Deep learning method and system for embryonic development multi-stage classification

Abstract

The invention discloses a deep learning method and system for multi-stage classification of embryonic development, and belongs to the field of multi-stage classification of embryonic development, and the method comprises the steps: carrying out the partitioning operation of a preprocessed embryo image, inputting the segmented image into a dual-branch local feature fusion module, and carrying out the local feature extraction, and obtaining a feature map containing fine-grained local features; inputting the feature map containing the fine-grained local features into a deep neural network module for global feature extraction and coding to obtain a final feature map; and inputting the final feature map into a classifier for development stage prediction, and outputting a predicted embryonic development stage. The advantages of the dual-branch local feature fusion module and the deep neural network module are combined to be used for classification of the embryo development stage, local and global information of the embryo image can be fully extracted, the accuracy of the classification stage is improved, and meanwhile a new thought is provided for application of deep learning in IVF.

Description

Technical Field

[0001] The present invention belongs to the field of multi-stage classification of embryonic development, and specifically relates to a deep learning method and system for multi-stage classification of embryonic development. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Infertility is a widespread global problem. According to the World Health Organization (WHO), approximately one-sixth of the world's population suffers from infertility.

[0004] At present, domestic and international guidelines and expert consensus point out that In Vitro Fertilization-Embryo Transfer (IVF-ET) is one of the most effective ways to solve the problem of infertility. According to statistics, the current success rate of IVF is usually only about 40%. During IVF treatment, in order to reduce the risks of pregnancy and fetal health caused by multiple pregnancies, doctors usually recommend transplanting only one embryo in one embryo transfer cycle. Therefore, how to select an embryo with the greatest implantation potential for implantation is crucial to improving the success rate of IVF.

[0005] The embryonic development process can be subdivided into 16 consecutive stages, including pPB2, pPNa, pPNf, p2, p3, p4, p5, p6, p7, p8, p9+, pM, pSB, pB, pEB and pHB. The fine-grained classification of embryonic development stages can better understand its developmental status, which is of great significance for selecting the most promising embryos for transplantation.

[0006] Currently in clinical practice, embryologists use their professional knowledge and experience to select embryos with high potential by observing the morphological characteristics and development speed of embryos. However, this method is easily affected by subjective factors, and there are problems such as inconsistent standards, large subjective errors, and heavy workload. At the same time, there are certain differences in technical levels between different regions and laboratories. Therefore, there is an urgent need for an objective and automated evaluation method to provide effective auxiliary support for embryologists, and at the same time provide advanced technology for laboratories in areas with insufficient medical levels, so as to promote efficient IVF treatment methods on a wider range and ultimately improve the success rate of IVF worldwide.

[0007] In recent years, with the rapid development of deep learning neural networks, the application of deep learning technology in the field of assisted reproduction has begun to attract attention from the academic community. Deep learning technology is expected to accurately classify the stages of embryonic development, thereby providing doctors and embryologists with rich information to help doctors select embryos with relatively good development status.

[0008] Although deep learning-based embryo classification methods have made significant progress in the field of assisted reproduction, these methods still face many challenges. Specifically, embryos show complex morphological and structural changes at different stages of development. Therefore, the key to improving the accuracy of multi-stage classification of embryo development lies in the accurate extraction of local and global features. However, existing deep learning-based related methods may have information loss problems during feature extraction, which in turn affects the accuracy of multi-stage classification of embryo development, and may affect the subsequent IVF success rate. Summary of the invention

[0009] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a deep learning method and system for multi-stage classification of embryonic development. The method and system combine a dual-branch local feature fusion module and a Transformer deep neural network module, and combine the advantages of these two modules for the classification of embryonic development stages. It can not only fully extract the local and global information of embryonic images and improve the accuracy of the classification stage, but also provide a new idea for the application of deep learning in IVF.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] In a first aspect, the present invention provides a deep learning method for multi-stage classification of embryonic development, comprising:

[0012] Acquiring an embryo image, and preprocessing the embryo image to obtain a preprocessed embryo image;

[0013] The preprocessed embryo image is divided into blocks and input into a dual-branch local feature fusion module for local feature extraction to obtain a feature map containing fine-grained local features.

[0014] Inputting the feature map containing fine-grained local features into a deep neural network module for global feature extraction and encoding to obtain a final feature map;

[0015] The final feature map is input into a classifier for developmental stage prediction, and the predicted embryonic developmental stage is output.

