Time-lapse image classifier, learning method for time-lapse image classifier, embryo classification device, and embryo classification method.

Abstract

To provide a time lapse image discriminator, a learning method of the time lapse image discriminator, and an embryo discrimination device and an embryo discrimination method that enhance discrimination precision for an embryo.SOLUTION: A time lapse image discriminator consists of an input layer to which a time lapse image obtained by imaging an embryo of a patient to be discriminated is input, an output layer to which a discrimination result of the embryo is output, and an intermediate layer which calculates the discrimination result to be output to the output layer by using the time lapse image input to the input layer and parent body information related to the parent body of the patient. The intermediate layer is learnt through a neural network, a learning time lapse image for a learning embryo of a learning patient is input to the input layer in the learning of the intermediate layer, and learning parent body information related to the parent body of the learning patient is input to the intermediate layer.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present invention relates to a time-lapse image discriminator, a learning method for the time-lapse image discriminator, an embryo discriminator, and an embryo discrimination method.

Background Art

[0002] The number of births by assisted reproductive technology (ART) in Japan has been increasing in recent years. The number of births by ART in 2019 exceeded 60,000, which corresponds to about 7% of the annual number of births. ART is an infertility treatment method in which a plurality of collected eggs are cultured to blastocysts in an embryo culture device (incubator), and the embryos selected as good embryos from the plurality of cultured embryos (fertilized eggs) are transplanted into the uterus.

[0003] The selection of good embryos is performed by an embryologist visually observing each embryo immediately after fertilization (early embryo) and 5 days after fertilization (blastocyst). The evaluation accuracy of the embryologist's visual selection of embryos depends on the knowledge and experience of the embryologist. Therefore, a uniform and highly accurate embryo evaluation method is required.

[0004] On the other hand, it is not uncommon for good embryos transplanted into the uterus and implanted to result in miscarriage. Many of the causes of miscarriage are chromosomal abnormalities (karyotype abnormalities). It is known that the probability of chromosomal abnormalities increases with the age of the patient. Therefore, for patients who have reached a certain age, evaluating chromosomal abnormalities is more important for the success or failure of obtaining a child from embryo transplantation than evaluating the quality of the embryo (whether it is a good embryo or not). Therefore, in recent years, preimplantation genetic testing for aneuploidy (PGT-A), which examines the chromosome number of embryos before transplantation into the uterus, has been performed for patients with habitual miscarriage or repeated ART failure cases. However, PGT-A has problems such as invasiveness to embryos and high examination costs.

[0005] To date, inventions have been proposed relating to embryo culture devices (time-lapse incubators) that generate time-lapse images by imaging embryos in culture at predetermined intervals, and to the evaluation of embryos using machine learning models based on time-lapse images (see, for example, Patent Documents 1 and 2, and Non-Patent Documents 1 and 3). [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Patent No. 6956302 specification [Patent Document 2] Patent No. 6414310 specification [Non-patent literature]

[0007] [Non-Patent Document 1] Jorgen Berntsen et.al., Robust and generalizable embryo selection based on artificial intelligence and time-lapse image sequences, PLOS ONE, 2022. [Non-Patent Document 2] Tran et.al., Deep learning as a predictive tool for fetal heart pregnancy following time-lapse incubation and blastocyst transfer, Human Reproduction, 2018. [Non-Patent Document 3] Khosravi et.al., Deep learning enables robust assessment and selection of human blastocyst after in vitro fertilization, NPJ Digital Medicine, 2019. [Overview of the project] [Problems that the invention aims to solve]

[0008] The present invention aims to provide a time-lapse image discriminator, a learning method for the time-lapse image discriminator, an embryo discrimination device, and an embryo discrimination method that improve the accuracy (prediction accuracy) of embryo discrimination. [Means for solving the problem]

