Semi-supervised three-dimensional medical image segmentation method, system, device and storage medium based on double correction mutual learning

Abstract

The application discloses a semi-supervised three-dimensional medical image segmentation method, system and device based on double correction mutual learning and a storage medium, belongs to the technical field of medical image processing, and solves the problems that existing methods have model cognitive bias and cannot obtain reliable pseudo labels. A labeled three-dimensional medical image is used to train two networks with different structures, a supervision loss of the labeled three-dimensional medical image is calculated, a prediction difference between the two networks with different structures is corrected, and two pre-training networks with different structures are obtained; a no-labeled three-dimensional medical image is used to generate a preliminary prediction pseudo mask corresponding to the no-labeled three-dimensional medical image through the two pre-training networks with different structures, the preliminary prediction pseudo mask is corrected, and a reliable pseudo label corresponding to the preliminary prediction pseudo mask is obtained; a cross pseudo supervision loss based on the no-labeled three-dimensional medical image is used to obtain final two networks with different structures; and a three-dimensional medical image to be processed is input into the final two networks with different structures to implement a three-dimensional medical image segmentation task.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a semi-supervised three-dimensional medical image segmentation method, system and device based on double correction mutual learning and a storage medium. BACKGROUND

[0002] Anatomical structure and lesion segmentation in medical images plays an important role in clinical disease diagnosis, radiotherapy planning, surgical guidance and prognosis evaluation. In recent years, supervised learning technology based on deep convolutional neural network has made remarkable achievements in medical image segmentation field. However, full supervised learning method needs to rely on a large number of pixel-level label images to train network model, which has certain limitations for clinical three-dimensional medical image segmentation task with a small amount of pixel-level annotation data. Semi-supervised learning can use a small amount of labeled data and a large amount of unlabeled data to train network model, effectively alleviating the dependence of model on pixel-level annotation data, and more in line with the needs of clinical application scenarios in three-dimensional medical image segmentation.

[0003] In semi-supervised medical image segmentation, the mainstream methods are: pseudo-label based method and consistency regularization method. The idea of pseudo-label is to use the pre-trained model of the labeled data, then use the model to generate the pseudo-label of the unlabeled data, and retrain the model with the unlabeled data. The idea of consistency regularization is to encourage the network to generate consistent predictions under different perspectives or perturbations. In addition, in order to improve the segmentation performance of the model, some methods introduce cross pseudo-supervision strategy between networks to combine the two methods. There are also some methods, for example, Chinese patent CN118710912A discloses a semi-supervised image segmentation method, which adopts mutual learning strategy in the segmentation process, so that the teacher model can learn knowledge from the labeled data like the student model, thereby alleviating the problem of error accumulation in the teacher model.

[0004] Although these methods have achieved certain results, they still face the same challenges. On the one hand, the segmentation performance of the model depends on the quality of the pseudo-label, so how to improve the quality of the pseudo-label and obtain reliable pseudo-label is a key problem to be solved. On the other hand, due to the cognitive bias of network feature learning, this bias will intensify during model training and is difficult to correct, so how to use the difference between the predictions of the networks to guide the self-correction of the network is another key problem to be solved. SUMMARY

[0005] The present application solves the problem that the existing method has model cognitive bias and cannot obtain reliable pseudo-label.

[0006] The semi-supervised three-dimensional medical image segmentation method based on double correction mutual learning according to the present application, specifically comprises:

[0007] preprocessing a three-dimensional medical image, the three-dimensional medical image comprising a labeled three-dimensional medical image and an unlabeled three-dimensional medical image;

[0008] training two networks with different structures respectively by using the labeled three-dimensional medical image, calculating a supervised loss of the labeled three-dimensional medical image, correcting a prediction difference between the two networks with different structures, and obtaining two pre-training networks with different structures respectively;

[0009] generating a preliminary prediction pseudo mask corresponding to the unlabeled three-dimensional medical image by using the two pre-training networks with different structures respectively, correcting the preliminary prediction pseudo masks generated by the two pre-training networks with different structures, and obtaining reliable pseudo labels corresponding to the unlabeled three-dimensional medical image;

[0010] implementing mutual learning between the two pre-training networks with different structures based on a cross pseudo supervision loss of the unlabeled three-dimensional medical image, updating parameters of the two pre-training networks with different structures, and obtaining two final networks with different structures respectively;

[0011] inputting a three-dimensional medical image to be processed into the two final networks with different structures respectively to implement a three-dimensional medical image segmentation task.

