Embryo development prediction classification system based on double-layer optimization framework

Abstract

The invention provides an embryonic development prediction and classification system based on a double-layer optimization framework. The embryonic development prediction and classification system comprises an upper-layer self-supervised training model and a lower-layer embryonic development prediction and classification model which share an encoder, training is carried out by adopting an embryo development image sequence; a model parameter # imgabs0 # of a lower-layer embryo image development prediction classification model is fixed, and a model parameter # imgabs1 # of an upper-layer self-supervised training model is optimized; fixing a model parameter # imgabs2 # to optimize the model parameter of the lower embryo image development prediction classification model; and repeating the steps until iteration is completed. The collaborative optimization strategy not only improves the precision of the model, but also enables the model to have stronger adaptive capacity when facing variable and complex medical data.

Description

Technical Field

[0001] The present invention relates to the technical field of embryonic development detection, and in particular to an embryonic development prediction and classification system based on a double-layer optimization framework. Background Art

[0002] In the process of in vitro fertilization (IVF), selecting the embryo with the best development potential for maternal implantation is one of the key steps to ensure a successful pregnancy. Since the developmental potential of each embryo is different, selecting the best embryo can significantly improve the implantation rate and pregnancy success rate. The quality of the embryo is affected by many factors, including the morphological characteristics of the embryo, chromosomal normality, genome stability, maternal hormone levels, and endometrial preparation. Among them, the morphological evaluation is mainly aimed at the analysis of embryo images obtained during embryonic development, which is one of the most widely used methods. By comprehensively evaluating the embryo images at different developmental stages, the embryos with the highest developmental potential can be selected for transplantation. Therefore, research on embryo implantation prediction provides a scientific basis for embryologists to select the most suitable embryos.

[0003] Morphological embryo assessment is generally performed by analyzing a series of embryo development images taken by a built-in CCD camera in a time-lapse incubator. These images record the various developmental stages of the embryo from fertilization to embryo transfer, helping embryologists observe the development of the embryo. By analyzing embryo images, the number of embryo cells, the synchronization of cell division, the symmetry of cell morphology, and the presence of embryo fragments can be evaluated. Morphological assessment is currently one of the most commonly used and intuitive embryo screening methods, which can provide a preliminary quality judgment for embryo selection. Traditionally, professionals need to manually conduct a systematic evaluation and analysis of a large number of images during embryo development. This is a very time-consuming task that is susceptible to human subjective bias. Especially when faced with a large number of embryo images, the efficiency and accuracy of manual evaluation often have certain limitations. In addition, some subtle morphological features may be overlooked during manual analysis. Although these features are not easy to detect, they may have an important impact on the developmental potential and implantation success rate of the embryo.

[0004] In recent years, with the rise of deep learning algorithms, embryo morphological assessment has been significantly optimized. Deep learning technology, especially image recognition algorithms such as convolutional neural networks (CNNs), has been widely used in embryo image analysis. These algorithms can automatically and accurately analyze and evaluate a large number of embryo images, thereby improving the efficiency and accuracy of screening. Compared with traditional manual evaluation, deep learning can identify more subtle morphological differences, even those details that are difficult to detect with the naked eye, thus providing a more reliable basis for embryo selection. By training embryo images at different developmental stages, deep learning models can gradually learn and identify key features of high-quality embryos, such as the regularity of cell division, the symmetry of cell morphology, and the degree of fragmentation. These features are often easily interfered by human factors in traditional manual evaluation, while deep learning can extract these potential rules from massive data and achieve more objective and consistent analysis. The advantage of this technology is that it not only improves the speed and efficiency of evaluation, but also provides more refined judgments, helping embryologists make more accurate decisions. At the same time, the introduction of deep learning technology also solves the problems of "fatigue" and "bias" in traditional evaluation methods. When faced with a large number of embryo images, manual evaluation is prone to loss of attention or evaluation bias due to long-term work, while AI algorithms can maintain consistent judgment standards at all times, thereby reducing human errors.

