Pronucleus and cleavage sphere target detection method based on multistage local attention mechanism

Abstract

The invention discloses a pronucleus and cleavage sphere target detection method based on a multistage local attention mechanism, and the method comprises the steps: 1, obtaining embryo images in 1-8 cell periods to construct a data set, and carrying out the preprocessing to construct a sample set; 2, constructing a network model for pronucleus and blastosphere target detection based on a multistage local attention mechanism, wherein the network model comprises a multistage local attention feature extraction network, a multi-scale feature fusion module, a feature pyramid aggregation module and a pronucleus and blastosphere detection head module; and 3, carrying out pronucleus and cleavage sphere detection by utilizing the trained model. According to the method, features are extracted through multistage local receptive fields, pixel-level details are reserved for pronucleus, local region semantic features are enhanced for overlapped blastocytes, and efficient technical support is provided for embryo quality evaluation.

Description

Technical Field

[0001] This invention relates to the field of pronuclear and blastomere target detection and deep learning technology, specifically to an optimized method for pronuclear and blastomere target detection based on a multi-level local attention mechanism, which is suitable for the accurate identification and localization of pronuclear and blastomere during embryonic development. Background Technology

[0002] Detection of pronuclei and blastomeres during early human embryonic development is a core technical means for assessing embryo quality and predicting developmental potential. The size, symmetry, and spatial location of the pronucleus, as well as the number, morphological regularity, and cleavage synchronicity of blastomeres, are directly related to embryo implantation rate and pregnancy success rate, and are key assessment parameters in clinical assisted reproduction and embryological research. Accurate detection of pronuclei and blastomeres can effectively identify pathological phenomena such as abnormal embryonic division and developmental arrest, providing objective evidence for embryo selection.

[0003] Currently, the interpretation of pronuclei and blastomeres in embryo images largely relies on manual work, which is not only time-consuming, labor-intensive, and inefficient, but also easily affected by the professional experience and subjective factors of the operators, resulting in poor consistency of interpretation results and making it difficult to meet the needs of large-scale clinical and scientific research applications.

[0004] Existing detection technologies are mainly divided into two categories: traditional machine vision algorithms and deep learning algorithms. Traditional algorithms, such as edge detection and ellipse fitting, have low computational costs, but they suffer from insufficient feature extraction capabilities and high false negative and false positive rates in complex scenarios where pronuclear targets overlap with blastomeres in embryonic images, resulting in detection accuracy that cannot meet practical needs. Deep learning algorithms have shown advantages in the field of biomedical image detection due to their autonomous feature learning capabilities, but existing general models have not been specifically optimized for the morphological features and detection pain points of pronuclei and blastomeres. The models have limited feature representation capabilities for embryonic images, making it difficult to balance detection accuracy and efficiency.

[0005] In summary, existing technologies cannot efficiently and accurately achieve simultaneous detection of embryonic pronuclei and blastomeres. There is an urgent need to design a deep learning detection architecture adapted to the characteristics of embryonic images to solve the technical problems of missed detection of small pronuclei and false detection of overlapping blastomeres, thereby improving the accuracy and stability of embryo detection. Summary of the Invention

[0006] This invention addresses the technical bottlenecks in existing prokaryotic and blastomere detection technologies, such as low detection accuracy and difficulty in simultaneously extracting prokaryotic features from small targets and identifying overlapping blastomere boundaries. It provides an optimized method for prokaryotic and blastomere target detection based on a multi-level local attention mechanism. By constructing multi-level local receptive fields for multi-scale feature mining, this method effectively enhances the model's ability to capture detailed features of prokaryotes and distinguish blastomere boundaries, thereby improving the overall detection accuracy of the model.

