Artificial Intelligence Evaluation Method for Embryo Development Based on Progressive Active-Passive Domain Adaptation

Abstract

The present invention provides an artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation. S1: Train a source domain model using a source domain dataset, where the source domain dataset is a sequence of embryonic images with high-quality and low-quality labels. S2: Use the trained source domain model to generate pseudo-labels for the embryonic image sequence dataset in the target domain, and initialize the target domain model using the source domain model and the pseudo-labels of the target domain data. S3: Use an active learning algorithm to select samples of embryonic image sequences according to the level of uncertainty, manually annotate the samples with high uncertainty, and update the target domain dataset. S4: Use the updated target domain dataset to train and optimize the target domain model. S5: Repeat steps S3 and S4 to obtain an optimized target domain model, and use the target domain model to perform artificial intelligence evaluation on embryonic images. It can better adapt to the feature distribution of the target domain and improve the generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of embryo development quality assessment, and particularly relates to an artificial intelligence assessment method for embryo development based on progressive active source-free domain adaptation. Background Art

[0002] With the development of assisted reproductive technology (ART), in vitro fertilization (IVF) technology has become a conventional means for treating infertility. However, the accuracy of embryo quality assessment is crucial for selecting the embryo most likely to successfully implant, directly affecting the success rate of IVF. In the development of computer vision and machine learning technologies, significant progress has been made in image sequence recognition and classification technologies. Especially in the field of medical image sequence analysis, the application of these technologies is crucial for improving the accuracy and efficiency of diagnosis. In this special field of embryo development assessment, it has become an important issue to judge the quality of embryo development by analyzing embryo image sequences.

[0003] Traditional embryo image sequence classification methods often rely on supervised learning, which requires a large amount of labeled data in the target domain. However, in practical applications, obtaining these labeled data is often expensive and time-consuming, and traditional image sequence analysis methods rely on rule-based feature extraction and classification algorithms, but these methods often cannot handle the complex and variable features in embryo image sequences. Especially in the case of cross-device or cross-experimental environments, there are often significant differences in the quality and manifestation forms of embryo image sequences, resulting in a decline in the performance of traditional models. To ensure that in different devices, especially in the scenarios of privacy protection and cross-time-zone incubator devices, how to train an efficient and accurate target domain model for embryo quality assessment is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] The present invention proposes an artificial intelligence assessment method for embryo development based on progressive active source-free domain adaptation to solve the technical problem of low performance of artificial intelligence assessment of embryo development in the scenario of cross-time-zone incubator devices.

[0005] To solve the above technical problem, the present invention provides an artificial intelligence assessment method for embryo development based on progressive active source-free domain adaptation, including the following steps:

[0006] S1: Training a source domain model using a source domain dataset, where the source domain dataset is an embryo image sequence with high-quality and low-quality labels;

[0007] S2: Generating pseudo-labels for the embryo image sequence dataset in the target domain using the trained source domain model, and initializing the target domain model using the source domain model and the pseudo-labels of the target domain data;

[0008] S3: Use the active learning algorithm to select embryo image sequence samples according to the level of uncertainty. For samples with high uncertainty, perform manual annotation and update the target domain dataset;

[0009] S4: Use the updated target domain dataset to train and optimize the target domain model;

[0010] S5: Repeat steps S3 and S4 to obtain the optimized target domain model, and use the target domain model to perform artificial intelligence evaluation on embryo images.

[0011] Preferably, in step S3, the active learning algorithm calculates the uncertainty of embryo image sequence samples by the entropy method, selects a specified proportion of samples with the highest entropy values for manual annotation, and superimposes the annotated data with samples with low uncertainty to form a new embryo target domain dataset.

[0012] Preferably, after extracting features from the target domain data through the initialized target domain model in step S2, first perform binary classification on the target domain data through the clustering module, divide the embryo development data into two categories: high development quality and low development quality, and generate pseudo-labels.

[0013] Preferably, the source domain model and / or the target domain model in step S2 is a feature extraction model using Video Mamba as the backbone network.

