A method for detecting overlapping blastomeres based on multi-dimensional feature fusion

Abstract

This invention discloses a method for detecting overlapping blastomeres based on multi-dimensional feature fusion. The steps include: 1. Obtaining embryo images from the 1-8 cell stage to construct a dataset and preprocessing them to build a sample set; 2. Separating the phase and amplitude components from 1-2 cell images randomly selected from the training set and performing phase fusion on the images; 3. Performing inverse Fourier transform on the fused phase and amplitude to obtain a new training sample set; 4. Constructing a network model for blastomere target detection; 5. Detecting blastomeres using the trained model. This invention enhances data by introducing interference with the phase information of the images. Simultaneously, it improves the ability to separate overlapping regions by performing fine-grained feature extraction and multi-scale global feature aggregation through multi-dimensional feature fusion (weighted summation of height, width, and channels), thereby significantly improving the accuracy of overlapping blastomere detection.

Description

Technical Field

[0001] This invention relates to the field of blastomere target detection, specifically to an image preprocessing method based on phase fusion and an improved method for an overlapping blastomere detection model based on multi-dimensional feature fusion. Background Technology

[0002] blastomere detection technology plays a crucial role in early human life development. This technology not only identifies abnormal or non-dividing embryos during development by analyzing the number of blastomeres, but also allows for pathological analysis, such as tracing the origins of newly formed blastomeres during subsequent divisions, using information about blastomeres. Therefore, the accurate and effective detection and identification of blastomeres has become one of the key technologies in this field. Furthermore, it is an indispensable part of embryo assessment, significantly helping embryologists evaluate embryo quality and developmental potential. According to the D3 embryo assessment system, from day 1 to day 3 after embryo formation (D1-D3), doctors can obtain key morphological information about the embryo, including important parameters such as the number and morphology of blastomeres. This data is one of the important bases for judging embryo quality.

[0003] In the medical field, particularly in embryological research, manually interpreting large numbers of images is a time-consuming and specialized task. Therefore, utilizing computer technology to build visual models to assist doctors in quickly and accurately detecting blastomeres has a profound impact on this research field. This technology can significantly improve efficiency, reduce doctors' workload, and enhance diagnostic accuracy.

[0004] Currently, mainstream methods for blastomere detection mainly rely on traditional schemes such as ellipse detection and arc segment detection. While these methods offer fast processing speeds and low computational resource consumption, their performance in practical applications is often unsatisfactory. This limitation becomes even more pronounced in the 4-cell stage and above, due to the increased number and high overlap of blastomeres.

[0005] In contrast, deep learning algorithms, due to their excellent generalization ability and powerful automatic feature learning capabilities, have been widely used in various object detection tasks and have achieved remarkable results. However, in the field of blastomere detection, deep learning algorithms face unique difficulties in dealing with the challenge of overlapping blastomeres. As the number of blastomeres increases, the overlap between them becomes more significant, and the morphological overlap patterns of blastomeres also exhibit diversity, posing a greater challenge to traditional detection algorithms. Furthermore, due to the semi-transparent nature of blastomeres, the overlap between them has unique characteristics compared to the complete occlusion caused by overlap in other scenes. This makes it difficult for current detection algorithms to adapt to scenes with highly overlapping blastomeres, and their detection results are often unsatisfactory.

[0006] First, the increased overlap of blastomeres makes feature extraction and boundary detection more complex, and traditional feature representation and detection methods struggle to effectively address this situation. Second, due to the diversity of blastomere morphological overlap patterns, existing algorithms face difficulties in capturing and distinguishing this diversity. Finally, existing algorithms may lack robustness to highly overlapping blastomeres, failing to accurately identify and locate overlapping blastomeres. Summary of the Invention

[0007] This invention aims to address the shortcomings of existing technologies by proposing a method for detecting overlapping blastomeres based on multi-dimensional feature fusion. This method aims to solve the problem of low accuracy in detecting blastomeres in highly overlapping environments, enhance the model's generalization ability to cope with blastomere occlusion, thereby maintaining high detection efficiency while improving detection accuracy, and further facilitating the application of blastomere detection in highly overlapping scenarios.

