Cell type classification / identification using deep learning of cell nuclear staining image

Abstract

The present invention addresses the problem of providing: a device, a method, and a program for easily, quickly, and accurately classifying and / or identifying a cell type of a living or fixed cell, particularly a cell of the nervous system, head, or the like; and a neural network training method for the device, the method, or the program.　In the present invention, a neural network is trained by using two-dimensional and black-and-white training image data based on captured images of stained cell nuclei and training data including information related to cell types corresponding to the cells, thereby making it possible to inexpensively, quickly, and easily classify / identify a cell type with high accuracy on the basis of the image data of the stained cell nuclei for cells contained in a specimen such as cultured cells and tissue sections.

Description

Cell type identification using deep learning in cell nuclear staining images

[0001] This invention relates to a rapid and simple cell type identification and recognition technique using a trained neural network trained on image data of stained cell nuclei. Specifically, it relates to an apparatus, method, program, and method for identifying and / or recognizing living and fixed cells or cell populations contained in a sample, particularly cells or cell populations of the nervous system and head. Multicellular organisms, including humans, maintain their survival through the function and coordination of many types of cells. Therefore, rapid and accurate identification of cell types is extremely important for understanding life phenomena and pathological conditions. However, existing methods for classifying cell types are complex, require many processes and expensive reagents, and are limited in the materials they can be applied to. The inventors therefore considered whether it would be possible to identify cell types using stained DNA images captured with a high-magnification objective lens. The inventors demonstrated that by training a neural network using only two-dimensional, black-and-white training image data based on images of cell nuclei stained with inexpensive and common dyes, and training data containing information about the corresponding cell types, it is possible to identify and distinguish cell types inexpensively, quickly, and easily, based on image data of stained cell nuclei, in both cultured cells and cells contained in tissue sections. Normally, the human eye cannot distinguish cell types from such stained nuclear images. However, as described in detail in the following examples, the inventors achieved an average accuracy of 99.61% in identifying four different types of living cultured cells, including nerve cells, using images of stained cell nuclei. Furthermore, they were able to identify four different types of cells, including nerve cells and cartilage, with an average accuracy of 96.94% in fixed tissue. These results were achieved using two-dimensional images of any single plane (cross-section) of the cell nucleus. Based on the above, it is suggested that in a wide range of cell types, the structure of intranuclear DNA contains cell-type-specific characteristics that can be deciphered with high accuracy using a trained neural network, and the present invention is based on this finding.

[0002] <The Importance of Cell Type Identification> In the field of clinical medicine, if the cell type of cancer cells can be easily and quickly detected in a diseased area, such as in the case of cancer, it is extremely effective for diagnosis and treatment, and is also useful for investigating the cause of the disease. The target cells in this case are mainly fixed cells. Furthermore, in the fields of clinical and basic medicine, if the cell type of cells involved in disease can be easily discovered and identified, it is possible to proliferate and purify the living cells that will be used for treatment, which is effective in verifying the efficacy of therapeutic drugs. In addition, by identifying the cell types that make up the tissue under study, it is possible to investigate the function of normal tissues and cells based on cellular diversity, making it extremely effective for basic research in the life sciences in general.

[0003] Furthermore, in recent years, organoids and nerve cells have been created from human iPS cells and ES cells, and research into the causes of neurological diseases such as Alzheimer's disease and Parkinson's disease, as well as research into therapeutic drugs, is actively progressing using these nerve cells in the process of differentiation or after differentiation. In this context, if it is possible to quickly and accurately identify and isolate nerve cells while they are still alive, it will be easier to investigate culture conditions for efficiently obtaining the target nerve cells, and it will be possible to more accurately investigate the effects of drugs on nerve cells. In particular, in the field of regenerative medicine that uses living adherent cells, there is a strong desire to quickly and accurately identify therapeutic cells and to identify those cells while they are adhered without impairing their original morphological characteristics.

[0004] <Methods for Identifying Cell Types> Conventional methods for identifying cell types include, in the case of fixed cells, attaching fixed tissue sections to a glass slide or similar and staining them. For example, this method uses hematoxylin-eosin (HE) staining to identify cell types, using morphological features of cells observable to the human eye, such as nuclear atypia, mitotic figures, invasion patterns, and cytoplasmic state. However, its application was limited, as it could not classify cell types with similar morphologies. Furthermore, if antibodies or RNA specific to the target cell type exist, there are methods for identification using immunohistochemistry with specific antibodies or in situ hybridization with specific RNA. Recently, with the development of next-generation sequencing technology, it has become possible to classify cell types based on gene expression status by comprehensively analyzing the mRNA expressed by individual cells. However, the existence of specific antibodies or RNA is extremely limited, and both methods require expensive reagents and numerous processes. Furthermore, when the target is living cells, if there is an antibody specific to the target cell type, cell sorting can be performed using that specific antibody, allowing for the detection and purification of a specific cell type, followed by culture. However, in the case of adherent cells, this requires enzymatic treatment of the target cells, detaching them from the culture dish and placing them in suspension, which carries the risk of altering the cell's properties. While separation based on cell size and density is sometimes possible, the applicable materials are limited. Moreover, all of these methods require expensive reagents and numerous processes, and take several hours or more.

[0005] <Cell Analysis Using Deep Learning> A method is known in which fixed cells, such as sections of cancerous tissue, are stained with HE and deep learning is used to determine the primary site of the tumor (Reference 1). Recently, when applied to image analysis, CNN (Convolutional Neural Network), one of the deep learning models, is widely used.

[0006] <Analysis using nuclear staining> There are blue fluorescent dyes such as DAPI (4',6-diamino-2-phenylindole) and Hoechst® 33342 that bind to the minor grooves of the DNA double helix structure, especially the adenine-thymine (AT) rich regions (References 2, 3). The technique of staining nuclear DNA using these dyes has a long history, and has been used since at least the 1980s to investigate the location of the nucleus or to analyze the cell cycle. DNA forms complexes with proteins such as histones and is distributed in the nucleus as chromatin. DAPI and Hoechst® 33342 can strongly stain heterochromatin regions where chromatin is condensed in a short time, from a few minutes to about 30 minutes (References 4, 5). However, this method has not been used to classify cell types such as adherent cells.

[0007] <Analysis of Nuclear DNA Arrangement> The cell nucleus is not merely a static storage place for DNA containing genetic information, but a highly ordered tissue that influences gene expression, DNA replication, and repair (References 6, 7). Recently, technological innovations such as high-throughput chromosome conformation capture (Hi-C) for analyzing the nuclear arrangement of DNA have revealed that the complex spatial structures of chromatin, such as loops, domains, and compartments, are related to gene regulation (References 8, 9). As described above, recent research has provided meaningful insights into the complexity of chromatin interactions and the spatial organization of chromosomes, and detailed analysis at the locus level has advanced significantly. However, the relationship between the arrangement of DNA in the nucleus and cell types from a macroscopic perspective remains largely unclear.

