Cell analysis method, device, system, and program, and method, device, and program for generating a trained artificial intelligence algorithm

Abstract

The present application aims to provide a cell analysis method, a cell analysis device, a cell analysis system, and a cell analysis program that make it easy to analyze more cells contained in a subject with high precision and at high speed, and a method, a device, and a program for generating a trained artificial intelligence algorithm. The problem is solved by a cell analysis method for analyzing cells using an artificial intelligence algorithm, the cell analysis method including: causing a sample containing cells to flow through a flow path, capturing an analysis target image generated by cells within the flow path, generating analysis data from the generated analysis target image, inputting the generated analysis data to the artificial intelligence algorithm, and generating data showing the properties of the cells contained in the analysis target image by the artificial intelligence algorithm.

Description

TECHNICAL FIELD

[0001] The present application relates to a cell analysis method, a cell analysis device, a cell analysis system, and a cell analysis program for analyzing cells, and a method, a device, and a program for generating a trained artificial intelligence algorithm for analyzing cells. BACKGROUND

[0002] In Patent Literature 1, a method is disclosed in which a microscope image subjected to a filtering process is applied to a machine learning model trained to determine the center and the boundary of a cell of a specific type, and an image of the determined cell is output while counting the cell.

[0003] PRIOR ART DOCUMENTS

[0004] PATENT LITERATURE

[0005] [Patent Literature 1] International Publication No. 2015 / 065697

[0006] SUMMARY OF THE INVENTION

[0007] PROBLEMS TO BE SOLVED BY THE INVENTION

[0008] In the examination of a patient who has a possibility of having a tumor, in order to determine the presence or absence of a tumor, the effect of an anticancer therapy, the presence or absence of recurrence, etc., it is necessary to grasp the proportion in which abnormal cells such as cells having chromosomal abnormalities or peripheral circulating tumor cells exist in a subject containing a plurality of cells.

[0009] There are also cases in which the number of abnormal cells contained in a subject is extremely small compared to the number of normal cells that should originally exist in the subject. Therefore, in order to detect abnormal cells contained in a subject, it is necessary to analyze more cells. However, the method described in Patent Literature 1 increases the number of cells to be determined as the microscope image is used, and the time required for obtaining the microscope image increases.

[0010] The present application aims to provide a cell analysis method, a cell analysis device, a cell analysis system, and a cell analysis program for analyzing more cells contained in a subject with high accuracy and at high speed, and a method, a device, and a program for generating a trained artificial intelligence algorithm.

[0011] MEANS FOR SOLVING THE PROBLEMS

[0012] One embodiment of the present application relates to a cell analysis method of analyzing cells using an artificial intelligence algorithm (60, 63, 97). The cell analysis method causes a sample (10) containing cells to flow through a flow path (111), captures an analysis target image (80, 85, 95) of the cells passing through the flow path (111), generates analysis data (82, 87, 96) from the generated analysis target image (80, 85, 95), inputs the generated analysis data to the artificial intelligence algorithm (60, 63, 97), and generates data (84, 88, 98) showing a property of the cells contained in the analysis target image (80, 85, 95) by the artificial intelligence algorithm.

[0013] One embodiment of the present application relates to a cell analysis device (400A, 200B, 200C) of analyzing cells using an artificial intelligence algorithm (60, 63, 97). The cell analysis device (400A, 200B, 200C) includes a control unit (40A, 20B, 20C) configured to cause a sample (10) containing cells to flow through a flow path (111), input analysis data (82, 87, 96) generated from an analysis target image (80, 85, 95) of the cells passing through the flow path (111) to the artificial intelligence algorithm (60, 63, 97), and generate data (84, 88, 98) showing a property of the cells contained in the analysis target image (80, 85, 95) by the artificial intelligence algorithm (60, 63, 97).

[0014] One embodiment of the present application relates to a cell analysis system (1000, 2000, 3000). The cell analysis system (1000, 2000, 3000) includes a flow cell (110) through which a sample (10) containing cells flows, a light source (120, 121, 122, 123) configured to irradiate the sample (10) flowing through the flow cell (110) with light, an imaging unit (160) configured to capture cells in the sample (10) irradiated with the light, and a control unit (40A, 20B, 20C). The control unit (40A, 20B, 20C) is configured to acquire an analysis target image (80, 85, 95) of the cells passing through the flow path (111) captured by the imaging unit (160), generate analysis data (82, 87, 96) from the analysis target image (80, 85, 95), input the analysis data (82, 87, 96) to an artificial intelligence algorithm (60, 63, 97), and generate data (84, 88, 98) showing a property of the cells contained in the analysis target image (80, 85, 95) by the artificial intelligence algorithm (60, 63, 97).

[0015] One embodiment of the present application relates to a cell analysis program that analyzes cells. The cell analysis program causes a computer to execute processing including a step of causing a sample (10) containing cells to flow through a flow path (111), a step (S22) of inputting, to an artificial intelligence algorithm (60, 63, 97), analysis data (82, 87, 96) generated from an analysis target image (80, 85, 95) that captures cells passing through the flow path (111), and a step (S32) of generating, by the artificial intelligence algorithm (60, 63, 97), data (84, 88, 98) that shows a property of a cell contained in the analysis target image (80, 85, 95).

[0016] The cell analysis device (400A, 200B, 200C), the cell analysis system (1000, 2000, 3000), and the cell analysis program make it easy to perform high-precision and high-speed analysis on more cells contained in a subject.

[0017] One embodiment of the present application relates to a method of generating an artificial intelligence algorithm (60, 63, 97) for analyzing cells. The method includes causing a sample (10) containing cells to flow through a flow path (111), inputting, to the artificial intelligence algorithm (50, 53, 94), training data (73, 78, 92) generated from a training image (70, 75, 90) that captures cells passing through the flow path (111), and a label (74P, 74N, 79P, 79N, 93P, 93N) that shows a property of a cell contained in the training image (70, 75, 90), and training the artificial intelligence algorithm (50, 53, 94).

[0018] One embodiment of the present application relates to a generation device (200A, 200B, 200C) of an artificial intelligence algorithm (60, 63, 97) for analyzing cells. The generation device (200A, 200B, 200C) includes a control unit (20A, 20B, 20C) that causes a sample (10) containing cells to flow through a flow path (111), inputs, to an artificial intelligence algorithm (50, 53, 94), training data (73, 78, 92) generated from a training image (70, 75, 90) that captures cells passing through the flow path (111), and a label (74P, 74N, 79P, 79N, 93P, 93N) that shows a property of a cell contained in the training image (70, 75, 90), and trains the artificial intelligence algorithm (50, 53, 94).

[0019] An embodiment of the present application relates to a generation program of a trained artificial intelligence algorithm (60, 63, 97) for analyzing cells. The above generation program causes a computer to execute a process including a step (S12) of flowing a sample (10) containing cells through a flow path (111), inputting, to an artificial intelligence algorithm (50, 53, 94), training data (73, 78, 92) generated from a training image (70, 75, 90) that captures a cell passing through the above flow path (111) and a label (74P, 74N, 79P, 79N, 93P, 93N) that shows a property of a cell contained in the training image (70, 75, 90), and a step (S12) of training the artificial intelligence algorithm (50, 53, 94).

[0020] The generation method, the generation device (200A, 200B, 200C), and the generation program of the trained artificial intelligence algorithm (60, 63, 97) can generate an artificial intelligence algorithm (60, 63, 97) that makes it easy to perform high-precision and high-speed analysis on more cells contained in a subject.

[0021]

Effects of the Invention

[0022] It is made easy to perform high-precision and high-speed analysis on more cells contained in a subject.

[0023]

Brief Description of the Drawings

[0024]

【 Figure 1 A method of generating training data for training a first artificial intelligence algorithm 50 for analyzing chromosomal abnormalities is shown. (A) shows a method of generating positive training data. (B) shows a method of generating negative training data.

[0025]

【 Figure 2 A method of generating training data for training a first artificial intelligence algorithm 50 for analyzing chromosomal abnormalities is shown.

[0026]

【 Figure 3 A method of generating analysis data for analyzing chromosomal abnormalities and a method of analyzing cells by a trained first artificial intelligence algorithm 60 are shown.

[0027]

【 Figure 4 Images of staining patterns of PML-RARA chimera gene positive cells using an imaging flow cytometer are shown. (A) shows images of channel 2 on the left and images of channel 2 on the right. (B) is a different cell from (A), showing images of channel 2 on the left and images of channel 2 on the right.

[0028]

【 Figure 5 A pattern example of fluorescent labeling is shown.

[0029]

【Figure 6 ] The fluorescently labeled pattern is displayed.

[0030] Figure 7 ] The fluorescently labeled pattern is displayed.

[0031] Figure 8 ] A method of generating training data for training a first artificial intelligence algorithm 53 for resolving peripheral circulating tumor cells is displayed.

[0032] Figure 9 ] A method of generating training data for training a first artificial intelligence algorithm 53 for resolving peripheral circulating tumor cells is displayed. (A) A method of generating positive training data is displayed. (B) A method of generating negative training data is displayed.

[0033] Figure 10 ] A method of generating training data for training a first artificial intelligence algorithm 53 for resolving peripheral circulating tumor cells is displayed.

[0034] Figure 11 ] A method of generating resolution data for resolving peripheral circulating tumor cells and a method of resolving cells by a trained first artificial intelligence algorithm 63 is displayed.

[0035] Figure 12 ] (A) A method of generating training data for training a second artificial intelligence algorithm 94 for resolving peripheral circulating tumor cells is displayed. (B) A method of generating resolution data and a method of resolving cells by a second artificial intelligence algorithm 97 is displayed.

[0036] Figure 13 ] A feature quantity for training a second artificial intelligence algorithm 94 is displayed.

[0037] Figure 14 ] A definition of a feature quantity for training a second artificial intelligence algorithm 94 is displayed. (A) Height and width are indicated. (B) Long axis and short axis are indicated. (C) Length, maximum thickness, and minimum thickness are indicated. (D) Aspect ratio, stretch length, and shape ratio are indicated. (E) Leaf symmetry pattern is indicated.

[0038] Figure 15 ] A hardware configuration of a cell resolution system 1000 is displayed.

[0039] Figure 16 ] A hardware configuration of a training device 200A, 200B, 200C is displayed.

[0040] Figure 17 ] A functional block of a training device 200A is displayed.

[0041] Figure 18 ​​​​​​​​​​​​(A) shows a flowchart of a training process of the first artificial intelligence algorithm. (B) shows a flowchart of a training process of the second artificial intelligence algorithm.

[0042]

Figure 19

[0043]

Figure 20

[0044]

Figure 21

[0045]

Figure 22

[0046]

Figure 23

[0047]

Figure 24

[0048]

Figure 25

[0049]

Figure 26

[0050]

Figure 27

[0051]

Figure 28

[0052]

Figure 29

DETAILED DESCRIPTION

[0053] Hereinafter, the outline and embodiments of the present application will be described in detail with reference to the accompanying drawings. Also, in the following description and drawings, the same symbols will be used to represent the same or similar constituent elements, and thus the description of the same or similar constituent elements will be omitted at times.

[0054] [1. Cell analysis method]

[0055] [1. Summary of cell analysis method]

[0056] The present embodiment relates to a cell analysis method for analyzing cells using an artificial intelligence algorithm. In the cell analysis method, an analysis target image of a cell that is an analysis target is captured, and a sample containing the cell is caused to flow through a flow path, and the cell passing through the flow path is captured. Analysis data for inputting to the artificial intelligence algorithm is generated from the captured analysis target image. When the analysis data is input to the artificial intelligence algorithm, data showing the properties of the cell contained in the analysis target image is generated by the artificial intelligence algorithm. The analysis target image is preferably captured one by one for the cells passing through the flow path.

[0057] In the present embodiment, the sample can be a sample prepared from a test body collected from a subject. The test body can contain, for example, a blood test body such as peripheral blood, venous blood, arterial blood, and the like, a urine test body, a test body of a body fluid other than blood and urine. As the body fluid other than blood and urine, bone marrow, ascites, pleural fluid, spinal fluid, and the like can be included. Sometimes, the body fluid other than blood and urine is simply referred to as "body fluid". The blood is preferably peripheral blood. For example, the blood can be peripheral blood collected using an anticoagulant such as sodium or potassium salt of ethylenediaminetetraacetic acid, sodium heparin, or the like.

[0058] Preparation of the sample from the test body can be performed according to a known method. For example, a tester recovers nucleated cells by performing centrifugal separation or the like using a cell separation medium such as Ficoll on a blood test body collected from a subject. For the recovery of nucleated cells, red blood cells or the like can also be hemolyzed by using a hemolytic agent instead of recovering nucleated cells by centrifugal separation, and nucleated cells are left. The target site of the recovered nucleated cells is labeled, preferably fluorescently labeled, by at least one selected from the group consisting of Fluorescence in situ hybridization (FISH) method, immunostaining method, and organelle staining method, or the like as described later, and a suspension of the labeled cells is used as a sample for, for example, an imaging flow cytometer, and imaging of cells that can be analysis targets is performed.

[0059] A plurality of cells can be contained in the sample. The number of cells contained in the sample is not particularly limited, and at least 10 2 or more, preferably 10 3 or more, more preferably 10 4 or more, still more preferably 10 5 or more, yet more preferably 10 6 or more. In addition, the plurality of cells can contain different types of cells.

[0060] In the present embodiment, the cell to be the analysis target is also referred to as an analysis target cell. The analysis target cell can be a cell contained in a test subject collected from a subject. The above-mentioned cell is preferably a nucleated cell. Among the cells, there can be a normal cell and an abnormal cell.