[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0009] The present invention provides a method for detecting overlapping blastomeres based on multi-dimensional feature fusion, characterized by the following steps:

[0010] Step 1: Obtain embryo image datasets at the 1-8 cell stage and perform preprocessing to obtain an embryo image sample set. , Let N represent the i-th embryo image sample, and N represent the total number of embryo image samples. The blastomeres defined in the middle are labeled to obtain The label makes The actual location of the j-th blastomere is denoted as ,in, , They are respectively The coordinates of the top left and top right corners of the j-th blastomere are: , These are the coordinates of the lower left and lower right corners of the j-th blastomere, respectively;

[0011] Step 2, for Perform a Fourier transform to obtain phase With amplitude Thus, by using equation (1), we can obtain new phase :

[0012] (1)

[0013] In equation (1), Represents the weight value. Indicates the first j Embryo image samples The phase;

[0014] Step 3, for new phase With amplitude Perform the inverse Fourier transform to obtain the new fourth... i Embryo image samples This constitutes a new set of embryo image samples. ;

[0015] Step 4: Construct a blastomere target detection network, including: backbone network, neck network, and head network;

[0016] Step 4.1, the backbone network... The corresponding deep fusion convolution feature map of the blastomere is obtained through processing. ;

[0017] Step 4.2, the neck network pair Processing is performed to obtain the first... i Global feature map of individual blastomeres ;

[0018] Step 4.3: The head network includes: M convolutional layers, H pooling layers, and a prediction head;

[0019] The input is fed into the head network and processed through M convolutional layers for feature extraction, resulting in the... i Convolutional feature map of individual blastomeres ;

[0020] After downsampling through H pooling layers, the th... i Feature map of blastomeres ;

[0021] Predicting head pairs After processing, we obtain Predicted coordinates of the bounding box of the j-th blastomere and the predicted category probability of blastomeres ;in, They are respectively The predicted coordinates of the top left and top right corners of the j-th blastomere are as follows: , They are respectively The predicted coordinates of the lower left and lower right corners of the j-th blastomere;

[0022] Step 4.4: Construct the total loss of the blastomere target detection network. ;

[0023] Step 5, based on A blastomere detection network was trained using gradient descent and the total loss was calculated. Update network parameters until total loss is reached. Once convergence or the maximum number of iterations is reached, training stops, thereby generating the blastomere detection model corresponding to the optimal parameters, which is used to detect blastomeres in the embryo image sample set of the test set.

[0024] The method for detecting overlapping blastomeres based on multi-dimensional feature fusion described in this invention is also characterized in that the backbone network in step 4.1 sequentially includes: a first convolutional module, S An enhanced convolutional feature fusion module, P Each enhanced downsampling feature fusion module includes: a convolutional module and a C2f_RFMGConv module; each enhanced downsampling feature fusion module includes: a downsampling module and a C2f_RFMGConv module.

[0025] Step 4.1.1 Feature extraction is performed in the first convolutional module of the input backbone to generate... Initial characteristics of blastomeres ;

[0026] Step 4.1.2, when s=1, Enter the first s In each enhanced convolutional feature fusion module, and after processing by a convolutional module, the result is... The s Standard convolution feature map of individual blastomeres After further processing by a C2f_RFMGConv module, the result is obtained. s Standard fusion feature map of individual blastomeres ;

[0027] Step 4.1.3, when p=1, Enter the first p In the enhanced downsampling feature fusion module, and after processing by a downsampling module, the first... p blastome convolution downsampling feature map After further processing by a C2f_RFMGConv module, the result is obtained. p Feature diagram of blastomer fusion ;

[0028] when p =2,3,…, P At that time, the first p- Feature map of fusion of blastomeres Enter the first p The process is carried out in the enhanced downsampling feature fusion module, thereby being processed by the first... P The final output of the enhanced downsampling feature fusion module is the [number]th [module name]. P Deep fusion convolution feature map of individual blastomeres .

