Pseudo-Labeling-Based Passive Unsupervised Domain Adaptation Method and Device for Medical Image Segmentation

Abstract

The present invention discloses a method and device for pseudo-label based unsupervised domain adaptation medical image segmentation, which relates to the field of computer technology. The method includes: a supervised source domain model training stage, an unsupervised adaptive target domain model training stage, and a segmentation stage. Specifically, it uses pseudo-label pixel-level uncertainty estimation to perform pixel-level self-labeling on source domain medical images to generate reliable pseudo-labels, and on this basis, uses a strong and weak enhanced teacher-student target domain model to make full use of high-quality pseudo-labels to train target domain medical images to adapt the network to the representation specific to target domain medical images. The present invention performs unsupervised segmentation on target domain data graphs on the premise of knowing the source domain model and unlabeled data in the target domain, does not require access to source domain data, and can reasonably utilize source domain model features, can well alleviate the problems existing in SFUDA, and still maintains accurate segmentation ability on the target domain when less data is available than the UDA method.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and particularly relates to a method and device for source-free unsupervised domain adaptation medical image segmentation based on pseudo-labels. Background Art

[0002] Currently, medical images are strictly regulated, and it is unethical for hospitals to disclose patients' privacy data. Therefore, 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 in the absence of available source domain data, that is, to solve the source-free unsupervised domain adaptation (SFUDA) problem of medical image segmentation of target domain data in the case where the source domain is inaccessible, is a major challenge.

[0003] Traditional UDA problems usually focus on using labeled images of the available source domain dataset to train the model with a large amount of diverse training data. However, this method ignores the rich features contained in the unlabeled images of the target domain dataset. The pseudo-label learning method provides a solution. By generating pseudo-labels for the unlabeled images of the target domain and using these pseudo-labels in the training process, the target domain data can be effectively utilized, thereby improving the generalization ability and performance of the model on 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 on the target domain. Without the need to access the source domain data, the unsupervised problem is transformed into a simpler semi-supervised problem. However, due to the significant distribution differences that may exist between the source domain and the target domain, directly using the source domain model to generate pseudo-labels for target domain images may have problems such as insufficient accuracy, high noise level, and limited generalization ability. This method may not be able to fully exploit the inherent features and structures of the target domain data, limiting the performance improvement of the model on the target domain. Therefore, how to make the distribution of pseudo-labels close to the true labels of the target domain has become the key to improving this type of method. Summary of the Invention

[0004] The purpose of this application is to propose a method and device for unsupervised domain adaptation medical image segmentation based on multi-scale features for the above-mentioned technical problems. Without accessing the source domain data and being able to reasonably utilize the source domain model features, it can perform unsupervised segmentation on the target domain data graph on the premise of knowing the source domain model and the unlabeled data of the target domain. It can well alleviate the problems existing in SFUDA and still maintain the accurate segmentation ability on the target domain when less data is available than the UDA method.

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

[0006] A supervised source domain model training stage, constructing an evidence segmentation network based on the U-net network model as the source domain model; using an uncertainty estimation method based on evidence theory to generate a mapping of pseudo-labels with uncertainty, information entropy, and relative feature distance to generate reliable pseudo-labels; using the first evidence loss and the first Dice loss between the predicted probability output by the evidence segmentation network and 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 stage;

[0007] An unsupervised adaptive target domain model training stage, constructing a target domain model composed of a teacher model and a student model initialized with the parameters of the trained source domain model; for the weakly augmented image of the unlabeled medical image in the given target domain, generating reliable pseudo-labels through the teacher model; using the strongly augmented image of the unlabeled medical image in the given target domain and the 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 learning progress of the student model; using the target domain model segmentation loss and the second evidence loss to train the target domain model in an unsupervised manner, promoting the student model to learn reliable pseudo-labels from the teacher model through the target domain model segmentation loss, and promoting the teacher model to generate potential evidence from the predicted labels of the student model through the second evidence loss;

[0008] A segmentation stage, using the trained target domain model to segment the medical image to be processed.

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

[0010] argmax ;

[0011] where is the predicted probability of pixel point i, is the predicted probability of the source domain model for pixel point i regarding the th class, and K represents the number of segmentation categories; , is the output of the evidence segmentation network for the i-th pixel of the source domain medical image label ;

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

[0013] ;

[0014] Among them, represents uncertainty; represents information entropy; represents the relative feature distance; is the uncertainty threshold, is the low entropy threshold, is the low distance threshold.

[0015] Preferably, the uncertainty is calculated as follows:

[0016] ;

[0017] Among them, the belief mass represents the degree of belief of the output of the source domain model for the i-th pixel of the source domain medical image label in the th class; is the total evidence of the source domain medical image label at the i-th pixel;

[0018] The information entropy is calculated as follows:

[0019] ;

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

[0021] ;

[0022] The pseudo centroid of the class feature is calculated as follows:

[0023] ;

[0024] Among them, represents the total number of pixels of the source domain medical image label; represents the feature map obtained by upsampling the feature map obtained by the last convolution and having the same dimension as the pseudo label.

[0025] Preferably, the first evidence loss includes the first cross-entropy loss and the loss of KL divergence , and is represented as follows:

[0026] ;

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

[0028] The first cross-entropy loss , is expressed as follows:

[0029] ;

[0030] Among them, K represents the number of segmentation categories; is the digamma function, is the prediction probability of the source domain model for pixel point i with respect to the th class; is the label of the source domain medical image at pixel point i with respect to the th class; is the label of the source domain medical image all the evidence at the i-th pixel; is the Dirichlet distribution parameter;

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

[0032] ;

[0033] Among them, is the prediction probability of pixel point i; is the Dirichlet parameter after removing non-misleading evidence from the prediction result in pixel point i, is the uniform Dirichlet distribution, represents the Dirichlet distribution after removing non-misleading evidence from the prediction result, is the label of pixel point i.

