Model generation program, model generation method, cell sorting device, and trained model

Abstract

To efficiently generate a model that can select desired cells from a plurality of cells included in a cell image.SOLUTION: A model generation program that causes a computer to generate a model for selecting desired cells from a cell image containing multiple cells, includes: a first step of cutting out a partial region containing the desired cells from the cell image; a second step of generating a partial region pasted image so as to partially include the partial region; a third step of annotating the desired cells included in the partial region; and a fourth step of generating the model through machine learning using the annotated partial region pasted image as training data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a model generation program, a model generation method, a cell sorting device, and a trained model. [Background technology]

[0002] To efficiently cultivate cells for use in regenerative medicine and cell-based medicines, it is desirable to select high-quality cells with high reproducibility. Visual selection by humans has low reproducibility because the results vary depending on the person and their level of experience.

[0003] A method for selecting high-quality cells using cell images has been devised. A method for selecting desired cells from multiple cells contained in a cell image has been devised, using a model trained by machine learning using annotated cell images as training data (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 6941123 Summary of the Invention [Problem to be solved by the invention]

[0005] In general, it is considered desirable for training data in machine learning to be diverse, so it is desirable to prepare many annotated cell images for model training. However, the task of preparing training data, which involves preparing a variety of cell images and annotating the cells that are most frequently found in the cell images, is time-consuming.

[0006] In light of the above circumstances, the present invention aims to provide a model generation program, a model generation method, a cell sorting device equipped with said model, and said model that can efficiently generate a model that can select a desired cell from multiple cells contained in a cell image. [Means for solving the problem]

[0007] In order to solve the above problems, the present invention proposes the following means. A model generation program according to a first aspect of the present invention is a program for causing a computer to generate a model for selecting a desired cell from a cell image containing a plurality of cells, and includes a first step of cutting out a partial region containing the cell from the cell image, a second step of generating a partial region pasted image in which the partial region is pasted so as to include a part of the partial region, a third step of annotating the cells contained in the partial region, and a fourth step of generating the model by machine learning using the partial region pasted image with the annotations added as training data.

[0008] A model generation method according to a second aspect of the present invention is a method for generating a model for selecting a desired cell from a cell image containing a plurality of cells, and includes the following steps: a first step of cutting out a partial region containing the cell from the cell image; a second step of generating a partial region pasted image that includes the partial region; a third step of annotating the cells contained in the partial region; and a fourth step of generating the model by machine learning using the annotated partial region pasted image as training data.

[0009] A cell sorting device according to a third aspect of the present invention comprises an image acquisition unit that acquires a cell image containing a plurality of cells from a microscope, a sorting unit having a model that selects desired cells from the cell image, and an output unit that outputs information related to the selected cells, wherein the model is an inference model generated by machine learning using training data to which annotations have been added for the cells contained in the processed image, and the processed image is a partial area pasted image that includes a partial area cut out from the cell image and containing the cells.

[0010] A trained model according to a fourth aspect of the present invention is a trained model for selecting a desired cell from a cell image containing multiple cells, and is generated by machine learning using training data to which annotations have been added for the cells contained in a processed image, and the processed image is a partial area pasted image in which a partial area containing the cell cut out from the cell image is pasted so as to include a part of the processed image. [Effects of the Invention]

[0011] According to the model generation program, model generation method, cell sorting device, and trained model of the present invention, a model capable of sorting a desired cell from a plurality of cells contained in a cell image can be efficiently generated. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram showing a cell sorting system according to a first embodiment. [Figure 2] FIG. 2 is a diagram illustrating a hardware configuration of the cell selection device. [Figure 3] 1 is a flowchart illustrating a model generation method. [Figure 4] FIG. 2 is a diagram illustrating the first and second steps in the model generation method. [Figure 5] FIG. 10 is a diagram illustrating a third step in the model generation method. DETAILED DESCRIPTION OF THE INVENTION

[0013] (First embodiment) A first embodiment of the present invention will be described with reference to FIGS.

[0014] [Cell Sorting System 400] FIG. 1 is a diagram showing a cell sorting system 400 according to this embodiment. The cell sorting system 400 is a system capable of sorting a desired cell from a plurality of cells. The cell sorting system 400 includes a cell sorter 100, a microscope 200, and a display device 300.

[0015] The microscope 200 observes an observation object such as a cell placed on a sample stage. The observed microscope image can be output to the cell sorting device 100 via wired or wireless communication. The microscope 200 has a function of magnifying the observation object, for example, by 100 times or more.

