A Classification Method and System for Embryonic Development Stages Based on Residual Sensing and Wavelet Coding

Abstract

The application provides an embryo development stage classification method and system based on residual perception and wavelet coding, and belongs to the embryo development stage classification field.The method comprises the following steps: acquiring an embryo image and performing pretreatment; performing a blocking operation on the pretreated embryo image, inputting the blocked image data into a residual perception attention module, and extracting a first feature map of the embryo image; inputting the first feature map into a wavelet position coding module, and extracting a second feature map containing frequency domain information; adding the first feature map and the second feature map and inputting the sum into a trained deep neural network, performing global feature extraction and coding, and obtaining a final feature map; inputting the final feature map into a classifier for development stage classification, and outputting a classification result of the embryo development stage. The method effectively overcomes the problem of spatial information loss in the downsampling method, enhances the sensitivity of the model to the frequency domain information, and improves the accuracy of the embryo development stage classification.

Description

Technical Field

[0001] This invention belongs to the field of embryonic development stage classification, and in particular relates to a method and system for embryonic development stage classification based on residual sensing and wavelet coding. Background Technology

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

[0003] Infertility is a reproductive health issue affecting millions of couples worldwide. Currently, in-vitro fertilization (IVF) has become one of the most effective assisted reproductive technologies (ART) for resolving infertility. This technology involves combining eggs and sperm in a laboratory to create embryos, which are then transferred back into the uterus. During ART, the health and viability of the embryo directly impact implantation success rates and pregnancy stability. Embryo development is a complex and intricate process, divided into 16 consecutive stages. This detailed classification system allows embryologists to more accurately assess the developmental status of embryos, thereby selecting the embryos with the highest implantation potential and ensuring the success of ART.

[0004] In recent years, with the rapid development of deep learning neural networks, research applying deep learning technology to the field of assisted reproduction has begun to receive attention from the academic community. Convolutional Neural Networks (CNNs) and Transformers, for example, have emerged as leaders in the field of assisted reproduction due to their superior feature extraction capabilities and modeling flexibility. However, these CNN-based models primarily focus on local feature extraction, potentially neglecting global contextual information related to the overall morphology, structure, and tissue characteristics of the embryo. This information is crucial for accurately identifying and classifying embryos at different developmental stages. Similarly, Transformer-based methods may lose key spatial information during spatial downsampling (i.e., when segmenting the image into fixed-size convolutional blocks). This missing information plays a decisive role in distinguishing embryo types, and its absence weakens the model's discriminative ability in embryo image classification.

[0005] Furthermore, current deep learning-based embryo classification methods primarily rely on spatial domain features, potentially failing to fully utilize valuable information in the frequency domain. This could lead to deep learning models missing subtle morphological changes and developmental features in embryo images, thereby reducing their effectiveness in distinguishing different embryonic developmental stages and ultimately affecting overall classification accuracy. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for classifying embryonic development stages based on residual perception and wavelet coding. Building upon the Transformer deep neural network, it combines residual perception attention and wavelet position coding for embryonic development stage classification. This not only effectively reduces the loss of important features during Transformer downsampling and maintains information integrity as much as possible, but also integrates frequency domain information from embryonic images, improving the model's ability to perceive details and differences in embryos at different developmental stages, thereby increasing the accuracy of embryonic development stage classification.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] The first aspect of this invention provides a method for classifying embryonic developmental stages based on residual sensing and wavelet coding;

[0009] Embryonic development stage classification methods based on residual sensing and wavelet coding include:

[0010] Acquire an embryo image, and preprocess the embryo image to obtain a preprocessed embryo image;

[0011] The preprocessed embryo image is segmented to obtain segmented image data. The segmented image data is then input into the residual perception attention module to extract the first feature map of the embryo image.

[0012] The first feature map is input into the wavelet position coding module to extract the second feature map containing frequency domain information;

[0013] The first feature map and the second feature map are added together and then input into the trained deep neural network for global feature extraction and encoding to obtain the final feature map.