[0005] For deep learning with traditional supervised learning, its method relies on task-specific label information during the training process. Although it can achieve excellent performance when the labeled data is rich, this method also has some limitations. First, the limitation of feature learning is a key issue. Supervised learning usually assumes that the features in the data can be directly used to solve the task objectives, but this assumption is often oversimplified in practical applications, and feature learning is greatly affected by the data distribution. Especially when the feature space is complex or the task objectives are unclear, traditional supervised learning often finds it difficult to effectively extract valuable features, resulting in the model's excessive dependence on training data and poor generalization ability. Secondly, the training process of traditional supervised learning is susceptible to overfitting. Since the task-specific objectives and loss functions only focus on the labels of the training data during optimization, the model may only achieve good performance on the training set, but has weak generalization ability on the test set or unknown data. When the distribution of training data changes greatly, traditional methods often cannot cope with new data well, thus affecting the stability and practicality of the model. In addition, although the combination of self-supervised learning and supervised learning can make up for the limitations of supervised learning to a certain extent, the two are often carried out independently and lack effective synergy. Summary of the invention

[0006] The present invention proposes an embryo development prediction and classification system based on a two-layer optimization framework to solve the technical problem that in traditional training methods, the self-supervised pre-training and deep learning model training processes are carried out independently and cannot form an effective synergistic effect, resulting in poor prediction accuracy of the deep learning model after training.

[0007] To solve the above technical problems, the present invention provides an embryo development prediction and classification system based on a two-layer optimization framework, wherein the system comprises an upper self-supervised training model and a lower embryo development prediction and classification model of a shared encoder; The system is trained using an embryonic development image sequence: the model parameters of the lower embryonic image development prediction classification model are fixed. , the model parameters of the upper self-supervised training model Optimize; then fix the model parameters Optimizing the model parameters of the lower embryo image development prediction and classification model; repeating the above steps until the iteration is completed.

[0008] Preferably, the encoder comprises a first encoder and a second encoder; The first encoder is composed of a module including a 3x3 convolutional neural network and a ReLU activation function; The second encoder uses ResNet50 as the basic network.

[0009] Preferably, a bilinear pooling operation is performed on the features extracted by the second encoder.

[0010] Preferably, the first decoder of the upper-layer self-supervised training model consists of three ConvLSTM modules.

[0011] Preferably, the second decoder of the lower-layer embryo image development prediction classification model first converts the high-dimensional feature map into a one-dimensional feature vector through a Flatten feature flattening operation, and then performs nonlinear transformation and weighted processing on the flattened features through a KAN module.

[0012] Preferably, the upper layer self-supervised training model is trained using a point-to-point contrast loss function.

[0013] Preferably, the contrast loss function The expression is: ; ; ; In the formula, Represents an embryo image sequence subsequence The number of samples; express The total number of feature points in ; Represents model predictions The number of samples in the time dimension; and Represent subsequences height and width in spatial dimensions; , both represent indexes, Represents a subsequence Middle The index of the samples, Used to represent the index number of another sample in calculating the point-to-point contrast loss; , both are used to represent indexes, Represents a subsequence Middle The index number of the feature point, Used to represent the index number of the feature point of another sample in calculating the point-to-point contrast loss; Represents a subsequence Middle Embryo samples The real embryonic development feature points at each position; Represents the model for subsequence Middle Embryo samples The predicted embryonic developmental features at each position; Representation calculation and The cosine similarity between ; Represents the temperature parameter.

[0014] Preferably, the cross entropy loss function The expression is: ; In the formula, Refers to a sequence of embryo images The corresponding label is passed through a one-hot encoder to obtain the encoded label; The model is a sequence The predicted probability of .

[0015] Preferably, before using the embryonic development image sequence for training, the embryonic development image sequence is resized and reduced.