[0012] Further, in an embodiment of the present application, the two networks with different structures are an uncertainty-aware three-dimensional segmentation network and a structure-enhanced three-dimensional segmentation network respectively;

[0013] The decoder end of the uncertainty-aware three-dimensional segmentation network adopts an uncertainty-aware module, and the uncertainty-aware module specifically comprises:

[0014]

[0015] wherein, u i is the overall uncertainty of the sample i, is the trust quality of the kth class in the sample i;

[0016] The encoder end of the structure-enhanced three-dimensional segmentation network adopts a structure-enhanced module, and the structure-enhanced module specifically comprises:

[0017]

[0018] wherein, f ψ , f x and are 1x1x1 convolutional layers respectively, x i and are feature maps of the ith layer of the encoder and the decoder respectively, and σ is a sigmod activation function. The i-th layer encoder feature map after enhancing the boundary contour information of the target region, and alpha is a feature map containing structural enhancement information.

[0019] Further, in an embodiment of the present application, the prediction difference between the two networks with different structures is corrected, specifically:

[0020] The two binary mask maps of the output predicted by the two networks with different structures are subjected to XOR operation to obtain the inconsistent part mask predicted by the two networks with different structures;

[0021] Based on the inconsistent part mask predicted by the two networks with different structures, the prediction difference region of the two networks with different structures is calculated;

[0022] The correction loss of the prediction difference region of the two networks with different structures and the corresponding real mask region is calculated to guide the two networks with different structures to correct the wrong prediction.

[0023] Further, in an embodiment of the present application, the inconsistent part mask predicted by the two networks with different structures is obtained, specifically:

[0024]

[0025] Wherein, and are the softmax outputs of the labeled three-dimensional medical image through the two networks with different structures, is the XOR operation, BINA is the binary operation, M diff is the inconsistent part mask predicted by the two networks with different structures after the labeled three-dimensional medical image passes through the two networks with different structures;

[0026] The prediction difference region of the two networks with different structures is calculated, specifically:

[0027]

[0028] Wherein, is the prediction difference region of the two networks with different structures, includes and CLIP is the operation of obtaining the prediction difference region corresponding to the inconsistent part mask region;

[0029] The correction loss of the prediction difference region of the two networks with different structures and the corresponding real mask region is calculated, specifically:

[0030]

[0031] Wherein, To predict the true mask image corresponding to the discrepancy region, MSE is the mean squared error loss function. The correction loss is used to compare the predicted difference regions of two networks with their corresponding ground truth mask regions.

[0032] Furthermore, in one embodiment of the present invention, the preliminary predicted pseudo-masks generated by two pre-trained networks with different structures are both corrected to obtain reliable pseudo-labels, specifically as follows:

[0033] The two pre-trained networks with different structures are a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network.

[0034] The pre-trained uncertainty-aware 3D segmentation network generates a preliminary prediction pseudo-mask, which is then corrected by uncertainty awareness to obtain a reliable pseudo-label.

[0035] The pre-trained structure-enhanced 3D segmentation network generates a preliminary predicted pseudo-mask, which is then corrected by dynamically generating a high-confidence mask to obtain a reliable pseudo-label.

[0036] Furthermore, in one embodiment of the present invention, the uncertainty perception correction specifically includes:

[0037]

[0038] in, Predictive probability map for label-free 3D medical images The binary image P obtained by the argmax function A These are reliable pseudo-labels generated after uncertainty-aware correction. For uncertainty graphs, This is an indicator function, and T = 0.2 is the threshold.

[0039] The aforementioned dynamic generation of high-confidence mask correction specifically refers to:

[0040]

[0041] Among them, C A and C B These are the confidence score maps predicted by a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network, respectively, where ∈ is a very small constant. and These are the probabilities of a pixel belonging to the k-th class. This is a high-confidence mask image. Predictive probability map for label-free 3D medical images The binary image P obtained by the argmax function B These are reliable pseudo-labels generated after dynamic generation and high-confidence mask correction.

[0042] Furthermore, in one embodiment of the present invention, the supervision loss specifically refers to:

[0043]

[0044] in, To monitor losses, The supervised loss for pre-training uncertainty-aware 3D segmentation networks, The supervised loss for the pre-trained structure-enhanced 3D segmentation network, The correction loss is used to compare the predicted difference regions of two networks with their corresponding ground truth mask regions. g represents the operation for calculating uncertainty. This is a preliminary predicted probability map obtained from a labeled 3D medical image using an uncertainty-aware 3D segmentation network. Y is the preliminary predicted probability map obtained from a structure-enhanced 3D segmentation network for a labeled 3D medical image;

[0045] The aforementioned cross-supervision loss is specifically as follows:

[0046]

[0047] in, For cross-supervision loss, P A For reliable pseudo-labels generated after uncertainty-aware correction, P B This is a reliable pseudo-label generated after dynamic high-confidence mask correction. This is the cross-entropy loss function.

[0048] The semi-supervised 3D medical image segmentation system based on dual-calibration mutual learning described in this invention is specifically as follows:

[0049] Preprocessing is performed on the three-dimensional medical images, which include labeled three-dimensional medical images and unlabeled three-dimensional medical images;

[0050] Two networks with different structures were trained using labeled 3D medical images. While calculating the supervised loss of the labeled 3D medical images, the prediction difference between the two networks with different structures was corrected, resulting in two pre-trained networks with different structures.

