Learning Method and Learning Device for Machine Learning Model

Abstract

Provided is a method for training a machine learning model, whereby more appropriate training can be performed for characteristic prediction after cell culturing. This method for training a machine learning model comprises: a first step for performing, at least three times, a cell culture experiment in which a plurality of cell images are captured during or after culturing of cells, and characteristic information of a plurality of cells in the plurality of cell images is acquired after the cell culturing; and a second step for using some or all of combination pairs of the characteristic information and the captured plurality of cell images as training data. In the second step, information indicating a size relationship of the plurality of characteristic information of the plurality of cell images is classified into at least the three classes of large, medium, and small, and the cell images in the large and small classes are trained as training data.

Description

Technical Field

[0001] The present invention relates to a method and apparatus for training a machine learning model for predicting the characteristics of cells after cell culture.

Background Art

[0002] For example, pluripotent stem cells are stem cells that can differentiate into most cells and are expected to be applied to drug discovery and regenerative medicine. However, the technology for culturing pluripotent stem cells and efficiently inducing differentiation into specific cells or cell tissues is still at the research level, and the establishment of an effective technology is awaited.

[0003] In order to establish an effective technology, it is necessary to confirm whether the cultured or differentiation-induced cells satisfy the expected characteristics.

[0004] Generally, evaluation of expression using undifferentiated markers or differentiation markers, evaluation of cells after differentiation induction by fluorescence microscopy or flow cytometry, etc. are performed.

[0005] In addition, there are those that use machine learning as shown in Patent Document 1.

Prior Art Documents

Patent Documents

[0006]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0007] The technique described in Patent Document 1 differentiates induced pluripotent stem cells into predetermined cells, acquires cell information of the cells induced to differentiate into the predetermined cells, and acquires teacher data including the cell information and physical property information of the cells based on measurement using an analyzer. Based on the teacher data, a learning model is generated that outputs the physical property information of the cells when the cell information of the cells induced to differentiate into the predetermined cells is input. This is a method for generating a learning model.

[0008] Cell information includes, for example, cell images after culturing. Physical property information includes, for example, cell viability.

[0009] In the technique described in Patent Document 1, when the cell information is only a plurality of unstained cell images and the physical property information is specific numerical information such as cell viability, although not described in Patent Document 1, an approach of preparing a plurality of data sets with the same label corresponding to the numerical information for all cell images and performing learning can be considered. It is considered that such an approach enables more specific and appropriate learning.

[0010] That is, when the physical property (characteristic) is specific numerical information, if a plurality of data sets with the same label corresponding to the numerical information can be prepared for all cell images and learning can be performed, learning can be specifically performed, and more accurate learning can be achieved.

[0011] However, with the above approach, there is a problem that many cell images with incorrect labels are generated, and accurate learning cannot be performed.

[0012] In images in a culture experiment where the physical property is at an intermediate level, it is considered that those with information acting positively on the physical property and those with information acting negatively on the physical property are evenly mixed. Therefore, if the same label is attached, many images with labels that do not match the information contained in the image, that is, incorrect labels, will be generated. Accurate learning is difficult for images with incorrect labels.

[0013] An object of the present invention is to provide a learning method and a learning device for a machine learning model that can solve the above-described problems and perform more appropriate learning for predicting characteristics after cell culture. **Means for Solving the Problems**

[0014] To achieve the above object, the present invention is configured as follows.

[0015] A learning method for a machine learning model includes: a first step of taking a plurality of cell images during or after culturing of cells, and performing at least three cell culture experiments for obtaining Information indicating the magnitude relationship of the characteristics of the cells after culture characteristics of a plurality of cells in the plurality of cell images after culturing the cells; and a second step of using a part or all of pairs obtained by combining the plurality of taken cell images and the as training data. The second step classifies the plurality of Information indicating the magnitude relationship in the plurality of cell images into at least three classes of large, medium, and small, and learns the cell images of the large and small classes as training data. Information indicating the magnitude relationship