[0061] The normal cell refers to a cell that is originally contained in the test subject corresponding to the part of the body from which the test subject is collected. The abnormal cell refers to a cell other than the normal cell. Among the abnormal cells, there can be a cell having a chromosomal abnormality and / or a tumor cell. Among them, the above-mentioned tumor cell is preferably a peripheral circulating tumor cell. More preferably, the peripheral circulating tumor cell does not refer to a hematopoietic tumor cell in which a tumor cell is present in the blood in a normal state, but refers to a tumor cell that circulates in the blood with a cell series other than the hematopoietic cell system as the origin. In the present specification, the tumor cell circulating in the periphery is also referred to as a circulating tumor cell (CTC).

[0062] The target site for detecting a chromosomal abnormality is the nucleus of the analysis target cell. As the chromosomal abnormality, for example, there can be a translocation, a deletion, an inversion, a duplication, and the like of a chromosome. As the cell having such a chromosomal abnormality, for example, there can be a cell appearing when a disease selected from leukemia such as myelodysplastic syndrome, acute myeloblastic leukemia, acute myeloblastic leukemia, acute promyelocytic leukemia, acute myelomonocytic leukemia, acute monocytic leukemia, erythroleukemia, acute megakaryocytic leukemia, acute myelogenous leukemia, acute lymphocytic leukemia, lymphoblastic leukemia, chronic myelogenous leukemia, or chronic lymphocytic leukemia, malignant lymphoma such as Hodgkin's lymphoma, non-Hodgkin's lymphoma, and multiple myeloma.

[0063] The chromosomal abnormality can be detected by a publicly known method such as the FISH method. In general, an examination item for detecting a chromosomal abnormality is set in correspondence with the kind of abnormal cell to be detected. In correspondence with which examination item is performed for the test subject, a gene or a locus to be the analysis target is set as an analysis item. In the detection of a chromosomal abnormality by the FISH method, by hybridizing a probe that specifically binds to a gene or a locus present in the nucleus of the cell to be the analysis target, an abnormality in the position or the number of chromosomes can be detected. The probe is labeled with a labeling substance. The labeling substance is preferably a fluorescent dye. The labeling substance differs in correspondence with the probe, and when the labeling substance is a fluorescent dye, fluorescent dyes having different wavelength regions of fluorescence are combined, and a plurality of genes or loci can be detected for one cell.

[0064] An abnormal cell is a cell that appears when a subject suffers from a specified disease, and includes, for example, a tumor cell such as a cancer cell, a leukemia cell, and the like. In the case of a hematopoietic organ, the specified disease is a disease selected from the group consisting of a leukemia such as myelodysplastic syndrome, acute myeloblastic leukemia, acute myeloblastic leukemia, acute promyelocytic leukemia, acute myelomonocytic leukemia, acute monocytic leukemia, erythroleukemia, acute megakaryocytic leukemia, acute myelogenous leukemia, acute lymphocytic leukemia, lymphoblastic leukemia, chronic myelogenous leukemia, or chronic lymphocytic leukemia, a malignant lymphoma such as Hodgkin's lymphoma, non-Hodgkin's lymphoma, and multiple myeloma. In addition, in the case of an organ other than a hematopoietic organ, the specified disease can be a digestive tract malignant tumor occurring in the upper pharynx, esophagus, stomach, duodenum, jejunum, ileum, cecum, appendix, ascending colon, transverse colon, descending colon, sigmoid colon, rectum, or anal portion, and the like; liver cancer; gallbladder cancer; bile duct cancer; pancreatic cancer; pancreatic duct cancer; a urinary organ malignant tumor occurring in the bladder, ureter, or kidney; a female reproductive organ malignant tumor occurring in the ovary, oviduct, uterus, and the like; breast cancer; prostate cancer; skin cancer; a central nervous system malignant tumor; a solid tumor such as a malignant tumor occurring in the bone soft tissue; and the like.

[0065] The detection of the abnormal cell can be performed using at least one selected from the group consisting of a bright field image, an immunostaining image for various antigens, and an organelle staining image in which an organelle is specifically stained.

[0066] The bright field image can be obtained by irradiating light to a cell, and taking a transmitted light or a reflected light from the cell. Preferably, the bright field image is preferably an image in which a phase difference of the cell is taken using a transmitted light.

[0067] The immunostaining image can be obtained by taking a cell in which a labeling substance is labeled by immunostaining using an antibody that binds to an antigen present in at least one of a target site in a cell or on a cell selected from the group consisting of a nucleus, cytoplasm, and cell surface. The labeling substance preferably uses a fluorescent dye as in the FISH method. The labeling substance differs in correspondence to the antigen, and when the labeling substance is a fluorescent dye, fluorescent dyes having different wavelength regions of fluorescence are combined, and a plurality of antigens can be detected for one cell.

[0068] The organelle-stained image can be obtained by taking a photograph of a cell that has been stained using a dye that selectively binds to a protein, sugar chain, lipid, or nucleic acid, etc. present in at least one of the nucleus, cytoplasm, and cell membrane, or a target site of the cell membrane. For example, as a staining dye specific to the nucleus, there are DNA-binding dyes such as Hoechst (trademark) 33342, Hoechst (trademark) 33258, 4', 6-diamidino-2-phenylindole (DAPI), Propidium Iodide (PI), ReadyProbes (trademark) nuclear stain reagent, and the like, histone-binding reagents such as CellLight (trademark) reagent, and the like. As a staining reagent specific to the nucleolus and RNA, there are SYTO (registered trademark) RNASelect (trademark) and the like that specifically bind to RNA. As a staining reagent specific to the cytoskeleton, there are, for example, fluorescently labeled phalloidin and the like. As a dye for staining organelles such as lysosomes, vesicles, Golgi apparatus, mitochondria, and the like, there are, for example, the CytoPainter series from Abeam plc (Cambridge, UK). These staining dyes or staining reagents are fluorescent dyes or reagents containing fluorescent dyes, and different wavelength regions of fluorescence can be selected in correspondence with the wavelength region of fluorescence of the fluorescent dyes used in other staining performed on the organelles or the cell as a whole.

[0069] When detecting abnormal cells, the inspection items are set in correspondence with the type of abnormal cells to be detected. In the inspection items, there can be included analysis items necessary for detecting abnormal cells. The analysis items can be set in correspondence with the above-described bright-field image, each antigen, and each organelle. The fluorescent dyes having different wavelength regions of fluorescence in each of the analysis items other than the bright field can detect different analysis items in one cell.

[0070] The analysis data for input to the artificial intelligence algorithm is obtained by the method described later. The data showing the properties of the cells contained in the analysis target image generated by the artificial intelligence algorithm is, for example, data showing whether the analysis target cell is normal or abnormal. More specifically, the data showing the properties of the cells contained in the analysis target image is data showing whether the analysis target cell is a cell having a chromosomal abnormality or a peripheral circulating tumor cell.

[0071] In this specification, for convenience, the "analysis target image" is sometimes referred to as the "analysis image", the "analysis data" is sometimes referred to as the "analysis data", the "training image" is sometimes referred to as the "training image", and the "training data" is sometimes referred to as the "training data". In addition, the "fluorescent image" refers to a training image photographed with fluorescence or an analysis image photographed with fluorescence.

[0072] [2. Cell analysis method using the first artificial intelligence algorithm]

[0073] For the training method of the first artificial intelligence algorithm 50, 53 and the analysis method of the cell using the trained first artificial intelligence algorithm 60, 63, the following is used. Figures 1 to 11 The first artificial intelligence algorithm 60, 63 can be a deep learning algorithm having a neural network structure. The neural network structure described above can be selected from a fully connected deep neural network (FC-DNN), a convolutional neural network (CNN), a recurrent neural network (RNN), and a combination of these. The convolutional neural network is preferred.

[0074] The artificial intelligence algorithm can use, for example, TensorFlow® provided by the Python company.

[0075] [2-1. Artificial intelligence algorithm for detecting chromosomal abnormalities]

[0076] The present embodiment relates to a training method of the first artificial intelligence algorithm 60 for detecting chromosomal abnormalities and an analysis method of the cell using the first artificial intelligence algorithm 60 for detecting chromosomal abnormalities. The term "training" or "training" is sometimes replaced by the term "generation" or "generation" is used.

[0077] [(1) Generation of training data]

[0078] The following is used. Figure 1 and Figure 2 The training method of the first artificial intelligence algorithm 50 for detecting chromosomal abnormalities will be described. In the training method of the first artificial intelligence algorithm 50 for detecting chromosomal abnormalities, the following is used. Figure 1 An example of an image showing FISH staining of a PML-RARA chimera gene formed by translocation of a PML gene as a transcription control factor located on the long arm of chromosome 15 (15q24.1) and a retinoic acid receptor alpha (RARA) gene located on the long arm of chromosome 17 (17q21.2) is shown.

[0079] As shown in Figure 1 The positive training data 73P and the negative training data 73N are each generated from a positive training image 70P that photographs a cell (hereinafter referred to as "first positive control cell") that is positive for chromosomal abnormalities and a negative training image 70N that photographs a cell (hereinafter referred to as "first negative control cell") that is negative for chromosomal abnormalities. Sometimes the positive training image 70P and the negative training image 70N are collectively referred to as the training image 70. In addition, sometimes the positive training data 73P and the negative training data 73N are collectively referred to as the training data 73.

[0080] In an example illustrating the detection of the PML-RARA chimeric gene, a first fluorescent dye is shown that causes the probe detecting the PML locus to bind to a greenish wavelength region, while a second fluorescent dye is shown that causes the probe detecting the RARA locus to bind to a reddish wavelength region, different from the first fluorescent dye. The nuclei of a first positive control cell and a first negative control cell can be labeled with the first and second fluorescent dyes respectively using a FISH method, using both probes that bind the first and second fluorescent dyes. Sometimes, labeling of the target site with the first fluorescent dye is referred to as first fluorescent labeling, and labeling of the target site with the second fluorescent dye is referred to as second fluorescent labeling.

[0081] A sample containing cells labeled with a first fluorescent marker and a second fluorescent marker can be analyzed in a cell imaging device such as an imaging flow cytometer to capture images of the cells. Multiple images of the cells can be captured for the same field of view of the same cells. Because the first and second fluorescent markers have different wavelength regions of fluorescence due to their respective fluorescent dyes, the first filter for transmitting light emitted from the first fluorescent dye and the second filter for transmitting light emitted from the second fluorescent dye are different. Therefore, the light transmitted through the first filter and the light transmitted through the second filter are received by the imaging unit 160 (described later) via their respective first and second channels, and captured as separate images of the same field of view of the same cells. That is, in the imaging unit 160, multiple images corresponding to the number of labeled substances on the labeled cells are acquired for the same field of view of the same cells.

[0082] Therefore, in Figure 1 In the example, such as Figure 1 As shown in (A), the positive training image 70P may contain a first positive training image 70PA (green fluorescent label) captured via the first channel and a second positive training image 70PB (red fluorescent label) captured via the second channel for the first positive control cell. The first positive training image 70PA and the second positive training image 70PB correspond to images of the same field of view for the same cells. Next, the brightness of the captured light in each pixel of the image is transformed into numerical training data 71PA and 71PB for the first and second positive cells, respectively.

[0083] The generation method of the 1st positive numerical training data 71PA will be described using the 1st positive training image 70PA. In order to extract a cell region, for example, each image captured in the imaging section 160 is subjected to trimming processing to a specified number of pixels, for example, 100 pixels in the vertical direction x 100 pixels in the horizontal direction, to generate a training image 70. At this time, the images obtained from each channel are trimmed so as to have the same field of view for one cell. The trimming processing can exemplify determining the center of gravity of the cell, and cutting out the cell region within a range of a specified number of pixels centered on the center of gravity. When the images of the cells flowing through the flow cell are captured, the positions of the cells in the images are sometimes different between the images, and by performing the trimming, more accurate training becomes possible. The 1st positive training image 70PA is represented as, for example, a 16-bit grayscale image. Thus, in each pixel, the luminance of the pixel can be represented by a value of 65,536 gradations of luminance from 1 to 65,536. As shown in Figure 1 (A), the value of the gradation of the luminance in each pixel of the 1st positive training image 70PA is the 1st positive numerical training data 71PA, which is represented by a row of numbers corresponding to each pixel.

[0084] Similarly to the 1st positive numerical training data 71PA, the 2nd positive numerical training data 71PB, which shows the luminance of the captured light in each pixel (pixel) within the image by a numerical value, can be generated from the 2nd positive training image 70PB.

[0085] Next, the 1st positive numerical training data 71PA and the 2nd positive numerical training data 71PB are integrated per pixel to generate positive integrated training data 72P. As shown in Figure 1 (A), the positive integrated training data 72P becomes row and column data in which the numerical value in each pixel of the 1st positive numerical training data 71PA and the value in the corresponding pixel of the 2nd positive numerical training data 71PB are displayed side by side.

[0086] Next, a label value 74P indicating that the positive integrated training data 72P originates from the 1st positive control cell is attached to the positive integrated training data 72P to generate label-attached positive integrated training data 73P. As the label indicating the 1st positive control cell, "2" is attached in Figure 1 (A).

[0087] The label-attached negative integrated training data 73N is also generated from the negative training image 70N, similarly to when the label-attached positive integrated training data 73P is generated.