[0029] Furthermore, the C2f_RFMGConv module in step 4.1.2 is composed of a second convolutional module, a feature map splitting module, a Bottleneck_RFMGConv module, and a third convolutional module in sequence;

[0030] Step 4.1.2.1 After processing by the second convolution module, the result is the first... s High-level convolutional feature map of individual blastomeres ;

[0031] Step 4.1.2.2 The input feature map is processed in the feature map splitting module to obtain the first... s Pre-convolutional feature map of individual blastomeres and the s Feature map of each blastomere after high-level convolution ;

[0032] Step 4.1.2.3, the convolution module in the Bottleneck_RFMGConv module... Feature extraction is performed to obtain the first... s Deep convolution feature map of individual blastomeres ;

[0033] The multi-dimensional feature fusion module in the Bottleneck_RFMGConv module utilizes equation (2) to... After feature generation processing, the first... s Global feature map of individual blastomeres And then After feature rearrangement processing, the rearrangement feature map of the s-th blastomeres is obtained. :

[0034] (2)

[0035] In equation (2), For activation function, For normalization processing, This is a two-dimensional convolution operation. For the weight tensor of the convolution kernel in the multi-dimensional feature fusion module, This represents the stride of the convolution operation;

[0036] Using equation (3), the s-th high-pooling feature is obtained respectively. The s-th width pooling feature and the s-th global average pooling feature Thus, by using equation (4), the first... s Feature diagram of blastomer fusion :

[0037] (3)

[0038] (4)

[0039] In equation (3), This is a two-dimensional adaptive average pooling operation. The pooling output size is defined by the dimension h (height) and the dimension w (width) as 1; h represents the height of the input feature map, and w represents the width of the input feature map. This represents the weight tensor of the convolution kernel in the average pooling module; This indicates that the input feature map is at the location Eigenvalues ​​at;

[0040] In equation (4), , , These represent the weight tensors in the convolution kernels used in the corresponding height pooling, width pooling, and global average pooling operations, respectively.

[0041] Step 4.1.2.4 After feature extraction processing by the third convolutional module, the first convolutional module is obtained. s Standard fusion feature map of individual blastomeres .

[0042] Furthermore, the neck network in step 4.2 sequentially includes: LThe system comprises an enhanced upsampling feature fusion module, a fourth convolutional module, a first C2f_RFMGConv module, and a C2fCIB module; wherein each enhanced upsampling feature fusion module includes: an upsampling module and a C2f_RFMGConv module;

[0043] Step 4.2.1, when hour, The first input to the neck network In the enhanced upsampling feature fusion module, and after processing by an upsampling module, the first... Sampling convolutional feature map of individual blastomeres ,Will Enter the number The process is performed in the C2f_RFMGConv module of the enhanced upsampling feature fusion module to obtain the first... Deep fusion feature map of individual blastomeres ;

[0044] when =2,3,…, L At that time, the first Deep fusion feature map of individual blastomeres Enter the first The process is carried out in the enhanced convolutional feature fusion module, thereby being processed by the first... L The final output of the enhanced upsampling feature fusion module is the [number]th [module name]. L Feature diagram of fusion of blastomeres ;

[0045] Step 4.2.2 The input is processed in the fourth convolutional module for feature extraction, resulting in the... i Postblastoma characteristic diagram ;

[0046] Step 4.2.3 The input is processed in the first C2f_RFMGConv module to obtain the... i Feature diagram of fusion of blastomeres ;

[0047] Step 4.2.3 After being processed in the C2fCIB module, the result is the first... i Global feature map of individual blastomeres .

