Machine learning-based cell quality analysis method and system therefor

Abstract

The present invention relates to a method and a system for early prediction of cell quality in an organoid production process. More specifically, the present invention relates to a method and a system for early prediction of the differentiation potential of cells using machine learning-based image analysis technology, thereby enabling selection of optimal cells for high-quality organoid production, and to a training method for a prediction model therefor.

Description

Machine learning-based cell quality analysis method, and system for the same

[0001] The present invention relates to a method and system for analyzing cell quality during an organoid production process. More specifically, the invention relates to a method and system that utilizes machine learning-based image analysis technology to acquire fluorescent images capable of evaluating cell quality without cell staining, further predict the differentiation potential of cells at an early stage, thereby enabling the selection of optimal cells for the production of high-quality organoids, and a method for learning a prediction model for this purpose.

[0002] With the recent advancement of stem cell technology, research on organoids, which mimic the structure and function of human organs, is actively underway.

[0003] However, the organoid production process faces the problem of significant quality variability between batches due to the complex and sensitive nature of the stem cell differentiation induction process. To obtain high-quality organoids, it is crucial to accurately assess the cellular state during intermediate differentiation stages and to select and culture only those cells with a high probability of successful differentiation.

[0004] Conventionally, methods such as immunofluorescence staining or flow cytometry have been primarily used to evaluate the differentiation efficiency of these cells. While these methods have the advantage of accurately confirming the expression of specific differentiation markers (e.g., SIX2), they have the disadvantage that cells cannot be reused in the culture process because they may die or be damaged during the fixation or staining process. Furthermore, these analytical processes are time-consuming and labor-intensive, which limits their applicability to real-time quality control.

[0005] Therefore, there is an urgent need for the development of new technologies that can objectively and quantitatively predict cell quality and differentiation potential at an early stage without destroying cells.

[0006] The present invention was designed to solve these problems.

[0007] For reference, this invention was supported by the following national research and development project.

[0008] [Project ID] 2410005238

[0009] [Assignment No.] 20012378

[0010] [Ministry Name] Ministry of Trade, Industry and Energy

[0011] [Name of Project Management (Specialized) Agency] Korea Institute of Industrial Technology Planning and Evaluation

[0012] [Research Project Name] Industrial Technology Alchemist Project

[0013] [Research Project Title] Development of Meta Soft Organ Module Fabrication Technology and Module Assembly Robot System

[0014] [Name of Project Performing Organization] Pohang University of Science and Technology Industry-Academic Cooperation Foundation

[0015] [Research Period] 2024.01.01 ~ 2024.12.31

[0016] The present invention has been devised to solve the problems of the prior art described above, and aims to provide a method for precisely predicting the differentiation state of cells using only images (observation images) that can be obtained without destroying or staining the cells.

[0017] In addition, the present invention aims to provide a system capable of generating a fluorescence image that enables quantitative analysis equivalent to that of actual fluorescence staining by utilizing deep learning-based image conversion technology, and thereby objectively calculating cell quality indicators (e.g., differentiation efficiency index).

[0018] In addition, the present invention aims to provide a model learning method capable of predicting and selecting final quality early using only images from the initial stage by backtracking and learning the differentiation potential in the initial differentiation stage based on the quality prediction results in the final differentiation stage, and a production process optimization method using the same.

[0019] The problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned problems will be clearly understood by those skilled in the art from the description below.

[0020] To solve the above problems, a method for a system according to the present invention to analyze cell quality may include: (a) a step of acquiring an observation image of any cell being cultured—said that the cell is in a state where it has not been stained—; (b) a step of generating a virtual fluorescent image based on the acquired observation image using an image conversion model—said that the image conversion model is a machine learning model trained to convert a brightfield image into a fluorescent image—; and (c) a step of calculating a quality index of the cell being cultured from the generated virtual fluorescent image.

[0021] In addition, in the above method, the cell quality index is a differentiation efficiency index (DEI), and the differentiation efficiency index may be characterized as being calculated as the ratio of the total intensity of the target differentiation marker channel to the total intensity of the cell nucleus indicator marker channel for pixels that exceed a preset brightness threshold in the fluorescence image.

[0022] In addition, in the above method, the image transformation model may be characterized as a model based on a Conditional Generative Adversarial Network (cGAN), comprising a generator that extracts features from the observed image to generate a fluorescent image, and a discriminator that determines the authenticity of the generated image.

[0023] A system for analyzing cell quality according to another embodiment of the present invention comprises a processor and a memory, wherein the memory stores instructions for executing a method for analyzing cell quality and the processor executes the instructions, and the method for analyzing cell quality may include: (a) a step of acquiring an observation image of any cell being cultured—the cell being in a state where it is not stained—; (b) a step of generating a virtual fluorescent image based on the acquired observation image using an image conversion model—the image conversion model being a machine learning model trained to convert a brightfield image into a fluorescent image—; and (c) a step of calculating a quality index of the cell being cultured from the generated virtual fluorescent image.

[0024] Additionally, a method for training an image conversion model for converting a bright-field image of a cell into a fluorescent image according to another embodiment of the present invention may include: (a) acquiring a training dataset in which a training observation image of any cell being cultured and a real fluorescent image obtained by staining a cell corresponding to the training observation image are mutually mapped; (b) inputting the training observation image into the image conversion model to generate a virtual fluorescent image corresponding to the morphological features of the training observation image; and (c) inputting a pair of the training observation image and the virtual fluorescent image, or a pair of the training observation image and the real fluorescent image, into the image conversion model to determine whether the input fluorescent image is a real image or a fake image; and (d) updating the parameters of the image conversion model.

[0025] Additionally, a system according to another embodiment of the present invention comprises a processor and a memory, wherein the memory stores instructions for executing a method for training an image conversion model, and the processor executes the instructions, and the method for training the image conversion model may include: (a) acquiring a training dataset in which a training observation image of any cell being cultured and a real fluorescent image obtained by staining a cell corresponding to the training observation image are mutually mapped; (b) inputting the training observation image into the image conversion model to generate a virtual fluorescent image corresponding to the morphological features of the training observation image; and (c) inputting a pair of the training observation image and the virtual fluorescent image, or a pair of the training observation image and the real fluorescent image, into the image conversion model to determine whether the input fluorescent image is a real image or a fake image; and (d) updating the parameters of the image conversion model.

[0026] Additionally, a method for predicting cell quality for organoid production early in a cell culture process according to another embodiment of the present invention may include: (a) acquiring a first observation image of a culture vessel containing cells at a first time point after inducing organoid differentiation; and (b) predicting the expected cell quality when the cells reach a second time point, which is a time point after the first time point, by utilizing an early prediction model with respect to the first observation image.

[0027] In addition, the cell quality in the above method may be characterized as being a differentiation potential grade expected when the second time point is reached.

[0028] In addition, the early prediction model in the above method may be characterized by being trained using training data that maps the quality of the cell measured at the second time point to the image at the first time point.

[0029] Additionally, the method may further include (c) a step of determining whether to continue culturing the cells in the culture vessel based on the predicted cell quality.

[0030] Additionally, the method may further include: (d) a step of acquiring a second observation image when the cell reaches the second time point; and (e) a step of calculating a cell quality index for the cell based on the second observation image.