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

[0035] ;

[0036] Among them, K represents the number of segmentation categories; is the label of the source domain medical image at pixel point i with respect to the th class; is the belief mass after passing through the softmax function of the softmax result.

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

[0038] ;

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

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

[0041]

[0042] Wherein, represents the second evidence loss; represents the segmentation loss of the target domain model; represents the prediction probability of the teacher model for the weakly augmented image, argmax represents the prediction label of the student model for the strongly augmented image, represents the prediction probability of the student model for the strongly augmented image, argmax represents the prediction label of the teacher model for the weakly augmented image.

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

[0044] ;

[0045] Wherein, represents the second cross-entropy loss, represents the second Dice loss, which are respectively as follows:

[0046] ;

[0047] ;

[0048] Wherein, represents the number of segmentation categories; corresponds to argmax and is the prediction label of the teacher model for the weakly augmented image at pixel point j; corresponds to and is the prediction probability of the student model for the strongly augmented image at pixel point j for the th class.

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

[0050] A 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, generate a mapping of pseudo-labels to uncertainty, information entropy, and relative feature distance to generate reliable pseudo-labels; use the first evidence loss and the first Dice loss between the predicted probability output by the evidence segmentation network and 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, transfer the parameters of the trained source domain model to the target domain model training stage;

[0051] An unsupervised adaptive target domain model training module is used to construct a target domain model composed of a teacher model and a student model initialized with the parameters of the trained source domain model; for the weakly augmented images of the unlabeled medical images in the given target domain, generate reliable pseudo-labels through the teacher model; use the strongly augmented images of the unlabeled medical images in the given target domain and the pseudo-labels generated by the teacher model as input data to train the student model; update the parameters of the teacher model in a moving average manner as the learning progress of the student model; use the target domain model segmentation loss and the second evidence loss to train the target domain model in an unsupervised manner, prompting the student model to learn reliable pseudo-labels from the teacher model through the target domain model segmentation loss, and prompting the teacher model to generate potential evidence from the predicted labels of the student model through the second evidence loss;

[0052] A segmentation module is used to segment the medical image 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, helps the network improve the classification accuracy by screening low-uncertainty samples, and then uses information entropy to assist in correcting the samples selected based on evidence theory, so as to obtain more accurate pseudo-labels;

[0055] (2) The weak and strong augmentation data augmentation strategy introduced in the present invention is used to improve the adaptability and generalization ability of the model in the face of new environments, and improve the performance of medical image segmentation. The consistency regularization of the weak and strong augmentation mechanism effectively promotes the model to learn more robust feature representations by simultaneously learning the weakly augmented and strongly augmented versions of the same image, mapping images that are visually different due to augmentation but have the same semantic content to the same feature space region, not only improving the discrimination of feature representations, but also enabling the model to more stably perform correct classification in the face of minor changes in medical images or inconsistent scanning quality. Brief Description of the Drawings

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0057] Figure 1 Schematic flowchart of the pseudo-label-based unsupervised domain adaptation medical image segmentation method for the embodiments of this application;

[0058] Figure 2 Schematic diagram of the overall model structure of the medical image segmentation model of the pseudo-label-based unsupervised domain adaptation medical image segmentation method for the embodiments of this application;

[0059] Figure 3 Segmentation results of the method of the embodiments of this application and the method of the prior art in fundus images;

[0060] Figure 4 Segmentation results of the method of the embodiments of this application and the method of the prior art in prostate images;

[0061] Figure 5 Ablation experiment results of the method of the embodiments of this application and the method of the prior art on the Drishti-GS fundus image dataset;

[0062] Figure 6 Ablation experiment results of the method of the embodiments of this application and the method of the prior art on the RIM-ONE_r3 fundus image dataset;

[0063] Figure 7 Comparison schematic diagram of high-noise pseudo-labels, uncertainty maps, pseudo-labels after reducing noise using the uncertainty estimation method, and image labels generated on the target domain image; where (a) is the high-noise pseudo-label directly generated on the target domain image; (b) is the uncertainty map; (c) is the pseudo-label after improving the quality using the method of the embodiments of this application; (d) is the real image label;

[0064] Figure 8 Schematic diagram of the pseudo-label-based unsupervised domain adaptation medical image segmentation device for the embodiments of this application. Detailed implementation manners

[0065] To make the objectives, technical solutions 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 a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

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

[0067] S101. In the supervised source domain model training stage, 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, generate the mapping of pseudo-labels with uncertainty, information entropy and relative feature distance, and generate reliable pseudo-labels; use the first evidence loss and the first Dice loss between the predicted probability output by the evidence segmentation network and 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, transfer the parameters of the trained source domain model to the target domain model training stage;

[0068] S102. In the unsupervised adaptive target domain model training stage, construct a target domain model composed of a teacher model and a student model initialized with the parameters of the trained source domain model; for the weakly augmented image of the unlabeled medical image in the given target domain, generate reliable pseudo-labels through the teacher model; use the strongly augmented image of the unlabeled medical image in the given target domain and the pseudo-labels generated by the teacher model as input data to train the student model; update the parameters of the teacher model in a moving average manner as the learning progress of the student model; use the target domain model segmentation loss and the second evidence loss to train the target domain model in an unsupervised manner, promote the student model to learn reliable pseudo-labels from the teacher model through the target domain model segmentation loss, and promote the teacher model to generate potential evidence from the predicted labels of the student model through the second evidence loss;

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

[0070] Specifically, in the source domain model training stage, the model is trained using the image labels from the source domain, thereby generating the segmentation output and evidence output of the model. The segmentation loss and evidence loss between the output of the source domain model and the labels are used to optimize the source domain model in a supervised manner. After the pre-training is completed, the source domain model parameters are passed to the target domain model training stage, i.e., the adaptation stage. In the target domain model training stage, only the medical image samples from the target domain and the model parameters passed from the source domain model training stage are allowed to be used for the adaptation of the 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 augmented images of the unlabeled images in the given target domain, the teacher model generates corresponding pseudo-labels and evidence vectors. This process ensures the reliability of the pseudo-labels generated by the teacher model through pixel-level uncertainty estimation. Then, the student target domain model makes full use of the high-quality pseudo-labels to train the strongly augmented version of the target domain data to adapt the network to the representation specific to the target domain data.

