Pseudo-label-based passive unsupervised domain adaptive medical image segmentation method and device

Abstract

The invention discloses a passive unsupervised domain adaptive medical image segmentation method and device based on pseudo labels, and relates to the technical field of computers, and the method comprises a supervised source domain model training stage, an unsupervised adaptive target domain model training stage and a segmentation stage. The method specifically comprises the following steps of: performing pixel-level self-marking on a source domain medical image by applying pseudo-tag pixel-level uncertainty estimation to generate a reliable pseudo-tag; and on this basis, a strong and weak enhanced teacher-student target domain model is used, and a high-quality pseudo tag is fully utilized to train a target domain medical image, so that the network adapts to representation specific to the target domain medical image. According to the method, unsupervised segmentation is carried out on the target domain data graph on the premise that the source domain model and the target domain unmarked data are known, source domain data do not need to be accessed, the features of the source domain model can be reasonably utilized, and the problem existing in SFUDA can be well relieved; and the accurate segmentation capability on the target domain is still maintained under the condition that less data is available compared with the UDA method.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a passive unsupervised domain adaptation medical image segmentation method and device based on pseudo labels. Background Art

[0002] At present, medical images are strictly regulated, and it is unethical for hospitals to disclose patient privacy data, so it is often difficult to obtain source domain data to solve the unsupervised domain adaptation (UDA) problem of medical images. This situation significantly hinders the practical application of UDA in medical image segmentation in the real world. How to solve the UDA problem of medical image segmentation when no source domain data is available, that is, to solve the problem of passive unsupervised domain adaptation (SFUDA) for medical image segmentation of target domain data when the source domain is inaccessible, is a major challenge.

[0003] Traditional UDA problems usually focus on leveraging labeled images from available source domain datasets and training models using large and diverse training data. However, this approach ignores the rich features contained in the unlabeled images in the target domain dataset. Pseudo-label learning methods provide a solution that can effectively utilize target domain data by generating pseudo-labels for unlabeled images in the target domain and using these pseudo-labels in the training process, thereby improving the generalization ability and performance of the model in the target domain. This method effectively utilizes the rich information of the target domain dataset, helps the model capture the feature distribution of the target domain, and then improves the performance in the target domain. Without access to the source domain data, it transfers the unsupervised problem to a simpler semi-supervised problem. However, due to the significant distribution difference between the source and target domains, directly using the pseudo-labels generated by the source domain model for the target domain images may have problems such as insufficient accuracy, high noise level, and limited generalization ability. This method may not fully explore the intrinsic features and structure of the target domain data, limiting the performance improvement of the model in the target domain. Therefore, how to make the distribution of pseudo-labels close to the true labels of the target domain becomes the key to improving such methods. Summary of the invention

[0004] The purpose of this application is to propose an unsupervised domain adaptation medical image segmentation method and device based on multi-scale features to address the above-mentioned technical problems. The method and device can perform unsupervised segmentation on the target domain data graph under the premise of known source domain model and unlabeled data in the target domain. It does not require access to the source domain data and can reasonably utilize the source domain model features. It can well alleviate the problems existing in SFUDA and still maintain accurate segmentation capabilities in the target domain when less data is available than the UDA method.

[0005] On the one hand, a pseudo-label based passive unsupervised domain adaptation medical image segmentation method, comprising:

[0006] In the supervised source domain model training phase, an evidence segmentation network based on the U-net network model is constructed as the source domain model; based on the uncertainty estimation method of evidence theory, a mapping between pseudo labels and uncertainty, information entropy and relative feature distance is generated to generate reliable pseudo labels; the source domain model is trained in a supervised manner using the predicted probability output by the evidence segmentation network and the first evidence loss and the first Dice loss between the medical image labels from the source domain; after the source domain model training is completed, the parameters of the trained source domain model are passed to the target domain model training phase;

[0007] In the unsupervised adaptive target domain model training phase, a target domain model consisting of a teacher model and a student model initialized with the parameters of the trained source domain model is constructed; for weakly enhanced images of unlabeled medical images of a given target domain, reliable pseudo-labels are generated through the teacher model; the student model is trained using strongly enhanced images of unlabeled medical images of a given target domain and the pseudo-labels generated by the teacher model as input data; the parameters of the teacher model are updated in a moving average manner as the student model learns; the target domain model segmentation loss and the second evidence loss are used to train the target domain model in an unsupervised manner, the target domain model segmentation loss is used to encourage the student model to learn reliable pseudo-labels from the teacher model, and the second evidence loss is used to encourage the teacher model to generate potential evidence from the predicted labels of the student model;

[0008] In the segmentation stage, the trained target domain model is used to segment the medical image to be processed.

[0009] Preferably, the medical image label of the source domain is the pseudo label of the i-th pixel It is expressed as follows:

[0010] argmax ;

[0011] in, is the predicted probability of pixel i, is the source domain model for pixel i about the The predicted probability of the class, K represents the number of segmentation categories; , It is the evidence segmentation network for the source domain medical image label The output of the i-th pixel;

[0012] The label selection mask for a binary vector is expressed as follows:

[0013] ;

[0014] in, Indicates uncertainty; represents information entropy; Represents relative feature distance; is the uncertainty threshold, is the low entropy threshold, is the low distance threshold.

[0015] Preferred, Uncertain The calculation formula is as follows:

[0016] ;

[0017] Among them, the quality of belief Represents the source domain model for the source domain medical image label The output of the i-th pixel is The degree of belief of the class; is the source domain medical image label All evidence at the ith pixel;

[0018] Information Entropy The calculation formula is as follows:

[0019] ;

[0020] The relative feature distance between each pixel i and its corresponding nearest class centroid feature It is expressed as follows:

[0021] ;

[0022] Class feature pseudo centroid The calculation formula is as follows:

[0023] ;

[0024] in, The total number of pixels representing the source domain medical image labels; It represents the feature map obtained by upsampling the feature map obtained by the last convolution, which has the same dimension as the pseudo-label.

[0025] Preferably, the first evidence loss Including the first cross entropy loss and KL divergence loss , which is expressed as follows:

[0026] ;

[0027] in, is the annealing attenuation coefficient, is the number of epochs for training, is the maximum number of epochs;

[0028] First cross entropy loss , which is expressed as follows:

[0029] ;

[0030] Among them, K represents the number of segmentation categories; is the digamma function, is the source domain model for pixel i about the The predicted probability of the class; is the source domain medical image label At pixel i about The label of the class; is the source domain medical image label All evidence at the ith pixel; is the Dirichlet distribution parameter;

[0031] KL divergence loss , which is expressed as follows:

[0032] ;

[0033] in, is the predicted probability of pixel i; is the Dirichlet parameter after removing non-misleading evidence from the prediction at pixel i, is a uniform Dirichlet distribution, represents the Dirichlet distribution after removing non-misleading evidence from the prediction results, is the label of pixel i.

[0034] Preferably, the first Dice loss , which is expressed as follows:

[0035] ;

[0036] Among them, K represents the number of segmentation categories; is the source domain medical image label At pixel i about The label of the class; For the quality of belief Softmax after softmax function result.