[0051] Unlabeled 3D medical images are processed by two pre-trained networks with different structures to generate corresponding preliminary prediction pseudomasks. The preliminary prediction pseudomasks generated by the two pre-trained networks with different structures are corrected to obtain corresponding reliable pseudo labels.

[0052] Based on the cross-pseudo-supervised loss of unlabeled 3D medical images, mutual learning between two pre-trained networks with different structures is achieved, the parameters of the two pre-trained networks with different structures are updated, and the final two networks with different structures are obtained respectively.

[0053] The three-dimensional medical images to be processed are input into two networks with different final structures to achieve the three-dimensional medical image segmentation task.

[0054] The electronic device of the present invention includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0055] Memory, used to store computer programs;

[0056] When a processor executes a program stored in memory, it implements the steps described in Method 1 above.

[0057] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the steps described above.

[0058] This invention solves the problems of existing methods having model cognitive bias and failing to obtain reliable pseudo-labels. Specific beneficial effects include:

[0059] 1. The semi-supervised 3D medical image segmentation method based on dual-correction mutual learning described in this invention relies on the quality of pseudo-labels for its segmentation performance, thus requiring improvement in pseudo-label quality. To address this technical problem, this invention corrects the generated preliminary predicted pseudo-mask to obtain reliable pseudo-labels for unlabeled 3D medical images processed through two networks with different structures.

[0060] 2. The semi-supervised 3D medical image segmentation method based on dual-calibration mutual learning described in this invention faces the challenge of cognitive bias in existing network feature learning. This bias tends to worsen and become difficult to correct during model training. Therefore, guiding the network to self-correct is another key issue to be addressed. To solve this problem, this invention utilizes a difference correction module to extract the predicted difference information of labeled images obtained from two networks with different structures. Based on the difference generated by this module, the two networks are guided to self-correct, reducing cognitive bias during model training.

[0061] 3. The semi-supervised 3D medical image segmentation method based on dual correction and mutual learning described in this invention improves the performance of semi-supervised 3D medical image segmentation by using dual correction and mutual learning between sub-networks through the supervision loss of labeled images and the cross pseudo-supervision loss of unlabeled images. Attached Figure Description

[0062] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0063] Figure 1 This is a flowchart illustrating the semi-supervised three-dimensional medical image segmentation method based on dual-calibration mutual learning as described in Implementation Method 1.

[0064] Figure 2 This is a schematic diagram of the network structure described in Implementation Method 2;

[0065] Figure 3 This is a schematic diagram comparing the model trained on the LA dataset with 20% labels as described in Implementation Method 5 with other methods. Detailed Implementation

[0066] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0067] Implementation Method 1: The semi-supervised 3D medical image segmentation method based on dual-calibration mutual learning described in this implementation method is as follows:

[0068] Preprocessing is performed on the three-dimensional medical images, which include labeled three-dimensional medical images and unlabeled three-dimensional medical images;

[0069] Two networks with different structures were trained using labeled 3D medical images. While calculating the supervised loss of the labeled 3D medical images, the prediction difference between the two networks with different structures was corrected, resulting in two pre-trained networks with different structures.

[0070] Unlabeled 3D medical images are processed by two pre-trained networks with different structures to generate corresponding preliminary prediction pseudomasks. The preliminary prediction pseudomasks generated by the two pre-trained networks with different structures are corrected to obtain corresponding reliable pseudo labels.

[0071] Based on the cross-pseudo-supervised loss of unlabeled 3D medical images, mutual learning between two pre-trained networks with different structures is achieved, the parameters of the two pre-trained networks with different structures are updated, and the final two networks with different structures are obtained respectively.

[0072] The three-dimensional medical images to be processed are input into two networks with different final structures to achieve the three-dimensional medical image segmentation task.

[0073] In existing technologies, on the one hand, the segmentation performance of the model depends on the quality of the pseudo-labels, so how to obtain reliable pseudo-labels is a key problem to be solved. On the other hand, due to the existence of cognitive biases in network feature learning, these biases will continue to intensify and are difficult to correct during model training, so how to guide the network to self-correct is another key problem to be solved.

[0074] To solve the above technical problems, such as Figure 1 As shown, this embodiment proposes a semi-supervised 3D medical image segmentation method based on dual-calibration mutual learning, specifically as follows:

[0075] Load the 3D medical image and preprocess it. Specifically, normalize the 3D medical image volume to zero mean and unit variance, and crop the 3D medical image volume with expanded edges according to the target region location to make its size 112×112×80.

[0076] The three-dimensional medical images include labeled three-dimensional medical images and unlabeled three-dimensional medical images;

[0077] Two networks with different structures were trained using labeled 3D medical images. While calculating the supervised loss of the labeled 3D medical images, the prediction difference between the two networks with different structures was corrected, resulting in two pre-trained networks with different structures.