[0014] The final feature map is input into a classifier to classify developmental stages, and the classification results of embryonic development stages are output.

[0015] As a further technical solution, the process of preprocessing the embryo image to obtain a preprocessed embryo image includes:

[0016] Scaling the embryo image reduces the pixel value;

[0017] Data augmentation is performed on the scaled embryo image using random rotation to obtain the scaled and augmented embryo image.

[0018] As a further technical solution, the preprocessed embryo image is segmented to obtain segmented image data. The segmented image data is then input into a residual perception attention module to extract the first feature map of the embryo image, including:

[0019] The preprocessed embryo images are downsampled to obtain downsampled feature maps;

[0020] The downsampled feature maps are input into the channel attention branch and spatial attention branch of the residual perception attention module, respectively, to obtain channel attention features and spatial attention features;

[0021] The channel attention features and spatial attention features are multiplied together and then fused with the segmented image data through skip connections to obtain the first feature map of the embryo image.

[0022] As a further technical solution, the process of obtaining spatial attention features using spatial attention branches includes:

[0023] The downsampled feature map is then subjected to mean and reshape operations to obtain the reshaped feature map.

[0024] Normalize and standardize each sample of the reshaped feature map to obtain a feature map with zero mean and unit variance;

[0025] The spatial attention mapping coefficients are obtained by multiplying the feature maps with zero mean and unit variance with learnable weight parameters and adding a bias term, and then passing the Sigmoid function.

[0026] The spatial attention mapping coefficients are multiplied element-wise with the segmented image data to obtain spatial attention features.

[0027] As a further technical solution, the first feature map is input into the wavelet position coding module to extract a second feature map containing frequency domain information, including:

[0028] The first feature map is processed by the frequency information extraction branch of the wavelet position coding module to obtain the first branch feature map with fused frequency domain information; the first feature map is processed by the feature enhancement convolution branch to obtain the second branch feature map with enhanced detail features; and the first feature map is processed by the detail preservation convolution branch to obtain the third branch feature map that preserves the original detail information.

[0029] The first branch feature map and the second branch feature map are fused by element-wise multiplication to obtain a fused feature map;

[0030] The fused feature map is fused with the third branch feature map through element-wise addition to obtain a second feature map containing frequency domain information.

[0031] As a further technical solution, the first feature map and the second feature map are added together and then input into a trained deep neural network for global feature extraction and encoding to obtain the final feature map, including:

[0032] The trained deep neural network uses a Transformer neural network to normalize the result of adding the first feature map and the second feature map.

[0033] The normalized result is input into the multi-head self-attention mechanism to calculate the attention score and obtain the output of the multi-head self-attention mechanism. The output of the multi-head self-attention mechanism is a weighted feature representation that integrates the attention information of different attention heads.

[0034] The weighted feature representation is nonlinearly transformed and modeled using a feedforward neural network to obtain the final feature map.

[0035] As a further technical solution, the final feature map is input into a classifier for developmental stage classification, and the classification result of the embryonic development stage is output, including:

[0036] The final feature map is linearly transformed through a fully connected layer to map the features to the output space, and the classification result of the embryonic development stage is output using the output layer.

[0037] A second aspect of the present invention provides a classification system for embryonic development stages based on residual sensing and wavelet coding.

[0038] An embryonic development stage classification system based on residual sensing and wavelet coding includes:

[0039] The image preprocessing module is configured to: acquire an embryo image, preprocess the embryo image, and obtain a preprocessed embryo image;

[0040] The feature map extraction module is configured to: perform a block operation on the preprocessed embryo image to obtain block image data; input the block image data into the residual perception attention module to extract the first feature map of the embryo image; and input the first feature map into the wavelet position coding module to extract the second feature map containing frequency domain information.