[0016] The display device 300 is a display device such as an LCD display, etc. The display device 300 can display image data output by the cell selection device 100.

[0017] [Cell sorting device 100] FIG. 2 is a diagram showing the hardware configuration of the cell selection apparatus 100. As shown in FIG. The cell sorting device 100 is a device that sorts desired cells from a cell image P containing multiple cells. The cell sorting device 100 is a programmable computer that includes a processor 11 such as a CPU, a memory 12, a recording unit 13, an input / output control unit 14, and the like. The functions of the cell sorting device 100 are realized by the processor 11 executing a program (such as a cell sorting program). At least some of the functions of the cell sorting device 100 may be realized by a dedicated logic circuit implemented in an ASIC or FPGA. The cell sorting device 100 includes an acquisition unit 1, a sorting unit 2, and an output unit 3.

[0018] The acquisition unit 1 acquires a cell image P containing a plurality of cells from the microscope 200. The image format of the cell image P is, for example, JPEG or TIFF. If the microscope 200 is capable of outputting still images or videos of the captured cells, the acquisition unit 1 acquires the cell image P from the output of the microscope 200. If the microscope 200 is not capable of outputting still images or videos of the captured cells, the acquisition unit 1 acquires the cell image P by capturing an observation image of the microscope 200.

[0019] When the microscope 200 is located at a location distant from the cell selection device 100, the acquisition unit 1 may acquire the cell image P via a recording medium such as a memory.

[0020] The output unit 3 outputs information about the cells selected by the selection unit 2 to a display device, etc. For example, the output unit 3 outputs to the display device 300 an image in which a marking image (such as a frame image or an arrow image) indicating the selected cells is superimposed on the cell image P.

[0021] The sorting unit 2 sorts desired cells from the cell image P. The sorting unit 2 has a model M that sorts desired cells from the cell image P. The model M is stored in the memory 12 or recording unit 13 that the cell sorting device 100 includes.

[0022] The cell sorting device 100 may further include components other than the processor 11, the memory 12, and the recording unit 13. For example, the cell sorting device 100 may further include an image calculation unit that performs part or all of the image processing and image recognition processing. By including the image calculation unit, the cell sorting device 100 can perform specific image processing and image recognition processing at high speed. The image calculation unit may be a calculator provided in a cloud server connected via the Internet.

[0023] The recording unit 13 is a non-volatile recording medium that stores the above-mentioned programs and data necessary for executing the programs. The recording unit 13 is configured, for example, with a writable non-volatile memory such as a ROM or flash memory, a portable medium such as a CD-ROM, or a storage device such as a hard disk or SSD built into a computer system. The recording unit 13 may also be a storage device provided in a cloud server connected via the Internet.

[0024] The above program may be provided by a "computer-readable recording medium" such as a flash memory. The program may be transmitted from a computer that stores the program to memory 12 or recording unit 13 via a transmission medium or by transmission waves in the transmission medium. A "transmission medium" that transmits the program is a medium that has the function of transmitting information. Media that have the function of transmitting information include networks (communication networks) such as the Internet and communication lines (communication lines) such as telephone lines. The above program may realize some of the functions described above. Furthermore, the above program may be a difference file (difference program). The functions described above may be realized by combining a program already recorded on the computer with a difference program.

[0025] [Model M] The model (trained model) M is an inference model generated by machine learning using training data to which annotations are added for cells included in the processed image. Here, the processed image is a partial region-pasted image Q, which will be described later. The model M is a machine learning model such as a neural network. The model M is not limited to a neural network, and may be any machine learning model that uses an image as input.

[0026] FIG. 3 is a flowchart showing a method for generating the model M. The model M is generated through a first step, a second step, a third step, a fourth step, and a fifth step. The model M may be generated in the cell sorting device 100 or in another device.

[0027] <First step> FIG. 4 is a diagram showing the first step and the second step. In the first step, a partial region R containing cells is cut out from the learning cell image PL, which is the cell image P used for learning. The partial region R is a region containing at least one cell to which annotation can be added. When cutting out multiple partial regions R in the first step, it is desirable to select and cut out regions with sparse cell populations and regions with dense cell populations in a balanced and equal manner. This is to avoid over-learning in the fourth step.

[0028] In the first step, partial regions R may be cut out from two or more different learning cell images PL. In this case, the multiple learning cell images PL are images observed under microscope 200 at the same magnification.

