JPH0469775A - Automatic cell classifying device - Google Patents

Abstract

PURPOSE:To execute the discriminating operation at a high speed by connecting a cell image input processing means with a hierarchical type network arithmetic means and a real time communication means and using the hierarchical arithmetic means at the time of execution of an automatic classification. CONSTITUTION:A cell image input processor 1 is constituted of an image input device for inputting a cell image in a sample, an image storage device 12 for storing an inputted cell image, an image process 13 for extracting a feature parameter from the inputted cell image, an arithmetic controller 14 for controlling these devices, and also, storing a result of extraction by the image processor 13, and an output device 15. Also, the cell image input processor 1 is connected with a hierarchical type network arithmetic unit 3 and a real time communication equipment 2 and the hierarchical type network arithmetic unit 3 is used at the time of execution (discrimination) of an automatic classification. In such a way, the discriminating operation can be executed at a high speed without changing an internal structure of the cell image input process.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a batan classification device using a hierarchical network, and particularly to an automatic cell classification device suitable for automatically classifying cells.

[Prior Art] As a conventional device of this kind, for example,
A device disclosed in Japanese Patent No. 29872 is known. This device identifies patterns with high accuracy by having a multi-stage identification process.

In addition, Japanese Patent Application Laid-open No. 59-114667 describes a region segmentation method using color, specifically, a method for dividing a blood cell image into white blood cell nuclei,
A region segmentation method suitable for segmentation into cytoplasmic, red blood cell and background regions has been proposed.

In the double identification method or device, the identification process is multi-stage and the amount of processing is large. The results were incorporated into a device that actually performed the identification calculations, and the identification processing was executed. Regarding this, for example, there is a patent application filed by the present applicant for an "automatic blood cell sorting device" (see Japanese Patent Application No. t17.l-268916). This device uses a neural network, which is capable of learning even when the number of samples is relatively small, as the discrimination logic.

[Problem to be solved by the invention]

In the conventional automatic cell classification device described in Section 1, another high-speed computer generates (learns) the discrimination logic, the result is stored in the device, and the device's arithmetic control computer executes the discrimination. However, when a hierarchical network is used as the identification logic, the amount of calculation is large, and in order to perform identification at high speed, the automatic cell classification apparatus itself requires a device for high-speed calculation. For this reason, there is a problem in that a conventional automatic cell classification device in which identification is performed by the device's arithmetic control computer cannot be used as is as an automatic cell classification device with built-in hierarchical network identification logic. Regarding this, T
ζ will be explained in more detail.

The hierarchical neural network is capable of learning and discrimination even when the distributions between categories intersect with each other, as shown in FIG. As the hierarchical neural network, for example, a network consisting of three layers such as an input layer, an intermediate layer, and an output layer as shown in FIG. 3 is used, and as the threshold function of the unit, for example, a sigmoid function is used. The intermediate layer may be multi-layered in film form; FIG. 3 shows an example of a negative layer. Also, the relationship between input and output is z,
= r<Σw, 1.・f(Σw, 1.-x,))
・−・(] ), l However, xl Two input parameters zk Output W I +l T W +l k Nail sharpening coefficient f゛()
- It is a threshold function, and the sigmoid function is a function that can be changed in the form of f(+1)=l/(1→exp(-u)).

``In the input layer of a hierarchical neural network such as a double series, units for the number of feature parameters, units for the number of categories in the output layer, and units in the middle layer corresponding to the number of units in both layers are set.'' , when the feature parameters of cells in a certain category are given to the units in the input layer, the unit I corresponding to the above category in the output layer is given Pavi' as teaching data, and the other units in the output layer are given P'O'' In order to reduce the error between the teacher's data and the output data, the so-called
A backpropagation method is used to change the weighting coefficients between network units. Regarding the backpropagation method described in 1., for example, "Neural Network Information Processing" published by Sangyo Tosho, 1988 is a good reference.

The above process is repeated to optimize the weighting coefficients.

