Embryo development quality evaluation method and system based on test time adaptive strategy

Abstract

The invention discloses an embryonic development quality evaluation method and system based on a test time adaptive strategy, and the method comprises the following steps: training a model through employing an embryonic image training set with a label, obtaining a source model, and initializing the parameters of a teacher-student model through employing the source model; respectively inputting a group of embryo image sequences to be evaluated into the teacher model and the student model, calculating KL divergence between prediction results output by the teacher model and the student model and Shannon entropy loss of the student model, and optimizing parameters of the student model by minimizing the sum of the KL divergence and the Shannon entropy loss; storing the features of the embryo image sequence and the Shannon entropy loss as a key value pair in a memory bank, comparing the Shannon entropy loss of the student model at the current time point and the Shannon entropy loss of the student model at the previous time point, and dynamically adjusting the learning rate of the student model according to the change of the Shannon entropy loss; and repeating the steps until all the to-be-evaluated embryo image sequences are predicted.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for evaluating embryo development quality based on a test time adaptive strategy. Background Art

[0002] In vitro fertilization (IVF) technology provides hope for many infertile couples to have children. After the fertilized egg develops to the embryo stage, embryo transfer is required, and the quality of the embryo is directly related to its subsequent development and success rate. Therefore, embryo quality assessment plays a vital role in the assisted reproductive process. With the advancement of computer technology and image sequence processing technology, it is particularly important to improve the accuracy of embryo image sequence classification, which directly affects the doctor's subsequent diagnosis and treatment.

[0003] Initially, the classification of embryo image sequences mainly relied on morphological observation, and the quality of embryos was judged by evaluating their appearance, cell number, cell uniformity and other features. However, this method that relies on subjective judgment has limitations, and inexperienced doctors may get wrong evaluation results due to differences in observation angles and subjective judgment criteria. With the rapid development of machine learning technology, machine learning-based methods have been widely used to process a large number of embryo image sequences. By extracting features and building models to classify and evaluate embryo quality, they have been generally recognized by industry insiders. In order to solve the problems of inconsistent data distribution and training data involving patient privacy, passive domain adaptation technology came into being. The goal of passive domain adaptation technology is to enable the model to have better performance in the target domain by learning the commonalities and differences between different domains. In the field of embryo image sequence quality assessment, the application of passive domain adaptation technology is mainly to adjust the model trained in one or more source domains through certain methods so that it can adapt to the data distribution of the target domain. When the source domain data cannot be obtained and only the target domain data is available, this method uses the source model obtained by training the source domain data to test the target domain, thereby protecting the privacy of patients in the quality assessment of embryo image sequences. Although passive domain adaptation technology can solve the problems of inconsistent data distribution and privacy protection, there are some challenges in its practical application. Passive domain adaptation training for embryo sequences is offline and requires the reception of a set of sequence images, which limits the real-time processing capability of the model. In order to be able to process the embryo image sequence received by the model in real time, the test time adaptive TTA technology came into being. TTA supports online testing and can process the image sequence when it is received, without waiting for the entire image sequence to be received.

[0004] However, in the test time adaptation TTA technology, as the model is constantly adjusted and updated during the test phase, the original structure of the model may change, resulting in catastrophic forgetting, which leads to a decrease in the classification evaluation performance of the model. This shows that although passive domain adaptation technology has made progress in some aspects, further improvements are still needed to meet the needs of real-time processing. Summary of the invention

[0005] The present invention proposes an embryo development quality assessment method and system based on a test time adaptive strategy, which solves the problem of catastrophic forgetting that may occur in the test time adaptive technology during online testing.

[0006] To solve the above technical problems, the present invention provides a method for evaluating embryo development quality based on a test time adaptive strategy, comprising the following steps: Step S1: training a model using a labeled embryo image training set to obtain a source model, and using the source model to initialize parameters of a teacher-student model; Step S2: inputting a set of embryo image sequences to be evaluated into the teacher model and the student model respectively, calculating the KL divergence between the prediction results output by the teacher model and the student model and the Shannon entropy loss of the student model, and optimizing the parameters of the student model by minimizing the sum of the KL divergence and the Shannon entropy loss; Step S3: storing the features of the embryo image sequence and the Shannon entropy loss as key-value pairs in a memory bank, comparing the Shannon entropy loss of the student model at the current time point and the previous time point, and dynamically adjusting the learning rate of the student model according to the change of the Shannon entropy loss; Step S4: Repeat steps S2 to S3 until prediction is completed for all embryo image sequences to be evaluated.