[0009] The time-lapse image discriminator according to the present invention comprises an input layer into which a time-lapse image of the embryo of a patient to be discriminated is input, an output layer that outputs the discrimination result of the embryo, and an intermediate layer that calculates the discrimination result to be output to the output layer using the time-lapse image input to the input layer and maternal information related to the patient's mother, wherein the intermediate layer is trained using a neural network, and in the training of the intermediate layer, a training time-lapse image of a training embryo of a training patient is input to the input layer, and training maternal information related to the training patient's mother is input to the intermediate layer. [Effects of the Invention]

[0010] This invention improves the accuracy of embryo discrimination. [Brief explanation of the drawing]

[0011] [Figure 1] This is a functional block diagram showing an embodiment of the embryo discrimination device according to the present invention. [Figure 2] This is a schematic diagram illustrating an example of a conventional neural network model. [Figure 3] This is a schematic diagram illustrating another example of a conventional neural network model. [Figure 4] This is a schematic diagram showing an example of a neural network model used as a time-lapse image discriminator according to the present invention. [Figure 5] This is a schematic diagram showing another example of a neural network model used as a time-lapse image discriminator. [Figure 6]It is a schematic diagram showing another example of the neural network model used as the above time-lapse image discriminator. [Figure 7] It is a schematic diagram showing another example of the neural network model used as the above time-lapse image discriminator.

Embodiments for Carrying Out the Invention

[0012] Embodiments of the time-lapse image discriminator, the learning method of the time-lapse image discriminator, the embryo discriminator, and the embryo discrimination method according to the present invention will be described below together with the drawings.

[0013] ●Configuration of Embryo Discriminator● FIG. 1 is a functional block diagram showing an embodiment of an embryo discriminator (hereinafter referred to as "this device") according to the present invention.

[0014] This device 1 is an embryo culture device (time-lapse incubator) that captures time-lapse images of embryos while culturing embryos of a patient to be discriminated. This device 1 also functions as a time-lapse image discrimination device in which the time-lapse image discriminator (hereinafter referred to as "this discriminator") according to the present invention operates.

[0015] This discriminator is a learned neural network model that has been machine-learned in advance using supervised learning data by the learning method of the time-lapse image discriminator (hereinafter referred to as "this learning method") according to the present invention. Details of this learning method will be described later. The neural network is, for example, a convolutional neural network, which is said to have high discrimination accuracy for images. The learning data is the time-lapse images of embryos for each learning patient, maternal information, and the progress information of embryos for each learning patient. The progress information of the embryo is the information to be discriminated (predicted) by this discriminator, and is, for example, the success or failure of embryo implantation, the presence or absence of chromosomal abnormalities (whether it is a euploid embryo).

[0016] This discriminator discriminates whether a good embryo with a high implantation rate or a chromosomal abnormality with a high miscarriage rate based on the time-lapse images and maternal information of the embryos (fertilized eggs) of the patient to be discriminated, that is, the patient who hopes to give birth after receiving assisted reproductive medicine.

[0017] In the following description, the object of discrimination by this discriminator is whether the embryo is a good embryo.

[0018] Note that instead of operating in a time-lapse incubator, this discriminator may operate in an information processing device such as a personal computer. The information processing device in which this discriminator operates functions as a time-lapse image discriminator. This discriminator cooperates with the hardware resources of the information processing device in which it operates to discriminate (predict the progress of) the embryos of the patient.

[0019] Here, the hardware resources of the information processing device in which this discriminator operates are, for example, processors such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), DSP (Digital Signal Processor), and storage media. This storage media stores this discriminator, the time-lapse images of the patient to be discriminated, maternal information, and the like. The processor in which this discriminator operates realizes each of the means (input unit, acquisition unit, discrimination unit, output unit) described later provided in this device using the information stored in the storage media provided in the information processing device in which this discriminator operates.

[0020] This device 1 includes a storage unit 2, an imaging unit 3, an input unit 4, an acquisition unit 5, a discrimination unit 6, and an output unit 7.

[0021] The storage unit 2 stores this discriminator and information used by this device 1 to implement the embryo discrimination method according to the present invention (hereinafter referred to as "this method"). The details of the information used by this device 1 to implement this method will be described later.