[0029] <Second process> The second step is to generate a partial area pasted image Q that is pasted so as to include the partial area R as a part of it. The partial area pasted image Q is an image that contains cells only in the area where the partial area R is arranged. Furthermore, the image size of the partial area pasted image Q is the same as the image size of the learning cell image PL. Therefore, the size of the partial area R relative to the partial area pasted image Q is the same as the size of the partial area R relative to the learning cell image PL.

[0030] The image size of the partial region pasted image Q does not have to be exactly the same as the image size of the learning cell image PL, and may differ by a few pixels. In other words, it is sufficient that the size of the partial region R in the partial region pasted image Q is approximately the same as the size of the partial region R in the learning cell image PL.

[0031] In the partial region pasted image Q, the area other than the partial region R (background area) is solid. The background color of the background area may be extracted from pixels selected from the partial region pasted image Q, or the average value of the color of the partial region pasted image Q may be assigned, or an RGB value may be specified. However, the area other than the partial region R in the partial region pasted image Q is not limited to being solid, and may include, for example, objects other than cells or noise. In other words, the area other than the partial region R in the partial region pasted image Q should not include anything that will affect the learning of cells.

[0032] In the second step, multiple partial region pasted images Q may be generated in which the same partial region R is arranged in different positions, as shown in Fig. 4. The image size of each partial region pasted image Q is the same as the image size of the learning cell image PL. In the second step, it is desirable to arrange the same partial region R in a random position determined using a random number or the like.

[0033] <Third step> FIG. 5 is a diagram showing the third step. The third step is to generate training data T in which annotations are added to cells included in the partial region R in the partial region pasted image Q. Annotations are, for example, labeling of cells in an area surrounded by a bounding box, and are information that serves as a reference when selecting desired cells. A bounding box is a means of identifying cells to be annotated by enclosing the cells included in the partial region R. The means of identifying cells to be annotated is not limited to a bounding box, and may be an arrow image, for example.

[0034] <Fourth process> In the fourth step, a model M is generated by machine learning using the training data T (annotated partial region-pasted image Q). Specifically, in the fourth step, a known machine learning algorithm is used to adjust the internal parameters of the model M so that the model M can output appropriate annotations for the input partial region-pasted image Q.

[0035] <Fifth process> The fifth step is to select desired cells using a trained model M for the cell image for inference PI, which is the cell image P used for inference. The fifth step is an inference test to confirm whether the reliability of the inference of the trained model M is ensured. The data used in the inference test is a combination of the cell image for inference PI and annotations.

[0036] The learning cell image PL and the inference cell image PI are images observed at the same magnification under the microscope 200. The image size of the inference cell image PI is the same as the image size of the learning cell image PL. If the reliability of the inference of the trained model M is not ensured, the fourth step and so on are performed again.

[0037] The learning cell image PL and the inference cell image PI may be images acquired from different types of microscopes 200. When the learning cell image PL and the inference cell image PI are images acquired from different types of microscopes 200, it is desirable that the learning cell image PL and the inference cell image PI are images observed at approximately the same magnification, and that the image size of the inference cell image PI is approximately the same as the image size of the learning cell image PL.

[0038] Next, the operation of the cell sorting system 400 will be described. The acquisition unit 1 acquires an inference-use cell image PI. The inference-use cell image PI is not for use in an inference test, and is an image to which no annotations have been added.

[0039] Next, the selection unit 2 selects desired cells from the cell images for inference PI using the trained model M. Specifically, the selection unit 2 selects cells for which the annotations estimated by the model M for the input cell images for inference PI include specific characteristics as desired cells.

[0040] Next, the output unit 3 outputs information about the desired cells selected by the selection unit 2 to the display device 300. For example, the output unit 3 outputs to the display device 300 an image in which a marking image (such as a frame image or an arrow image) indicating the selected cells is superimposed on the inference-use cell image PI.

[0041] According to the cell sorting system 400 and model generation method of this embodiment, it is possible to efficiently generate a model M that can select a desired cell from multiple cells included in an inference-use cell image PI. If the learning cell image PL instead of the partial region pasted image Q is used to train the model M, the model M will be trained assuming that unannotated cells are not targets for detection, which requires annotating many of the cells included in the learning cell image PL, which takes time and effort. On the other hand, if the partial region pasted image Q instead of the learning cell image PL is used to train the model M, the cells that need to be annotated are limited, thereby reducing the time and effort required for annotation.