[0007] Preferably, the expression for calculating the KL divergence in step S2 is: ; In the formula, is the KL divergence between the teacher model and the student model; T ( x ) is the probability distribution of the embryo image sequence output by the teacher model belonging to each category; S ( x ) is the probability distribution of the embryo image sequence output by the student model belonging to each category.

[0008] Preferably, the expression for calculating the Shannon entropy loss in step S2 is: ; In the formula, is the Shannon entropy loss of the student model; is a random variablex Value probability; is a random variable x Total number of; is the base of the logarithm.

[0009] Preferably, the learning rate of the student model in step S3 is expressed as: ; In the formula, is the learning rate of the student model; is the initial learning rate of the student model; is the base of the logarithmic function; is the decay rate; is the current iteration number; , The time points i and time point i Shannon entropy loss of the student model when -1.

[0010] Preferably, in step S4, after the student model undergoes a set number of self-trainings, the teacher model is updated according to the performance of the student model: ; In the formula, , are the teacher model at time points i and time point i Parameters when -1; is the attenuation factor; For the student model at the time point i Parameters when .

[0011] Preferably, the method for obtaining the embryo image training set in step S1 comprises the following steps: Step S11: collecting a plurality of continuously shot embryo images, adjusting each embryo image to the same size, and forming an embryo image sequence; Step S12: analyzing the embryo image sequence, evaluating the contrast between the embryo and the background in all embryo images, taking the embryo images with contrast less than a set threshold as low-contrast images, and performing contrast enhancement on the low-contrast images; Step S13: Mark the developmental stages of the embryos in all embryo images to obtain an embryo image training set.

[0012] Preferably, the teacher model and the student model output the probability distribution of the embryo image sequence belonging to each category in step S2, comprising the following steps: Step S21: evenly divide each embryo image in the embryo image sequence into a number of image blocks; Step S22: stretching the matrix corresponding to each image block into a one-dimensional vector, and adding a position code and a time code to the one-dimensional vector; Step S23: extracting the features of all image blocks through the self-attention layer and the feed-forward network layer, performing weighted summation on the features of all image blocks, and obtaining the overall features of the embryo image sequence; Step S24: input the overall features into a multi-layer perceptron, and the multi-layer perceptron outputs the probability that the embryo image sequence belongs to each category.

[0013] Preferably, the memory bank follows a first-in-first-out rule.

[0014] Preferably, in step S4, when a new embryo image sequence arrives, the parameters of the teacher-student model are initialized using the source model.

[0015] The present invention also provides an embryo development quality assessment system based on a test time adaptive strategy, which is implemented based on the above-mentioned embryo development quality assessment method based on a test time adaptive strategy, and includes: a source model training module, an online prediction module, a loss value memory module and a model update module; The source model training module: uses a labeled embryo image training set to train an initial model to obtain a source model, and the parameters of the source model are used to initialize the parameters of the teacher-student model; The online prediction module: receives the embryo image sequence to be evaluated as the input of the teacher-student model, and performs prediction using the current model parameters; The loss value memory module is used to store and manage the Shannon entropy loss value of the student model during the model prediction process, and to store the features of each embryo image sequence and the corresponding Shannon entropy loss as a key-value pair in the memory bank; The model updating module adjusts the parameters of the student model according to the prediction results output by the online prediction module and the Shannon entropy loss value, and updates the parameters of the teacher model through the exponential moving average, so that the teacher and student models can learn from each other.

[0016] The benefits of the present invention include at least: 1. The features of the embryo image sequence and the Shannon entropy loss are stored in the memory bank, which effectively preserves historical information, helps the model refer to previous experience in subsequent training, and enhances the stability and generalization ability of the model; 2. Through knowledge distillation, the student model can learn richer embryo image feature representations from the teacher model, thereby significantly improving the classification accuracy of the student model for embryo image sequences; 3. Through the adaptive strategy of dynamically adjusting the learning rate, the student model reduces the learning rate when the classification performance of the embryo image sequence decreases, effectively avoiding overfitting. At the same time, the learning rate is maintained when the model performance improves, promoting faster convergence of the model to more accurately identify and classify the embryo image sequence; 4. Supports online processing of embryo image sequences to be evaluated, and can evaluate embryo image sequences immediately upon receipt without waiting for all data to be received. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention; Figure 2 A flowchart of a teacher model and a student model processing an embryo image sequence according to an embodiment of the present invention; Figure 3 A test time adaptation flow chart of an embodiment of the present invention; Figure 4 is a schematic diagram of the structure of a memory bank according to an embodiment of the present invention; Figure 5 Schematic diagram of the adaptation process of the online test 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] like Figure 1 As shown, an embodiment of the present invention provides a method for evaluating embryo development quality based on a test time adaptive strategy, comprising the following steps: Step S1: Use the labeled embryo image training set to train the model to obtain the source model, and use the source model to initialize the parameters of the teacher-student model.

