron deficient, chronic disease, B-thalassemia, megaloblastic, hemoglobin SS, hemoglobin SC, and spherocytic. A specific classification technique to produce a similarity measure for the sample blood taken from the patient which is compared to recognized categories of anemia and normal blood is shown in FIG. 14. Sixteen of the 24 parameters can be thought of as defining a sixteen variable space or sixteen-space. Values for the sixteen different parameters would define a vector having sixteen components, one for each parameter. Thus, when a sample of blood taken from a patient is analyzed, the sixteen parameters calculated therefrom would define a vector Y having sixteen components (y.sub.1 . . . y.sub.i . . . y.sub.16). Similarly, analysis of samples from the eight previously mentioned types of blood, normal and anemic, would define eight vectors, W.sub.i,1 to W.sub.i,18. Each component of the vector for a category of anemic or normal blood is determined by obtaining blood with prior knowledge of the anemic condition and measuring mean parameters over a plurality of such bloods for the sixteen parameters for that particular category. The vector Y representing the parameter values calculated for the sample of blood taken from the patient may be compared to the vectors representing the mean of various categories of anemic and normal blood. The vector that the vector Y most closely resembles, i.e., is the closest to in the sixteen-space would determine the classification of the patient's blood.

The first step in the anemia classification logic of FIG. 14 is to normalize each parameter value to produce the sixteen components of the Y vector. Thus a parameter value X.sub.1 which represents, the mean cell area of biconcave cells in the patient's blood sample is normalized by a logic subsection 320 to produce the first component of the Y vector, Y.sub.1. The logic subsection 320 in normalizing the parameter value X.sub.1 subtracts the mean value a.sub.1 from X.sub.1 and divides by the standard deviation b.sub.1 of the distribution of X, with respect to the biconcave cells. The distribution for each of the sixteen parameters has been determined with the mean a.sub.i and the standard deviation b.sub.i for each of the sixteen parameters as set forth in a Table 6 below:

There are sixteen logic sections represented by a logic section 322 which normalizes the parameter x.sub.i to produce one of the sixteen components y.sub.i of the vector Y. The vector Y representing the sixteen parameter values for the sample of blood taken from the patient is compared to the eight vectors W.sub.i,1 to W.sub.i,8 representing the parameter values for each of the eight categories of blood to determine the proper classification for the patient's blood.

Accordingly there are provided eight logic sections represented by a logic section 324 having the sixteen components y.sub.1 -y.sub.16 of the vector Y as inputs. In addition, each of these logic sections has the sixteen components of a vector representing the parameter values of one of the categories of blood. For example, the first logic section 326 compares the parameter values for the patient's blood with the parameter values for normal blood. In this connection, the vector W.sub.i,1 represents the sixteen components of the vector for normal blood. The two vectors, Y and W.sub.i,1 are compared by the logic section 326 wherein the standard distance formula is used to calculate the distance between the two vectors to produce a distance D1. Referring to Table 1, it is seen that the patient's blood parameter vector has a distance of 0.9 to the normal blood parameter vector. In a like manner, the patient's parameter vector is compared to the parameter vectors for the seven categories of anemia, as seen in Table 1. The patient's blood parameter vector has a distance of 4.2 to the iron deficient category of anemia, a distance of 2.5 to the chronic disease category, and so on.

The sixteen components of the parameter vector for each of the eight categories of anemic and normal blood are set forth in a Table 7 below:

Referring back to FIG. 3, upon completion of the anemia classification (operation 10), the master control logic proceeds to print the results (operation 102) of the analysis and similarity comparison or classification. One example of a print-out by the preferred method and apparatus has already been given as Table 1. The print-out in Table 1 indicates that the sample of blood analyzed is closest to normal based on the features analyzed. Two more examples are given in Tables 8 and 9 respectively, with Table 8 indicating hemoglobin SS anemia and Table 9 indicating .beta. thalassemia.

Two examples of results of red blood cell analysis with the present invention will be listed below in Tables 8 and 9.

From the foregoing, it will be seen that the present invention allows a new and improved analysis of red blood cells heretofore not possible (even with the system disclosed in the above-identified co-pending application) and the discovery that the red blood cell population carries sufficient information to diagnose many anemias without resort to other conventional tests. That is, subtle, slight and early changes in either cell production or destruction may now be discovered, such as at the incipiency of an anemia because of the ability to measure accurately the hemoglobin content in individual blood cells, the cell shape, the cell size variations in size of central pallor, as well as the count of individual cell sizes, and an understanding of the total red cell population in blood samples that was heretofore not possible. With the continual process of red blood cell production and destruction over a 120 day life span for each cell, the high percentage of old normal red blood cells would mask the smaller number of new red cells being produced at the incipiency of a particular anemia. For instance, a chronic disease anemia such as caused by an infection, cancer or tuberculosis, may cause the new red blood cells being produced to be smaller in size with larger, central pallors than is normal and with decreased hemoglobin content. Naturally, early detection of a chronic disease kind of anemia would be most helpful in the treatment of this particular anemia.