[0016] A learning device includes: a first data storage unit that stores a plurality of cell images taken during or after culturing of cells and characteristics of cells in the plurality of cell images; a data selection unit that classifies the plurality of Information indicating the magnitude relationship of the characteristics of the cells after culture stored in the first data storage unit into at least three classes of large, medium, and small, and selects the cell images of the large and small classes; a second data storage unit that stores the cell images of the large and small classes selected by the data selection unit together with the Information indicating the magnitude relationship ; and a machine learning model that learns the cell images of the large and small classes stored in the second data storage unit as training data. Information indicating the magnitude relationship **Advantages of the Invention**

[0017] It is possible to provide a learning method and a learning device for a machine learning model that can perform more appropriate learning for predicting characteristics after cell culture. **Brief Description of the Drawings**

[0018]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Mode for Carrying Out the Invention

[0019] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

Examples

[0020] (Example 1) Example 1 of the present invention will be described with reference to FIGS. 1 to 2.

[0021] FIG. 1 is a block diagram showing a learning device 100 according to Example 1. In FIG. 1, the learning device 100 includes a data storage unit 101 (first data storage unit) that stores acquisition information in three or more cell culture experiments, a machine learning model 107, a data storage unit 105 that stores machine learning data to be input to the machine learning model 107, and a data selection unit 104 that selects the data in the data storage unit 101 and outputs it to the data storage unit 105 (second data storage unit).

[0022] The learning device 100 may be a general-purpose computer such as a PC or a dedicated device. When the learning device 100 is a general-purpose computer, the data storage units 101 and 105 are memories or storages.

[0023] The acquired information 1010, 1020, ··· 1030 in the cell culture experiments includes cell image groups 1011, 1021, ··· 1031 during or after culture and characteristic information 1012, 1022, ··· 1032 after culture in each cell culture experiment. The characteristic information 1012, 1022, ··· 1032 is information representing the characteristics of the cells after culture, such as the differentiation induction efficiency, and is information having information indicating a magnitude relationship.

[0024] The information representing the characteristics of the cells after culture may be numerical information representing a ratio or information representing a degree such as large, medium, and small. However, it is information for which the order between the characteristic information can be determined.

[0025] The data selection unit 104 refers to the characteristic information 1012, 1022, ··· 1032 stored in the data storage unit 101 and the characteristic information of cell culture experiments other than the above included in the data storage unit 101, selects a cell image group excluding the cell image group of the cell culture experiment having intermediate characteristic information therein, and sets the cell image group of the cell culture experiment having characteristic information larger than the above intermediate characteristic information as machine learning data 1050 corresponding to the label 1053 corresponding to the characteristic information, and sets the cell image group of the cell culture experiment having characteristic information smaller than the above intermediate characteristic information as machine learning data 1060 corresponding to the label 1063 corresponding to the characteristic information, and stores them in the data storage unit 105. The machine learning data 1050 includes cell image groups 1051 ··· 1052, and the machine learning data 1060 includes cell image groups 1061 ··· 1062.

[0026] The data selection unit 104 may be realized by hardware, may be realized by software, or the same operation may be performed manually.

[0027] An example of realizing the data selection unit 104 by software is shown in FIG. 2. FIG. 2 is an operation flowchart showing an example of the data selection unit 104.

[0028] In FIG. 2, when the process starts (step 201), first, threshold value A and threshold value B where A > B are determined from the characteristic information of all cell culture experiments stored in data storage unit 101 (step 202). In step 202, it is assumed that the characteristic information is numerical.

[0029] Next, it is determined whether there is an unselected cell culture experiment (step 203). In step 203, if there is no unselected cell culture experiment, the process ends (step 209), and if there is an unselected cell culture experiment, the process proceeds to step 204.

[0030] In step 204, an unselected cell culture experiment is selected. Then, it is determined whether the characteristic information of the selected cell culture experiment is A or more (step 205).

[0031] In step 205, if the characteristic information is A or more, the cell image group of that cell culture experiment is stored in the cell image groups 1051 ··· 1052 of the machine learning data 1050 in data storage unit 105 (step 206), and the process returns to step 203.

