Best focus embryo image screening method based on ConvNeXt network fused with position information

Abstract

This invention discloses a method for selecting optimally focused embryonic images based on a ConvNeXt network with fused location information. The method includes: 1) combining two grayscale embryonic images to create dual-channel combined data that conforms to the model input pattern; 2) subtracting the data from the two channels of the dual-channel combined data to obtain a difference image and downsampling it; 3) using a convolutional neural network to extract the overall features of the difference image, obtaining the corresponding feature matrix; 4) incorporating the prior information of the location index of the original embryonic image corresponding to the dual-channel combined data into the feature matrix; 5) constructing a loss function and training a binary classification model for optimally focused images; and 7) using the trained binary classification model to process a set of embryonic images to achieve the selection of optimally focused images. This invention can effectively analyze the focus of embryonic images, demonstrates excellent performance in the task of selecting optimally focused embryonic images, and has a certain degree of interpretability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for selecting the best-focused embryo images based on a supervised learning algorithm, which aims to locate the best-focused embryo image from a set of embryo images. Background Technology

[0002] In the rapid development of assisted reproductive technology, in-vitro fertilization (IVF) and embryo culture have become particularly popular. However, embryo quality assessment remains a crucial and complex task. The accuracy of the assessment directly affects the success rate of embryo transfer. Therefore, how to accurately determine the embryonic development stage using scientific and objective methods to select high-quality embryos for optimal transfer results is one of the core issues in current assisted reproductive technology research.

[0003] In assisted reproductive technology (ART) practice, high-quality embryo images are fundamental to embryo assessment. Physicians need to observe the morphological characteristics of embryo images to determine its developmental status, including cell division, morphological integrity, and internal structure. However, obtaining high-quality images is not easy. Factors affecting image quality include the performance of the microscopic imaging equipment, the stability of sample preparation, and the complexity of the imaging environment. Clear embryo images with appropriate contrast provide physicians with reliable information, while blurry or significantly inaccurate images can lead to uncertainty and potential misjudgments. Therefore, improving the efficiency and accuracy of embryo image acquisition and screening has become a crucial step in enhancing the quality of embryo assessment.

[0004] In current embryo image acquisition processes, time-series incubators have become a commonly used tool. The microscope within the incubator typically takes multiple automatic images of the embryo sample at set time intervals, acquiring images from different focal planes in each shot. This multi-focal-plane imaging method mitigates the shortcomings of manual focusing to some extent, especially reducing image incompleteness caused by microscope focusing errors. However, this technology also presents new challenges: each embryo sample folder usually contains a large number of out-of-focus images that cannot be directly used for evaluation. Physicians need to manually select the embryo images at the optimal focal plane—the best-focused embryo images—from the folders for subsequent analysis and decision-making. This manual selection process is time-consuming and labor-intensive, especially when dealing with large volumes of embryo data, where the workload increases exponentially. Summary of the Invention

[0005] This invention addresses the difficulties encountered in the aforementioned scenario by proposing an optimally focused embryo image selection method based on a ConvNeXt network with fused location information. This method aims to automatically analyze the focus of embryo images, thereby effectively selecting the optimally focused embryo images from a set of embryo images and making the results highly interpretable.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] This invention discloses an optimal focusing embryo image screening method based on a ConvNeXt network with fused location information, characterized by the following steps:

[0008] Step 1: Obtain an RGB three-channel embryo image dataset ,make The Middle Embryo image data are denoted as ,and ,in, express The first in Zhang embryo image, express The number of embryo images in the middle It is the height of an embryo image. It is the width of an embryo image; (Note: The original text appears to be incomplete and contains several grammatical errors. for The first in The image shows an embryo, and it is the best-focused image. ;

[0009] Step 2: After grayscale conversion, the first Group of grayscale embryo image data ,in, express Grayscale embryo images, express Grayscale embryo images;

[0010] Step 3: Put Separately with the previous A grayscale embryo image By combining them, we obtain the first channel category. Group embryo image dual-channel combined data ,in, For the first channel category The first in the group Combined data of two channels of embryo images, , and , Indicates the preceding The first grayscale embryo image Several grayscale embryo images, among which... The data for the first channel consists of grayscale images. constitute; The data for the second channel is derived from the grayscale image. constitute; The actual channel category label is denoted as ;