[0037] Preferably, the parameter update method of the teacher model is expressed as follows:

[0038] ;

[0039] in, are the parameters of the teacher model, are the parameters of the student model, is the update rate.

[0040] Preferably, the overall loss after the target domain model segmentation loss and the second evidence loss are combined , which is expressed as follows:

[0041]

[0042] in, Indicates loss of secondary evidence; represents the target domain model segmentation loss; represents the predicted probability of the teacher model for the weakly enhanced image, argmax represents the predicted label of the student model for the strongly enhanced image, represents the predicted probability of the student model for the strongly enhanced image, argmax represents the predicted label of the teacher model for the weakly enhanced image.

[0043] Preferably, the target domain model segmentation loss , which is expressed as follows:

[0044] ;

[0045] in, represents the second cross entropy loss, Represents the second Dice loss, which are as follows:

[0046] ;

[0047] ;

[0048] in, Indicates the number of segmentation categories; with argmax Correspondingly, is the predicted label of the teacher model for the weakly enhanced image at pixel j; and Correspondingly, the student model is the strong enhancement image at pixel j about the The predicted probability of the class.

[0049] On the other hand, a pseudo-label-based passive unsupervised domain adaptation medical image segmentation device comprises:

[0050] The supervised source domain model training module is used to construct an evidence segmentation network based on the U-net network model as the source domain model; based on the uncertainty estimation method of evidence theory, a mapping between pseudo labels and uncertainty, information entropy and relative feature distance is generated to generate reliable pseudo labels; the source domain model is trained in a supervised manner using the predicted probability output by the evidence segmentation network and the first evidence loss and the first Dice loss between the medical image labels from the source domain; after the source domain model training is completed, the parameters of the trained source domain model are passed to the target domain model training stage;

[0051] The unsupervised adaptive target domain model training module is used to construct a target domain model consisting of a teacher model and a student model initialized with the parameters of the trained source domain model; for weakly enhanced images of unlabeled medical images of a given target domain, reliable pseudo labels are generated through the teacher model; the student model is trained using strongly enhanced images of unlabeled medical images of a given target domain and the pseudo labels generated by the teacher model as input data; the parameters of the teacher model are updated in a moving average manner as the student model learns; the target domain model is trained in an unsupervised manner using the target domain model segmentation loss and the second evidence loss, the student model is prompted to learn reliable pseudo labels from the teacher model through the target domain model segmentation loss, and the teacher model is prompted to generate potential evidence from the predicted labels of the student model through the second evidence loss;

[0052] The segmentation module is used to segment the medical images to be processed using the trained target domain model.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] (1) The present invention quantifies the uncertainty of source domain samples by using evidence theory, and helps the network improve classification accuracy by screening low-uncertainty samples. Then, the samples selected based on evidence theory are assisted in correction by information entropy, thereby obtaining more accurate pseudo labels.

[0055] (2) The data augmentation strategy of weak-strong enhancement introduced in the present invention is used to improve the adaptability and generalization ability of the model in the face of new environments, and to improve the performance of medical image segmentation. The consistency regularization of the strong-weak enhancement mechanism effectively promotes the model to learn a more robust feature representation by simultaneously learning the weakly enhanced and strongly enhanced versions of the same image. Images that have consistent semantic content despite visual differences due to enhancement are mapped to the same feature space region, which not only improves the discrimination of the feature representation, but also enables the model to more stably perform correct classification when faced with slight changes in medical images or inconsistent scanning quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 A schematic diagram of a flow chart of a pseudo-label-based passive unsupervised domain adaptation medical image segmentation method according to an embodiment of the present application;

[0058] Figure 2 A schematic diagram of the overall model structure of a medical image segmentation model of a pseudo-label-based passive unsupervised domain adaptation medical image segmentation method according to an embodiment of the present application;

[0059] Figure 3 The segmentation results of the embodiment method of the present application and the prior art method in the fundus image;

[0060] Figure 4 The segmentation results of the embodiment method of the present application and the prior art method in the prostate image;

[0061] Figure 5 The ablation experiment results of the embodiment method of the present application and the prior art method on the Drishti-GS fundus image dataset;

[0062] Figure 6 The ablation experiment results of the embodiment method of the present application and the prior art method on the RIM-ONE_r3 fundus image dataset;

[0063] Figure 7 Schematic diagram of a pseudo label with high noise generated on a target domain image, an uncertainty map, a pseudo label after noise reduction using an uncertainty estimation method, and an image label comparison; wherein (a) is a pseudo label with high noise generated directly on a target domain image; (b) is an uncertainty map; (c) is a pseudo label after quality improvement using the method of an embodiment of the present application; and (d) is a real image label;

[0064] Figure 8 Schematic diagram of a pseudo-label-based passive unsupervised domain adaptation medical image segmentation device according to an embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] like Figure 1 As shown, this embodiment provides a source-free unsupervised domain adaptation medical image segmentation method based on pseudo-label (Source-free Unsupervised Domain Adaptation Medical Image Segmentation Method Based on Pseudo-label Adaptive Optimization, AOPL), including the following steps:

[0067] S101, supervised source domain model training phase, constructing an evidence segmentation network based on the U-net network model as the source domain model; generating a mapping between pseudo labels and uncertainties, information entropy and relative feature distances based on the uncertainty estimation method of evidence theory, and generating reliable pseudo labels; using the predicted probability output by the evidence segmentation network and the first evidence loss and the first Dice loss between the medical image labels from the source domain to train the source domain model in a supervised manner; after the source domain model training is completed, passing the parameters of the trained source domain model to the target domain model training phase;

[0068] S102, unsupervised adaptive target domain model training phase, constructing a target domain model consisting of a teacher model and a student model initialized with parameters of a trained source domain model; generating reliable pseudo labels through the teacher model for weakly enhanced images of unlabeled medical images of a given target domain; using strongly enhanced images of unlabeled medical images of a given target domain and pseudo labels generated by the teacher model as input data to train the student model; updating the parameters of the teacher model in a moving average manner as the student model learns; using the target domain model segmentation loss and the second evidence loss to train the target domain model in an unsupervised manner, using the target domain model segmentation loss to encourage the student model to learn reliable pseudo labels from the teacher model, and using the second evidence loss to encourage the teacher model to generate potential evidence from the predicted labels of the student model;

[0069] S103, the segmentation stage, uses the trained target domain model to segment the medical image to be processed.

[0070] Specifically, in the source domain model training phase, the model is trained using image labels from the source domain to generate segmentation output and evidence output of the model. The source domain model is optimized in a supervised manner using the segmentation loss and evidence loss between the output of the source domain model and the label. After pre-training, the source domain model parameters are passed to the target domain model training phase, i.e., the adaptation phase. In the target domain model training phase, only medical image samples from the target domain and model parameters passed from the source domain model training phase are allowed to be used for the adaptation of a given pre-trained model. The target domain model consists of a teacher model and a student model initialized with the source domain model parameters. For the weakly enhanced image of the unlabeled image of the given target domain, the teacher model is used to generate the corresponding pseudo-label and evidence vector. This process ensures that the pseudo-label generated by the teacher model is reliable through pixel-level uncertainty estimation. Then, the student target domain model is used to fully utilize the high-quality pseudo-label to train the strongly enhanced version of the target domain data to adapt the network to the representation specific to the target domain data.