[0088] As Figure 1As shown in (B), the negative training image 70N contains a first negative training image 70NA, captured via the first channel, showing a green first fluorescent label, and a second negative training image 70NB, captured via the second channel, showing a red second fluorescent label. The imaging and retouching, and the numericalization of the light intensity in each pixel, are performed in the same manner as when obtaining the first positive numerical training data 71PA from the first positive training image 70PA. The first negative numerical training data 71NA, which displays the numerical value of the light intensity captured in each pixel within the image, can be generated from the first negative training image 70NA using the same method as the first positive numerical training data 71PA.

[0089] Similarly, second negative numerical training data 71NB can be generated from the second negative training image 70NB, which displays the brightness of the captured light in each pixel of the image in numerical form.

[0090] like Figure 1 As shown in (B), based on the method for generating positive integrated training data 72P, the first negative numerical training data 71NA and the second negative numerical training data 71NB are integrated for each pixel to generate negative integrated training data 72N. Figure 1 As shown in (B), the negative integrated training data 72N becomes row and column data in which the values ​​in each pixel of the first negative numerical training data 71NA and the values ​​in each corresponding pixel of the second negative numerical training data 71NB are displayed side by side.

[0091] Next, a label value 74N indicating that this negative integration training data 72N originates from the first negative control cell is attached to the negative integration training data 72N, generating a label attached to the negative integration training data 73N. As a label indicating that it is the first negative control cell, in... Figure 2 (B) Appendix “1”.

[0092] exist Figure 3A method of displaying the label-attached positive integrated training data 73P and the label-attached negative integrated training data 73N generated in the first artificial intelligence algorithm 50. The number of nodes of the input layer 50a in the first artificial intelligence algorithm 50 having a neural network structure corresponds to the product of the number of pixels of the training image 70 (10,000 in the above example) and the number of channels for one cell (2 channels of the green channel and the red channel in the above example). Data corresponding to the positive integrated training data 72P of the label-attached positive integrated training data 73P is input to the input layer 50a of the neural network. The label value 74P corresponding to the data input to the input layer 50a is input to the output layer 50b of the neural network. In addition, data corresponding to the negative integrated training data 72N of the label-attached negative integrated training data 73N is input to the input layer 50a of the neural network. The label value 74N corresponding to the data input to the input layer 50a is input to the output layer 50b of the neural network. Each weight in the intermediate layer 50c of the neural network is calculated from these inputs, the first artificial intelligence algorithm 50 is trained, and the trained first artificial intelligence algorithm 60 is generated.

[0093] (2) Generation of data for analysis and cell analysis

[0094] Using Figure 3 For photographing the cells flowing through the flow cell 110, the integrated analysis data 82 is generated from the generated analysis image 80, and the cell analysis method using the trained first artificial intelligence algorithm 60 is described. The analysis image 80 can be photographed in the same manner as the method of photographing the training image 70.

[0095] As Figure 3As shown, cells flowing through flow cell 110 are captured by camera unit 160 to generate resolved images 80. By capturing cells flowing through flow cell 110, a large number of resolved images 80 can be generated in a short time, making it possible to resolve a large number of cells in a short time. In cases where the number of abnormal cells in a test subject is much smaller than the number of normal cells that should be present in the test subject, this resolution method, which can resolve a large number of cells in a short time, can suppress the neglect of abnormal cells. The resolved images 80 include a first resolved image 80A with a green first fluorescent label captured via the first channel and a second resolved image 80B with a red second fluorescent label captured via the second channel for the cells to be resolved. The imaging and trimming, and the numerical conversion of the light intensity in each pixel are the same as when obtaining the first positive value training data 71PA from the first positive fluorescent label image 70PA. As mentioned above, the positions of cells in the images of cells flowing through the flow cell are sometimes different between images; by performing trimming, higher precision resolution becomes possible. The first numerical analysis data 81A, which displays the brightness of the captured light in each pixel of the image, can be generated from the first analytical image 80A using the same method as the first positive numerical training data 71PA.

[0096] Similarly, second numerical analysis data 81B, which displays the brightness of the captured light in each pixel of the image, can be generated from the second analyzed image 80B.

[0097] like Figure 3 As shown, based on the method for generating positive integrated training data 72P, integrated parsing data 82 is generated by integrating the first numerical parsing data 81A and the second numerical parsing data 81B for each pixel. Figure 3 As shown, the integrated parsing data 82 is displayed side-by-side with the values ​​of each pixel in the first numerical parsing data 81A and the corresponding values ​​in each pixel in the second numerical parsing data 81B.

[0098] like Figure 3 As shown, the generated integrated parsing data 82 is input into the input layer 60a of the neural network within the first trained artificial intelligence algorithm 60. The values ​​contained in the input integrated parsing data 82 are processed through the intermediate layer 60c of the neural network and output from the output layer 60b of the neural network a label value 84 indicating whether the parsed cell has a chromosomal abnormality. Figure 4 In the example shown, when the parsed cell is determined to have no chromosomal abnormalities, the label value is "1", and when it is determined to have chromosomal abnormalities, the label value is "2". Alternatively, labels such as "none", "present", "normal", or "abnormal" can be output instead of the label value.

[0099]

(3) Other components

[0100] i.In the present embodiment, sometimes an Extended Depth of Field (EDF) filter for enlarging the depth of field when photographing cells is used in an imaging flow cytometer, and the depth of focus of the photographed image is restored to provide the examiner with a cell image. However, the training image 70 and the analysis image 80 used in the present embodiment are preferably images that have not been subjected to restoration processing for images photographed using the EDF filter. An example of an image that has not been subjected to restoration processing is shown in Figure 4 . Figure 4 Cells that showed positive for the PML-RARA chimeric gene are shown. (A) and (B) are images of different cells. Figure 4 (A) and Figure 4 (B) on the left show images photographed for the first fluorescent marker. Figure 4 (A) and Figure 5 The images on the right of (A) and (B) are images photographed for the second fluorescent marker for the same cells as those on the left, in the same field of view as the images on the left.

[0101] ii.From the training image 70 and the analysis image 80, images that are out of focus can be excluded. Whether the focus of an image is in focus or not can be determined when the difference in luminance from an adjacent pixel is calculated for each pixel, and when the gradient of the difference in luminance is not extreme in the entire image. An image that is out of focus can be determined when the gradient of the difference in luminance is extreme in the entire image.

[0102] iii.The training image 70 and the analysis image 80 used in the present embodiment are exemplarily adjusted in pixel number so as to be 100 pixels in the vertical direction x 100 pixels in the horizontal direction, but the size of the image is not limited thereto. The number of pixels can be appropriately set to be 50 to 500 pixels in the vertical direction and 50 to 500 pixels in the horizontal direction. The number of pixels in the vertical direction and the number of pixels in the horizontal direction are not necessarily the same. However, the training image 70 used for training the first artificial intelligence algorithm 50 and the analysis image 80 used for generating the integrated analysis data 82 input to the first artificial intelligence algorithm 60 trained using the training image 70 described above are preferably the same in pixel number, and the number of pixels in the vertical direction and the number of pixels in the horizontal direction are also the same.

[0103] iv.In the present embodiment, the training image 70 and the analysis image 80 use a 16-bit grayscale image. However, the color scale of the luminance can be 8 bits, 32 bits, or the like, in addition to 16 bits. In the present embodiment, the values of the luminance expressed in 16 bits (65,536 color scales) are used directly as each numerical training data 71PA, 71PB, 71NA, 71NB, but low-dimensionalization processing that integrates the values in a certain color scale can be performed, and the values after low-dimensionalization can be used as each numerical training data 71PA, 71PB, 71NA, 71NB. At this time, the same processing is preferably performed for the training image 70 and the analysis image 80.

[0104] v. The chromosomal abnormality detectable in the present embodiment is not limited to the PML-RARA chimeric gene. For example, the BCR / ABL fusion gene, the AML1 / ETO (MTG8) fusion gene (t(8; (21)), the PML / RARα fusion gene (t(15; (17)), the AML1 (21q22) translocation, the MLL (11q23) translocation, the TEL (12p13) translocation, the TEL / AML1 fusion gene (t(12; (21)), the IgH (14q32) translocation, the CCND1 (BCL1) / IgH fusion gene (t(11; (14)), the BCL2 (18q21) translocation, the IgH / MAF fusion gene (t(14; (16)), the IgH / BCL2 fusion gene (t(14; (18)), the c-myc / IgH fusion gene (t(8; (14)), the FGFR3 / IgH fusion gene (t(4; (14)), the BCL6 (3q27) translocation, the c-myc (8q24) translocation, the MALT1 (18q21) translocation, the API2 / MALT1 fusion gene (t(11; (18) translocation), the TCF3 / PBX1 fusion gene (t(1; (19) translocation), the EWSR1 (22q12) translocation, the PDGFRβ (5q32) translocation, and the like can be detected.

[0105] In addition, various changes can be involved in the translocation. Figure 6 and Figure 6 Examples of the fluorescent labels showing a typical positive pattern (main pattern) of the BCR / ABL fusion gene. In the state of superimposing the first fluorescent label image and the second fluorescent label image, the number of the first fluorescent labels in the negative example when using the ES probe is 2, the number of the second fluorescent labels is 2, and the number of the fusion fluorescent labels is 0. The number of the first fluorescent labels in the typical positive pattern example when using the ES probe is 1, the number of the second fluorescent labels is 2, and the number of the fusion fluorescent labels is 1. In the state of superimposing the first fluorescent label image and the second fluorescent label image when using the DF probe, the number of the first fluorescent labels in the negative pattern is 2, the number of the second fluorescent labels is 2, and the number of the fusion fluorescent labels is 0. The number of the first fluorescent labels in the typical positive pattern example when using the DF probe is 1, the number of the second fluorescent labels is 1, and the number of the fusion fluorescent labels is 2.

[0106] Figure 7is an example of fluorescent labeling of the atypical positive pattern of the BCR / ABL fusion gene. One example of the atypical positive pattern is a secondary BCR / ABL pattern, and three first fluorescent labels are also detected with the ES probe because the breakpoint of the BCR gene is in a relatively upstream position of the BCR gene. Another example of the atypical positive pattern is an example in which a part of the binding region of the probe targeting the ABL gene of chromosome 9 is deleted, and only one fusion fluorescent label is detected when the DF probe is used in dependence thereon, although two fusion fluorescent labels should have been detected. In addition, another example of the atypical positive pattern is an example in which a part of the binding region of the probe targeting the ABL gene of chromosome 9 and a part of the binding region of the probe targeting the BCR gene of chromosome 22 are deleted. Only one fusion fluorescent label is detected when the DF probe is used in dependence thereon, although two fusion fluorescent labels should have been detected.

[0107] In Figure 7 An example of a reference pattern showing a negative pattern and a positive pattern when a chromosomal abnormality associated with the ALK locus is detected. In the negative pattern, two fusion fluorescent labels exist because the ALK gene is not truncated. On the other hand, in the positive pattern, the fusion fluorescent label becomes only one (when only one of the alleles is truncated), or becomes unrecognizable (when both of the alleles are truncated) because the ALK gene is truncated. This negative pattern and positive pattern are the same for the ROS1 gene and the RET gene in addition to the ALK gene.

[0108] Further, in Figure 7 An example of a reference pattern showing a chromosomal abnormality of deletion of the long arm of chromosome 5 (5q). For example, designed in a manner in which the first fluorescent label probe binds to the long arm of chromosome 5 and the second fluorescent label probe binds to the centromere of chromosome 5. In the negative pattern, because the number of the centromere of chromosome 5 and the number of the long arm of chromosome 5 are the same, the first fluorescent label and the second fluorescent label exist each in two, reflecting the same number of chromosomes. In the positive pattern, deletion of the long arm of one or both of the chromosomes 5 occurs, and the number of the first fluorescent label becomes only one or zero. Deletion of the short arm or the long arm of other chromosomes is also the same for this negative pattern and positive pattern. As an example of deletion of the long arm of other chromosomes, deletion of the long arm of chromosome 7 and deletion of the long arm of chromosome 20 can be given. In addition, further, as an example showing the same positive pattern and negative pattern, 7q31 (deletion), p16 (9p21 deletion analysis), IRF-1 (5q31) deletion, D20S108 (20q12) deletion, D13S319 (13q14) deletion, 4q12 deletion, ATM (11q22.3) deletion, p53 (17p13.1) deletion, and the like can be given.

[0109] Further, in Figures 8 to 10Example of display of trisomy of chromosome No. 8. First fluorescent label probe binds to centromere of chromosome No. 8, for example. First fluorescent label of positive pattern becomes 3. First fluorescent label of negative pattern becomes 2. Such fluorescent label pattern of trisomy of chromosome No. 12 is also the same. Further, in monosomy of chromosome No. 7, when first fluorescent label probe which binds to centromere of chromosome No. 7 is used, for example, first fluorescent label of positive pattern becomes 1. First fluorescent label of negative pattern becomes 2.

[0110] [2-2. Artificial intelligence algorithm for detecting peripheral circulating tumor cells]

[0111] The present embodiment relates to a training method of the first artificial intelligence algorithm 63 for detecting peripheral circulating tumor cells and a cell analysis method using the first artificial intelligence algorithm 63 for detecting peripheral circulating tumor cells. The term "training" or "train" is sometimes replaced with the term "generate" or "generate" in use.

[0112] [(1) Generation of training data]

[0113] The training method of the first artificial intelligence algorithm 53 for detecting peripheral circulating tumor cells is described using Figure 8 , for example.