[0011] Step 4: Separately and after A grayscale embryo image By combining them, we obtain the second channel category. Group embryo image dual-channel combined data ,in, For the second channel category The first in the group Two-channel combined data of embryo images, in which... , and , After indicating The first grayscale embryo image Several grayscale embryo images, among which... The data for the first channel consists of grayscale images. constitute; The data for the second channel is derived from the grayscale image. constitute; The actual channel category label is denoted as ;

[0012] Step 5: Construct a ConvNeXt network that integrates location information, and then... and Process them separately to obtain the corresponding results. Binary classification probability and Binary classification probability ,in, and They are The probability of belonging to channel category 1 and channel category 2, and and They are The probability of belonging to channel category 1 and channel category 2;

[0013] Step 6: Construct the total loss function of the ConvNeXt network using equation (1) :

[0014] (1)

[0015] In equation (1), Denotes the first loss function. This represents the second loss function;

[0016] Step 7: Minimize the total loss function using the backpropagation algorithm and optimizer. The ConvNeXt network, which integrates location information, is used to train and update its parameters. This process continues until a preset maximum number of training epochs or a total loss function is reached. Upon convergence, training is stopped, and the best-focused image binary classification model is obtained after training, which is used to select the best-focused image.

[0017] The present invention discloses an optimal focusing embryo image screening method based on a ConvNeXt network with fused location information, characterized in that the ConvNeXt network with fused location information in step 5 includes: an image difference module. One ConvNeXt module and a fully connected layer;

[0018] Step 5.1: The image difference module will... After subtracting the data from the two channels, we obtain the first channel category. The first in the group Differential images ;

[0019] Step 5.2: Differential Image via The ConvNeXt module processes the data, ultimately yielding the first channel category. The first in the group A process The first ConvNeXt module processed the [number]th Feature matrices ;

[0020] Step 5.3: For Perform a flattening operation to obtain a flattened one-dimensional vector. ,in, Represents a one-dimensional vector The length, and ;Will The data is fed into a fully connected layer for processing to obtain dual-channel combined data. Binary classification probability ,in, and They are The probability of belonging to channel category 1 and channel category 2;

[0021] Step 5.4: Following the process in steps 5.1-5.3, combine the dual-channel data. The input location information is processed in a ConvNeXt network to obtain... Binary classification probability ,in and They are The probability of belonging to channel category 1 and channel category 2.

[0022] Furthermore, in step 5.2... The ConvNeXt module includes: the first The downsampling module, the first The location information fusion module, the first The nth depthwise convolutional module, the nth The first layer normalization module, the first The first layer scaling module, the first The path discard module, the first One residual connection module; ;

[0023] when At that time, the difference image Enter the first In the first ConvNeXt module, by the first Each downsampling module is scaled according to the scaling ratio. right Perform downsampling to obtain the first channel category's... The first in the group The first one after being processed by the downsampling module downsampled differential images ,in, Indicates the height of the downsampled difference image. , This represents the width of the downsampled differential image. , ;

[0024] No. Each deep convolutional module Perform channel expansion processing to obtain the first channel category's... The first in the group The first one after being processed by the depthwise convolution module A depthwise convolutional feature matrix ,in, The number of channels in the feature matrix. ;

[0025] No. The location information fusion module calculates Location information and to By expanding the dimensions, we obtain the first... Extended location information Then Extended to Having the same number of channels , obtained the The location information after further expansion Thus and After adding them, we get the first channel category. The first in the group The first fused location information Location information feature matrix ;

[0026] No. Each layer normalization module calculates The Middle The average of each channel and variance and obtained Specifically, the calculation of the first channel category... The first in the group After normalization of the first layer The first channel Feature matrices Thus, the first channel category can be obtained. The first in the group After normalization of the first layer Each layer of normalized feature matrix ,in, It is a constant to prevent the denominator from being zero. ;

[0027] No. Each layer scaling module applies the feature matrix Perform channel scaling, reducing its number of channels from Scaled down to 1, thus obtaining the first channel category's... The first in the group The first one after being processed by the layer scaling module Each layer scaled feature matrix Specifically, Through the The result is obtained by weighted summation along the channel dimension: .in It is the first The learnable weights corresponding to each channel ;

[0028] No. The path discarding module uses a Bernoulli distribution to generate a path that is similar to... Random tensors with the same dimensions and shape and will Each element in the array is divided by the set path drop probability. Then, the first channel category is obtained. The first in the group The first one after the above operations Discarding tensors along each path Thus, the first channel category is obtained. The first in the group The first one after being processed by the path discarding module Discard feature matrix of each path ,in , ; This represents the element-wise multiplication of two feature matrices with the same shape.