[0041] The global feature extraction module is configured to: add the first feature map and the second feature map together and input them into the trained deep neural network to perform global feature extraction and encoding to obtain the final feature map;

[0042] The stage classification module is configured to input the final feature map into a classifier to classify the developmental stages and output the classification results of the embryonic development stages.

[0043] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the embryonic development stage classification method based on residual sensing and wavelet coding as described in the first aspect of the present invention.

[0044] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the embryonic development stage classification method based on residual sensing and wavelet coding as described in the first aspect of the present invention.

[0045] The above one or more technical solutions have the following beneficial effects:

[0046] (1) The residual perception attention module in this invention can minimize the information loss caused by the reduction in resolution during downsampling and efficiently retain the key feature information in the embryo image; the wavelet position coding module can enhance the model's ability to perceive the local fine structure and overall morphological features in the embryo image; the encoder in the Transformer module extracts global context features from the input feature map through multi-head self-attention and feedforward neural network. By combining the advantages of these three modules for the classification of embryonic development stages, the problem of spatial information loss in downsampling methods can be effectively overcome, and the model's sensitivity to frequency domain information can be enhanced.

[0047] (2) The residual perception attention module in this invention includes channel attention branch, spatial attention branch and skip connection branch. The channel attention branch enables the model to identify and enhance important features in each channel, effectively preventing information loss; the spatial attention branch helps the network focus on key regions in the embryo image; and the skip connection can directly pass high-resolution features to subsequent network layers, thereby preserving key information.

[0048] (3) The wavelet position coding module in this invention includes a frequency information extraction branch, a feature enhancement convolution branch, and a detail preservation convolution branch. The frequency information extraction branch uses wavelet transform to extract low-frequency and high-frequency components from the embryo feature map; the feature enhancement convolution branch is used to enhance complex detail features in the image; and the detail preservation convolution branch can preserve the original detail information of the input features.

[0049] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0050] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0051] Figure 1 This is a flowchart of the method in the first embodiment.

[0052] Figure 2 This is an overall architecture diagram of the method used in the first embodiment.

[0053] Figure 3 This is a structural diagram of the residual perception attention module in the first embodiment.

[0054] Figure 4 This is a structural diagram of the wavelet position coding module in the first embodiment.

[0055] Figure 5 This is a structural diagram of the deep neural network module in the first embodiment.

[0056] Figure 6 This is a structural diagram of the multi-head self-attention mechanism in the deep neural network module of the first embodiment.

[0057] Figure 7 This is a system structure diagram of the second embodiment. Detailed Implementation

[0058] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0059] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0060] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0061] Example 1

[0062] This embodiment discloses a method for classifying embryonic developmental stages based on residual sensing and wavelet coding;

[0063] Combination Figure 1 and Figure 2 Methods for classifying embryonic developmental stages based on residual sensing and wavelet coding include:

[0064] Step S1: Obtain an embryo image and preprocess the embryo image to obtain a preprocessed embryo image;

[0065] Step S2: Perform block segmentation on the preprocessed embryo image to obtain block image data. Input the block image data into the residual perception attention module to extract the first feature map of the embryo image.

[0066] Step S3: Input the first feature map into the wavelet position coding module to extract the second feature map containing frequency domain information;

[0067] Step S4: Add the first feature map and the second feature map together and input them into the trained deep neural network to perform global feature extraction and encoding to obtain the final feature map;

[0068] Step S5: Input the final feature map into the classifier for developmental stage classification and output the embryonic developmental stage classification result.

[0069] Specifically, it also includes the following:

[0070] In step S1, the preprocessing of the input embryo image is as follows: First, the embryo image is scaled from 500×500 pixels to 224×224 pixels to fit the network architecture. Then, the scaled embryo image is subjected to data augmentation and random rotation to obtain the scaled and augmented embryo image EM.

[0071] In step S2, the preprocessed embryo image EM undergoes convolution block processing to obtain block-based image data. The segmented embryo image data is then input into the residual perception attention module, which outputs the first feature map of the embryo image.