[0022] The storage unit of this device 1 may include, for example, portable storage media such as HDDs (Hard Disk Drives), SSDs (Solid State Drives), and flash memory, as well as other non-temporary recording media, or RAM (Random Access Memory) and other temporary recording media.

[0023] The imaging unit 3 captures time-lapse images of embryos being cultured in the device 1 (cultured embryos from fertilization to the blastocyst stage). The imaging unit 3 captures time-lapse images at predetermined time intervals (for example, every 10 minutes). The captured time-lapse images are stored in the storage unit 2. Multiple time-lapse images stored in the storage unit 2 are displayed on the device 1's display (not shown) in the order they were captured, as if in a video. The operator of the device 1 (for example, an embryologist) can check the changes in the displayed time-lapse images over time.

[0024] Input unit 4 receives maternal information of the patient to be identified. Input unit 4 is an information input means provided by the device 1, such as a keyboard, mouse, or touch panel. Input unit 4 is operated by the operator of the device 1. Maternal information of the patient includes, for example, information about the patient's age (information indicating age or age group) and information about the patient's physical condition (for example, medical history). Maternal information entered from input unit 4 is stored in storage unit 2.

[0025] The acquisition unit 5 acquires the time-lapse images of the patient's embryo and maternal information stored in the memory unit 2.

[0026] The discrimination unit 6 identifies the patient's embryo based on the time-lapse image of the patient's embryo and maternal information acquired by the acquisition unit 5 from the memory unit 2, and the discrimination unit stored in the memory unit 2.

[0027] The output unit 7 outputs the discrimination result determined by the discrimination unit 6. The methods of outputting the discrimination result by the output unit 7 include, for example, storing the discrimination result in the storage unit 2, displaying the discrimination result on the display (not shown) of the device 1, or transmitting the discrimination result to an information processing device (not shown) connected to the device 1 via a communication network.

[0028] ●Traditional neural network models● Figure 2 is a schematic diagram illustrating an example of a conventional neural network model.

[0029] A neural network consists of an input layer, hidden layers, and an output layer. In this case, the neural network is a classifier that uses time-lapse images of the embryos of the patient being classified to determine whether or not an embryo is a good embryo. That is, the input layer is the layer to which the time-lapse images are input. The output layer is the layer to which the classification result (whether or not the embryo is a good embryo) is output. The hidden layers are the layers that connect the input layer and the output layer. The hidden layers are a collection of multiple individual hidden layers. An individual hidden layer contains n individual hidden layers, arranged from the input layer to the output layer: the first individual hidden layer, the second individual hidden layer, the third individual hidden layer, ..., the (n-1)th individual hidden layer, the nth individual hidden layer, and so on.

[0030] In this diagram, neurons (units) that make up the input layer are shown as black circles, neurons that make up the hidden layer are shown as white circles (solid lines), and neurons that make up the output layer are shown as white circles (dashed lines).

[0031] The number of neurons in the input layer corresponds to the resolution of the time-lapse image. If the resolution of the time-lapse image is 1280 x 1024 pixels, the number of neurons in the input layer is 1,310,720 (= 1280 x 1024). In other words, each neuron in the input layer holds information about the brightness value and RGB values ​​for each pixel that makes up the time-lapse image.

[0032] Furthermore, the neurons constituting the input layer may include neurons corresponding to so-called biases. In this case, the number of neurons constituting the input layer is the number corresponding to the resolution of the time-lapse image plus the number of biases.

[0033] The number of neurons in the output layer corresponds to the number of classification results. Here, the classification result is whether or not it is a good embryo. That is, the number of neurons in the output layer is 2 ("good embryo" and "not a good embryo").

[0034] The number of individual hidden layers that make up the hidden layer, the number of neurons that make up each individual hidden layer, the weight of each edge between neurons from the input layer to the output layer, and the bias value are all set appropriately during the neural network learning process based on evaluation values ​​such as AUC (Area Under the Curve).