[0029] No. Each residual connection module discards the path feature matrix. Compared with downsampled differential image Add the corresponding elements together to obtain the first channel category. The first in the group The first one after processing by the residual connection module Each residual feature matrix ,in, ;

[0030] when At that time, the first channel category will be... The first in the group The first one after the The first ConvNeXt module processed the [number]th Feature matrices Enter the number The first ConvNeXt module is used for processing, thus ultimately obtaining the first channel category. The first in the group The first one after the The first ConvNeXt module processed the [number]th Feature matrices ,in, express height, express The width.

[0031] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the optimal focused embryo image screening method, and the processor is configured to execute the program stored in the memory.

[0032] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the optimal focused embryo image screening method.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. This invention employs techniques such as image grayscale conversion, downsampling, and image differencing to effectively reduce the computational overhead of the model. Specifically, image grayscale conversion converts color images into grayscale images, reducing the dimensionality and complexity of image data, lowering the computational load for subsequent processing, while preserving the basic structure and texture information of the image, laying the foundation for subsequent feature extraction. Downsampling further reduces the size of image data by decreasing the image resolution, accelerating the model's processing speed, improving its operating efficiency, and enabling the model to process more embryo image data in a shorter time, thereby improving overall screening efficiency.

[0035] 2. This invention cleverly utilizes image differencing technology to highlight the difference in focus between two embryo images. By performing differencing operations on the two embryo images, the changes in focus can be clearly displayed. This difference provides important features for model training, thereby improving the model's ability to extract effective features from dual-channel combined data. This improved feature extraction capability enables the model to more accurately select the best-focused embryo image when processing complex multi-focal plane embryo image data.

[0036] 3. This invention utilizes the ConvNeXt network to extract data features, significantly improving the performance of embryo image data processing and analysis. The ConvNeXt network extracts features from the input embryo image layer by layer through depthwise convolution operations. Each convolutional operation captures different features in the embryo image data. This multi-layered feature extraction method enables the ConvNeXt network to comprehensively analyze embryo image data. Furthermore, this network incorporates techniques such as layer normalization, residual connections, and path dropping to further enhance the network's training effectiveness and stability.

[0037] 4. This invention enhances the model's ability to fully extract prior knowledge of embryonic image data by incorporating prior information about embryonic image location into the model, thereby improving the final focusing accuracy. This results in superior performance in the optimal focusing embryonic image selection task and strong interpretability. Specifically, when processing embryonic image data, this method considers not only the features of individual embryonic images but also the relationships between images and positional index information, thus more accurately determining the optimal focal plane. This method improves focusing accuracy, makes the selection results more reliable, and also enhances the model's interpretability. Attached Figure Description

[0038] Figure 1 A schematic diagram of the dataset construction;

[0039] Figure 2 This is an overall flowchart of the method of the present invention;

[0040] Figure 3 This is a detailed diagram of the testing process for a set of embryo images. Detailed Implementation

[0041] In this example, a method for selecting optimal-focused embryo images based on a ConvNeXt network with fused location information uses ConvNeXt as the base model. The location information of the embryo images is incorporated into the model's training process. Furthermore, based on the idea of ​​comparing neighboring images, this method automatically selects a set of embryo images to determine the optimal-focused embryo image. Specifically, firstly, for the creation of the training dataset, all embryo images are converted to grayscale, and then... Figure 1 As shown, this invention first manually determines the best-focused embryo image from a set of embryo images in the training set, and combines it with other images in the same set. If the combined image is positioned before the best-focused image, it is designated as the first channel category, i.e., category 1; otherwise, it is designated as the second channel category, i.e., category 2. Secondly, as... Figure 2 As shown, to enhance the model's ability to fully extract data features, this invention incorporates the prior positional information of the two combined embryo images into the model during the training phase. Finally, as... Figure 3 As shown, during the testing phase, this invention combines adjacent images from a set of embryo images, processes them using a trained model, and automatically filters them round by round, discarding embryo images that do not meet the model's expectations, until the last embryo image is output, which is the optimal focused embryo image, thus achieving the selection of the optimal focused embryo image. Specifically, the method is carried out according to the following steps:

[0042] Step 1: Obtain an RGB three-channel embryo image dataset ,make The Middle Embryo image data are denoted as ,and ,in, express The first in Zhang embryo image, express The number of embryo images in the middle It is the height of an embryo image. It is the width of an embryo image; (Note: The original text appears to be incomplete and contains several grammatical errors. for The first in The image shows an embryo, and it is the best-focused image. .