[0031] In addition, the above method may further include the step (e), (e1) generating a fluorescent image corresponding to the second observation image by utilizing an image conversion model—the image conversion model being trained to convert the observation image into a fluorescent image—based on the second observation image; and (e2) analyzing the fluorescent image to calculate a cell quality index (differentiation efficiency index).

[0032] In addition, the above method may be characterized in that the first observation image and the second observation image are live-cell bright-field images of unstained cells.

[0033] Meanwhile, a system for predicting cell quality early according to another embodiment of the present invention includes a processor and a memory, wherein the memory stores instructions for executing a method for predicting cell quality early, and the processor executes the instructions, and the method for predicting cell quality early may include: (a) a step of acquiring a first observation image of a culture vessel containing cells at a first time point after organoid differentiation induction; and (b) a step of predicting the expected cell quality when the cells reach a second time point, which is a time point after the first time point, by utilizing an early prediction model with respect to the first observation image.

[0034] In addition, the system is characterized by acquiring the first observation image from an imaging module that automatically scans a cell culture vessel at predetermined time intervals to generate an observation image, and the system may also be characterized by outputting the cell quality predicted from the image acquired at the first time point to a display unit.

[0035] Meanwhile, a method for training an early prediction model according to another embodiment of the present invention may include the step of training an early prediction model to predict a cell quality index using only an observation image at a first time point using a training dataset; wherein the training dataset is characterized by being generated through the following steps: (a) acquiring a first observation image at a first time point and a second observation image at a second time point that is a predetermined time after the first time point, for a cell that is differentiating in a culture space; (b) generating a virtual fluorescence image from the second observation image at the second time point using an image conversion model; (c) analyzing the virtual fluorescence image to calculate a cell quality index for each cell; (d) determining the quality of an arbitrary cell as one of a plurality of grades according to the calculated cell quality index; and (e) constructing a training dataset by mapping the grade of the cell determined at the second time point to the first observation image at the first time point.

[0036] In addition, the above method may be characterized by using a back-labeling method that maps the grade result of the second time point to the first observation image of the first time point.

[0037] Meanwhile, a system for training an early prediction model according to another embodiment of the present invention comprises a processor and a memory, wherein the memory stores instructions for executing a method for training an early prediction model, and the processor executes said instructions, and the method for training the early prediction model includes the step of training an early prediction model to predict a cell quality indicator using only an observation image at a first time point using a training dataset; and the training dataset comprises: (a) a step of acquiring a first observation image at a first time point and a second observation image at a second time point that is a predetermined time after the first time point for a cell that is differentiating in a culture space; (b) a step of generating a virtual fluorescence image from the second observation image at the second time point using an image conversion model; (c) a step of analyzing the virtual fluorescence image to calculate a cell quality indicator for each cell; (d) a step of determining the quality of an arbitrary cell as one of a plurality of grades according to the calculated cell quality indicator; and (e) a step of constructing a training dataset by mapping the cell grade determined at the second time point to the first observation image at the first time point. It can be characterized as being generated through.

[0038] According to the present invention, there is an effect of being able to accurately predict the differentiation state of cells in a non-invasive manner without destroying or staining the cells.

[0039] In addition, according to the present invention, cells used for quality evaluation can be utilized in subsequent processes without being discarded, thereby preventing waste of cell resources and maximizing production yield.

[0040] In addition, according to the present invention, by generating a virtual fluorescent image through deep learning-based image conversion technology and quantitatively analyzing it, the subjectivity of the conventional visual observation method can be eliminated, and objective and highly reproducible quality evaluation results can be provided.

[0041] In addition, according to the present invention, there is an effect of improving the standardization of the organoid production process and quality uniformity.

[0042] In addition, according to the present invention, differentiation potential in the initial differentiation stage can be predicted early based on quality results in the final differentiation stage, and thus cell batches with low differentiation potential or cell batches with high differentiation potential can be selected early, thereby reducing unnecessary consumption of culture media and reagents and saving time and costs in the overall production process.

[0043] In addition, according to the present invention, it is possible to efficiently construct a high-accuracy prediction model by applying a batch image mapping and learning technique at the well level without complex and high-cost technologies such as individual cell tracking.

[0044] The effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art from the description in the claims.

[0045] Figure 1 is a flowchart schematically showing the overall flow of a method for predicting cell quality using an image conversion model according to one embodiment of the present invention.

[0046] FIG. 2 is a block diagram illustrating the internal structure and operating principle of an image conversion model (generator and discriminator) according to an embodiment of the present invention.

[0047] Figure 3 is an example diagram illustrating the concept of calculating the Differentiation Efficiency Index (DEI) defined in the present invention.

[0048] FIG. 4 is a flowchart showing the process of training an image conversion model according to one embodiment of the present invention in steps.

[0049] FIG. 5 is a conceptual diagram illustrating the specific learning process of a deep learning model trained to convert a bright-field image into a fluorescence image according to an embodiment of the present invention.

[0050] FIG. 6 is a diagram showing an overall overview of a cell quality prediction system and a cell culture timeline according to one embodiment of the present invention.

[0051] FIG. 7 shows the process of an early quality prediction method according to another embodiment of the present invention in sequence.

[0052] Figure 8 is an example diagram illustrating the concept of predicting cell quality early and determining whether to continue culture based on an observation image at the first time point.

[0053] FIG. 9 is a diagram showing an example of a display screen that provides a cell quality prediction result to a user according to one embodiment of the present invention.

[0054] Figure 10 is a flowchart showing the step-by-step process of generating a training dataset to train an early prediction model.

[0055] Figure 11 is a graph and chart showing the criteria for determining cell quality grades (Low, Mid, High) according to the differentiation efficiency index (DEI) and the correlation with the organoid formation rate.

[0056] Figure 12 is a conceptual diagram illustrating the correlation between how an image transformation model is utilized to generate training data for an early prediction model.

[0057] Figure 13 is a flowchart illustrating the flow of a cell quality prediction method including a first prediction through an early prediction model and a second verification through an image transformation model.

[0058] Figure 14 is a flowchart that comprehensively illustrates the learning, prediction, and validation processes of an early prediction model based on logistic regression.

[0059] Hereinafter, some embodiments of the present disclosure will be described in detail with reference to the exemplary drawings. In assigning reference numerals to the components of each drawing, the same components may have the same reference numeral as much as possible, even if they are shown in different drawings. Furthermore, in describing the embodiments, if it is determined that a detailed description of related known components or functions may obscure the essence of the technical concept, such detailed description may be omitted. Where terms such as "comprising," "having," or "consisting of" are used in this specification, other parts may be added unless "only" is used. Where a component is expressed in the singular, it may include a plural unless otherwise specified.

[0060] Additionally, terms such as first, second, A, B, (a), (b), etc., may be used to describe the components of the present disclosure. These terms are used merely to distinguish the components from other components, and the nature, order, sequence, or number of the components are not limited by such terms.

[0061] In describing the positional relationship of components, where it is stated that two or more components are "connected," "combined," or "joined," it should be understood that while the two or more components may be directly "connected," "combined," or "joined," they may also be "connected," "combined," or "joined" with other components "intervened." Here, the other components may be included in one or more of the two or more components that are "connected," "combined," or "joined" with one another.