[0032] In step 205, if the characteristic information is less than A, it is determined whether the characteristic information of the selected cell culture experiment is B or less (step 207).

[0033] In step 207, if the characteristic information is B or less, the cell image group of that cell culture experiment is stored in the cell image groups 1061 ··· 1062 of the machine learning data 1060 in data storage unit 105 (step 208), and the process returns to step 203. In step 207, if the characteristic information is not B or less, the process returns to step 203.

[0034] The processes from step 204 to step 208 are performed on the acquisition information 1010, 1020 ··· 1030 in all cell culture experiments stored in data storage means 101.

[0035] In the example shown in FIG. 2, it is assumed that the characteristic information after cell culture is numerical. However, when the characteristic information is qualitative information such as "large", "medium", or "small", the cell image group of the cell culture experiment in which the characteristic information is "large" may be stored in the machine learning data 1050 of the data storage unit 105, and the cell image group of the cell culture experiment in which the characteristic information is "small" may be stored in the machine learning data 1060 of the data storage unit 105.

[0036] When there are four or more types of qualitative characteristic information, one or more pieces of characteristic information with a continuous magnitude relationship are determined from the remaining characteristic information excluding the characteristic information corresponding to the maximum and the characteristic information corresponding to the minimum. The cell image group of the cell culture experiment with characteristic information larger than that is stored in the machine learning data 1050 of the data storage unit 105, and the cell image group of the cell culture experiment with characteristic information smaller than that is stored in the machine learning data 1060 of the data storage unit 105.

[0037] The machine learning model 107 performs learning using the data with the label 1053 attached as the teacher label to a part or all of the cell image group of the machine learning data 1050 in the data storage unit 105 and the data with the label 1063 attached as the teacher label to a part or all of the cell image group of the machine learning data 1060 in the data storage unit 105.

[0038] It is considered that the cell images of the cell culture experiment with intermediate characteristic information contain more characteristic information in the cell culture experiment with smaller characteristic information than the cell images of the cell culture experiment with larger characteristic information.

[0039] Also, it is considered that the cell images of the cell culture experiment with intermediate characteristic information contain more characteristic information in the cell culture experiment with larger characteristic information than the cell images of the cell culture experiment with smaller characteristic information.

[0040] Therefore, as described above, by preventing the characteristic information from including cell images of intermediate cell culture experiments in the training data, the chance of data with incorrect labels being added to the training data can be reduced. As a result, the possibility of realizing a more appropriate machine learning model can be increased.

[0041] In addition, using the machine learning model learned by the above learning method or learning device, the cell images obtained in the cell culture experiment are inferred, the label corresponding to the cell image with a high confidence level output by the machine learning model is selected, and by learning to classify into two classes using some or all of the labels corresponding to the cell images with a high confidence level, the chance of data with incorrect labels being added to the training data can be further reduced. As a result, the possibility of realizing a more appropriate machine learning model can be increased.

[0042] That is, according to Example 1 of the present invention, it is possible to provide a learning method and a learning device for a machine learning model capable of performing more appropriate learning for predicting characteristics after cell culture.

[0043] The learning method and learning device according to Example 1 of the present invention include: a first step of performing at least three cell culture experiments of photographing a plurality of cell images during or after culturing and acquiring characteristic information of a plurality of cells in the plurality of cell images after culturing the cells; and a second step of using a part or all of a pair combining the cell images obtained by photographing and the characteristic information of the corresponding cells as training data. In the second step, information indicating the magnitude relationship of the characteristic information of the plurality of cell images obtained by the cell culture experiment is classified into at least three classes of large, medium, and small in ascending order according to the information indicating the magnitude relationship, and the cell images of the large and small classes are used as training data for learning. This is for cell culture experiments excluding at least one cell culture experiment in which the characteristic information is continuous and which is neither the cell culture experiment with the maximum characteristic information nor the cell culture experiment with the minimum characteristic information. For the cell culture experiments, it is also possible to classify them into two classes: a class of all cell culture experiments having characteristic information larger than the characteristic information of the excluded cell culture experiment and a class of all cell culture experiments having characteristic information smaller than the characteristic information of the excluded cell culture experiment, and use the cell images of the classified cell experiments as training data for learning.