[0016] The beneficial effects of the present invention at least include: the present invention adopts a double-layer optimization strategy, so that the fine-tuning stage of the upper self-supervised training model and the lower embryonic development prediction classification model can be collaboratively optimized, rather than independently performed. This collaborative optimization method avoids the performance bottleneck between pre-training and fine-tuning in the traditional method. Through self-supervised learning, the model can more effectively capture the key features and potential patterns in the embryonic development process, help the model to mine the intrinsic relationship and structural information in the data, so that the model can better understand the internal feature structure of the data when processing complex medical images, thereby improving the learning efficiency and feature expression ability of the model. In addition, the generalized features learned by the upper self-supervised training model can help the fine-tuning stage of the lower embryonic development prediction classification model to adapt to the task faster, improve the adaptability to new samples, and further enhance the robustness and generalization ability of the model. This collaborative optimization strategy not only improves the accuracy of the model, but also makes it have stronger adaptability when facing changeable and complex medical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of the system structure of an embodiment of the present invention; Figure 2 A schematic diagram of a double-layer optimization process according to an embodiment of the present invention; Figure 3 Schematic diagram of the architecture of a network encoder and decoder according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0019] Example 1 like Figure 1 As shown, an embodiment of the present invention provides an embryo development prediction and classification system based on a two-layer optimization framework, the system comprising an upper-layer self-supervised training model of a shared encoder and a lower-layer embryo development prediction and classification model.

[0020] Specifically, this embodiment mainly uses image data of the entire embryonic development process recorded in a time-lapse incubator to assist embryologists in conducting predictive research before embryo transplantation. The embryo image data used comes from multiple different reproductive centers, which provides a rich and diverse data source for this embodiment and helps to demonstrate the accuracy and reliability of the system proposed in this embodiment.

[0021] In order to ensure the accuracy and representativeness of the image data, all embryo images in the experiments were captured in the incubator using time-lapse photography technology. This technology can continuously record the development process of the embryo without disturbing its normal growth and development, thereby obtaining high-quality, high-definition embryo development image data. The dataset covers all key stages of embryonic development, from cell division of the fertilized egg, differentiation of each cell of the embryo, to the complete development process of the blastocyst stage. The total number of embryo images included in this process is between 300 and 500. These images provide rich time series data for the study, allowing researchers to accurately track the changes in embryos at different stages of development, and further lay a solid data foundation for the development of image-based automated evaluation and prediction methods. These data not only provide support for clinical practice in the field of assisted reproduction, but also provide valuable reference materials for the application of deep learning algorithms in embryo quality assessment.

[0022] The embryo image data set collected in the present invention is analyzed and evaluated by multiple embryologists, and the image sequence of the embryonic development process in the data set is distinguished by the principle of minority obeys majority as the label of the model in the present invention. The data set contains 2403 groups of embryonic development whole process image sequences. What experiment adopts in the present invention is to divide training set and test set by the way of ten-fold cross validation.

[0023] The system provided by the present embodiment comprises two parts: upper self-supervision training model and lower embryonic development prediction classification model.Wherein the upper self-supervision training model and lower embryonic development prediction classification model share the same encoder, which provides a basis for collaborative training.What the encoder shared by two models in the present embodiment adopts is optimized convolutional module and ResNet50 network structure, including but not limited to LeNet, AlexNet, VGGNet, ResNet series, EfficientNet and Mamba and Transformer model etc.Wherein the decoder of the upper self-supervision training model adopts the network architecture of optimized ConvLSTM module stacking, including but not limited to VAE, GANs and time series generation model etc.The lower embryonic development prediction classification model adopts optimized KAN module, including but not limited to MLP, CNN and capsule network etc.

[0024] The two-level optimization process is as follows Figure 2 As shown, the parameters of the network model in this embodiment are recorded as , the parameters involved in the upper self-supervised training model are recorded as , the parameters involved in the lower embryonic development prediction classification model are recorded as At the beginning of training, the model is initialized. The first step is to fix the parameters in the lower embryonic development prediction classification model. , perform the upper self-supervised training model parameters Optimize and initially obtain the best parameters in the self-supervised task ; The second step is to use the best parameters in the self-supervised task Fixed constraints to optimize parameters in the underlying embryonic development prediction classification model , to obtain the best parameters in the lower embryonic development prediction classification model ; The third step is to obtain the best parameters in the classification task in the second step Fix and repeat the first step to optimize the parameters in the upstream self-supervised task; perform double-layer optimization in this way until a fixed number of iterations is reached, or the performance of the validation set does not improve significantly in several consecutive iterations, such as less than 0.1%, then it can be considered that the model has reached the optimal state, and the optimization is terminated to finally obtain the optimal parameters of the model. .