[0007] The present invention provides a method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism, characterized by the following steps: Step 1: Obtain the preprocessed embryo image dataset and annotate the bounding boxes of the pronucleus and blastomere regions in each embryo image to construct the embryo image sample set Pic. Let pic be any preprocessed embryo image sample, and let the bounding box containing the j-th target in pic be denoted as . ,make Let be the true category of the j-th embryo target in pic, including: pronucleus, blastomeres; Step 2: Construct an embryo target detection network, including a backbone network, a neck network, and a head network, and process the pic to obtain the desired result. Step 2.1: The backbone network processes the pic to obtain the embryo deep fusion convolutional feature map corresponding to the pic. ; Step 2.2, the neck network pair The process was performed to obtain a global feature map of the embryo. ; Step 2.3, the head network pair The process is performed to obtain the bounding box sequence of the target embryo. and the predicted category probability sequence of embryo targets ; Step 3: Construct the label boxes and their true classes, and the bounding boxes and their predicted class probabilities to build the total loss of the embryo target detection network. ; Step 4: Based on the training set, train the detection network using gradient descent and calculate the total loss. To update network parameters until After convergence or reaching the maximum number of iterations, training stops, and a detection model with optimal parameters is generated for detecting pronuclei and overlapping blastomeres in embryo images.

[0008] The method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism described in this invention is characterized in that the backbone network in step 2.1 includes, in sequence: a first convolutional module, S C2f feature fusion modules, P SCDown downsampling modules, a multi-level local attention fusion module, an SPPF module, and a PSA attention module. Step 2.1.1: In the first convolutional module of the pic input backbone, feature extraction is performed to generate the initial feature map of the embryo. ; Step 2.1.2 The feature maps are processed layer by layer through S C2f feature fusion modules, resulting in the output of feature maps at S scales. ;in, This represents the feature map at the k-th scale output by the k-th C2f feature fusion module; Step 2.1.3, Feature map at the S-th scale After processing by P SCDown downsampling modules, the Pth embryo downsampled feature map is output by the Pth SCDown downsampling module. ; Step 2.1.4 After being processed in the multi-level local attention fusion module, the embryonic attention fusion feature map is obtained. ; Step 2.1.5 After being processed sequentially by the SPPF module and the PSA module, the deep fused convolutional feature map of the embryo is output. .

[0009] Furthermore, the multi-level local attention fusion module in step 2.1.4 consists of a convolutional adaptation layer, a multi-level local feature extraction module, an attention-weighted fusion layer, and a feature restoration layer, in sequence. Step 2.1.4.1 Channel dimension alignment and feature dimension compression are performed in the input convolutional adaptation layer to obtain a high-level convolutional feature map of the embryo with uniform dimensions. ; Step 2.1.4.2 The input is processed by a multi-level local feature extraction module, and then through N parallel branches for feature extraction, resulting in embryonic local feature maps at N scales. ,in, This represents the local feature map of the embryo at the nth scale, where each branch includes: a branch convolutional layer with receptive field kernels of different sizes, a batch normalization layer, and a SiLU activation function; Step 2.1.4.3 Global average pooling is performed on the input attention weights in the learning layer, resulting in N global feature vectors. , … ,… },in, This represents the nth global eigenvector; Will{ , … ,… The input is processed in a single-hidden-layer neural network, and the output is N normalized attention weight tensors. , … … },in, Let n represent the nth attention weight tensor, and ; Step 2.1.4.4: Attention-weighted fusion layer uses equation (1) to obtain the local fusion feature map of the embryo. ; (1) In equation (1), The SiLU activation function is used. For bias terms; Steps 2.1.4.5 After dimensionality reduction by the feature reduction layer, the embryo attention fusion feature map is obtained. .

[0010] Furthermore, the neck network in step 2.2 sequentially includes: L upsampling modules, a Concat module, a fourth convolution module, a C2f feature fusion module, and a C2fCIB module; Step 2.2.1, when When =1, The first input to the neck network The process is performed in the upsampling module to obtain the first... Upsampled convolutional feature map of each embryo ,Will With the Feature maps at various scales The input is processed in the Concat module to obtain the first... Embryo splicing feature map ; when When =2,3,…,L, the first Embryo splicing feature map Enter the first The process is performed in the upsampling module to obtain the first... Feature map of embryos , and then with The input is concatenated using the Concat module, and the output is the final concatenated feature map of the Lth embryo. ; Step 2.2.2 The feature map of the embryo is obtained by extracting features from the fourth convolutional module. ; Step 2.2.3 The input is processed in the C2f feature fusion module to obtain the embryo fusion feature map. ; Step 2.2.4 After being processed in the C2fCIB module, the global feature map is obtained. .