[0071] The following will detail a pseudo-label-based source-free unsupervised domain adaptation medical image segmentation method in two parts: pseudo-label generation based on pixel-level uncertainty estimation and teacher-student target domain models based on strong and weak augmentations. The overall model structure is as 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 through uncertainty estimation based on the evidence theory is used to generate reliable pseudo-labels. The pseudo-labels directly generated by the source domain model may not only fail to bring a positive boost 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 on the target domain. How to reduce the unreliability of the pseudo-labels is the key to solving this problem. For this purpose, a method of uncertainty estimation based on the evidence theory is introduced. Through uncertainty estimation, the pseudo-labels are analyzed, and the unreliable pseudo-label predictions are discarded, and the pseudo-labels with high credibility are selected.

[0074] In medical image segmentation tasks, for the uncertainty estimation of pseudo-labels, early uncertainty quantification methods based on Bayesian theory mainly dealt with uncertainties related to class probabilities. However, these methods usually may lead to inaccurate estimations due to insufficient data or imperfect models. There are also some methods such as Monte Carlo inference. Although it can provide a quantitative estimation of uncertainty, it usually requires multiple forward propagations, with high computational costs and difficulty in adapting to dynamically changing data distributions. To solve these problems, the Dempster-Shafer theory is introduced into the field of medical image SFUDA segmentation. As an extension of Bayesian theory, the Dempster-Shafer theory provides a novel way of uncertainty quantification. It can express uncertainty by aggregating evidence from data without prior knowledge. The Dempster-Shafer theory expresses uncertainty by collecting the amount of evidence that supports classifying a sample into a specific class. The Dempster-Shafer theory can quantify the uncertainty of each pixel point, not just the uncertainty of the overall image, which is crucial for the accurate segmentation of medical images. The Dempster-Shafer theory uncertainty estimation method evaluates uncertainty by integrating all necessary information in one forward pass of the model, thus greatly reducing the computational complexity. This method avoids the need for multiple inferences, thus reducing the consumption of computational resources. At the same time, it also avoids the instability of prediction results caused by multiple random Dropout operations in the Monte Carlo method. More importantly, the Dempster-Shafer theory does not require modifying the model structure, such as adding a Dropout layer, thus avoiding unknown impacts on the model performance. By obtaining uncertainty estimation with only one forward pass, the Dempster-Shafer theory method improves the efficiency of the model while maintaining the stability and accuracy of prediction. This makes the Dempster-Shafer theory an efficient and reliable uncertainty estimation method, especially suitable for application scenarios with limited computational resources or high requirements for prediction stability. By using the Dempster-Shafer theory to quantify the uncertainty of target domain samples and helping the network improve the classification accuracy by screening low-uncertainty samples, and then using information entropy to assist in correcting the samples selected based on the Dempster-Shafer theory, more accurate pseudo-labels can be obtained.

[0075] In this embodiment, an Evidence Segmentation Network (ESN) is designed based on the U-net network model. The purpose of this network design is to generate reliable pseudo-labels through the uncertainty estimation of the Dempster-Shafer theory, and at the same time generate pseudo-labels and uncertainty maps. By estimating uncertainty, highly credible pseudo-labels are generated, thus enhancing the ability to generate reliable pseudo-labels from unlabeled target domain data.

[0076] The main difficulty of the Dempster-Shafer theory lies in obtaining evidence. In this embodiment, a transformation function with input is proposed to obtain the evidence vector of the sample, as shown in the following formula.

[0077] ;

[0078] Among them, is the evidence vector, represents the transformation function, is the output of the evidence segmentation network for the i-th pixel of the source domain medical image label ; is a scaling factor, and its value is the reciprocal of the segmentation category. Subjective logic proposes a logical reasoning framework based on belief mass and uncertainty. In subjective logic, belief mass represents the degree of belief in a certain proposition or event, while uncertainty represents the degree of uncertainty about this degree of belief. Here, represents the output of the model for at the i-th pixel in the -th class, and represents the uncertainty predicted by the model for the i-th pixel of . Then, according to subjective logic, the following formula is obtained.

[0079] ;

[0080] represents the number of segmentation categories. Since the belief mass represents the degree of belief of the output of the model for the i-th pixel of in the -th class, and the evidence vector represents the degree of support for the output of the i-th pixel of in the -th class. Then, the following formula can be defined to calculate .

[0081] ;

[0082] Among them, is the total evidence of at the i-th pixel. Based on the above formula, the calculation formula for the uncertainty is as follows:

[0083] .

[0084] Therefore, based on the above formula, it is concluded that the uncertainty is inversely proportional to the total evidence. When there is no evidence, the belief mass of each category is zero, and at this time 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, during the training process, evidence analysis calculates the uncertainty of the pixel prediction according to the total evidence amount of the pixel points, and optimizes the generation of pseudo-labels specific to the target domain based on these confidence distributions to optimize the quality of the pseudo-labels.

[0085] Selecting pixels with low uncertainty through the evidence theory can improve the quality of pseudo-labels. To further enhance the quality of pseudo-labels, in this section, the sample images selected based on the evidence theory are assisted and corrected by information entropy and the feature distance method based on clustering, so as to obtain more accurate pseudo-labels. Information entropy is calculated as shown in the following formula:

[0086] .

