Discrimination method using deep learning

Abstract

To provide a method which makes up for shortage of teacher data with regard to unusual examples in which teacher data is insufficient in a deep learning.SOLUTION: The present invention relates to a method for determining, by using a deep learning, the group of input data which has a feature common with any of at least two different groups A when there are n number of groups A and m number of groups X and the teacher data of the groups X is less than that of the groups A. The method includes: a data input step of inputting data; an initial determination step of determining whether the input data is of a group A or of a group X having a feature common with the group A by using a first learned model; a final determination step of determining whether the input data is of a group A or of a group X by using a second learned model different from the first learned model; and a determination result output step of outputting the result of determination by the determination step.SELECTED DRAWING: Figure 6

Description

Technical Field

[0001] The present invention relates to a discrimination method using deep learning.

Background Art

[0002] Deep learning, which is a type of machine learning, adopts a method of learning feature quantities that become the points of discrimination by the deep learning program itself, as opposed to general machine learning that requires specifying parameters to focus on. Therefore, the quantity and quality of the exemplary data (teacher data) are very important.

[0003] When performing image discrimination, it is common to use deep learning, and among them, it is common to use a Convolutional Neural Network (CNN) using a method called the error backpropagation method.

[0004] A deep learning device using the error backpropagation method inputs a certain teacher data into a neural network, and calculates an error based on the output value of the CNN and the correct value of the teacher data. The deep learning device inputs the calculated error to the nodes of the output layer in the CNN, and updates the weight count regarding the nodes of the output layer so that the input error becomes smaller. Then, the deep learning device repeats the process of updating the weight count regarding the nodes of each layer so that the error of the output value output from the nodes of the lower layer located in front of the output layer becomes smaller. This operation is equivalent to creating a plurality of small image filters for determining how similar an object example is, and is equivalent to learning which part of the target image can be discriminated more carefully using those filters.

[0005] The fields where deep learning is used have conventionally been mainly applied to what is called big data analysis, which processes a large amount of data. However, in recent years, attempts have been made to use it for the discrimination of rare examples. In the case of rare examples, it is often impossible to collect sufficient teacher data. Therefore, techniques such as data augmentation called data boosting by rotation or reduction, and transfer learning that uses pre-trained data used in other learning have been devised and used. However, data boosting by rotation or reduction is only equivalent to eliminating errors caused by rotation or reduction, and does not fundamentally solve the shortage of teacher data. Summary of the Invention Problems to be Solved by the Invention

[0006] In deep learning, it is an object to provide a method for supplementing insufficient teacher data for rare examples with insufficient teacher data. Means for Solving the Problems

[0007] The present invention made to solve the above problems includes the following inventions: That is, one aspect of the present invention is A method for discriminating the classification of input data, where there are n classifications A and m classifications X as the classifications (n is an integer of 2 or more, m is an integer of 1 or more, provided that m ≤ nC2), the classification X has relatively fewer teacher data than the classification A, and the classification X has features in common with any one of two or more different classifications A.

[0008] As a method for discriminating the classification of input data, it specifically has the following steps. (1) A data input step of inputting data. (2) An initial discrimination step of discriminating whether the input data is an arbitrary classification A or a classification X having features in common with the classification A using a first pre-trained model. (3) A final discrimination step of discriminating whether the input data belongs to the arbitrary classification A or the classification X using a second learned model different from the first learned model. The second learned model is characterized in that part or all of the teacher data of classification A having features common to the classification X different from the arbitrary classification A is diverted as the teacher data of the classification X. (4) A discrimination result output step of outputting the discrimination result of the discrimination step.

Effects of the Invention

[0009] The present invention enables rare examples with insufficient teacher data to supplement the lack of teacher data and improve the discrimination accuracy.