[0011] Furthermore, the head network includes: a multi-scale feature fusion layer, M convolutional layers, N upsampling layers, N-1 downsampling layers, and a v10Detect prediction head; Step 2.3.1 The data is input into the head network and processed through M convolutional layers for feature extraction to obtain the embryonic convolutional feature map. ; Step 2.3.2 After performing scale adaptation processing on N upsampling and N-1 downsampling layers sequentially, the embryo multi-scale feature map is obtained. ; Step 2.3.3, v10Detect predicts head pairs After processing, the bounding box sequence of the target embryo is obtained. and the predicted category probability sequence of embryo targets .

[0012] Furthermore, step 3 includes: Step 3.1: Construct the bounding box CIoU loss using equation (2). : (2) In equation (2), J represents the predicted total number of embryos. The function for calculating the intersection-union ratio is... express The bounding box containing the target of the j-th embryo in the model. These are the weighting coefficients. Aspect ratio consistency parameter; Step 3.2: Construct the cross-entropy classification loss using equation (3). : (3) In equation (3), This represents the predicted category probability of the j-th embryo target; Step 3.3: Construct the total loss using equation (4). : (4) In equation (4), This is the loss weighting coefficient.

[0013] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing the prokaryotic and blastomere target detection method, and the processor is configured to execute the program stored in the memory.

[0014] The present invention discloses a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the prokaryotic and blastomere target detection method.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention employs a multi-level local attention method, designing multi-level local receptive fields for small target pronuclei and overlapping blastomeres. Attention weights are calculated only for local regions of the target. Without increasing the overall complexity of the network, this avoids the pronucleus features being submerged by background noise during global feature fusion, while also reducing the amount of attention computation, significantly improving the pronucleus detection accuracy, model training efficiency, and accuracy. 2. This invention designs a weighted loss strategy for biological target detection. By setting the weight coefficient of the bounding box CIoU loss in the total loss, the positioning accuracy of pronuclei and blastomeres is prioritized. At the same time, the aspect ratio of blastomeres is introduced as a prior in the CIoU loss to optimize the bounding box regression effect of overlapping blastomeres. This effectively solves the core detection pain points of pronucleus positioning offset and low fitting accuracy of overlapping blastomeres. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the training method for a prokaryotic and blastomere detection model according to an embodiment of the present invention. Figure 2 This is a structural diagram of an embryo detection model according to an embodiment of the present invention; Figure 3 This is a structural diagram of the MLABlock module according to an embodiment of the present invention; Figure 4 This is a diagram showing the detection effect of pronuclei and blastomeres according to an embodiment of the present invention. Detailed Implementation

[0017] In this embodiment, an optimization method for prokaryotic and blastomere target detection based on a multi-level local attention mechanism is described, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the preprocessed dataset of embryo images with 1-4 cells and more, and annotate the bounding boxes of the pronucleus and blastomere regions in each embryo image to construct the embryo image sample set Pic. Let pic be any preprocessed embryo image sample, and let the bounding box containing the j-th target in pic be denoted as . ,and ,in, , These are the top-left and top-right coordinates of the label box containing the actual location of the j-th embryo in the pic. , These are the coordinates of the bottom left and bottom right corners of the label box containing the actual location of the j-th embryo target in the pic. The true category of the j-th embryo target in pic includes: pronucleus, blastomere; in this embodiment, samples collected in the time difference incubator of Chongqing Maternal and Child Health Hospital are used. The dataset contains images of the 1-8 cell stage, and the image size is 800×800.

[0018] Step 2: In this embodiment, the embryo target detection network is constructed as follows: Figure 2 As shown, it includes: backbone network, neck network and head network; Step 2.1: The backbone network processes the pic to obtain the embryo deep fusion convolutional feature map corresponding to the pic. (3×3 kernel, 2 stride).

[0019] Step 2.1.1: In the first convolutional module of the pic input backbone, feature extraction is performed to generate the initial feature map of the embryo. ; Step 2.1.2 The feature maps are processed layer by layer through S C2f feature fusion modules, resulting in the output of feature maps at S scales. ;in, This represents the feature map at the k-th scale output by the k-th C2f feature fusion module.