[0044] Also, in the learning method and learning device according to Example 1, a plurality of cell images are photographed during or after culturing, and at least three cell culture experiments of acquiring characteristic information after culturing are performed. A part or all of the pairs of the obtained cell images and the characteristic information are used as training data for a machine learning model. Specific thresholds A and B (A > B) are determined. When the maximum value of the characteristic information is C and the minimum value is D, it may be classified into two classes: a class of all cell experiments in which the characteristic information is greater than or equal to min(A, C) and a class of all cell experiments in which the characteristic information is less than or equal to max(B, D), and then learned. That is, it may be classified into two classes: a class of all cell images in which the characteristic information ranges from the threshold A or more to the maximum value C and a class of all cell images in which the characteristic information ranges from the threshold B or less to the minimum value D, and the cell images of the two classes may be used as training data for learning.

[0045] (Example 2) Next, Example 2 of the present invention will be described with reference to FIG. 3.

[0046] FIG. 3 is a block diagram showing a learning device 300 according to Example 2. In FIG. 3, the learning device 300 includes a machine learning model 307, data storage units 305 (third data storage unit) and 315 (fourth data storage unit) that store machine learning data to be input to the machine learning model 307, and a data storage unit 309 (fifth data storage unit) that stores the inference result of the machine learning model 307. Based on the inference result of the data storage unit 309, a data selection unit 308 that selects the cell image group of the data storage unit 305 and stores it in the data storage unit 315.

[0047] The learning device 300 may be a general-purpose computer such as a PC or a dedicated device. When the learning device 300 is a general-purpose computer, the data storage units 305 and 315 are memories or storages.

[0048] Stored in the data storage unit 305 are machine learning data 3050 corresponding to the label 3053 corresponding to the characteristic information and data 3060 corresponding to the label 3063 corresponding to the characteristic information, similar to the data storage unit 105. The machine learning data 3050 includes a cell image group 3051 ··· 3052, and the data 3060 includes a cell image group 3061 ··· 3062.

[0049] The machine learning model 307 performs learning using, as training data, data in which the label 3053 corresponding to the characteristic information is attached as a teacher label to a part or all of the cell image group 3051 ··· 3052 of the data 3050 in the data storage unit 305, and data in which the label 3063 corresponding to the characteristic information is attached as a teacher label to a part or all of the cell image group 3061 ··· 3062 of the data 3060 in the data storage unit 305.

[0050] Next, the learned machine learning model 307 infers each of all the cell image groups 3051···3052, 3061···3062 in the data storage unit 305, and obtains the confidence level regarding the correct label for each cell image. This inference result is stored in the data storage unit 309.

[0051] Based on the inference results stored in the data storage unit 309, the data selection unit 308 selects cell images with a high confidence level regarding the correct label for each of the cell images stored in the data storage unit 305, and stores them in the data storage unit 315 as data corresponding to the correct label.

[0052] The data selection unit 308 may be implemented in hardware, may be implemented in software, or the same operation may be performed manually.

[0053] Furthermore, the machine learning model 307 uses data in which labels 3053 are attached as teacher labels to some or all of the cell image groups 3151···3152 of the machine learning data 3150 in the data storage unit 315, and data in which labels 3063 are attached as teacher labels to some or all of the cell image groups 3161···3162 of the machine learning data 3160 in the data storage unit 315 for learning.

[0054] In the above-described Example 2, an example using one machine learning model was shown. However, a plurality of machine learning models learned by changing training data, parameters, initial values of weights which are internal parameters, etc. may be created, and their inference results may be comprehensively judged to select the data to be stored in the data storage unit 315.

[0055] As described above, according to Example 2, all the cell image groups 3051···3052, 3061···3062 stored in the data storage unit 305 are inferred, based on the inference results, cell image groups with a high confidence level are selected, and learning is performed using the data with teacher labels among the selected cell image groups.