Brief Description of the Drawings

[0010]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Embodiments for Carrying Out the Invention

[0011] Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings. However, the embodiments described below are merely examples of forms to which the technical idea of the present invention is applied, and the present invention is not limited to the content of these embodiments. The present invention is applicable to one-dimensional data as well, but since deep learning generally often uses image data (hereinafter sometimes simply referred to as an image), the description will be given by taking an image as an example. (Classification A and Classification X) Classification A refers to a classification with sufficient teacher data, and Classification X refers to a classification with relatively fewer teacher data than Classification A. There are n or more (n is an integer of 2 or more) of Classification A and m or more (m is an integer of 1 or more, provided that m ≤ nC2) of Classification X, and Classification X has two or more different common features with any of Classification A. FIG. 1 shows a Venn diagram representing the relationship in the case where there are 2 of Classification A and 1 of Classification X. It can be seen that Classification X1 has common features (F1, F2) different from Classification A1 and Classification A2 respectively. It can also be seen that Classification A1 does not have F2 and Classification A2 does not have F1. F1 and F2 may be any differences in features, such as being a cell, round, monochrome, etc. (Deep Learning) FIG. 2 shows a flowchart of learning in deep learning. First, teacher data for learning is acquired for the number of classifications to be discriminated (step P11), and using this, a learned model for distinguishing images of each classification is created (step P12). This learned model has already been pre-examined (hereinafter sometimes referred to as labeling) as to which classification it usually belongs to. This learned model, in the case of a CNN, is a plurality of image filters created by learning and the weighting values indicating how much the pixels of the image filter are to be adopted in the determination. In the present invention, the description is made on the premise of a CNN, but it is not limited to a CNN as long as the principle is similar.

[0012] Next, obtain the data to be discriminated (hereinafter sometimes referred to as discrimination data) (step P13). The discrimination data does not have to be labeled and can be any image, but in practical use, it is desirable that it corresponds to any of the classifications included in the learned model. Then, use the learned model to discriminate the obtained discrimination data (step P14). At this time, the raw data obtained by the discrimination is the degree of coincidence indicating how much it matches each classification filter, and finally, it is determined that the discrimination data belongs to the classification with the highest degree of coincidence, and the determination result is output (step P15). In this description, the process until the creation of the learned model in the first half is called the learning stage, and the process from the acquisition of discrimination data for discrimination to the output of the result is called the discrimination stage. (Tree model) Fig. 3 shows a tree when the discrimination of three classifications, classification A1, classification A2, and classification X1, is performed in one-stage discrimination. In this case, in the discrimination unit, each filter of classification A1, classification A2, and classification X1 is sequentially applied to the input discrimination data, and the degree of coincidence is calculated for each filter. Then, finally, the discriminated discrimination data is determined to be the classification with the highest degree of coincidence.

[0013] When using the filter of classification A2, the degree of coincidence of the data belonging to classification X1 having the same characteristics as the data belonging to classification A2 is the second highest, and the data belonging to classification A1, which has the least similar characteristics to the data belonging to classification A2, has the lowest degree of coincidence. At this time, if there is definitely no data belonging to classification A2 in the discrimination data, then what is selected using the filter learned for classification A2 will definitely be classification X1.

[0014] For example, when using the data belonging to classification A1 and the data belonging to classification X1 as teacher data for learning, it is considered that the learning device learns more about the common items (area F1) of classification A1 and classification X1. Also, when using the data belonging to classification A2 and the data belonging to classification X1 as teacher data for learning, it is considered that the learning device learns more about the common items (area F2) of classification A2 and classification X1.

[0015] Fig. 4 shows a tree when the discrimination of three classifications, i.e., classification A1, classification A2, and classification X1, is performed in two steps. S12 represents an initial discrimination step described later, and S14 represents a final discrimination step. First, in S12, classification A1 or classification X1 is discriminated from classification A2, and in S14, classification A1 is discriminated from classification X1. In S12, if the discrimination data is classification A1 or classification X1, the discrimination is performed in S14, and if it is discriminated as classification A2, the process ends as it is. In S14, an image determined to be classification A1 or classification X1 is received, and it is discriminated whether it is classification A1 or classification X1, and then the process ends. (Discrimination device) Fig. 5 is a functional block diagram showing an example of the schematic configuration of a discrimination device 1 that can be used in the present invention. The discrimination device 1 includes an input unit 10, a main control unit 20, and a display unit 30. The main control unit 20 includes a discrimination unit 40, a learning model 50, a learning unit 60, and a discrimination result output unit 70. The discrimination unit 40 includes a first discrimination unit 41 and a second discrimination unit 42. Further, the learning unit 60 functions as a learning device for constructing a learned model used for discrimination, and includes a first learning unit 61 and a second learning unit 62. The learning model 50 includes a first learned model 51 and a second learned model 52.