[0016] According to a further technical solution, the dual-branch local feature fusion module includes a first branch module and a second branch module connected in parallel, wherein the first branch module captures the local spatial information of the input image, and the second branch module effectively extracts channel features.

[0017] A further technical solution is to connect the fully connected layer and the local self-attention module in sequence to obtain the first branch module, and connect the fully connected layer and the deep convolution module in sequence to obtain the second branch module.

[0018] A further technical solution is that a cross-branch interaction fusion module is connected between the local self-attention module and the deep convolution module to realize cross-branch feature interaction and fusion operations.

[0019] A further technical solution is that the cross-branch interactive fusion module includes a parallel channel interaction module and a spatial interaction module, and the features extracted by the deep convolution module are interacted to the local self-attention module through the channel interaction module, and the spatial relationship is transferred from the local self-attention module to the deep convolution module through the spatial interaction module.

[0020] A further technical solution is to connect the pooling layer, 1×1 convolution layer, batch normalization, GELU activation function, 1×1 convolution layer and sigmoid function in sequence to obtain a channel interaction module, and to connect the 1×1 convolution layer, batch normalization, GELU activation function, 1×1 convolution layer and sigmoid function in sequence to obtain a spatial interaction module.

[0021] According to a further technical solution, the deep neural network module is specifically a Transformer encoder module.

[0022] In a second aspect, the present invention provides a deep learning system for multi-stage classification of embryonic development, comprising:

[0023] An image acquisition module is configured to: acquire an embryo image, preprocess the embryo image, and obtain a preprocessed embryo image;

[0024] A local feature extraction module is configured to: perform a block operation on the preprocessed embryo image, and input the preprocessed embryo image into a dual-branch local feature fusion module for local feature extraction to obtain a feature map containing fine-grained local features;

[0025] A global feature extraction module is configured to: input the feature map containing fine-grained local features into a deep neural network module for global feature extraction and encoding to obtain a final feature map;

[0026] The stage prediction module is configured to: input the final feature map into a classifier to perform developmental stage prediction, and output the predicted embryonic development stage.

[0027] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a deep learning method for multi-stage classification of embryonic development as described in the first aspect.

[0028] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a deep learning method for multi-stage classification of embryonic development as described in the first aspect are implemented.

[0029] One or more of the above technical solutions have the following beneficial effects:

[0030] The present invention combines a dual-branch local feature fusion module and a Transformer deep neural network module, and the Transformer deep neural network module is a global feature extraction module. The dual-branch local feature fusion module is composed of a local self-attention module, a deep convolution module, and a cross-branch interactive fusion module, which is used to effectively extract local features from embryo images; the encoder in the Transformer module extracts global context features from the input feature map through multi-head self-attention and feedforward neural network. The present invention combines the advantages of these two modules for the classification of embryonic development stages, which can not only fully extract the local and global information of embryonic images and improve the accuracy of the classification stage, but also provides a new idea for the application of deep learning in IVF.

[0031] The present invention designs a new Transformer-based dual-branch local feature fusion network, which can extract local and global features from input embryo images, effectively improving the accuracy of multi-stage classification of embryonic development.

[0032] The present invention proposes a new dual-branch local feature fusion module for local feature extraction. The dual-branch local feature fusion module mainly includes a local self-attention module, a deep convolution module and a cross-branch interaction fusion module. The local self-attention module can capture the local spatial information of the input image, the deep convolution module can effectively extract channel features, and the cross-branch interaction and fusion module processes the feature interaction between the two branches. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0034] Figure 1 is a flow chart of a classification method according to an embodiment of the present invention;

[0035] Figure 2 This is an overall architecture diagram of a dual-branch local feature fusion network based on Transformer in an embodiment of the present invention;

[0036] Figure 3 (a) is a structural diagram of a dual-branch local feature fusion module according to an embodiment of the present invention. Figure 3 (b) is the local self-attention module diagram in the dual-branch local feature fusion module. Figure 3 (c) is the deep convolution module diagram in the dual-branch local feature fusion module. Figure 3 (d) is the cross-branch interactive fusion module diagram in the dual-branch local feature fusion module;

[0037] Figure 4 (a) is a structural diagram of a deep neural network module according to an embodiment of the present invention. Figure 4 (b) Diagram of the multi-head self-attention mechanism in the deep neural network module. DETAILED DESCRIPTION

[0038] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0039] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0040] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0041] Embodiment 1

[0042] like Figure 1 As shown, this embodiment discloses a deep learning method for multi-stage classification of embryonic development, which includes the following steps:

[0043] Step 1: acquiring an embryo image, and preprocessing the embryo image to obtain a preprocessed embryo image;

[0044] In this embodiment, the specific process of preprocessing the input embryo image is as follows: first, the embryo image is scaled from 500×500 pixels to 224×224 pixels to adapt to the network architecture, and then the scaled embryo image is data enhanced and randomly rotated to obtain the scaled and enhanced embryo image E.