[0114] Figure 8 A pre-processing method for an image captured by the imaging section 160 is shown. Figure 8 (A) shows an image captured before pre-processing. Pre-processing is a trimming process for making training images 75 and analysis images 85 the same size, and can be performed on all images used as training images 75 or analysis images 85. In Figure 8 (a) and (b) in (A) are images of the same cell, but the channels at the time of imaging are different. In Figure 8 (A), (c) is an image of a different cell from (a). (c) and (d) are images of the same cell, but the channels at the time of imaging are different. As Figure 8 As shown in (a) and (c) of (A), the size of the image at the time of imaging the cell is sometimes different. In addition, the size of the cell itself also differs depending on the cell. Therefore, it is preferable to trim the acquired image in a manner that reflects the size of the cell and becomes the same image size. In Figure 8 the example shown in (A), the center of gravity of the nucleus of the cell within the image is taken as the center, and positions 16 pixels apart in the vertical and horizontal directions from the above-mentioned center are set as trimming positions. The image cut out by trimming is shown in Figure 8 (B). Figure 8 (B) (a) is an image cut out from Figure 8 (A) (a), Figure 8 (B) (b) is an image cut out fromFigure 8 (A) (b) the cut-out image, Figure 8 (B) (c) is from Figure 8 (A) (c) the cut-out image, Figure 8 (B) (d) is from Figure 8 (A) (d) the cut-out image. Figure 9 Each image of (B) becomes 32 pixels in the vertical direction x 32 pixels in the horizontal direction. The center of gravity of the nucleus can be determined using, for example, analysis software (IDEAS) attached to an imaging flow cytometer (ImageStream Mark II, Luminex).

[0115] In Figure 10 and Figure 9 a training method of the first artificial intelligence algorithm 53 is shown.

[0116] As shown in Figure 9 , positive integrated training data 78P and negative integrated training data 78N are each generated from a positive training image 75P that photographs a peripheral circulating tumor cell (hereinafter referred to as "second positive control cell") and a negative training image 75N that photographs a cell other than a peripheral circulating tumor cell (hereinafter referred to as "second negative control cell"). The positive training image 75P and the negative training image 75N are sometimes collectively referred to as a training image 75. In addition, the positive integrated training data 78P and the negative integrated training data 78N are sometimes collectively referred to as training data 78.

[0117] When detecting a peripheral circulating tumor cell, in an image photographed by the imaging section 160, a bright field image and a fluorescent image can be included. The bright field image can be an image that photographs a phase difference of a cell. This imaging can be performed by, for example, the first channel. The fluorescent image is an image that photographs a fluorescent marker that marks a target site in a cell by immunostaining or organelle staining. The fluorescent marker is performed by a fluorescent dye having a wavelength region of fluorescence that is different for each antigen and / or for each organelle.

[0118] For example, when the first antigen is labeled with a first fluorescent dye that emits a wavelength region of fluorescence of green color, by causing an antibody that directly or indirectly binds to the first antigen to bind to the first fluorescent dye, the first fluorescent dye can be labeled to the first antigen.

[0119] When an antibody that binds to the second antigen is caused to bind to a second fluorescent dye that emits a wavelength region of fluorescence of red color that is different from the first fluorescent dye, by causing an antibody that directly or indirectly binds to the second antigen to bind to the second fluorescent dye, the second fluorescent dye can be labeled to the second antigen.

[0120] When an antibody that binds to the third antigen binds to a third fluorescent dye that has a yellowish wavelength region different from the first and second fluorescent dyes, the third antigen can be labeled with the third fluorescent dye by binding the antibody that binds directly or indirectly to the third antigen.

[0121] In this way, by labeling each antigen and / or each organelle with a fluorescent dye having a different wavelength region of fluorescence, the first to the Xth fluorescent labels can be labeled with fluorescent dyes having different wavelength regions of fluorescence.

[0122] A sample containing cells with fluorescent labels of the first to the Xth generation can be captured by imaging in a cell imaging device such as an imaging flow cytometer, and images of the cells can be obtained. Multiple images of the cells can be obtained for the same field of view of the same cells. Since the fluorescent labels of the first to the Xth generation have different wavelength regions of fluorescence due to the different fluorescent dyes, the filters used to transmit light emitted from each fluorescent dye are different for each dye. Furthermore, it is necessary to use a filter different from the filter that transmits light from the fluorescent dye for bright-field images. Therefore, the light transmitted through each filter is received by the imaging unit 160 (described later) through its respective channel, and is captured as separate images of the same field of view of the same cells. That is, in the imaging unit 160, multiple images corresponding to the number of labeled cells plus the number of bright-field images are obtained for the same field of view of the same cells.

[0123] exist Figure 9 In the example shown, in Figure 9 In (A) and (B), channel 1 (Ch1) displays a bright-field camera image. Figure 9 In (A) and (B), the second channel (Ch2), the third channel (Ch3), ... the Xth channel (ChX) refer to the channels for capturing images of different labeled substances.

[0124] exist Figure 9 In the example, such as Figure 9 As shown in (A), the positive training image 75P may contain the first positive training image 75P1 captured via the first channel for the second positive control cell, the second positive training image 75P2 captured via the second channel for the first fluorescent label, the third positive training image 75P3 captured via the third channel for the second fluorescent label, and the Xth positive training images 75Px captured via the Xth channel for each fluorescent label. The first positive training images 75P1 to the Xth positive training images 75Px correspond to images of the same field of view for the same cells. Next, the brightness of the captured light in each pixel of the image is transformed into the first positive numerical training data 76P1 to the Xth positive numerical training data 76Px, expressed in numerical form.

[0125] The generation method of the first positive numerical training data 76P1 will be described using the first positive training image 75P1. Each image photographed in the imaging section 160 is subjected to the above-described preprocessing, for example, trimmed to 32 pixels in the vertical direction 32 x 32 pixels in the horizontal direction, to become a training image 75. The first positive training image 75P1 is represented as, for example, a 16-bit gray scale image. Thus, in each pixel, the luminance of the pixel can be represented by a value of 65,536 gradations of luminance from 1 to 65,536. As shown in Figure 9 (A), the value of the gradation of luminance in each pixel of the first positive training image 75P1 is the first positive numerical training data 76P1, represented as a table of rows and columns of numbers corresponding to each pixel.

[0126] Similarly to the first positive numerical training data 76P1, the second positive numerical training data 76P2 to the Xth positive numerical training data 76Px can be generated from the second positive training image 75P2 to the Xth positive training image 75Px, representing the luminance of the photographed light in each pixel (pixel) in the image by a numerical value.

[0127] Next, the first positive numerical training data 76P1 to the Xth positive numerical training data 76Px are integrated per pixel to generate positive integrated training data 77P. As shown in Figure 9 (A), the positive integrated training data 77P becomes a table of rows and columns in which the numerical value in each pixel of the first positive numerical training data 76P1 is displayed side by side with the value in the corresponding pixel of the second positive numerical training data 76P2 to the Xth positive numerical training data 76Px.

[0128] Next, a label value 79P indicating that this positive integrated training data 77P is derived from the second positive control cell is attached to the positive integrated training data 77P to generate label-attached positive integrated training data 78P. As the label indicating that it is the second positive control cell, "2" is attached in Figure 9 (A).

[0129] From the negative training image 75N, label-attached negative integrated training data 78N is also generated similarly to when the label-attached positive integrated training data 78P is generated.

[0130] As shown in Figure 9As shown in (B), the negative training image 75N contains, similarly to the positive training image 75P, the first negative training images 75N1 to Xth negative training images 75Nx obtained from the images through the first channel to the Xth. The numerical representation of the light intensity in each pixel is the same as when obtaining the first positive numerical training data 76P1 to Xth positive numerical training data 76Px from the first positive training images 75PA to the Xth positive training images 75Px. The first negative numerical training data 76N1, which represents the light intensity captured in each pixel (pixel) of the image, can be generated from the first negative training image 75N1 using the same method as the first positive numerical training data 76P1.

[0131] Similarly, negative numerical training data 76N2 to Xth can be generated from the second negative training image 75N2 to the Xth negative training image 75Nx, which express the brightness of the captured light in each pixel (pixel) in the image in numerical terms.

[0132] like Figure 9 As shown in (B), based on the method for generating positive integrated training data 77P, negative integrated training data 77N is generated by integrating the first negative numerical training data 76N1 to the Xth negative numerical training data 76Nx for each pixel. Figure 10 As shown in (B), the negative integrated training data 77N becomes the row and column data in which the values ​​in each pixel of the first negative numerical training data 76N1 are displayed side by side with the values ​​in each pixel of the second negative numerical training data 76N2 to the Xth negative numerical training data 76Nx.

[0133] Next, a label value 79N indicating that this negative integration training data 77N originates from the second negative control cell is attached to the negative integration training data 77N, generating a label attached to the negative integration training data 78N. This label represents the second negative control cell. Figure 11 (B) Appendix “1”.

[0134] exist Figure 11The method of inputting label-attached positive integration training data 78P and label-attached negative integration training data 78N into the first artificial intelligence algorithm 53 is shown. The number of nodes in the input layer 53a of the first artificial intelligence algorithm 53, which has a neural network structure, corresponds to the product of the number of pixels in the training image 75 (32×32 = 1024 in the example above) and the number of channels for one cell (X channels from 1 to X in the example above). Data corresponding to the positive integration training data 77P of the label-attached positive integration training data 78P is input to the input layer 53a of the neural network. The label value 79P corresponding to the data input to the input layer 53a is input to the output layer 53b of the neural network. Additionally, data corresponding to the negative integration training data 77N of the label-attached negative integration training data 78N is input to the input layer 53a of the neural network. The label value 79N corresponding to the data input to the input layer 53a is input to the output layer 53b of the neural network. The weights in the intermediate layer 53c of the neural network are calculated from these inputs, the first artificial intelligence algorithm 53 is trained, and the trained first artificial intelligence algorithm 63 is generated.

[0135]

(2) Generation of data for parsing

[0136] use Figure 11 The method for generating the integrated parsed data 72 after parsing image 85 and the cell parsing method using the trained first artificial intelligence algorithm 63 will be described. The parsed image 85 can be captured and preprocessed in the same way as the training image 75.

[0137] like Figure 11 As shown, the resolved image 85 includes a first resolved image 85T1, taken via the first channel as a bright-field image for the target cell, and second resolved images 85T2 to Xth resolved images 85Tx, taken via the second to Xth channels for fluorescent labels. The imaging and preprocessing, and the numerical representation of the light intensity in each pixel, are the same as when obtaining the first positive numerical training data 76P1 from the first positive training image 75P1. The first numerical resolved data 86T1, which represents the light intensity of each pixel in the image in numerical terms, can be generated from the first resolved image 85T1 using the same method as the first positive numerical training data 76P1.

[0138] Similarly, from the second resolved image 85T2 to the Xth resolved image 85Tx, the second numerical resolved data 86T2 to the Xth numerical resolved data 86Tx, which represent the brightness of the captured light in each pixel (pixel) within the image, can be generated.

[0139] like Figure 11As shown, the cells flowing through the flow cell 110 are imaged according to the method of generating positive integrated training data 77P, and first to Xth numerical analysis data 86T1 to 86Tx are generated per pixel, and integrated analysis data 87 is generated. As shown in Figure 11 As shown, the integrated analysis data 87 is tabular data in which the values in each pixel of the first numerical analysis data 86T1 are displayed side by side with the values in the corresponding pixels of the second to Xth numerical analysis data 86T2 to 86Tx.

[0140] As shown, the cells flowing through the flow cell 110 are imaged according to the method of generating positive integrated training data 77P, and first to Xth numerical analysis data 86T1 to 86Tx are generated per pixel, and integrated analysis data 87 is generated. As shown in Figure 12 As shown, the cells flowing through the flow cell 110 are imaged according to the method of generating positive integrated training data 77P, and first to Xth numerical analysis data 86T1 to 86Tx are generated per pixel, and integrated analysis data 87 is generated. As shown in Figure 13 In the example shown in FIG. 8, when the analysis target cell is determined not to be a peripheral circulating tumor cell, the label value "1" is output, and when it is determined to be a peripheral circulating tumor cell, the label value "2" is output. Instead of the label value, labels such as "none", "yes", "normal", and "abnormal" can also be output.

[0141]

(3) Other configurations

[0142] i. The training images 75 and analysis images 85 used in the present embodiment are preferably images that are not subjected to restoration processing, for images taken using an EDF filter.

[0143] ii. Images in which the focus is not in agreement can be excluded from the training images 75 and analysis images 85 at the time of imaging.

[0144] iii.In the present embodiment, the training image 75 and the analysis image 85 are illustratively trimmed to have a pixel number of 32 pixels vertically by 32 pixels horizontally, but are not limited as long as the size of the image is such that the entire cell is contained within the image. The pixel number in the vertical direction and the pixel number in the horizontal direction of the image are not necessarily the same. However, the training image 75 used for training the first artificial intelligence algorithm 53 and the analysis image 85 used for generating the integrated analysis data 87 input to the first artificial intelligence algorithm 63 trained using the training image 75 described above are preferably the same pixel number, and the pixel number in the vertical direction and the pixel number in the horizontal direction are also the same.

[0145] iv.In the present embodiment, the training image 70 and the analysis image 80 use a 16-bit grayscale image. However, the color scale of the luminance can be 8 bits, 32 bits, or the like in addition to 16 bits. In the present embodiment, the values of the luminance expressed in 16 bits (65,536 color scales) are directly used as the numerical value training data 76P1 to 76Px and the numerical value training data 76N1 to 76Nx, but low-dimensionalization processing that integrates these values in a certain color scale range can be performed, and the values after low-dimensionalization can be used as the numerical value training data 76P1 to 76Px and the numerical value training data 76N1 to 76Nx. At this time, it is preferable to perform the same processing on the training image 70 and the analysis image 80.