[0043] Step 2: After grayscale conversion, the first Group of grayscale embryo image data ,in, express Grayscale embryo images, express Grayscale embryo images; converting images to grayscale can simplify calculations, reduce computational load, thereby improving processing efficiency and accelerating algorithm execution.

[0044] Step 3: Put Separately with the previous A grayscale embryo image By combining them, we obtain the first channel category. Group embryo image dual-channel combined data ,in, For the first channel category The first in the group Combined data of two channels of embryo images, , and , Indicates the preceding The first grayscale embryo image Several grayscale embryo images, among which... The data for the first channel consists of grayscale images. constitute; The data for the second channel is derived from the grayscale image. constitute; The actual channel category label is denoted as .

[0045] Step 4: Separately and after A grayscale embryo image By combining them, we obtain the second channel category. Group embryo image dual-channel combined data ,in, For the second channel category The first in the group Two-channel combined data of embryo images, in which... , and , After indicating The first grayscale embryo image Several grayscale embryo images, among which... The data for the first channel consists of grayscale images. Composition; The data for the second channel consists of grayscale images. constitute; The actual channel category label is denoted as .

[0046] Step 5: Construct a ConvNeXt network that fuses location information, including: an image difference module. One ConvNeXt module and a fully connected layer;

[0047] Step 5.1: The image difference module will... After subtracting the data from the two channels, we obtain the first channel category. The first in the group Differential images Image differencing can suppress some spatially independent noise while enhancing the contrast between different regions in an image, making the target area of ​​interest more prominent. Furthermore, image differencing transforms dual-channel data into single-channel data, reducing the computational cost of the model.

[0048] Step 5.2: Differential Image via The ConvNeXt module processes the data, ultimately yielding the first channel category. The first in the group A process The first ConvNeXt module processed the [number]th Feature matrices Among them, the first The ConvNeXt module includes: the first The downsampling module, the first The location information fusion module, the first The nth depthwise convolutional module, the nth The first layer normalization module, the first The first layer scaling module, the first The path discard module, the first One residual connection module; ;

[0049] when At that time, the difference image Enter the first In the first ConvNeXt module, by the first Each downsampling module is scaled according to the scaling ratio. right Perform downsampling to obtain the first channel category's... The first in the group The first one after being processed by the downsampling module downsampled differential images ,in, Indicates the height of the downsampled difference image. , This represents the width of the downsampled differential image. ,and Downsampling can reduce the computational overhead of the network.

[0050] No. Each deep convolutional module Perform channel expansion processing to obtain the first channel category's... The first in the group The first one after being processed by the depthwise convolution module depthwise convolutional matrices ,in, The number of channels in the feature matrix. Deep convolution can expand the receptive field and extract multi-level features, thereby enhancing the model's expressive power.

[0051] No. The location information fusion module calculates Location information and to By expanding the dimensions, we obtain the first... Extended location information Then Extended to Having the same number of channels , obtained the The location information after further expansion Thus and After adding them, we get the first channel category. The first in the group The first fused location information Location information feature matrix The fusion of location information helps the model to better extract the correlation between two embryo images by utilizing the sequential correlation information of embryo images, thereby improving the model's feature extraction capability for embryo image data.

[0052] No. Each layer normalization module calculates The Middle The average of each channel and variance and obtained Specifically, the calculation of the first channel category... The first in the group After normalization of the first layer The first channel Feature matrices Thus, the first channel category can be obtained. The first in the group After normalization of the first layer Each layer of normalized feature matrix ,in, It is a constant to prevent the denominator from being zero. The layer normalization module normalizes the input data for each layer, which speeds up training, enhances model stability, and enables the network to learn feature representations more effectively, avoiding gradient vanishing or exploding problems.