[0008] <Conventional Cell Type Identification Techniques> Conventionally, for suspension cells such as leukocytes of the immune system, a widely known method for identifying cells with high accuracy was to bind them to a fluorescently labeled antibody against an antigen marker on the cell surface and then use a cell sorter or the like. In the case of T cells, a type of leukocyte, a technique for distinguishing them using phase image data, intensity image data, and superimposed image data, and deep learning with CNN (Patent Document 1) was also known. Furthermore, it was known that a method for distinguishing blood cells from tissue cells using a spectral distribution of wavelengths from 500 nm to 800 nm (Patent Document 2), and that adherent cultured cells could be identified as normal cells or cancerous cells using a microscope that quantifies the phase difference of laser light passing through cells (Patent Document 3).

[0009] <Machine Learning-Based Identification Technology Using Nuclear Stained Images> Conventionally, a technique for identifying epithelial cells, fibroblasts, etc., by deep learning on images of colon cancer tissue sections stained with hematoxylin-eosin (HE) has been known (Non-Patent Literature 1). However, techniques for identifying living cells rather than tissue sections are limited, and even with expensive equipment such as cell sorters, the target cells are limited to suspension cells such as hematopoietic cells that have easily distinguishable surface antigens. Recently, it has been reported that deep learning using CNNs, etc., with a massive dataset of nuclear-stained images using Hoechst® 33342, etc., can be applied to identification technology that also targets adherent cells. Specifically, several techniques have been reported, including one that uses human lung cancer-derived cells stained with Hoechst® 33342 to train a CNN to identify virus-infected and uninfected cells and predict virus-infected cells (Non-Patent Literature 2), another that automatically detects cell nuclei from images of human osteosarcoma cells stained with Hoechst® 33342 and provides annotated datasets for training (Non-Patent Literature 3), and a third that observes the nuclei of various living cells stained with Hoechst® 33342 using a confocal microscope and provides excellent indicators of cellular responses such as early apoptosis and necrosis based on their morphology, as well as characterizing the effects of small molecules on the health of cells (Non-Patent Literature 4). However, these are not cell type identification techniques, but rather techniques for distinguishing diseased cells caused by viral infection or aging from normal cells. On the other hand, in the case of suspension cells such as blood cells, there have been reports of attempts to distinguish lymphocyte subtypes (CD3, CD8-expressing T cells, CD20-expressing B cells) using CNN learning on tissue slides stained with Hoechst® 33342 (Non-Patent Documents 5, 6). However, these techniques use a medium-magnification objective lens of 20x, and are only capable of identifying subtypes of suspension cells, and moreover, specific lymphocytes.Furthermore, according to Non-Patent Literature 6, even when distinguishing between two types, CD3-positive and non-CD3-positive, the accuracy is only about 80%, which is insufficient for regenerative medicine, where it is necessary to accurately identify and distinguish the cell type of various types of living cells.

[0010] Therefore, there has been a need for a simple, rapid, and accurate method of identifying and distinguishing cell types that can be applied to a wide range of living cell types. In particular, there is an urgent need to develop a method that can be used for adherent cells such as nerve cells with the high precision required for regenerative medicine.

[0011] Japanese Patent Publication No. 2022-546396, Japanese Patent Publication No. 2020-034551, International Publication No. 2015 / 076311 (WO2015 / 076311)

[0012] K.Sirinukunwattana,et al., IEEE Trans Med Imaging, 35, 1196-1206, 2016V.Andriasyan,et al., iScience, 24, 102543,2021M.Arvidsson,et al., Data in Brief, 46, 108769, 2023A.Tjaden,et al., Molecules, 27,1439, 2022J.Cooper, et al., arXiv, 2107.04388, 2021J.Cooper, et al., Cancers, 14, 5957, 2022

[0013] The first objective of the present invention is to provide a device, method, program, and neural network training method for simple, rapid, and accurate identification and / or identification of cell types, which can be applied to a wide range of cell types of living cells, such as adherent cells like nerve cells and stem cells.

[0014] A second objective of the present invention is to provide an apparatus, method, program, and neural network training method for accurately and rapidly identifying and / or distinguishing cell types such as nerve cells in a fixed tissue sample, or each cell type in a sample containing multiple cell types.

[0015] The present invention aims to provide a device, method, and program for simple, rapid, and accurate identification and / or classification of cell types, applicable to a wide range of cell types, including adherent cells such as nerve cells and stem cells. The inventors focused on the distribution of DNA-binding dyes within the nucleus at the macro level and hypothesized that cell types could be identified simply and rapidly. As a result of diligent research, the inventors completed the present invention by applying a simple yet powerful approach: instead of focusing on the surface structure of cells, such as antigen markers, they trained a neural network with training data containing learning image data based on high-magnification lens images of stained cell nuclei, and information on the cell types corresponding to each image, to accurately identify and classify the cell types of cells in a sample based on image data of cells contained in the sample. In an embodiment illustrating one aspect of the present invention, a convolutional neural network (CNN) (References 10, 11) was trained using image data captured by a two-dimensional confocal microscope of cell nuclei stained with DAPI and / or Hoechst® 33342, from multiple model cells selected from cultured cells or tissue sections of nerve-related cells. In this embodiment, the inventors selected live cultured cells of four different cell types as target models and acquired image data using a 160x high-magnification objective lens after staining with Hoechst® 33342. The CNN model was then adjusted to enable the extraction of complex staining pattern features from each image, and the training image data was increased before training with 8-fold cross-validation. Testing with images not used for training showed that cell type identification was successful with high accuracy. In addition, four different cell types were selected from fixed tissue sections as target models, and after staining with DAPI, training image data was acquired in the same manner as above to construct a CNN model. As a result, we succeeded in identifying the cell type of the tested cells with high accuracy, similar to the case with cultured cells.

