JP3114277B2 - Construction method of automatic cell classification logic - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern classification device using a hierarchical network, and more particularly to a cell classification device.

[0002]

2. Description of the Related Art As described in Japanese Patent Publication No. 58-29872, in the conventional cell classification (mainly leukocyte classification), an identification logic in which a statistical identification function is combined in multiple stages based on cell characteristic parameters. Alternatively, a branching identification logic is used.

[0003]

In the conventional method of constructing an automatic cell classification logic, a reference learning sample is set and an identification logic is generated (learned). However, if the cell characteristic parameters vary from the initially set range due to differences in sample preparation conditions compared to the reference learning sample, the classification logic cannot be adapted, and the classification logic needs to be changed.

An object of the present invention is to provide a method of constructing an automatic cell classification logic corresponding to a facility for preparing a sample by making use of the feature of a hierarchical network in which classification logic can be easily constructed.

[0005]

In order to achieve the above object, in the method of constructing an automatic cell classification logic of the present invention, a hierarchical neural network is used, and a sample for each facility is added to a reference learning sample to add By creating a learning sample corresponding to a facility, an automatic cell classification logic corresponding to the facility is constructed.

[0006]

In the present invention, a hierarchical neural network is used as the identification logic, and a hierarchical neural network operation device is used for learning and real-time identification.

As a hierarchical neural network,
For example, a network composed of three layers of an input layer, an intermediate layer, and an output layer as shown in FIG. The intermediate layer is generally a multilayer, but an example is shown here. The relationship between input and output is

[0008]

(Equation 1)

And the sigmoid function is

[0010]

(Equation 2)

Is a function given in the form

[0012] In the input layer, units for the number of feature parameters,
The output layer is provided with units corresponding to the number of categories, and the intermediate layer is provided with a number of units corresponding to the number of units in the above two layers. Are assigned as “1” as teacher data to the unit corresponding to “1” and “0” to other units in the output layer, so that an error between the teacher data and the output data is reduced so-called back propagation method (for example, neural network information). Processing: Industrial books, (1
988)) to change the weighting factor between network units. The above processing is repeated to optimize the weight coefficient. This process is learning, and identification is performed based on the weighting factor after learning. Specifically, from the output value Z _k obtained by Equation 1,

[0013]

(Equation 3)

The following m is obtained, and the category to which the input cell belongs is set as the m-th category for identification.

In the above process, the weighting factors between the units are completely different depending on the setting of the learning sample. Therefore, the identification logic can be constructed by the configuration of the learning sample. In the present invention, since the learning sample of the institution is added to the reference sample, it is possible to construct an identification logic that follows the variation of the characteristic parameter of the cell unique to the cell sample created in the institution.

[0016]

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a diagram showing the configuration of the apparatus. As for a learning sample, an output of a learning reference sample memory 1 and an output of a learning target sample memory 2 are guided to a sampling circuit 4 by a selection circuit 3, and a sampling circuit is provided. 4 is stored in the learning sample memory 5. The learning reference sample memory 1 stores the characteristic parameters normalized by the control device 7 in advance. However, another form of the learning reference sample memory may be a read-only memory. On the other hand, the learning target sample memory 2 is configured to be writable by the control device 7 in order to store samples for each facility. Here, the facility indicates a place where a cell sample to be provided for identification is prepared, and is, for example, a hospital or an inspection center. Even if a cell sample is prepared according to a common preparation standard, the optimal logical formula for cell classification is different for each facility because there are differences in the conditions and environment of sample preparation equipment and preparation techniques for each facility. It will be different. Selection circuit 3
Selects the input of the sampling circuit 4 in accordance with an instruction from the control circuit 6, and switches the output between the learning reference sample memory 1 and the learning target sample memory 2. The sampling circuit 4 receives the output of the selection circuit 3 as an input, and
The data of the cells belonging to the category to be sampled is suppressed by the instruction, and the cells not to be sampled are output as they are and stored in the learning sample memory 5. Thereafter, the output of the learning sample memory 5 is provided to a learning device. Learning reference sample memory 1,
Learning target sample memory 2, learning sample memory 5
Has the same configuration, and is composed of normalized characteristic parameters of cells and teacher data. As the characteristic parameters, for the cell nucleus, the area, perimeter, shape (area perimeter ratio, nucleus thickness, unevenness, number of nuclei, etc.), concentration,
Cell cytoplasm such as color tone and texture includes area, density, color tone, texture, granules, etc., and nuclear / cytoplasmic early regression information includes area and density.

FIG. 3 is a block diagram of the sampling circuit of FIG.

The output of the selection circuit 3 is divided into characteristic parameters and teacher data by a distributor 40 and held in latches 41 and 42, respectively. The output of the latch 41 becomes the input of the gate circuit 49. On the other hand, the output of the latch 42 holding the teacher data becomes the input of the decoder 43 and the input of the gate circuit 49. The output of the latch 42 is decoded by the decoder 43, and the output corresponding to the cell category is turned on. The output of the decoder 43 is a counter 4
4 to 47, the counter is preset in advance by the output from the control circuit 6 in FIG. The counters preset here correspond to cells of the category to be sampled. For example, if the counter 46 is a counter to be sampled, the decoder 43
Is counted down from the preset value of the counter 46, and the output of the counter 46 is turned on at intervals of the preset value. When any of the counters 44 to 47 is turned on, the gate circuit 48 is turned on and the gate circuit 49 is turned on to pass the outputs of the latches 41 and 42.
When all the outputs of the counters 44 to 47 are off, the gate circuits 48 and 49 are both off, and the outputs of the latches 41 and 42 do not pass through the gate circuit 49. Only the data that has passed through the gate circuit 49 is stored in the learning sample memory of FIG. Further, the present embodiment can be realized only by a general-purpose computer.

[0020]

According to the present invention, more stable cell classification logic can be constructed by adding a facility-compatible sample to a reference sample in learning of a hierarchical neural network, and the number of learning samples increases by adding two types of samples. Even so, the number of cell images belonging to the category with a high appearance frequency can be sampled, so that the learning sample can be adjusted and the logic of automatic cell classification can be easily constructed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a system configuration according to an embodiment of the present invention.

FIG. 2 is a conceptual diagram illustrating a hierarchical neural network according to an embodiment.

FIG. 3 is a detailed block diagram illustrating an example of a sampling circuit in FIG. 1;

[Explanation of symbols]

 1, 2, 5: learning sample memory, 3: selection circuit, 4: sampling circuit, 6: control circuit, 7: control device, 40: distributor, 41, 42 ... latch 43: decoder, 44-47 ... Counter, 48, 49 ... gate circuit.

Continuation of the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G01N 33/48 G01N 15/10 G06F 15/18 520

Claims (4)

(57) [Claims] 1. A cell having a hierarchical neural network.
Generates the logic required for automatic cell sorting with an automatic sorter
In the construction of automatic cell classification logic
A learning sample corresponding to each facility is added to the learning sample.
To generate the automatic cell classification logic at each facility.
And a method for constructing an automatic cell classification logic. 2. Construction of the automatic cell classification logic according to claim 1.
In the method, when adding the learning sample for each facility,
To select training samples that fall into each category.
Of automatic cell classification logic characterized by adding
Law. (3) Cell with hierarchical neural network
In the automatic classifier, each sample is applied to the reference learning sample.
A means for adding a learning sample for each facility;
Automatic cell classification theory that generates the logic required for automatic cell classification
Automatic cell sorter, comprising:
Place. (4) The automatic cell sorting apparatus according to claim 3.
The means for adding the learning sample includes
Have means to select training samples classified as
An automatic cell sorting apparatus characterized by the above-mentioned.