[0062] In describing the temporal flow relationship regarding components, methods of operation, or methods of production, for example, when the temporal or sequential relationship is described using "after," "following," "next," or "before," it may include cases where the relationship is not continuous unless "immediately" or "directly" is used.

[0063] Meanwhile, where numerical values ​​or corresponding information regarding a component are mentioned, even without a separate explicit description, the numerical values ​​or corresponding information may be interpreted as including a margin of error that may occur due to various factors.

[0064] Furthermore, in this detailed description, terms such as embedding and vector may be used interchangeably; an embedding can be defined as a vector or set of vectors that mathematically represents the characteristics of data, such as text or images, in a high-dimensional space. Depending on the context, the term embedding may refer to the computational operation of performing embedding, or to the result obtained from such operation.

[0065]

[0066] FIG. 1 schematically illustrates a machine learning-based cell quality analysis method according to the present invention. The cell quality analysis method according to the present invention is characterized by inputting a bright-field image obtained without destroying or staining the cell into a machine learning-based image transformation model (10) to generate a virtual fluorescent image at a level equivalent to that obtained by performing actual fluorescent staining, and quantitatively analyzing the image to evaluate the quality of the cell. To this end, the present invention utilizes a conditional generative adversarial network (cGAN)-based image transformation model.

[0067] Referring to the drawings, the cell quality analysis method may first include the step of acquiring an observation image (S11). An observation image refers to an image of unfixed cells or cells that have not undergone fluorescent staining treatment. Conventional fixation or staining treatments kill cells, making subsequent follow-up observation and evaluation of the cells impossible; however, unlike conventional methods, the present invention enables observation and quality evaluation while maintaining the cells in a living state.

[0068] Meanwhile, observed images may contain morphological information about the cells. That is, they may include structural and morphological information such as cell density, aggregation, texture, and boundary sharpness, and this morphological information can be utilized as key basic data for evaluating cell quality.

[0069] While there are no methodological limitations on the process of acquiring observation images itself, as long as living cells can be photographed while containing morphological information, this is generally accomplished by photographing the culture vessel in which the cells are being cultured using a microscope. Specifically, this process may include photographing the entire well of the culture vessel as a grid pattern using a microscope, extracting only the cell-containing regions to observe the cells within the wells intensively, correcting the image to enhance sharpness, and resizing the image for input into an image conversion model.

[0070] After step S11, a step (S13) of generating a virtual fluorescence image by inputting the observation image into an image conversion model may be executed. A virtual fluorescence image refers to an image generated through an artificial intelligence model using an observation image (brightfield image) in a state where the cell has not been physically destroyed or chemically stained as input data.

[0071] The image conversion model used in step S11 includes a generator (101) and a discriminator (103) as shown in FIG. 2, and the image conversion model is trained to predict the expression location and intensity of a specific biomarker within a cell from the morphological characteristics of the cell by learning from a pair of input brightfield images and corresponding actual fluorescence images.

[0072] Referring to FIG. 2, the generator (101) can be designed based on a U-Net structure to extract features of an input brightfield image and convert them into a fluorescence image domain. In particular, it can be designed to have an architecture in which a plurality of residual blocks, preferably eight residual blocks, are added to the bottleneck section of the U-Net to learn complex mapping relationships while preserving the morphological features and texture information of the cell.

[0073] The discriminator (103) evaluates how similar the generated fluorescent image is to the actual image. The discriminator (103) can adopt a PatchGAN structure and be trained to evaluate whether the local details and textures are similar to the actual dyed image by dividing the image into small patches.

[0074] For reference, as will be explained in more detail later, the training of the above image transformation model is performed using an adversarial learning method in which a generator (101) and a discriminator (103) compete. The objective function can be defined as a weighted sum of an L1 reconstruction loss that minimizes the pixel-unit difference between the generated image and the actual image, and an adversarial loss that deceives the discriminator.

[0075] After step S13, a step of calculating the quality index of the cultured cells from the virtual fluorescent image generated in the preceding step, or a step of predicting cell quality (S15) may be performed.

[0076] The cell quality index is a numerical representation of how well a cell corresponding to a previously input observation image can differentiate, and preferably refers to the Differentiation Efficiency Index (DEI). The Differentiation Efficiency Index is a numerical representation of how well a cell has differentiated into a progenitor cell of a target organ, and the system according to the present invention can be implemented to calculate the Differentiation Efficiency Index from a previously generated virtual fluorescence image.

[0077] The differentiation efficiency index can be calculated as the ratio of the total fluorescence intensity of the target marker to the total fluorescence intensity of the entire cell nucleus for valid pixels that exceed a preset brightness threshold within the fluorescence image.

[0078] Figure 3 is a diagram illustrating the concept and calculation principle of the differentiation efficiency index. Referring to Figure 3, the virtual fluorescent image generated through the image conversion model can be largely composed of a cell nucleus indicator marker channel (e.g., DAPI) and a target differentiation marker channel (e.g., SIX2). To aid in understanding the invention, the explanation is based on DAPI and SIX2. DAPI is a dye that binds to DNA in the cell nucleus and emits blue fluorescence; it can be used to indicate the distribution and total quantity of all cells in the image and can serve as the denominator that serves as the standard when calculating differentiation efficiency. Additionally, SIX2 is a protein marker that is specifically expressed when a cell becomes a kidney progenitor cell (NPC) during the kidney organoid differentiation process; it is used to indicate the quantity of cells that have actually undergone differentiation in the image and can serve as the numerator when calculating differentiation efficiency.

[0079] The system (100) according to the present invention can calculate a differentiation efficiency index by comparing and analyzing two channels of images on a pixel-by-pixel basis as shown on the right side of FIG. 3. Specifically, the differentiation efficiency index can be calculated by (i) a process of selecting valid pixels that exceed a preset brightness threshold to remove background noise, and (ii) a process of calculating the ratio of the total fluorescence intensity of the target differentiation marker (SIX2) to the total fluorescence intensity of the entire cell nucleus (DAPI) for the selected pixels.

[0080] In addition, the formula for calculating the differentiation efficiency index can be expressed as the following mathematical formula 1.

[0081]

[0082] (WxW: pixel unit size of the fluorescent image; I i,j (SIX2): SIX2 channel pixel intensity at position (i,j); I i,j (DAPI): DAPI channel pixel intensity at position (i, j); α: threshold for specifying brightness for background noise removal)

[0083] With reference to Figures 1 to 3 above, a machine learning-based cell quality analysis method was examined.

[0084]

[0085] Figure 4 illustrates the method of training the previously mentioned image conversion model in sequence.

[0086] Referring to the drawings, the training of the image transformation model may begin with the step (S20) of first obtaining a training dataset, more precisely, a training observation image of a random cell being cultured and a training dataset in which the actual fluorescence image obtained by staining the cell corresponding to the training observation image is mutually mapped. Examples of actual fluorescence images include DAPI (cell nucleus) or SIX2 (kidney precursor marker) stained images, which correspond to the Ground Truth for model training.