[0016] In other words, the present invention is as follows: [1] An apparatus for identifying and / or specifying the cell type of cells contained in a sample, comprising: an input unit for sample image data based on images of stained cell nuclei of cells contained in the sample; an analysis unit for causing the trained neural network, which has been trained using training data including training image data based on images of stained cell nuclei of a plurality of cell types taken using a high-magnification objective lens and information on the cell type corresponding to each image, to calculate information on the cell type corresponding to the input sample image data; and an output unit for outputting the calculated information on the cell type. [2] The apparatus according to [1], wherein the trained neural network is a convolutional neural network (CNN). [3] The apparatus according to [1], wherein the cells used in the training image data and / or the cells contained in the sample are living cells. [4] The apparatus according to [1] above, wherein the cell nuclei of the cells used in the training image data and the cells contained in the sample are cell nuclei stained with DAPI (4',6-diamino-2-phenylindole) and / or Hoechst® 33342. [5] The apparatus according to [1] above, wherein the training image data and / or sample image data are based on images captured by a high-resolution two-dimensional confocal microscope equipped with a high-magnification objective lens of 160x or more. [6] The apparatus according to [1] above, wherein the training image data and / or sample image data are obtained by rotating, enlarging / reducing, adjusting brightness, framing, denoising, increasing resolution or other image processing on captured images, or by synthesizing from multiple images captured under different conditions and then performing image processing.[7] A method for identifying and / or determining the cell type of cells contained in a sample, comprising the steps of: inputting sample image data based on images of stained cell nuclei of cells contained in the sample to a trained neural network which has been trained using training data that includes training image data based on images of stained cell nuclei taken with a high-magnification objective lens for each of several types of cells, and information about the cell type corresponding to each image; and causing the trained neural network to calculate information about the cell type corresponding to the input sample image data, and outputting the calculation result. [8] A computer program for identifying and / or distinguishing the cell types of cells contained in a sample, wherein the computer program is a trained neural network, which has been trained using training data that includes training image data based on images of stained cell nuclei taken with a high-magnification objective lens for each of several types of cells, and information about the cell types corresponding to each image, and the computer program causes the trained neural network to receive sample image data based on images of stained cell nuclei of cells contained in the sample as input, calculate information about the cell types corresponding to the input sample image data, and output the calculation result. [9] A method for training a neural network for classifying cells contained in a sample, the training method comprising the step of training a neural network using training image data based on images of stained cell nuclei taken with a high-magnification objective lens for each of several types of cells, and training data that includes information about the cell types corresponding to each image.

[0017] According to the present invention, cells that could not be detected due to the lack of specific detection reagents can be easily identified and identified in a short time, making it a widely applicable and important tool for pathological diagnosis. Furthermore, because it can identify and identify living cells, it opens the way for the analysis of patient-derived cells and cells differentiated from human iPS cells, and for the search for effective therapeutic drugs using these cells. Moreover, it becomes possible to identify and identify only therapeutically effective cells from a population of patient-derived cells or cells differentiated from human iPS cells for transplantation and other treatments, which could bring about a major revolution in medicine.

[0018] Furthermore, according to the present invention, it has become possible to classify cell types using only a two-dimensional image taken of any single plane (cross-section) of the cell nucleus. This suggests that even a part of the three-dimensional structure of the cell nucleus may have characteristics specific to each cell type.

[0019] The inventors' approach also demonstrated effectiveness in the analysis of live-cell imaging data. This enabled the identification and recognition of cell types without the need for fixation procedures that could potentially damage the original nuclear structure. This opens new avenues for research analyzing the dynamic changes in nuclear structure during processes such as cell differentiation and disease progression (Reference 12). Combining live-cell imaging with the inventors' approach can provide valuable insights into how nuclear DNA arrangement changes over time in response to various stimuli, such as drug administration, enabling rigorous analysis of drug efficacy. Furthermore, it is expected that incorporating super-resolution microscopy and electron microscopy techniques will further improve the resolution and accuracy of this method, enabling more detailed identification and recognition of cell types (Reference 13).

[0020] In particular, in the field of handling nerve cells, we have been able to provide a foundational technology for accurately and rapidly identifying, separating, and isolating nerve cells from progenitor cells in the process of differentiation while they are still alive and healthy, which has been a major challenge in the past.

[0021] Furthermore, the contributions of this invention extend beyond cell biology. The ability to rapidly and accurately identify cell types using simple DNA staining methods has the potential to significantly advance various fields. In the field of developmental biology, it is expected to help elucidate the complex processes of cell fate determination. In disease research, it will enable faster and more accurate diagnosis, accelerate the development of targeted therapies, and lead to a deeper understanding of disease mechanisms. The versatility and simplicity of this invention make it well-suited for automation and high-throughput analysis, positioning it as a powerful tool for clinical application and ultimately contributing to advancements in regenerative medicine, transplantation medicine, and personalized medicine.

[0022] Nuclear features under different objective lenses: Left figure, schematic diagram of the cerebrum and section of a 13.5-day-old (E13.5) embryonic mouse. Right figure, images of DAPI-stained nuclei taken with 20x, 63x, and 160x objective lenses. White dashed lines indicate areas photographed at high magnification. The detail of the DNA staining pattern increases with increasing magnification. Scale bars: 20x = 50 μm, 63x = 50 μm, 160x = 20 μm. CNN model overview: An overview of the convolutional neural network (CNN) used in this study is shown. Multiple convolutional layers and a maximum pooling layer were used for feature extraction, with a dropout layer in between to mitigate overfitting and improve generalization ability. The final layer (Fully Connected Layers) applies a softmax function to predict the probabilities of four cell types. Types of cultured cells used to train the CNN model: Left figure, schematic diagram of neuronal differentiation from P19 mouse embryonic tumor cells (P19 cells). P19 cells differentiate into neurons through cell aggregation, retinoic acid treatment, and subsequent culture. Right figure, representative phase-contrast or fluorescence images (Hoechst, NeuO) of P19 cells (P19), neurons, cultured astrocytes from neonatal mice, and NIH3T3 mouse fibroblasts (NIH3T3). The nuclei were stained with Hoechst® 33342. Neurons were visualized by staining with the neuronal cell marker NeuO. Typical example of cultured cell nuclei used to train the CNN model: The nucleus of each cultured cell was placed in the center of the black background image. Scale bar: 2 μm. Confusion matrix showing model performance validated with images not used for training: The values ​​in the figure represent the average classification accuracy (%) of eight models (Tables 1-1 to 1-4) independently trained using 8-fold cross-validation. Each cell type was tested using 200 different nuclear images. Receiver operational characteristic curve (ROC) and area under the curve (AUC) for a typical example of the eight models (Model 6 in Table 1-4): AUC is micro-averaged, TPR represents the true positive rate, and FPR represents the false positive rate. The 95% confidence interval (CI) for AUC is shown in the lower right corner of the graph. All eight models achieved similarly high AUC values ​​(see Tables 1-1 to 1-4).Typical phase-contrast (Phase) or fluorescence (Hoechst, NeuO) images of ES cell-derived neurons: Nuclei were stained with Hoechst® 33342. Neurons were visualized by staining with the neuronal marker NeuO. Typical example of ES cell-derived neuronal nuclei used to test the CNN model: The nucleus of each cultured cell is placed in the center of the black background image. Scale bar: 2 μm. Classification of ES cell-derived neurons using the CNN model: A matrix showing performance, tested using 200 different nuclear images of ES cell-derived neurons. The values ​​in the figure represent the average classification accuracy (%) of eight models independently trained using 8-fold cross-validation. Fixed cerebral sections used to train the CNN model: Left figure, schematic diagram of the cerebrum and section of a 13.5-day-old (E13.5) mouse; right figure, typical example of a fluorescence image of a section. Cells were stained with DAPI or the neuron marker anti-TBR1 antibody (TBR1), the neural stem cell (NPC) marker anti-SOX2 antibody (SOX2), and EdU, which labels cells undergoing DNA replication. Scale bar, 50 μm. Typical high-magnification image of a cerebral section: Asterisks indicate cells double-positive for SOX2 and EdU. Cells undergoing DNA replication were excluded from the training dataset because their DNA arrangement differs from that of normal cells. Fixed head section used to train the CNN model: Left figure, schematic diagram of a 15.5-day-old (E15.5) mouse head and a typical example of a coronal section stained with Alcian blue. Scale bar, 500 μm. Center figure, high-magnification image of the area enclosed by the frame in the left figure. The upper and lower frames indicate the lens and nasal cartilage, respectively. Scale bar, 200 μm. High-magnification images of the areas indicated by the arrows in the right and center figures (Lens at the top, Cartilage at the bottom). Only EdU-negative nuclei were used for training in chondrocytes and lens cells. Scale bar: 10 μm. Typical example of fixed cell nuclei used for training the CNN model: The nucleus of each cell is placed in the center of the black background image. Scale bar: 2 μm. Confusion matrix showing the performance of the model, validated using images not used for training: The values ​​in the figure represent the average classification accuracy (%) of eight models (Tables 2-1 to 2-4) independently trained using 8-fold cross-validation.For testing neural stem cells (NPCs), nerve cells (Neurons), cartilage cells (Cartilage), and lens cells (Lens), 201, 201, 213, and 206 nuclear images were used, respectively. ROC curves and AUC for typical examples of the eight models (Model 7 in Table 2-4) are shown: AUC is micro-averaged, TPR represents the true positive rate, and FPR represents the false positive rate. The 95% CI for AUC is shown in the lower right corner of the graph. All eight models achieved similarly high AUC values ​​(see Tables 2-1 to 2-4).