[0014] Preferably, the entropy value of the sample in step S2 is calculated by the following formula:

[0015] ;

[0016] where ( ) represents the probability that the sample belongs to the category ; represents the parameter of Tsallis entropy; represents the total number of categories.

[0017] Preferably, in step S4, use the updated embryo target domain dataset to retrain the target domain model to iterate the new embryo target domain dataset, and judge the training result through the KL divergence and the binary cross-entropy loss function of the target domain.

[0018] Preferably, the expression of the KL divergence is:

[0019] ;

[0020] where is the total number of categories, is the index of the category, =1 indicates poor embryo development, =2 indicates good embryo development, is the true probability distribution of the category, and is the probability distribution predicted by the model =1 or =2.

[0021] Preferably, the expression of the binary cross-entropy loss function for the target domain is:

[0022] ;

[0023] In the formula, represents the total number of samples in the target domain; is the pseudo-label of the target domain sample; represents the probability that the model predicts the sample as a poor embryo.

[0024] Preferably, the weights of the target domain model are penalized by L2 regularization, and the expression of the L2 regularization is:

[0025] ;

[0026] In the formula, is the binary cross-entropy loss; is the L2 norm of all weight parameters; is the hyperparameter of the regularization strength; represents the th weight in the model.

[0027] Preferably, when the target domain model performs binary classification in step S2, the binary cross-entropy loss function of the target domain is used to measure the difference between the probability distribution predicted by the model and the actual label in the binary classification task.

[0028] To solve the problem that the target domain model trained by active passive domain adaptation has poor performance in matching the data features of the target domain, the present invention adopts the passive domain adaptation technology (SFDA). SFDA allows the model to be trained on the source domain. After obtaining the source domain model, the source domain model is used to generate pseudo-labels for the target domain data, and the pseudo-labels are used to retrain the source domain model on the target domain data, thereby obtaining the target domain model. This method is particularly suitable for fields where it is difficult to obtain labeled data. The core of the traditional SFDA technology is to train a source domain model using source domain data, and then generate pseudo-labels on the target domain with the trained source domain model, and then use the auxiliary information of the pseudo-labels on the unlabeled target domain to achieve domain adaptation. However, for scenarios such as embryo quality assessment that require precise classification but are limited by data annotation costs and privacy protection, the traditional SFDA cannot effectively adapt the feature distribution of the source domain to the target domain completely. On the other hand, due to the lack of source domain data and its corresponding target supervision information in the SFDA method, the trained target domain model cannot fully adapt to the target domain data, and there are still certain limitations in the improvement of the adaptation performance between the source domain and the target domain in the current state-of-the-art SFDA technology.

[0029] To further improve the adaptability of the model on cross-time-lapse incubator equipment, the present invention adopts the active passive domain adaptation technology (ASFDA). Without using source domain data, some embryo image sequence samples with high uncertainty are selected through the use of an active learning algorithm for supervised learning. Among these selected embryo image sequence samples with high uncertainty, the features of the samples only contain the embryo features in the source domain data. Due to the lack of embryo data features in the target domain, when these embryo samples with high uncertainty containing source domain data features are trained with the remaining samples at one time and then used on the target domain, it will lead to poor performance in the matching performance of the trained target domain model with the target domain data features. The present invention overcomes the above defects by designing a progressive passive domain adaptation technology.

[0030] The beneficial effects of the present invention at least include:

[0031] 1) Reduce the annotation workload: In practical applications, obtaining labeled data is costly. The progressive method can more effectively utilize limited annotation resources by progressively selecting the most valuable samples for annotation. Spreading the resource requirements of these most valuable samples instead of processing all valuable samples at one time greatly reduces the workload of one-time annotation.

[0032] 2) Improving the model generalization ability: The progressive method reduces the risk of the model overfitting to the initial samples by continuously introducing new uncertain samples. In each iteration, based on the prediction results and uncertainty estimates of the model on the latest data samples, it is gradually optimized and progressively improved in performance, rather than relying on the initial uncertainty estimates, so as to better adapt to the feature distribution of the target domain and improve the generalization ability.