[0146] 【3. Cell analysis method using the second artificial intelligence algorithm】

[0147] The training method of the second artificial intelligence algorithm 94 and the cell analysis method using the trained second artificial intelligence algorithm 97 are described using Figure 12 and Figure 13 . The second artificial intelligence algorithm 94, 97 is an algorithm other than a deep learning algorithm that can have a neural network structure. The second artificial intelligence algorithm 94 is trained using the feature amount determined by the user extracted from the second positive control cell or the second negative control cell described above and the label showing the property of the second positive control cell or the second negative control cell corresponding to the extracted feature amount as training data. In addition, the trained second artificial intelligence algorithm 97 extracts a feature amount corresponding to the feature amount extracted when generating the training data from the analysis target image, and generates data showing the property of the cell using the same as analysis data.

[0148] In this embodiment, algorithms that can be used as the second artificial intelligence algorithms 94 and 97 include random forest, gradient boosting, support vector machine (SVM), relevance vector machine (RVM), naive Bayes, logistic regression, fed forward neural network, deep learning, K-nearest neighbor method, AdaBoost, Bagging, C4.5, kernel approximation, probabilistic gradient descent (SGD) classifier, Lasso, ridge regression, Elastic Net, SGD regression, kernel regression, Lowess regression, matrix factorization, non-negative matrix factorization, kernel matrix factorization, interpolation, kernel smoothing, and coordinated selection. Random forest or gradient boosting are preferred as the second artificial intelligence algorithms 94 and 97.

[0149] As the second artificial intelligence algorithm 94, 97, it can be used, for example, from Python.

[0150] The word "training" or "training" is sometimes replaced by "generating" or "generating".

[0151]

(1) Generation of training data

[0152] like Figure 13 As shown in (A), in this embodiment, positive training data 91A and negative training data 91B are generated from the positive training image 90A of the second positive control cell used in 2-2. and the negative training image 90B of the second negative control cell used in 2-2., respectively. Sometimes, the positive training image 90A and the negative training image 90B are collectively referred to as training image 90. Additionally, sometimes, the positive training data 91A and the negative training data 91B are collectively referred to as training data 91.

[0153] When detecting peripheral circulating tumor cells, the image captured by the imaging unit 160 (described later) used as training image 90 can be a bright-field image and / or a fluorescence image, similar to 2-2 above. Bright-field images can capture the phase difference of cells. Training image 90 can be obtained in the same way as 2-2.(1) above.

[0154] As training data 91, examples can be found as follows: Figure 13 The characteristic quantities shown. Figure 14 The features shown can be categorized into five classes. These classes are information about cell size, cell location, cell shape, cell texture, and light intensity obtained from the cell image. Details of the features contained in each class are as follows: Figure 14The characteristic amounts shown can use one or more than two combinations. Among the characteristic amounts, at least one information selected from information on the size of the cell is preferable. More preferably, among the characteristic amounts, information on the area of the cell is preferable. These characteristic amounts can be determined using, for example, the analysis software (IDEAS) described above.

[0155] In Figure 13 Explanations of representative characteristic amounts are shown. Figure 14 (A) Explanations of the height and the width in Figure 13 The height (height) exemplarily means the length of the long side (one side in the case of a square) of the smallest quadrangle (preferably a right-angled rectangle, or a square) that can enclose the cell on the image. The width (width) exemplarily means the length of the short side (one side in the case of a square) of the above-mentioned quadrangle. Figure 14 (B) Explanations of the major axis (major axis) and the minor axis (minor axis) in Figure 14 The major axis exemplarily means the long diameter of the ellipse (preferably a right-angled ellipse) that can enclose the cell on the image, the center of gravity of which overlaps with the center of gravity of the cell, and the smallest ellipse that can enclose the cell. The minor axis means the short diameter of the above-mentioned ellipse. The minor axis exemplarily means the short diameter of the above-mentioned ellipse.

[0156] In Figure 14 (C) Explanations of the length (length) of the cell, the maximum thickness (maximum thickness), and the minimum thickness (minimum thickness) are shown. The length of the cell is different from the height shown in Figure 14 (A) and means the length of the longest line when a line segment connecting the tip of one side and the tip of the other side of the cell on the image is assumed. The maximum thickness means the length of the longest line segment when a line segment and a straight line that represent the length are assumed from the line segments on the inside divided by the contour line of the cell. The minimum thickness means the length of the shortest line segment when a line segment and a straight line that represent the length are assumed from the line segments on the inside divided by the contour line of the cell.

[0157] In Figure 12 (D) Explanations of the aspect ratio (aspect ratio), the extension (extension), and the shape ratio (shape ratio) are shown. The aspect ratio means the value obtained by dividing the length of the minor axis by the length of the major axis. The extension means the value obtained by dividing the value of the height by the value of the width. The shape ratio means the value obtained by dividing the value of the minimum thickness by the value of the maximum thickness.

[0158] In Figure 12 (E) Explanations of the leaf symmetry (lobation) are shown. In Figure 12 (E) Examples of 2-leaf symmetry (2 minutes leaf), 3-leaf symmetry (3 minutes leaf), and 4-leaf symmetry (4 minutes leaf) are shown. The lobation means that one cell is divided into leaves.

[0159] As Figure 12As shown in (A), the positive training data 91A is combined with, for example, "2" as a label value indicating that it is derived from the 2nd positive control cell, as the label-attached positive training data 92A, and is input to the 2nd artificial intelligence algorithm 94. The negative training data 91B is combined with, for example, "1" as a label value indicating that it is derived from the 2nd negative control cell, as the label-attached negative training data 92B, and is input to the 2nd artificial intelligence algorithm 94. The 2nd artificial intelligence algorithm 94 is trained by the label-attached positive training data 92A and the label-attached negative training data 92B.

[0160] wherein, in the case of Figures 15 to 25 In the case of the example shown in 2-2 above, only the bright field image is displayed, and in the case of using a plurality of channels as shown in 2-2 above, when a plurality of fluorescent markers are imaged, the positive training data 91A and the negative training data 91B are obtained for each channel, the label-attached positive training data 92A and the label-attached negative training data 92B are generated for each, and the 2nd artificial intelligence algorithm 94 is input.

[0161] wherein the label-attached positive training data 92A and the label-attached negative training data 92B are collectively referred to as the training data 92.

[0162] The 2nd artificial intelligence algorithm 94 is trained by the training data 92, and a trained 2nd artificial intelligence algorithm 97 is generated.

[0163]

(2) Generation of data for analysis and cell analysis

[0164] In the case of Figure 15 (B) shows a cell analysis method using the trained 2nd artificial intelligence algorithm 97, in which the 3rd analysis image 95 of the cells flowing through the flow cell 110 is imaged, and the analysis data 96 is generated. The trained 2nd artificial intelligence algorithm 97 generates data 98 showing the characteristics of the cells to be analyzed using the analysis data 96. As shown in Figure 16 (B) shows a cell analysis method using the trained 2nd artificial intelligence algorithm 97, in which the 3rd analysis image 95 of the cells flowing through the flow cell 110 is imaged, and the analysis data 96 is generated. The trained 2nd artificial intelligence algorithm 97 generates data 98 showing the characteristics of the cells to be analyzed using the analysis data 96. As shown in

[0165] By inputting the analysis data 96 to the trained 2nd artificial intelligence algorithm 97, data 98 showing whether the cells to be analyzed are peripheral circulating tumor cells is generated as data showing normal or abnormality of the cells. For example, when it is determined that the cells to be analyzed are not peripheral circulating tumor cells, "1" is output as a label value, and when it is determined that they are peripheral circulating tumor cells, "2" is output as a label value. Instead of the label value, a label such as "none", "yes", "normal", or "abnormal" can be output.

[0166] 【4. Cell analysis system】

[0167] Hereinafter, the following will be used Figure 18 The cell analysis system 1000, 2000, 3000 related to the first to third embodiments will be described. In the following description, the first artificial intelligence algorithm 50, the first artificial intelligence algorithm 53, and the second artificial intelligence algorithm 94 are sometimes referred to as "artificial intelligence algorithm" without distinction.

[0168] 【4-1. First embodiment of the cell analysis system】

[0169] In Figure 17 The configuration of the hardware of the cell analysis system 1000 related to the first embodiment is shown. The cell analysis system 1000 can be provided with a training device 200A for training the artificial intelligence algorithm 94, a cell camera 100A, and a cell analysis device 400A. The cell camera 100A and the cell analysis device 400A are communicably connected. In addition, the training device 200A and the cell analysis device 400A can be connected by a wired or wireless network.

[0170] 【4-1-1. Training device】

[0171]

(1) Configuration of hardware

[0172] The configuration of the hardware of the training device 200A will be described using Figure 18 The configuration of the hardware of the training device 200A will be described using

[0173] The control section 20A has a CPU (Central Processing Unit) 21 that performs data processing described later, a memory 22 used in a work area of the data processing, a storage section 23 that records programs and processing data described later, a bus 24 that transmits data between the sections, an interface (I / F) section 25 that performs input and output of data with external machines, and a GPU (Graphics Processing Unit) 29. An input section 26 and an output section 27 are connected to the control section 20A via the I / F section 25. Illustratively, the input section 26 is an input device such as a keyboard or a mouse, and the output section 27 is a display device such as a liquid crystal display. The GPU 29 functions as an accelerator that assists the calculation processing (for example, parallel calculation processing) performed by the CPU 21. In the following description, the processing performed by the CPU 21 refers to processing performed by the CPU 21 using the GPU 29 as an accelerator. Among them, a chip that is preferable for calculation of a neural network can be mounted instead of the GPU 29. As such a chip, for example, a FPGA (Field-Programmable Gate Array), an ASIC (Application specific integrated circuit), a Myriad X (Intel), and the like can be given.

[0174] In addition, the control section 20A pre-records, for example, in an executable form, a training program for training an artificial intelligence algorithm and an artificial intelligence algorithm before training in the storage section 23 in order to perform processing of each step described later in Figure 18

[0175] In the following description, the processing performed by the control section 20A refers to processing performed by the CPU 21 or the CPU 21 and the GPU 29 based on the programs and the artificial intelligence algorithm stored in the storage section 23 or the memory 22, unless otherwise specified. The CPU 21 temporarily stores necessary data (intermediate data during processing, and the like) in the memory 22 as a work area, and appropriately records data for long-term storage of calculation results and the like in the storage section 23.

[0176]

(2) Function Configuration of Training Device

[0177] The function configuration of the training device 200A is shown in Figure 18 The training device 200A has a training data generation section 201, a training data input section 202, an algorithm update section 203, a training data database (DB) 204, and an algorithm database (DB) 205. Figure 18 ​Step S11 shown in (A) and Figure 18 Step S111 shown in (B) corresponds to the training data generation section 201. Figure 18 Step S12 shown in (A) and Figure 18 Step S112 shown in (B) corresponds to the training data input section 202. Figure 18 Step S14 shown in (A) corresponds to the algorithm update section 203.

[0178] The training images 70PA, 70PB, 70NA, 70NB, 75P1 to 75Px, 75N1 to 75Nx, 90A, 90B are pre-acquired from the cell camera 100A by the cell analysis device 400A, and are pre-stored in the storage section 23 or the memory 22 of the control section 20A of the training device 200A. The training device 200A can acquire the training images 70PA, 70PB, 70NA, 70NB, 75P1 to 75Px, 75N1 to 75Nx, 90A, 90B from the cell analysis device 400A through a network, or can acquire them through a medium drive D98. The training data database (DB) 204 stores the generated training data 73, 78, 92. The artificial intelligence algorithm before training is pre-stored in the algorithm database 205. The trained first artificial intelligence algorithm 60 can be recorded in the algorithm database 205 in correspondence with the inspection items and analysis items for checking chromosomal abnormalities. The trained first artificial intelligence algorithm 63 can be recorded in the algorithm database 205 in correspondence with the inspection items and analysis items for checking peripheral circulating tumor cells. The trained second artificial intelligence algorithm 97 can be recorded in the algorithm database 205 in correspondence with the items of the feature amounts to be input.

[0179]

(3) Training processing

[0180] The control section 20A of the training device 200A performs Figure 18 the training processing shown in (C).

[0181] First, the CPU 21 of the control section 20A acquires the training images 70PA, 70PB, 70NA, 70NB stored in the storage section 23 or the memory 22 in accordance with a request for starting processing by the user. The training images 70PA, 70PB, 70NA, 70NB are used for training the first artificial intelligence algorithm 50, the training images 75P1 to 75Px, 75N1 to 75Nx are used for training the first artificial intelligence algorithm 53, and the training images 90A, 90B are used for training the second artificial intelligence algorithm 94, respectively.

[0182]

i. Training processing of the first artificial intelligence algorithm 50

[0183] The control section 20A in Figure 18 In step Sll of (A), the control section 20A generates positive integrated training data 72P from the positive training images 70PA, 70PB, and negative integrated training data 72N from the negative training images 70NA, 70NB. The control section 20A generates label-attached positive integrated training data 73P or label-attached negative integrated training data 73N by attaching a respective label value 74P or label value 74N corresponding to the positive integrated training data 72P and the negative integrated training data 72N. The label-attached positive integrated training data 73P or the label-attached negative integrated training data 73N is recorded in the storage section 23 as training data 73. The method of generating the label-attached positive integrated training data 73P and the label-attached negative integrated training data 73N is as explained in 2-1 above.

[0184] Next, the control section 20A in Figure 18 In step S12 of (A), the control section 20A inputs the generated label-attached positive integrated training data 73P and the label-attached negative integrated training data 73N to the first artificial intelligence algorithm 50, and trains the first artificial intelligence algorithm 50. The training result of the first artificial intelligence algorithm 50 is accumulated to the degree of training using a plurality of the label-attached positive integrated training data 73P and the label-attached negative integrated training data 73N.