[0048] Furthermore, step 4.4 includes:

[0049] Step 4.4.1: Construct the bounding box loss using equation (5). :

[0050] (5)

[0051] In equation (5), J represents The predicted total number of blastomeres;

[0052] Step 4.4.2: Construct the cross-entropy loss using equation (6). :

[0053] (6)

[0054] Step 4.4.3: Construct the total loss using equation (7). :

[0055] (7).

[0056] 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 overlapping blastomere target detection method, and the processor is configured to execute the program stored in the memory.

[0057] 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 overlapping blastomere target detection method.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] 1. This invention uses phase fusion to augment the training set. Phase fusion is used to simulate occlusion interference between blastomeres in the multicellular stage, which greatly enriches the dataset without changing the network complexity and improves the generalization of the blastomere detection model.

[0060] 2. This invention proposes the C2f_RFMGConv module, which preserves and highlights details at specific locations through multi-dimensional feature fusion (weighted summation of height, width, and channels). In the overlapping region of blastomeres, the module captures subtle differences through a fine-grained feature generation mechanism, enabling the extraction of local and global features at higher resolution. The refined feature extraction process makes the generated feature weights more accurate, helping to suppress background interference and thus improving the ability of the blastomere detection network to separate overlapping regions. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the training method of the overlapping blastomere detection model according to an embodiment of the present invention;

[0062] Figure 2 This is a schematic flowchart of the phase fusion algorithm according to an embodiment of the present invention;

[0063] Figure 3 This is a structural diagram of the overlapping blastomere detection model according to an embodiment of the present invention;

[0064] Figure 4 This is a structural diagram of the C2f_RFMGConv module according to an embodiment of the present invention;

[0065] Figure 5 This is a diagram illustrating the detection effect of overlapping blastomeres according to an embodiment of the present invention. Detailed Implementation

[0066] In this embodiment, a method for detecting overlapping blastomeres based on multi-dimensional feature fusion is described, such as... Figure 1 As shown, it includes the following steps:

[0067] Step 1: Obtain embryo image datasets at the 1-8 cell stage and perform preprocessing to obtain an embryo image sample set. , Let N represent the i-th embryo image sample, and N represent the total number of embryo image samples. The blastomeres defined in the middle are labeled to obtain The label makes The actual location of the j-th blastomere is denoted as ,in, , They are respectively The coordinates of the top left and top right corners of the j-th blastomere are: , These are the coordinates of the lower left and lower right corners of the j-th blastomere, respectively. In this embodiment, samples collected from the time difference incubator at Chongqing Maternal and Child Health Hospital are used. The training set contains images of the 1-4 cell stage, and the test set contains images of the 1-8 cell stage. The image size is 800×800.

[0068] Step 2, as follows Figure 2 As shown, images from the training sample set at the 1-2 cell stage are selected. ,right Perform a Fourier transform to obtain phase With amplitude Thus, by using equation (1), we can obtain new phase :

[0069] (1)

[0070] in Represents the weight value. Indicates the first j Embryo image samples The phase. In this embodiment, only 50% of the 1-2 cell training set images are selected for phase fusion data augmentation, where the weight values ​​are... It is a random value between 0.01 and 0.2.

[0071] Step 3, for new phase With amplitude Perform the inverse Fourier transform to obtain the new fourth... i Embryo image samples This constitutes a new set of embryo image samples. ;

[0072] Step 4: Construct a blastomere target detection network, including: a backbone network, a neck network, and a head network, such as... Figure 3 As shown;

[0073] Step 4.1: The backbone network includes, in sequence: the first convolutional module, S An enhanced convolutional feature fusion module, P There are two enhanced downsampling feature fusion modules; wherein, each enhanced convolutional feature fusion module includes: a convolutional module and a C2f_RFMGConv module, and each enhanced downsampling feature fusion module includes: a downsampling module and a C2f_RFMGConv module; in this embodiment, there are two enhanced convolutional feature fusion modules and two enhanced downsampling feature fusion modules.