[0023] 1. Cells (Cell Populations) Targeted by the Present Invention In the present invention, the cell nucleus of living cells or living cells contained in a cell population is stained, and image data based on the stained image is obtained, or fixed cells or tissue sections are attached to a glass slide or the like, and the cell nucleus is stained and observed. As for the former, living cells, it can be applied to suspension cells such as blood cells and immune cells, but it is effective to apply it to adherent cells such as nerve cells, chondrocytes, or tumor cells. Typical cells include nerve cells, especially brain nerve cells. For example, it can be used to non-invasively identify and isolate nerve cells, chondrocytes, etc. from cell populations that are in the process of differentiating or have differentiated from stem cells such as human iPS cells for use in regenerative medicine, or when isolating cancerous cells from normal adherent cells in the process of differentiating. Typical cells of the latter are inflammatory cells and cancer cells in excised pathological tissue, and the tissue sections containing such cells are tissue sections collected as specimens.

[0024] 2. Regarding Biological Microscopes Microscopes can be broadly divided into biological microscopes and metallurgical microscopes. In this invention, since we observe nuclear DNA in thin sections of living cells or fixed biological tissue, a biological microscope is suitably used. In this invention, the most preferred microscope is a confocal microscope used to acquire the main image. However, images acquired under different conditions can also be obtained by using other biological microscopes in combination or as an auxiliary, and these images can be incorporated by combining them with the main image. This will be briefly explained below.

[0025] (Confocal Microscope) This is a type of fluorescence microscopy that uses laser light as the light source to obtain images from which fluorescence generated outside the focal plane has been removed. By continuously observing tomographic images of a specimen, a three-dimensional microscopic image can be created. This confocal microscope is the most suitable microscope for observing nuclear DNA in this invention.

[0026] (Bright-field microscopy / optical microscope) Among the observation methods for observing living cells or thin sections of fixed biological tissue, bright-field microscopy using an optical microscope is the most common method for stained specimens. (Phase-contrast microscopy / phase-contrast microscope) Phase-contrast microscopy is one of the observation methods suitable for observing colorless, transparent specimens and living cells. It involves attaching a phase-contrast objective lens and condenser to the microscope. When illumination light passes through the specimen, the difference in the optical path between the diffracted light path and the direct light path from the illumination source is used to create a contrast of light and dark in the image of a transparent specimen. (Differential interference contrast microscopy / differential interference contrast microscope) Similarly, differential interference contrast microscopy is another observation method suitable for observing colorless, transparent specimens and living cells. It utilizes the phase difference that occurs in the sloped parts of the specimen when illumination light passes through the specimen to create a contrast of light and dark in the image of a transparent specimen. For observation, a differential interference prism and polarizer are attached. (Relief Contrast Observation Method / Relief Contrast Observation Microscope) Relief contrast observation is a suitable observation method for observing living cells in birefringent containers such as plastic petri dishes. It utilizes the change in refractive index due to the unevenness of the specimen as illumination light passes through the specimen to create a contrast of light and dark in the image of a transparent specimen. For observation, an objective lens and condenser for relief contrast observation are attached. (Fluorescence Observation Method / Fluorescence Microscope, Confocal Microscope) Fluorescence observation is used to observe cells and tissues stained and labeled with fluorescent dyes, or specimens with autofluorescence. A high-pressure mercury lamp or laser is used as the light source to irradiate the specimen with excitation light to cause it to emit fluorescence. (Single-Molecule Analysis Method / Total Internal Reflection Microscope) Total internal reflection microscopy is also a type of fluorescence observation method, but it is widely used to observe the dynamics of molecules and intracellular organelles occurring very close to the adhesion surface between cells and coverslips, and fluorescence originating from single molecules on the substrate surface. This fluorescence microscope utilizes evanescent light—the small amount of light that seeps to the opposite side when light undergoes total internal reflection at an interface with different refractive indices—as excitation light, thereby illuminating only the immediate vicinity of the interface. Because a non-diffusing light source is required, laser beams are typically used.(Other) In addition, transmission electron microscopes, which illuminate a specimen with an electron beam and observe the transmitted electrons at a magnified view, and scanning electron microscopes and scanning probe microscopes, which illuminate a specimen with an electron beam and observe the image obtained from the reflected secondary electrons, can also be used.

[0027] 3. About the Objective Lens The objective lens is the part that first creates the image of the observed object and is the most important instrument that determines the optical performance of the microscope. Magnification (M) is the magnification of the intermediate image, which is an inverted real image relative to the specimen. There are very low magnifications (2.5x or less), low magnifications (4x to 10x), medium magnifications (20x to 50x), and high magnifications (up to 100x or more). In this invention, high magnifications, preferably (60x) or (63x) or higher, and more preferably (100x) or (120x) or higher, can be used. Furthermore, when using a confocal microscope, among the high magnifications, objective lenses of (160x) or higher, which are usually only used in total internal reflection microscopes, are particularly preferred. Numerical Aperture (NA) is an element that determines the resolution, depth of field, and brightness of the image. The larger the numerical aperture, the higher the resolution and the brighter the image that can be observed. And, the higher the magnification of the objective lens, the larger the numerical aperture. Working distance (WD) is the distance from the front of the objective lens to the specimen surface when in focus. The working distance decreases with larger numerical apertures of the objective lens. Parfocal distance (PFD) is the distance from the barrel surface of the objective lens to the specimen surface when in focus. There are standards set by the International Organization for Standardization (ISO) for parfocal distance, such as 45 mm, 60 mm, or 75 mm. Parfocal distance is constant regardless of the magnification of the objective lens, so low-magnification objective lenses with a short overall length have a long working distance, while high-magnification objective lenses with a long overall length have a short working distance.