[0045] Step 2: The preprocessed embryo image is divided into blocks and input into a dual-branch local feature fusion module for local feature extraction to obtain a feature map containing fine-grained local features;

[0046] In this embodiment, if Figure 2 As shown in FIG. 1 , the preprocessed embryo image E is processed by convolution and block processing, and the blocked embryo image is input into multiple dual-branch local feature fusion modules to output a feature map containing fine-grained local features.

[0047] like Figure 3 As shown in (a), the dual-branch local feature fusion module includes a first branch module, a second branch module and a cross-branch interactive fusion module. The local self-attention module in the first branch module can capture the local spatial information of the input image, the deep convolution module in the second branch module can effectively extract channel features, and the cross-branch interactive fusion module processes the feature interaction between the two branches.

[0048] Specifically, the first branch module and the second branch module are connected in parallel, and the input image is input into the first branch module and the second branch module for processing after normalization. The first branch module includes a fully connected layer and a local self-attention module. Preferably, the fully connected layer and the local self-attention module are connected in sequence to obtain the first branch module. The second branch module includes a fully connected layer and a deep convolution module. Preferably, the fully connected layer and the deep convolution module are connected in sequence to obtain the second branch module.

[0049] Furthermore, the input image is normalized before being input to the first branch module and the second branch module; similarly, the input is normalized before being input to the local self-attention module, the deep convolution module, and the multi-layer perceptron. Performing normalization can improve the convergence speed of the model, increase the stability and generalization ability of the model, prevent the problem of gradient disappearance or explosion, and balance the impact of different features on the model.

[0050] In the first branch module, the input image first passes through the fully connected layer, outputting the input tensor of the local self-attention module, marked as like Figure 3 (b) First, a 1×1 convolution increases the number of channels and generates a new representation Then will It is divided into three tensors: Q, K, and V. For tensors Q and K, the attention score is calculated by Hadamard product, and then the position encoding operation of the attention score is implemented through a 3×3 dilated convolution with a dilation rate of 1, a GELU activation function, and a 1×1 2D convolution. The position-encoded attention score is then passed through the softmax function and instance normalization (IN) to generate the IN attention (In_Attn) matrix. The specific expression is:

[0051] In_Attn=IN(softmax(Conv(Q*K T )))

[0052] Wherein, * represents the Hadamard product.

[0053] Next, the channel capacity of In_Attn is expanded through the Ghost head module. In the Ghost head module, In_Attn is first flattened along the second dimension, and then two learnable matrices ghost_mul and ghost_add are introduced, and then Hadamard product and element-by-element addition operations are performed on In_Attn and the two learnable matrices. The specific formula is:

[0054] G(In_Attn)=In_Attn*ghost_mul+ghost_add

[0055] Wherein, * represents the Hadamard product.

[0056] For the tensor V, it is first processed by the GLU activation function, and then the processing results of Q, K and V are aggregated through the Hadamard product. The GLU activation function can be implemented using this formula, specifically:

[0057] GLU(V)=σ(VW1+b)⊙σ(VW2+c)

[0058] Among them, W1 and W2 represent the learnable matrix weights, b and c represent the biases, σ represents the sigmoid function, and ⊙ is the element-by-element multiplication operation.

[0059] In the second branch module, the input image first passes through the fully connected layer and normalization operation, and outputs the input tensor of the deep convolution module, marked as like Figure 3 (c) First, an independent 3×3 convolution operation is performed on each channel dimension, and then a 1×1 point-by-point convolution is performed on the output results of the 3×3 convolution to perform a linear combination operation, and finally batch normalization (BN) and GELU activation function operations are performed to generate the output of the deep convolution module.

[0060] like Figure 3 As shown in (a), a cross-branch interaction fusion module is connected between the local self-attention module and the deep convolution module. In order to enhance the modeling capabilities of the local self-attention module and the deep convolution module in the spatial and channel dimensions and effectively integrate the features extracted from the two branch modules, i LoSA and X i DW A cross-branch interactive fusion module is added during the processing to realize cross-branch feature interaction and fusion operations.