[0016] The discrimination device 1 is a device that discriminates the classification of the input data. (Input unit) The teacher data input unit 10 receives the input of the teacher data of the image to be discriminated. The teacher data input unit 11 receives the input of the above-mentioned image by reading the teacher data file stored in the storage medium or receiving the image from another device via a wired or wireless network. The teacher data input unit 10 transmits the received image to the main control unit 20. The teacher data can be used forever as long as there is no problem with the learned data created after one-time learning. Once the learned model is completed, there is no need to input new data thereafter. The discrimination data input unit 11 receives the input of the image to be discriminated, and each time an image to be determined is input, the received image is transmitted to the discrimination unit 40. (Image) The teacher data input unit 11 inputs data that has been pre-labeled with information indicating whether it belongs to classification A1, classification A2, or classification X1. The discrimination data input unit 11 inputs discrimination data. It is not specified which of classification A1, classification A2, or classification X1 this discrimination data belongs to, and it is discriminated by the discrimination unit 40.

[0017] The image is not limited to an image obtained by a specific imaging method, and images obtained by a plurality of imaging methods may be combined and used. For example, in the case of a microscopic observation image, two or more observation images selected from bright-field observation images, dark-field observation images, phase-contrast observation images, and fluorescence observation images can be combined and used. Also, when using a combination of multiple images, the multiple images can be synthesized into one image. For example, it may be synthesized into one image by performing an overlay process, or they may be arranged side by side and combined without overlaying to form one image.

[0018] Also, for each image, as preprocessing, contrast adjustment and adjustment of another channel based on a specific channel may be performed. For example, a process of normalizing the luminance values of other channels based on the luminance value of a specific channel may be performed.

[0019] Examples of the object of the image include cells or fine particles similar to cells. (Trained model) The first trained model 51 used in the first discrimination unit 41 takes as input teacher data for the first discrimination that has been input from the teacher data input unit 10 of the discrimination device 1 and preprocessed by the teacher data generation unit 63, and is constructed by the first learning unit 61. Specifically, this first trained model is a set of a trained model (filter) created using data belonging to classification A1 and classification X1 as teacher data, and a trained model created using data belonging to classification A2 as teacher data.

[0020] On the other hand, the second pre-trained model 52 used in the second discrimination unit 42 is constructed by the second learning unit 62 with the input of teacher data for the second discrimination, which is input from the teacher data input unit 10 of the discrimination device 1 and pre-processed by the teacher data generation unit 63. Specifically, this second pre-trained model is a set of a pre-trained model created from the teacher data of classification A1 and a pre-trained model created from the teacher data of classification X1 and classification A2.

[0021] The first pre-trained model 51 and the second pre-trained model 52 are models constructed by training a CNN having a plurality of convolutional layers using the above-described teacher data, respectively. (Teacher data generation unit) The teacher data generation unit 63 creates teacher data to be used in the first learning unit 61 and the second learning unit 62 from the image data input to the teacher data input unit 10. Some pre-processing operations that enable efficient learning are performed by this teacher data generation unit 63 according to the situation. For example, when there is a large variation in the number of data, the number of data of the cell type with the smallest number existing in the input data is adjusted, or operations such as data augmentation that increases the number of learning images are performed by rotating and inverting the images.

[0022] It will also be set here which teacher image is applied to the pre-trained model of which classification, as also described in the item of the learning model, but this function may be set as a program that is performed by the learning unit according to the situation. (First learning unit and second learning unit) The first learning unit 61 constructs the first pre-trained model 51 by using the teacher data generated by the teacher data generation unit 63 and by a known machine learning method. On the other hand, the second learning unit 62 constructs the second pre-trained model 52 by using the teacher data for the second learning unit 62 generated by the teacher data generation unit 63 and by a known deep learning method. As described above, the first learning unit 61 and the second learning unit 62 make the CNN learn the correspondence between the input images and their label information.