[0035] Figure 3 is a schematic diagram illustrating another example of a conventional neural network model. The figure shows that neurons, indicated by black squares, have been added to the input layer of the model shown in Figure 2. These black square neurons contain maternal information of the patient being identified. In other words, the difference between the model shown in Figure 3 and the model shown in Figure 2 is whether or not maternal information is included in the input layer (model in Figure 3) or not (model in Figure 2). That is, the model shown in Figure 2 determines whether or not a patient's embryo is a good embryo based solely on time-lapse images of the patient's embryo. On the other hand, the model shown in Figure 3 determines whether or not a patient's embryo is a good embryo based on both time-lapse images of the patient's embryo and maternal information.

[0036] The number of neurons constituting the input layer of the model shown in Figure 3 is the sum of the number of black circles (corresponding to the resolution of the time-lapse image) and the number of black squares (corresponding to 1) that represent the parent information (without considering bias). Here, as mentioned above, when the resolution of the time-lapse image is 1280 × 1024 pixels, the number of black circles is 1,310,720. On the other hand, the number of black squares is 1.

[0037] Thus, when time-lapse images and parental information are input to the input layer for discrimination, the amount of information in the parental information is overwhelmingly less than the amount of information in the time-lapse images. Therefore, in the learning process of the neural network, the parental information is not easily reflected in the learning results. Similarly, in the discrimination process using a pre-trained neural network, the parental information is not easily reflected in the discrimination results.

[0038] On the other hand, in determining the success of embryo transfer and live birth, the presence or absence of chromosomal abnormalities is as important as the quality of the embryo (whether it is a good quality embryo or not). Embryo quality and the presence or absence of chromosomal abnormalities can be influenced by the maternal condition, such as the patient's age and medical history. Therefore, it is desirable that information related to the patient's maternal condition be taken into account when using neural networks for embryo selection.

[0039] ●Configuration of the time-lapse image discriminator● As explained below, this classifier improves the accuracy of classification by incorporating both time-lapse images and parental information, which contains significantly less information than time-lapse images, into the learning and classification results.

[0040] Figure 4 is a schematic diagram showing an example of a neural network model used as this discriminator. The figure shows that neurons indicated by black squares have been added to some of the individual intermediate layers (the (n-1)th individual intermediate layer) of the model shown in Figure 2. These black square neurons are the same as those shown in Figure 3 and contain maternal information of the patient being identified. In other words, the difference between the model shown in Figure 4 and the model shown in Figure 3 is whether the neurons containing maternal information are included in the input layer (model in Figure 3) or in the intermediate layer (model in Figure 4). That is, the model shown in Figure 3 determines whether a patient's embryo is a good embryo or not based on the time-lapse image of the patient's embryo input to the input layer and the maternal information. On the other hand, the model shown in Figure 4 determines whether a patient's embryo is a good embryo or not based on the time-lapse image of the patient's embryo input to the input layer and the maternal information of the patient input to the intermediate layer (the (n-1)th individual intermediate layer).

[0041] Here, in the neural network model used as the main classifier in Figure 4, during the learning process using training data, time-lapse images of embryos for each training patient are input to the input layer, and maternal information for each training patient is input to the (n-1)th individual hidden layer, and the model is learned. Similarly, in the neural network model used as the classifier in Figure 4, during the classification process using data from the patient to be classified, a time-lapse image of the patient's embryo is input to the input layer, and the patient's maternal information is input to the (n-1)th individual hidden layer, and then classification is performed.

[0042] Thus, in the learning and classification processes of the neural network model used as this classifier, inputting maternal information, which has significantly less information than time-lapse images, into the intermediate layer allows maternal information to be more easily reflected in the learning and classification results compared to inputting maternal information into the input layer along with time-lapse images (model in Figure 3). This is because when maternal information is input into the input layer along with time-lapse images, the content of the maternal information is treated to the same extent as one pixel of the time-lapse image. On the other hand, when maternal information is input into individual intermediate layers that make up the intermediate layer, particularly individual intermediate layers close to the output layer where the information from the time-lapse images input into the input layer is aggregated, as in this classifier, the content of the maternal information is emphasized and reflected in the learning and classification results. In other words, this classifier obtains learning and classification results that reflect maternal information, which affects the success or failure of embryo transfer to live birth. In short, this classifier can improve the accuracy of embryo classification.