With the present invention and for the first time, a total spectrum or galaxy of cells (i.e., the red cell population) may be automatically analyzed on an individual cell basis and over a sufficient number of cells and with sufficient accuracy to detect a dispersion of distribution indicative of a particular kind of anemia. Take, for example, a blood sample from a person suffering from an iron deficiency anemia. Typically, in such an instance, the usual values of dispersion, e.g., EV1 will be increased from a normal blood value of about 30 to values such as 45 or higher, and likewise the usual value of dispersion for EV2 will be increased from about 2 for a normal blood to about 3 or 4 or higher for an anemia such as an iron deficiency anemia. These slightly higher values of dispersion of distribution indicate that the normal cell population is changed because of the addition of these additional cells having small size, a large central pallor, and low amounts of hemoglobin. Thus, the diagnostician will see that the normal closely packed cell population has been expanded by these new cells formed after the onset of the anemia. The other values, such as location of the central tendency of the population' s dispersion of distribution, which location is defined by the MCA and MCH values, may also have shifted because of the new anemia cells present. It is to be understood that the parameters used herein, such as MCA, MCH, EV1, EV2, skewness, etc., along with the central pallor descriptors, have been experimentally found to be most powerful (at this time) in analysis of the anemias illustrated and described herein. With further investigation, it may be that other parameters and/or measures may be used to describe and define the red cell population but such changes will still be within the purview of the appended claims and this invention. By way of example, the covariance, or correlation coefficients could be used to describe the red cell population as well as other measures, which are used to define a distribution and dispersion, and which could be used in lieu of the terms used herein in detail to describe the population of red blood cells.

Further, the present invention has been described principally in connection with the variables of mean cell hemoglobin and mean cell size although other variables such as pallor size and mean cell area have been tried and could be used. The particular parameters used and the names thereof may also be changed from that described herein and still fall within the purview of the present invention and the claims hereinafter recited.

The present invention is not to be construed as being limited to the classifying of cells into mutually exclusive subpopulations prior to making analysis of the dispersion of distribution of the red blood cells. For instance, it is possible to measure the characteristics of each of the red blood cells as to size, shape, hemoglobin, central pallor, etc., and then to make a multivariant dispersion distribution analysis without having a separate analysis of the biconcave cells as described herein. The classifying of the cells into a biconcave subpopulation and into other well known subpopulations is done because it is thought to be helpful to the diagnostician. Also, the classifying and reporting of subpopulations of biconcave cells, spherocytes, elongated cells, target cells, and irregular cells may be eliminated. The latter has been included merely as an aid to the diagnostician. Moreover, the listing of a plurality of anemias could be eliminated with the analysis and report being made only for one or more specific anemias thought to be most likely for the patient, or only that the blood sample appears to be normal. On the other hand, in a screening process of large numbers of blood samples each from a different person, it may be more helpful to include other anemias in addition to the seven anemias listed herein. The invention is thought to provide a particularly powerful tool for the screening of blood samples as well as for verification of anemias where a visual examination or other tests leave the diagnostician to suspect the presence of an anemia.

Of the seven listed anemias, the hardest anemias to distinguish from each other are the iron deficiency, chronic disease and B-Thalassemia. The accuracy of the diagnosis of these three anemias from one another is thought to be about 80% accurate with the existing equipment described herein. It is thought that the other anemias listed herein can be diagnosed with almost 100% accuracy. Generally speaking, when it has been found that the closest anemia listed was one of the three anemias of iron deficiency, chronic disease, B-Thalassemia and that closest anemia was not verified that the actual verified anemia will then be the next closest one listed of these three anemias. For this reason and other reasons, it is preferred to quantify the closeness of several anemias so that, if the first anemia is not verified, then the second closer anemia can be next chosen and examined for verification.

It will be recognized by hematologists skilled in the art that the diagnosis of some anemias, such as iron deficiency or chronic disease anemia, are most difficult today with conventional equipment even with all the information of other tests available to the hematologist. The present invention should provide a very useful tool for verification of a particular anemia when the other conventional tests which have been used and need to be verified.

Although the term "anemia" has been used extensively in this description, it should be noted that the term has been used in the general sense and the present invention may be used in a detection of other red cell disorders or pathologies such as hereditary elliptocytosis, for example, or others.

Although there has been described herein the use of first and second microcomputers, it is to be understood that only one larger computer could be used or hard wired logic could be used. On the other hand, more than one additional microcomputer may be added with each simultaneously measuring characteristics of a different red blood cell and each classifying different red blood cells into subpopulations. Thus, it is considered that one or more additional microcomputer may be used than described herein to expedite the system.

The above-described description and drawings provide a clear understanding of the invention and an enabling disclosure to persons skilled in the art. A specific example of the preferred equipment practicing the invention herein described is as follows: In the preferred embodiment, the master control logic 28 and the image processing logic 22 which carry out the flow diagram of FIG. 3 comprise two microcomputers such as the Digital Equipment Corporation LSI/11 microprocessors. In this embodiment, the listings shown in Appendix 1 and Appendix 2 can be used for implementing the flow diagram of FIG. 3 of the master control logic and image processing logic, respectively, using the features described herein together with the threshold values and parameter descriptors that have been previously set forth. The listing is of the binary load module as would be loaded into the microcomputer from the RT11 operating system monitor furnished by Digital Equipment Corp., with the listings being produced by the standard DUMP program with the /W/N option. ##SPC1## ##SPC2## ##SPC3## ##SPC4## ##SPC5## ##SPC6## ##SPC7## ##SPC8## ##SPC9## ##SPC10## ##SPC11##