[0087] After step S20, a step (S22) can be executed in which a training observation image is input into an image conversion model, or more precisely, into a generator to generate a virtual fluorescence image. As previously explained, the generator corresponds to a configuration that generates a virtual fluorescence image by learning the morphological characteristics of the cell while preserving the context of the original observation image.

[0088] After step S22, a step (S24) may be executed in which a pair of the training observation image and the virtual fluorescence image, or a pair of the training observation image and the actual fluorescence image, is input to the discriminator to determine whether the input fluorescence image is a real image or a fake image. For example, on the left side of FIG. 5, a process is illustrated in which a virtual fluorescence image (Predicted SIX2 or DAPI) generated by inputting an observation image (brightfield image of the NPC stage) into the generator is input to the discriminator in pair with the observation image and is determined to be Fake, and on the right side of FIG. 5, a process is illustrated in which a real fluorescence image (True SIX2 or DAPI) is input to the discriminator in pair with the observation image and is determined to be Real. Step S24 can be understood as a step in which a plurality of training data sets undergo a process of being determined in this manner.

[0089] Meanwhile, after step S24, a step (S26) of updating the parameters of the image transformation model according to the discrimination result may be executed. The parameters of the image transformation model can be understood as a collective term for variables representing the connection strength or activation threshold between each node within the artificial intelligence model. That is, they can be defined as values ​​directly involved in the computational process in which a machine learning model extracts features from training data and transforms input data into desired output data, and in the process of training a general machine learning model, they are updated in a direction that minimizes the loss function value.

[0090] Based on the above, the training of the image transformation model is characterized by an adversarial process in which the generator and the discriminator compete against each other. Specifically, the generator is trained to produce virtual fluorescent images sophisticated enough for the discriminator to determine them as real (minimizing generation loss), and the discriminator is trained to accurately distinguish between the fake images created by the generator and the actual ground truth images (minimizing discrimination loss). During this process, an L1 loss function is also optimized to reduce pixel-level differences between the generated images and the actual images, thereby updating the parameters of the image transformation model.

[0091] We also examined the method of training an image conversion model with reference to Figures 4 and 5.

[0092] For reference, the cell quality analysis method of FIG. 1 and the method of training the image conversion model of FIG. 4 described above can all be executed by a device or a set of devices equipped with a central processing unit and memory, and in this detailed description, such a device or a set of devices will be referred to as a system (100).

[0093] The central processing unit may also be referred to as a controller, microcontroller, microprocessor, or microcomputer. Additionally, the central processing unit can be implemented using hardware, firmware, software, or a combination thereof. When implemented using hardware, it may be an ASIC (Application Specific Integrated Circuit), DSP (Digital Signal Processor), DSPD (Digital Signal Processing Device), PLD (Programmable Logic Device), FPGA (Field Programmable Gate Array), etc. When implemented using firmware or software, the firmware or software may be configured to include modules, procedures, or functions that perform the above functions or operations. In addition, memory can be implemented as ROM (Read Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, SRAM (Static RAM), HDD (Hard Disk Drive), SSD (Solid State Drive), etc.

[0094] In addition, the types of systems may include portable terminals such as smartphones, PDAs, and tablet PCs, as well as terminals permanently placed in a specific location, such as desktop PCs and servers.

[0095] When the above system (100) is implemented as a server, the server may be a device that stores and executes a set of instructions, that is, a program, for actually implementing the method according to the present invention. The form of the server may be at least one server PC managed by a specific user, or it may be a cloud server provided by another company, that is, a cloud server that a user can use after signing up.

[0096] In addition, depending on the case, the method according to the present invention may be executed on a clustering system composed of multiple computing devices rather than a single system (100). Within the clustering system, multiple computing devices required to execute the crack management support method may be configured to perform different operations. Furthermore, in addition to the aforementioned server and clustering system, the system (100) mentioned in this detailed description may also be implemented in an on-device form.

[0097] Meanwhile, the display unit mentioned in this detailed description may include all devices that display various content using a user interface and / or a graphical user interface stored in memory under the control of a central processing unit. Here, the content displayed on the display unit may include various text or image data (including various information data) and menu screens including data such as icons, list menus, and combo boxes. Additionally, the display unit may be a touch screen. The display unit may include a Liquid Crystal Display (LCD), a Thin Film Transistor-Liquid Crystal Display (TFT LCD), an Organic Light-Emitting Diode (OLED), a Flexible Display, a 3D Display, an e-ink Display, etc., and the technology used in the display unit is not limited to those exemplified above.

[0098]

[0099] Meanwhile, the image transformation model described above is characterized by the ability to analyze cell quality at the same point in time by generating a virtual fluorescence image for an observation image acquired at a specific point in time. By utilizing this image transformation model, another machine learning-based model, an early prediction model of cell quality, can be implemented.

[0100] Below, we will explain a method for predicting cell quality early using the above cell quality early prediction model (abbreviated as 'early prediction model').

[0101] FIG. 6 is a schematic diagram for understanding the overall concept of a machine learning-based cell quality early prediction system (100) according to another embodiment of the present invention and the flow of applying it to an organoid differentiation process.

[0102] For example, the production process of kidney organoids is known to include a series of steps starting from human induced pluripotent stem cells (Day 0), passing through late primitive striatum (Day 4) and kidney progenitor cells (Day 9), then into tubular aggregates (Day 11) and finally kidney organoids (Day 24-26). The system (100) according to the present invention aims to increase the efficiency of the organoid production process by predicting whether cells can be cultured while maintaining good quality continuously, based on images obtainable at a specific point in time during the early stages of this differentiation process. In particular, the system (100) according to the present invention aims to enable the prediction of cell quality based on observation images from the fourth day of differentiation (Day 4).

[0103] Referring to FIG. 6, the system (100) can obtain an image (observation image) of the early differentiation stage (e.g., late primitive striatal stage of Day 4), that is, before differentiation into full-fledged elongation progenitor cells, at a time when morphological features that can determine whether future differentiation will be successful appear.

[0104] Afterward, the system (100) can analyze the input image and calculate the quality grade (differentiation potential) that the cell is expected to have when it reaches a future point in time (e.g., the elongation progenitor cell stage of Day 9).

[0105] Additionally, the system (100) can provide the previously predicted results to the user through a network and / or display unit, thereby enabling the user to decide whether to continue or stop the culture of a specific cell during the early stages of differentiation.

[0106] For reference, FIG. 6 conceptually illustrates an example in which an observation image acquired on Day 4 is input into the system (100) and the cell quality of Day 9 is predicted therefrom, and it is understood that this is intended to easily explain an embodiment of the invention. That is, it is understood that the invention described herein is not necessarily limited to a specific time point in which the observation image is acquired, and that the observation image can be acquired at any time point where the morphological characteristics of the cell appear.

[0107] Meanwhile, as briefly mentioned earlier, morphological characteristics of cells may vary depending on the type of cell being cultured and may refer to data that quantifies the visual, spatial, and structural characteristics formed by cells or cell clusters within a culture vessel. Specifically, this may include cell density, aggregation, spacing between clusters, boundary clarity, or cytoplasmic texture.