[0053] No. Each layer scaling module applies the feature matrix Perform channel scaling, reducing its number of channels from Scaled down to 1, thus obtaining the first channel category's... The first in the group The first one after being processed by the layer scaling module Each layer scaled feature matrix Specifically, Through the The result is obtained by weighted summation along the channel dimension: .in, It is the first The learnable weights corresponding to each channel The layer scaling module introduces learnable weights to the residual connections, which can appropriately scale and adjust the gradients, making the gradients more stable during backpropagation and effectively alleviating the gradient vanishing problem, thus enabling the network to be trained more smoothly.

[0054] No. The path discarding module uses a Bernoulli distribution to generate a path that is similar to... Random tensors with the same dimensions and shape and will Each element in the array is divided by the set path drop probability. Then, the first channel category is obtained. The first in the group The first one after the above operations Discarding tensors along each path Thus, the first channel category is obtained. The first in the group The first one after being processed by the path discarding module Discard feature matrix of each path ,in, ,and ; This indicates that corresponding elements of two feature matrices with the same shape are multiplied together; the path discarding module can select to discard some paths according to the actual situation, which can reduce the model's overfitting to the training data.

[0055] No. Each residual connection module discards the path feature matrix. Compared with downsampled differential image Add the corresponding elements together to obtain the first channel category. The first in the group The first one after processing by the residual connection module Each residual feature matrix ,in, Residual connections can fuse features from different levels, enabling the network to utilize both shallow, low-level features and deep, high-level features simultaneously, which improves the model's feature extraction capabilities.

[0056] when At that time, the first channel category will be... The first in the group The first one after the The first ConvNeXt module processed the [number]th Feature matrices Enter the number The first ConvNeXt module is used for processing, thus ultimately obtaining the first channel category. The first in the group The first one after the The first ConvNeXt module processed the [number]th Feature matrices ,in, express height, express The width.

[0057] Step 5.3: For Perform a flattening operation to obtain a flattened one-dimensional vector. ,in, Represents a one-dimensional vector The length, and .Will The data is fed into a fully connected layer for processing to obtain dual-channel combined data. Binary classification probability ,in, and They are The probability of belonging to channel category 1 and channel category 2;

[0058] Step 6: Following the process in Step 5, The input location information is processed in a ConvNeXt network. Binary classification probability ,in and They are The probability of belonging to channel category 1 and channel category 2.

[0059] Step 7: Construct the total loss function of the ConvNeXt network using equation (1) :

[0060] (1)

[0061] In equation (1), Denotes the first loss function. This represents the second loss function.

[0062] Step 8: Use the backpropagation algorithm and optimizer to minimize the total loss function. The ConvNeXt network, which integrates location information, is used to train and update its parameters. This process continues until a preset maximum number of training epochs or a total loss function is reached. Upon convergence, training is stopped, and the best-focused image binary classification model is obtained after training, which is used to select the best-focused image.

[0063] Step 9, Model Testing:

[0064] For the Group of embryo images After being grayscaled, it becomes The process of selecting the best-focused embryo images is carried out according to the following steps.

[0065] Step 9.1: Set variables as well as Used to record the embryo images and corresponding indexes selected in each round of model screening in steps 5 to 6.

[0066] Step 9.2: Initialize variables as well as .

[0067] Step 9.3: Transfer variables By combining two adjacent embryo images in sequence, we obtain dual-channel data suitable for the model: The dual-channel data is processed using the model in steps 5 and 6 to obtain the corresponding category 1 or 2. If it is category 1, it is retained in the form of... In (Data from the second channel) to a variable At the same time, index Record to variable If it is category 2, then retain. In (Data from the first channel) to a variable At the same time Record to variable .

[0068] Step 9.4: Remove variables Redundant images are deleted when the same image is saved twice. The image index in the database was also reduced accordingly.

[0069] Step 9.5: Hypothesize variables Updated to ,variable Updated to .

[0070] Step 9.6: Update the variables By combining two adjacent embryo images in sequence, we can obtain dual-channel combined data suitable for the model again: The combined dual-channel data is processed sequentially using the model in steps 5 and 6 to obtain the corresponding category 1 or 2, and then the variables are updated again accordingly. and .

[0071] Step 9.7: Repeat steps 9.4 to 9.6.

[0072] Step 9.8: When the variable If there is only one embryo image, stop step 9.7. At this point, you can start from the variable. The corresponding image index is obtained from the [image index], and this index is the [image index]. Group of embryo images Index of best-focused embryo images.