[0025] The double-layer optimization framework adopted by the present invention effectively improves the model performance by combining the advantages of self-supervised learning and supervised learning. In the self-supervised learning stage, the model learns effective feature representations in the embryonic development process through unlabeled embryo image data, increases the generalization ability of the model on this task and improves its robustness. Then, the downstream fine-tuning stage uses the labeled embryo classification label data to help the model better understand the underlying structure of the data and optimize it. In this way, the feature representation of the embryonic development process has been significantly improved, and it can perform better in the downstream embryo implantation prediction task. Self-supervised learning enables the model to extract useful information from large-scale unlabeled data, while fine-tuning allows the model to be adjusted more accurately in specific tasks, thereby enhancing the generalization ability and task adaptability of the model. In general, the double-layer optimization framework not only improves the effect of pre-training and fine-tuning by efficiently combining self-supervised learning and supervised learning, but also enables the model to achieve better performance in an environment where data is scarce and tasks are diverse.

[0026] In order to further illustrate the effectiveness of the proposed embryo implantation prediction based on the double-layer optimization strategy, the present invention conducted a large number of ablation experiments and comparative experiments.

[0027] In the pre-training stage of self-supervised learning, the model learns the deep feature representation of embryo images through unlabeled data. Through the two-layer optimization strategy, the model not only optimizes the initial parameters of the feature extraction network during the pre-training process, but also adjusts the loss function of the downstream classification task. This two-layer optimization process enables the model to gradually improve the effect of the pre-training and fine-tuning stages, and can effectively capture the key features in the embryonic development process, thereby improving the accuracy of embryo implantation prediction. In the experiment of the present invention, the Adam optimizer is used for parameter optimization, and the learning rate is set to 0.005, the batch size is 16, and the number of training rounds is 60 times. The Adam optimizer combines weight decay and L2 regularization technology to effectively solve the problems of slow network convergence and parameter overfitting, thereby speeding up the training process of the model and improving the generalization ability.

[0028] Table 1 shows the accuracy of different method models in embryo implantation prediction tasks. From the experimental results in the table, it can be found that the system proposed in the present invention is 3.28% higher than the LeNet model in embryo implantation prediction tasks, 1.51% higher than the best model ResNet152 in the ResNet series network, and 2.23% higher than the VGGNet network in terms of accuracy. The results show that the research method based on double-layer optimization proposed in the present invention has better performance.

[0029] Table 1

[0030] In summary, the present invention adopts a double-layer optimization strategy, so that the self-supervised pre-training of upstream tasks and the fine-tuning stage of downstream embryo implantation prediction research can be collaboratively optimized, rather than independently performed. This collaborative optimization method avoids the performance bottleneck between pre-training and fine-tuning in the traditional method. Through self-supervised learning, the model can more effectively capture the key features and potential patterns in the embryonic development process, help the model to mine the intrinsic relationship and structural information in the data, so that the model can better understand the internal feature structure of the data when processing complex medical images, thereby improving the learning efficiency and feature expression ability of the model. In addition, the generalized features learned in the self-supervised pre-training stage of the upstream can help the fine-tuning stage of the downstream embryo implantation prediction task to adapt to the task faster, improve the adaptability to new samples, and further enhance the robustness and generalization ability of the model. This collaborative optimization strategy not only improves the accuracy of the model, but also makes it have stronger adaptability when facing changeable and complex medical data.

[0031] Example 2 Based on Example 1, this example improves the encoder and decoder of the upper self-supervised training model and the lower embryonic development prediction and classification model.

[0032] Specifically, the encoder shared by the upper self-supervised training model and the lower embryonic development prediction classification model includes a first encoder and a second encoder, and its structure is as follows: Figure 3 shown.

[0033] In the present embodiment, the design of the encoder is intended to fully extract and model the real developmental features in the embryo image, and provide accurate feature representation for subsequent prediction and analysis tasks. The first encoder is mainly composed of a module comprising a 3x3 convolutional neural network and a ReLU activation function, and its core goal is to extract real developmental features from the embryo image, and as a supervision source for predicting embryo development features in the self-supervised task. The encoder accurately extracts local features through the operation of a small convolution kernel, and simultaneously combines the ReLU activation function to enhance the nonlinear expression ability of the model, thereby enhancing the recognition and modeling capabilities of embryo development features.