This process is ``learning,'' and ``identification'' is performed based on the weighting coefficients after ``learning. Specifically, m = (klmax Zk + k
"Identification" is performed by determining m, which is 1 to nk), and setting the category to which the input cell belongs to the m-th category.

The problem here is that the operations in the two processes of ``learning'' and ``identification'' are both product-sum operations and sigmoid function operations, and as the number of units increases, the operations This means that the amount will increase dramatically. Conventionally, a hierarchical neural network calculation device that executes the operations described in -1- at high speed has been used only for "learning". However, in order to perform "identification" at high speed with a cell image input processing device, it is necessary to incorporate a separate high-speed calculation device, and there is also the problem that conventional automatic cell classification devices cannot be used as is. There is.

The present invention has been made in view of the above circumstances, and its purpose is to create a hierarchical network that solves the above-mentioned problems in the conventional technology and that performs learning and identification using a hierarchical network at high speed. The arithmetic device is used in real time in conjunction with a conventional automatic cell classification device (hereinafter referred to as "cell image input processing device") that performs identification using the device's arithmetic control computer, even during identification calculations. In this way, it is an object of the present invention to provide an automatic cell classification device that can perform identification calculations at high speed by using a conventional cell image input processing device as is.

[Means for Solving the Problems] The above-mentioned object of the present invention is to provide an automatic cell classification device using a cell image input device having an image input means, an image recording means, and an image processing means, wherein the cell image input processing j11] means is connected to the hierarchical network calculation means by the real [1 layer communication means, and when performing automatic classification (identification), 1) u
4. [Operation] The automatic cell classification device according to the present invention uses a hierarchical neural network as one discrimination logic. , and its “
A hierarchical neural network calculation device is used for "learning" and "identification" in real time. As the hierarchical neural network, a network consisting of three layers, an input layer, an intermediate layer, and an output layer, as shown in FIG. 3, is used, and the sigmoid function is used as the threshold function of the unit. The intermediate layer is: - Multi-layered in membrane form, or here a layer.

In the automatic cell classification device according to the present invention, learning is performed using a hierarchical neural network computing device that performs f-operations at high speed, and the hierarchical neural network computing device is used for real-time processing or It is also used for identification. As a result, in order to perform only identification at high speed with the cell image input processing device, there is no need to incorporate a separate high-speed calculation device, and the conventional automatic cell classification device can be used as is.

[Example] Hereinafter, an example of the present invention will be described in detail based on the drawings.

FIG. 1 is a diagram showing the configuration of the apparatus. In the figure, 1 is a cell image input processing device, 2 is a real-time communication device, and 3 is a hierarchical neural network high-speed calculation device. The cell image input processing device 1 includes an image input device WI 1 for inputting a cell image in a specimen, an image storage device 12 for storing the input cell image, and extracts feature parameters from the input cell image. Image processing device +3. - consists of an arithmetic control device 14 and an output layer +7115 that control each device described in 3 and 3 and store the extraction results by the image processing device 13, and the hierarchical neural network high-speed arithmetic device 3 includes a Calculation Iku 31 and high-speed calculation device 3
2 is found as 0111.

The operation of the automatic cell sorting device of this embodiment configured as described above will be described below.

First, a cell image within a specimen is input using the image input device 11 of the cell image input processing device 1, and is stored in the image storage device 2. At the same time, feature parameters are extracted using the image processing device 13, and the extraction results are It is stored in the arithmetic and control unit j4.

After repeating the above process for a fixed number of cells in the sample, the characteristic parameters for all input cells stored in the memory of the arithmetic and control unit 14 or external memory are transmitted via the real-time communication device 2. Then, the hierarchical neural network work is transferred to the high-speed arithmetic unit 3.

In addition, the identification results of the song samples, which have been subjected to identification calculations by the hierarchical neural network high-speed calculation device 3, are sent to the calculation control device 1 of the cell image input processing device 1 via the real-time communication device 2.
4 and outputted by the output device 15. Thereafter, while the hierarchical Nicole network 1 to the work high-speed calculation device 3 are executing identification calculations, the cell image input processing device 1 performs the following steps.
Process the next specimen.