[0074] Step 4.1.1 Feature extraction is performed in the first convolutional module of the input backbone to generate... Initial characteristics of blastomeres In this embodiment, the size of the convolution kernel is 3×3, and the stride is 2.

[0075] Step 4.1.2, when s=1, Enter the first s In each enhanced convolutional feature fusion module, and after processing by a convolutional module, the result is... The s Standard convolution feature map of individual blastomeres After further processing by a C2f_RFMGConv module, the result is obtained. s Standard fusion feature map of individual blastomeres ;

[0076] like Figure 4As shown, the C2f_RFMGConv module consists of a second convolutional module, a feature map splitting module, a Bottleneck_RFMGConv module, and a third convolutional module. In this embodiment, the Bottleneck_RFMGConv module introduces a gating mechanism: when the input is True, a residual connection is introduced; if the input is False, no residual is introduced. In the experiment, all inputs were set to True.

[0077] Step 4.1.2.1 After processing by the second convolution module, the result is the first... s High-level convolutional feature map of individual blastomeres ;

[0078] Step 4.1.2.2 The input feature map is processed in the feature map splitting module to obtain the first... s Pre-convolutional feature map of individual blastomeres and the s Feature map of each blastomere after high-level convolution .

[0079] Step 4.1.2.3, the convolution module in the Bottleneck_RFMGConv module... Feature extraction is performed to obtain the first... s Deep convolution feature map of individual blastomeres ;

[0080] The multi-dimensional feature fusion module in the Bottleneck_RFMGConv module utilizes equation (2) to... After feature generation processing, the first... s Global feature map of individual blastomeres And then After feature rearrangement processing, the rearrangement feature map of the s-th blastomeres is obtained. And the s-th highly pooled feature is obtained using equation (3). The s-th width pooling feature and the s-th global average pooling feature Thus, by using equation (4), the first... s Feature diagram of blastomer fusion :

[0081] (2)

[0082] (3)

[0083] (4)

[0084] In equation (2), For activation function, To normalize the input features, This is a two-dimensional convolution operation. For the weight tensor of the convolution kernel in the multi-dimensional feature fusion module, This represents the stride of the convolution operation;

[0085] In equation (3), This is a two-dimensional adaptive average pooling operation. The pooling output size is defined by the dimension h (height) and the dimension w (width) as 1; h represents the height of the input feature map, and w represents the width of the input feature map. This represents the weight tensor of the convolution kernel in the average pooling module; This indicates that the input feature map is at the location The value at;

[0086] In equation (4), , , These represent the weight tensors in the convolution kernels used in the corresponding height pooling, width pooling, and global average pooling operations, respectively; in this embodiment... The value is 2.

[0087] Step 4.1.2.4 After feature extraction processing by the third convolutional module, the first convolutional module is obtained. s Standard fusion feature map of individual blastomeres .

[0088] Step 4.1.3, when p=1, Enter the first p In the enhanced downsampling feature fusion module, and after processing by a downsampling module, the first... p blastome convolution downsampling feature map After further processing by a C2f_RFMGConv module, the result is obtained. p Feature diagram of blastomer fusion ;

[0089] when p =2,3,…, P At that time, the first p- Feature map of fusion of blastomeres Enter the first p The process is carried out in the enhanced downsampling feature fusion module, thereby being processed by the first... P The final output of the enhanced downsampling feature fusion module is the [number]th [module name]. P Deep fusion convolution feature map of individual blastomeres .

[0090] Step 4.2, the neck network sequentially includes:L The system comprises two enhanced upsampling feature fusion modules: a fourth convolutional module, a first C2f_RFMGConv module, and a C2fCIB module. Each enhanced upsampling feature fusion module includes an upsampling module and a C2f_RFMGConv module. In this embodiment, two enhanced upsampling feature fusion modules are included.