[0087] Low-entropy pseudo-labels can be screened out through information entropy, thereby further improving the quality of pseudo-labels. The segmentation problem has structural characteristics, and the segmentation regions of the same category are usually highly correlated. To make full use of this phenomenon to assist in correcting pseudo-labels, a feature distance method based on clustering is proposed. By optimizing the clustering process, the accuracy of pseudo-label generation is improved. The feature distance method based on clustering uses clustering, that is, the class feature pseudo-centroid, to identify and eliminate potential unreliable pseudo-labels, ensuring that the selected pseudo-labels are closer to the feature of their class prototype and improving the quality of pseudo-labels. In the feature distance method based on clustering, first, the class feature pseudo-centroid needs to be calculated. The calculation of the class feature pseudo-centroid is shown in the following formula.

[0088] ;

[0089] where represents the total number of pixels in the label of the source domain medical image; represents the feature map obtained by upsampling the feature map obtained from the last convolution and having the same dimension as the pseudo-label. The relative feature distance between each pixel i and the feature of its corresponding nearest class centroid is calculated by the following formula:

[0090] .

[0091] To use the uncertainty map to filter out noisy pseudo-labels, use information entropy to filter out high-entropy pseudo-labels, and use the relative feature distance to screen out low-distance samples to assist in correcting pseudo-labels, a binary vector label selection mask is defined. The value of the pseudo-label of pixel i is determined by , and as follows:

[0092] ;

[0093] where is the uncertainty threshold, is the low-entropy threshold, is the low-distance threshold. In this embodiment, Set to 0.05, Set to the median of the entropy over all pixels, Set to the median of the relative feature distances over all pixels. If the uncertainty is low enough and the sample has low entropy and low distance, use to select the corresponding pseudo-label; otherwise, the pseudo-label is considered noise and excluded from the loss calculation. Therefore, in the generation of source model pseudo-labels, through the continuous training on the model by Dempster-Shafer theory analysis, the calculations of uncertainty, information entropy, and relative feature distance can be performed periodically to dynamically screen out new high-confidence samples. For the prediction probability of pixel i, where , is the prediction probability of the model for pixel i with respect to the -th class, then the pseudo-label of pixel i argmax .

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

[0095] To quantify the confidence and transform the generated evidence into a probability framework for decision-making, the Dirichlet distribution is used to represent the probability distribution of belief masses. Define the parameters of the Dirichlet distribution and for the prediction probability of pixel i, there is the following Dirichlet distribution.

[0096] ;

[0097] where, is the K-dimensional multinomial beta function. The parameters of this Dirichlet distribution can interpret the evidence for each class. When 's prediction result is related to one of the K classes, the corresponding Dirichlet distribution parameter is incremented to update the Dirichlet distribution with the new observation. Since and follow this Dirichlet distribution, for deep learning work based on quantifying uncertainty, the cross-entropy loss is introduced, as shown in the following equation.

[0098] ;

[0099] where, is the digamma function, is at pixel i with respect to the The label of the class. By optimizing this loss, the model will generate as much evidence as possible for the correct label. However, for uncertain or mispredicted labels, the model may also generate evidence, which makes the uncertain part more difficult to predict. To penalize the uncertainty of the model on unknown labels and thus encourage the model to generate more reliable predictions, the KL divergence between the Dirichlet distribution output by the model and the uniform distribution 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 case of uncertainty or misclassification, so as to improve the overall robustness and generalization ability of the model. Therefore, the loss of KL divergence is specifically shown as the following formula.

[0100] ;

[0101] Among them, is the Dirichlet parameter after removing non-misleading evidence from the prediction result for pixel i, is the uniform Dirichlet distribution, represents the Dirichlet distribution after removing non-misleading evidence from the prediction result. To obtain more evidence, optimize the first evidence loss generated by the model as shown in the following formula.

[0102] ;

[0103] Among them, is the annealing decay coefficient, is the number of epochs of training, is the maximum number of epochs. Through the evidence loss to generate more evidence to reduce uncertainty. But it mainly focuses on the prediction accuracy at the single-pixel level and ignores the crucial spatial relationships between pixels in image segmentation. In image segmentation tasks, the spatial continuity and mutual relationships between pixels are crucial for obtaining high-quality segmentation results. Relying solely on cannot fully capture these relationships, which will lead to inaccurate segmentation boundaries or ignore some fine structures. Therefore, to make up for this, the first Dice loss is introduced as a supplement. Different from the traditional Dice loss, regarding the belief quality after passing through the softmax function of softmax result as , is defined as the following formula:

[0104] .

[0105] In summary, the total loss of the evidence segmentation network of the source domain model proposed in this embodiment is defined as shown in the following formula:

[0106] .

[0107] During the process of optimizing the loss of the evidence segmentation network source model , the model performance is improved through the collaborative optimization of the loss function and the loss function. First, by optimizing the loss function for generating evidence, it is ensured that the model can generate sufficient evidence for each sample, enhancing the confidence and accuracy of the model when processing data. In , the based on KL divergence is used to guide the model to generate less evidence for incorrect samples, which helps the model distinguish samples with insufficient model predictions, so that these samples with higher uncertainty can be given more attention in subsequent training. Next, by optimizing the Dice loss , the spatial relationship between different predicted pixels is strengthened, which is crucial for maintaining the coherence and accuracy of the prediction. Through the correlation between predicted pixels, it can be ensured that the segmentation boundary is more precise, thus being more in line with the real image structure in details. During the process of optimizing , not only the performance of the evidence segmentation network in predicting a single pixel is improved, but also the global consistency and accuracy of the prediction are ensured.