[0072] Among them, combined Figure 3 The residual perception attention module includes a first branch, a second branch, and a third branch. The first branch is a channel attention branch, which enables the model to identify and enhance important features in each channel; the second branch is a spatial attention branch, which helps the network focus on key regions in the image, preserving key spatial information even when the resolution is reduced; the third branch is a skip connection branch, which can directly pass high-resolution features to subsequent network layers without downsampling, thereby effectively preserving key information and reducing information loss caused by reduced resolution.

[0073] Specifically, the first and second branches are connected in parallel, and the input image data is divided into blocks. After downsampling, the downsampled feature map is obtained. The input signals are then fed into the first and second branches for processing. The first branch includes adaptive average pooling, reshaping, normalization, standardization, learnable weight multiplication, bias term addition, and a sigmoid activation function. The second branch includes meanization, reshaping, normalization, standardization, learnable weight multiplication, bias term addition, and a sigmoid activation function, and then jumps to the third branch to perform element-wise multiplication of the input signals.

[0074] Furthermore, the output features of the first branch are multiplied by the output features of the second branch and added to the input features of the third branch to obtain the first feature map of the embryo image.

[0075] In the channel attention branch, the downsampled feature map First, an adaptive average pooling layer is used to perform global averaging along the channel dimension, thereby obtaining the feature map. The specific expression for the calculation process of adaptive average pooling is as follows:

[0076]

[0077] Where H and W represent feature maps respectively. Height and width. Representation of feature map The element located in row j and column k.

[0078] For feature maps Perform a reshape operation to obtain Where B is the batch size and C is the number of input feature channels. Subsequently, the data for each channel undergoes normalization and standardization to obtain feature maps with zero mean and unit variance. and To ensure the consistency and stability of the channel data distribution, the standardized channel features are multiplied by learnable weights, and a bias term is added. Then, a non-linear transformation is performed using the sigmoid activation function to obtain the channel attention features for this branch. .

[0079] In the spatial attention branch, the downsampled feature map First, a mean operation is performed along the channel dimension to obtain the feature map. This reduces the spatial size of the data. The specific expression for this mean operation is:

[0080]

[0081] in, Represents the feature map during downsampling. The element of the nth channel.

[0082] The feature map is reshaped using the reshape operation. By reorganizing and changing its dimensions, we can obtain... Then, each sample is normalized and standardized to obtain a feature map with zero mean and unit variance. and To ensure the consistency and stability of the feature distribution, the processed features are multiplied by learnable weight parameters and a bias term is added. Then, they are processed using the sigmoid function to obtain spatial attention mapping coefficients. These coefficients reflect the importance and influence of each spatial location in the entire feature map, thus helping the model better utilize the spatial structure information of the input data. Finally, the spatial attention mapping coefficients are applied to the block-wise image data. Perform element-wise multiplication to obtain the spatial attention features of the spatial attention branch. , where B, H and W represent the batch normalization size, height and width of the input features, respectively.

[0083] In the skip connection branch, the channel attention features and spatial attention features are multiplied and then combined with the segmented image data again through the skip connection. The images are fused to obtain the first feature map of the embryo image. This jump connection preserves the integrity of the original input features while retaining the channel and spatial characteristics, ensuring the effective transmission of information.

[0084] In step S3, the first feature map extracted by the residual perception attention module is... The data is input into the wavelet position encoding module to capture subtle structural changes in the embryo feature map.

[0085] Among them, such as Figure 4 As shown, the wavelet positional encoding module also includes a first branch, a second branch, and a third branch. The first branch is the frequency information extraction branch, which can provide structural features at different scales, enabling the model to analyze embryo images more comprehensively. The second branch is the feature-enhancing convolution branch, which enhances the model's ability to extract complex information from images. The third branch is the detail-preserving convolution branch, which can preserve the original details of the input features, avoiding the loss of detail caused by downsampling, nonlinear activation, or multi-layer convolution.