[0073] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0074] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0075] The following is the implementation process of this embodiment:

[0076] 1. Data collection and processing stage:

[0077] (1) Preparation of embryo images:

[0078] The embryo images used in this embodiment were acquired by the Time Difference Incubator (TLS301) developed by Wuhan Huchuang United Technology Co., Ltd. The 2000 sets of embryo images used in the experiment came from this device, with each set containing 7 embryo images.

[0079] In research on automated embryo image focusing techniques, color information is not a decisive factor, and grayscale images are sufficient to provide the main visual cues required for automatic focusing. To reduce the complexity of data processing, this invention converts color embryo images into grayscale images.

[0080] (2) Preparation of training data:

[0081] In a set of embryo images, the optimally focused embryo image differs to varying degrees from other embryo images in the same group. This difference forms the basis for selecting the optimally focused embryo image. Figure 1 As shown, the optimally focused embryo image in this group of embryo images has a position index of 4. Combining it with embryo images with position indices 1, 2, and 3 yields data belonging to category 1; combining it with images with position indices 5, 6, and 7 yields data belonging to category 2.

[0082] 2. Model training phase:

[0083] (1) Selection of network model:

[0084] To select the best-focused embryo image from a set of embryo images, this invention designs a ConvNeXt network that incorporates embryo image location information. The original ConvNeXt modifies the convolutional network, referencing some key designs of the Transformer, such as: a deeper network hierarchy and the use of LayerNorm instead of the traditional BatchNorm. Furthermore, it introduces new designs, such as depthwise convolution, which reduces the number of parameters and computational complexity. It also employs the GELU activation function instead of the traditional ReLU activation function, providing smoother gradients. ConvNeXt also uses larger convolutional kernels (e.g., 7×7), allowing the network to capture a larger receptive field and enhancing its ability to model global information.

[0085] However, for a set of embryo images, since the feature relationships between the two channels of the dual-channel combined data are strongly correlated with the intra-group position index of the corresponding original embryo image, the lack of explicit position information in ConvNeXt, which does not incorporate position index information, may lead to misclassification of certain details by the model. Therefore, incorporating the position index information of embryo images into the ConvNeXt network has the potential to improve the accuracy of classification tasks. Fully utilizing this position information can help the model better understand the feature distribution and class discrimination of the data.

[0086] like Figure 2 As shown, two embryo images (including one optimally focused image) are first grayscaled and then combined. The combined image can be considered as a single data set with two channels. Then, the difference image between the two images is calculated, which involves subtracting the two channel data. This is done to highlight the morphological differences between the optimally focused image and other images, allowing the model to better extract features from the constructed data while reducing computational cost. The difference image is then fed into a ConvNeXt module that integrates positional information for training. In each ConvNeXt module, the present invention uses the positional index information of the two original embryo images (including one optimally focused image) corresponding to the difference image as prior knowledge and integrates it into the network, thereby enabling the model to better extract features between the two embryo images.

[0087] (3) Model training settings:

[0088] In the experiment, this invention used an NVIDIA GeForce RTX 4090 GPU with 24GB of video memory. During model training, the epoch was set to 100, the batch size to 1, and the learning rate to 0.001. Training was conducted on a Linux platform using a virtual Python environment set up with Anaconda 24.1.2 and PyCharm 2024.1, with Python version 3.8.

[0089] 3. Evaluation indicators:

[0090] This invention focuses on "focus accuracy ( ")" is used as an evaluation indicator, and its calculation is based on the following formula:

[0091] (2)

[0092] (3)

[0093] in, This refers to an optimal focused embryo image selection method based on the ConvNeXt network that integrates location information and incorporates the idea of ​​combining adjacent embryo images. It is the first The first group of embryo images Zhang Image It refers to the The inference and prediction results obtained after processing this group of embryo images. It is the location index of the best-focused embryo image in this group of embryo images.