[0061] like Figure 3 As shown in (d), the cross-branch interaction fusion module includes a parallel channel interaction module and a spatial interaction module, through which the features extracted by the deep convolution module are interacted to the local self-attention module, and the spatial relationship is transferred from the local self-attention module to the deep convolution module through the spatial interaction module. The channel interaction module consists of a pooling layer, two 1×1 convolution layers and a sigmoid function. Preferably, the pooling layer, the 1×1 convolution layer, the 1×1 convolution layer and the sigmoid function are connected in sequence to obtain the channel interaction module. The two 1×1 convolution layers are connected in sequence by batch normalization (BN) and GELU activation function. The spatial interaction module has a similar composition structure to the channel interaction module, and consists of two 1×1 convolution layers and a sigmoid function. Preferably, the 1×1 convolution layer, the 1×1 convolution layer and the sigmoid function are connected in sequence to obtain the spatial interaction module. The two 1×1 convolution layers are connected in sequence by BN and GELU activation functions.

[0062] Specifically, the features extracted by the deep convolution module are element-wise multiplied with the V tensor in the local self-attention module through the channel interaction module, while the spatial interaction module transfers the spatial relationship from the local self-attention module to the deep convolution module.

[0063] Further, if Figure 3As shown in (a), the feature maps processed by the local self-attention module and the deep convolution module are spliced ​​along the channel dimension. Subsequently, the cascade features are processed by the linear layer, i.e., the fully connected layer, and input into the multi-layer perceptron (MLP) for processing, and the result is used as the final result of the dual-branch local feature fusion module, i.e., the feature map containing fine-grained local features. Among them, the MLP consists of two convolutional layers with a kernel size of 1×1 and a GELU activation function. Furthermore, regularization processing, i.e., dropout operation, is performed after the fully connected layer and the multi-layer perceptron to prevent overfitting and improve the generalization ability of the model.

[0064] Step 3: Input the feature map containing fine-grained local features into a deep neural network module for global feature extraction and encoding to obtain a final feature map;

[0065] In this embodiment, if Figure 2 As shown in the figure, the feature map containing fine-grained local features is processed by convolution blocks, and the blocked feature map is input into multiple Transformer encoders to output the final feature map, which is linearly transformed through multiple fully connected layers to output the predicted embryonic development stage.

[0066] In this embodiment, if Figure 4 As shown in (a), the deep neural network module is specifically a Transformer-based global feature extraction encoding module (also called a Transformer encoder module). The encoder includes a multi-head self-attention mechanism and a feedforward neural network connected in sequence.

[0067] The feedforward neural network is composed of two fully connected layers separated by a GELU activation function and a dropout layer. Preferably, the fully connected layer, the GELU activation function, the dropout layer, the fully connected layer and the dropout layer are sequentially connected to obtain a feedforward neural network.

[0068] Before the feature map is input into the multi-head self-attention mechanism and the feedforward neural network, the input feature map is normalized.

[0069] like Figure 4 As shown in (b), the input feature map is a feature map containing fine-grained local features. The specific steps in the encoder are as follows: In the multi-head self-attention mechanism, the input feature map is first linearly transformed to generate three vectors Q, K, and V. Then the Q, K, and V vectors are equally divided into h parts along the channel dimension, respectively denoted as Q i ,K i ,V i ,(i=1,2,...,h), where h is the number of attention heads in the multi-head self-attention mechanism.

[0070] Then Q i and K i Perform element-wise multiplication and softmax operation to calculate the attention score and add the attention score and V i Aggregate to get the feature map of the i-th attention head. Finally, concatenate the outputs of all attention heads and process them through a linear layer to get the output of the multi-head self-attention mechanism.

[0071] like Figure 4 As shown in (a), the feedforward neural network further extracts features by performing nonlinear transformation and modeling on the mapping in the multi-head self-attention mechanism. After the feedforward neural network is processed, a more abstract feature representation is finally obtained, providing a more powerful feature expression for subsequent tasks.

[0072] Step 4: input the final feature map into a classifier to predict the developmental stage, and output the predicted embryonic developmental stage.

[0073] In this embodiment, if Figure 2 As shown in the figure, the final feature map is linearly transformed through a series of fully connected layers, and the attention weights are calculated through the softmax function, and matrix multiplication is performed with the final feature map for weighted aggregation. In the last layer, the fully connected layer maps the features to the output space and finally outputs the predicted embryonic development stage.

[0074] The above-mentioned dual-branch local feature fusion module and the Transformer-based global feature extraction encoding module constitute a new Transformer-based dual-branch local feature fusion network, which can extract local and global features from the input embryo image, effectively improving the accuracy of multi-stage classification of embryonic development.