[0056] Therefore, it is possible to provide a learning method and a learning apparatus for a machine learning model capable of further performing appropriate learning for predicting characteristics after cell culture.

[0057] The learning apparatus 300 according to Example 2 of the present invention can operate separately from the learning apparatus 100 according to Example 1. Also, the learning apparatus 300 according to Example 2 of the present invention can operate together with the learning apparatus 100 according to Example 1.

[0058] (Example 3) Next, Example 3 of the present invention will be described with reference to FIG. 4.

[0059] FIG. 4 is a block diagram showing a cell characteristic prediction apparatus 400 according to Example 3.

[0060] In FIG. 4, the prediction apparatus 400 includes a machine learning model 401, a storage 402, an arithmetic processing unit 403 such as a CPU or a GPU, a memory 404, and an internal bus 405.

[0061] Also, the prediction apparatus 400 can perform input / output of data using an external display device 406 and an input device 407, and can execute processes such as prediction of cell characteristics. The machine learning model 401 is a machine learning model learned using the learning method according to Example 1 or Example 2 of the present invention. Therefore, a detailed description of the machine learning model 401 will be omitted.

[0062] The procedure for predicting cell characteristics is shown below.

[0063] First, store a group of cell images during or after cell culture in the storage 402. Next, use the machine learning model 401, the arithmetic processing unit 403, and the memory 404 to infer the group of cell images, and store the inference result in the storage 402. The arithmetic processing unit 403 refers to the inference result stored in the storage 402 and predicts the characteristic information of the group of cell images. Alternatively, an operator (not shown) uses the input device 407 and the display device 406 to refer to the inference result stored in the storage 402 and predicts the characteristic information of the group of cell images.

[0064] For example, if the group of cell images consists of 10 images, and 8 cell images are inferred to have the characteristic information "large" and 2 cell images are inferred to have the characteristic information "small", the characteristic information after culture in the cell culture experiment corresponding to this group of cell images is predicted to be "large".

[0065] The prediction device 400 according to Example 3 includes a storage 402 that stores a plurality of cell images taken during or after cell culture in a cell culture experiment and a plurality of characteristic information of the inferred cell images, a machine learning model 401 that learns the plurality of cell images stored in the storage 402 as training data, an arithmetic processing unit 403 that infers the plurality of characteristic information of the plurality of cell images stored in the storage 402, and a memory 404 for the arithmetic processing unit 403 to perform operations such as arithmetic operations. The arithmetic processing unit 403 predicts the characteristic information of the cells after the cell culture experiment based on the plurality of characteristic information stored in the storage 402 or the memory 404.

[0066] According to Example 3, it is possible to provide a prediction device including a machine learning model capable of performing appropriate learning for predicting characteristics after cell culture.

[0067] (Example 4) Next, Example 4 of the present invention will be described with reference to FIG. 5.

[0068] FIG. 5 is an explanatory diagram of a cell culture apparatus 501 according to Example 4. In FIG. 5, the cell culture apparatus 501 includes a prediction apparatus 502 having the same configuration as the prediction apparatus 400 according to Example 3, and a cell culture unit 501A. The cell culture unit 501A and the prediction apparatus 502 transmit and receive information to and from each other. The prediction apparatus 502 includes a machine learning model capable of performing appropriate learning in the same manner as the prediction apparatus 400.

[0069] The cell culture unit 501A acquires a group of cell images during culture with a camera (not shown) and transmits them to the prediction apparatus 502. The prediction apparatus 502 predicts the characteristic information after culture.

[0070] If the prediction result of the prediction apparatus 502 is good, the operator continues the cell culture in the cell culture unit 501A. When the prediction result of the prediction apparatus 502 is not good, the operator stops the cell culture in the cell culture unit 501A or takes measures to make the culture good.

[0071] According to Example 4, it is possible to provide a cell culture apparatus including a prediction apparatus 502 capable of performing appropriate prediction regarding the prediction of characteristics after cell culture, and capable of determining whether the cell culture is good or not.