FIG. 4 shows the flow of this process, and so-called pipeline processing is performed. Various execution forms can be considered for this pipeline processing, but by performing multitask processing in the arithmetic and control unit (4) of the cell image input processing device 1, it is possible to carry out the cell image input processing of the next specimen and the previous one. It is also possible to output the specimen identification results in parallel, and furthermore, it is also possible to transfer the characteristic parameters and transfer/output the identification results for multiple shelves at once.

Regarding the configuration of the image input device M112 image processing device 13 described in L, for example, Japanese Patent Laid-Open No. 591.14.
The device shown in FIG. 5 of the 667 publication can be used.

The characteristic parameters that are input to the hierarchical neural network high-speed calculation device 3 include, for cell nuclei, area 2 perimeter, shape (area perimeter ratio).

Regarding cell cytoplasm, area, density, two tones, texture, etc.
Regarding texture, granule, etc., nucleus/cytoplasm sum pair information includes area, concentration, etc., and a sigmoid function, a linear function, etc. can be applied as a threshold function.

In the hierarchical neural network high-speed calculation device 3 of this embodiment, nuclear area, nuclear circumference, nuclear average concentration (red, green, blue)
, nuclear texture 1 (red, green), nuclear texture 2 (red, green), cytoplasm area, cytoplasm concentration (red, green, blue), etc.
The number of units in the input layer is set to 13, corresponding to the three feature parameters, and the number of units in the output layer is set to 30, corresponding to the number of categories of mature cells, immature cells, and artifacts.

Further, the intermediate layer is experimentally set to have 50 to 100 units per layer, and a sigmoid function is used as the threshold function.

However, the configuration of the network is not necessarily limited to this. 1. In the hierarchical neural network high-speed calculation device 3, all input cells transferred from the calculation control device 14 via the real-time communication device 2 Using the feature parameters of (
1) Calculate the formula and obtain the output value Zk (k = 1−nk,
n), : number of categories), m = (k
Imax Zk+ k= 1~nk)
(:3) m ≧ T
(/1.), the input cell belongs to the m-th [1 category. When m(T (5)), the input cell is an unknown cell. For example, the value of T is 0.8. use

The hierarchical neural network high-speed calculation device 3 of this embodiment can also be realized using only a high-speed general-purpose computer. Furthermore, it is also possible to use the output means of the hierarchical neural network high-speed calculation device 3 to output the identification results.

[Effects of the Invention] As described above in detail, according to the present invention, in an automatic cell classification apparatus using a cell image input device having an image input means, an image recording means, and an image processing means, the cell image input processing is performed. The internal structure of the cell image input processing device is improved by connecting the means to the hierarchical network calculating means and the real-time communication means and using the hierarchical network calculating means when performing automatic classification (identification). It is possible to speed up the identification calculation without changing the
This has the remarkable effect of easily realizing an automatic cell classification device capable of high-accuracy and high-speed identification.

[Brief explanation of the drawing]

Figure 1 is an example of the system configuration of an automatic cell classification device, Figure 2 is a diagram showing an example of distribution of feature parameters, Figure 3 is an explanatory diagram of a hierarchical neural network, and Figure 4 is an example of a pipeline processing flow. It is a diagram. 1: Cell image input processing device, 2: Real-time communication device, 3: Hierarchical network calculation device, 11: Image input device, 12
: Image storage device, 13: Image processing device, 14: Arithmetic control device, 15: Output device, 31: General purpose computer, 32: High speed arithmetic device.

Claims (1)

[Scope of Claims] 1. In an automatic cell classification device using a cell image input device having an image input means, an image recording means, and an image processing means, the cell image input processing means is combined with a hierarchical network calculation means and a real-time An automatic cell classification device, characterized in that it is connected by a communication means and uses the hierarchical network calculation means when performing automatic classification (identification). 2. The automatic cell classification device according to claim 1, wherein the cell image input processing for the next specimen and the identification processing for the specimen whose cell image input processing has been completed are performed in parallel.