[0108] For reference, such morphological features of the cell can be defined as descriptors or feature vectors extracted by computer vision algorithms, and as will be explained again later, in this invention, morphological features are extracted and analyzed using SIFT (Scale-Invariant Feature Transform) or ORB (Oriented FAST and Rotated BRIEF) algorithms.

[0109]

[0110] The present invention will be described in more detail below with reference to the drawings, including Figure 7.

[0111] FIG. 7 illustrates a flowchart of a method for early prediction of cell quality according to one embodiment of the present invention.

[0112] Referring to FIG. 7, the prediction method according to the present invention may include a first observation image acquisition step (S110) and a cell quality prediction step (S120), and subsequently may additionally include a step (S130) for determining whether to continue cell culture.

[0113] First, step S110 is a step in which the system (100) obtains a first observation image of a cell that has reached a first time point, which is an early stage after organoid differentiation induction.

[0114] The first time point mentioned in this detailed description refers to the time point at which the morphological features of the cell (progenitor cell) are observed, and the morphological features of the cell may refer to data quantifying the visual, spatial, and structural characteristics formed by the cell or cell cluster within the culture vessel. Specifically, this may include cell density, aggregation, spacing between clusters, boundary clarity, or cytoplasmic texture.

[0115] Meanwhile, the above-mentioned first point in time may be defined not only as whether morphological characteristics of the cell are observed, but also as the point in time when a biomarker specifically expressed in the progenitor cells of the organ is detected or expressed above a threshold value.

[0116] Consequently, the first point in time in the present invention can be defined as a point in time when a differentiated cell exhibits characteristics as a progenitor cell of a specific organ and can be observed, and more precisely, as a point in time when the morphological characteristics of the cell and / or biomarker expression characteristics are observed.

[0117] Meanwhile, the first time point in the cell culture environment for producing kidney organoids, which is the main embodiment mentioned to aid in understanding the present invention, is preferably the late primitive stage before the cells differentiate into kidney progenitor cells, for example, day 4 after differentiation induction. At this time point, the cells are in a state where the morphological characteristics of the cells mentioned earlier can be extracted, and the first observation image may be an image obtained without staining or destroying the cells. Preferably, the first observation image may be a bright-field image captured by an imaging module (200; see FIG. 4), and may also be an image obtained by scanning the entire well plate, which is the cell culture space.

[0118] After step S110, the system (100) inputs the acquired first observation image into an early prediction model to predict the cell quality (S120) when the cell reaches a second time point in the future.

[0119] Here, the second time point refers to a specific time point after the first time point, and may refer to the kidney progenitor cell stage (Day 9), which is known to allow quality determination through the expression of actual differentiation markers during the culture process of kidney organoid cells. The early prediction model may be a model trained to receive image features from the first time point (Day 4) as input and classify the expected cell quality (e.g., differentiation potential) at the second time point (Day 9) into grades.

[0120] FIG. 8 is a conceptual diagram showing a first observation image of a cell obtained on the 4th day of differentiation, which is a first time point, according to one embodiment of the present invention, and a cell quality result predicted by an early prediction model based thereon.

[0121] Referring to FIG. 8, the image on the left is a first observation image (bright-field image) taken during the late primitive striatum stage, which is the fourth day after inducing kidney organoid differentiation. As can be seen in the drawing, the cells at this point in time do not yet have a clear shape of kidney progenitor cells and appear as fine clusters that are difficult to determine with the naked eye whether differentiation will be successful in the future. The system (100) can be implemented to predict the quality grade of the corresponding cells as High by extracting a morphological feature vector from this first observation image and inputting it into a pre-trained early prediction model.

[0122] FIG. 9 illustrates a system (100) according to the present invention outputting the result calculated by step S120 above through a display unit (300). Referring to FIG. 9, the system (100) of the present invention can be implemented to visually display cell-specific quality information analyzed corresponding to each cell culture space location in a well plate on the display unit (300).

[0123] The result screen (a) of FIG. 9 is an example showing the analysis results produced by the system (100), displaying specific numerical values ​​(e.g., 70.32%, 16.23%, etc.) of the differentiation efficiency index (DEI) for each cell calculated based on the first observation image obtained at the initial stage of differentiation. To explain this process more precisely, when the first observation image is obtained (S11) at the initial stage of differentiation, a virtual fluorescent image is generated (S13) based on it, and the differentiation efficiency index (DEI) is calculated from the virtual fluorescent image generated in this way and can be output through the display unit (300). From the perspective of the user (experimenter), one can experience being able to confirm the degree of differentiation of each cell numerically, as shown in the result screen (a), using only the image (first observation image) observed at the initial stage of differentiation.

[0124] Meanwhile, result screen (b) of FIG. 9 is another example of a result screen provided to a user (experimenter). While result screen (a) displays the differentiation efficiency index as a numerical value, result screen (b) is an example that shows a grade prediction result by assigning a grade according to the range to which the numerical value belongs. (Result screen (a) and result screen (b) are result values ​​from separate experiments, respectively.) Specifically, by displaying grades such as High for cells with a very high success rate of organoid formation, Mid for cells capable of organoid formation but not of very good quality, and Low for cells with a high probability of organoid formation failure, the user can intuitively understand the status of the cells being cultured.

[0125] Meanwhile, the prediction in the above S120 step has technical significance in that it not only classifies the cell state at the current time point (first time point) but also predicts whether culture will be successful at a future time point (second time point). Generally, cells in the early stages of culture appear as simple clusters externally, making it difficult to determine the success of final differentiation with the naked eye. However, the present invention enables the early prediction model to statistically learn the potential correlation between the fine morphological features of the first time point and the quality indicators of the second time point through accumulated learning data, thereby making it possible to predict whether culture will be successful at the second time point solely through image analysis at the first time point.

[0126] Furthermore, in this detailed description, to aid in understanding the invention, the time point prior to the progenitor cell (first time point) and the time point of the progenitor cell (second time point) are explained as examples; however, it is understood that the definition of the first time point described above and the definition of the second time point to be described later are not limited to the above embodiments. That is, it is understood that among the various embodiments to which the present invention can be applied, there may exist embodiments where the first time point is defined as the time when the progenitor cell is generated and the second time point is defined as the time when the organoid is formed, or embodiments where the first time point is defined as the time of the stage prior to the progenitor cell and the second time point is defined as the time when the organoid is formed, and various other embodiments may exist.

[0127] Meanwhile, after step S120, the system (100) may execute a step (S130) of determining whether to continue culturing cells in the culture space based on the quality or grade of the cells predicted in the preceding step.

[0128] This step can be implemented in an automatic manner in which the system (100) automatically determines whether to continue culturing the cells according to a predetermined algorithm, or in a manual manner in which the system (100) receives input from a user regarding whether to continue culturing a specific cell.

[0129] In the case of an automatic method, the system (100) can be implemented to decide to continue the culture process for cells with a predicted grade of high or medium, and to stop or discard the culture for cells with a grade of low.

[0130] In the case of a manual method, the system (100) receives input from a user who has checked the quality of the cells through the display unit (300) regarding whether to continue culture for each cell, and can be implemented to proceed to a subsequent process (continuation or discontinuation of culture) based on the input.