[0075] To verify the effectiveness of the present invention, the present invention used a public embryo image dataset for experiments and compared it with a separate Transformer model. The results showed that the method disclosed in the present invention has better classification effect and can effectively improve the accuracy of multi-stage classification of embryonic development.

[0076] Embodiment 2

[0077] This embodiment provides a deep learning system for multi-stage classification of embryonic development, including:

[0078] An image acquisition module is configured to: acquire an embryo image, preprocess the embryo image, and obtain a preprocessed embryo image;

[0079] A local feature extraction module is configured to: perform a block operation on the preprocessed embryo image, and input the preprocessed embryo image into a dual-branch local feature fusion module for local feature extraction to obtain a feature map containing fine-grained local features;

[0080] A global feature extraction module is configured to: input the feature map containing fine-grained local features into a deep neural network module for global feature extraction and encoding to obtain a final feature map;

[0081] The stage prediction module is configured to: input the final feature map into a classifier to perform developmental stage prediction, and output the predicted embryonic development stage.

[0082] Embodiment 3

[0083] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of embodiment 1 when executing the program.

[0084] Embodiment 4

[0085] The purpose of this embodiment is to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the steps of the method of embodiment 1 are performed.

[0086] The steps involved in the apparatus of the above embodiments 3 and 4 correspond to the method embodiment 1, and the specific implementation method can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0087] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0089] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A deep learning method for multi-stage classification of embryonic development, characterized in that: include: Acquiring an embryo image, and preprocessing the embryo image to obtain a preprocessed embryo image; The preprocessed embryo image is divided into blocks and input into a dual-branch local feature fusion module for local feature extraction to obtain a feature map containing fine-grained local features. Inputting the feature map containing fine-grained local features into a deep neural network module for global feature extraction and encoding to obtain a final feature map; The final feature map is input into a classifier for developmental stage prediction, and the predicted embryonic developmental stage is output.

2. A deep learning method for multi-stage classification of embryonic development as claimed in claim 1, characterized in that: The dual-branch local feature fusion module includes a first branch module and a second branch module connected in parallel, the first branch module captures local spatial information of the input image, and the second branch module effectively extracts channel features.

3. A deep learning method for multi-stage classification of embryonic development as claimed in claim 2, characterized in that: The fully connected layer and the local self-attention module are connected in sequence to obtain the first branch module, and the fully connected layer and the deep convolution module are connected in sequence to obtain the second branch module.

4. A deep learning method for multi-stage classification of embryonic development as claimed in claim 3, characterized in that: A cross-branch interaction fusion module is connected between the local self-attention module and the deep convolution module to realize cross-branch feature interaction and fusion operations.

5. A deep learning method for multi-stage classification of embryonic development as claimed in claim 4, characterized in that: The cross-branch interactive fusion module includes a parallel channel interaction module and a spatial interaction module. The features extracted by the deep convolution module are interacted to the local self-attention module through the channel interaction module, and the spatial relationship is transferred from the local self-attention module to the deep convolution module through the spatial interaction module.

6. A deep learning method for multi-stage classification of embryonic development as claimed in claim 5, characterized in that: The pooling layer, 1×1 convolution layer, batch normalization, GELU activation function, 1×1 convolution layer and sigmoid function are connected in sequence to obtain the channel interaction module, and the 1×1 convolution layer, batch normalization, GELU activation function, 1×1 convolution layer and sigmoid function are connected in sequence to obtain the spatial interaction module.

7. A deep learning method for multi-stage classification of embryonic development as claimed in claim 1, characterized in that: The deep neural network module is specifically a Transformer encoder module.

8. A deep learning system for multi-stage classification of embryonic development, characterized in that: include: An image acquisition module is configured to: acquire an embryo image, preprocess the embryo image, and obtain a preprocessed embryo image; A local feature extraction module is configured to: perform a block operation on the preprocessed embryo image, and input the preprocessed embryo image into a dual-branch local feature fusion module for local feature extraction to obtain a feature map containing fine-grained local features; A global feature extraction module is configured to: input the feature map containing fine-grained local features into a deep neural network module for global feature extraction and encoding to obtain a final feature map; The stage prediction module is configured to: input the final feature map into a classifier to perform developmental stage prediction, and output the predicted embryonic development stage.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in a deep learning method for multi-stage classification of embryonic development as described in any one of claims 1-7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the deep learning method for multi-stage classification of embryonic development as described in any one of claims 1-7 are implemented.

Images