[0042] In a situation where cells to be used in regenerative medicine or cell-based pharmaceuticals are selected from a plurality of cells, a training cell image PL, which is an observation of candidate cells to be selected, contains a large number of cells. By subdividing and cutting out partial regions R from the training cell image PL, multiple pieces of training data T can be obtained from a single training cell image PL. Therefore, with the model generation method according to this embodiment, sufficient training data T can be prepared by simply preparing a few training cell images PL, and a model M can be generated efficiently.

[0043] A typical trained model for object detection determines the relative position and relative size of a target object within a detection target image, so the training data used to train the model must include targets with different relative sizes and positions. Meanwhile, in situations where cells to be used in regenerative medicine or cell therapy are selected from multiple cells, cell images P observed with the same (similar) microscope 200 are used for both training and inference, so the relative sizes of the target cells are often the same. Therefore, the model generation method of this embodiment omits cells with different relative sizes from the training data T used to train the model. The model generation method of this embodiment can efficiently generate a model M that can be used in the above-described situations by omitting cells with different relative sizes from the training data T used to train the model and limiting it to cells with different relative positions.

[0044] The cell sorting system 400 and model generation method of this embodiment allow researchers and other users to easily generate training data T with flexible annotations, and efficiently generate a model M capable of selecting desired cells that meets the user's needs. If model M can efficiently select the user's desired cells, a robot or other device can automatically collect and culture the selected cells, eliminating the need for manual cell identification and selection. Furthermore, if model M can efficiently select the user's desired cells, the selected cells can be automatically held and injected or extracted with test solutions, eliminating the need for manual cell identification and selection. Automating time-consuming tasks that previously required precise manual labor can be useful for improving culture efficiency and lab automation. Furthermore, automation eliminates cell contamination, which previously occurred when humans were involved in culture, and is expected to improve the success rate of culture and reduce equipment costs such as clean rooms and air showers. [Example]

[0045] The present invention will be described in detail below based on examples, but the technical scope of the present invention is not limited to these examples.

[0046] As an example, 21 partial region-pasted images Q were generated from each of two different learning cell images PL, resulting in a total of 42 partial region-pasted images Q. When generating the partial region-pasted images Q from the learning cell images PL, the partial regions R were placed at random positions determined using random numbers. Annotations were added to the 42 partial region-pasted images Q to create training data T. The annotation criterion was a reference value that evaluated the degree to which "cells are separated, their outlines are roughly closed, and they are not angular." A model M was generated using the training data T generated in this way.

[0047] When an inference test was conducted using inference cell images PI, which were different from the learning cell images PL, the model M of the cell sorting device 100 was able to select cells that conformed to the annotation criteria with an accuracy of 81%. Conventional methods cannot even perform learning with only two or so learning cell images PL due to a lack of training data. However, the model generation method of the present invention can efficiently generate a model M capable of selection with an accuracy sufficient to meet the needs of, for example, a site that handles cells in the field of regenerative medicine, such as "I want to extract 10 to 20 separated cells."

[0048] Although the first embodiment of the present invention has been described above in detail with reference to the drawings, the specific configuration is not limited to this embodiment and includes design modifications within the scope of the present invention. Furthermore, the components shown in the above embodiment and modifications can be configured by appropriately combining them. [Industrial Applicability]

[0049] The present invention can be used in a method for selecting desired cells and a system for selecting desired cells. In addition to selecting desired cells, the present invention can also be used in a system for sorting food or processed products on a factory lane or a system for sorting vegetables or fruits from fields, etc.