[0033] 3) Protecting privacy and data security: Since the source domain data is not required, this method has advantages in terms of privacy protection and data security.

[0034] Through the evaluation method of the present invention, a new technical means is provided for the IVF field, significantly improving the accuracy and efficiency of embryo quality evaluation, and further improving the success rate of IVF. Description of the Drawings

[0035] Figure 1 It is a schematic flowchart of the method of the embodiment of the present invention;

[0036] Figure 2 It is a schematic diagram of the comparison between good and bad embryos of the embodiment of the present invention;

[0037] Figure 3 It is a schematic diagram of initializing the source domain model with unlabeled source domain data of the embodiment of the present invention;

[0038] Figure 4 It is a schematic diagram of active learning for sample selection of the embodiment of the present invention;

[0039] Figure 5 It is a schematic diagram of the final model classification result of the embodiment of the present invention. Detailed Embodiments

[0040] Next, in combination with the drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0041] This embodiment aggregates one hundred thousand embryo image sequences from multiple reproductive medicine centers, and these image sequences are all from time-lapse incubators. This huge dataset provides solid data support, but during the process of constructing the dataset, we faced significant challenges. Due to the limited resources of professional embryologists, they need to spend a large amount of time annotating these image sequences. This is not only a time-consuming and labor-intensive task, but also means that they are unable to engage in other more valuable professional activities during this period. Therefore, there is an urgent need for a more efficient solution to reduce the image sequence annotation work. Exploring how to use limited annotated data and a large number of unannotated image sequences for effective model training has become a new challenge faced by this invention.

[0042] Figure 2 Shows a comparison chart of the quality of embryo development during the development process. Through Figure 2 It can be seen from the left figure that high-quality embryos usually have regular morphologies, consistent cell sizes, and few cell fragments. Figure 2 Some fragments can be seen in the right figure, while Figure 2 there are fewer fragments in the embryos in the left figure, which will be a key feature for using deep learning algorithms to judge the quality of embryo development.

[0043] For early embryos, the number of blastomeres and the development rate are the most important indicators for embryo grading. And the training of deep learning models for embryo image sequences requires a large amount of accurately annotated data, which greatly increases the workload of embryologists. At the same time, the annotation of a large amount of data also increases the risk of leakage of the privacy information of subjects in the source domain.

[0044] Therefore, as Figure 1 shown, the embodiment of this invention provides an artificial intelligence evaluation method for embryo development based on progressive active source-free domain adaptation.

[0045] 1. Data collection and preprocessing stage

[0046] 1) Preparation of embryo image sequences

[0047] The embryo image sequence data used in this embodiment is sourced from multiple reproductive medical institutions, which have accumulated profound expertise in assisted reproductive technology and are equipped with state-of-the-art equipment. The embryo image sequences collected in the experiment utilize time-lapse photography technology and are taken in an incubator. This technology can continuously track the development process of embryos without interfering with their natural growth. The collected image sequence dataset widely covers all stages of embryo development from the single-cell stage to the blastocyst stage and includes image sequences of empty culture dishes. These image sequences are evaluated by multiple embryology experts and classified and stored in two different folders according to the quality of embryo development. All source domain image sequences are incorporated into the training set. After training the source domain model, only the source domain model needs to be adapted to the target domain data to obtain the target domain model, which can effectively protect the privacy information of source domain subjects. The leave-one-out method is used, with 80% of the data for the training set and 20% for the test set. The training set is used during the training process, and finally, the test set is used for evaluation. This fully utilizes all the data for learning during the training process and simultaneously evaluates the performance of the model through the test set. This data partitioning method helps improve the performance of the model and its generalization ability for new data.