[0086] In the frequency information extraction branch, the first feature map of the input... Perform a reshape operation to make it into Then, the discrete wavelet transform (DWT) is applied to it for frequency domain decomposition, resulting in four wavelet sub-bands. In this embodiment, the Haar wavelet is used for the DWT transform.

[0087] Specifically, firstly, low-pass and high-pass filters are used along the row direction to... Decomposed into two sub-bands and Then, for the two sub-bands respectively and Further low-frequency and high-frequency filtering is applied, resulting in four wavelet sub-bands. These four wavelet sub-bands respectively capture the low-frequency to high-frequency signals of the overall structure. and low-frequency to high-frequency signals of large-scale information And high-frequency-low-frequency signals that capture fine-grained details. High-frequency signals and edge information Subsequently, these sub-bands are spliced ​​along the channel dimension to form a new feature map. Next, regarding Applying a 3×3 two-dimensional convolution to enhance feature representation yields... In order to effectively reintegrate frequency domain information into the feature map, for Perform inverse discrete wavelet transform (iDWT) to obtain feature maps that fuse spatial and frequency domain information. Finally, the feature map that integrates spatial and frequency domain information is analyzed. The first branch feature map with fused frequency domain information is obtained by performing a reshape operation. This is to ensure that it is suitable for processing subsequent one-dimensional convolutional layers.

[0088] In the feature enhancement convolution branch and the detail preservation convolution branch, the first feature map of the input is first processed. Perform the flattening operation.

[0089] The feature enhancement convolution branch consists of flattening operations, one-dimensional convolution, BN, and ReLU activation functions, used to enhance complex detail features in the image, yielding the following results: , and And output the feature map of the second branch. Subsequently, the feature map of the second branch... With the first branch feature map Perform element-wise multiplication to obtain the fused feature map. .

[0090] The detail-preserving convolutional branch consists of a flattening operation and a single-layer one-dimensional convolution with a kernel size of 1, used to preserve the original detail information of the input features, resulting in a third branch feature map that retains the original detail information. .

[0091] Furthermore, the feature maps will be fused. With the third branch feature map By fusing elements-wise, followed by one-dimensional convolution and reshape operations, a second feature map containing frequency domain information is obtained. .

[0092] In step S4, the first feature map and the second feature map are added together and then input into the trained deep neural network for global feature extraction and encoding to obtain the final feature map.

[0093] like Figure 5 As shown, the trained deep neural network uses a Transformer neural network, specifically a Transformer-based global feature extraction and encoding module (also known as a Transformer encoder module). The encoder comprises a multi-head self-attention mechanism and a feedforward neural network connected sequentially.

[0094] A feedforward neural network consists 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 connected in the following order to obtain the feedforward neural network.

[0095] Before inputting the feature map obtained by adding the first feature map and the second feature map into the multi-head self-attention mechanism and the feedforward neural network, the input feature map is normalized.

[0096] like Figure 6 As shown, the specific steps in the encoder for obtaining the feature map after adding the inputs are as follows: In the multi-head self-attention mechanism, the input feature map first undergoes a linear transformation 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, denoted as h, ... , where h is the number of attention heads in the multi-head self-attention mechanism.

[0097] Then on and Perform element-wise multiplication, scaling, and softmax operations to calculate the attention score, and then sum the attention score with... The feature map of the i-th attention head is obtained by aggregation. Finally, the outputs of all attention heads are concatenated and processed through a linear layer to obtain the output of the multi-head self-attention mechanism, which is a weighted feature representation that integrates the attention information from different attention heads.

[0098] Feedforward neural networks further extract features by performing nonlinear transformations and modeling on the weighted feature representation of the multi-head self-attention mechanism. After processing by the feedforward neural network, a more abstract feature representation is obtained, resulting in the final feature map.