[0071] The following is a detailed description of a passive unsupervised domain adaptation medical image segmentation method based on pseudo labels, which is generally divided into two parts: pseudo label generation based on pixel-level uncertainty estimation and teacher-student target domain model based on strong and weak enhancement. The overall model structure is as follows: Figure 2 shown.

[0072] (1) Pseudo-label generation based on pixel-level uncertainty estimation

[0073] In this embodiment, a method for pixel-level pseudo-label denoising is performed through uncertainty estimation based on evidence theory to generate reliable pseudo-labels. Pseudo-labels generated directly through the source domain model may not only fail to bring positive improvements to the target domain model training, but also cause the target domain model to only learn the features or patterns of the source domain features, thereby reducing its accuracy and generalization ability in the target domain. How to reduce the unreliability of pseudo-labels is the key to solving this problem. To this end, a method of uncertainty estimation based on evidence theory is introduced. Through uncertainty estimation, pseudo-labels are analyzed, unreliable pseudo-label predictions are discarded, and pseudo-labels with high credibility are screened out.

[0074] In the medical image segmentation task, for the uncertainty estimation of pseudo labels, early uncertainty quantification methods based on Bayesian theory mainly deal with the uncertainty related to the category probability, but these methods may often lead to inaccurate estimates due to insufficient data or imperfect models. There are also some methods such as Monte Carlo inference, which can provide quantitative estimates of uncertainty, but usually require multiple forward propagations, are computationally expensive, and are difficult to adapt to dynamically changing data distributions. In order to solve these problems, evidence theory is introduced into the field of medical image SFUDA segmentation. As an extension of Bayesian theory, evidence theory provides a novel uncertainty quantification method that can express uncertainty by aggregating evidence from data without prior knowledge. Evidence theory expresses uncertainty by collecting data to support the amount of evidence that classifies a sample into a specific category. Evidence theory can quantify the uncertainty of each pixel, not just the uncertainty of the entire image, which is crucial for accurate segmentation of medical images. Evidence theory uncertainty estimation methods evaluate uncertainty by integrating all necessary information in one forward pass of the model, thereby greatly reducing computational complexity. This approach avoids the need for multiple inferences, thus reducing the consumption of computing resources, while also avoiding the instability of prediction results caused by multiple random Dropout operations in the Monte Carlo method. More importantly, evidence theory does not require modifying the structure of the model, such as adding Dropout layers, thus avoiding possible unknown effects on model performance. By obtaining uncertainty estimates with only one forward pass, the evidence theory approach improves the efficiency of the model while maintaining the stability and accuracy of the predictions. This makes evidence theory an efficient and reliable uncertainty estimation method, particularly suitable for application scenarios with limited computing resources or high requirements for prediction stability. By using evidence theory to quantify the uncertainty of target domain samples and helping the network improve classification accuracy by screening low-uncertainty samples, the samples selected based on evidence theory are then assisted in correction by information entropy to obtain more accurate pseudo-labels.

[0075] This embodiment designs an evidence segmentation network (ESN) based on the U-net network model. The design purpose of this network is to generate reliable pseudo labels through uncertainty estimation of evidence theory, and to generate pseudo labels and uncertainty mapping at the same time. The uncertainty estimation is used to generate pseudo labels with high credibility, thereby enhancing the ability to generate reliable pseudo labels from unlabeled target domain data.

[0076] The main difficulty of evidence theory is to obtain evidence. This embodiment proposes a transformation function with input to obtain the evidence vector of the sample, which is specifically shown in the following formula.

[0077] ;

[0078] in, is the evidence vector, represents the transformation function, It is the evidence segmentation network for the source domain medical image label The output of the ith pixel, is a scaling factor whose value is the inverse of the segmentation category. Subjective logic proposes a logical reasoning framework based on belief quality and uncertainty. In subjective logic, belief quality indicates the degree of belief in a proposition or event, while uncertainty indicates the degree of uncertainty about the degree of belief. Represents the model pair The output of the i-th pixel is The quality of the belief of the class, Represents the model pair The uncertainty of the prediction of the i-th pixel is obtained according to subjective logic.

[0079] ;

[0080] Represents the number of segmentation categories. Due to the belief quality Represents the model pair The output of the i-th pixel is The degree of belief of the class, and the evidence vector Express The output of the i-th pixel is The support level of the class. Then we can define the following calculation .

[0081] ;

[0082] in, for The total evidence for the i-th pixel is based on the uncertainty of the above formula The calculation formula is as follows:

[0083] .

[0084] Therefore, based on the above formula, uncertainty is inversely proportional to the total evidence. When there is no evidence, the belief quality of each category is zero, and the uncertainty is equal to 1. On the contrary, if the total amount of evidence obtained is large, the uncertainty is correspondingly small, which means that the model has a high confidence in the prediction. Therefore, in the training process, evidence analysis calculates the uncertainty of the pixel prediction based on the total amount of evidence for the pixel. , and optimizes the generation of pseudo labels specific to the target domain based on these confidence distributions to optimize the quality of pseudo labels.

[0085] The quality of pseudo labels can be improved by filtering out pixels with low uncertainty through evidence theory. In order to further improve the quality of pseudo labels, this section uses information entropy and clustering-based feature distance methods to assist in correcting sample images selected based on evidence theory, thereby obtaining more accurate pseudo labels. The calculation of is as follows:

[0086] .

[0087] Information entropy can be used to screen out pseudo-labels with low entropy, thereby further improving the quality of pseudo-labels. Segmentation problems have structural characteristics, and segmented areas of the same category are usually highly correlated. In order to make full use of this phenomenon to assist in correcting pseudo-labels, a clustering-based feature distance method is proposed to improve the accuracy of pseudo-label generation by optimizing the clustering process. The clustering-based feature distance method uses clustering, that is, class feature pseudo-centroids, to identify and eliminate potential unreliable pseudo-labels, ensuring that the selected pseudo-label is closer to the feature distance of its class prototype to improve the quality of the pseudo-label. In the clustering-based feature distance method, it is first necessary to calculate the class feature pseudo-centroid. , class feature pseudo centroid The calculation of is shown below.

[0088] ;

[0089] in, The total number of pixels representing the source domain medical image labels; Represents the feature map obtained by the last convolution after upsampling the feature map with the same dimension as the pseudo label. The relative feature distance between each pixel i and the centroid feature of its corresponding nearest category Use the following formula to calculate:

[0090] .