[0043] In the neural network model used as this discriminator, the parent information is input to at least one of the multiple individual hidden layers that make up the hidden layer.

[0044] Figure 5-7 is a schematic diagram showing another example of a neural network model used as this discriminator.

[0045] Figure 5 shows that common parental information is input to two individual intermediate layers (the (n-1)th individual intermediate layer and the nth individual intermediate layer) during the learning and discrimination processes. The number of individual intermediate layers to which the common parental information is input, and the positions within the individual intermediate layers to which the parental information is input (the arrangement of neurons in the time-lapse images that make up the individual intermediate layers and neurons in the parental information: the figure shows that the parental information is input to the beginning (top of the page) of each individual intermediate layer) are set as appropriate during the learning process.

[0046] Figure 6 shows that during the learning and classification processes, common parental information is input to two locations within a single individual intermediate layer (the (n-1)th individual intermediate layer). The number of parental information points input to a single individual intermediate layer, and the position within the individual intermediate layer where the parental information is input, are set as appropriate during the learning process.

[0047] Figure 7 shows that during the learning and classification processes, common parental information is input to all individual intermediate layers that make up the intermediate layer, as well as to the input layer. The number of parental information items input to each individual intermediate layer, and the position of the parental information within the individual intermediate layer, are set as appropriate during the learning process.

[0048] In the neural network model used as this classifier, the individual hidden layers that receive maternal information may be selected by the classifier based on the patient's time-lapse images and maternal information. The method for selecting the individual hidden layers that receive maternal information is set appropriately during the learning process using training data. That is, for example, if during the learning process it is evaluated that age has a small influence on the classification result for young people and a large influence for middle-aged and elderly people, the individual hidden layers that receive maternal information may be selected according to the patient's age. In other words, for example, as the patient's age increases, the individual hidden layers that receive maternal information may be selected that are closer to the output layer.

[0049] ●Summary● According to the embodiments described above, in the learning process of the neural network model used as the classifier, maternal information, which contains significantly less information than time-lapse images, is input to one of the multiple individual intermediate layers that make up the intermediate layer. Compared to the case where maternal information is input to the input layer along with time-lapse images, maternal information is more easily reflected in the learning results and classification results. In other words, this classifier and device obtain learning results and classification results that reflect maternal information, which affects the success or failure of embryo transfer to live birth. In short, this classifier and device can improve the accuracy (prediction accuracy) of embryo classification.

[0050] As mentioned above, this discriminator and device can determine (predict) whether an embryo is a good quality embryo with a high implantation rate, or whether it has a chromosomal abnormality with a high miscarriage rate, based on time-lapse images of the patient's embryo and maternal information. In particular, this discriminator and device, which can determine whether or not an embryo has a chromosomal abnormality, can serve as an alternative to PGT-A, which was previously the only available testing method. As mentioned above, PGT-A has drawbacks such as being invasive to the embryo and having high testing costs. By using this discriminator and device as an alternative to PGT-A, the number of cases requiring PGT-A can be reduced.