[0034] As the backbone structure of the network, the second encoder is not only a shared part in the classification task and the self-supervision task, but also the core feature extraction module of the model. The second encoder first uses ResNet50 as the basic network, and uses its powerful representation ability to extract local and global features in the embryo image. ResNet50 can effectively capture the deep semantic information in the image through multi-level convolution and residual connection, thereby generating a feature map containing multi-level embryo development information. This feature map not only retains the detailed features of the image in space, but also encodes rich high-order semantic information in the channel dimension, providing a solid foundation for subsequent fine-grained feature analysis and task modeling. After feature extraction, the second encoder further introduces the bilinear pooling operation to capture the high-order interaction features between different local regions in the embryo image. Bilinear pooling generates high-order feature representations through outer product calculations, which can effectively model the complex relationships between local regions, thereby enhancing the model's ability to express fine-grained features. This method not only improves the model's sensitivity to local details, but also captures the subtle differences between different parts or regions, which is the key to distinguishing information at different stages during embryonic development. In addition, in the bilinear pooling operation, the second encoder generates a global feature map and an attention map in parallel. The global feature map is extracted by ResNet50 and provides rich overall semantic information; the attention map is generated by the convolutional layer, reflecting the importance of each region in the image. This attention mechanism can guide the network to focus on key areas related to embryonic development and implantation prediction. In subsequent processing, the feature map and the attention map are combined to further optimize the feature expression of the model by means of weighted feature maps. This combination enables the network to automatically identify and strengthen the focus on key information in the embryonic development process, while suppressing the influence of irrelevant or interfering areas. This not only improves the network's ability to capture fine-grained features, but also provides stronger support for subsequent downstream tasks such as embryo implantation success rate prediction. Overall, through the multi-level feature extraction of ResNet50, the high-order feature modeling of bilinear pooling, and the introduction of the attention mechanism, the second encoder deeply represents the embryonic image in both spatial and semantic dimensions, laying the foundation for the excellent performance of the entire model in the classification and prediction of embryonic development stages.

[0035] In the present embodiment, the design of the decoder is intended to achieve the prediction of future developmental stages and the classification of embryonic development prediction stages by making full use of the embryonic development feature information extracted by the encoder. The task of the first decoder of the upper self-supervised training model is to extract useful information from the spatiotemporal feature representation obtained from the encoder to predict the future stages in the embryonic development process. For this reason, the present embodiment adopts an encoder architecture consisting of three ConvLSTM modules. Each ConvLSTM module combines convolution operation and LSTM structure, and can simultaneously learn spatial features and time dependencies when processing time series data. These three continuous ConvLSTM modules extract higher-level spatiotemporal features layer by layer, so that the feature representation of the final output contains rich spatiotemporal information, which can provide sufficient support for subsequent predictions. In this way, the first decoder can not only capture the fine-grained features of the current embryonic development stage, but also effectively perform temporal modeling to predict the future development process of the embryo.

[0036] The second decoder of the lower-level embryo image development prediction and classification model focuses on predicting and classifying using the embryo development feature information extracted by the encoder, and finally completes the classification task of whether different embryo development processes are suitable for further implantation into the mother. The workflow of the second decoder first converts the high-dimensional feature map into a one-dimensional feature vector through the Flatten feature flattening operation, so that the subsequent classification task can be processed in the simplified feature space. This operation integrates the features of the spatial and temporal dimensions into a compact one-dimensional vector, providing an effective input for the classification model. Next, the second decoder performs nonlinear transformation and weighting on these flattened features through the KAN (Key Attention Network) module. The KAN module uses a learnable edge activation function, which breaks through the limitation of the fixed activation function in the traditional multi-layer perceptron (MLP), so that the activation function of each node can be adaptively adjusted according to the input features, thereby effectively enhancing the network's adaptability to complex embryo development patterns. By introducing this flexible activation mechanism, the KAN module enables the decoder to more accurately learn the features that are critical to the classification task and automatically focus on the most distinguishing areas in the embryonic development process. This design enables the second decoder to not only complete complex multi-class classification tasks, but also show higher accuracy and robustness when dealing with fine-grained classification. In general, the two decoders are designed according to different tasks by introducing multi-layer ConvLSTM modules, Flatten feature flattening operations, and learnable edge activation functions of the KAN module, so that the model can achieve excellent performance in both time series prediction and classification tasks. The first decoder effectively uses the spatiotemporal features extracted by the encoder to predict future developmental stages, while the second decoder achieves accurate classification of embryonic development stages through efficient feature selection and weighting mechanisms. Through this design that combines time series modeling with classification tasks, the model can better understand and predict complex patterns in embryonic development.