[0091] Step 4.2.1, when hour, The first input to the neck network In the enhanced upsampling feature fusion module, and after processing by an upsampling module, the first... Sampling convolutional feature map of individual blastomeres ,Will Enter the number The process is performed in the C2f_RFMGConv module of the enhanced upsampling feature fusion module to obtain the first... Deep fusion feature map of individual blastomeres In this embodiment, the nearest neighbor interpolation method is used for upsampling. The width magnification factor of the two upsampling modules is 2, while the height remains unchanged.

[0092] when =2,3,…, L At that time, the first Deep fusion feature map of individual blastomeres Enter the first The process is carried out in the enhanced convolutional feature fusion module, thereby being processed by the first... L The final output of the enhanced upsampling feature fusion module is the [number]th [module name]. L Feature diagram of fusion of blastomeres ;

[0093] Step 4.2.2 The input is processed in the fourth convolutional module for feature extraction, resulting in the... i Postblastoma characteristic diagram ;

[0094] Step 4.2.3 The input is processed in the first C2f_RFMGConv module to obtain the... i Feature diagram of fusion of blastomeres ;

[0095] Step 4.2.3 After being processed in the C2fCIB module, the result is the first... i Global feature map of individual blastomeres In this embodiment, the C2fCIB module outputs 3 channels, and all Boolean parameters are set to True.

[0096] Step 4.3: The head network includes: M convolutional layers, H pooling layers, and a prediction head; in this embodiment, it includes 4 convolutional layers and 1 pooling layer.

[0097] The input is fed into the head network and processed through M convolutional layers for feature extraction, resulting in the... i Convolutional feature map of individual blastomeres ;

[0098] After downsampling through H pooling layers, the th... i Feature map of blastomeres ;

[0099] Predicting head pairs After processing, we obtain Predicted coordinates of the bounding box of the j-th blastomere and the predicted category probability of blastomeres ;in, They are respectively The predicted coordinates of the top left and top right corners of the j-th blastomere are as follows: , They are respectively The predicted coordinates of the lower left and lower right corners of the j-th blastomere.

[0100] Step 4.4: Construct the total loss of the blastomere target detection network. ;

[0101] Step 4.4.1: Construct the bounding box loss using equation (5). :

[0102] (5)

[0103] In equation (5), J represents The predicted total number of blastomeres;

[0104] Step 4.4.2: Construct the cross-entropy loss using equation (6). :

[0105] (6)

[0106] Step 4.4.3: Construct the total loss using equation (7). :

[0107] (7)

[0108] Step 5, based on A blastomere detection network was trained using gradient descent and the total loss was calculated. Update network parameters until total loss is reached. Once convergence or the maximum number of iterations is reached, training stops, thereby generating the blastomere detection model corresponding to the optimal parameters, which is used to detect blastomeres in the embryo image sample set of the test set.

[0109] 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. Early stopping was implemented if no improvement in the test loss was observed on the validation set for 15 consecutive epochs. The batch size was set to 16.

[0110] 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.

[0111] 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.

[0112] 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. True positive (TP), false negative (FN), and recall (R) were used as evaluation indicators. True positive (TP) represents the number of blastomeres that were correctly detected, false negative (FN) represents the number of blastomeres that were not detected, also known as the number of missed detections, and recall (R) represents the percentage of blastomeres that were correctly detected.

[0113] Table 1. Number of cell images of each type in the embryo dataset (per image)

[0114]

[0115] The trained network was tested on the test set, and its detection performance was compared with that of Transformer, YOLOv5, and YOLOv10 networks. It demonstrated significantly higher detection accuracy in 4-8 cells, as shown in Table 2. Typical blastomere detection images are shown below. Figure 5 As shown.

[0116] Table 2 Comparison of experimental results on cells 1-8

[0117]

[0118] Table 2 and Figure 5 The combined results show that this invention can significantly improve the detection of overlapping blastomeres. Ours represents the network traffic prediction model proposed in this invention. Transformer, YOLOv5, and YOLOv10 represent different network models. Comparative experiments show that the overlapping blastomeres detection model proposed in this invention achieves the best results in terms of evaluation metrics. This invention has good application potential and prospects, providing methods and ideas for future practical applications in embryo assessment.