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

[0109] A common challenge in current SFUDA field work is the instability during the training process. In existing methods, the model often highly depends on its self-generated prediction labels that may contain high noise, which can cause instability in the training process. To improve the stability of the training process and continuously enhance the quality of pseudo-labels, in this embodiment, an average teacher model is introduced to simulate unsupervised domain adaptation as supervised learning, and a Teacher Student Model Based on Strong Weak Data augmentation (TSDA) is proposed based on the average teacher model. A dual-model system is constructed, which includes a teacher model and a student model. The teacher model is inherited from the ESN model trained on the source dataset, while the student model adopts a relatively simple U-net model (for the student model, it is supervised-trained on the target domain data and the pseudo-labels of the target domain data provided by the teacher model, and does not need to additionally use the evidence theory uncertainty estimation in ESN, which can save computing resources). The model parameters are all 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 the images with weak data augmentation through evidence theory uncertainty estimation, while the student model is trained with the strongly augmented version of the same image and the reliable pseudo-labels from the teacher model. Among them, the teacher model is not directly trained through traditional backpropagation, but indirectly by updating the moving average of its weights to the weights of the student model, so as to resist the direct impact of noisy pseudo-labels and gradually accumulate the new knowledge learned by the student model.

[0110] The purpose of this embodiment to propose the TSDA framework is to ensure the robustness of the model training process and prevent the continuous accumulation of errors during training. It includes two key models: one is the guiding teacher model , and the other is the student model that performs the learning task , the main responsibility of the teacher model is to provide high-quality pseudo-label guidance for the student model through its predictions, while the student model is dedicated to learning these pseudo-labels to adapt to the target task. The introduced weak-strong augmentation data augmentation strategy is used to enhance the model's resilience and generalization ability in the face of new environments and improve the performance of medical image segmentation. The consistency regularization of the weak-strong augmentation mechanism effectively promotes the model to learn more robust feature representations by simultaneously learning the weakly augmented and strongly augmented versions of the same image. Mapping images that, although visually different due to augmentation, have consistent semantic content to the same region of the feature space not only improves the discriminability of the feature representation but also enables the model to classify correctly more stably in the face of minor changes in medical images or inconsistent scanning quality. Specifically, for each input target-domain medical image, first, mild data augmentation techniques such as image flipping and scaling are applied to generate a weakly augmented version of the image . At the same time, to increase the model's resilience to noise, a strongly augmented version with more drastic transformations is also created , including but not limited to techniques such as random erasing, contrast adjustment, and adding impulse noise. During this process, for each pixel j belonging to the weakly augmented image , a corresponding reliable pseudo-label is generated through the teacher model using the evidence theory-based uncertainty estimation method ESN . Then, these pseudo-labels are used to supervised train the student model , with the input being the corresponding strongly augmented image . This design not only enables the teacher model to utilize its knowledge on the source data to generate more accurate guidance labels but also allows the student model to learn under more challenging conditions through the strongly augmented input, thereby enhancing its generalization and adaptation abilities to unknown data.

[0111] In addition, instead of directly training the teacher model, the weights of the teacher model are updated in a moving average manner (EMA) as the student model progresses in learning. This design helps to balance the stability in the model training process and the accumulation of new knowledge, thus avoiding potential training instability problems caused by the model overfitting to the noisy pseudo-labels of the student model. The update of the teacher model parameters is as follows.

[0112] ;

[0113] where 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 a second segmentation loss function and a second evidence loss function. To combine the advantages of the cross-entropy loss and the Dice loss, a second segmentation loss is designed. The formula is as follows.

[0115] ;

[0116] Where, represents the number of segmentation categories; corresponds to argmax and is the predicted label of the teacher model for the weakly augmented image at pixel point j; corresponds to and is the predicted probability of the student model for the strongly augmented image at pixel point j for the th class.

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

[0118] ;

[0119] Where, is the predicted probability of the teacher model for the weakly augmented 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 augmented data is the predicted label of the teacher model for the weakly augmented data. By combining the overall loss into two parts: the second segmentation loss and the second evidence loss, the student model can learn reliable pseudo-labels from the teacher model through the second segmentation loss , and the teacher model can generate potential evidence from the predicted labels of the student model through the second evidence loss . 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 progress of the student model in turn obtains a better teacher model through EMA, enabling the teacher model to generate more evidence and ensuring the reliability of the pseudo-labels. The two cooperate with each other to overcome the domain difference.

[0120] The following will verify the pseudo-label-based source-free unsupervised domain adaptation medical image segmentation method of this embodiment through experiments, including the following steps.

[0121] (1) Dataset.

[0122] In the field of medical image segmentation, a medical image database is the cornerstone of medical image segmentation. Due to the difficulties in data acquisition and high requirements for data privacy, it is impossible to provide sufficient samples for medical image segmentation research. Therefore, the research and development in medical image segmentation are challenged by small data samples. In this embodiment, the public fundus image dataset and prostate image dataset are mainly used for experiments. There are mainly three fundus image datasets, namely Drishti-GS, RIM-ONE_r3, and REFUGE challenge training sets, and three prostate image datasets, namely NCI, I2CVB, and PROMISE12. These datasets will be introduced separately below.

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

[0124] RIM-ONE_r3 dataset: It is a public fundus image dataset focusing on ophthalmic research, aiming to promote the development of glaucoma detection and optic disc and optic cup segmentation technologies. This dataset collected 159 color fundus 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 ophthalmic experts, including the boundaries of the optic disc and optic cup.

[0125] REFUGE challenge dataset: The 2018 REFUGE challenge provided a set of annotated fundus images for training and validating computer vision and machine learning algorithms. The training set includes 400 color fundus images and corresponding segmentation annotations (320 images for training and 80 images for testing). The REFUGE dataset contains high-quality fundus color photos, which are usually used for the diagnosis and monitoring of glaucoma. These images can show the detailed structures of the retina, optic disc, and macula, which are crucial for the diagnosis of glaucoma. The images in the dataset are attached with expert-annotated information, including the boundary localization of the optic disc and optic cup. These are key indicators for evaluating optic nerve head damage and glaucoma. By calculating the cup-to-disc ratio, glaucoma diagnosis can be assisted.