[0048] 2) Preprocessing of embryo image sequences

[0049] In the field of embryo image sequence development quality assessment, the color information of embryo image sequences is usually not the core of the analysis, and grayscale image sequences are sufficient to provide the key visual information required to judge the quality of embryo development. To simplify the processing flow and make full use of the advantages of grayscale image sequences, the cv2.cvtColor() function in the OpenCV library is used to convert the embryo image sequences into grayscale format. This conversion process not only removes the interference of color but also retains the texture and contour information of the embryo image sequences, which is crucial for judging the quality of embryo development. Further, to meet the input requirements of the neural network model and optimize the efficiency of data processing, these grayscale image sequences are specially processed. Specifically, the embryo image sequences are first adjusted to the required size for network input using the cv.resize() function in OpenCV. Then, the embryo image sequences are normalized, and the pixel values of the resized embryo image sequences are scaled to the range of 0 to 1. Data augmentation techniques are also applied to increase data diversity and improve the generalization ability of the model.

[0050] 2. Model training stage

[0051] 1) Construction of the embryo source domain model

[0052] The embryo image sequence data of this embodiment is captured in an incubator through time-lapse photography technology. The image sequences of the same embryo at all time periods are integrated together to be equivalent to a video. The characteristics of its embryo image sequence include spatial characteristics and time characteristics. It is difficult for the 2D convolutional layer used in traditional network models to effectively extract the 3D characteristics of the embryo image sequence. And although ordinary 3D convolutional networks have shown powerful performance in fields such as video understanding and medical image sequence analysis, in practical applications, problems such as overfitting, gradient problems, interpretability, data dependence, computational resource consumption, and parameter and storage resource consumption still need to be considered. To solve this problem, Video Mamba is adopted as the backbone network in this embodiment. Video Mamba is a video understanding architecture based on the State Space Model (SSM), and it has shown efficient and powerful performance in the field of video understanding. The 3D convolutional layer therein can process data with a time dimension. By applying 3D convolutional operations to the embryo image sequences at each stage, time and space information can be fused, so as to extract the action characteristics of the embryo image subsequence in time and space. Furthermore, through the source domain data of the embryo image sequence, a source domain model is trained using the Video Mamba network model, as Figure 3 shown.

[0053] 2) Construction of the embryo target domain model

[0054] After initializing the source domain model, pseudo-labels are generated on the embryo target domain dataset through the source domain model, and the target domain model is initialized using the source domain model and the pseudo-labels of the target domain data. Subsequently, an active learning algorithm is used to select embryo image sequence samples according to uncertainty through the entropy method. Although the existing ASFDA algorithm breaks through the performance bottleneck by actively exploring a small number of actively labeled target samples, there are still challenges in the alignment of embryo image sequence sample distributions, the handling of class and quantity imbalances, and the exploration of difficult samples, especially the insufficient exploration of difficult embryo samples and mildly difficult embryo samples. For difficult embryo samples, their quantity is relatively small compared to easy samples. This imbalance may cause the model to overly focus on simple embryo samples during training, while ignoring those truly challenging difficult embryo samples, thus affecting the convergence and accuracy of the model. And for mildly difficult embryo samples, the learning difficulty of these samples is between non-difficult embryo samples and difficult embryo samples and cannot be fully explored either. At the same time, in order to avoid poor performance in the matching of the trained target domain model with the embryo characteristics in the target domain data. The embodiment of the present invention introduces a method for progressively selecting samples, as Figure 4As shown, after extracting features from the target domain data through the initialization model, the target domain data is first classified by Mean Shift clustering, and the embryonic development data is divided into two categories: high development quality and low development quality. By analyzing the attributes and distributions of the samples in each cluster, it is possible to determine which features are related to embryonic development quality and classify accordingly. Subsequently, the Tsallis entropy of the samples is evaluated to determine the uncertainty of the samples. The calculation formula is as follows:

[0055] (1)

[0056] where ( ) is the probability that the sample belongs to the category , is the parameter of the Tsallis entropy, called the topological parameter, which adjusts the calculation method of the entropy. The classical Shannon entropy is the limit when →1, is the total number of categories.