Claims

1. A method for detecting overlapping blastomere objects based on multi-dimensional feature fusion, characterized in that, is carried out as follows: Step 1, after obtaining and preprocessing the embryo image dataset of the 1-8 cell stage, an embryo image sample set is obtained , represents the i-th embryo image sample, and N represents the total number of embryo image samples, and the label of the i-th embryo image sample is represented as , and the label of the i-th embryo image sample is represented as , and the label of the i-th embryo image sample is represented as , and the label of the i-th embryo image sample is represented as , wherein , , and the label of the i-th embryo image sample is represented as , and the label of the i-th embryo image sample is represented as , , and the label of the i-th embryo image sample is represented as Step 2, Fourier transform of gives the phase and the amplitude of so that a new phase of is obtained using equation (1) (1) In formula (1), representing a weight value, denotes the phase of the jthembryo image sample . Step 3, for new phase With amplitude Perform an inverse Fourier transform to obtain a new i-th embryo image sample. This constitutes a new set of embryo image samples. ; Step 4, build a blastomere target detection network, including: a backbone network, a neck network and a head network; Step 4.1, the backbone network pair is processed to obtain the corresponding deep layer fusion convolution feature map of the blastomere ; Step 4.2, the neck network pair processing to obtain the ith blastomere global feature map ; Step 4.3, the head network comprises: M convolutional layers, H pooling layers, a prediction head; After inputting into the head network and performing feature extraction processing through M convolutional layers, the i-th blastomere convolutional feature map is obtained ; After the down-sampling processing in the H input pooling layers, the down-sampled feature map of the i-th blastomere is obtained ; Predicting head pairs After processing, we obtain Predicted coordinates of the bounding box of the j-th blastomere and the predicted category probability of blastomeres ;in, They are respectively The predicted coordinates of the top left and top right corners of the j-th blastomere are as follows: , They are respectively The predicted coordinates of the lower left and lower right corners of the j-th blastomere; Step 4.4, Constructing the total loss of the blastomere target detection network ; Step 5, based on , the gradient descent method is used to train the blastomere target detection network, and the total loss is calculated to update the network parameters until the total loss converges or reaches the maximum number of iterations, and the training is stopped, thereby generating a blastomere detection model corresponding to the optimal parameters, which is used to detect blastomeres in the test set of embryo image samples. 2.The method of claim 1, wherein, The backbone network in the step 4.1 comprises in sequence: a first convolutional module, S enhanced convolutional feature fusion modules, P enhanced down-sampling feature fusion modules; wherein each enhanced convolutional feature fusion module comprises: a convolutional module, a C2f_RFMGConv module, and each enhanced down-sampling feature fusion module comprises: a down-sampling module, a C2f_RFMGConv module; Step 4.1.1, feature extraction is performed in the first convolution module of the input backbone to generate the initial feature map of the blastomere ; Step 4.1.2, when s = 1, The s-th cleavage ball standard convolution feature map of the s-th enhanced convolution feature fusion module is obtained after the s-th enhanced convolution feature fusion module is input and processed by a convolution module The s-th cleavage ball standard fusion feature map is obtained after the s-th cleavage ball standard convolution feature map is processed by a C2f_RFMGConv module ;​ Step 4.1.3, when p = 1, The p-th blastomere convolution down-sampling feature map is obtained after the p-th enhanced down-sampling feature fusion module is input and processed by a down-sampling module The p-th blastomere fusion feature map is obtained after the p-th blastomere convolution down-sampling feature map is processed by a C2f_RFMGConv module ; When p = 2, 3, …, P, the (p-1)th blastomere fusion feature map is fused The pth enhanced down-sampling feature fusion module is input for processing, so that the final pth blastomere deep fusion convolution feature map is output from the pth enhanced down-sampling feature fusion module . 3.The method of claim 2, wherein, The C2f_RFMGConv module in the step 4.1.2 is sequentially composed of a second convolutional module, a feature map splitting module, a Bottleneck_RFMGConv module and a third convolutional module; Step 4.1.2.1、 After processing by the second convolution module, the high-level convolution feature map of the s-th blastomere is obtained ; Step 4.1.2.2, The high-level pre-convolutional feature map of the s-th blastomere and the high-level post-convolutional feature map of the s-th blastomere are obtained by processing in the input feature map splitting module and the high-level post-convolutional feature map of the s-th blastomere ; Step 4.1.2.