[0094] when When this occurs, the present invention refers to it as a correct inference prediction. Based on the above ideas, "Focusing on Accuracy ( The calculation method for “)” is as follows: In the test samples, the number of correct inference predictions made by the method proposed in this invention was: ,but ;

[0095] 4. Model testing phase:

[0096] During the testing phase, this invention designed a method based on combining adjacent embryo images, followed by model classification and comparison to obtain the optimal focused embryo image. This method uses a trained ConvNeXt binary classification model that fuses location information as the selector, sequentially selecting images from adjacent embryo images that meet the model's expectations, thereby achieving the final selection of the optimal focused embryo image. Figure 3As shown, a set of embryo images is processed by combining adjacent images after grayscale conversion to form a data format that can be processed by a binary classification model. After processing by the model, images that meet the model's expectations are selected. After deduplication, these images are combined again with adjacent images in a new round, and then fed into the model for processing. Through multiple rounds of comparison, the best-focused embryo image that meets the model's expectations can be obtained.

[0097] Furthermore, this invention employs a 5-fold cross-validation method. The embryo image dataset is divided into five non-overlapping subsets. Each subset is used alternately as the test set, while the remaining four subsets are used for training. This method allows the invention to undergo five independent experiments, thus comprehensively evaluating its performance. Table 2 shows the accuracy obtained using 5-fold cross-validation. The proposed method achieves a maximum accuracy of 97.25% in the second fold, with an overall average accuracy of 96.25%.

[0098] 5. Experimental Results:

[0099] The optimal focusing embryo image screening method based on the ConvNeXt network with fused location information uses ConvNeXt as the base model. The main improvement is that the location information of the embryo image is incorporated into the feature extraction of the ConvNeXt block. To this end, this invention sets up two sets of comparative experiments. One set is trained and tested using the original ConvNeXt network, and the other set is trained and tested using the ConvNeXt network with fused location information. This demonstrates the superior performance of the improved network. This invention uses 2000 sets of samples for comparative experiments and employs a 5-fold cross-validation method for verification. The results of the comparative experiments are shown in Tables 1 and 2:

[0100] Table 1 shows the experimental results of the original ConvNeXt;

[0101]

[0102] Table 2 shows the experimental results of the ConvNeXt network that integrates location information;

[0103]

[0104] Overall, the method proposed in this invention shows a certain degree of improvement over the original method after verification through a 5-fold cross-validation experiment.

Claims

1. A method for optimal focusing embryo image screening based on a ConvNeXt network with fused location information, characterized in that, The procedure is as follows: Step 1: Obtain an RGB three-channel embryo image dataset ,make The Middle Embryo image data are denoted as ,and ,in, express The first in Zhang embryo image, express The number of embryo images in the middle It is the height of an embryo image. It is the width of an embryo image; (Note: The original text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.) for The first in The image shows an embryo, and it is the best-focused image. ; Step 2: After grayscale conversion, the first Group of grayscale embryo image data ,in, express Grayscale embryo images, express Grayscale embryo images; Step 3: Put Separately with the previous A grayscale embryo image By combining them, we obtain the first channel category. Group embryo image dual-channel combined data ,in, For the first channel category The first in the group Combined data of two channels of embryo images, , and , Indicates the preceding The first grayscale embryo image Several grayscale embryo images, among which... The data for the first channel consists of grayscale images. constitute; The data for the second channel is derived from the grayscale image. constitute; The actual channel category label is denoted as ; Step 4: Separately and after A grayscale embryo image By combining them, we obtain the second channel category. Group embryo image dual-channel combined data ,in, For the second channel category The first in the group Two-channel combined data of embryo images, in which... , and , After indicating The first grayscale embryo image Several grayscale embryo images, among which... The data for the first channel consists of grayscale images. constitute; The data for the second channel is derived from the grayscale image. constitute; The actual channel category label is denoted as ; Step 5: Construct a ConvNeXt network that integrates location information, and then... and Process them separately to obtain the corresponding results. Binary classification probability and Binary classification probability ,in, and They are The probability of belonging to channel category 1 and channel category 2, and and They are The probability of belonging to the first channel category and the second channel category; where the location information is the location information fusion module in the a-th ConvNeXt module of the ConvNeXt network for the first channel category. The first in the group The first one after being processed by the depthwise convolution module A depthwise convolutional feature matrix After calculation, the first channel category is obtained as follows: The first in the group The first grayscale embryo image Location information Where k represents the image number corresponding to the best-focused image, and x is the number of the best-focused image. The sequence number of any grayscale embryo image in the group; 'a' represents the sequence number of the ConvNeXt module in the ConvNeXt network; Step 6: Construct the total loss function of the ConvNeXt network using equation (1) : (1) In equation (1), Denotes the first loss function. This represents the second loss function; Step 7: Minimize the total loss function using the backpropagation algorithm and optimizer. The ConvNeXt network, which integrates location information, is used to train and update its parameters. This process continues until a preset maximum number of training epochs or a total loss function is reached. Upon convergence, training is stopped, and the best-focused image binary classification model is obtained after training, which is used to select the best-focused image.