[0050] The present invention includes the following technical ideas. (Additional note 1) A program that causes a computer to generate a model for selecting a desired cell from a cell image containing a plurality of cells, a first step of cutting out a partial region including the cell from the cell image; a second step of generating a partial area pasted image pasted so as to include the partial area; a third step of annotating the cells included in the partial region; a fourth step of generating the model by machine learning using the partial region-pasted image to which the annotations have been added as training data; Equipped with Model generator. (Additional note 2) the partial region pasted image includes the cells only in the region where the partial region is arranged; Item 1. A model generation program according to item 1. (Additional note 3) The partial area pasted image has a plain area other than the partial area. Item 1. A model generation program according to item 1. (Additional note 4) The size of the partial region relative to the partial region pasted image is the same as the size of the partial region relative to the cell image; Item 3. A model generation program according to any one of items 1 to 3. (Additional note 5) The image size of the partial region pasted image is the same as the image size of the cell image. Item 3. A model generation program according to any one of items 1 to 3. (Additional note 6) the first step comprises cutting out the partial regions from each of the plurality of cell images; The plurality of cell images are images observed at the same magnification. 6. A model generation program according to any one of claims 1 to 5. (Additional note 7) the second step generates a plurality of partial area pasted images in which the same partial area is arranged at different positions; 7. A model generation program according to any one of claims 1 to 6. (Additional note 8) a fifth step of selecting the desired cell using the model for the cell image for inference, which is the cell image used for inference; The learning cell image, which is the cell image used for learning, and the inference cell image are images observed at the same magnification. 8. A model generation program according to any one of claims 1 to 7. (Additional note 9) The image size of the cell image for inference is the same as the image size of the cell image for learning. Item 8. A model generation program according to item 8. (Additional note 10) A method for generating a model for selecting a desired cell from a cell image containing a plurality of cells, comprising: a first step of cutting out a partial region including the cell from the cell image; a second step of generating a partial area pasted image pasted so as to include the partial area; a third step of annotating the cells included in the partial region; a fourth step of generating the model by machine learning using the partial region-pasted image to which the annotations have been added as training data; Equipped with Model generation method. (Additional note 11) the partial region pasted image includes the cells only in the region where the partial region is arranged; Item 11. A model generation method according to item 10. (Additional note 12) The partial area pasted image has a plain area other than the partial area. Item 11. A model generation method according to item 10. (Additional note 13) The size of the partial region relative to the partial region pasted image is the same as the size of the partial region relative to the cell image; 13. The model generation method according to any one of claims 10 to 12. (Additional note 14) The image size of the partial region pasted image is the same as the image size of the cell image. 13. The model generation method according to any one of claims 10 to 12. (Additional note 15) the first step comprises cutting out the partial regions from each of the plurality of cell images; The plurality of cell images are images observed at the same magnification. 15. The model generation method according to any one of claims 10 to 14. (Additional note 16) the second step generates a plurality of partial area pasted images in which the same partial area is arranged at different positions; 16. A model generation method according to any one of claims 10 to 15. (Additional note 17) a fifth step of selecting the desired cell using the model for the cell image for inference, which is the cell image used for inference; The learning cell image, which is the cell image used for learning, and the inference cell image are images observed at the same magnification. 17. A model generation method according to any one of claims 10 to 16. (Additional note 18) The image size of the cell image for inference is the same as the image size of the cell image for learning. Item 18. A model generation method according to item 17. (Additional note 19) an image acquisition unit that acquires a cell image including a plurality of cells from a microscope; a sorting unit having a model for sorting desired cells from the cell image; an output unit that outputs information about the selected cells; Equipped with the model is an inference model generated by machine learning using training data to which annotations are added for the cells included in the processed image; The processed image is a partial area pasted image pasted so as to include a partial area including the cell cut out from the cell image. Cell sorting device. (Additional note 20) the partial region pasted image includes the cells only in the region where the partial region is arranged; 20. The cell sorting device according to claim 19. (Additional note 21) The partial area pasted image has a plain area other than the partial area. 20. The cell sorting device according to claim 19. (Additional note 22) The size of the partial region relative to the partial region pasted image is the same as the size of the partial region relative to the cell image; 22. The cell sorting device according to any one of claims 19 to 21. (Additional note 23) The image size of the partial region pasted image is the same as the image size of the cell image. 22. The cell sorting device according to any one of claims 19 to 21. (Additional note 24) the selection unit selects the desired cells from the cell image for inference, which is the cell image used for inference, using the model; The learning cell image, which is the cell image used for learning, and the inference cell image are images observed at the same magnification. 24. The cell sorting device according to any one of claims 19 to 23. (Additional note 25) The image size of the cell image for inference is the same as the image size of the cell image for learning. Item 25. The cell sorting device according to claim 24. (Additional note 26) A trained model for selecting a desired cell from a cell image containing a plurality of cells, The processed image is generated by machine learning using training data to which annotations have been added for the cells included in the processed image, The processed image is a partial area pasted image pasted so as to include a partial area including the cell cut out from the cell image. Trained model. (Additional note 27) the partial region pasted image includes the cells only in the region where the partial region is arranged; The trained model described in Appendix 26. (Additional note 28) The partial area pasted image has a plain area other than the partial area. The trained model described in Appendix 26. (Additional note 29) The size of the partial region relative to the partial region pasted image is the same as the size of the partial region relative to the cell image; The trained model according to any one of supplementary items 26 to 28. (Additional note 30) The image size of the partial region pasted image is the same as the image size of the cell image. The trained model described in Appendix 26. [Explanation of symbols]