[0126] NCI-ISBI2013 Dataset: The NCI-ISBI2013 dataset is a prostate magnetic resonance imaging (MRI) data challenge jointly launched by the Cancer Imaging Program of the National Cancer Institute (NCI) and the International Society for Biomedical Imaging (ISBI) in 2013. The training set of the dataset for this challenge includes 60 cases, half of which are from the 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 endocavitary coil was used and the imaging was performed with Philips equipment. The other half of the cases were provided by the 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 was used and the imaging was performed with Siemens equipment.

[0127] I2CVB Dataset: This dataset is from the Initiative for Collaborative Computer Vision 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 endocavitary coil was used and the imaging was performed with Siemens equipment.

[0128] PROMISE1 Dataset: This dataset is from Prostate MR Image Segmentation 2012 (PROMISE12), provided by the UCL institution, the BIDMC institution and the HK institution, with 13, 12 and 12 cases respectively. The dataset used in the experiment of this example is the data provided by the HK institution, 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. An endocavitary coil was used and the imaging was performed with Siemens equipment.

[0129] (2) Data Preprocessing

[0130] Specifically, there are the following problems if the entire medical image is directly used for training: Background information in the original image, such as edge regions and non-target structures, may introduce noise, affecting the model's ability to identify key features and resulting in a decrease in the model's resolution. Therefore, this paper introduces the concept of ROI (Region of Interest). The ROI image reduces background interference by focusing on the segmentation region, enabling the model to more attentively learn the details of these key regions, thereby improving the accuracy of segmentation.

[0131] Due to the limitation of the small data volume of the medical image dataset, the model directly trained with the medical image dataset is prone to problems such as overfitting and poor generalization ability. To improve the model training effect, this embodiment adopts a series of data augmentation strategies to simulate different visual effects and expand the training dataset. By image inversion, rotation, and adjustment of brightness and contrast, the diversity of samples can be artificially increased, which helps the model to have better generalization ability in actual applications. After these enhanced images, for fundus images, they are fixed and scaled to dimensions, and for the prostate image dataset, each sample is adjusted to dimensions in the axial plane, ensuring the consistency of data during the training process. At the same time, random flipping and normalization processing are performed on the images, so that the pixel values of each channel have a unified average value and standard deviation, further reducing the risk of overfitting during the model training process. Through these preprocessing steps, the number of images in the training dataset increases and the quality improves, ultimately solving the limitation of the small data volume of the medical image dataset that leads to difficulties in training the model.

[0132] (3) Evaluation metrics

[0133] 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 metrics in image segmentation evaluation: Dice coefficient (Dice) and average surface distance (ASD). These metrics together provide a comprehensive evaluation for the quantification of the 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 prediction and the true segmentation to the sum of their respective areas. The value range of the Dice coefficient is from 0 to 1, and the higher the value, the closer the segmentation result is to the true situation, that is, the better the segmentation effect. The calculation formula is as follows.

[0134] ;

[0135] where A is the set of predicted segmentation results, B is the set of true segmentation results, 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 all points on the predicted segmentation to the true segmentation surface, and the average shortest distance from all points on the true segmentation to 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 respectively represent the point sets on the predicted segmentation surface and the ground truth segmentation surface. denotes the point and the point The Euclidean distance between. 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 detailed training parameters of this experiment are as follows:

[0142] Set the uncertainty threshold to 0.05, the information entropy threshold is set to the median of the entropy on all pixel points, and the relative feature distance threshold is set to the median of the relative feature distance on all pixel points. The initial learning rate is set to 0.001 and decreased every 10 epochs . The batchsize is set to 8, and the source domain model is obtained after 100 epochs of training. In the unsupervised domain adaptation stage, first initialize the teacher and student models from the source model. Update the teacher model parameters from the student model parameters using EMA, and the model EMA update rate is set to 0.98. Set the learning rate to 5e -4 The final target domain student model is obtained after 50 epochs. The model is trained using the Adam optimizer with momentum 0.9 and 0.99.

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

[0144]

[0145] (5) Experimental design

[0146] The experiment in 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 by the present invention, and whether it has advantages compared with other UDA medical image segmentation methods, and the results can be intuitively seen through the visualized segmentation results.

[0148] Second, in the SFUDA medical image segmentation scenario, the impact of the ESA and TSDA modules proposed by the present invention on the model performance. In this embodiment, under the condition of using the same dataset and training process, the effectiveness of each module is verified by designing four groups of ablation experiments.

[0149] Third, visually verify the effect of the proposed AOPL method on improving the quality of pseudo-labels through uncertainty estimation. The effectiveness of AOPL in improving the quality of pseudo-labels is visually verified by presenting and analyzing high-noise pseudo-labels, uncertainty maps, and pseudo-labels obtained by the AOPL method.

[0150] (6) Experimental Results and Analysis

[0151] (6.1) Analysis of Fundus Image Segmentation Experiment

[0152] Specifically, publicly available fundus image datasets are mainly used for experiments, including Drishti-GS, RIM-ONE_r3, and REFUGE challenge training sets, three fundus image datasets. To verify the segmentation performance of the model on the target domain dataset when passive domain data is available, 2 domain transfer cases are 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. To investigate the effectiveness of the model in this embodiment, an algorithm comparison experiment is carried out. The comparison experiment is divided into: an experiment comparison of active unsupervised domain adaptation methods to test whether the method of the present invention is inferior to the active domain adaptation method in performance; an experiment comparison with the SFUDA method to test whether the method of the present invention achieves the best results compared with similar methods. In the comparison experiment of active domain adaptation methods, the BEAL method and the pOSAL method are used as comparison algorithms. In the comparison experiment of the SFUDA method, the DPL method, the OS method, and the SFDA method are used as comparison algorithms. Among them, DPL is a passive unsupervised domain adaptation method of a pseudo-label generation method using Monte Carlo for uncertainty estimation, OS is a passive unsupervised domain adaptation method using batch normalization layer adaptation 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 transfer method that directly predicts the target domain result using the benchmark U-net model trained from the source domain data. Each method is run 5 times, and then the average value of the evaluation criteria for measuring the segmentation effect described above 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 terms of both the Dice coefficient and the ASD distance. It exceeds other UDA methods in the optic disc Dice coefficient and the optic cup Dice coefficient. It is worth noting that AOPL achieved the best results in the optic cup ASD, and is also better than other UDA methods in the optic disc ASD.