[0078] Unlabeled 3D medical images are processed by two pre-trained networks with different structures to generate corresponding preliminary prediction pseudomasks. The preliminary prediction pseudomasks generated by the two pre-trained networks with different structures are corrected to obtain corresponding reliable pseudo labels.

[0079] Based on the cross-pseudo-supervised loss of unlabeled 3D medical images, mutual learning between two pre-trained networks with different structures is achieved, the parameters of the two pre-trained networks with different structures are updated, and the final two networks with different structures are obtained respectively.

[0080] The three-dimensional medical image is fed into two pre-trained networks with different structures. By automatically extracting features from the image, the final segmentation result of the three-dimensional medical image is obtained.

[0081] This implementation utilizes labeled 3D medical images to correct prediction discrepancies between two networks with different structures, achieving network self-correction and reducing cognitive bias issues in the model. Furthermore, by correcting the generated preliminary prediction pseudo-masks, reliable pseudo-labels are obtained for unlabeled 3D medical images processed by the two networks with different structures, enabling reliable mutual learning between the two networks and improving the performance of semi-supervised 3D medical image segmentation.

[0082] Implementation Method 2: This implementation method further defines the semi-supervised 3D medical image segmentation method based on dual-correction mutual learning described in Implementation Method 1. The two networks with different structures are an uncertainty-aware 3D segmentation network and a structure-enhanced 3D segmentation network.

[0083] The decoder of the uncertainty-aware 3D segmentation network employs an uncertainty-aware module, which specifically includes:

[0084]

[0085] Among them, u i Let i be the population uncertainty of sample i. Let k be the trust quality of class k in sample i;

[0086] The encoder end of the structure-enhanced 3D segmentation network employs a structure enhancement module, which specifically includes:

[0087]

[0088] Among them, f ψ f x and They are 1×1×1 convolutional layers, x i and Let be the feature maps of the encoder and decoder at the i-th layer, respectively, and σ be the sigmoid activation function. The i-th layer encoder feature map is used to enhance the boundary contour information of the target region, and α is a feature map containing structural enhancement information.

[0089] In this embodiment, such as Figure 2 As shown, the two networks with different structures are the uncertainty-aware 3D segmentation network 3D-ResVnet and the structure-enhanced 3D segmentation network Vnet.

[0090] The aforementioned uncertainty-aware 3D segmentation network is specifically as follows:

[0091] An uncertainty-aware module is used at the decoder end of the uncertainty-aware 3D segmentation network to simultaneously output the network's predicted segmentation probability map and the corresponding uncertainty map.

[0092] The uncertainty perception module is specifically as follows:

[0093] For an image segmentation task containing K classes (including background), the trust quality and overall uncertainty are used to describe it, specifically:

[0094]

[0095] Among them, u i The population uncertainty (u) of sample i i ≥0), Trust quality of class k in sample i

[0096] The u i and Specifically:

[0097]

[0098] in, For the generated k-class evidence vector e i , S i The total evidence for the prediction of K classes for sample i;

[0099] Evidence refers to the degree of support for classifying a particular pixel in an image into a certain class. The total evidence is inversely proportional to the overall uncertainty; if there is no evidence, then... Then the trust quality of each class prediction is 0, and the overall uncertainty is 1. Conversely, if the total evidence is large, the overall uncertainty is u. i The probability will be very small, and the network's predictions have a higher confidence level;

[0100] The evidence vector e i Specifically:

[0101]

[0102] in, The sample i is processed by an uncertainty-aware 3D segmentation network to obtain the preliminary prediction result. τ is the scaling parameter, and 0 < τ < 1. tanh is the hyperbolic tangent function, and the range is defined as [-1, 1].

[0103] For the quality of trust The parameters of the Dirichlet distribution function are respectively Correspondingly, specifically:

[0104]

[0105] in,

[0106] The structure-enhanced 3D segmentation network is specifically as follows:

[0107] In the encoder of the structure-enhanced 3D segmentation network, a structure enhancement module is employed to enhance the structural information in the feature map and highlight the boundary contours of the target region in the feature map. Specifically:

[0108]

[0109] Among them, f ψ f x and They are 1×1×1 convolutional layers, x i and Let be the feature maps of the encoder and decoder at the i-th layer, respectively, and σ be the sigmoid activation function. The i-th layer encoder feature map is used to enhance the boundary contour information of the target region, and α is a feature map containing structural enhancement information.

[0110] Implementation Method 3: This implementation method further defines the semi-supervised 3D medical image segmentation method based on dual-correction mutual learning described in Implementation Method 1. Specifically, the correction of prediction differences between two networks with different structures involves:

[0111] The binarized mask images predicted by two networks with different structures are XORed to obtain the mask of the inconsistent part predicted by the two networks with different structures.

[0112] Based on the mask of the inconsistency in the predictions between two networks with different structures, the prediction difference region between the two networks with different structures is calculated.