[0051] Furthermore, in the embodiments described above, the learning data for this classifier consisted of time-lapse images and maternal information of the embryo for each patient being studied, as well as progress information of the embryo for each patient being studied. Alternatively, the learning data for this classifier may also include the PGT-A test results (number of chromosomes in the embryo) of the embryo for each patient being studied. That is, this classifier may be trained using time-lapse images, maternal information, and PGT-A test results of the embryo for each patient being studied, as well as progress information of the embryo for each patient being studied. In this case, the classifier determines whether the embryo is a good embryo or not, or whether it has a chromosomal abnormality or not, based on the time-lapse images, maternal information, and PGT-A test results of the embryo of the patient to be classified. Here, in the learning process or classification process of this classifier, the layer to which the PGT-A test results are input is the same intermediate layer (including the input layer in Figure 7) as the intermediate layer to which maternal information is input, or a different intermediate layer (including the input layer in Figure 7) from the intermediate layer to which maternal information is input. The position within the individual intermediate layer where the PGT-A test results are input is set appropriately during the learning process of this classifier, similar to the determination of the position within the individual intermediate layer where the aforementioned parental information is input.

[0052] ●Features of this discriminator, this learning method, this device, and this method● The characteristics of the classifier, learning method, apparatus, and method described above are summarized below.

[0053] ●Features of this discriminator This discriminator is An input layer receives time-lapse images of the embryos of the patients to be identified, An output layer on which the embryo discrimination result is output, An intermediate layer that calculates the discrimination result output to the output layer using the time-lapse image input to the input layer and maternal information related to the patient's mother, A time-lapse image discriminator comprising, The aforementioned intermediate layer is trained using a neural network. In the learning of the aforementioned intermediate layer, A learning time-lapse image of a learning embryo from a learning patient is input to the input layer. The learning maternal information related to the maternal organism of the learning patient is input to the intermediate layer. It is characterized by the following:

[0054] In this discriminator, The maternal information includes information regarding the patient's age. But that's fine.

[0055] In this discriminator, The aforementioned intermediate layer includes a plurality of individual intermediate layers, In the learning of the intermediate layer, the learning matrix information is input to at least one of the individual intermediate layers among the plurality of individual intermediate layers. It can be anything.

[0056] In this discriminator, Among the multiple individual intermediate layers, the individual intermediate layer into which the maternal information of the patient to be identified is input is selected based on the time-lapse image or maternal information of the patient to be identified. It can be anything.

[0057] In this discriminator, In the learning of the intermediate layer, the learning matrix information is input to the input layer. It can be anything.

[0058] ●Features of this learning method This learning method, An input layer receives time-lapse images of the embryos of the patients to be identified, An output layer on which the embryo discrimination result is output, An intermediate layer that calculates the discrimination result output to the output layer using the time-lapse image input to the input layer and maternal information related to the patient's mother, A learning method for a time-lapse image discriminator, comprising: The aforementioned intermediate layer is trained using a neural network. In the learning of the aforementioned intermediate layer, The steps include inputting learning time-lapse images of learning embryos from learning patients into the input layer, The steps include inputting learning maternal information related to the maternal organism of the learning patient into the intermediate layer, Having, It is characterized by the following:

[0059] ●Features of this device This device A storage unit that stores a discriminator learned based on learning time-lapse images for each learning patient and learning parent information related to the parent for each learning patient, An acquisition unit that acquires time-lapse images of the embryo of the patient to be identified and maternal information related to the patient's mother, A discrimination unit that discriminates the embryo based on the time-lapse image acquired by the acquisition unit, the maternal information, and the discriminator, An output unit that outputs the discrimination result determined by the discrimination unit, It has, The aforementioned discriminator is this discriminator, It is characterized by the following:

[0060] ●Features of this method This method, A storage unit that stores a discriminator learned based on learning time-lapse images for each learning patient and learning parent information related to the parent for each learning patient. An embryo discrimination method performed on an embryo discrimination device equipped with, The embryo discrimination device, An acquisition step to obtain a time-lapse image of the embryo of the patient to be identified and maternal information related to the patient's mother, A discrimination step in which the embryo is identified based on the time-lapse image acquired in the acquisition step, the maternal information, and the discriminator, An output step which outputs the discrimination result determined in the discrimination step, It has, The aforementioned discriminator is this discriminator, It is characterized by the following: [Explanation of Symbols]

[0061] 1 Embryo identification device 2 Storage section 3. Imaging Unit 4 Input section 5 Acquisition part 6 Discrimination section 7 Output section