[0023] Note that the first learned model 51 and the second learned model 52 may be constructed by another learning device including a first learning unit 61, a second learning unit 62, and a teacher data generation unit 63, which exists independently of the discrimination device 1. When the learning device exists independently of the discrimination device 1, the discrimination device 1 can use each learned model in the discrimination device 1 by reading each learned model stored in the storage medium or receiving each learned model from another device via a wired or wireless network. (First Discrimination Unit) The first discrimination unit 41 inputs discrimination data into the first learned model 51, and determines whether the discrimination data input to the discrimination data input unit 11 is classified as A1 or X1, or classified as A2, from the output result of the first learned model 51. Specifically, the one with the highest probability is adopted as the discrimination result. When it is determined to be classified as A1 or X1, it is sent to the second discrimination unit 42. When it is determined to be classified as A2, the discrimination result is sent to the discrimination result output unit 70. (Second Discrimination Unit) The second discrimination unit 42 inputs an image into the second learned model 52 only when the discrimination result indicating classification A1 or X1 is obtained in the first discrimination unit 41, and determines whether it is classified as A1 or X1 from the second learned model 52. Specifically, the one with the highest probability is adopted as the discrimination result. The second discrimination unit 42 sends the discrimination result to the discrimination result output unit 70. (Display Unit) The display unit 30 is a device that displays the final discrimination result output from the discrimination result output unit 70. The display unit 30 is a display device that displays the final discrimination result as image data or character data. Note that the display unit is not limited to being provided in the discrimination device 1, and may be provided as an external device connectable to the discrimination device 1. (Method of the Present Invention) Next, an example of the flow when performing the method of the present invention using the discrimination device 1 will be described with reference to FIG. 6.

[0024] First, the discrimination unit 40 acquires an image input by the user's input operation via the teacher data input unit 11 (step S11: data input step). Here, the image to be discriminated is acquired by the same processing operation as the images included in the teacher data used for constructing the first learned model 51 and the second learned model.

[0025] Next, the first discrimination unit 41 inputs the image to the first learned model 51 and acquires the output from the first learned model 51. Then, from the output result, it is discriminated whether the image indicates classification A1 or classification X1, or classification A2 (step S12: initial discrimination step). When the discrimination result is that the image indicates classification A1 or classification X1 (yes in step S13), this time the second discrimination unit 42 inputs the same image as the image input to the first learned model 51 to the second learned model 52 and acquires the output from the second learned model 52. Then, from the output result, it is discriminated whether the image is classification A1 or classification X1 (step S14: final discrimination step), and the discrimination result is transmitted to the discrimination result output unit 70. On the other hand, when the discrimination in the first discrimination unit 41 results in the discrimination that the image indicates classification A2 (no in step S13), the discrimination result is transmitted to the discrimination result output unit 70. When the discrimination result is sent from the second discrimination unit 42, the discrimination result output unit 70 outputs the discrimination result from the second discrimination unit 42 to the display unit 30, and when the discrimination result that the image is classification A2 is sent from the first discrimination unit 41, the discrimination result from the first discrimination unit 41 is output to the display unit 30 (step S15: discrimination result output step).

[0026] As described above, in the discrimination process using the discrimination device 1, the discrimination of the image is performed in two-stage discrimination steps. Thereby, it is possible to make good use of the teacher data of classification A2 for the shortage of teacher data of the target classification X1 and improve the learning accuracy without degrading the learning accuracy. Since the teacher data of classification A2 is intended to supplement the shortage of teacher data of classification X1, part or all of it may be diverted so as to be equivalent to the number of teacher data of classifications A1 and A2. (Method 2 of the present invention) An application example of the present invention will be described when there are three Classifications A (Classification A1, A2, A3) and three Classifications X (Classification X1, X2, X3). For example, it is assumed that each classification has a relationship as represented by the Venn diagram in FIG. 7. In this case, a decision tree as shown in FIG. 8 is considered appropriate. In step S21, a determination is made for Classification A1 or Classification X1, Classification A2 or Classification X2, and Classification A3 or Classification X3 (initial determination step).

[0027] Next, in step S22, a determination is made between Classification A1 and Classification X1 (first final determination step). At this time, since the teacher data of Classification A2 having characteristics similar to those of Classification X1 can be diverted as the teacher data of Classification X1, a decrease in accuracy due to insufficient teacher data does not occur during the learning stage of Classification X1.

[0028] Also, in step S23, a determination is made between Classification A2 and Classification X2 (second final determination step). At this time, the teacher data of Classification A3 having characteristics similar to those of Classification X2 can be diverted as the teacher data of Classification X2.

[0029] Similarly, in step S24, a determination is made between Classification A3 and Classification X3 (third final determination step), and the teacher data of Classification A1 having characteristics similar to those of Classification X3 can be diverted as the teacher data of Classification X3.

[0030] Also, when there are common characteristics even in Classifications X1 and X3 as in the Venn diagram of FIG. 9, a decision tree as shown in FIG. 10 is considered more appropriate. In step S31, a determination is made for Classification A1 or Classification X1 or Classification X3, Classification A2 or Classification X2, and Classification A3 (initial determination step).