[0113] Calculate the correction loss of the predicted difference region and its corresponding true mask region for two networks with different structures, and guide the two networks with different structures to correct the erroneous predictions.

[0114] In this embodiment, obtaining the mask of the inconsistency portion predicted between two networks with different structures specifically involves:

[0115]

[0116] in, and These are the softmax outputs of labeled 3D medical images processed by two networks with different structures. For XOR operation, BINA is the binarization operation, and M is the XOR operation. diff The mask for the inconsistent portions of a labeled 3D medical image after prediction by two networks with different structures;

[0117] The calculation of the prediction difference region between the two networks with different structures is specifically as follows:

[0118]

[0119] in, This represents the region of difference in predictions between two networks with different structures. include and CLIP is an operation to obtain the predicted difference region corresponding to the masked region of the inconsistency.

[0120] The calculation of the correction loss for the predicted difference region and its corresponding ground truth mask region between two networks with different structures is specifically as follows:

[0121]

[0122] in, To predict the true mask image corresponding to the discrepancy region, MSE is the mean squared error loss function. The correction loss is used to compare the predicted difference regions of two networks with their corresponding ground truth mask regions.

[0123] In this embodiment, the correction of prediction differences between two networks with different structures specifically includes:

[0124] The binarized mask images predicted by the uncertainty-aware 3D segmentation network and the structure-enhanced 3D segmentation network are XORed to obtain the mask M of the inconsistency between their predictions. diff Specifically:

[0125]

[0126] in, and These are the softmax outputs of labeled 3D medical images processed by an uncertainty-aware 3D segmentation network and a structure-enhanced 3D segmentation network, respectively. For XOR operation, BINA is the binarization operation, and M is the XOR operation. diff The mask for inconsistent parts predicted by a labeled 3D medical image after passing through an uncertainty-aware 3D segmentation network and a structure-enhanced 3D segmentation network.

[0127] The predicted difference region is calculated based on the inconsistency mask predicted by the uncertainty-aware 3D segmentation network and the structure-enhanced 3D segmentation network, specifically as follows:

[0128]

[0129] in, The predicted difference region between uncertainty-aware 3D segmentation networks and structure-enhanced 3D segmentation networks. include and CLIP is an operation to obtain the predicted difference region corresponding to the masked region of the inconsistency.

[0130] To correct the prediction discrepancy between the uncertainty-aware 3D segmentation network and the structure-enhanced 3D segmentation network, the correction loss between the predicted discrepancy region and its corresponding ground truth mask region is calculated. The network is guided to correct potential mispredictions, specifically as follows:

[0131]

[0132] in, To predict the true mask image corresponding to the difference region, MSE is the mean squared error loss function.

[0133] Implementation Method Four: This implementation method further defines the semi-supervised 3D medical image segmentation method based on dual-correction mutual learning described in Implementation Method One. The preliminary predicted pseudo-masks generated by two pre-trained networks with different structures are corrected to obtain corresponding reliable pseudo-labels. Specifically:

[0134] The two pre-trained networks with different structures are a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network.

[0135] The pre-trained uncertainty-aware 3D segmentation network generates a preliminary prediction pseudo-mask, which is then corrected by uncertainty awareness to obtain a reliable pseudo-label.

[0136] The pre-trained structure-enhanced 3D segmentation network generates a preliminary predicted pseudo-mask, which is then corrected by dynamically generating a high-confidence mask to obtain a reliable pseudo-label.

[0137] In this embodiment, the uncertainty perception correction specifically refers to:

[0138]

[0139] in, Predictive probability map for label-free 3D medical images The binary image P obtained by the argmax function A These are reliable pseudo-labels generated after uncertainty-aware correction. For uncertainty graphs, This is an indicator function, and T = 0.2 is the threshold.

[0140] The aforementioned dynamic generation of high-confidence mask correction specifically refers to:

[0141]

[0142] Among them, C A and C B These are the confidence score maps predicted by a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network, respectively, where ∈ is a very small constant. and These are the probabilities of a pixel belonging to the k-th class. This is a high-confidence mask image. Predictive probability map for label-free 3D medical images The binary image P obtained by the argmax function B These are reliable pseudo-labels generated after dynamic generation and high-confidence mask correction.

[0143] In this embodiment, the uncertainty-aware 3D segmentation network and the structure-enhanced 3D segmentation network are trained on labeled 3D medical images to generate a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network, respectively.

[0144] The pre-trained uncertainty-aware 3D segmentation network generates a preliminary prediction pseudo-mask, which is then corrected by uncertainty awareness to obtain a reliable pseudo-label.

[0145] The uncertainty map generated by the pre-trained uncertainty perception module filters out high-uncertainty components, improving the quality of pseudo-labels generated by the network. Specifically:

[0146]

[0147] in, Predictive probability map for label-free 3D medical images The binary image P obtained by the argmax function A These are reliable pseudo-labels generated after uncertainty-aware correction. For uncertainty graphs, This is an indicator function, and T = 0.2 is the threshold.