[0185] Next, the control section 20A in Figure 18 In step S13 of (A), the control section 20A determines whether or not the training result is accumulated to a predetermined specified trial number. When the training result is accumulated to the specified trial number ("YES" case), the control section 20A proceeds to the processing of step S14, and when the training result is not accumulated to the specified trial number ("NO" case), the control section 20A proceeds to the processing of step S15.

[0186] In step S14 when the training result is accumulated to the specified trial number, the control section 20A updates the weight w (combines the weight w) of the first artificial intelligence algorithm 50 using the training result accumulated in step S12.

[0187] Next, the control section 20A in step S15 determines whether or not the first artificial intelligence algorithm 50 is trained with a predetermined number of the label-attached positive integrated training data 73P and the label-attached negative integrated training data 73N. When trained with the predetermined number of the label-attached positive integrated training data 73P and the label-attached negative integrated training data 73N ("YES" case), the training processing is ended. The control section 20A stores the trained first artificial intelligence algorithm 60 in the storage section 23.

[0188] When the first artificial intelligence algorithm 50 is not trained with a specified number of label-attached positive integrated training data 73P and label-attached negative integrated training data 73N (in the case of "NO"), the control unit 20A proceeds to steps S15 to S16 and performs the processing of steps S11 to S15 for the following positive training images 70PA, 70PB and negative training images 70NA, 70NB.

[0189] [ii. Training and processing of the first artificial intelligence algorithm 53]

[0190] Control unit 20A in Figure 18 In step S11 of (A), positive integrated training data 77P is generated from positive training images 75P1 to 75Px, and negative integrated training data 77N is generated from negative training images 75N1 to 75Nx. The control unit 20A generates label-attached positive integrated training data 78P or label-attached negative integrated training data 78N based on the respective label values ​​79P and 79N of the positive integrated training data 77P and negative integrated training data 77N. The label-attached positive integrated training data 78P or label-attached negative integrated training data 78N is recorded as training data 78 in the storage unit 23. The method for generating label-attached positive integrated training data 78P and label-attached negative integrated training data 78N is as described in section 2-2 above.

[0191] Next, control unit 20A in Figure 18 In step S12 of (A), the generated label-attached positive integrated training data 78P and label-attached negative integrated training data 78N are input into the first artificial intelligence algorithm 53 to train the first artificial intelligence algorithm 53. The training results of the first artificial intelligence algorithm 53 are accumulated to the training degree using multiple label-attached positive integrated training data 78P and label-attached negative integrated training data 78N.

[0192] Next, control unit 20A in Figure 19 In step S13, it is determined whether the training result has been accumulated for a predetermined number of trials. If the training result has been accumulated for the specified number of trials ("YES"), the control unit 20A proceeds to step S14; if the training result has not been accumulated for the specified number of trials ("NO"), the control unit 20A proceeds to step S15.

[0193] When the training results are accumulated for a specified number of trials, in step S14, the control unit 20A uses the training results accumulated in step S12 to update the weight w of the first artificial intelligence algorithm 53 (combined with the weight w).

[0194] Next, the control section 20A determines whether or not the first artificial intelligence algorithm 53 is trained with the predetermined number of the label-attached positive integrated training data 78P and the label-attached negative integrated training data 78N in step S15. When the first artificial intelligence algorithm 53 is trained with the predetermined number of the label-attached positive integrated training data 78P and the label-attached negative integrated training data 78N ("YES" case), the training processing is ended. The control section 20A stores the trained first artificial intelligence algorithm 63 in the storage section 23.

[0195] When the first artificial intelligence algorithm 53 is not trained with the predetermined number of the label-attached positive integrated training data 78P and the label-attached negative integrated training data 78N ("NO" case), the control section 20A proceeds to step S15 to step S16, and performs the processing of step S11 to step S15 for the following positive training images 75P1 to 75Px and negative training images 75N1 to 75Nx.

[0196] [III. Training processing of the second artificial intelligence algorithm 94]

[0197] The control section 20A determines whether or not the second artificial intelligence algorithm 94 is trained with the predetermined number of the label-attached positive training data 92A and the label-attached negative training data 92B in step S113. When the second artificial intelligence algorithm 94 is trained with the predetermined number of the label-attached positive training data 92A and the label-attached negative training data 92B ("YES" case), the training processing is ended. The control section 20A stores the trained second artificial intelligence algorithm 97 in the storage section 23. Figure 19 (B), the control section 20A generates the positive training data 91A from the positive training images 90A and the negative training data 91B from the negative training images 90B. The control section 20A generates the label-attached positive training data 92A or the label-attached negative training data 92B with the respective label values 93P or 93N corresponding to the positive training data 91A and the negative training data 91B. The label-attached positive training data 92A or the label-attached negative training data 92B is recorded in the storage section 23 as the training data 92. The generation method of the label-attached positive training data 92A and the label-attached negative training data 92B is explained in the above 3.

[0198] Next, the control section 20A determines whether or not the second artificial intelligence algorithm 94 is trained with the predetermined number of the label-attached positive training data 92A and the label-attached negative training data 92B in step S112 of (B). When the second artificial intelligence algorithm 94 is trained with the predetermined number of the label-attached positive training data 92A and the label-attached negative training data 92B ("YES" case), the training processing is ended. The control section 20A stores the trained second artificial intelligence algorithm 97 in the storage section 23. Figure 19

[0199] Next, the control section 20A determines whether or not the second artificial intelligence algorithm 94 is trained with the predetermined number of the label-attached positive training data 92A and the label-attached negative training data 92B in step S113. When the second artificial intelligence algorithm 94 is trained with the predetermined number of the label-attached positive training data 92A and the label-attached negative training data 92B ("YES" case), the training processing is ended. The control section 20A stores the trained second artificial intelligence algorithm 97 in the storage section 23.

[0200] ​When the second artificial intelligence algorithm 94 is not trained with a specified number of labeled positive training data 92A and labeled negative training data 92B (in the case of "NO"), the control unit 20A proceeds to steps S113 to S114 and performs the processing of steps S111 to S113 on the following positive training image 90A and negative training image 90B.

[0201]

(4) Training Procedure

[0202] This embodiment includes a computer program for training an artificial intelligence algorithm, which causes the computer to perform the processes of steps S11 to S16 or S111 to S114.

[0203] Furthermore, one embodiment of this invention relates to a program article, such as a storage medium, that stores the aforementioned computer program. That is, the computer program can be stored in a semiconductor memory element such as a hard disk or flash memory, or a storage medium such as an optical disc. The form in which the program is recorded to the storage medium is not limited as long as the training device 200A can read the program. The recording to the storage medium is preferably non-volatile.

[0204] Here, "program" refers not only to programs that can be directly executed by the CPU, but also includes the concepts of source programs, compressed programs, encrypted programs, etc.

[0205] 【4-1-2. Cellular Imaging Device】

[0206] The configuration of the cell imaging device 100A for capturing training images 70, 75, 90 and / or resolving images 80, 85, 95 is shown in the figure. Figure 20 . Figure 21 The cell imaging device 100A shown illustrates an imaging flow cytometer. The operation of the imaging device 100A is controlled by the cell analysis device 400A.

[0207] For example, as described above, a target site is detected using one or more fluorescent dyes in a chromosomal abnormality or a peripheral circulating tumor cell. Preferably, the FISH method detects a target site on a first chromosome and a target site on a second chromosome using two or more fluorescent dyes ("first" and "second" of "modified chromosome" do not refer to the chromosome number, but are inclusive numerical concepts). For example, a probe that hybridizes to a PML locus is a nucleic acid having a sequence complementary to a base sequence of the PML locus, which is labeled with a first fluorescent dye that emits first fluorescence of wavelength λ21 by irradiation of light of wavelength λ11. By using this probe, the PML locus is labeled with the first fluorescent dye. A probe that hybridizes to a RARA locus is a nucleic acid having a sequence complementary to a base sequence of the RARA locus, which is labeled with a second fluorescent dye that emits second fluorescence of wavelength λ22 by irradiation of light of wavelength λ12. By using this probe, the RARA locus is labeled with the second fluorescent dye. A nucleus is stained with a nuclear staining dye that emits third fluorescence of wavelength λ23 by irradiation of light of wavelength λ13. Wavelength λ11, wavelength λ12, and wavelength λ13 are so-called excitation light. In addition, wavelength λ114 is light irradiated from a halogen lamp or the like for performing bright field observation.

[0208] The cell imaging apparatus 100A includes a flow cell 110, light sources 120 to 123, condenser lenses 130 to 133, dichroic mirrors 140 to 141, a condenser lens 150, an optical unit 151, a condenser lens 152, and an imaging section 160. The sample 10 is caused to flow through a flow path 111 of the flow cell 110.

[0209] The light sources 120 to 123 irradiate light to the sample 10 flowing through the flow cell 110 from the bottom to the top. The light sources 120 to 123 are constituted by, for example, semiconductor laser sources. The light sources 120 to 123 each emit light of wavelengths λ11 to λ14.

[0210] The condenser lenses 130 to 133 each condense the light of wavelengths λ11 to λ14 emitted from the light sources 120 to 123. The dichroic mirror 140 transmits the light of wavelength λ11 and refracts the light of wavelength λ12. The dichroic mirror 141 transmits the light of wavelengths λ11 and λ12 and refracts the light of wavelength λ13. Thereby, the light of wavelengths λ11 to λ14 is irradiated to the sample 10 flowing through the flow path 111 of the flow cell 110. Further, the number of the semiconductor laser sources included in the cell imaging apparatus 100A is not limited as long as it is one or more. The number of the semiconductor laser sources can be selected from, for example, one, two, three, four, five, or six.

[0211] When light of wavelengths λ11 to λ13 is irradiated to the sample 10 flowing through the flow cell 110, fluorescence occurs from the fluorescent dye labeling the cells flowing through the flow path 111. Specifically, when light of wavelength λ11 is irradiated to the first fluorescent dye labeling the PML locus, first fluorescence of wavelength λ21 occurs from the first fluorescent dye. When light of wavelength λ12 is irradiated to the second fluorescent dye labeling the RARA locus, second fluorescence of wavelength λ22 occurs from the second fluorescent dye. When light of wavelength λ13 is irradiated to the nuclear-staining dye staining the nucleus, third fluorescence of wavelength λ23 occurs from the nuclear-staining dye. When light of wavelength λ14 is irradiated to the sample 10 flowing through the flow cell 110, the light of wavelength λ14 transmits through the cells. The transmitted light of wavelength λ14 transmitted through the cells is used in generation of the bright-field image. For example, in an embodiment, the first fluorescence is a wavelength region of green light, the second fluorescence is a wavelength region of red light, and the third fluorescence is a wavelength region of blue light.

[0212] The condenser lens 150 condenses the first to third fluorescences occurring from the sample 10 flowing through the flow path 111 of the flow cell 110 and the transmitted light transmitted through the sample 10 flowing through the flow path 111 of the flow cell 110. The optical unit 151 has a configuration of four dichroic mirror combinations. The four dichroic mirrors of the optical unit 151 reflect the first to third fluorescences and the transmitted light at slightly different angles from each other, and separate on the light-receiving surface of the imaging section 160. The condenser lens 152 condenses the first to third fluorescences and the transmitted light.

[0213] The imaging section 160 is configured of a TDI (Time Delay Integration) imaging device. The imaging section 160 photographs the first to third fluorescences and the transmitted light, and outputs fluorescent images corresponding to the first to third fluorescences and a bright-field image corresponding to the transmitted light as imaging signals to the cell analysis device 400A. The photographed images can be color images or gray-scale images.

[0214] In addition, the cell imaging device 100A can also be provided with the pretreatment device 300 as needed.

[0215] The pretreatment device 300 samples a portion of a subject, and performs FISH, immunostaining, or organelle staining on cells contained in the subject, to prepare the sample 10.

[0216] 【4-1-3. Cell analysis device】

[0217]

(1) Configuration of hardware

[0218] Use Figure 21The configuration of the hardware of the cell analysis device 400A will be described. The cell analysis device 400A is communicably connected with the cell camera device 100A. The cell analysis device 400A includes a control section 40A, an input section 46, and an output section 47. In addition, the cell analysis device 400A is connectable with the network 99.

[0219] The configuration of the control section 40A is the same as that of the control section 20A of the training device 200A. Here, the CPU 21, the memory 22, the storage section 23, the bus 24, the I / F section 25, and the GPU 29 in the control section 20A of the training device 200A are each read as a CPU 41, a memory 42, a storage section 43, a bus 44, an I / F section 45, and a GPU 49, respectively. However, the storage section 43 stores the trained artificial intelligence algorithms 60, 63, and 94 generated by the training device 200A, which are acquired by the CPU 41 from the I / F section 45 via the network 99 or via the medium drive D98.

[0220] The analysis images 80, 85, and 95 acquired by the cell camera device 100A are stored in the storage section 43 or the memory 42 of the control section 40A of the cell analysis device 400A.

[0221] (2) Functional Configuration of Cell Analysis Device

[0222] In Figure 21 The functional configuration of the cell analysis device 400A will be described. The cell analysis device 400A includes an analysis data generation section 401, an analysis data input section 402, an analysis section 403, an analysis data database (DB) 404, and an algorithm database (DB) 405. Figure 21 Step S21 shown in the flowchart of FIG. 19 corresponds to the analysis data generation section 401. Figure 21 Step S22 shown in the flowchart of FIG. 19 corresponds to the analysis data input section 402. Figure 21 Step S23 shown in the flowchart of FIG. 19 corresponds to the analysis section 403. The analysis data database 404 stores the analysis data 82, 88, and 96.