2. The optimal focusing embryo image screening method based on a ConvNeXt network with fused location information as described in claim 1, characterized in that, The ConvNeXt network that fuses location information in step 5 includes: an image difference module, A ConvNeXt module and a fully connected layer; the value of A is an integer greater than 1; Step 5.1: The image difference module will... After subtracting the data from the two channels, we obtain the first channel category. The first in the group Differential images ; Step 5.2: Differential Image through The ConvNeXt module processes the data to obtain the first channel category. The first in the group The processed first Feature matrices ; Step 5.3: For Perform a flattening operation to obtain a flattened one-dimensional vector. ,in, Represents a one-dimensional vector The length, and ;Will The data is fed into a fully connected layer for processing to obtain dual-channel combined data. Binary classification probability ,in, and They are The probability of belonging to channel category 1 and channel category 2; Step 5.4: Following the process in steps 5.1-5.3, combine the dual-channel data. The input location information is processed in a ConvNeXt network to obtain... Binary classification probability ,in and They are The probability of belonging to channel category 1 and channel category 2.

3. The optimal focusing embryo image screening method based on a ConvNeXt network with fused location information according to claim 2, characterized in that, In step 5.2, the first The ConvNeXt module includes: the first The downsampling module, the first The nth depthwise convolutional module, the nth The location information fusion module, the first The first layer normalization module, the first The first layer scaling module, the first The path discard module, the first One residual connection module; ; when At that time, the difference image Enter the first In the first ConvNeXt module, by the first Each downsampling module is scaled according to the scaling ratio. right Perform downsampling to obtain the first channel category's... The first in the group The first one after being processed by the downsampling module downsampled differential images ,in, Indicates the height of the downsampled difference image. , This represents the width of the downsampled differential image. , ; No. Each deep convolutional module Perform channel expansion processing to obtain the first channel category's... The first in the group The first one after being processed by the depthwise convolution module A depthwise convolutional feature matrix ,in, The number of channels in the feature matrix. ; No. The location information fusion module calculates Location information and to Expanding the dimensions, we get the first... Extended location information Then Extended to Having the same number of channels , obtained the The location information after further expansion Thus and After adding them, we get the first channel category. The first in the group The first fused location information Location information feature matrix ; No. Each layer normalization module calculates The Middle The average of each channel and variance And calculate the first channel category's... The first in the group After normalization of the first layer The first channel Feature matrices Thus, the first channel category is obtained. The first in the group After normalization of the first layer Each layer of normalized feature matrix ,in, It is a constant to prevent the denominator from being zero. ; No. Each layer scaling module applies the feature matrix Perform channel scaling, reducing its number of channels from Scaled down to 1, thus obtaining the first channel category's... The first in the group The first one after being processed by the layer scaling module Each layer scaled feature matrix ,in, It is the first The weights to be learned for each channel. ;and ; No. The path discarding module uses a Bernoulli distribution to generate a path that is similar to... Random tensors with the same dimensions and shape and will Each element in the array is divided by the set path drop probability. Then, the first channel category is obtained. The first in the group The first one after the above operations Discarding tensors along each path Thus, the first channel category is obtained. The first in the group The first one after being processed by the path discarding module Discard feature matrix of each path ,in, ,and ; No. Each residual connection module discards the path feature matrix. Compared with downsampled differential image Add the corresponding elements together to obtain the first channel category. The first in the group The first one after processing by the residual connection module Each residual feature matrix ,in, ; when At that time, the first channel category will be... The first in the group The first one after the The first ConvNeXt module after processing Feature matrices Enter the number The first ConvNeXt module is used for processing, thus ultimately obtaining the first channel category. The first in the group The first one after the The first ConvNeXt module after processing Feature matrices ,in, express height, express The width.

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the optimal focused embryo image screening method according to any one of claims 1-3, the processor being configured to execute the program stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the optimal focused embryo image screening method according to any one of claims 1-3.

Images