[0148] The pre-trained, structure-enhanced 3D segmentation network generates a preliminary predicted pseudo-mask, which is then corrected by dynamically generated high-confidence masks to obtain a reliable pseudo-label. Specifically:

[0149] The confidence scores of the predicted outputs of the pre-trained uncertainty-aware 3D segmentation network and the pre-trained structure-enhanced 3D segmentation network are calculated separately. By comparing the confidence scores of the two networks, a high-confidence mask image of the pre-trained structure-enhanced 3D segmentation network is dynamically generated. Specifically:

[0150]

[0151]

[0152] Among them, C A and C B These are the confidence score maps predicted by a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network, respectively, where ∈ is a very small constant. and These are the probabilities of a pixel belonging to the k-th class. This is a high-confidence mask image. Predictive probability map for label-free 3D medical images The binary image P obtained by the argmax function B These are reliable pseudo-labels generated after dynamic generation and high-confidence mask correction.

[0153] Therefore, in this embodiment, the unlabeled image is processed by two pre-trained networks with different structures to generate two preliminary predicted pseudo-masks, which are then corrected by uncertainty awareness and by dynamically generated high-confidence masks to obtain reliable pseudo-labels.

[0154] Implementation Method 5: This implementation method further defines the semi-supervised 3D medical image segmentation method based on dual-correction mutual learning described in Implementation Method 1. The supervised loss is specifically:

[0155]

[0156] in, To monitor losses, The supervised loss for pre-training uncertainty-aware 3D segmentation networks, The supervised loss for the pre-trained structure-enhanced 3D segmentation network, The correction loss is used to compare the predicted difference regions of two networks with their corresponding ground truth mask regions. g represents the operation for calculating uncertainty. This is a preliminary predicted probability map obtained from a labeled 3D medical image using an uncertainty-aware 3D segmentation network. Y is the preliminary predicted probability map obtained from a structure-enhanced 3D segmentation network for a labeled 3D medical image;

[0157] The aforementioned cross-supervision loss is specifically as follows:

[0158]

[0159] in, For cross-supervision loss, P A For reliable pseudo-labels generated after uncertainty-aware correction, P B This is a reliable pseudo-label generated after dynamic high-confidence mask correction. This is the cross-entropy loss function.

[0160] In this implementation, the total loss of the calculation model includes two parts: supervision loss and cross-pseudo-supervision loss (consistency loss), specifically:

[0161]

[0162] Where λ is the balance monitoring loss. and consistency loss The parameters are obtained using the Gaussian ramp function. Gradually adjust the weights, λ max =0.1, where t is the current training epoch number. max It is the total number of training epochs;

[0163] The aforementioned monitoring loss specifically refers to:

[0164]

[0165] in, The supervised loss for pre-training uncertainty-aware 3D segmentation networks, The supervised loss for the pre-trained structure-enhanced 3D segmentation network, The correction loss is used to compare the predicted difference regions of two networks with their corresponding ground truth mask regions. g represents the operation for calculating uncertainty. This is a preliminary predicted probability map obtained from a labeled 3D medical image using an uncertainty-aware 3D segmentation network. Y is the preliminary predicted probability map obtained from a structure-enhanced 3D segmentation network for a labeled 3D medical image;

[0166]

[0167] in, The Dice loss function, To extract evidence of loss;

[0168] The loss is calculated based on the Dirichlet distribution function and cross-entropy. Loss, by minimizing The loss-encouraging network generates evidence for positive samples of different classes, while penalizing the evidence for negative samples through Kullback–Leibler (KL) divergence loss, specifically:

[0169]

[0170] Wherein, B(α) i ) is a parameter of α i k-dimensional polynomial beta function, S k Let Γ(·) be a k-dimensional simplex, where ψ(·) is the digamma function and Γ(·) is the gamma function. β is used to balance the two losses. t is the index of the current iteration, t max It is the maximum number of iterations, D(p) i |1) is a Dirichlet distribution with uniform parameters;

[0171] The loss of trust quality is specifically as follows:

[0172]

[0173] in, To improve the quality of belief b i The simplex obtained through softmax transformation;

[0174] The aforementioned cross-supervision loss is specifically as follows:

[0175]

[0176] in, For cross-supervision loss, P A For reliable pseudo-labels generated after uncertainty-aware correction, P B This is a reliable pseudo-label generated after dynamic high-confidence mask correction. This represents the cross-entropy loss.

[0177] Therefore, this implementation achieves mutual learning between the two networks by calculating the supervised loss of labeled 3D medical images and the cross-pseudo-supervised loss of unlabeled 3D medical images.