[0223] The trained first artificial intelligence algorithm 60 can be recorded in the algorithm database 405 in correspondence with the inspection items and the analysis items for checking chromosomal abnormalities. The trained first artificial intelligence algorithm 63 can be recorded in the algorithm database 405 in correspondence with the inspection items and the analysis items for checking peripheral circulating tumor cells. The trained second artificial intelligence algorithm 97 can be recorded in the algorithm database 405 in correspondence with the items of the feature quantities to be input.

[0224] (3) Cell Analysis Process

[0225] The control section 40A of the cell analysis device 400A performs Figure 21The cell analysis process shown in FIG. 2-1 becomes easy to perform with high precision and at high speed.

[0226] The CPU 41 of the control section 40A starts the cell analysis process according to a request from the user to start the process or as a trigger to start the analysis by the cell camera 100A.

[0227] [ i. Cell analysis process by the first artificial intelligence algorithm 60 ]

[0228] The control section 40A generates the integrated analysis data 82 from the analysis images 80A, 80B in step S21 shown in FIG. 2-1. The method of generating the integrated analysis data 82 is as explained in the above 2-1. The control section 40A stores the generated integrated analysis data 82 in the storage section 43 or the memory 42. Figure 21 The control section 40A inputs the integrated analysis data 82 generated in step S21 to the first artificial intelligence algorithm 60 in step S22 shown in FIG. 2-1 by calling the trained first artificial intelligence algorithm 60 stored in the storage section 43 to the memory 42.

[0229] Figure 21 The control section 40A judges the properties of the analysis target cells in the analysis images 80A, 80B using the first artificial intelligence algorithm 60 in step S23 shown in FIG. 2-1, and stores the label values 84 of the judgment results in the storage section 43 or the memory 42. The judgment method is as explained in the above 2-1.

[0230] The control section 40A judges whether all the analysis images 80A, 80B are judged in step S24 shown in FIG. 2-1, and proceeds to step S25 to store the judgment results of the label values 84 corresponding to the judgment results in the storage section 43 while outputting the judgment results to the output section when all the analysis images 80A, 80B are judged (“YES” case). In step S24, when all the analysis images 80A, 80B are not judged (“NO” case), the control section 40A updates the analysis images 80A, 80B in step S26, and repeats steps S21 to S24 to judge all the analysis images 80A, 80B. The judgment results can be the label values themselves, or labels corresponding to each label value such as “yes”, “no”, “abnormal”, “normal”, etc. Figure 21 [ ii. Cell analysis process by the first artificial intelligence algorithm 63 ]

[0231] Figure 21 The control section 40A judges whether all the analysis images 80A, 80B are judged in step S24 shown in FIG. 2-1, and proceeds to step S25 to store the judgment results of the label values 84 corresponding to the judgment results in the storage section 43 while outputting the judgment results to the output section when all the analysis images 80A, 80B are judged (“YES” case). In step S24, when all the analysis images 80A, 80B are not judged (“NO” case), the control section 40A updates the analysis images 80A, 80B in step S26, and repeats steps S21 to S24 to judge all the analysis images 80A, 80B. The judgment results can be the label values themselves, or labels corresponding to each label value such as “yes”, “no”, “abnormal”, “normal”, etc.

[0232] [ ii. Cell analysis process by the first artificial intelligence algorithm 63 ]

[0233] The control section 40A judges whether all the analysis images 80A, 80B are judged in step S24 shown in FIG. 2-1, and proceeds to step S25 to store the judgment results of the label values 84 corresponding to the judgment results in the storage section 43 while outputting the judgment results to the output section when all the analysis images 80A, 80B are judged (“YES” case). In step S24, when all the analysis images 80A, 80B are not judged (“NO” case), the control section 40A updates the analysis images 80A, 80B in step S26, and repeats steps S21 to S24 to judge all the analysis images 80A, 80B. The judgment results can be the label values themselves, or labels corresponding to each label value such as “yes”, “no”, “abnormal”, “normal”, etc. Figure 21 ​​In step S21 shown in FIG. 2, integrated analysis data 87 is generated from the analysis images 85T1 to 85Tx. The method of generating the integrated analysis data 87 is as explained in the above 2-2. The control section 40A stores the generated integrated analysis data 87 in the storage section 43 or the memory 42.

[0234] In step S22 shown in FIG. 2, the control section 40A acquires the trained first artificial intelligence algorithm 63 accommodated in the storage section 43 by inviting the first artificial intelligence algorithm 63 to the memory 42, and inputs the integrated analysis data 87 generated in step S21 to the first artificial intelligence algorithm 63. Figure 21

[0235] In step S23 shown in FIG. 2, the control section 40A judges the properties of the analyzed cells in the analysis images 85T1 to 85Tx using the first artificial intelligence algorithm 63, and stores the label values 88 of the judgment results in the storage section 43 or the memory 42. The judgment method is as explained in the above 2-2. Figure 21

[0236] In step S24 shown in FIG. 2, the control section 40A judges whether all of the analysis images 85T1 to 85Tx are judged, and when all of the analysis images 85T1 to 85Tx are judged ("YES" case), proceeds to step S25, and stores the judgment results of the label values 88 corresponding to the judgment results in the storage section 43 while outputting the judgment results to the output section. In step S24, when all of the analysis images 85T1 to 85Tx are not judged ("NO" case), the control section 40A updates the analysis images 85T1 to 85Tx in step S26, and repeats steps S21 to S24 to judge all of the analysis images 85T1 to 85Tx. The judgment results can be the label values themselves, or labels corresponding to each label value such as "present", "absent", "abnormal", "normal", and the like. Figure 21

[0237]

iii. Cell analysis processing by second artificial intelligence algorithm 97

[0238] In step S21 shown in FIG. 3, the control section 40A generates analysis data 96 from the analysis image 95. The method of generating the analysis data 96 is as explained in the above 3. The control section 40A stores the generated analysis data 96 in the storage section 43 or the memory 42. Figure 21 In step S22 shown in FIG. 3, the control section 40A acquires the trained second artificial intelligence algorithm 97 accommodated in the storage section 43 by inviting the second artificial intelligence algorithm 97 to the memory 42, and inputs the analysis data 96 generated in step S21 to the second artificial intelligence algorithm 97.

[0239] Figure 22 In step S23 shown in FIG. 3, the control section 40A judges the properties of the analyzed cells in the analysis image 95 using the second artificial intelligence algorithm 97, and stores the label values 98 of the judgment results in the storage section 43 or the memory 42. The judgment method is as explained in the above 3.

[0240] In step S24 shown in FIG. 3, the control section 40A judges whether all of the analysis images 95 are judged, and when all of the analysis images 95 are judged ("YES" case), proceeds to step S25, and stores the judgment results of the label values 98 corresponding to the judgment results in the storage section 43 while outputting the judgment results to the output section. In step S24, when all of the analysis images 95 are not judged ("NO" case), the control section 40A updates the analysis image 95 in step S26, and repeats steps S21 to S24 to judge all of the analysis images 95. The judgment results can be the label values themselves, or labels corresponding to each label value such as "present", "absent", "abnormal", "normal", and the like.​​​​Figure 19 In step S23 shown, the second artificial intelligence algorithm 97 is used to determine the characteristics of the target cells in the analyzed image 95, and the determination result is stored in the storage unit 43 or the memory 42. The determination method is as described in section 3 above.

[0241] Control unit 40A in Figure 23 In step S24, it is determined whether all parsed images 95 have been evaluated. If all parsed images 95 have been evaluated (in the case of "YES"), the process proceeds to step S25, where the label value 98 of the evaluation result is stored in the storage unit 43, and the evaluation result is output to the output unit. In step S24, if not all parsed images 95 have been evaluated (in the case of "NO"), the control unit 40A updates the parsed images 95 in step S26, and repeats steps S21 to S24 until all parsed images 95 are evaluated. The evaluation result can be the label value itself, or a label corresponding to each label value such as "present," "absent," "abnormal," or "normal."

[0242]

(4) Cell analysis procedure

[0243] This embodiment includes a computer program for performing cell analysis, which causes the computer to execute the processes of steps S21 to S26.

[0244] Furthermore, one embodiment of this invention relates to a program article, such as a storage medium, that stores the aforementioned computer program. That is, the computer program can be stored in a semiconductor memory element such as a hard disk or flash memory, or a storage medium such as an optical disc. The form in which the program is recorded to the storage medium is not limited as long as the training device 200A can read the program. The recording to the storage medium is preferably non-volatile.

[0245] Among them, "program" refers not only to programs that can be directly executed by the CPU, but also includes the concepts of source programs, compressed programs, encrypted programs, etc.

[0246] [4-2. Second Implementation of the Cell Analysis System]

[0247] The cell analysis system 2000 involved in the second embodiment, such as Figure 17As shown, the training / analysis device 200B is provided with the cell imaging device 100A and trains the artificial intelligence algorithm and analyzes the cells. The cell analysis system 1000 according to the first embodiment is configured to train the artificial intelligence algorithm and analyze the cells by different computers. In the second embodiment, one computer is configured to train the artificial intelligence algorithm and analyze the cells. The training / analysis device 200B acquires the training images 70PA, 70PB, 70NA, 70NB, 75P1 to 75Px, 75N1 to 75Nx, 90A, 90B and the analysis images 80A, 80B, 85T1 to 85Tx, 95 from the cell imaging device 100A.

[0248] The hardware configuration of the training / analysis device 200B is basically the same as that of the cell analysis device 400A shown in Figure 20 The hardware configuration of the training / analysis device 200B is basically the same as that of the cell analysis device 400A shown in Figure 18 The functions of the training / analysis device 200B will be described. The training / analysis device 200B is provided with a training data generation section 201, a training data input section 202, an algorithm update section 203, an analysis data generation section 401, an analysis data input section 402, an analysis section 403, a training data database (DB) 204, and an algorithm database (DB) 205. The respective functional configurations are basically the same as those of the cell analysis device 400A shown in Figure 18 Figure 18 The hardware configuration of the training / analysis device 200B is basically the same as that of the cell analysis device 400A shown in Figure 18 The step S11 shown in (A) and Figure 18 The step S111 shown in (B) corresponds to the training data generation section 201. Figure 21 The step S12 shown in (A) and Figure 21 The step S112 shown in (B) corresponds to the training data input section 202. Figure 21 The step S14 shown in (A) corresponds to the algorithm update section 203. Figure 24 The step S21 shown in (A) corresponds to the analysis data generation section 401. Figure 19 The step S22 shown in (A) corresponds to the analysis data input section 402. Figure 25 The step S23 shown in (A) corresponds to the analysis section 403.

[0249] The training process and the cell analysis process are described in the above 4-1. However, various data generated in the process are stored in the storage section 23 or the memory 22 of the training / analysis device 200B.

[0250] 【4-3. The third embodiment of the cell analysis system】

[0251] The cell analysis system 3000 according to the third embodiment is configured as shown in​Figure 23 As shown, the system includes a cell imaging device 100B, a training device 200C for training an artificial intelligence algorithm and performing cell analysis, a cell imaging device 100A, and an image acquisition device 400B for acquiring images from the cell imaging device 100A. In the cell analysis system 1000 according to the first embodiment, the training of the artificial intelligence algorithm and the analysis of cells are performed by different computers. In the third embodiment, the training device 200C is an example of performing both the training of the artificial intelligence algorithm and the analysis of cells. The training device 200C acquires training images 70PA, 70PB, 70NA, 70NB, 75P1-75Px, 75N1-75Nx, 90A, and 90B from the cell imaging device 100B, and acquires analyzed images 80A, 80B, 85T1-85Tx, and 95 from the image acquisition device 400B.

[0252] The hardware configuration of the training device 200C and the image acquisition device 400B Figure 17 The cell analysis device 400A shown is the same. Using... Figure 20 Explain the functions of training device 200C. The functional structure of training device 200C and... Figure 18 The training / analysis device 200B shown also includes a training data generation unit 201, a training data input unit 202, an algorithm update unit 203, an analysis data generation unit 401, an analysis data input unit 402, an analysis unit 403, a training data database (DB) 204, and an algorithm database (DB) 205. The functional configurations are similar to... Figure 18 and Figure 18 The configuration shown is basically the same. In this embodiment, the training data database (DB) 204 stores training data 73, 78, 92 and parsed data 82, 88, 96. Figure 18 Step S11 shown in (A) and Figure 18 Step S111 shown in (B) corresponds to the training data generation unit 201. Figure 21 Step S12 shown in (A) and Figure 21 Step S112 shown in (B) corresponds to the training data input unit 202. Figure 21 Step S14 shown in (A) corresponds to the algorithm update unit 203. Figure 26 The step S21 shown corresponds to the parsing data generation unit 401. Figure 26 The step S22 shown corresponds to the parsing data input unit 402. Figure 26 The step S23 shown corresponds to the parsing unit 403.

[0253] The training and cell analysis processes are described in section 4-1 above. However, the various data generated during the processing are stored in the storage unit 23 or memory 22 of the training 200C.

[0254] 【5. Other】

[0255] The present application is not construed in a manner limited to the above-described embodiments.

[0256] For example, in the above-described embodiments, different images of the same field of view of the same cell are used in the generation of the training data and the analysis data, but one training data can be generated from one cell image, and one analysis data can be generated from one cell image.