[0031] Next, in step S32, classification A1 or classification X1 is discriminated from classification X3 (intermediate discrimination step). In the learning process of classification X3, the lack of teacher data can be eliminated by diverting the teacher data of classification A3. When there are three or more classifications X (classification X1, X2, X3) similar to classification A1, it may be discriminated between classification A1 or classification X1 and classification X2 or X3, or it may be discriminated between classification A1 or classification X1 or classification X2 and classification X3, and then between classification A1 or classification X1 and classification X3. The discrimination process may be repeated until there is only one classification X similar to classification A1. In any case, since the number of learned models required is equal to the number of steps, the learning unit and the discrimination unit also need to be increased accordingly.

[0032] In step S33, classification A1 is discriminated from classification X1 (first final discrimination step). In the learning process of classification X1, the lack of teacher data for classification X1 can be eliminated by diverting the teacher data of classification A2 to the teacher data.

[0033] In step S34, classification A2 is discriminated from classification X2 (second final discrimination step). By putting the teacher data of classification A3 into the teacher data of classification X2, the lack of teacher data for classification X2 can be solved.

[0034] The above description is merely an example when the number of classifications and the number of judgment steps increase. Depending on the nature of the common features in the actual judgment, the classification A to be diverted and the number of multi-stages may become more complex. (Example of software implementation) The control block of the discrimination device 1 (main control unit 20, particularly discrimination unit 40, learning unit 60, and discrimination result output unit 70) may be realized by a logic circuit (hardware) formed in an integrated circuit (IC chip) or the like, or may be realized by software.

[0035] In the latter case, the discrimination device 1 includes a computer that executes instructions of a program, which is software for realizing each function. This computer includes, for example, at least one processor (control device) and at least one computer-readable recording medium that stores the above program. Then, in the above computer, when the above processor reads and executes the above program from the above recording medium, the object of the present invention is achieved. As the above processor, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) can be used. As the above recording medium, a "non-transitory tangible medium", for example, in addition to a ROM (Read Only Memory), a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, etc. can be used. Further, it may further include a RAM (Random Access Memory) for expanding the above program. Further, the above program may be supplied to the above computer via any transmission medium (communication network, broadcast wave, etc.) capable of transmitting the program. Note that one aspect of the present invention can also be realized in the form of a data signal embedded in a carrier wave, in which the above program is embodied by electronic transmission.

Example

[0036] The results of applying the present invention to the discrimination between CK-positive circulating cancer cells (hereinafter, CK-positive CTCs), which are one of rare cells, CK-negative circulating cancer cells (CK-negative CTCs) that have characteristics similar to CK-positive CTCs but for which relatively many sample images are available, and debris that resembles cells but only CK fluoresces (hereinafter, CK-positive debris) are shown. The imaging conditions for the learning data images and discrimination data used were as follows: a 30-micron well structure was created on a silicon substrate, the target cells or debris were placed in this well, and a total of four images, namely a bright-field image and three fluorescence images, were observed using a fluorescence microscope. This group of four images was used as one set (the size of each image was 64×64 pixels) for discrimination by deep learning. In this fluorescence observation, the fluorescence emissions of CK (a marker indicating the possibility of epithelial-mesenchymal transition), DAPI (a marker that stains cell nuclei), and CD45 (a marker for determining whether it is a white blood cell) were measured, and based on their luminance values, each classification and labeling were performed manually by experts in advance. CK-positive CTCs that were positive for both DAPI and CK and negative for CD45 were assigned to classification X1, CK-negative CTCs that were only positive for DAPI and negative for CK and CD45 were assigned to classification A1, and images of CK-positive debris that were only positive for CK and negative for DAPI and CD45 were assigned to classification A2. Figure 11 shows example images of each cell and the corresponding relationship of the classifications.

[0037] The learning and discrimination program was created in the programming language Python, and a two-stage discrimination program with the functions shown in the tree of Figure 4 was independently constructed using a deep learning library developed by Preferred Networks, Inc., and learning and discrimination were performed. In order to compare the performance of the image discrimination method of the present invention, the discrimination program used in this study was discriminated using the same parameters without changing any conditions except for the number and quality of the learning data. In addition, in order to compare with the two-stage discrimination method, a study was also conducted using a one-stage discrimination program equivalent to the tree of Figure 3. Also in that case, except for the part of the source code that makes the difference between the two-stage and one-stage, the content and parameters of the program were studied under the same conditions.