[0091] In order to use uncertainty graph to filter noisy pseudo labels, use information entropy to filter high entropy pseudo labels, and use relative feature distance to filter low distance samples to assist in correcting pseudo labels, a binary vector is defined. The label selection mask of pixel i is given by , and Decide as follows:

[0092] ;

[0093] in, is the uncertainty threshold, is the low entropy threshold, is the low distance threshold. In this embodiment, Set to 0.05, Set as the median entropy of all pixels, Set to the median of the relative feature distances of all pixels. If the uncertainty is low enough and it is a low entropy, low distance sample, use to select the corresponding pseudo-label; otherwise, the pseudo-label is regarded as noise and excluded from the loss calculation. Therefore, in the source model pseudo-label generation, through the continuous training of the model through evidence theory analysis, the calculation of uncertainty, information entropy and relative feature distance can be performed periodically to dynamically screen out new high-confidence samples. For the predicted probability of pixel i ,in , is the model's calculation of pixel i about the The predicted probability of the class, then the pseudo label of pixel i argmax .

[0094] The generation of pseudo-labels is a dynamic process. As the model performance improves, the quality of pseudo-labels can be continuously optimized to further improve learning efficiency and model accuracy.

[0095] In order to quantify the degree of trust and transform the generated evidence into a probabilistic framework that can be used for decision making, the Dirichlet distribution is used to represent the probability distribution of belief quality. The parameters of the Dirichlet distribution are defined as And the predicted probability for pixel i , there is the following Dirichlet distribution.

[0096] ;

[0097] in, is a multinomial Beta equation in dimension K. The parameters of this Dirichlet distribution can explain the evidence for each category when When the prediction result of is related to one of the K categories, the corresponding Dirichlet distribution parameter is increased to update the Dirichlet distribution to the new observation. and Following this Dirichlet distribution, deep learning work based on quantifying uncertainty introduces cross entropy loss , as shown in the following formula.

[0098] ;

[0099] in, is the digamma function, for At pixel i about Class label. By optimizing this loss, the model will try to generate more evidence for the confirmed label. However, for uncertain or incorrectly predicted labels, the model may also generate evidence, which will make the uncertain part more difficult to predict. In order to penalize the uncertainty of the model on unknown labels, thereby prompting the model to generate more reliable predictions. The KL divergence between the Dirichlet distribution and the uniform distribution of the model output is added to the loss function to regularize the model output. This can encourage the model to generate high-confidence evidence for the correct label, and reduce the output evidence to zero in the case of uncertainty or misclassification, so as to improve the robustness and generalization ability of the overall model, so the loss of KL divergence The details are shown in the following formula.

[0100] ;

[0101] in, is the Dirichlet parameter after removing non-misleading evidence from the prediction at pixel i, is a uniform Dirichlet distribution, represents the Dirichlet distribution after removing non-misleading evidence from the prediction results. In order to obtain more evidence, the first evidence loss generated by the optimization model is As shown in the following formula.

[0102] ;

[0103] in, is the annealing attenuation coefficient, is the number of epochs for training, is the maximum number of epochs. To generate more evidence to reduce uncertainty. However, it mainly focuses on the prediction accuracy at the single pixel level and ignores the spatial relationship between pixels, which is crucial in image segmentation. In the image segmentation task, the spatial continuity and mutual relationship between pixels are crucial to obtain high-quality segmentation results. Failure to fully capture these relationships will result in inaccurate segmentation boundaries or neglect of some subtle structures. Therefore, in order to compensate for this, the first Dice loss is introduced as a supplement. Different from the traditional Dice loss, the belief quality Softmax after softmax function The results are regarded as , Defined as:

[0104] .

[0105] In summary, the total loss of the evidence segmentation network of the source domain model proposed in this embodiment is The definition is as follows:

[0106] .

[0107] Loss of the source model in optimizing the evidence segmentation network In the process, through The loss function and The loss function is optimized to improve the model performance. First, the loss function of generating evidence is optimized. , ensuring that the model can generate sufficient evidence for each sample and enhancing the confidence and accuracy of the model when processing data. In the KL divergence based Guiding the model to generate less evidence for incorrect samples helps the model distinguish samples that are under-predicted by the model, so that more attention can be paid to these samples with higher uncertainty in subsequent training. Next, by optimizing the Dice loss Strengthening the spatial relationship between different predicted pixels is crucial to maintaining the consistency and accuracy of the prediction. By predicting the correlation between pixels, it can ensure that the segmentation boundaries are more precise, thus more in line with the real image structure in detail. The process not only improves the performance of the evidence segmentation network in predicting a single pixel, but also ensures the global consistency and accuracy of the prediction.

[0108] (2) Teacher-student target domain model based on strong and weak reinforcement

[0109] The challenge currently faced by the SFUDA field is the instability in the training process. In existing methods, the model is often highly dependent on its self-generated prediction labels that may contain high noise, but this will cause instability in the training process. In order to improve the stability of the training process and continuously improve the quality of pseudo-labels, this embodiment introduces the average teacher model to simulate unsupervised domain adaptation as supervised learning. Based on the average teacher model, a teacher student model based on strong-weak data augmentation (TSDA) is proposed to build a dual model system, including a teacher model and a student model. The teacher model inherits from the ESN model trained with the source data set, while the student model adopts a relatively simple U-net model (for the student model, it is supervised on the target domain data and the target domain data pseudo-label provided by the teacher model, and does not need to use the additional evidence theory uncertainty estimation in ESN, which can save computing resources), and the model parameters are initialized by the source model. In this collaborative framework, the role of the teacher model is to provide high-quality pseudo-labels, guide the student model to gradually adapt to the data distribution of the target domain, and generate accurate prediction results. Specifically, the teacher model generates reliable pseudo-labels from weakly data-augmented images through uncertainty estimation based on evidence theory, while the student model is trained with a strongly data-augmented version of the same image and reliable pseudo-labels from the teacher model as input. The teacher model is not trained directly through traditional back-propagation, but indirectly by updating its weights to the moving average of the student model weights, in order to resist the direct impact of noisy pseudo-labels and gradually accumulate new knowledge learned by the student model.

[0110] The purpose of the TSDA framework proposed in this embodiment is to ensure the robustness of the model training process and prevent the accumulation of errors during the training process. It contains two key models: one is the teacher model that plays a guiding role. , and the other is a student model that performs learning tasks , the main responsibility of the teacher model is to provide high-quality pseudo-label guidance to the student model through its predictions, while the student model is committed to learning these pseudo-labels to adapt to the target task. The data augmentation strategy introduced by weak and strong enhancement is used to improve the model's adaptability and generalization ability in the face of new environments, and improve the performance of medical image segmentation. The consistency regularization of the strong and weak enhancement mechanism effectively promotes the model to learn more robust feature representations by simultaneously learning weakly enhanced and strongly enhanced versions of the same image. Mapping images that are visually different due to enhancement but have consistent semantic content to the same feature space region not only improves the discriminability of feature representations, but also enables the model to more stably perform correct classification when faced with slight changes in medical images or inconsistent scanning quality. Specifically, for each input target domain medical image, slight data augmentation techniques, such as image flipping and scaling, are first applied to generate a weakly enhanced version of the image. At the same time, in order to increase the model's resilience against noise, a more drastically transformed strongly enhanced version is also created. In this process, the image is weakly enhanced. For each pixel j Through the teacher model Generate corresponding reliable pseudo labels using the uncertainty estimation method ESN based on evidence theory These pseudo labels are then used to train the student model in a supervised manner. , the input is the corresponding strongly enhanced image This design not only enables teacher models to leverage their knowledge on source data to generate more accurate guidance labels, but also enables student models to learn under more challenging conditions through strong reinforcement inputs, thereby enhancing their generalization and adaptability to unknown data.