[0131] With reference to FIGS. 6 to 9, we have examined the method for early prediction of cell quality according to the present invention and the system (100) for the same.

[0132]

[0133] Meanwhile, in the previous explanation of step S120, it was stated that the early prediction model is a model trained to predict cell quality from the first observation image. Below, we will explain how to train the early prediction model, or more precisely, the process of generating the dataset required to train the early prediction model.

[0134] FIG. 10 is a flowchart illustrating the process of constructing a dataset for training an early prediction model according to an embodiment of the present invention. The method for constructing a training dataset according to the present invention basically uses a method of backtracking and labeling the results of a precise analysis of a future point in time (second point in time) onto an image of a past point in time (first point in time), and the specific steps are as follows. For reference, the dataset construction process illustrated in FIG. 10 can be executed by any computing device equipped with a central processing unit and memory, and is not necessarily executed by the system (100) introduced above; however, for convenience, the situation in which it is executed by the system (100) will be assumed and explained here.

[0135] Referring to FIG. 10, the process of constructing a learning dataset first involves the step (S200) of acquiring a first observation image at a first time point and a second observation image at a second time point. The first observation image and the second observation image may be acquired by the imaging module (200) mentioned above, but are not necessarily limited thereto. Additionally, the observation images may preferably be images acquired in a time-series manner for the same culture vessel after inducing organoid differentiation within a cell culture vessel for the purpose of collecting learning data.

[0136] As a specific example, the system (100) may be implemented to acquire and store a first observation image of the corresponding well when it reaches a first time point (e.g., Day 4), which is an early stage of differentiation, and then to acquire a second observation image of the same well when it reaches a second time point (e.g., Day 9), which is an late stage of differentiation. At this time, the images of the first time point and the second time point do not need to be matched by tracking the location of individual cells, and as long as the information that they originated from the same well is maintained, the individual cells captured at the first time point and the second time point can be mapped to be the same.

[0137] For reference, the description of Fig. 7 above explained how the first time point can be defined, and similarly, here we will explain how the second time point mentioned in this detailed description can be defined.

[0138] The second time point can be defined as the time point at which the organoid characteristics of the target organ are expressed, and this time point can be defined by various methods, such as i) analyzing whether a biomarker is detected, ii) analyzing whether a gene is expressed, iii) analyzing whether the function of the organ is expressed, and iv) analyzing morphological characteristics, either individually or in combination.

[0139] i) Regarding the method of analyzing whether biomarkers are detected, whether a kidney organoid has been successfully formed can be determined at a second time point by observing whether the podocyte marker PODXL, the proximal tubule marker LTL, or the distal tubule marker ECAD is expressed in addition to SIX2, and whether the expression level is above a preset value; and whether a brain organoid has been successfully formed can be determined at a second time point by observing whether markers such as SOX2, Nestin, TUJ1, MAP2, and GFAP are expressed, and whether the expression level is above a preset value. In addition, markers such as Villin, MUC2, and LYZ can be used to determine whether an intestinal organoid has been successfully formed, and markers such as Albumin and HNF4α can be used to determine whether a liver organoid has been successfully formed.

[0140] Next, regarding ii) the method of analyzing gene expression, whether an organoid has been successfully formed can be confirmed by utilizing the results of gene expression analysis (qPCR or RNA-seq), and a second time point can also be determined through this. That is, the second time point can be determined by analyzing RNA extracted from cells at any given time point to measure the expression levels of organ-specific genes. For example, when the expression levels of the WT1 and NPHS1 genes in the kidney, the MAP2 and TUBB3 genes in the brain, and the VILLIN gene in the intestine are above a preset value, it can be confirmed that the corresponding time point corresponds to the second time point, and furthermore, the quality (grade) of the cells can also be classified at that second time point.

[0141] Next, regarding the method of analyzing whether the function of the organ is expressed, whether the organoid has been successfully formed can be confirmed by evaluating whether the cells perform the function of the actual organ, and through this, the second time point can be determined. For example, in the case of a kidney organoid, the success of the organoid formation and the second time point can be determined by confirming the Dextran transport function; in the case of a brain organoid, by confirming that neural signal transmission occurs through Calcium imaging or MEA (multi-electrode array); and in the case of an intestinal organoid, by performing a Permeability assay.

[0142] In addition, iv) a method of analyzing morphological features can also be used to confirm whether the organoid has been successfully formed. That is, a second time point can be determined by confirming whether a lumen has formed in a bright-field image at an arbitrary time point and whether tissue has formed through H&E staining.

[0143] Returning to the description of FIG. 10, after step S200, the system (100) can generate a virtual fluorescent image (S210) by inputting the second observation image of the second time point (Day 9) obtained into a pre-trained image conversion model. Here, the image conversion model is a model that receives the observation image and generates a fluorescent image including a DAPI channel representing the cell nucleus and a target marker (e.g., SIX2) channel representing the differentiation marker, and the image conversion model may also be trained to generate a virtual fluorescent image from any observation image.

[0144] Meanwhile, after step S210, the system (100) can analyze the fluorescent image generated in the preceding step in pixel units to calculate a cell quality index, more precisely a differentiation efficiency index (S220).

[0145] After step S220, the system (100) may execute a step (S230) of determining the grade of the cell according to the cell quality indicator. That is, the system (100) may be implemented to determine the final quality grade of the corresponding well by comparing the calculated cell quality indicator with a preset reference value. For example, it may be implemented to classify it as High grade if the differentiation efficiency index is above a first threshold, as Low grade if it is below a second threshold, and as Mid grade if it is in between. Here, the threshold may be a value set by experimentally deriving the range where the organoid formation success rate changes rapidly.

[0146] In this regard, Figure 11 is a diagram showing the cell quality grade classification criteria and the experimental results that serve as the basis for setting the criteria.

[0147] The left side of Figure 11 shows an immunofluorescence staining image of an organoid formed when cells classified into each grade (Low, Mid, High) were actually differentiated. Upon examination, it was observed that in the Low grade cells, the organoid structure was hardly formed or the expression of kidney tissue markers was minimal. In the Mid grade cells, the organoid structure was partially formed, but the shape was incomplete or the uniformity of marker expression tended to be poor. In the High grade cells, the core structure of the kidney organoid was formed distinctly and complexly, and the marker expression was also strong and clear.

[0148] The graph in the upper right of Figure 11 shows the correlation between the differentiation efficiency index (X-axis) and the actual number of organoids formed (Y-axis). In the range where the differentiation efficiency index was less than approximately 34% (blue dots), the number of organoids formed converged to 0, indicating a significantly low differentiation success rate. In the range where the differentiation efficiency index was between approximately 34% and 58% (pink dots), the number of organoids formed was distributed from 1 to 3, indicating uncertainty regarding the success of differentiation. In the range where the differentiation efficiency index was approximately 58% or higher (red dots), the number of organoids formed converged to 4, indicating that stable organoid formation occurred with a very high probability.