[0037] Example 3 Based on Example 1, this example constructs a loss function for the training process of the upper self-supervised training model.

[0038] Specifically, the goal of the loss function in the upper self-supervised training model is to predict the effective embryonic development feature representation of the next subsequence stage based on the current embryonic image subsequence. This process uses the developmental features of different subsequences in the embryonic image as the supervision source of the self-supervised learning task, and does not need to rely on artificial labels. Therefore, the inner loss function used in this embodiment is a point-to-point contrast loss function. : ; ; in, Indicates the current subsequence The number of embryo image samples in Represents a subsequence The total number of feature points in is calculated as follows: , Represents the predicted subsequence The number of samples in the time dimension, and Respectively represent the predicted subsequence height and width in spatial dimensions; , represents two indexes, used to traverse all embryo image samples in the embryo subsequence, represents the embryo image sample currently considered, Used to represent the index of another sample in calculating the point-to-point contrast loss. When , it means that the two are the same sample. When , it means that the two are not the same sample, and their corresponding feature points belong to the negative sample pair; , represents two indexes, used to traverse the feature points of all embryo image samples in the embryo subsequence, represents the feature point index of the embryo image sample currently considered, Used to represent the feature point index of another sample in calculating the point-to-point contrast loss. When calculating the loss of each feature point, and Will Jointly decide which feature points are positive sample pairs and which are negative sample pairs; Represents a subsequence Previous Of the embryo samples The true embryonic developmental characteristics at each position; Represents the model for Of the embryo samples Predicted embryonic developmental features at each position; Representation calculation and The cosine similarity between ; Represents the temperature parameter, which is used to adjust the scale of similarity and affect the sensitivity of the point-to-point contrast loss function. By adopting the point-to-point contrast loss function, the model can effectively capture the developmental characteristics of the embryo at different time points and improve the model's ability to distinguish between good and bad embryo development processes.

[0039] Example 4 Based on Example 1, this example constructs a loss function for the training process of the lower embryo development prediction classification model.

[0040] The goal of the lower embryonic development prediction classification model loss function is to fine-tune the model so that it can better complete the specific downstream classification tasks. The cross entropy loss function is favored because of its wide applicability and multiple advantages in classification problems. In this embodiment, the function maximizes the log likelihood in the embryo image by minimizing the loss value, thereby improving the accuracy of the model in predicting the quality of the embryonic development process. It has high sensitivity, can quickly respond to changes in model performance, and accelerate the convergence process of the model. In addition, the cross entropy loss function is easy to calculate and has good numerical stability and probabilistic interpretability, which makes it the preferred loss function in this embodiment. , and its calculation formula is as follows: ; in, refers to a set of embryo images The corresponding label is passed through the one-hot encoder to obtain the encoded label. refers to a set of embryo images The predicted probability obtained by the network model of the supervised task part.

[0041] Example 5 This embodiment pre-processes the embryo image sequence based on the embodiment 1.