[0051] 400 Cell Sorting System 300 display device 200 Microscope 100 Cell sorter 1 Acquisition part 2. Sorting Department 3 Output section M model (pre-trained model) P cell image Cell images for PI inference PL learning cell images Q Partial region pasted image R subregion T Teacher data

Claims

1. A program that causes a computer to generate a model for selecting a desired cell from a cell image containing a plurality of cells, a first step of cutting out a partial region including the cell from the cell image; a second step of generating a partial area pasted image pasted so as to include the partial area; a third step of annotating the cells included in the partial region; a fourth step of generating the model by machine learning using the partial region-pasted image to which the annotations have been added as training data; Equipped with Model generator.

2. the partial region pasted image includes the cells only in the region where the partial region is arranged; The model generation program according to claim 1 .

3. The partial area pasted image has a plain area other than the partial area. The model generation program according to claim 1 .

4. The size of the partial region relative to the partial region pasted image is the same as the size of the partial region relative to the cell image; The model generation program according to claim 1 .

5. The image size of the partial region pasted image is the same as the image size of the cell image. The model generation program according to claim 1 .

6. the first step comprises cutting out the partial regions from each of the plurality of cell images; The plurality of cell images are images observed at the same magnification. The model generation program according to any one of claims 1 to 5.

7. the second step generates a plurality of partial area pasted images in which the same partial area is arranged at different positions; The model generation program according to any one of claims 1 to 5.

8. a fifth step of selecting the desired cell using the model for the cell image for inference, which is the cell image used for inference; The learning cell image, which is the cell image used for learning, and the inference cell image are images observed at the same magnification. The model generation program according to any one of claims 1 to 5.

9. The image size of the cell image for inference is the same as the image size of the cell image for learning. The model generation program according to claim 8.

10. A method for generating a model for selecting a desired cell from a cell image containing a plurality of cells, comprising: a first step of cutting out a partial region including the cell from the cell image; a second step of generating a partial area pasted image pasted so as to include the partial area; a third step of annotating the cells included in the partial region; a fourth step of generating the model by machine learning using the partial region-pasted image to which the annotations have been added as training data; Equipped with Model generation method.

11. an image acquisition unit that acquires a cell image including a plurality of cells from a microscope; a sorting unit having a model for sorting desired cells from the cell image; an output unit that outputs information about the selected cells; Equipped with the model is an inference model generated by machine learning using training data to which annotations are added for the cells included in the processed image; The processed image is a partial area pasted image pasted so as to include a partial area including the cell cut out from the cell image. Cell sorting device.

12. the partial region pasted image includes the cells only in the region where the partial region is arranged; The cell sorting device according to claim 11.

13. The partial area pasted image has a plain area other than the partial area. The cell sorting device according to claim 11.

14. The size of the partial region relative to the partial region pasted image is the same as the size of the partial region relative to the cell image; The cell sorting device according to claim 11.

15. The image size of the partial region pasted image is the same as the image size of the cell image. The cell sorting device according to claim 11.

16. the selection unit selects the desired cells from the cell image for inference, which is the cell image used for inference, using the model; The learning cell image, which is the cell image used for learning, and the inference cell image are images observed at the same magnification. The cell sorting device according to any one of claims 11 to 15.

17. The image size of the cell image for inference is the same as the image size of the cell image for learning. The cell sorting device according to claim 16.

18. A trained model for selecting a desired cell from a cell image containing a plurality of cells, The processed image is generated by machine learning using training data to which annotations have been added for the cells included in the processed image, The processed image is a partial area pasted image pasted so as to include a partial area including the cell cut out from the cell image. Trained model.

19. the partial region pasted image includes the cells only in the region where the partial region is arranged; The trained model of claim 18.

20. The partial area pasted image has a plain area other than the partial area. The trained model of claim 18.

21. The size of the partial region relative to the partial region pasted image is the same as the size of the partial region relative to the cell image; The trained model of claim 18.

22. The image size of the partial region pasted image is the same as the image size of the cell image. The trained model of claim 18.

Images