[0072] (Example 5) Next, Example 5 of the present invention will be described with reference to FIG. 6.

[0073] FIG. 6 is an explanatory diagram of a cell culture apparatus 600 according to Example 5. In FIG. 6, the cell culture apparatus 600 includes the learning apparatus 100 of Example 1 and the cell culture unit 501A of Example 4. The cell culture unit 501A and the learning apparatus 100 transmit and receive information to and from each other.

[0074] According to Example 5, it is possible to provide a cell culture apparatus having a learning apparatus for a machine learning model capable of performing more appropriate learning regarding the prediction of characteristics after cell culture.

[0075] In the above-described Examples 1 to 5, an example of classifying cell images into three classes was shown. However, the present invention is also applicable to the case of classifying into four or more classes. For example, when classifying into four classes, they are classified into large, first medium, second medium, and small. And in the learning stage, for example, it is preferable to learn using cell images of large and small classes as training data. In another example, the class of "large and first medium" may be regarded as a single group, and learning may be performed using cell images of the single-group class and cell images of the small class. Or learning may be performed using cell images of large and the class of "second medium and small".

[0076] In other words, "learning using the above cell images of large and small classes as training data" means that "among the plurality of classified classes, the class having the largest characteristic information or a continuous group of classes including the largest class is regarded as the 'large class', and the class having the smallest characteristic information or a continuous group of classes including the smallest class among the plurality of classified classes is regarded as the'small class'. After that, in a situation where there is at least one or more classes that do not belong to either the large or small class groups between the class group constituting the 'large class' and the class group constituting the'small class', it can be said that it is 'learning using the above cell images of the 'large class' and'small class' regarded as such as training data'.

[0077] When classifying into five or more classes, they are classified into large, first medium, second medium, third medium, and small, and it is preferable to learn using cell images of large and small classes, or cell images of "large and first medium" and small classes, or cell images of "large, first medium, and second medium" and small classes, or cell images of "large and first medium" and the class of "third medium and small", or cell images of large and the class of "second medium, third medium, and small", or cell images of large and the class of "third medium and small" as training data.

Explanation of Reference Signs

[0078] 100, 300... learning devices, 101, 105, 309... data storage units, 104... data selection unit, 107, 307, 401... machine learning models, 305, 315... data storage units, 308... data selection unit, 400, 502... prediction devices, 402... storage, 403... arithmetic processing unit, 404... memory, 405... internal bus, 406... display device, 407... input device, 501, 600... cell culture devices, 501A... cell culture section, 1010, 1020, 1030... acquired information in cell culture experiments, 1011, 1021, 1031, 1051, 1052, 1061, 1062, 3051, 3052, 3061, 3062, 3151, 3152, 3161, 3162... cell image groups, 1012, 1022, 1032... characteristic information, 1050, 1060, 3050, 3060, 3150, 3160... machine learning data, 1053, 1063, 3053, 3063... labels corresponding to the characteristic information

Claims

1. A first step of taking a plurality of cell images during or after culturing cells, and after culturing the cells, performing at least three cell culture experiments to obtain information indicating the magnitude relationship of the characteristics of the cells in the plurality of cell images; A second step of using part or all of the pairs obtained by combining the plurality of captured cell images and the information indicating the magnitude relationship as training data; comprising; The second step is as follows: classifying the information indicating the magnitude relationships of the plurality of cell images into at least three classes of large, medium, and small; A method for training a machine learning model, characterized by learning the cell images of the large and small classes as training data.

2. In the method for training a machine learning model according to Claim 1, The second step is as follows: determining specific thresholds A and B (A > B), when the maximum value of the information indicating the magnitude relationship is C and the minimum value is D, the class of all the cell images from the information indicating the magnitude relationship being equal to or greater than the threshold A to the maximum value C, and the class of all the cell images from the information indicating the magnitude relationship being equal to or less than the threshold B to the minimum value D are classified into two classes, and learning the cell images of the two classes as training data. A method for training a machine learning model, characterized by this.