[0099] In step S5, the final feature map is input into the classifier for developmental stage classification, and the classification result of the embryonic development stage is output.

[0100] In this embodiment, the final feature map is linearly transformed through a series of fully connected layers to map the features to the output space, and the predicted embryonic development stage is output using the output layer.

[0101] The residual perception attention module and wavelet position coding module mentioned above can overcome the problem of spatial information loss in downsampling methods, enhance the model's sensitivity to frequency domain information, and thus effectively improve the model's classification performance during embryonic development.

[0102] To verify the effectiveness of the present invention, an experiment was conducted using a publicly available embryo image dataset. As shown in Table 1, the results show that the classification accuracy of the method disclosed in this invention is 71.0%, which is better than the classification accuracy of 66.7% of the Transformer model alone. The method disclosed in this invention can effectively improve the classification accuracy of embryonic development stages.

[0103] Table 1 Performance comparison between the method of the present invention and the Transformer model

[0104]

[0105] The formulas for calculating accuracy (ACC), precision (P), recall (R), and F1 score are as follows:

[0106] Accuracy = (TP + TN) / (TP + TN + FP + FN)

[0107] Precision = TP / (TP + FP)

[0108] Recall rate = TP / (TP + FN)

[0109] F1 score = (2 × precision × recall) / (precision + recall)

[0110] In this context, TP, FN, FP, and TN represent true examples, false counterexamples, false positive examples, and true counterexamples, respectively.

[0111] Example 2

[0112] This embodiment discloses an embryonic development stage classification system based on residual sensing and wavelet coding;

[0113] like Figure 7 As shown, the embryonic development stage classification system based on residual sensing and wavelet coding includes:

[0114] The image preprocessing module is configured to: acquire an embryo image, preprocess the embryo image, and obtain a preprocessed embryo image;

[0115] The feature map extraction module is configured to: perform a block operation on the preprocessed embryo image to obtain block image data; input the block image data into the residual perception attention module to extract the first feature map of the embryo image; and input the first feature map into the wavelet position coding module to extract the second feature map containing frequency domain information.

[0116] The global feature extraction module is configured to: add the first feature map and the second feature map together and input them into the trained deep neural network to perform global feature extraction and encoding to obtain the final feature map;

[0117] The stage classification module is configured to input the final feature map into a classifier for developmental stage classification and output the embryonic development stage classification result.

[0118] Example 3

[0119] The purpose of this embodiment is to provide a computer-readable storage medium.

[0120] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the embryonic development stage classification method based on residual sensing and wavelet coding as described in Example 1.

[0121] Example 4

[0122] The purpose of this embodiment is to provide an electronic device.

[0123] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the embryonic development stage classification method based on residual sensing and wavelet coding as described in Embodiment 1.

[0124] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section 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 as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0125] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0126] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this 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 without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for classifying the stage of embryonic development based on residual perception and wavelet coding, characterized in that, The method comprises the following steps: An embryo image is acquired, and the embryo image is preprocessed to obtain a preprocessed embryo image; The preprocessed embryo image is subjected to a blocking operation to obtain blocked image data, and the blocked image data is input into a residual perception attention module to extract a first feature map of the embryo image. Specifically, the preprocessed embryo image is down-sampled to obtain a down-sampled feature map; the down-sampled feature map is input into a channel attention branch and a spatial attention branch of the residual perception attention module to obtain a channel attention feature and a spatial attention feature; The channel attention feature and the spatial attention feature are multiplied, and are fused with the blocked image data through a skip connection to obtain the first feature map of the embryo image; The first feature map is input into a wavelet position encoding module to extract a second feature map containing frequency domain information. Specifically, a frequency information extraction branch of the wavelet position encoding module processes the first feature map to obtain a first branch feature map fused with frequency domain information; a feature enhancement convolution branch processes the first feature map to obtain an enhanced detail feature second branch feature map; a detail retention convolution branch processes the first feature map to obtain a third branch feature map retaining original detail information; the first branch feature map and the second branch feature map are fused through element-wise multiplication to obtain a fusion feature map; the fusion feature map and the third branch feature map are fused through element-wise addition to obtain the second feature map containing frequency domain information; The first feature map and the second feature map are added and input into a trained deep neural network to perform global feature extraction and coding to obtain a final feature map; The final feature map is input into a classifier for development stage classification, and a classification result of an embryo development stage is output.