[0020] Specifically, the embryo image training set used in the embodiment of the present invention comes from multiple reproductive medical centers. Reproductive medical centers usually use time-difference incubators to capture embryo images. This method can continuously record the development of embryos without interfering with the normal development of embryos. The captured images comprehensively cover all stages of embryo development, and each set of images contains multiple sequences of embryo images taken continuously.

[0021] Each embryo image is adjusted to the same size, the contrast between the embryo and the background in all embryo images is evaluated, and the embryo images with contrast less than the set threshold are regarded as low-contrast images, and the contrast of low-contrast images is enhanced. Then the developmental stage of the embryo in all embryo images is marked to obtain the embryo image training set. These preprocessing operations can effectively improve the robustness of the model, so that the model can perform effective reasoning and classification under various test conditions.

[0022] The source model of the embodiment of the present invention is the video vision Transformer model ViViT, which is a neural network model based on the self-attention mechanism. It performs well in the embryonic image sequence classification task and has good scalability, generalization ability and other characteristics. The teacher model in the teacher-student model is a large model, and the student model is a small model, which adopts a similar network structure.

[0023] Step S2: Input a set of embryo image sequences to be evaluated into the teacher model and the student model respectively, calculate the KL divergence between the prediction results output by the teacher model and the student model and the Shannon entropy loss of the student model, and optimize the parameters of the student model by minimizing the sum of the KL divergence and the Shannon entropy loss.

[0024] Specifically, Figure 2 Schematic diagram of the process of outputting the probability distribution of embryo image sequences belonging to various categories for the teacher model and the student model, where CLS represents the classification label, including the following steps: Step S21: Divide each embryo image in the embryo image sequence evenly into An image block of pixels.

[0025] Step S22: stretching the matrix corresponding to each image block into a one-dimensional vector, thereby converting the two-dimensional image data into one-dimensional sequence data, and adding position coding and time coding to the one-dimensional vector.

[0026] Specifically, position coding based on sine and cosine functions is added to each image block to make up for the problem that the Transformer network is insensitive to position information, so that the model can learn the relative position relationship between image blocks, which helps the model understand the spatial structure of the embryo image sequence. After linear embedding and position coding of the image blocks, they are sent to the SpatialTransformer encoder to add time coding to each image.

[0027] Step S23: extract the features of all image blocks through the self-attention layer and the feedforward network layer, perform weighted summation on the features of all image blocks, and obtain the overall features of the embryo image sequence.

[0028] Specifically, the image block sequence with position encoding and time encoding is fed into the Temporal Transformer encoder, which includes multiple self-attention layers and feedforward networks to capture the dependencies between image blocks and extract the key features of image blocks. The multiple heads of the multi-head self-attention mechanism process the image block sequence in parallel as independent attention modules, calculate the attention score, and dynamically adjust the contribution of each image block to the final feature representation.

[0029] The feedforward neural network performs further nonlinear transformation on the features output by the self-attention layer, extracts information from the image block features, integrates features such as cell morphology and density, and generates more representative features.

[0030] After being processed by multiple Transformer layers, the features of the image blocks are fully extracted and transformed, and the features of all image blocks are fused by weighted averaging or splicing, integrating the local image block features into the overall features of the embryo image sequence.

[0031] Step S24: The overall features are input into a multi-layer perceptron, and the multi-layer perceptron outputs the probability that the embryo image sequence belongs to each category.

[0032] Specifically, the overall features are input into the multi-layer perceptron MLP Head, which contains multiple fully connected layers. The first fully connected layer performs linear transformation and nonlinear activation on the input features, and the second fully connected layer continues to perform similar operations on the output of the first fully connected layer and maps it to the final output dimension. Finally, the probability corresponding to each category is used to determine the category to which the embryo image sequence belongs, completing the classification task of the embryo image sequence.