[0020] Step 2.1.3, Feature map at the S-th scale After processing by P SCDown downsampling modules, the Pth embryo downsampled feature map is output by the Pth SCDown downsampling module. ; Step 2.1.4 After being processed in the multi-level local attention fusion module, the embryonic attention fusion feature map is obtained. A schematic diagram of the multi-level local attention fusion module is shown below. Figure 3 As shown.

[0021] Step 2.1.4.1 Channel dimension alignment and feature dimension compression are performed in the input convolutional adaptation layer (1×1 convolution, stride 1) to obtain a high-level convolutional feature map of the embryo with uniform dimensions. ; Step 2.1.4.2 The input is processed by a multi-level local feature extraction module, and then through N parallel branches for feature extraction, resulting in embryonic local feature maps at N scales. ,in, This represents the local feature map of the embryo at the nth scale. Each branch includes: a branch convolutional layer with receptive field convolutional kernels of different sizes, a batch normalization layer, and a SiLU activation function. In this embodiment, features are extracted using 1×1, 3×3, and 5×5 receptive field convolutional kernels to obtain the local feature maps at the 1st, 2nd, and 3rd scales, respectively.

[0022] Step 2.1.4.3 Global average pooling is performed on the input attention weights in the learning layer, resulting in N global feature vectors. , … ,… },in, This represents the nth global eigenvector; Will{ , … ,… The input is processed in a single-hidden-layer neural network, and the output is N normalized attention weight tensors. , … … },in, Let n represent the nth attention weight tensor, and .

[0023] Step 2.1.4.4: Attention-weighted fusion layer uses equation (1) to obtain the local fusion feature map of the embryo. ; (1) In equation (1), The SiLU activation function is used. This is a bias term.

[0024] Steps 2.1.4.5 After dimensionality reduction by the feature reduction layer, the embryo attention fusion feature map is obtained. .

[0025] Step 2.1.5 After being processed sequentially by the SPPF module and the PSA module, the deep fused convolutional feature map of the embryo is output. .

[0026] Step 2.2, Neck network Processing was performed to obtain a global feature map of the embryo. ; Step 2.2.1, when When =1, The first input to the neck network The process is performed in the upsampling module to obtain the first... Upsampled convolutional feature map of each embryo ,Will With the Feature maps at various scales The input is processed in the Concat module to obtain the first... Embryo splicing feature map ; when When =2,3,…,L, the first Embryo splicing feature map Enter the first The process is performed in the upsampling module to obtain the first... Feature map of embryos , and then with The input is concatenated using the Concat module, and the output is the final concatenated feature map of the Lth embryo. .

[0027] Step 2.2.2 The feature map of the embryo is obtained by extracting features from the fourth convolutional module. ; Step 2.2.3 The input is processed in the C2f feature fusion module to obtain the embryo fusion feature map. ; Step 2.2.4 After being processed in the C2fCIB module, the global feature map is obtained. In this embodiment, the C2fCIB module outputs 3 channels, and all Boolean parameters are set to True.

[0028] Step 2.3, Head Network Pair Processing is performed to obtain the bounding box containing the embryo target. ; Step 2.3.1 The data is input into the head network and processed through M convolutional layers for feature extraction to obtain the embryonic convolutional feature map. .

[0029] Step 2.3.2 After performing scale adaptation processing on N upsampling and N-1 downsampling layers sequentially, the embryo multi-scale feature map is obtained. In this embodiment, there is one upsampling layer and two downsampling layers. Step 2.3.3, v10Detect predicts head pairs After processing, the predicted coordinates of the bounding box containing the embryo target are obtained. and the predicted category probability of embryo targets .

[0030] Step 3: Based on the training set, train the detection network using gradient descent and calculate the total loss. To update network parameters until After convergence or reaching the maximum number of iterations, training is stopped, and the detection model corresponding to the optimal parameters is generated, which is used to detect pronuclei and overlapping blastomeres in embryo images of the test set; Step 3.1: Construct the bounding box CIoU loss using equation (2). : (2) In equation (2), J represents the predicted total number of embryos. The function for calculating the intersection-union ratio is... express The bounding box containing the target of the j-th embryo in the dataset, and , , These are the top-left and top-right corner coordinates of the bounding box containing the j-th embryo target in the pic, respectively. , These are the coordinates of the bottom left and bottom right corners of the bounding box containing the j-th embryo target in the pic. These are the weighting coefficients. This is a parameter for aspect ratio consistency.