3. The convolution module pair in the Bottleneck_RFMGConv module performs feature extraction to obtain the s-th blastomere deep convolution feature map ; and ; The multi-dimensional feature fusion module in the Bottleneck_RFMGConv module utilizes formula (2) on After the feature generation processing, the s-th blastomere global feature map is obtained After the feature rearrangement processing, the s-th blastomere rearranged feature map is obtained :​ (2) In formula (2), is an activation function, is a normalization processing, is a two-dimensional convolution operation, is a weight tensor of a convolution kernel in a multi-dimensional feature fusion module, is a stride of the convolution operation; The s-th high-pooling feature is obtained by using formula (3) respectively The s-th width-pooling feature The s-th global average-pooling feature The s-th blastomere fusion feature map is obtained by using formula (4) : (3) (4) In formula (3), is a two-dimensional adaptive average pooling operation, h represents the height of the input feature map, and w represents the width of the input feature map, represents a weight tensor of a convolution kernel in the average pooling module; represents a feature value of the input feature map at position in the input feature map. In formula (4), , , respectively represent the weight tensors in the convolution kernels used in the corresponding high-level pooling, wide-level pooling, and global average pooling operations. Step 4.1.2.4、 After the feature extraction processing through the third convolution module, the standard fusion feature map of the s-th blastomere is obtained . 4.The method of claim 3, wherein, The neck network in the step 4.2 comprises in sequence: L enhanced up-sampling feature fusion modules, a fourth convolutional module, a first C2f_RFMGConv module and a C2fCIB module; wherein each enhanced up-sampling feature fusion module comprises: an up-sampling module and a C2f_RFMGConv module; Step 4.2.1, when hour, The first input to the neck network In the enhanced upsampling feature fusion module, and after processing by an upsampling module, the first... Sampling convolutional feature map of individual blastomeres ,Will Enter the number The process is performed in the C2f_RFMGConv module of the enhanced upsampling feature fusion module to obtain the first... Deep fusion feature map of individual blastomeres ; When = 2, 3, …, L, the deep layer fusion feature map of the i-th blastomere is input into the i-th enhanced convolution feature fusion module for processing, so that the final i-th blastomere fusion feature map is output by the L-th enhanced up-sampling feature fusion module ;​​​ Step 4.2.2、 input the fourth convolutional module for feature extraction to obtain the i-th post-cleavage blastomere feature map ; Step 4.2.3, Input the first C2f RFMGConv module for processing to obtain the i-th blastomere fusion feature map ; Step 4.2.3, After inputting into the C2fCIB module for processing, the global feature map of the ith blastomere is obtained .

5. The method according to claim 4, wherein, The step 4.4 comprises: Step 4.4.1, constructing the bounding box loss with formula (5) : (5) In formula (5), J represents the total number of blastomeres predicted in the middle Step 4.4.2, constructing the cross-entropy loss with formula (6) : (6) Step 4.4.3, building the total loss with formula (7) : (7)。 6. An electronic device comprising a memory and a processor, characterized in that The memory is configured to store a program supporting the processor to execute the overlapping blastomere target detection method in any one of claims 1-5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to execute the steps of the overlapping blastomere target detection method in any one of claims 1-5.

Images