[0111] In addition, the weights of the teacher model are updated in a moving average (EMA) manner as the student model learns, rather than training the teacher model directly. This design helps balance the stability of the model during training and the accumulation of new knowledge, thereby avoiding potential training instability caused by over-adaptation of the noisy pseudo-labels of the student model. The teacher model parameter update is shown below.

[0112] ;

[0113] in, are the parameters of the teacher model, are the parameters of the student model, is the update rate.

[0114] The student model is supervised and trained on the target domain data and pseudo labels. The overall loss function of the student model consists of two parts: the second segmentation loss function and the second evidence loss function. In order to combine the advantages of cross entropy loss and Dice loss, the second segmentation loss function is designed. The formula is as follows.

[0115] ;

[0116] in, Indicates the number of segmentation categories; with argmax Correspondingly, is the predicted label of the teacher model for the weakly enhanced image at pixel j; and Correspondingly, the student model is the strong enhancement image at pixel j about the The predicted probability of the class.

[0117] In summary, the overall loss of the target domain model training phase proposed in this embodiment is As follows:

[0118] ;

[0119] in, is the predicted probability of the teacher model for the weakly enhanced data, is the predicted label of the student model for the strongly augmented data, is the predicted probability of the student model for the strongly enhanced data is the predicted label of the teacher model for the weakly enhanced data. By combining the overall loss into the second segmentation loss and the second evidence loss, the second segmentation loss can be used to Encourage the student model to learn reliable pseudo-labels from the teacher model through the second evidence loss This encourages the teacher model to generate potential evidence from the predicted labels of the student model. This is an iterative process in which the teacher model continuously provides high-quality pseudo labels to improve the student model's understanding of real data; the improvement of the student model in turn obtains a better teacher model through EMA, allowing the teacher model to generate more evidence and ensure the reliability of pseudo labels. The two work together to overcome domain differences.

[0120] The pseudo-label-based passive unsupervised domain adaptation medical image segmentation method of this embodiment is verified by experiments, including the following steps.

[0121] (1) Dataset.

[0122] In the field of medical image segmentation, medical image database is the cornerstone of medical image segmentation. Due to the difficulty of data collection and high data privacy requirements, it is impossible to provide sufficient samples for the realization of medical image segmentation research. Therefore, the research and development of medical image segmentation is challenged by the small data samples. This embodiment mainly uses public fundus image datasets and prostate image datasets for experiments. There are mainly three fundus image datasets, namely Drishti-GS, RIM-ONE_r3 and REFUGE challenge training set, and three prostate image datasets, namely NCI, I2CVB and PROMISE12. These datasets are introduced below.

[0123] Drishti-GS dataset: The dataset was developed by a team of researchers from India to provide high-quality fundus images to support automatic segmentation of the optic disc and optic cup. It provides 101 2049 color fundus images and corresponding annotations (the training set includes 50 images and the test set includes 51 images).

[0124] RIM-ONE_r3 dataset: is a public fundus image dataset focusing on dry eye research, aiming to promote the development of glaucoma detection and optic disc and optic cup segmentation technology. This dataset collects 159 images from several European hospitals. The training set includes 99 images and the test set includes 60 images. Each image has been accurately annotated by ophthalmologists, including the boundaries of the optic disc and optic cup.

[0125] REFUGE Challenge Dataset: The 2018 REFUGE Challenge provides a set of annotated fundus images for training and validating computer vision and machine learning algorithms. The training set includes 400 The REFUGE dataset contains high-quality color fundus images that are commonly used for the diagnosis and monitoring of glaucoma. These images can show the detailed structure of the retina, optic disc, and macula, which is crucial for the diagnosis of glaucoma. The images in the dataset are accompanied by expert annotations, including the boundary positioning of the optic disc and optic cup, which are key indicators for evaluating optic nerve head damage and glaucoma. By calculating the cup-to-disc ratio, it can assist in the diagnosis of glaucoma.

[0126] NCI-ISBI2013 Dataset: The NCI-ISBI2013 dataset is a prostate magnetic resonance imaging (MRI) data challenge jointly initiated by the Cancer Imaging Program of the National Cancer Institute (NCI) of the United States and the International Society of Biomedical Imaging (ISBI) in 2013. The training set of the challenge dataset includes 60 cases, half of which are from Boston Medical Center (BMC), with a magnetic field strength of 1.5T, a resolution of 0.4 mm in the transverse plane and 3 mm in the sagittal plane. An endoscopic coil is used, and imaging is performed by Philips equipment. The other half of the cases are provided by Radboud University Medical Center (RUNMC) in Nijmegen, the Netherlands, with a magnetic field strength of 3T, a resolution of 0.6-0.625 mm in the transverse plane and 3.6-4 mm in the sagittal plane. A surface coil is used, and imaging is performed by Siemens equipment.

[0127] I2CVB dataset: This dataset comes from the Initiative for Collaborative Computer Visior Benchmarking (I2CVB), provided by the HCRUDB institution, and contains 19 cases. The magnetic field strength is 3T, the resolution is 0.67-0.79 mm in the transverse plane and 1.25 mm in the sagittal plane. No endoscopic coil is used, and imaging is performed by Siemens equipment.

[0128] PROMISE1 dataset: This dataset is derived from ProstateMRImageSegmentation2012 (PROMISE12), provided by UCL, BIDMC and HK, with 13, 12 and 12 cases respectively. The dataset used in the experiment of this embodiment is the data provided by HK, which contains 12 cases, with a magnetic field strength of 1.5T, a resolution of 0.625 mm in the transverse plane and 3.6 mm in the sagittal plane, using an endoscopic coil and imaging by Siemens equipment.

[0129] (2) Data preprocessing

[0130] Specifically, if the entire medical image is used directly for training, there are the following problems: the background information of the original image, such as edge areas and non-target structures, may introduce noise, affecting the model's ability to recognize key features, resulting in a decrease in the model's resolution. To this end, this paper introduces the concept of ROI (region of interest). The ROI image focuses on the segmented area and reduces background interference, allowing the model to learn the details of these key areas more attentively, thereby improving the accuracy of segmentation.