Claims

1. A computer An input layer receives time-lapse images of the embryos of the patients to be identified, An output layer on which the embryo discrimination result is output, An intermediate layer that calculates the discrimination result output to the output layer using the time-lapse image input to the input layer and maternal information related to the maternal body of the patient to be discriminated, A time-lapse image discriminator program for which the following functions: The input layer, the hidden layer, and the output layer constitute a trained neural network. The aforementioned trained neural network is A learning time-lapse image of a learning embryo from a learning patient is input to the input layer. The learning maternal information related to the maternal organism of the learning patient is input to the intermediate layer. This is how it is learned, The aforementioned intermediate layer includes a plurality of individual intermediate layers, In the trained neural network, the maternal information of the patient to be identified is input to at least one of the multiple individual intermediate layers. The individual intermediate layer into which the maternal information of the patient to be identified is input is configured to be selected from among a plurality of individual intermediate layers based on the time-lapse image or maternal information of the patient to be identified. Time-lapse image recognition program.

2. The maternal information is information regarding the patient's age. A time-lapse image discriminator program according to claim 1.

3. In the training of the aforementioned trained neural network, the training data is the data input to the input layer. A time-lapse image discriminator program according to claim 1.

4. An embryo discrimination device that discriminates embryos using time-lapse images, Processor and A memory unit that stores the trained neural network, An acquisition unit that acquires time-lapse images of the embryo of the patient to be identified and maternal information related to the patient's mother, A discrimination unit that identifies the embryo based on the time-lapse image acquired by the acquisition unit, the maternal information, and the trained neural network, An output unit that outputs the discrimination result determined by the discrimination unit, It has, The aforementioned trained neural network is An input layer receives time-lapse images of the embryos of the patients to be identified, An output layer on which the embryo discrimination result is output, An intermediate layer that calculates the discrimination result output to the output layer using the time-lapse image input to the input layer and maternal information related to the maternal body of the patient to be discriminated, Composed of, Learning is performed by inputting learning time-lapse images of a learning patient's learning embryo into the input layer, and learning maternal information related to the learning patient's mother into the intermediate layer. The aforementioned intermediate layer includes a plurality of individual intermediate layers, In the trained neural network, the maternal information of the patient to be identified is input to at least one of the multiple individual intermediate layers. The individual intermediate layer into which the maternal information of the patient to be identified is input is configured to be selected from among a plurality of individual intermediate layers based on the time-lapse image or maternal information of the patient to be identified. Embryo identification device.

5. The maternal information is information relating to the patient's age, The embryo discrimination device according to claim 4.

6. A method for identifying an embryo using time-lapse images, performed by a computer, The aforementioned computer, Processor and A memory unit that stores the trained neural network, It has, The aforementioned trained neural network is It consists of an input layer, a hidden layer, and an output layer. Learning is performed by inputting learning time-lapse images of a learning patient's learning embryo into the input layer, and learning maternal information related to the learning patient's mother into the intermediate layer. The aforementioned intermediate layer includes a plurality of individual intermediate layers, The aforementioned computer, An acquisition step to obtain a time-lapse image of the embryo of the patient to be identified and maternal information related to the mother of the patient to be identified, The steps include inputting the acquired time-lapse image to the input layer, The steps include inputting the acquired maternal information of the patient to be identified into at least one of the multiple individual intermediate layers, In the intermediate layer, a discrimination step is performed in which the time-lapse image input to the input layer and maternal information related to the mother of the patient to be discriminated are used to calculate a discrimination result to be output to the output layer, thereby discriminating the embryo. An output step which outputs the discrimination result determined in the discrimination step, Includes, Here, the individual intermediate layer into which the maternal information of the patient to be identified is input is selected from among a plurality of individual intermediate layers based on the time-lapse image or maternal information of the patient to be identified. Embryo identification method.

7. The maternal information is information relating to the patient's age, The embryo discrimination method according to claim 6.

Images