3. In the method for training a machine learning model according to Claim 1, A method for training a machine learning model, characterized by selecting the cell images with high confidence from the training data and storing them in a data storage unit.

4. A first data storage unit for storing a plurality of cell images taken during or after culturing cells and information indicating the magnitude relationship of the characteristics of the cells after culturing these plurality of cell images; A data selection unit that classifies the information indicating the plurality of magnitude relationships stored in the first data storage unit into at least three classes of large, medium, and small, and selects the cell images of the large and small classes; A second data storage unit that stores the cell images of the large and small classes selected by the data selection unit together with the information indicating the magnitude relationship; A machine learning model that learns the cell images of the large and small classes stored in the second data storage unit as training data; A learning device, characterized by comprising.

5. In the learning device according to Claim 4, The data selection unit determines specific thresholds A and B (A > B), sets the maximum value of the information indicating the magnitude relationship as C, and the minimum value as D. When the information indicating the magnitude relationship is within the range from the threshold A or higher to the maximum value C for all the cell images, and within the range from the threshold B or lower to the minimum value D for all the cell images, it classifies the cell images into two classes, and uses the cell images of these two classes as the cell images of the large and small classes. A learning device characterized by this is provided.

6. A third data storage unit that classifies information indicating the magnitude relationship of cells after culturing of a plurality of cell images taken during or after culturing of cells and the cells of these plurality of cell images into at least three classes of large, medium, and small, and stores the cell images of the large and small classes together with the information indicating the magnitude relationship; A machine learning model that learns, using the cell images of the large and small classes stored in the third data storage unit as training data, and selects cell images with high confidence; A fifth data storage unit that stores the confidence levels of each of the plurality of cell images obtained by the machine learning model; A data selection unit that selects the cell images with high confidence levels stored in the fifth data storage unit; A fourth storage unit that stores the cell images selected by the data selection unit together with the information indicating the magnitude relationship; A learning device comprising the above, wherein the machine learning model further learns using the cell images stored in the fourth storage unit as training data.

7. A storage that stores a plurality of cell images taken during or after culturing of cells in a cell culture experiment, and information indicating the magnitude relationship of the characteristics of the cells after culturing of the plurality of inferred cell images; A machine learning model that classifies information indicating the magnitude relationship of the characteristics of the cells after culturing of the plurality of cell images into at least three classes of large, medium, and small, and learns using the cell images of the large and small classes as training data, and the machine learning model that learns using the plurality of cell images stored in the storage as training data; A memory for performing operations such as calculations; An arithmetic processing unit that predicts the information indicating the magnitude relationship after the cell culture experiment based on the information indicating the magnitude relationship stored in the storage or the memory; A prediction device characterized by comprising the above.

8. A cell culture device comprising the learning device according to claim 4 and a cell culture unit.

9. A cell culture device comprising the prediction device according to claim 7 and a cell culture unit.

10. In the method for learning a machine learning model according to claim 1, The information indicating the magnitude relationship of the characteristics of the cells after the culture is the differentiation induction efficiency. A method for learning a machine learning model characterized by this.

11. In the learning device according to claim 4 or 6, The information indicating the magnitude relationship of the characteristics of the cells after the culture is the differentiation induction efficiency. A learning device for a machine learning model characterized by this.

12. In the prediction device according to claim 7, The information indicating the magnitude relationship of the characteristics of the cells after the culture is the differentiation induction efficiency. A prediction device characterized by this.

13. In the method for learning a machine learning model according to claim 1, The information indicating the magnitude relationship of the characteristics of the cells after the culture is the cell survival rate. A method for learning a machine learning model characterized by this.

14. In the learning device according to claim 4 or 6, The information indicating the magnitude relationship of the characteristics of the cells after the culture is the cell survival rate. A learning device for a machine learning model characterized by this.

15. In the prediction device according to claim 7, The information indicating the magnitude relationship of the characteristics of the cells after the culture is the cell survival rate. A prediction device characterized by this.

Images