[0028] 4. Staining Methods for Nuclear DNA To stain nuclear DNA, there are nonspecific physical adsorption methods to the proteins that make up chromosomes, such as acetic carmine, acetic orcein, and Giemsa orcein. However, these staining methods also stain things other than chromosomes. Therefore, the following three types are generally used for DNA-specific staining: (1) "Feulgen staining," which chemically stains aldehyde groups, can stain chromosomal DNA almost quantitatively. (2) This is a dye known to specifically adsorb to AT base pairs, and chromosomal regions containing many AT base pairs are stained intensely. Examples include DAPI, Hoechst (Hoechst® 33342), and quinacrine mustard. DAPI and Hoechst are particularly preferred, and in this example, Hoechst and DAPI are used. (3) This is a dye that specifically adsorbs to CG base pairs, and chromosomal regions containing many CG base pairs are stained intensely. Examples include chromomycin A3 and mithramycin.

[0029] Furthermore, there is ethidium bromide (EtBr), which can be inserted into DNA double strands and increase fluorescence intensity by approximately 20 times. However, EtBr has been shown to be mutagenic, so it is not used in this invention, which also targets living cells. However, it can be used in combination with a reagent that removes the mutagenicity of EtBr (Ethidium Bromide Destroyer: manufactured by ALPHAGEN).

[0030] 5. Regarding captured images, it is also preferable to obtain image data by applying image processing techniques to captured images obtained with the above-mentioned biological microscope, such as rotation, enlargement / reduction, brightness adjustment, framing, noise reduction, resolution enhancement, or other image processing. Alternatively, it is also preferable to obtain image data by applying image processing techniques to multiple images taken with the same biological microscope under different conditions such as focal length and light source, or to multiple captured images taken with different microscopes.

[0031] Hereinafter, the present invention will be described more specifically by way of examples. However, the scope of the present invention is not intended to be limited to the following examples. Other terms and concepts in the present invention are based on the meanings of terms commonly used in the art. The techniques used to implement the present invention can be easily and surely implemented by those skilled in the art based on known documents, etc., except for the techniques whose sources are specifically indicated. Also, various analyses were performed by applying the methods described in the instruction manuals, catalogs, etc. of the analytical instruments, reagents, and kits used. The contents described in the technical documents, patent gazettes, and patent application specifications cited in this specification are incorporated herein by reference as the contents of the present invention.

[0032] (Experimental Example 1) Typical Embodiment of the Present Invention A typical embodiment of the present invention relates to a method for quickly, simply, and accurately identifying and / or determining cell types by performing deep learning using a CNN on stained images of nuclear DNA in cells of the nervous system, head, etc. The specific procedure is generally as follows. Here, cells of the nervous system and head are used for explanation, but the present invention is not limited to only these cell types.

[0033] (1-1) Preparation of Experimental Materials Used in Examples (1-1-1) Mice All experiments were performed using ICR mice purchased from CLEA Japan, Inc. (Tokyo). All experimental procedures using these mice were carried out in accordance with the guidelines established by the Animal Experiment Review Committee (Chiba University).

[0034] (1-1-2) Antibodies The following detection antibodies were used. Primary antibodies: Goat anti-SOX2 antibody (1:800, R&D systems, AF2018), Rabbit anti-TBR1 antibody (1:400, abcam, ab31904). Secondary antibodies: Donkey anti-goat IgG antibody conjugated with DyLight488 (1:250, abcam, ab96935), ＴＭ Donkey anti-goat IgG antibody conjugated with DyLight650 (1:250, abcam, ab96938), ＴＭ Donkey anti-rabbit IgG antibody conjugated with DyLight550 (1:250, abcam, ab96920).

[0035] (1-2) Method for preparing the cells used in the examples (culture method) NIH3T3 cells were maintained in Dulbecco's modified Eagle medium (DMEM) (Gibco®, 10566016) supplemented with 10% fetal bovine serum (Gibco®, 10439024), 100 U / ml penicillin, and 100 μg / ml streptomycin (Wako, 168-23191), and cultured at 37°C and 5% CO2. ２ It was cultured in [a specific medium].

[0036] P19 cells were maintained in DMEM supplemented with 10% fetal bovine serum, 100 U / ml penicillin, and 100 μg / ml streptomycin. The cells were maintained in DMEM supplemented with 100 U / ml penicillin and 100 μg / ml streptomycin at 37°C and 5% CO2. ２ Cells were cultured in the following manner. Neuronal differentiation of the cells was performed using a modified method from (Reference 14). Specifically, cells were plated in a petri dish (Falcon®, 351029) and cultured in the aforementioned growth medium supplemented with 1 μM retinoic acid (RA) (Sigma-Aldrich, R2625). The medium was changed after 2 days, and after another 2 days, the cell aggregates were collected, washed with D-PBS (Wako, 045-29795), and then treated with 0.25% Trypsin-EDTA solution (Gibco®, 25200056) at room temperature for 3 minutes. The aggregates were gently dissociated into single cells by pipetting. GlutaMAX ＴＭ The mixture was plated onto a polyethyleneimine (PEI, Sigma-Aldrich, P3143) coated glass base dish (IWAKI, 3970-035) in Neurobasal medium (Gibco®, 21103049) containing supplement (Gibco®, 35050061), B-27® Supplement (Gibco®, 17504044), and brain-derived neurotrophic factor 10 ng / mL (Wako, 020-12913), and heated at 37°C in 5% CO2. ２It was cultured for 10 - 12 days. The PEI-coated dish was prepared by treating a glass-based dish with 0.1% PEI in 0.15 M borate buffer (pH 8.3) at 4 °C overnight and washing it 5 times with sterile water. The medium was changed at 2-day intervals.