[0033] like Figure 3 As shown in the figure, for a set of embryo image sequences, they are input into the teacher model for testing. The teacher model outputs the probability distribution of each embryo image sequence belonging to each category. These probability distributions contain the teacher model's high-level feature understanding and classification knowledge of the embryo image sequence. The student model also receives this set of embryo image sequences as input, and then outputs its own predicted probability distribution. The feature extraction and classification capabilities learned by the teacher model are imparted to the student model through knowledge distillation. Specifically, the KL divergence is used to measure the difference between the prediction results of the student model and the teacher model, and then knowledge distillation is achieved by minimizing the KL divergence. The calculation formula of KL divergence is: ; In the formula, is the KL divergence between the teacher model and the student model; T ( x ) is the probability distribution of the embryo image sequence output by the teacher model belonging to each category; S (x ) is the probability distribution of the embryo image sequence output by the student model belonging to each category.

[0034] The embodiment of the present invention uses Shannon entropy as the loss function of the student model. Shannon entropy can be used as a measure of the complexity of a system. If the system is more complex and there are more types of different situations, the information entropy of the system is greater. The calculation formula of Shannon entropy is: ; In the formula, is the Shannon entropy loss of the student model; is a random variable x Value probability; is a random variable x Total number of; is the base of the logarithm.

[0035] When the uncertainty of the model prediction results is greater, that is, the more uniform the probability distribution is, the greater the value of Shannon entropy; conversely, when the model's predicted probability for a certain category approaches 1, and the probability of other categories approaches 0, Shannon entropy approaches 0. For example, when embryos are divided into two categories of high quality and low quality, if the model predicts that the probability of an embryo being high quality is 0.9 and the probability of being low quality is 0.1, then the Shannon entropy is relatively low; if the model predicts that the probability of both high quality and low quality is 0.5, then the Shannon entropy is higher, indicating that the uncertainty of the model prediction is greater. Therefore, choosing Shannon entropy as the loss function can directly measure the difference between the model prediction results and the ideal deterministic classification. By minimizing Shannon entropy, the uncertainty of model prediction can be reduced, thereby improving the accuracy of model classification.

[0036] Step S3: Store the features of the embryo image sequence and the Shannon entropy loss as key-value pairs in the memory bank, compare the Shannon entropy loss of the student model at the current time point and the previous time point, and dynamically adjust the learning rate of the student model according to the change of the Shannon entropy loss.

[0037] Specifically, in order to avoid catastrophic forgetting of the student model during the updating process, the memory bank is combined with the adaptive learning rate of the student model.

[0038] The memory bank is usually a repository that stores a large amount of data features or model intermediate results. In the context of embryo classification, it can store the feature representation of the embryo image sequence processed before, the prediction results of the model for these image sequences, or other relevant information. The memory bank of the embodiment of the present invention stores the features of the samples and their corresponding Shannon entropy loss values, and the memory bank is updated by continuously storing new sample features and Shannon entropy loss values. The memory bank follows the first-in-first-out rule, fills up the memory bank first, and then each time node is a sample feature out, a sample feature in, and the test phase adaptation is not started until the memory bank is filled.

[0039] First, the embryo image sequence is input into the student model, and the feature representation of the image sequence is obtained through the feature extraction layer of the student model. These feature representations are stored in the memory bank. At the same time, the model obtains the predicted category probability distribution through the classification head, and then calculates the Shannon entropy loss value of the current prediction result. Figure 4 As shown, in the embodiment of the present invention, the features of the image sequence are used as keys and the Shannon entropy loss values ​​are used as values. The image sequence features and the corresponding Shannon entropy loss values ​​are stored in the memory in the form of key-value pairs to facilitate subsequent retrieval and use.

[0040] Compare the Shannon entropy loss value of the student model at the current time point with the Shannon entropy loss value at the previous time point. If the Shannon entropy loss value increases, it means that the performance of the student model has decreased, and the learning rate of the student model is updated; if the Shannon entropy loss value decreases, it means that the performance of the current student model is better, and the current learning rate continues to be used. The expression for updating the learning rate is: ; In the formula, is the learning rate of the student model; is the initial learning rate of the student model; is the base of the logarithmic function; is the decay rate; is the current iteration number; , The time points i and time point i Shannon entropy loss of the student model when -1.