[0131] Due to the limitation of the small amount of medical image datasets, the model trained directly with medical image datasets is prone to overfitting and poor generalization. In order to improve the model training effect, this embodiment adopts a series of data enhancement strategies to simulate different visual effects and expand the training dataset. By inverting and rotating the image, adjusting the brightness and contrast, the diversity of samples can be artificially increased, which helps the model have better generalization ability in practical applications. After these enhanced images, for fundus images, by fixing the zoom to For the prostate image dataset, each sample is resized to The size of the image is 1000×10, ensuring the consistency of the data during the training process. At the same time, the image is randomly flipped and standardized so that the pixel values ​​of each channel have a uniform mean and standard deviation, further reducing the risk of overfitting during model training. Through these preprocessing steps, the number of images in the training dataset is increased and the quality is improved, ultimately solving the limitation of the difficulty of training models caused by the small amount of data in the medical image dataset.

[0132] (3) Evaluation indicators

[0133] In order to objectively evaluate the segmentation performance of the target domain model on the medical image dataset in the experiment, this embodiment selects two widely used standard indicators in image segmentation evaluation: Dice coefficient (Dice) and average surface distance (ASD). These indicators together provide a comprehensive evaluation for the quantification of segmentation effect. The Dice coefficient (Dice) compares the similarity between the predicted segmentation result and the true segmentation label. It calculates the ratio of twice the intersection area of ​​the predicted and true segmentation to the sum of their respective areas. The value range of the Dice coefficient is from 0 to 1. The higher the value, the closer the segmentation result is to the actual situation, that is, the better the segmentation effect. The calculation formula is as follows.

[0134] ;

[0135] Among them, A is the predicted segmentation result set, B is the actual segmentation result set, Represents the number of elements in the intersection of set A and set B. represents the number of elements in set A. Represents the number of elements in set B.

[0136] The average surface distance (ASD) measures the average distance between the predicted segmentation and the true segmentation surface. It calculates the average shortest distance from the predicted segmentation to all points on the true segmentation surface, and the average shortest distance from the true segmentation to all points on the predicted segmentation surface. The smaller the ASD value, the closer the segmentation result is to the true situation, that is, the higher the segmentation quality. The calculation formula is as follows.

[0137] ;

[0138] Among them, X1 and Y1 represent the point sets on the predicted segmentation surface and the true segmentation surface respectively. Indicate point and Point ASD calculates the average shortest distance between all pairs of points from X1 to Y1 and from Y1 to X1.

[0139] (4) Experimental environment and training configuration information

[0140] The software and hardware experimental environment in this experiment is shown in Table 1.

[0141] The training parameters of this experiment are detailed as follows:

[0142] The uncertainty threshold Set to 0.05, the information entropy threshold Set to the median entropy of all pixels, relative feature distance threshold Set to the median of the relative feature distances of all pixels. The initial learning rate is set to 0.001 and is reduced every 10 epochs. The batch size is set to 8, and the source domain model is obtained after 100 epochs of training. In the passive domain adaptation stage, the teacher and student models are first initialized from the source model. EMA is used to update the teacher model parameters from the student model parameters, and the model EMA update rate Set to 0.98. Set the learning rate to 5e -4 After 50 epochs, the final target domain student model is obtained. The model is trained using the Adam optimizer with momentum of 0.9 and 0.99.

[0143] Table 1 Experimental software and hardware environment configuration;

[0144]

[0145] (5) Experimental design

[0146] The experiment of this embodiment is mainly to explore the following questions:

[0147] First, in the SFUDA medical image segmentation scenario, what is the accuracy of the AOPL method proposed in the present invention, and whether it has any advantages compared with other UDA medical image segmentation methods, and the results can be intuitively seen through visual segmentation results.

[0148] Second, in the SFUDA medical image segmentation scenario, the impact of the ESA and TSDA modules proposed in the present invention on the model performance. In this embodiment, four groups of ablation experiments are designed to verify the effectiveness of each module using the same data set and training process.

[0149] Third, we visually verify the effect of the proposed AOPL method on improving the quality of pseudo labels through uncertainty estimation. We visually verify the effectiveness of AOPL in improving the quality of pseudo labels by displaying and analyzing high-noise pseudo labels, uncertainty graphs, and pseudo labels obtained by the AOPL method.

[0150] (6) Experimental results and analysis

[0151] (6.1) Fundus image segmentation experiment analysis

[0152] Specifically, experiments were conducted using publicly available fundus image datasets, including Drishti-GS, RIM-ONE_r3, and REFUGE challenge training sets. In order to verify the segmentation performance of the model on the target domain dataset when no source domain data is available, two domain transfer cases were designed, including: (1) setting the REFUGE challenge training set as the source domain dataset and the Drishti-GS dataset as the target domain dataset. (2) setting the REFUGE challenge training set as the source domain dataset and the RIM-ONE_r3 dataset as the target domain dataset. In order to examine the effectiveness of the model of this embodiment, algorithm comparison experiments were conducted. The comparison experiments were divided into: active unsupervised domain adaptation method experimental comparison to test whether the performance of the method of the present invention is insufficient compared with the active domain adaptation method; and SFUDA method experimental comparison to test whether the method of the present invention achieves the best results compared with similar methods. In the active domain adaptation method comparison experiment, the BEAL method and the pOSAL method were used as comparison algorithms. In the SFUDA method comparison experiment, the DPL method, the OS method, and the SFDA method were used as comparison algorithms. Among them, DPL is a passive unsupervised domain adaptation method that uses Monte Carlo for uncertainty estimation of pseudo-label generation, OS is a passive unsupervised domain adaptation method that uses the adaptation of batch normalization layers to align domains, and SFDA is a passive unsupervised domain adaptation method that recovers and retains source domain knowledge from a pre-trained source model and then extracts target domain information for self-supervised training. Similarly, adding Baseline represents a direct migration method that predicts target domain results directly from a baseline U-net model trained directly from source domain data. Each method is run 5 times, and then the average of the aforementioned evaluation criteria for measuring segmentation effects is summarized.

[0153] Table 2 shows the experimental results of different algorithms on the REFUGE challenge training set → Drishti-GS dataset. As can be seen from Table 2, the AOPL of the present invention performs better than other SFUDA methods in both Dice coefficient and ASD distance. It exceeds other UDA methods in optic disc Dice coefficient and optic cup Dice coefficient. It is worth noting that the best results are achieved in optic cup ASD and AOPL, and better than other UDA methods in optic disc ASD.

[0154] Table 2 Comparison of results of different algorithms on REFUGE challenge training set → Drishti-GS dataset;

[0155]

[0156] Table 3 shows the experimental results of different algorithms on the REFUGE challenge training set → RIM-ONE_r3 dataset. As can be seen from Table 3, the AOPL method performs better than other SFUDA methods in all evaluation indicators, and performs better than all other methods on the optic disc ASD. It is second only to the BEAL method in terms of the cup Dice coefficient and cup ASD distance. Combining the evaluation indicators of the above two domain transfer experiments, it can be seen that the AOPL method can not only solve the SFUDA problem in fundus image segmentation well when there is no source domain data available, but also outperforms some UDA methods, which proves the effectiveness of AOPL.