[0037] Mouse embryonic stem cells (ES cells) were obtained from the RIKEN BioResource Research Center (AES0139: EB3, Japan) (References 15, 16). The cells were added to a gelatin-coated dish with 1% FBS, 10% KnockOut Serum Replacement (Gibco®, 10828028), 1 mM sodium pyruvate (Gibco®, 11360070), 1×Non-Essential Amino Acids (Gibco®, 11140050), 0.1 mM 2-mercaptoethanol (Gibco®, 21985023), 1000 U / mL recombinant mouse LIF (Sigma-Aldrich, ESG1106), 50 U / ml penicillin, and 50 μg / mL streptomycin, and cultured at 37 °C, 5% CO ２ For neuronal differentiation of the cells, neuronal cell differentiation medium RHB-A® (Takara, Y40001) was used. The cells were plated at a density of 1x10 ４ cells / cm ２ and cultured at 37 °C, 5% CO ２ for 10 - 12 days. The gelatin-coated dish was prepared by treating it with a 0.1% gelatin solution (Sigma-Aldrich, ES-006-B) at 37 °C for 30 minutes and washing it with D-PBS immediately before use. During neuronal cell differentiation, the medium was changed at 1 - 2-day intervals.

[0038] To visualize P19 and ES cell-derived neurons in a live state, NeuroFluor® NeuO (Veritas, ST-01801) was administered to undifferentiated or differentiated cells. Specifically, the cells were incubated for 1 hour in each differentiation medium containing 125 nM NeuO and washed three times with the medium. All types of cells were counterstained with Hoechst® 33342 (Invitrogen, H3570) at 5 μg / ml for 15 minutes, washed three times with the medium, and then incubated for an additional 1-2 hours. Using a stage incubator set (Tokai Hit, STXG-GSI2X-SET), the cells were observed at 37 °C, 5% CO ２ 2.

[0039] For astrocytes, for culturing, the hippocampus was removed from 1-day-old mice, treated with 0.25% trypsin (Gibco®, 15090046), and dissociated by pipetting. The dissociated cells were seeded in a cell culture dish (Falcon®, 353002) and cultured at 3 × 10 ４ cells / cm ２ 2 in astrocyte medium (Gibco®, A1261301) containing 1% penicillin-streptomycin. When the culture reached 90-95% confluence, subculture was performed, and the resuspended cells were seeded in a glass-bottom dish (Iwaki, 3970-035) coated with poly-D-lysine (Merck, P7280 and P6407) at 37 °C, 5% CO ２ 2.

[0040] (Experimental Example 2) Method for preparing tissue sections (staining method) For EdU (5-ethynyl 2'-deoxyuridine) labeling, the body weight of the mice was measured, and 2 hours before tissue collection, PBS (137 mM NaCl, 8.1 mM Na ２ HPO ４ , 2.6 mM KCl, 1.47 mM KH ２ PO ４EdU (100 mg / kg body weight) dissolved in pH 7.4 was injected. E13.5 heads were fixed for 2 hours, and E15.5 heads for 3 hours. All samples were fixed on ice with 4% paraformaldehyde in PBS (PFA). The fixed tissue was then replaced overnight with PBS containing 30% sucrose, embedded in O.C.T. compound (Sakura FineTech Japan, 4583), and 10-12 μm sections were cut onto glass slides using a CM3050S cryostat (Leica Biosystems). Before staining, the sections were incubated in TE buffer (10 mM Tris, 1 mM EDTA, pH 9.0) at 95°C for 1 minute to retrieve the antigen. Next, the sections were treated with blocking buffer (0.5% Skim milk, 0.25% Gelatin, 0.5% Triton-X 100 in TBS (150 mM NaCl, 20 mM Tris, pH 7.6)) at room temperature for 1 hour. Afterward, the sections were incubated overnight at 4°C in blocking buffer with the primary antibody. After washing three times with TBS, EdU staining was performed using the Click-IT® EdU Imaging Kit according to the manufacturer's instructions (Thermo Fisher Scientific, C10340 or C10637). A wash was then performed once with TBS in normal serum washing buffer (NSB) (1% Donkey Serum (Sigma, S30-100ML, 0.05% Triton-X 100 in TBS). After washing four times with TBS, the sections were incubated for 90 minutes in blocking buffer with a secondary antibody containing DAPI (Roche, 10236276001, 4 μg / mL). Cartilage and lens were identified based on anatomical location and morphology without antibody staining. After washing twice with NSB and five times with TBS, coverslips (Matsunami Glass, thickness No. 1S) were mounted on VECTASHIELD® Vibrance Antifade Mounting Medium (Vector Laboratories, H-1700).

[0041] To visualize the nasal cartilage, sections of E15.5 mouse heads were first incubated in 3% acetic acid for 3 minutes, followed by staining with 0.1% Alcian blue (Merck, A5268-10G) in 3% acetic acid (pH 2.5) for 5 minutes. The sections were then washed once with 3% acetic acid, five times with distilled water, and once with TBS, before being mounted on VECTASHIELD® Vibrance.

[0042] (Experimental Example 3) Method for Acquiring Fluorescence and Phase Contrast Images Fluorescence images of cultured cells or tissue sections were acquired using a confocal microscope (Leica, TCS SP8) equipped with a 20x objective lens (NA 0.70), a 63x oil immersion objective lens (NA 1.40), a 160x oil immersion objective lens (NA 1.43), a hybrid detector (HyD), and 405nm, 488nm, 552nm, and 638nm lasers (Figure 1). Fluorescence and phase contrast images of cultured cells were acquired using a confocal microscope (Olympus, FV10i-LIV) equipped with a 10x objective lens (NA 0.40), or an inverted microscope (Olympus, CKX41) equipped with a 40x objective lens (NA 0.55). To obtain black and white images (8-bit, Z-axis spacing 0.56 μm) of tissue sections or cultured cells, a 160x objective lens was used and the images were recorded at 4096 × 4096 pixels.

[0043] (Experimental Example 4) Method for obtaining DNA stained images Multiple nuclei were manually cropped from images acquired by sliding along the Z-axis. For E13.5 cerebrum, neural stem cell nuclei (SOX2 positive and EdU negative) and nerve cell nuclei (TBR1 positive) stained with DAPI were cropped. For E15.5 head, chondrocyte nuclei (EdU negative) and lens cell nuclei (EdU negative) stained with DAPI were cropped. For cultured cells, images of P19 cells, astrocytes, or NIH3T3 cell nuclei stained with Hoechst® 33342, and nerve cell nuclei (NeuO positive) differentiated from P19 cells or ES cells were cropped.

[0044] The extracted images were placed in the center of a black background. Furthermore, new images were created by rotating, scaling, and adjusting the brightness of each image using Keras's ImageDataGenerator, a Python deep learning library, and these were used as the training image dataset.

[0045] (Experimental Example 5) Model Construction and Interpretability CNN is a type of deep learning widely used for image analysis (References 17, 18). The inventors constructed each model by combining multiple convolutional layers (kernel size 3 × 3), a pooling layer (MaxPooling, pool size 2 × 2), and a subsequent dropout layer (Figure 2). Specifically, two convolutional layers were stacked, followed by a pooling layer and a dropout layer. Then, a unit consisting of three convolutional layers, a pooling layer, and a dropout layer was repeated three times. After that, the input image was flattened, and then a fully connected (FC) layer was adopted. An activation function (ReLU) was used for each convolutional layer and FC layer except for the output layer. The output layer used a softmax activation function. The model parameters were a learning rate of 1 × 10⁻⁶. －３ , damping rate 1 × 10 －３ The model was optimized using the Adam Optimizer. Multiclass cross-entropy was used as the loss function and the model was trained for 100 epochs. The model was trained on 7 of the 8 datasets, and evaluated on the remaining dataset using the Scikit-learn Python package in an 8-fold cross-validation (Reference 19). This process was repeated 8 times, with each dataset used once as the validation set, and an averaged performance metric was calculated.