[0057] After calculating the Tsallis entropy of all unlabeled samples, a threshold can be set to select embryo samples with high uncertainty whose entropy values exceed the threshold and the number is α , and select the top α / 10 embryo samples with the highest entropy values for manual annotation. Subsequently, update the training set, stack these new labeled data with the remaining samples with low uncertainty and retrain the model. In the initial stage of training, the number of iterations for each epoch is set to 10. After obtaining the latest optimized model after this epoch is executed, use this latest optimized model to determine the uncertainty of the embryo image sequence samples using the entropy method. Select the top α / 10 of the embryo image sequences with high uncertainty for separate annotation, and stack them in the previously labeled dataset. Subsequently, use the most optimized model trained through epoch = 10 last time to train again, and iterate in this way. Finally, observe the performance of the model during training, including the decreasing trend of the loss function, and terminate the iteration. By iterating in this way multiple times, the performance of the model is gradually improved.

[0058] In this embodiment, for the judgment of the final model effect, a method of calculating the KL divergence and the binary cross-entropy loss function of the target domain is adopted. Introducing the KL divergence is to measure the gap between the predicted probability distribution and the true label distribution, and evaluate the difference between probability distributions. The calculation formula is as follows:

[0059] (2)

[0060] where, is the total number of categories, is the index of the category, = 1 indicates poor embryo development, = 2 indicates good embryo development, is the category of the true probability distribution, is the model prediction = 1 or = 2 of the probability distribution.

[0061] 3) Selection of the loss function

[0062] In this embodiment, for the binary classification problem of judging the quality of the embryo development image sequence, the binary cross-entropy loss function (Binary Cross-Entropy Loss) of the target domain is adopted, which is used to measure the difference between the model prediction probability distribution and the actual label in the binary classification task. The calculation formula is as follows:

[0063] (3)

[0064] Among them, yi is the actual label (1 or 2) of the i th sample, representing poor or good development quality respectively, is the probability that the model predicts the sample as the positive class (excellent embryo), = 1 - is the probability that the model predicts the sample as the negative class (poor embryo), is the total number of samples.

[0065] For each embryo sample xi , the model will predict the probability that it belongs to category 2 (good quality) or category 1 (poor quality). Then, the binary cross-entropy loss will calculate the error between this prediction and the actual label yi . If , indicating excellent embryo development quality, the loss is calculated as . If , indicating poor embryo development quality, the loss is calculated as ( ).

[0066] In the initial training, the initial model is only trained on the embryo target domain data, and the binary cross-entropy loss of the target domain samples is calculated. As the training progresses and the uncertain samples are selected, at this time, the calculation of the loss function depends not only on the true labels of the target domain but also on the pseudo-labels or the generated adversarial samples of the target domain samples. In the target domain samples, the binary cross-entropy loss is calculated in the following way:

[0067] (4)

[0068] Among them, is the pseudo-label of the target domain sample, is the binary cross-entropy loss value of the target domain, is the total number of samples in the target domain.

[0069] To avoid overfitting of the model or loss of generalization ability on the target domain. In this embodiment, L2 regularization (L2 Regularization) is introduced to penalize the weights of the model and avoid the model from being too complex. Its calculation formula is as follows:

[0070] (5)

[0071] Among them, is the binary cross-entropy loss, is the L2 norm of all weight parameters, is the hyperparameter of the regularization strength, which determines the penalty degree of L2 regularization, represents the th weight in the model.

[0072] 3) Model training settings

[0073] The embodiments of the present invention adopt a progressive active-passive domain adaptation method for the task of judging the quality of embryo development image sequences. In the experiment, the Adam+SGD optimizer is used for parameter optimization. The Adam (Adaptive Moment Estimation) optimizer adaptively adjusts the learning rate of each parameter and uses the first moment (mean) and second moment (variance) of the gradient for optimization. Adam usually shows a faster convergence speed in the early stage of training embryo image sequences and is particularly suitable for large-scale embryo image sequence data and complex network structures. The SGD (Stochastic Gradient Descent) optimizer is a more basic optimization method that calculates the mean of each gradient and updates the parameters. The learning rate of SGD is usually small, which can effectively avoid overfitting when selecting high-uncertainty embryo image sequence samples and can finely adjust the model parameters in the later stage of convergence. The method of combining these two optimizers gives full play to the advantages of the Adam optimizer, converges quickly in the initial stage of training, and then gradually switches to SGD to further fine-tune the model, avoid overfitting and find the global optimal solution. For the switching strategy of the two optimizers during the training process, a switching epoch number is set according to the stability of the binary cross-entropy loss function. When switching to SGD, to avoid "jumping" updates, the learning rate needs to be reduced to 0.0001, and the learning rate scheduler can be used to further reduce the learning rate. The Adam optimizer incorporates weight decay and L2 regularization, effectively alleviating the problems of slow convergence and parameter overfitting in network training. The parameters are iteratively updated by the SGD method to minimize the loss function, optimizing the training process of the embryo target domain model. This optimization strategy can train a network model with excellent performance and is particularly suitable for the embryo image sequence classification task. The model can accurately distinguish well-developed embryo image sequences from poorly developed embryo image sequences, providing accurate data for further analysis and research. The model of the present invention is implemented using the Pytorch framework.