[0031] Step 3.2: Construct the cross-entropy classification loss using equation (3). : (3) In equation (3), This represents the predicted category probability of the j-th embryo target; Step 3.3: Construct the total loss using equation (4). : (4) In equation (4), This is the loss weighting coefficient.

[0032] The experiment was conducted on a 64-bit Ubuntu 18.04 operating system, using PyTorch v1.8.0 as the deep learning framework. The CPU was an AMD(R) Epyc 7r32 48-Core, and the GPU was an NVIDIA GeForce RTX4090. The Adam optimizer was used during training, with an initial learning rate of 0.01 and weight decay of 0.0005. The batch size was set to 16.

[0033] 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.

[0034] 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.

[0035] The experiment used embryo images collected at Chongqing Maternal and Child Health Hospital as the dataset, and the number of cells of each type in the dataset is shown in Table 1. Prokaryotic mAP50, blastomeric mAP50, and total mAP50 were used as evaluation indicators. Prokaryotic mAP50 represents the accuracy rate of pronuclear detection in all data, blastomeric mAP50 represents the accuracy rate of blastomeres detection in all data, and total mAP50 represents the accuracy rate of both pronuclear and blastomeres detection in all data.

[0036] Table 1. Number of cell images of each type in the embryo dataset (per image) Table 2 Comparison of experimental results on cells 1-4 and above. Figure 4 Table 2 shows the pronucleus and blastomere detection performance of various models in embodiments of the present invention. The results indicate that the present invention significantly improves the detection performance of pronucleus and blastomere. Ours represents the network traffic prediction model proposed in this invention. SSD, YOLOv5, and YOLOv10 represent different network models. Comparative experiments show that the pronucleus and blastomere detection model proposed in this invention achieves the best results in terms of evaluation metrics. This invention has good application potential and provides methods and ideas for future practical applications in embryo assessment.

Claims

1. A method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism, characterized in that, The procedure is as follows: Step 1: Obtain the preprocessed embryo image dataset and annotate the bounding boxes of the pronucleus and blastomere regions in each embryo image to construct the embryo image sample set Pic. Let pic be any preprocessed embryo image sample, and let the bounding box containing the j-th target in pic be denoted as . ,make Let be the true category of the j-th embryo target in pic, including: pronucleus, blastomeres; Step 2: Construct an embryo target detection network, including a backbone network, a neck network, and a head network, and process the pic to obtain the desired result. Step 2.1: The backbone network processes the pic to obtain the embryo deep fusion convolutional feature map corresponding to the pic. ; Step 2.2, the neck network pair The process was performed to obtain a global feature map of the embryo. ; Step 2.3, the head network pair The process is performed to obtain the bounding box sequence of the target embryo. and the predicted category probability sequence of embryo targets ; Step 3: Construct the label boxes and their true classes, and the bounding boxes and their predicted class probabilities to build the total loss of the embryo target detection network. ; Step 4: Based on the training set, train the detection network using gradient descent and calculate the total loss. To update network parameters until After convergence or reaching the maximum number of iterations, training stops, and a detection model with optimal parameters is generated for detecting pronuclei and overlapping blastomeres in embryo images.

2. The method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism according to claim 1, characterized in that, The backbone network in step 2.1 includes, in sequence: a first convolutional module, S C2f feature fusion modules, P SCDown downsampling modules, a multi-level local attention fusion module, an SPPF module, and a PSA attention module; Step 2.1.1: In the first convolutional module of the pic input backbone, feature extraction is performed to generate the initial feature map of the embryo. ; Step 2.1.2 The feature maps are processed layer by layer through S C2f feature fusion modules, resulting in the output of feature maps at S scales. ;in, This represents the feature map at the k-th scale output by the k-th C2f feature fusion module; Step 2.1.3, Feature map at the S-th scale After processing by P SCDown downsampling modules, the Pth embryo downsampled feature map is output by the Pth SCDown downsampling module. ; Step 2.1.4 After being processed in the multi-level local attention fusion module, the embryonic attention fusion feature map is obtained. ; Step 2.1.5 After being processed sequentially by the SPPF module and the PSA module, the deep fused convolutional feature map of the embryo is output. .