[0157] Table 3 Comparison of results of different algorithms on REFUGE challenge training set → RIM-ONE_r3 dataset;

[0158]

[0159] (6.2) Experimental analysis of prostate image segmentation

[0160] In order to verify the universality of the AOPL method on medical image datasets, three prostate image datasets, NCI-ISBI 2013, I2CVB and PROMISE12, were further verified. Design: (1) I2CVB dataset is the source domain, and NCI-ISBI 2013 is the target domain; (2) I2CVB dataset is the source domain, and PROMISE12 is the target domain. In this embodiment, only the source domain model and the target domain unlabeled data are available. The comparison method is the same as above, and the BEAL method, pOSAL method, DPL method, OS method, SFDA method and Baseline are used as comparison algorithms. The evaluation criteria are also compared with the average value of five experiments.

[0161] Table 4 shows the experimental results of different algorithms on the I2CVB dataset → NCI-ISBI 2013 dataset. It can be seen that the Dice coefficient and ASD distance of the proposed method are significantly better than those of other SFUDA methods. The Dice coefficient and ASD distance are improved by 0.72 and reduced by 1.05 respectively compared with the second-best DPL method. In comparison with the UDA method, it is better than pOSAL and close to BEAL in Dice, with a difference of only 0.67.

[0162] Table 4 Comparison of results of different algorithms on I2CVB dataset → NCI-ISBI 2013 dataset;

[0163]

[0164] Table 5 shows the experimental results of different algorithms on the I2CVB dataset → PROMISE12 dataset.

[0165] Table 5 Comparison of results of different algorithms on I2CVB dataset → ROMISE12 dataset;

[0166]

[0167] It can be seen that among the various methods of SFUDA, the overall performance is also lower than the domain transfer results of the I2CVB dataset → NCI-ISBI2013, which further shows that the size of the domain shift seriously affects the model segmentation performance. The AOPL method also performs best among all SFUDA methods. In the comparison of UDA methods, the Dice coefficient and ASD distance are both better than other methods and very close. Combining the performance of the AOPL method in the two domain transfer experiments of the prostate dataset, it can be seen that AOPL can not only solve the SFUDA problem in the prostate dataset very well, but also exceeds the performance of some other UDA methods. It proves that the AOPL method can still effectively solve the domain adaptation problem when using less data.

[0168] Figure 3 and Figure 4 The experimental results of different algorithms in the case of fundus image and prostate image domain adaptation are shown respectively. It can be seen from the figure that if no method is used to shorten the domain offset and direct migration is performed, very poor segmentation results will be obtained. Both the SFUDA method and the UDA method shorten the domain offset and the segmentation ability is significantly improved, but the segmentation shape and boundary of the OS and SFDA methods are not very good, while the AOPL method of the present invention performs well in both segmentation shape and boundary, and the effect exceeds that of other SFUDA methods, and the prediction results are close to the true label.

[0169] From the above, the AOPL method proposed in the present invention has been verified by comparative experiments and shows significant advantages in the SFUDA scenarios of fundus image datasets and prostate image datasets. The performance improvement of the AOPL method compared with other SFUDA technologies is demonstrated by comparing the Dice score and the ASD distance. In addition, the superiority of the segmentation effect of the AOPL method can also be clearly seen from the visualized image of the segmentation result. In summary, the AOPL method effectively solves the SFUDA problem in the field of medical image segmentation through the advantages of uncertainty estimation and consistency learning of evidence theory.

[0170] (6.3) Ablation experiment

[0171] This experiment conducts ablation experiments on the SFUDA case of the REFUGE challenge training set → Drishti-GS dataset to verify the effectiveness of the ESN method (evidence segmentation network) and TSDA method (teacher-student model based on strong-weak data augmentation) proposed in this chapter.

[0172] Table 6 Comparison of results in ablation experiments on the REFUGE challenge training set → Drishti-GS dataset;

[0173]

[0174] Table 6 shows the performance of different strategies on the REFUGE challenge training set ( Figure 5 )→Drishti-GS( Figure 6 ) dataset. It can be seen from the table that compared with the Baseline method of direct migration, the Dice coefficient and ASD distance of ESN and TSDA have been greatly improved. The segmentation performance of ESN is more improved than that of TSDA. This is because ESN can eliminate the prediction area with high uncertainty through the uncertainty estimation of evidence theory, and obtain prediction labels with better segmentation results. Although TSDA can improve the segmentation performance by emphasizing the consistency of prediction results with consistency loss through strong and weak data enhancement, due to the domain offset in the target domain, the generated prediction results have a lot of noise. At this time, the teacher-student model will accumulate errors due to continuous learning of high-noise prediction labels, resulting in a decrease in segmentation performance. Therefore, AOPL (Baseline + ESN + TSDA) combines the advantages of ESN and TSDA, while producing high-quality pseudo labels, and continuously optimizes the segmentation process through the teacher-student model. In this process, low-uncertainty segmentation areas are continuously screened, thereby achieving good segmentation performance.

[0175] (6.4) Other experiments

[0176] In order to further demonstrate the effectiveness of the evidence theory uncertainty estimation method proposed in this invention, the pseudo labels with high noise generated directly on the target domain image during the model training process, the uncertainty map, the pseudo labels and image labels after the noise is reduced by the uncertainty estimation method are shown below. Figure 7 (a) represents the high-noise pseudo-label generated on the target domain image. Figure 7 (b) represents the uncertainty diagram, Figure 7 (c) represents the pseudo-label after the quality is improved by the method of the present invention. Figure 7 (d) represents the image label. It can be seen from the figure that the predicted pseudo-label directly obtained on the target domain image has obvious noise problems compared with the real label. The highlighted area in the uncertainty map indicates the high uncertainty area, which usually corresponds to the segmentation boundary and background. When the background is incorrectly classified as a segmentation label, the uncertainty is very high. At this time, it can be well reclassified as the background to improve the quality of the pseudo-label. Although the uncertainty of the segmentation boundary is also high, it is more difficult to distinguish which category the boundary belongs to. It can be seen from the figure that the pseudo-label obtained after processing by the AOPL method can effectively handle the boundary, and its quality is significantly improved, which is close to the real image label.

[0177] like Figure 8 As shown, this embodiment also discloses a passive unsupervised domain adaptation medical image segmentation device based on pseudo labels, comprising:

[0178] The supervised source domain model training module 801 is used to construct an evidence segmentation network based on a U-net network model as a source domain model; generate a mapping between pseudo labels and uncertainties, information entropy and relative feature distances based on an uncertainty estimation method based on evidence theory, and generate reliable pseudo labels; use the predicted probability output by the evidence segmentation network and the first evidence loss and the first Dice loss between medical image labels from the source domain to train the source domain model in a supervised manner; after the source domain model training is completed, pass the parameters of the trained source domain model to the target domain model training stage;

[0179] The unsupervised adaptive target domain model training module 802 is used to construct a target domain model composed of a teacher model and a student model initialized with parameters of a trained source domain model; for a weakly enhanced image of an unlabeled medical image of a given target domain, a reliable pseudo-label is generated by the teacher model; the student model is trained using the strongly enhanced image of the unlabeled medical image of the given target domain and the pseudo-label generated by the teacher model as input data; the parameters of the teacher model are updated in a moving average manner as the learning progress of the student model; the target domain model is trained in an unsupervised manner using the target domain model segmentation loss and the second evidence loss, the student model is urged to learn reliable pseudo-labels from the teacher model through the target domain model segmentation loss, and the teacher model is urged to generate potential evidence from the predicted labels of the student model through the second evidence loss;

[0180] The segmentation module 803 is used to segment the medical image to be processed using the trained target domain model.