[0074] In the embryo sample selection process of the present invention, a progressive active learning algorithm is used to select uncertain samples, which improves the convergence speed and accuracy compared with the traditional progressive sample selection.

[0075] In addition, to further verify the effectiveness of the method adopted by the present invention, the impacts of four different active learning algorithms on the experimental results are compared, and the capabilities of three currently popular models in processing embryo development image sequences are compared: VideoMamba, CNN3D, and CNN+Transformer. The results are shown in Table 1.

[0076] Table 1

[0077]

[0078] Table 1 shows the performance comparison of four different active learning algorithms in the task of embryo image sequence recognition. From the perspective of accuracy, the algorithm of progressive active learning for selecting uncertain samples performs the best. It achieves an accuracy of 89.6% in the evaluation of the quality of embryo development image sequences, indicating that this algorithm has high precision in recognizing embryo image sequences. The convergence rate of this algorithm is described as exponential, meaning that as the algorithm iterates, its performance improves very quickly.

[0079] In contrast, the accuracy of the progressive sample selection algorithm is 83.1%. Although it is lower than the algorithm for selecting uncertain samples, it still shows good recognition ability. The convergence rate of this algorithm is linear, which may mean that the speed of its performance improvement is relatively stable, but it may not be as fast as the algorithm with exponential convergence.

[0080] The accuracies of the CCRL and QUIRE algorithms are similar, 82.4% and 82.5% respectively, and their performances in the recognition task are relatively close. The convergence rates of these two algorithms are not as efficient as the algorithm with exponential convergence. In terms of annotation cost, both CCRL and QUIRE have achieved a "substantial reduction", indicating that using these two algorithms can reduce the manual annotation workload while maintaining a high recognition accuracy.

[0081] Through the above comparison, the progressive active learning algorithm has obvious advantages in selecting uncertain samples, especially in terms of convergence rate and accuracy in the big data environment. Compared with the traditional progressive sample selection method and some other active learning methods, the progressive active learning algorithm can make more effective use of annotation resources, improve learning efficiency and model performance.

[0082] Table 2 shows the recognition accuracies obtained by different models in the evaluation of the quality of embryo image sequences. We compared the performances of VideoMamba, 3D convolutional network (CNN 3D), and the model combining CNN and Transformer in processing embryo development image sequences, focusing on two key metrics, accuracy and F1-score, to evaluate the effectiveness of each model in this field.

[0083] Table 2

[0084]

[0085] First, the Video Mamba model showed the best performance in terms of accuracy, reaching 93.2%. This result is significantly higher than 91.8% of CNN 3D and 92.6% of CNN and Transformer. This difference indicates that Video Mamba has an obvious advantage in the ability to recognize and classify embryo development image sequences, being able to capture important features in the image sequence more accurately. This excellent performance may be related to the network structure of Video Mamba and its way of processing time series data, enabling it to better analyze dynamic biological processes.

[0086] In terms of the F1 score, Video Mamba also achieved an excellent result of 92.6%, leading 91.5% of CNN and Transformer and 90.1% of CNN 3D. The F1 score comprehensively considers the accuracy and recall rate of the model, reflecting the model's ability to handle sample imbalance. The high F1 score of Video Mamba indicates that while ensuring high accuracy, the model can still effectively identify positive samples, reducing the situation of missed reports, and thus performing more robustly in the classification task of embryo development.