[0149] Referring to the experimental results in Fig. 11, the aforementioned cell quality grades can be set as shown in the lower right corner of Fig. 11. That is, they can be set to be classified as Low grade, requiring cessation of culture when the differentiation efficiency index is less than 34.90%; Mid grade, requiring careful observation when it is between 31.02% and less than 58.81%; and High grade, allowing for continued culture when it is 56.13% or higher. For reference, in the table of Figure 11, the minimum and maximum values ​​of the data observed in the actual experiment are reflected as they are, and the lowest differentiation efficiency index among the samples of the group with 4 organoids formed was 56.13%, and the highest differentiation efficiency index among the samples of the group with 1 to 3 organoids formed was 58.81%, so the interval between 56.13 and 58.81 is shown to overlap. This interval can be expressed in various ways depending on the user's needs, such as being defined as a "Mid-High grade where 4 organoids may be formed or only 1 to 3 may be formed."

[0150] In other words, in the lower right corner of Figure 11, cell quality grades are divided into only three categories—Low, Mid, and High—based on the differentiation efficiency index. However, this represents an example where three sections were divided for convenience during the experiment, and it is understood that the number of cell quality grades, i.e., the sections of the differentiation efficiency index, can be varied as needed. In other words, it is understood that the main technical concept is not limited to the absolute numerical value of the differentiation efficiency index, but rather that the differentiation potential of cells can be predicted and graded by comparing the calculated index with a predetermined standard (threshold value or section).

[0151] Furthermore, the range of the differentiation efficiency index used to distinguish the above grades is not a limitation of the present invention, and the example of FIG. 11 is intended to demonstrate that the range of the differentiation efficiency index can be statistically derived to establish cell quality grades. In addition, these numerical threshold values ​​are not merely a simple statistical arrangement, but rather represent a section with critical significance where the organoid formation success rate changes rapidly. The section where the organoid formation success rate changes rapidly can be defined as a section corresponding to Low where only 0 to 5% of the total number of organoids that can be formed are formed normally, a section corresponding to High where 70% to 100% of the total number of organoids that can be formed are formed normally, and a section existing between these sections corresponding to Mid.

[0152] Meanwhile, after step S230, the system (100) can map the grade (e.g., High) of the second time point (Day 9) determined in the preceding step to the first observation image of the first time point (Day 4), which was taken in the same well, as the correct label. Through this, training data consisting of pairs such as [Input: Day 4 image, Output: High] can be generated.

[0153] For reference, Figure 12 is illustrated to provide an intuitive understanding of the training dataset required to train an early prediction model. In the figure, it can be seen that the training dataset is composed of pairs of observed images at a first time point (first observed image) and cell quality indices (e.g., DEI or grade) calculated based on virtual fluorescence images at a second time point. Prior to explaining Figure 6, it was mentioned that an early prediction model can be implemented by utilizing an image conversion model. This implies that virtual fluorescence images generated by the image conversion model are utilized to train the early prediction model.

[0154] On the other hand, the system (100) may increase the amount of data by using the entire well image as a single piece of data, or by dividing the well image into a number of small patches and then assigning the same grade to all patches.

[0155] The dataset constructed in this way is used to train an early prediction model, which can improve the model's performance in predicting future differentiation results based solely on early morphological features.

[0156] With reference to Figure 10 above, we have examined the process of training an early prediction model, that is, a model that predicts cell quality from observation images of the early stages of cells, and the process of generating a training dataset necessary for training it.

[0157]

[0158] FIG. 13 illustrates a method for early prediction of cell quality according to another embodiment of the present invention in sequence. While the embodiment described in FIG. 7 proceeds only up to the step of predicting cell quality from a first observation image, the embodiment of FIG. 13 is an embodiment that combines early prediction using an early prediction model with analysis and feedback using an image conversion model.

[0159] Referring to FIG. 13, the method of the present embodiment is largely divided into an early prediction and screening step (S110 to S130) and a quality confirmation and verification step (S140 to S150).

[0160] First, regarding steps S110 through S130, since these steps are substantially the same as those described in FIG. 7 above, a detailed explanation will be omitted. Briefly, the system (100) acquires a first observation image at a first time point (e.g., Day 4) (S110), inputs it into an early prediction model to predict a quality grade at a future time point (S120), and, based on the prediction result, selects cells early to determine whether to stop culture (S130).

[0161] After step S130, the system (100) can acquire a second observation image (S140) at a second time point (e.g., Day 9) for the cells determined to be suitable for continued culture, and can analyze the acquired second observation image to calculate a cell quality index (e.g., differentiation efficiency index) indicating how well the cells have actually differentiated.

[0162] Looking closely at step S150, it may include step S1501, in which an image conversion model generates a fluorescent image, and step S1502, in which the fluorescent image is analyzed to calculate cell quality indicators. Specifically, step S1501 is a step of inputting a second observed image into a pre-trained image conversion model to generate a fluorescent image, and step S1502 is a step of analyzing the generated fluorescent image pixel by pixel to calculate cell quality indicators, such as the ratio of differentiated cells to total cells.

[0163] As such, the embodiment illustrated in FIG. 13 can be understood as an integrated quality management method capable of performing both rapid decision-making in the initial stage and quality verification in the later stage in a single cycle process, and in particular, the cell quality indicators calculated in step S1502 can be utilized as feedback data to train or improve the early prediction model in step S120.

[0164]

[0165] Figure 14 is a diagram illustrating the embodiment of Figure 13 described above, along with images obtained from actual experiments, and shows the flow of the entire system in which the learning, prediction, and verification processes of a machine learning model are integrated.

[0166] Referring to FIG. 14, the system (100) is largely composed of three stages: a learning process, a prediction process, and a verification process, which can be structured in a cyclic manner as seen in the drawing.

[0167] First, the learning process represents the process of training an early prediction model. In this process, as previously explained, back-labeled image streams from past experimental data may be input, and it is assumed that these data are classified into Low, Mid, and High according to differentiation potential. The system (100) trains an early prediction model using this labeled data, thereby enabling the model to improve its performance in distinguishing each grade.

[0168] Next, the prediction process corresponds to steps S110 to S120 and represents the application step in the actual organoid production process. When a new first observation image (bright-field imagestream) acquired at the first time point (Day 4) is input, the previously trained early prediction model operates to predict the grade of the cell, and in the example of the diagram, it can be seen that the input image is predicted to be grade 'Low'.

[0169] Finally, the verification process corresponds to steps S140–S150 and can be understood as a step to confirm the accuracy of the prediction or to obtain new training data. In this process, the cells classified in the prediction process are cultured until the second time point (Day 9), and the second observation image (bright-field image) obtained is input into an image conversion model to generate a virtual fluorescence image. Through the differentiation efficiency index (e.g., 2.89%) calculated from this, it can be verified that the cells are actually of the 'Low' grade.

[0170] As suggested by the progress flow of Figure 14 (learning process → prediction process → verification process), the results confirmed during the verification process (Ground Truth) can be included back into the training dataset and used to retrain the early prediction model or verify its performance.

[0171]

[0172] The method for predicting cell quality early and the system for such prediction according to the present invention have been examined above. Meanwhile, the present invention is not limited to the specific embodiments and applications described above, and various modifications are possible by those skilled in the art without departing from the essence of the invention as claimed in the claims. Furthermore, such modifications should not be understood as being distinct from the technical spirit or perspective of the present invention.