[0154] Table 2 Comparison of the results of different algorithms on the 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 metrics, and is better than all other methods in the optic disc ASD. It is second only to the BEAL method in the optic cup Dice coefficient and the optic cup ASD distance. Considering the evaluation metrics of the above two domain transfer experiments, it can be seen that the AOPL method can not only well solve the SFUDA problem in fundus image segmentation without available source domain data, but also outperform some UDA methods, which proves the effectiveness of AOPL.

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

[0158]

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

[0160] To verify the generality of the AOPL method on medical image datasets, three prostate image datasets, namely NCI-ISBI 2013, I2CVB, and PROMISE12, were further verified. Design: (1) The I2CVB dataset was used as the source domain and NCI-ISBI 2013 as the target domain; (2) The I2CVB dataset was used as the source domain and PROMISE12 as the target domain for two domain transfer situations. In this embodiment, only the source domain model and unlabeled data in the target domain are available. The comparison methods are the same as above, and the BEAL method, pOSAL method, DPL method, OS method, SFDA method, and Baseline are used as comparison algorithms. The average value of five experiments was also used as the comparison evaluation criterion.

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

[0162] Table 4 Comparison of the results of different algorithms on the 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 the results of different algorithms on the I2CVB dataset → ROMISE12 dataset;

[0166]

[0167] It can be seen that among the various methods of SFUDA, the overall performance is also less than the domain transfer results of the I2CVB dataset → NCI-ISBI 2013, further indicating that the magnitude of domain shift seriously affects the model segmentation performance. The AOPL method also performs the best among all SFUDA methods. In comparison with the UDA method, both the Dice coefficient and ASD distance are better than other methods and are very close. Considering the performance of the AOPL method in the two domain transfer experiments on the prostate dataset, it can be seen that AOPL can not only well solve the SFUDA problem in the prostate dataset, but also exceed the performance of some other UDA methods. It is proved that the AOPL method can still effectively solve the domain adaptation problem with less data used.

[0168] Figure 3 and Figure 4 respectively show the experimental results of different algorithms under the domain adaptation of fundus images and prostate images. It can be seen from the figure that if no method is used to shorten the domain shift and direct migration is carried out, very poor segmentation results will be obtained. Whether it is the SFUDA method or the UDA method, the domain shift is shortened and the segmentation ability is significantly improved. However, the segmentation shapes and boundaries of the OS and SFDA methods are not very good, while the AOPL method of the present invention performs well in terms of segmentation shape and boundary, and the effect exceeds other SFUDA methods, and the prediction results are already close to the true labels.

[0169] As described above, the AOPL method proposed by the present invention, through comparative experiments, shows significant advantages in the SFUDA scenario of fundus image datasets and prostate image datasets. Through the comparison of Dice scores and ASD distances, it is proved that the AOPL method has improved performance compared with other SFUDA techniques. In addition, the superiority of the segmentation effect of the AOPL method can also be clearly seen from the visualization images of the segmentation results. To sum up, the AOPL method effectively solves the SFUDA problem in the field of medical image segmentation by taking advantage of the uncertainty estimation of evidence theory and consistency learning.

[0170] (6.3)Ablation experiment

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

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

[0173]

[0174] Table 6 shows the segmentation results on the REFUGE challenge training set ( Figure 5 ), → Drishti-GS ( Figure 6 ) dataset under different strategies. It can be seen that, compared with the directly transferred Baseline method, the Dice coefficients and ASD distances of ESN and TSDA have been greatly improved. The segmentation performance of ESN has a greater improvement than that of TSDA because ESN can eliminate high-uncertainty prediction regions through the uncertainty estimation of evidence theory and obtain better prediction labels for the segmentation results. Although TSDA can improve the segmentation performance by emphasizing the consistency of prediction results through strong-weak data augmentation with consistency loss, due to the domain shift in the target domain, there are a large number of noises in the generated prediction results. At this time, the teacher-student model will continuously accumulate errors because it keeps learning high-noise prediction labels, resulting in a decline in segmentation performance. Therefore, AOPL (Baseline + ESN + TSDA) combines the advantages of ESN and TSDA, produces high-quality pseudo-labels, and continuously optimizes the segmentation process through the teacher-student model, continuously screening low-uncertainty segmentation regions during this process, thereby achieving good segmentation performance.

[0175] (6.4)Other experiments

[0176] To further demonstrate the effectiveness of the evidence theory uncertainty estimation method proposed in the present invention, the following shows the high-noise pseudo-labels, uncertainty maps, pseudo-labels after reducing noise using the uncertainty estimation method, and image labels directly generated on the target domain images during the model training process, where Figure 7 (a) represents the high-noise pseudo-labels generated on the target domain images, Figure 7 (b) represents the uncertainty maps, Figure 7 (c) represents the pseudo-labels with improved quality using the method of the present invention, Figure 7 (d) represents the image labels. It can be seen from the figure that there are obvious noise problems in the predicted pseudo-labels directly obtained on the target domain images compared with the true labels. The highlighted areas in the uncertainty maps represent high-uncertainty regions, which usually correspond to segmentation boundaries and backgrounds. When the background is misclassified as a segmentation label, the uncertainty is very high, and at this time, it can be well reclassified as the background to improve the quality of the pseudo-labels. Although the uncertainty of the segmentation boundaries 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-labels obtained after being processed by the AOPL method can effectively handle the boundaries, and their quality is significantly improved and is already close to the true image labels.