[0087] The present invention makes full use of the relevant information contained in the embryo image sequence, adopts the VideoMamba network model, and combines the progressive use of the active learning algorithm to select uncertain samples and source-free domain adaptation to evaluate the quality of the embryo development image sequence as Figure 5 shown. Through a large number of experiments, it is shown that the method proposed by the present invention not only protects the privacy of the subjects, but also solves the challenge that the training of deep learning models requires a large amount of accurately labeled data, greatly reducing the labeling cost.

[0088] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, 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 relatively specific and detailed, but it cannot be understood as a limitation on the scope of the present invention. As long as the combinations of these technical features do not conflict, they should be considered as within the scope described in this specification.

[0089] It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. An artificial intelligence evaluation method for embryo development based on progressive active-passive domain adaptation, characterized in that: S1: Train a source domain model using a source domain dataset, where the source domain dataset is a sequence of embryo images with high-quality and low-quality labels; S2: Use the trained source domain model to generate pseudo-labels for the embryo image sequence dataset in the target domain, and initialize the target domain model using the source domain model and the pseudo-labels of the target domain data; S3: Use an active learning algorithm to select embryo image sequence samples according to the level of uncertainty. For samples with high uncertainty, perform manual annotation and update the target domain dataset; S4: Use the updated target domain dataset to train and optimize the target domain model; S5: Repeat steps S3 and S4 to obtain an optimized target domain model, and use the target domain model to perform artificial intelligence evaluation on embryo images; In step S3, the active learning algorithm calculates the uncertainty of embryo image sequence samples by the entropy method, selects a specified proportion of samples with the highest entropy values for manual annotation, and superimposes the annotated data with samples with low uncertainty to form a new embryo target domain dataset.

2. The artificial intelligence evaluation method for embryonic development based on progressive active and passive domain adaptation according to claim 1, wherein: In step S2, after extracting features from the target domain data through the initialized target domain model, first perform binary classification on the target domain data through a clustering module, divide the embryo development data into two categories: high development quality and low development quality, and generate pseudo-labels.

3. An artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation according to claim 2, characterized in that: In step S2, the source domain model or / and the target domain model is a feature extraction model using Video Mamba as the backbone network.

4. The artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation according to claim 1, characterized in that: In step S2, the entropy value of the sample is calculated by the following formula: ； In the formula, ( ) represents the probability that the sample belongs to the category ; represents the parameter of the Tsallis entropy; represents the total number of categories.

5. The artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation according to claim 1, wherein: In step S4, use the updated embryo target domain dataset to retrain the target domain model to iterate the new embryo target domain dataset, and judge the training result through the KL divergence and the binary cross-entropy loss function of the target domain.

6. The artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation according to claim 5, wherein: The KL divergence is expressed as: ； Wherein, is the total number of categories, is the index of the category, = 1 indicates poor embryo development, = 2 indicates good embryo development, is the category of the true probability distribution, is the probability distribution predicted by the model = 1 or = 2 of the probability distribution.

7. An artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation according to claim 5, characterized in that: The expression of the binary cross-entropy loss function of the target domain is: ； wherein, represents the total number of samples in the target domain; is the pseudo-label of the target domain sample; represents the probability that the model predicts the sample as a poor embryo.

8. The artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation according to claim 1, characterized in that: Penalize the weights of the target domain model through L2 regularization, and the expression of the L2 regularization is: ； Wherein, is the binary cross-entropy loss; is the L2 norm of all weight parameters; is the hyperparameter of the regularization strength; represents the th weight in the model.

9. An artificial intelligence evaluation method for embryonic development based on progressive active-passive domain adaptation according to claim 2, characterized in that: In step S2, when the target domain model performs binary classification, use the binary cross-entropy loss function of the target domain to measure the difference between the model prediction probability distribution and the actual label in the binary classification task.

Images