Claims

1. In the method by which the system analyzes cell quality, (a) A step of acquiring an observation image of any cell in culture - said cell is in an unstained state; (b) a step of generating a virtual fluorescent image based on the acquired observation image using an image conversion model - the image conversion model is a machine learning model trained to convert a brightfield image into a fluorescent image -; and (c) a step of calculating a quality index of the cultured cells from the virtual fluorescent image generated above; including, Cell quality analysis method.

2. In Paragraph 1, The quality index of the above cells is the differentiation efficiency index (DEI), and The above differentiation efficiency index is, Characterized by being calculated as the ratio of the total intensity of the target differentiation marker channel to the total intensity of the cell nucleus indication marker channel for pixels exceeding a preset brightness threshold within the above-mentioned fluorescent image. Cell quality analysis method.

3. In Paragraph 1, The above image transformation model is a model based on a Conditional Generative Adversarial Network (cGAN), characterized by being trained to include a generator that extracts features from the observed image to generate a fluorescent image, and a discriminator that determines the authenticity of the generated image. Cell quality analysis method.

4. A system for analyzing cell quality, comprising a processor and memory, wherein The above memory stores instructions for executing a method for analyzing cell quality, and the processor executes the said instructions, The method for analyzing the cell quality mentioned above is, (a) A step of acquiring an observation image of any cell in culture - said cell is in an unstained state; (b) a step of generating a virtual fluorescent image based on the acquired observation image using an image conversion model - the image conversion model is a machine learning model trained to convert a brightfield image into a fluorescent image -; and (c) a step of calculating a quality index of the cultured cells from the virtual fluorescent image generated above; including, A system for analyzing cell quality.

5. A method for a system to train an image conversion model for converting a brightfield image of a cell into a fluorescent image, (a) a step of obtaining a training dataset in which a training observation image of any cell being cultured and an actual fluorescence image obtained by staining the cell corresponding to the training observation image are mutually mapped; (b) inputting the training observation image into the image conversion model to generate a virtual fluorescent image corresponding to the morphological features of the training observation image; and (c) inputting a pair of the training observation image and the virtual fluorescence image, or a pair of the training observation image and the actual fluorescence image, into the image conversion model to determine whether the input fluorescence image is a real image or a fake image; (d) a step of updating the parameters of the image transformation model; including, Method for training an image conversion model.

6. A system for training an image conversion model, comprising a processor and memory, The above memory stores instructions for executing a method of training an image conversion model, and the processor executes the said instructions, The method for training the above image transformation model is, (a) a step of obtaining a training dataset in which a training observation image of any cell being cultured and an actual fluorescence image obtained by staining the cell corresponding to the training observation image are mutually mapped; (b) inputting the training observation image into the image conversion model to generate a virtual fluorescent image corresponding to the morphological features of the training observation image; and (c) inputting a pair of the training observation image and the virtual fluorescence image, or a pair of the training observation image and the actual fluorescence image, into the image conversion model to determine whether the input fluorescence image is a real image or a fake image; (d) a step of updating the parameters of the image transformation model; including, A system for training an image conversion model.

7. A system for a method of predicting cell quality for organoid production early in a cell culture process, (a) a step of acquiring a first observation image of a culture vessel containing cells at a first time point after organoid differentiation induction; and (b) a step of predicting the expected cell quality when the cell reaches a second time point, which is a time point after the first time point, using an early prediction model for the first observation image; including, Method for predicting cell quality early.

8. In Paragraph 7, The cell quality is characterized as being a differentiation potential grade expected when the second time point is reached. Method for predicting cell quality early.

9. In Paragraph 8, The above early prediction model is, Characterized by the quality of the cell measured at the second time point being learned using learning data mapped to the image at the first time point. Method for predicting cell quality early.

10. In Paragraph 7, (c) a step of determining whether to continue culturing the cells in the culture vessel based on the predicted cell quality above; further comprising, Method for predicting cell quality early.

11. In Paragraph 7, (d) a step of acquiring a second observation image when the cell reaches the second time point; (e) a step of calculating a cell quality index for the cell based on the second observation image above; including, Method for predicting cell quality early.

12. In Paragraph 11, The above step (e) is, (e1) a step of generating a fluorescent image corresponding to the second observed image by utilizing an image conversion model—the image conversion model being trained to convert the observed image into a fluorescent image—based on the second observed image; and (e2) a step of analyzing the above fluorescent image to calculate a cell quality indicator (differentiation efficiency index); further comprising, Method for predicting cell quality early.

13. In Paragraph 11, The first observation image and the second observation image are characterized as being live-cell bright-field images of unstained cells. Method for predicting cell quality early.

14. A system for predicting cell quality early, comprising a processor and memory, wherein The above memory stores instructions for executing a method to predict cell quality early, and the processor executes the said instructions, The method for predicting the above cell quality early is, (a) a step of acquiring a first observation image of a culture vessel containing cells at a first time point after organoid differentiation induction; and (b) a step of predicting the expected cell quality when the cell reaches a second time point, which is a time point after the first time point, using an early prediction model with respect to the first observation image; comprising A system for predicting cell quality early.

15. In Paragraph 14, The above system is characterized by acquiring the first observation image from an imaging module that automatically scans a cell culture vessel at predetermined time intervals to generate an observation image. In addition, the system is characterized by outputting the cell quality predicted from the image acquired at the first time point to a display unit. A system for predicting cell quality early.

16. In the method by which the system trains an early prediction model, The method includes the step of training an early prediction model using a training dataset to predict cell quality indicators using only the observed images at a first time point; The above training dataset is, (a) A step of obtaining a first observation image at a first time point and a second observation image at a second time point, which is a predetermined time after the first time point, for a cell that is differentiating in a culture space; (b) Generate a virtual fluorescence image from the second observation image at the second time point using an image transformation model, and (c) a step of calculating cell quality indicators for each cell by analyzing the virtual fluorescence image above; (d) A step of determining the quality of any cell as one of a plurality of grades according to the cell quality indicator calculated above; (e) a step of constructing a training dataset by mapping the cell grade determined at the second time point to the first observation image at the first time point; Characterized by being generated through, Method for training an early prediction model.

17. In Paragraph 16, The above step (e) is, Characterized by using a back-labeling method that maps the grade result of the second time point to the first observation image of the first time point. Method for training an early prediction model.

18. A system for training an early prediction model, comprising a processor and memory, wherein The above memory stores instructions for executing a method for training an early prediction model, and the processor executes the said instructions, The method for training the above-mentioned early prediction model includes the step of training the early prediction model to predict cell quality indicators using only the observed images at a first time point, using a training dataset; and The above training dataset is, (a) A step of obtaining a first observation image at a first time point and a second observation image at a second time point, which is a predetermined time after the first time point, for a cell that is differentiating in a culture space; (b) Generate a virtual fluorescence image from the second observation image at the second time point using an image transformation model, and (c) a step of calculating cell quality indicators for each cell by analyzing the virtual fluorescence image above; (d) A step of determining the quality of any cell as one of a plurality of grades according to the cell quality indicator calculated above; (e) a step of constructing a training dataset by mapping the cell grade determined at the second time point to the first observation image at the first time point; Characterized by being generated through, A system for training an early prediction model.

Images