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

[0178] A supervised source domain model training module 801, which is used to construct an evidence segmentation network based on the U-net network model as the source domain model; based on the evidence theory uncertainty estimation method, generate the mapping of pseudo-labels with uncertainty, information entropy, and relative feature distance, and generate reliable pseudo-labels; use the first evidence loss and the first Dice loss between the predicted probability output by the evidence segmentation network and 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, transfer 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 the parameters of the trained source domain model; for the weakly augmented images of the unlabeled medical images in the given target domain, reliable pseudo-labels are generated through the teacher model; the strongly augmented images of the unlabeled medical images in the given target domain and the pseudo-labels generated by the teacher model are used as input data to train the student model; 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 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;

[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 pseudo-label-based source-free unsupervised domain adaptation medical image segmentation device is the same as that of a pseudo-label-based source-free unsupervised domain adaptation medical image segmentation method, and will not be repeated in this embodiment.

[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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and 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; The medical image label of the source domain is the pseudo label of the i-th pixel It is expressed as follows: in, is the predicted probability of pixel i, is the predicted probability of the source domain model for pixel i about the kth category, K represents the number of segmentation categories; p i =F(x i ), F(x i ) is the output of the evidence segmentation network for the i-th pixel of the source domain medical image label x; m i ∈{0,1} is a label selection mask for a binary vector, expressed as follows: m i =1[u i <η1]×1[ent i <η2]×1[d i <η3]; Among them, u i Indicates uncertainty; i represents information entropy; d i Represents the relative feature distance; η1 is the uncertainty threshold, η2 is the low entropy threshold, and η3 is the low distance threshold; The first Dice loss is L eDice , which is expressed as follows: Among them, K represents the number of segmentation categories; is the label of the source domain medical image label x at pixel i about the kth category; Belief quality b i After the softmax function, softmax(b i )result.

2. The pseudo-label based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: Uncertainty i The calculation formula is as follows: Among them, the quality of belief represents the degree of belief of the source domain model on the output of the i-th pixel of the source domain medical image label x in the k-th category; s i is all the evidence for the source domain medical image label x at the i-th pixel; Information entropy i The calculation formula is as follows: The relative feature distance d between each pixel i and its corresponding nearest class centroid feature i It is expressed as follows: Class feature pseudo centroid z k The calculation formula is as follows: Where N represents the total number of pixels of the source domain medical image label; f i 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.

3. The pseudo-label based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: First evidence loss L evg Including the first cross entropy loss L ece and KL divergence loss L KL , which is expressed as follows: L evg =L ece +λ t L KL ; Among them, λ t =min(1.0,t / T max ) is the annealing attenuation coefficient, t is the number of epochs for training, T max is the maximum number of epochs; The first cross entropy loss L ece , which is expressed as follows: Where K represents the number of segmentation categories; ψ(·) is the double gamma function, is the predicted probability of the source domain model for pixel i regarding the kth category; is the label of the source domain medical image label x at pixel i about the kth category; S i is all the evidence for the source domain medical image label x at the i-th pixel; is the Dirichlet distribution parameter; KL divergence loss L KL , which is expressed as follows: Among them, p i is the predicted probability of pixel i; is the Dirichlet parameter after removing non-misleading evidence from the prediction at pixel i, D(p i |1) is a uniform Dirichlet distribution, represents the Dirichlet distribution after removing non-misleading evidence from the prediction results, y i is the label of pixel i.

4. 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, θ are the parameters of the student model, and ρ is the update rate.

5. The pseudo-label-based passive unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that: The overall loss L after combining the target domain model segmentation loss and the second evidence loss tar , which is expressed as follows: Among them, L′ evg Indicates the loss of second evidence; L sc represents the target domain model segmentation loss; M t (x weak ) represents the predicted probability of the teacher model for the weakly enhanced image, argmax(M s (x strong )) represents the predicted label of the student model for the strongly enhanced image, M s (x strong ) represents the predicted probability of the student model for the strongly enhanced image, argmax(M t (x weak )) represents the predicted label of the teacher model for the weakly enhanced image.

6. The pseudo-label-based passive unsupervised domain adaptation medical image segmentation method according to claim 5, characterized in that: Target domain model segmentation loss L sc , which is expressed as follows: Among them, L ce represents the second cross entropy loss, L Dice Represents the second Dice loss, which are as follows: Among them, K represents the number of segmentation categories; With argmax(M t (x weak )) corresponds to the predicted label of the teacher model for the weakly enhanced image at pixel j; With M s (x strong ) corresponds to the predicted probability of the student model for the strongly enhanced image at pixel j regarding the kth class.

7. 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. A segmentation module, used to segment the medical image to be processed using the trained target domain model; The medical image label of the source domain is the pseudo label of the i-th pixel It is expressed as follows: in, is the predicted probability of pixel i, is the predicted probability of the source domain model for pixel i about the kth category, K represents the number of segmentation categories; p i =F(x i ), F(x i ) is the output of the evidence segmentation network for the i-th pixel of the source domain medical image label x; m i ∈{0,1} is a label selection mask for a binary vector, expressed as follows: m i =1[u i <η1]×1[ent i <η2]×1[d i <η3]; Among them, u i Indicates uncertainty; i represents information entropy; d i Represents the relative feature distance; η1 is the uncertainty threshold, η2 is the low entropy threshold, and η3 is the low distance threshold; The first Dice loss is L eDice , which is expressed as follows: Among them, K represents the number of segmentation categories; is the label of the source domain medical image label x at pixel i about the kth category; Belief quality b i After the softmax function, softmax(b i )result.

Images