[0046] (Experimental Example 6) In the statistical analysis multi-class classification, the argmax of the predicted probability for each class was defined as the predicted class. To evaluate classification performance, the one-vs-rest approach was used to calculate precision, recall, F-score, and area under the ROC curve (AUC). Furthermore, the micro-mean, macro-mean, and weighted mean of each indicator were calculated using R (The R Foundation; https: / / www.r-project.org / ). Nonparametric bootstrapping with 2,000 samples was used to calculate the 95% confidence interval (CI).

[0047] (Example 1) Preparation of a CNN model using live cultured cells: (1-1) Selection and preparation of CNN model cells: In this example, the following four types of mouse cells were selected and cultured as CNN model cells. The first was P19 embryonic tumor cells (labeled "P19" in the figure), the second was P19-derived nerve cells differentiated by retinoic acid treatment (labeled "Neuron"), the third was cultured astrocytes collected from the hippocampus (labeled "Astrocyte"), and the fourth was NIH3T3 fibroblasts (labeled "NIH3T3") (Figure 3). Here, P19 cells are a model cell line derived from teratomas and are characterized by high proliferative capacity and the ability to differentiate into nerve cells, etc. (Reference 14). Nuclear DNA was stained with Hoechst® 33342, a membrane-permeable DNA fluorescent dye, and images were acquired using a confocal microscope with a 160x objective lens (Figure 4). In addition, to confirm differentiation into nerve cells, the cells were stained with NeuO (Reference 20), a nerve cell marker.

[0048] For the following deep learning, a CNN model was adjusted to enable the extraction of complex features from cell nucleus image data (Figure 2). As described in Experimental Example 5, the model configuration combines 64 convolutional layers with a 3x3 kernel size to capture the complex spatial patterns of stained DNA. During training, an image dataset was created with a similar number of images for each cell type. To mitigate overfitting and improve generalization ability, the image data for training was increased by randomly rotating, resizing, and adjusting the brightness of each image.

[0049] (1-2) Validation of classification accuracy with the CNN model: In this example, we employed eight-fold cross-validation, training the model on seven of the eight image datasets and evaluating it on the remaining one. This process was repeated eight times, using each image dataset once as the validation set, to reliably evaluate how well the model generalizes to unknown image data. The high accuracy of nearly 100% was consistent across all eight models, demonstrating the robustness of the models (Tables 1-1 to 1-4).

[0050]

[0051] Tables 1-1 to 1-4 above show the statistics for eight CNN models used to classify cell types in living cells. These tables show the performance metrics for eight CNN models (Models 1-8) trained using eight-fold cross-validation. A total of 15,015 P19 cells, 15,015 neurons, 14,205 astrocytes, and 15,105 NIH3T3 nuclei images were used to build the models. For each model, precision, recall, F-score, and area under the ROC curve (AUC) were calculated using a one-vs-rest approach, where each cell type is classified relative to all other cell types. These metrics were calculated for each individual cell type and then combined across all cell types using three different averaging methods (micro-average, macro-average, and weighted average). The 95% confidence interval (CI) for each indicator was estimated using the bootstrap method and is shown below the corresponding value.

[0052] (1-3) Model validation with external datasets: To further evaluate the ability to generalize beyond the training data, performance on independent datasets was assessed. The models were tested on external test datasets, and the average performance of all eight models was calculated. Surprisingly, each model maintained high accuracy, achieving average values ​​of 100% in P19 cells, 99.88% in neurons, 99.06% in astrocytes, and 99.50% in NIH3T3 cells (Figure 5). The AUC was 0.9999854, further demonstrating the outstanding performance of this model (Figure 6).

[0053] (1-4) Validation of the model using cell-derived data different from the training data: To investigate the versatility of this model, we examined whether it could identify neurons differentiated from mouse ES cells instead of P19 cells (Figures 7 and 8). The stained images of ES cell-derived neuronal nuclei were identified with a high accuracy of 99.25%, even though they were not used in training (Figure 9). This suggests that this CNN model can extract common features of neurons.

[0054] (Example 2) Verification of the versatility of the CNN model using fixed tissue sections: (2-1) Enrichment of CNN model data by cell types derived from fixed tissue sections: To evaluate whether the CNN model, adjusted in the same manner as in (Example 1), can be implemented in other biological tissues, we attempted to classify cell types in the mouse head. The inventors focused on the E13.5 mouse cerebral cortex and targeted neural stem cells (SOX2-positive) (indicated as "NPC" in the figure) and nerve cells (TBR1-positive) (indicated as "Neuron" in the figure) (Figure 10, Reference 22). In this process, EdU labeling was used to identify nuclei during DNA replication, and nuclei with DNA in an abnormal state during DNA replication were excluded from the training image dataset (Figure 11). In addition, to evaluate whether the CNN model could clearly capture the nuclear characteristics of nerve cells and neural stem cells, two cell types were added from the heads of E15.5 mice: nasal chondrocytes ("Cartilage") and ocular lens cells ("Lens") (Figure 12). Chondrocytes could be identified using Alcian blue, a blue dye that stains acidic mucopolysaccharides such as chondroitin sulfate and hyaluronic acid (Reference 21). Lens cells could be identified as lenses within the eyeball (Reference 22). All nuclear DNA was stained with DAPI, which is widely used in histological studies (Figure 13).

[0055] (2-2) Verification of classification accuracy with the new CNN model: Using the same configuration as the cultured cell-derived CNN model obtained in (Example 1), eight-fold cross-validation was performed. As a result, all eight models consistently achieved an accuracy of nearly 100%, similar to the case of cultured cells, demonstrating robust model performance (Tables 2-1 to 2-4).

[0056]

[0057] Table 2 shows the statistics for eight CNN models used to classify cell types in fixed tissue. This table shows the performance metrics for eight CNN models (Models 1-8) trained using eight-fold cross-validation. For model construction, images of 18,480 chondrocytes, 18,120 lens cells, 18,150 nerve cells, and 18,120 neural stem cell nuclei were used. For each model, precision, recall, F-score, and area under the ROC curve (AUC) were calculated using a one-vs-rest approach, where each cell type is classified relative to all other cell types. These metrics were calculated for each individual cell type and then combined across all cell types using three different averaging methods (micro-average, macro-average, and weighted average). The 95% confidence interval (CI) for each metric was estimated using the bootstrap method and is shown below the corresponding value.