2. The method for classifying the stage of embryonic development based on residual perception and wavelet coding according to claim 1, wherein, The process of preprocessing the embryo image to obtain the preprocessed embryo image comprises: The embryo image is subjected to a scaling operation to reduce pixel values; The scaled embryo image is subjected to a data enhancement operation, and a scaled and enhanced embryo image is obtained through random rotation.

3. The method for classification of embryonic development stage based on residual perception and wavelet coding as claimed in claim 1, wherein, The process of obtaining the spatial attention feature by using the spatial attention branch comprises: The down-sampled feature map is subjected to a mean operation and a reshape operation in sequence to obtain a reshaped feature map; Each sample of the reshaped feature map is subjected to normalization and standardization processing to obtain a feature map with zero mean and unit variance; The feature map with zero mean and unit variance is multiplied by a learnable weight parameter, and a bias term is added to obtain a spatial attention mapping coefficient through a Sigmoid function; The spatial attention mapping coefficient is subjected to element-wise multiplication operation with the blocked image data to obtain the spatial attention feature.

4. The method for classifying the stage of embryonic development based on residual perception and wavelet coding according to claim 1, wherein, The first feature map and the second feature map are added and input into the trained deep neural network to perform global feature extraction and coding to obtain the final feature map, which comprises: The trained deep neural network adopts a Transformer neural network, and the result of adding the first feature map and the second feature map is subjected to a normalization operation; The normalized result is input into a multi-head self-attention mechanism, attention scores are calculated, and an output of the multi-head self-attention mechanism is obtained, which is a weighted feature representation that fuses information focused by different attention heads; The weighted feature representation is nonlinearly transformed and modeled by using a feedforward neural network, and a final feature map is obtained.

5. The method for classification of embryonic development stage based on residual perception and wavelet coding as claimed in claim 1, wherein, The final feature map is input into a classifier for development stage classification, and a classification result of an embryonic development stage is output, including: The final feature map is linearly transformed through a fully connected layer to map the feature to an output space, and an output layer is used to output the classification result of the embryonic development stage.

6. Embryonic development stage classification system based on residual perception and wavelet coding, characterized in that, The image preprocessing module is configured to: The feature map extraction module is configured to: The channel attention feature and the spatial attention feature are multiplied and fused with the blocked image data through a skip connection to obtain the first feature map of the embryonic image. The first feature map is input into a wavelet position encoding module to extract a second feature map containing frequency domain information. Specifically, the frequency information extraction branch of the wavelet position encoding module is used to process the first feature map to obtain a first branch feature map that fuses frequency domain information; the feature enhancement convolution branch is used to process the first feature map to obtain an enhanced detail feature second branch feature map; the detail preservation convolution branch is used to process the first feature map to obtain a third branch feature map that preserves original detail information; the first branch feature map and the second branch feature map are fused through element-wise multiplication to obtain a fusion feature map; and the fusion feature map and the third branch feature map are fused through element-wise addition to obtain the second feature map containing frequency domain information. The global feature extraction module is configured to: The stage classification module is configured to: The program is executed by the processor to implement the steps in the embryonic development stage classification method based on residual perception and wavelet coding in any one of claims 1-5.

7. A computer-readable storage medium having stored thereon a program, characterized in that, The processor executes the program to implement the steps in the embryonic development stage classification method based on residual perception and wavelet coding in any one of claims 1-5.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized by ​

Images