3. The method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism according to claim 2, characterized in that, The multi-level local attention fusion module in step 2.1.4 consists of a convolutional adaptation layer, a multi-level local feature extraction module, an attention-weighted fusion layer, and a feature restoration layer. Step 2.1.4.1 Channel dimension alignment and feature dimension compression are performed in the input convolutional adaptation layer to obtain a high-level convolutional feature map of the embryo with uniform dimensions. ; Step 2.1.4.2 The input is processed by a multi-level local feature extraction module, and then through N parallel branches for feature extraction, resulting in embryonic local feature maps at N scales. ,in, This represents the local feature map of the embryo at the nth scale, where each branch includes: branch convolutional layers with receptive field kernels of different sizes, batch normalization layers, and SiLU activation functions; Step 2.1.4.3 Global average pooling is performed on the input attention weights in the learning layer, resulting in N global feature vectors. , … ,… },in, This represents the nth global eigenvector; Will{ , … ,… The input is processed in a single-hidden-layer neural network, and the output is N normalized attention weight tensors. , … … },in, Let n represent the nth attention weight tensor, and ; Step 2.1.4.4: Attention-weighted fusion layer uses equation (1) to obtain the local fusion feature map of the embryo. ; (1) In equation (1), The SiLU activation function is used. For bias terms; Steps 2.1.4.5 After dimensionality reduction by the feature reduction layer, the embryo attention fusion feature map is obtained. .

4. The method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism according to claim 3, characterized in that, The neck network in step 2.2 includes, in sequence: L upsampling modules, a Concat module, a fourth convolutional module, a C2f feature fusion module, and a C2fCIB module; Step 2.2.1, when When =1, The first input to the neck network The process is performed in the upsampling module to obtain the first... Upsampled convolutional feature map of each embryo ,Will With the Feature maps at various scales The input is processed in the Concat module to obtain the first... Embryo splicing feature map ; when When =2,3,…,L, the first Embryo splicing feature map Enter the first The process is performed in the upsampling module to obtain the first... Feature map of embryos , and then with The input is concatenated using the Concat module, and the output is the final concatenated feature map of the Lth embryo. ; Step 2.2.2 The feature map of the embryo is obtained by extracting features from the fourth convolutional module. ; Step 2.2.3 The input is processed in the C2f feature fusion module to obtain the embryo fusion feature map. ; Step 2.2.4 After being processed in the C2fCIB module, the global feature map is obtained. .

5. The method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism according to claim 4, characterized in that, The head network includes: a multi-scale feature fusion layer, M convolutional layers, N upsampling layers, N-1 downsampling layers, and a v10Detect prediction head; Step 2.3.1 The data is input into the head network and processed through M convolutional layers for feature extraction to obtain the embryonic convolutional feature map. ; Step 2.3.2 After performing scale adaptation processing on N upsampling and N-1 downsampling layers sequentially, the embryo multi-scale feature map is obtained. ; Step 2.3.3, v10Detect predicts head pairs After processing, the bounding box sequence of the target embryo is obtained. and the predicted category probability sequence of embryo targets .

6. The method for detecting prokaryotic and blastomere targets based on a multi-level local attention mechanism according to claim 1, characterized in that, Step 3 includes: Step 3.1: Construct the bounding box CIoU loss using equation (2). : (2) In equation (2), J represents the predicted total number of embryos. The function for calculating the intersection-union ratio is... express The bounding box containing the target of the j-th embryo in the model. These are the weighting coefficients. Aspect ratio consistency parameter; Step 3.2: Construct the cross-entropy classification loss using equation (3). : (3) In equation (3), This represents the predicted category probability of the j-th embryo target; Step 3.3: Construct the total loss using equation (4). : (4) In equation (4), This is the loss weighting coefficient.

7. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the prokaryotic and blastomere target detection methods of claims 1-5, and the processor is configured to execute the programs stored in the memory.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the prokaryotic and blastomere target detection method according to any one of claims 1-5.

Images