[0257] In addition, in the above-described embodiments, the analysis data is generated from a plurality of images of different light wavelength regions of the same field of view of one cell, but the plurality of images can be obtained by photographing one cell a plurality of times by another method. For example, the analysis data can be generated from a plurality of images obtained by photographing one cell from different angles, or the analysis data can be generated from a plurality of images obtained by photographing one cell at different timings.

[0258] In addition, in the above-described embodiments, the normality or abnormality of the cell is determined, but the type of the cell or the morphology of the cell can be determined.

[0259]

Embodiments

[0260] The embodiments will be described in more detail using the embodiments. However, the present application is not construed in a manner limited to the embodiments.

[0261]

I. Detection of Peripheral Circulating Tumor Cells

[0262] 【1. Data Acquisition Method】

[0263] As a model sample of CTC and blood cells, breast cancer cell line MCF7 and peripheral blood mononuclear cells PBMC (Peripheral Blood Mononuclear Cells) were used. After the cells were stained with a Hoechst reagent, they were supplied to an imaging flow cytometer (ImageStream Mark II, Luminex) to obtain bright field images and nuclear staining images. The conditions of the imaging flow cytometer were set to magnification: 40x, flow rate: Medium, with EDF filter.

[0264] 【2. Analysis】

[0265] 【2-1. Analysis Example by Deep Learning Algorithm】

[0266]

(1) Artificial Intelligence Algorithm

[0267] The language and library used are Python 3.7.3, TensorFlow 2.0a Keras. As an artificial intelligence algorithm, a folding neural network (CNN) is used.

[0268]

(2) Data set

[0269] Details of the data set are shown in Figure 26 (A). Furthermore, in order to use two images of bright field images and nuclear staining images for each 1 cell, the number of images used in the analysis was twice the number of cells. The image size was trimmed to be 32 x 32 pixels. At this time, the cell was cut out in such a way that the center of gravity of the nucleus became the center of the image.

[0270] Training data and analysis data were generated according to the analysis method of peripheral circulating tumor cells using the first artificial intelligence algorithm 63 described herein.

[0271]

(3) Results

[0272] The two-class discrimination of MCF7 and PBMC was performed. First, a discrimination model was created using the training data set. In Figure 13 (B) shows the relationship between the number of Epochs (number of learning times) and the accuracy (correct answer rate). The accuracy reached approximately 100% in less than 10 Epochs. The accuracy was investigated using the model at the 50th Epoch. When the model at the 50th Epoch was used, the accuracy of the training data set was 99.19%, and the accuracy of the validation data set was 99.10%, which was a very good result.

[0273] In Figure 27 (C) shows examples of images in which the answer was correct. Nuc indicates nuclear staining, and BF indicates a phase contrast image.

[0274] 【2-2. Analysis example by machine learning other than deep learning】

[0275]

(1) Artificial intelligence algorithm

[0276] The language and libraries used were Python 3.7.3 and scikit-learn. As the artificial intelligence algorithm, random forest and gradient boosting were used.

[0277]

(2) Data set

[0278] For each of the bright field images and nuclear staining images of the data set shown in Figure 28 (A), the data set defining the 70 kinds of feature amounts (140 kinds of feature amounts combining the bright field images and nuclear staining images) shown in Figure 28 was generated using the analysis software (IDEAS) attached to the imaging flow cytometer.

[0279]

(3) Results

[0280] Discrimination of Class 2 of MCF7 and PBMC was performed. Using the data set described above, a discrimination model was made by random forest and gradient boosting. The hit rate when using each model is shown in Table 1. Figure 29 The hit rate was 99.9% or more in total, and the hit rate was very good for both random forest and gradient boosting.

[0281]

II. Detection of Chromosomal Abnormal Cells

[0282] 【1. Investigation 1】

[0283]

(1). Artificial intelligence algorithm

[0284] Python 3.7.3 and TensorFlow 2.0a were used for language and libraries. As an artificial intelligence algorithm, a convolutional neural network (CNN) was used. Training was performed for 50 times.

[0285]

(2). Data acquisition method

[0286] The PML-RARA chimera gene-positive cells were supplied to the imaging flow cytometer MI-1000 to acquire images of channel 2 (green) and channel 4 (red). The photographing was performed at 60 times magnification with an EDF filter.

[0287] From the image group of channel 2 (green) and channel 4 (red) of the negative control cells judged as having no chromosomal abnormality (G2R2F0) by a known method, negative integrated training data was generated according to the chromosomal abnormal cell analysis method using the first artificial intelligence algorithm 60 described herein. In the negative integrated training data, a "nega label" indicating that the chromosomal abnormality is negative was attached, and the label-attached negative integrated training data was generated. Similarly, from the image group of channel 2 and channel 4 of the positive control cells judged as having a chromosomal abnormality (G3R3F2) by a known method, positive integrated training data was generated. In the positive integrated training data, a "posi label" indicating that the chromosomal abnormality is positive was attached, and the label-attached positive integrated training data was generated. Here, for example, "G" and "R" of G2R2F0 indicate the channel number, and "F" indicates the fusion signal. The numbers indicate the number of signals in one cell.

[0288] The label-attached negative integrated training data was prepared in 3741 groups, and the label-attached positive integrated training data was prepared in 2052 groups, and 3475 groups corresponding to 6 times thereof were used as training data. In addition, 1737 groups corresponding to 3 times were used as test data, and 581 groups corresponding to 1 time were used as validation data.

[0289]

(3) Results

[0290] The hit rate was 100%. In addition, in ​(A) shows the change in loss rate with an increase in the number of Epochs. A decrease in the loss rate was confirmed with an increase in the number of Epochs. In addition, in ​ (B) shows the change in the positive answer rate with an increase in the number of Epochs. The positive answer rate improved with an increase in the number of Epochs.

[0291] 【2. Investigation 2】

[0292]

(1) Artificial intelligence algorithm

[0293] The language and libraries used Python 3.7.3, TensorFlow 2.0a. As an artificial intelligence algorithm, a folding neural network (CNN) was used. Training was performed to 100 times.

[0294]

(2) Data

[0295] The PML-RARA chimera gene-positive subject 3 samples (sample ID: 03-532, 04-785, 11-563) were supplied to the imaging flow cytometer MI-1000 to obtain images of channel 2 (green) and channel 4 (red). The magnification was 60x, and the images were taken with an EDF filter. From the image set of channel 2 and channel 4 of the cells judged to be free of chromosomal abnormalities (G2R2F0) by a known method, negative integration training data was generated according to the method described in the specification. The negative integration training data was attached with a "nega label" showing that the chromosomal abnormality was negative, and the label-attached negative integration training data was generated. Similarly, from the image set of channel 2 and channel 4 of the cells judged to be chromosomal abnormalities (G3R3F2) by a known method, positive integration training data was generated. The positive integration training data was attached with a "posi label" showing that the chromosomal abnormality was positive, and the label-attached positive integration training data was generated. Among them, for example, "G" and "R" of G2R2F0 indicate the channel number, and "F" indicates the fusion signal. The numbers indicate the number of signals in one cell.

[0296] Using the captured images of these subjects, detection of the PML-RARA chimera gene by a deep learning algorithm was attempted. The investigation was performed with 20537 as the number of training data and 5867 as the number of analysis data.

[0297]

(3) Results

[0298] The judgment results in each subject are shown in ​(A) to (C). (A) shows the inference result of test body No. 04-785, (B) shows the inference result of test body No. 03-352, and (C) shows the inference result of test body No. 11-563. The inference result is that the proportion of cells that are correctly judged as positive or negative among the entire analysis data is 92%. In addition, the correct answer rate of each test body is about 90% and no bias is seen. Furthermore, the proportion that becomes false positive or false negative is also 3 to 6% and no bias is seen. From the results, it is considered that a model that does not involve bias of each test body and bias of positive or negative can be generated.

[0299] [Explanation of symbols]

[0300] 10: sample

[0301] 20A, 20B, 20C: control unit

[0302] 50, 53, 60, 63, 94, 97: artificial intelligence algorithm

[0303] 70, 75, 90: training image

[0304] 73, 78, 92: training data

[0305] 74N, 74P: label

[0306] 79N, 79P: label

[0307] 80, 85, 95: analysis target image

[0308] 82, 87, 96: analysis data

[0309] 84: data

[0310] 88: data

[0311] 93N, 93P: label

[0312] 98: data

[0313] 110: flow cell

[0314] 111: flow path

[0315] 120, 121, 122, 123: light source

[0316] 160: imaging unit

[0317] 200A: generation device

[0318] 200B, 200C, 400A: cell analysis device

[0319] 1000, 2000, 3000: cell analysis system

Claims

1. A cell analysis method for analyzing cells using an artificial intelligence algorithm, comprising: passing a sample containing cells in which a target site is labeled through a flow path, capturing a plurality of analysis target images of the same cell by cells passing through the flow path, in mutually different wavelength regions in the same field of view, integrating the plurality of analysis target images generated to generate analysis data, inputting the analysis data generated to the artificial intelligence algorithm, and generating data showing a property of a cell contained in the analysis target image by the artificial intelligence algorithm, wherein the data showing the property of the cell is data showing whether or not it is a cell having a chromosomal abnormality, wherein the target site includes a first gene and a second gene in a nucleus of the cell, wherein the first gene is labeled with a first fluorescent label, and wherein the second gene is labeled with a second fluorescent label different from the first fluorescent label in a fluorescent wavelength region, wherein the plurality of analysis target images include: a first fluorescent image of the first fluorescent label present in the nucleus of the cell, and a second fluorescent image of the second fluorescent label present in the nucleus of the cell, and wherein the analysis data includes tabular data integrating: data showing the brightness of each pixel of the first fluorescent image, and data showing the brightness of each pixel of the second fluorescent image. 2.The cell analysis method according to claim 1, wherein the target site is labeled by an in situ hybridization method. 3.The cell analysis method according to claim 1, wherein generating the analysis target image includes a trimming process of extracting a cell region from an image captured of the cell. 4.The cell analysis method according to claim 1, wherein the artificial intelligence algorithm is a deep learning algorithm having a neural network structure. 5.A cell analysis apparatus for analyzing cells using an artificial intelligence algorithm, comprising a control unit configured to: pass a sample containing cells in which a target site is labeled through a flow path, capture a plurality of analysis target images of the same cell by cells passing through the flow path, in mutually different wavelength regions in the same field of view, integrate the plurality of analysis target images generated to generate analysis data, input the analysis data generated to the artificial intelligence algorithm, generate data showing a property of a cell contained in the analysis target image by the artificial intelligence algorithm, wherein the data showing the property of the cell is data showing whether or not it is a cell having a chromosomal abnormality, wherein the target site includes a first gene and a second gene in a nucleus of the cell, wherein the first gene is labeled with a first fluorescent label, and wherein the second gene is labeled with a second fluorescent label different from the first fluorescent label in a fluorescent wavelength region, wherein the plurality of analysis target images include: a first fluorescent image of the first fluorescent label present in the nucleus of the cell, and a second fluorescent image of the second fluorescent label present in the nucleus of the cell, and The analysis data contains tabular data in which the following data are integrated: data showing the brightness of each pixel of the first fluorescent image, and data showing the brightness of each pixel of the second fluorescent image.

6. A cell analysis system comprising: a flow cell through which a sample containing cells in which a target site is labeled flows, a light source for irradiating a sample flowing through the flow cell with light, an imaging unit that images cells in the sample irradiated with the light, and a control unit, wherein the imaging unit images cells flowing through the flow cell in mutually different wavelength regions of the same field of view to generate a plurality of analysis target images of the same cell, wherein the control unit is configured in such a manner that: the plurality of analysis target images generated by the imaging unit are acquired, the plurality of generated analysis target images are integrated to generate analysis data, the generated analysis data is input to an artificial intelligence algorithm, data showing the properties of cells contained in the analysis target images are generated by the artificial intelligence algorithm, the data showing the properties of the cells are data showing whether or not the cells have chromosomal abnormalities, the target site includes a first gene and a second gene in the nucleus of the cells, the first gene is labeled with a first fluorescent label, and the second gene is labeled with a second fluorescent label that is different from the first fluorescent label in the fluorescent wavelength region, the plurality of analysis target images include: a first fluorescent image of the first fluorescent label present in the nucleus of the cells obtained by imaging the cells, and a second fluorescent image of the second fluorescent label present in the nucleus of the cells obtained by imaging the cells, and wherein the analysis data contains tabular data in which the following data are integrated: data showing the brightness of each pixel of the first fluorescent image, and data showing the brightness of each pixel of the second fluorescent image.

7. A cell analysis program for analyzing cells, which causes a computer to execute processing including the steps of: causing a sample containing cells in which a target site is labeled to flow through a flow path, imaging cells passing through the flow path in mutually different wavelength regions of the same field of view to generate a plurality of analysis target images of the same cell, integrating the plurality of generated analysis target images to generate analysis data, inputting the generated analysis data to an artificial intelligence algorithm, and generating data showing the properties of cells contained in the analysis target images by the artificial intelligence algorithm, wherein the data showing the properties of the cells are data showing whether or not the cells have chromosomal abnormalities, the target site includes a first gene and a second gene in the nucleus of the cells, the first gene is labeled with a first fluorescent label, and the second gene is labeled with a second fluorescent label that is different from the first fluorescent label in the fluorescent wavelength region, the plurality of analysis target images include: a first fluorescent image of the first fluorescent label present in the nucleus of the cells obtained by imaging the cells, and a second fluorescent image of the second fluorescent label present in the nucleus of the cells obtained by imaging the cells, and ​ wherein the analysis data includes tabular data integrating the following data: data showing the brightness of each pixel of the first fluorescent image, and data showing the brightness of each pixel of the second fluorescent image.

Images