[0038] The number of data used was 200 for classification A1 and classification A2, and 110 for classification X1. Regarding classification A1 and classification A2, they were divided into 100 for learning and 100 for discrimination. For classification X1, 10 were used for learning and 100 for discrimination. To grasp the variation in discrimination, the above examination was conducted 3 times from the point of dividing into learning images and discrimination images to the point of discrimination, and the average value and error were calculated. In these 3 times of learning and determination, the combination of learning images and discrimination images was adjusted so as to be different each time. The discrimination results are shown in Table 1.

[0039]

Table 1

[0040] The results in column (1) of Table 1 are the results of one-stage discrimination, the results in column (2) are the results of two-stage discrimination without using the data transfer method of the present invention, and the results in column (3) are the results using the data transfer method of the present invention. By using the present invention, the correct answer rate (positive medium rate) which was 44.0 ± 7.9% in one-stage discrimination and 27.3 ± 3.8% in two-stage discrimination without diversion of learning data increased to 71.7 ± 2.1%. This is because in the one-stage discrimination method and the two-stage discrimination method without using the present invention, there were only 10 images of classification X1, and it was not possible to well learn the features that would distinguish them from other cells. However, by using the present invention, the learning data increased to 110, and more emphasis was placed on learning the point that DAPI, which is a common feature, is positive, resulting in an improved discrimination result.

Explanation of Signs

[0041] 1 Discrimination device 10 Teacher data input unit 11 Discrimination data input unit 20 Main control unit 30 Display unit 40 Discrimination unit 41 First discrimination unit 42 Second discrimination unit 50 Learning model 51 First learned model 52 Second learned model 60 Learning unit (learning device) 61 First learning unit 62 Second learning unit 63 Teacher data generation unit 70 Discrimination result output unit

Claims

1. A method for determining the classification of input data, comprising: As the classifications, there are n classifications A and m classifications X (n is an integer of 2 or more, m is an integer of 1 or more, provided that m ≤ nC2), The number of teacher data for the classification X is relatively smaller than that for the classification A, and the classification X has two or more different features in common with any one of the classifications A, The method includes: A data input step of inputting data; An initial determination step of determining, using a first learned model, whether the input data is any classification A or a classification X having a feature in common with the classification A; In the initial determination step, when there are two or more classifications X having a feature in common with any classification A, An intermediate determination step of repeatedly performing, using a learned model different from the first learned model, a step of determining one of the classifications X and the other classifications until there is only one classification X having a feature in common with the any classification A, by diverting part or all of the teacher data of the classification A having a feature in common with another classification X different from the any classification A as the teacher data of the classification X; A final determination step of determining, using a second learned model different from the first learned model, whether the input data is the any classification A or the classification X; A determination result output step of outputting the determination result of the final determination step; And The method is characterized in that the second learned model diverts part or all of the teacher data of the classification A having a feature in common with a classification X different from the any classification A as the teacher data of the classification X.

2. A method for determining the classification of input data, comprising: As the classifications, there are n classifications A and m classifications X (n is an integer of 2 or more, m is an integer of 1 or more, provided that m ≤ nC2), The number of teacher data for the classification X is relatively smaller than that for the classification A, and the classification X has two or more different features in common with any one of the classifications A, The method includes: A data input step of inputting data; An initial determination step of determining, using a first learned model, whether the input data is any classification A or has a feature in common with the classification A; And In the initial determination step, when there are two or more classifications X having a feature in common with any classification A, An intermediate discrimination step of discriminating any one of the classifications X from the other classifications using a learned model different from the first learned model that transfers part or all of the teacher data of classification A having features common to the other classification X different from the arbitrary classification A to the teacher data of the other classification X; A final discrimination step of discriminating whether the input data is the arbitrary classification A or the classification X using a second learned model different from the first learned model; A discrimination result output step of outputting the discrimination result of the final discrimination step; comprising: The method, wherein the second learned model transfers part or all of the teacher data of classification A having features common to the classification X different from the arbitrary classification A to the teacher data of the classification X.

3. The method according to claim 1 or 2, wherein the data is an image and the learning method is a convolutional neural network.

4. The method according to claim 3, wherein the image is a bright-field image or a fluorescence image of cells or microparticles similar to cells.

Images