[0178] To better illustrate the semi-supervised 3D medical image segmentation method based on dual-calibration mutual learning as described in any of embodiments one through five, the following examples provide a detailed description:

[0179] To verify the effectiveness and superiority of the method in this embodiment, additional methods such as MT, UA-MT, SASSNet, DTC, MCNet, and MCF were selected for parallel comparison. Quantitative comparisons were performed on 20% labeled images of the LA left atrial segmentation dataset, and the Dice, Jaccard, 95HD, and ASD segmentation evaluation metrics for each method were calculated. The segmentation results are shown below. Figure 3 As shown in Table 1, the comparison results are as follows. It can be seen from the table that in the method of this embodiment, the Dice and Jaccard indices are higher than those of similar comparative methods, indicating that the prediction result of the segmentation method in this embodiment is closest to the actual result. The 95HD and ASD indices of the method of this embodiment are lower than those of similar comparative methods, indicating that the distance between the prediction result and the actual result of the segmentation method in this embodiment is the smallest and the similarity is the highest.

[0180] Table 1

[0181]

[0182] In summary, this embodiment designs a semi-supervised 3D medical image segmentation scheme with dual correction and mutual learning, namely, using network prediction difference correction and pseudo-label correction, which effectively alleviates the problems of model cognitive bias and low-quality pseudo-label filtering in existing semi-supervised medical image segmentation. Through the advantages of dual correction, a reliable pseudo-mask is generated to guide network mutual learning, thereby improving the performance of semi-supervised 3D medical image segmentation.

[0183] Implementation Method Six: The semi-supervised 3D medical image segmentation system based on dual-calibration mutual learning described in this implementation method is specifically as follows:

[0184] Preprocessing is performed on the three-dimensional medical images, which include labeled three-dimensional medical images and unlabeled three-dimensional medical images;

[0185] Two networks with different structures were trained using labeled 3D medical images. While calculating the supervised loss of the labeled 3D medical images, the prediction difference between the two networks with different structures was corrected, resulting in two pre-trained networks with different structures.

[0186] Unlabeled 3D medical images are processed by two pre-trained networks with different structures to generate corresponding preliminary prediction pseudomasks. The preliminary prediction pseudomasks generated by the two pre-trained networks with different structures are corrected to obtain corresponding reliable pseudo labels.

[0187] Based on the cross-pseudo-supervised loss of unlabeled 3D medical images, mutual learning between two pre-trained networks with different structures is achieved, the parameters of the two pre-trained networks with different structures are updated, and the final two networks with different structures are obtained respectively.

[0188] The three-dimensional medical images to be processed are input into two networks with different final structures to achieve the three-dimensional medical image segmentation task.

[0189] Implementation Method Seven: An electronic device according to this implementation method includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0190] Memory, used to store computer programs;

[0191] When a processor executes a program stored in memory, it implements the steps of any one of embodiments one through five.

[0192] Implementation Method 8: A computer-readable storage medium according to this implementation method, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of any one of the methods described in Implementation Methods 1-5.

[0193] The foregoing has provided a detailed description of the semi-supervised 3D medical image segmentation method, system, device, and storage medium based on dual-correction mutual learning proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A semi-supervised three-dimensional medical image segmentation method based on dual-calibration mutual learning, characterized in that, Specifically: Preprocessing is performed on the three-dimensional medical images, which include labeled three-dimensional medical images and unlabeled three-dimensional medical images; Two networks with different structures were trained using labeled 3D medical images. While calculating the supervised loss of the labeled 3D medical images, the prediction difference between the two networks with different structures was corrected, resulting in two pre-trained networks with different structures. Unlabeled 3D medical images are processed by two pre-trained networks with different structures to generate corresponding preliminary prediction pseudomasks. The preliminary prediction pseudomasks generated by the two pre-trained networks with different structures are corrected to obtain corresponding reliable pseudo labels. Based on the cross-pseudo-supervised loss of unlabeled 3D medical images, mutual learning between two pre-trained networks with different structures is achieved, the parameters of the two pre-trained networks with different structures are updated, and the final two networks with different structures are obtained respectively. The three-dimensional medical images to be processed are input into two final networks with different structures to achieve the three-dimensional medical image segmentation task; Both pre-trained networks with different structures generate their corresponding preliminary predicted pseudomasks, which are then corrected to obtain their corresponding reliable pseudolabels, as follows: The two pre-trained networks with different structures are a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network. The pre-trained uncertainty-aware 3D segmentation network generates a preliminary prediction pseudo-mask, which is then corrected by uncertainty awareness to obtain a reliable pseudo-label. The pre-trained structure-enhanced 3D segmentation network generates a preliminary predicted pseudo-mask, which is then corrected by dynamically generated high-confidence masks to obtain a reliable pseudo-label. The aforementioned uncertainty perception correction specifically includes: ； ； in, Predictive probability map for label-free 3D medical images The binary image obtained by the argmax function. These are reliable pseudo-labels generated after uncertainty-aware correction. For uncertainty graphs, This is an indicator function, and T=0.2 is the threshold. The aforementioned dynamic generation of high-confidence mask correction specifically refers to: ； ； ； ； in, and The images show the confidence scores predicted by a pre-trained uncertainty-aware 3D segmentation network and a pre-trained structure-enhanced 3D segmentation network, respectively. It is a very small constant. and All pixels belong to the first The probability of a class This is a high-confidence mask image. Predictive probability map for label-free 3D medical images The binary image obtained by the argmax function. These are reliable pseudo-labels generated after dynamic generation and high-confidence mask correction.