[0181] The specific implementation of each module of a passive unsupervised domain adaptive medical image segmentation device based on pseudo labels is the same as the passive unsupervised domain adaptive medical image segmentation method based on pseudo labels, and the present embodiment will not repeat the description.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A passive unsupervised domain adaptation medical image segmentation method based on pseudo labels, characterized in that: include: In the supervised source domain model training phase, an evidence segmentation network based on the U-net network model is constructed as the source domain model; Based on the uncertainty estimation method of evidence theory, a mapping between pseudo labels and uncertainty, information entropy and relative feature distance is generated to generate reliable pseudo labels; the source domain model is trained in a supervised manner using the predicted probability output by the evidence segmentation network and the first evidence loss and the first Dice loss between the medical image labels from the source domain; after the source domain model training is completed, the parameters of the trained source domain model are passed to the target domain model training stage; In the unsupervised adaptive target domain model training phase, a target domain model consisting of a teacher model and a student model initialized with the parameters of the trained source domain model is constructed; for weakly enhanced images of unlabeled medical images of a given target domain, reliable pseudo labels are generated through the teacher model; and the student model is trained using strongly enhanced images of unlabeled medical images of a given target domain and the pseudo labels generated by the teacher model as input data; As the student model progresses, the parameters of the teacher model are updated in a moving average manner. The target domain model is trained in an unsupervised manner using the target domain model segmentation loss and the second evidence loss. The target domain model segmentation loss is used to encourage the student model to learn reliable pseudo labels from the teacher model, and the second evidence loss is used to encourage the teacher model to generate potential evidence from the predicted labels of the student model. In the segmentation stage, the trained target domain model is used to segment the medical image to be processed.

2. The pseudo-label based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: The medical image label of the source domain is the pseudo label of the i-th pixel It is expressed as follows: argmax ; in, is the predicted probability of pixel i, is the source domain model for pixel i about the The predicted probability of the class, K represents the number of segmentation categories; , It is the evidence segmentation network for the source domain medical image label The output of the i-th pixel; The label selection mask for a binary vector is expressed as follows: ； in, Indicates uncertainty; represents information entropy; Represents relative feature distance; is the uncertainty threshold, is the low entropy threshold, is the low distance threshold.

3. The pseudo-label based passive unsupervised domain adaptation medical image segmentation method according to claim 2, characterized in that: Uncertainty The calculation formula is as follows: ； Among them, the quality of belief Represents the source domain model for the source domain medical image label The output of the i-th pixel is The degree of belief of the class; is the source domain medical image label All evidence at the ith pixel; Information Entropy The calculation formula is as follows: ； The relative feature distance between each pixel i and its corresponding nearest class centroid feature It is expressed as follows: ； Class feature pseudo centroid The calculation formula is as follows: ； in, The total number of pixels representing the source domain medical image labels; It represents the feature map obtained by upsampling the feature map obtained by the last convolution, which has the same dimension as the pseudo-label.

4. The pseudo-label based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: First evidence loss Including the first cross entropy loss and KL divergence loss , which is expressed as follows: ； in, is the annealing attenuation coefficient, is the number of epochs for training, is the maximum number of epochs; First cross entropy loss , which is expressed as follows: ； Among them, K represents the number of segmentation categories; is the digamma function, is the source domain model for pixel i about the The predicted probability of the class; is the source domain medical image label At pixel i about The label of the class; is the source domain medical image label All evidence at the ith pixel; is the Dirichlet distribution parameter; KL divergence loss , which is expressed as follows: ； in, is the predicted probability of pixel i; is the Dirichlet parameter after removing non-misleading evidence from the prediction at pixel i, is a uniform Dirichlet distribution, represents the Dirichlet distribution after removing non-misleading evidence from the prediction results, is the label of pixel i.

5. The pseudo-label-based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: First Dice Loss , which is expressed as follows: ； Among them, K represents the number of segmentation categories; is the source domain medical image label At pixel i about The label of the class; For the quality of belief Softmax after softmax function result.

6. The pseudo-label based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: The parameter update method of the teacher model is expressed as follows: ； in, are the parameters of the teacher model, are the parameters of the student model, is the update rate.

7. The pseudo-label based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: The overall loss after combining the target domain model segmentation loss and the second evidence loss , which is expressed as follows: ； in, Indicates loss of secondary evidence; represents the target domain model segmentation loss; represents the predicted probability of the teacher model for the weakly enhanced image, argmax represents the predicted label of the student model for the strongly enhanced image, represents the predicted probability of the student model for the strongly enhanced image, argmax represents the predicted label of the teacher model for the weakly enhanced image.

8. The pseudo-label-based passive unsupervised domain adaptation medical image segmentation method according to claim 7, characterized in that: Target domain model segmentation loss , which is expressed as follows: ； in, represents the second cross entropy loss, Represents the second Dice loss, which are as follows: ； ； in, Indicates the number of segmentation categories; with argmax Correspondingly, is the predicted label of the teacher model for the weakly enhanced image at pixel j; and Correspondingly, the student model is the strong enhancement image at pixel j about the The predicted probability of the class.

9. A passive unsupervised domain adaptation medical image segmentation device based on pseudo labels, characterized in that: include: A supervised source domain model training module is used to build an evidence segmentation network based on the U-net network model as the source domain model; Based on the uncertainty estimation method of evidence theory, a mapping between pseudo labels and uncertainty, information entropy and relative feature distance is generated to generate reliable pseudo labels; the source domain model is trained in a supervised manner using the predicted probability output by the evidence segmentation network and the first evidence loss and the first Dice loss between the medical image labels from the source domain; after the source domain model training is completed, the parameters of the trained source domain model are passed to the target domain model training stage; The unsupervised adaptive target domain model training module is used to construct a target domain model consisting of a teacher model and a student model initialized with the parameters of the trained source domain model; for weakly enhanced images of unlabeled medical images of a given target domain, reliable pseudo labels are generated through the teacher model; and the student model is trained using strongly enhanced images of unlabeled medical images of a given target domain and the pseudo labels generated by the teacher model as input data; As the student model progresses, the parameters of the teacher model are updated in a moving average manner. The target domain model is trained in an unsupervised manner using the target domain model segmentation loss and the second evidence loss. The target domain model segmentation loss is used to encourage the student model to learn reliable pseudo labels from the teacher model, and the second evidence loss is used to encourage the teacher model to generate potential evidence from the predicted labels of the student model. The segmentation module is used to segment the medical image to be processed using the trained target domain model.

Images