[0042] The original size of the embryo image data collected by the present invention is 500x500 pixels. In order to better adapt to the input requirements of the deep learning model, the size of each image is reduced to 224x224. This adjustment ensures that the image meets the standard input size of the neural network, especially when using the pre-trained model, the input is usually required to be 224x224 pixels. In addition, reducing the image size helps to reduce the consumption of computing resources, speed up model training and reasoning speed, thereby improving overall computing efficiency. Smaller input images reduce memory usage and computing overhead, effectively improving training efficiency without sacrificing key information. During the adjustment process, a scaling technique that maintains the image aspect ratio is adopted to ensure that the core information of the embryo image will not be lost. Through these processes, the model can focus on the key areas of the image, avoid interference from irrelevant backgrounds, thereby improving classification accuracy. At the same time, the unified image size also simplifies subsequent data processing and batch processing processes, especially under the processing of large-scale data sets, the size standardization to 224x224 can optimize model training, reduce the complexity in training, and improve the generalization ability of the model. Through these preprocessing operations, this embodiment can not only effectively improve the accuracy of embryo image classification, but also significantly reduce the computing cost, providing more efficient and stable support for subsequent intelligent diagnosis and medical decision-making.

[0043] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. Only the preferred embodiments of the present invention are expressed. The description is more specific and detailed, but it cannot be understood as limiting the scope of the present invention. As long as there is no contradiction in the combination of these technical features, they should be considered as within the scope of this specification.

[0044] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. An embryonic development prediction and classification system based on a two-layer optimization framework, characterized in that: The system includes an upper self-supervised training model and a lower embryonic development prediction classification model that share an encoder; The system is trained using an embryonic development image sequence: the model parameters of the lower embryonic image development prediction classification model are fixed. , the model parameters of the upper self-supervised training model Optimize; then fix the model parameters Optimizing the model parameters of the lower embryo image development prediction and classification model; repeating the above steps until the iteration is completed.

2. The embryonic development prediction and classification system based on a double-layer optimization framework according to claim 1, characterized in that: The encoder includes a first encoder and a second encoder; The first encoder is composed of a module including a 3x3 convolutional neural network and a ReLU activation function; The second encoder uses ResNet50 as the basic network.

3. The embryonic development prediction classification system based on a double-layer optimization framework according to claim 2, characterized in that: A bilinear pooling operation is performed on the features extracted by the second encoder.

4. The embryonic development prediction classification system based on a double-layer optimization framework according to claim 1, characterized in that: The first decoder of the upper self-supervised training model consists of three ConvLSTM modules.

5. The embryonic development prediction classification system based on a double-layer optimization framework according to claim 1, characterized in that: The second decoder of the lower embryo image development prediction classification model first converts the high-dimensional feature map into a one-dimensional feature vector through a Flatten feature flattening operation, and then performs nonlinear transformation and weighted processing on the flattened features through a KAN module.

6. The embryonic development prediction and classification system based on a double-layer optimization framework according to claim 1, characterized in that: The upper layer self-supervised training model is trained using a point-to-point contrast loss function.

7. The embryonic development prediction and classification system based on a double-layer optimization framework according to claim 6, characterized in that: The contrast loss function The expression is: ； ； ； In the formula, Represents an embryo image sequence subsequence The number of samples; express The total number of feature points in ; Representing model predictions The number of samples in the time dimension; and Represent subsequences height and width in spatial dimensions; , both represent indexes, Represents a subsequence Middle The index of the samples, Used to represent the index number of another sample in calculating the point-to-point contrast loss; , both are used to represent indexes, Represents a subsequence Middle The index number of the feature point, Used to represent the index number of the feature point of another sample in calculating the point-to-point contrast loss; Represents a subsequence Middle Embryo samples The real embryonic development feature points at each position; Represents the model for subsequence Middle Embryo samples The predicted embryonic developmental features at each position; Representation calculation and The cosine similarity between ; Represents the temperature parameter.

8. The embryonic development prediction and classification system based on a double-layer optimization framework according to claim 1, characterized in that: The lower embryonic development prediction classification model is optimized using a cross entropy loss function.

9. The embryonic development prediction and classification system based on a double-layer optimization framework according to claim 8, characterized in that: The cross entropy loss function The expression is: ； In the formula, Represents a sequence of embryo images The corresponding label is passed through a one-hot encoder to obtain the encoded label; express The predicted probability obtained by the upper self-supervised training model.

10. The embryo development prediction and classification system based on a double-layer optimization framework according to claim 1, characterized in that: Before using the embryonic development image sequence for training, the size of the embryonic development image sequence is unified and reduced.

Images