[0041] After minimizing the sum of KL divergence and Shannon entropy loss, the parameters of the student model are updated through back propagation. During the back propagation process, the weight matrix, bias and other parameters in the Transformer layer of the student model are adjusted according to the gradient information of the loss and the current learning rate. In this way, the model gradually optimizes its own parameters while considering the current sample information, combining the classification of similar samples in the memory bank and the dynamically changing learning rate, thereby improving the model's ability to classify embryonic image sequences.

[0042] After the student model has undergone a certain period of self-training, the teacher model is updated using an exponential moving average: ; In the formula, , are the teacher model at time points i and time point i Parameters when -1; is the attenuation factor, which is used to balance the weights of the teacher model and the student model. It is 0.99, which can stabilize the model training and maintain the smoothness of the teacher model; For the student model at the time point i Parameters when .

[0043] Through the process of mutual learning between the teacher and student models, we can fully utilize the prior knowledge of the teacher model and the learning potential of the student model, improve the overall performance and generalization ability of the model in the embryo classification problem, reduce dependence on a large amount of labeled data, and gradually optimize the model's understanding of the characteristics of the embryo image sequence and classification decision-making ability.

[0044] Step S4: Repeat steps S2 to S3 until prediction is completed for all embryo image sequences to be evaluated.

[0045] like Figure 5 The figure shows the adaptive process of online testing implemented in the embodiment of the present invention. The student model after the source model is initialized is recorded as f 0, select the embryo image sequence 1 to be evaluated as f 0 input, the source model is trained using the embryo image sequence 1 to be evaluated, and after completing one training cycle epoch, the parameters of the student model are updated to obtain the updated model f s,1 Then, using f s,1 The embryo image sequence 1 to be evaluated is predicted and the corresponding prediction results are output. Next, the next embryo image sequence 2 to be evaluated is introduced and input into the model obtained in the previous step. f s,1 In the process, the model adjusts and updates parameters again based on the new input to obtain a further updated model f s,2 Repeat the above operation process, that is, input the current embryo image sequence to be evaluated into the model obtained by the last update f s,i-1 In the example, the model is updated to be after one epoch of training. f s,i, and predict the embryo image sequence to be evaluated. This cycle repeats until all embryo image sequences to be evaluated are predicted.

[0046] When a new batch of embryo image sequences arrives, the parameters of the student model are initialized to f 0 parameter setting to adapt to the new data distribution characteristics. During the entire online test process, the model is iteratively updated and adapted by continuously using the newly input test samples to achieve real-time classification of the continuously input embryo image sequence. Moreover, in this process, each classification result can be collected, sorted and analyzed, so as to achieve effective management and utilization of the classification results. The memory library effectively guarantees the stability and reliability of the model in long-term continuous operation and in the face of different batches of data input, so that it can always maintain relatively good classification performance and adaptability in a constantly changing data environment.

[0047] The method of the present invention adopts a mutual learning strategy between teachers and students. In the back propagation of the student model, the memory bank is combined with the learning rate of the model, and the learning rate is dynamically adjusted to avoid catastrophic forgetting of the model, thereby improving the stability and generalization of the model. By using a pre-trained model for testing, the number of accesses to the source domain data is greatly reduced, protecting the privacy of patients.

[0048] An embodiment of the present invention also provides an embryo development quality assessment system based on a test time adaptive strategy, which is implemented based on the above-mentioned embryo development quality assessment method based on a test time adaptive strategy, and includes: a source model training module, an online prediction module, a loss value memory module and a model update module.

[0049] The source model training module uses the labeled embryo image training set to train the initial model to obtain the source model, and the parameters of the source model are used to initialize the parameters of the teacher-student model.

[0050] The online prediction module is used to receive the embryo image sequence to be evaluated as the input of the teacher-student model and make predictions using the current model parameters.

[0051] The loss value memory module is used to store and manage the Shannon entropy loss value of the student model during the model prediction process, and saves the features of each embryo image sequence and the corresponding Shannon entropy loss as a key-value pair in the memory bank.

[0052] The model update module adjusts the parameters of the student model according to the prediction results output by the online prediction module and the Shannon entropy loss value, and updates the parameters of the teacher model through the exponential moving average, so that the teacher and student models can learn from each other.