[0058] (2-3) Model validation with external image datasets: Next, the model was tested with external image datasets. The mean accuracy for neural stem cells ("NPC"), nerve cells ("Neuron"), chondrocytes ("Cartilage"), and lens cells ("Lens") was 97.95%, 96.95%, 98.94%, and 93.93%, respectively (Figure 14). The AUC was 0.9986569, showing performance comparable to that of cultured cells (Figure 15).

[0059] <References> 1. Ming Y.Lu, Tiffany Y.Chen, Drew F.K.Williarnson, Melissa Zhao, Maha Shady, Jana Lipkova & Faisal Mahmood. AI-based pathology predicts origins for cancers of unknown primary. Nature 594, 106-110 (2021). 2. Kubista, M., Akerman, B. & Norden, B. Characterization of interaction between DNA and 4',6-diamidino-2-phenylindole by optical spectroscopy. Biochemistry 26, 4545-4553 (1987). 3. Martin, R. F. & Anderson, R. F. Pulse radiolysis studies indicate that electron transfer is involved in radioprotection by Hoechst 33342 and methylproamine. International Journal of Radiation Oncology Biology Physics 42, 827-831 (1998). 4. Niki Kourmouli et al. Heterochromatin and tri-methylated lysine 20 of histone H4 in animals. Journal of Cell Science 117, 2491-2501 (2004). 5. Bucevicius, J., Keller-Findeisen, J., Gilat, T., Hell, S. W. & Lukinavicius, G. Rhodamine-Hoechst positional isomers for highly efficient staining of heterochromatin. Chemical Science 10, 1962 (2019). 6. Bohrer, C. H. & Larson, D.R. The Stochastic Genome and Its Role in Gene Expression. Cold Spring Harbor Perspectives in Biology 13 : a040386 (2021). 7.Misteli, T. Beyond the sequence: cellular organization of genome function. Cell 128, 787-800 (2007). 8.Gibcus, J. H. & Dekker, J. The hierarchy of the 3D genome. Molecular Cell 49, 773-782 (2013). 9.Lieberman-Aiden, E. et al. Comprehensive mapping of long-range interactions reveals folding principles of the human genome. Science 326, 289-293 (2009). 10.LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436-444 (2015). 11.Litjens, G. et al. A survey on deep learning in medical image analysis. Medical Image Analysis 42, 60-88 (2017). 12.Dekker, J. et al. The 4D nucleome project. Nature 549, 219-226 (2017). 13.Nozaki, T. et al. Dynamic Organization of Chromatin Domains Revealed by Super-Resolution Live-Cell Imaging. Molecular Cell 67, 282-293.e287 (2017). 14.Jones-Villeneuve, E. M., Rudnicki, M. A., Harris, J. F. & McBurney, M. W.Retinoic acid-induced neural differentiation of embryonal carcinoma cells. Molecular and Cellular Biology 3, 2271-2279 (1983). 15.Ogawa, K., Matsui, H., Ohtsuka, S. & Niwa, H. A novel mechanism for regulating clonal propagation of mouse ES cells. Genes to Cells : devoted to molecular & cellular mechanisms 9, 471-477 (2004). 16.Niwa, H., Masui, S., Chambers, I., Smith, A. G. & Miyazaki, J. Phenotypic complementation establishes requirements for specific POU domain and generic transactivation function of Oct-3 / 4 in embryonic stem cells. Molecular and Cellular Biology 22, 1526-1536 (2002). 17.Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 542, 115-118 (2017). 18.Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H. & Aerts, H. Artificial intelligence in radiology. Nature Reviews Cancer 18, 500-510 (2018). 19.Fabian Pedregosa, G. V., Alexandre Gramfort , Vincent Michel , Bertrand Thirion Scikit-learn: Machine Learning in Python.Journal of Machine Learning Research 12, 2825-2830 (2011). 20.Er, J. C. et al. NeuO: a fluorescent chemical probe for live neuron labeling. Angewandte Chemie (International ed. in English) 54, 2442-2446 (2015). 21.McBratney-Owen, B., Iseki, S., Bamforth, S. D., Olsen, B. R. & Morriss-Kay, G. M. Development and tissue origins of the mammalian cranial base. Developmental Biology 322, 121-132 (2008). 22.Cvekl, A. & Ashery-Padan, R. The cellular and molecular mechanisms of vertebrate lens development. Development 141, 4432-4447 (2014).

Claims

1. A device for identifying and / or determining the cell type of cells contained in a sample, comprising: an input unit for sample image data based on images of stained cell nuclei of cells contained in the sample; an analysis unit that causes a trained neural network, which has been trained using training data including training image data based on images of stained cell nuclei of multiple types of cells taken with a high-magnification objective lens and information on the cell type corresponding to each image, to calculate information on the cell type corresponding to the input sample image data; and an output unit that outputs the calculated information on the cell type.

2. The apparatus according to claim 1, wherein the trained neural network is a convolutional neural network (CNN).

3. The apparatus according to claim 1, wherein the cells used in the training image data and / or the cells contained in the sample are living cells.

4. The apparatus according to claim 1, wherein the cell nuclei of the cells used in the training image data and the cells contained in the sample are cell nuclei stained with DAPI (4',6-diamino-2-phenylindole) and / or Hoechst® 33342.

5. The apparatus according to claim 1, wherein the learning image data and / or sample image data are based on images captured by a high-resolution two-dimensional confocal microscope equipped with a high-magnification objective lens of 160x or more.

6. The apparatus according to claim 1, wherein the learning image data and / or sample image data are obtained by rotating, scaling, adjusting brightness, framing, denoising, increasing resolution or other image processing on captured images, or by synthesizing and processing multiple captured images under different conditions.

7. A method for identifying and / or determining the cell type of cells contained in a sample, comprising the steps of: inputting sample image data based on images of stained cell nuclei of cells contained in the sample to a trained neural network which has been trained using training data that includes training image data based on images of stained cell nuclei taken with a high-magnification objective lens for each of several types of cells, and information about the cell type corresponding to each image; and causing the trained neural network to calculate information about the cell type corresponding to the input sample image data, and outputting the calculation result.

8. A computer program for identifying and / or determining the cell types of cells contained in a sample, wherein the computer program is a trained neural network, which has been trained using training data that includes training image data based on images of stained cell nuclei taken with a high-magnification objective lens for each of several types of cells, and information about the cell types corresponding to each image, and the computer program causes the trained neural network to receive sample image data based on images of stained cell nuclei of cells contained in the sample as input, calculate information about the cell types corresponding to the input sample image data, and output the calculation results.

9. A method for training a neural network to identify and / or distinguish the cell types of cells contained in a sample, comprising the step of training the neural network using training image data based on images of stained cell nuclei taken with a high-magnification objective lens for each of several types of cells, and training data containing information about the cell type corresponding to each image.

Images