2. The semi-supervised three-dimensional medical image segmentation method based on dual-calibration mutual learning according to claim 1, characterized in that, The two networks with different structures are an uncertainty-aware 3D segmentation network and a structure-enhanced 3D segmentation network. The decoder of the uncertainty-aware 3D segmentation network employs an uncertainty-aware module, which specifically includes: ； in, For the sample The overall uncertainty For the sample middle The quality of trust in the class; The encoder end of the structure-enhanced 3D segmentation network employs a structure enhancement module, which specifically includes: ； ； in, 、 and Each is a 1×1×1 convolutional layer ， and The first Feature maps of layer encoders and decoders, For sigmoid activation function, To enhance the boundary contour information of the target area, the first Layer encoder feature map, This is a feature map that contains structural enhancement information.

3. The semi-supervised three-dimensional medical image segmentation method based on dual-calibration mutual learning according to claim 1, characterized in that, The method for correcting the prediction discrepancy between two networks with different structures is as follows: The binarized mask images predicted by two networks with different structures are XORed to obtain the mask of the inconsistent part predicted by the two networks with different structures. Based on the mask of the inconsistency in the predictions between two networks with different structures, the prediction difference region between the two networks with different structures is calculated. Calculate the correction loss of the predicted difference region and its corresponding true mask region for two networks with different structures, and guide the two networks with different structures to correct the erroneous predictions.

4. The semi-supervised three-dimensional medical image segmentation method based on dual-calibration mutual learning according to claim 3, characterized in that, The method for obtaining the mask of the inconsistency in predictions between two networks with different structures is as follows: ； in, and These are the softmax outputs of labeled 3D medical images processed by two networks with different structures. For XOR operation, For binarization operation, The mask for the inconsistent portions of a labeled 3D medical image after prediction by two networks with different structures; The calculation of the prediction difference region between the two networks with different structures is specifically as follows: ； in, This represents the region of difference in predictions between two networks with different structures. include and , To obtain the predicted difference region operation corresponding to the inconsistent mask region; The calculation of the correction loss for the predicted difference region and its corresponding ground truth mask region between two networks with different structures is specifically as follows: ； in, To predict the true mask image corresponding to the difference region, Let the mean squared error loss function be . The correction loss is used to compare the predicted difference regions of two networks with their corresponding ground truth mask regions.

5. The semi-supervised three-dimensional medical image segmentation method based on dual-calibration mutual learning according to claim 1, characterized in that, The aforementioned monitoring loss specifically refers to: ； in, To monitor losses, The supervised loss for pre-training uncertainty-aware 3D segmentation networks, The supervised loss for the pre-trained structure-enhanced 3D segmentation network, The correction loss is used to compare the predicted difference regions of two networks with their corresponding ground truth mask regions. , To calculate uncertain operations, This is a preliminary predicted probability map obtained from a labeled 3D medical image using an uncertainty-aware 3D segmentation network. Y is the preliminary predicted probability map obtained from a structure-enhanced 3D segmentation network for a labeled 3D medical image; The aforementioned cross-supervision loss is specifically as follows: ； in, For the loss of cross-monitoring, These are reliable pseudo-labels generated after uncertainty-aware correction. This is a reliable pseudo-label generated after dynamic high-confidence mask correction. This is the cross-entropy loss function.

6. A semi-supervised three-dimensional medical image segmentation system based on dual-calibration mutual learning, wherein the system is used to implement the semi-supervised three-dimensional medical image segmentation method based on dual-calibration mutual learning as described in claim 1, characterized in that, Specifically: Preprocessing is performed on the three-dimensional medical images, which include labeled three-dimensional medical images and unlabeled three-dimensional medical images; Two networks with different structures were trained using labeled 3D medical images. While calculating the supervised loss of the labeled 3D medical images, the prediction difference between the two networks with different structures was corrected, resulting in two pre-trained networks with different structures. Unlabeled 3D medical images are processed by two pre-trained networks with different structures to generate corresponding preliminary prediction pseudomasks. The preliminary prediction pseudomasks generated by the two pre-trained networks with different structures are corrected to obtain corresponding reliable pseudo labels. Based on the cross-pseudo-supervised loss of unlabeled 3D medical images, mutual learning between two pre-trained networks with different structures is achieved, the parameters of the two pre-trained networks with different structures are updated, and the final two networks with different structures are obtained respectively. The three-dimensional medical images to be processed are input into two networks with different final structures to achieve the three-dimensional medical image segmentation task.

7. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method described in any one of claims 1-5.

Images