[0053] 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.

[0054] 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. A method for evaluating embryo development quality based on a test time adaptive strategy, characterized in that: The following steps are involved: Step S1: training a model using a labeled embryo image training set to obtain a source model, and using the source model to initialize parameters of a teacher-student model; Step S2: inputting a set of embryo image sequences to be evaluated into the teacher model and the student model respectively, calculating the KL divergence between the prediction results output by the teacher model and the student model and the Shannon entropy loss of the student model, and optimizing the parameters of the student model by minimizing the sum of the KL divergence and the Shannon entropy loss; Step S3: storing the features of the embryo image sequence and the Shannon entropy loss as key-value pairs in a memory bank, comparing the Shannon entropy loss of the student model at the current time point and the previous time point, and dynamically adjusting the learning rate of the student model according to the change of the Shannon entropy loss; Step S4: Repeat steps S2 to S3 until prediction is completed for all embryo image sequences to be evaluated.

2. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: The expression for calculating the KL divergence in step S2 is: ； In the formula, is the KL divergence between the teacher model and the student model; T ( x ) is the probability distribution of the embryo image sequence output by the teacher model belonging to each category; S ( x ) is the probability distribution of the embryo image sequence output by the student model belonging to each category.

3. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: The expression for calculating the Shannon entropy loss in step S2 is: ； In the formula, is the Shannon entropy loss of the student model; is a random variable x Value The probability of is a random variable x Total number of; is the base of the logarithm.

4. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: The expression of the learning rate of the student model in step S3 is: ； In the formula, is the learning rate of the student model; is the initial learning rate of the student model; is the base of the logarithmic function; is the decay rate; is the current iteration number; , The time points i and time point i Shannon entropy loss of the student model when -1.

5. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: In step S4, after the student model has undergone a set number of self-training, the teacher model is updated according to the performance of the student model: ； In the formula, , are the teacher model at time points i and time point i Parameters when -1; is the attenuation factor; For the student model at the time point i Parameters when .

6. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: The method for obtaining the embryo image training set in step S1 comprises the following steps: Step S11: collecting a plurality of continuously shot embryo images, adjusting each embryo image to the same size, and forming an embryo image sequence; Step S12: analyzing the embryo image sequence, evaluating the contrast between the embryo and the background in all embryo images, taking the embryo images with contrast less than a set threshold as low-contrast images, and performing contrast enhancement on the low-contrast images; Step S13: Mark the developmental stages of the embryos in all embryo images to obtain an embryo image training set.

7. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: The teacher model and the student model output the probability distribution of the embryo image sequence belonging to each category in step S2, including the following steps: Step S21: evenly divide each embryo image in the embryo image sequence into a number of image blocks; Step S22: stretching the matrix corresponding to each image block into a one-dimensional vector, and adding a position code and a time code to the one-dimensional vector; Step S23: extracting the features of all image blocks through the self-attention layer and the feed-forward network layer, performing weighted summation on the features of all image blocks, and obtaining the overall features of the embryo image sequence; Step S24: input the overall features into a multi-layer perceptron, and the multi-layer perceptron outputs the probability that the embryo image sequence belongs to each category.

8. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: The memory bank follows a first-in-first-out rule.

9. The method for evaluating embryo development quality based on a test time adaptive strategy according to claim 1, characterized in that: In step S4, when a new embryo image sequence arrives, the parameters of the teacher-student model are initialized using the source model.

10. An embryo development quality assessment system based on a test time adaptive strategy, implemented based on an embryo development quality assessment method based on a test time adaptive strategy as claimed in any one of claims 1 to 9, characterized in that: include: Source model training module, online prediction module, loss value memory module and model update module; The source model training module: uses a labeled embryo image training set to train an initial model to obtain a source model, and the parameters of the source model are used to initialize the parameters of the teacher-student model; The online prediction module: receives the embryo image sequence to be evaluated as the input of the teacher-student model, and performs prediction using the current model parameters; The loss value memory module is used to store and manage the Shannon entropy loss value of the student model during the model prediction process, and to store the features of each embryo image sequence and the corresponding Shannon entropy loss as a key-value pair in the memory bank; The model updating module adjusts the parameters of the student model according to the prediction results output by the online prediction module and the Shannon entropy loss value, and updates the parameters of the teacher model through the exponential moving average, so that the teacher and student models can learn from each other.

Images