A semi-supervised domain generalization medical image segmentation method and system

Abstract

The application belongs to the field of image segmentation, and provides a semi-supervised domain generalization medical image segmentation method and system, which comprises obtaining an original medical image and pre-processing to obtain a pre-processed medical image; based on the pre-processed medical image, a pre-trained semi-supervised domain generalization medical image segmentation model is used for image segmentation; a style enhancement method is used for style mixing and style disturbance operation on the low-frequency information of the features, so that the image style is diversified and enhanced while the semantic information is unchanged. At the same time, a style enhancement decoder is introduced to generate a style enhancement image, so as to improve the explainability of the style enhancement process. In addition, through the confidence-aware consistency loss, the pseudo label of the teacher model which is more confident about the unlabeled data from the unknown domain is selected, so as to better guide the consistency regularization process to improve the generalization performance of the model in the unknown domain.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, specifically relating to a semi-supervised domain generalized medical image segmentation method and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Deep learning-based medical image segmentation algorithms perform exceptionally well when the training and test sets of medical images are of the same distribution, with segmentation performance even comparable to that of experienced clinicians. However, medical images are typically acquired from different clinical centers and are affected by factors such as scanning equipment and parameter settings. This results in significant distributional differences in the acquired medical images, known as domain shift. Consequently, models trained in the source domain often produce poor segmentation results when applied to unknown target domains. Therefore, the task of domain generalization has been proposed to overcome the domain shift problem. The goal of domain generalization is to learn a model from the source domain and apply it directly to unseen target domains without additional training.

[0004] How to generalize a model trained in the source domain to the target domain is a pressing problem. There are three main categories of methods: The first category is data augmentation-based methods, which propose hybrid task sampling to enhance the diversity of task-level training samples. Another approach is to directly use CycleGAN to transform images from one specific domain to other domains for data augmentation. The second category is methods that learn domain-invariant features. These methods utilize disentangled networks to decompose data into cross-domain invariant structural content information and domain-specific style information. The cross-domain invariant information is further used in subsequent segmentation tasks. Disentangled models trained adversarially filter out style information, making the segmentation model unaffected by style changes, thereby improving the model's generalization performance. The third category is domain-adaptive feature calibration methods. These methods propose a dynamic convolutional head that adapts the model's convolutional parameters to the unseen target domain by predicting the similarity of each domain. A domain-oriented feature embedding framework is introduced to dynamically update prior knowledge specific to the domain, making semantic features more discriminative. However, the above methods require fully labeled datasets for each domain, which may not be readily available in clinical practice.

[0005] Therefore, some researchers have proposed semi-supervised domain generalization medical image segmentation, aiming to achieve generalization on unknown domains by training with a small amount of labeled source domain data and a large amount of unlabeled source domain data. For example, anatomical and style information obtained using a disentangled model is utilized. Subsequently, based on the disentangled model, a semi-supervised meta-learning training strategy is used to further improve the generalization of the model segmentation. However, the meta-learning process is very time-consuming because it requires considering all potential subsets of assignments during meta-training and meta-testing. Data augmentation is an effective method to improve model robustness; it transforms or perturbs source domain data to simulate unseen variations in data distribution. Existing techniques propose a confidence-cross pseudo-label method, which augments source data by mixing MRI images in the Fourier domain and uses pseudo-labels to utilize unlabeled data. However, using linear operations on the low-frequency spectrum of the image for style augmentation limits the diversity of synthetic variations and cannot effectively diversify the augmentation of data styles.

[0006] In summary, existing methods for generalizing medical image segmentation in semi-supervised domains suffer from several drawbacks. The meta-learning process is very time-consuming because it requires considering all potential allocation subsets during meta-training and meta-testing. Furthermore, the use of linear operations on the low-frequency spectrum of the image for style enhancement limits the diversity of synthetic variations.

[0007] To address the semi-supervised domain generalization problem in medical image segmentation, a common approach is to use data augmentation to obtain more diverse data to simulate possible data distributions in the unknown domain, thereby improving the model's generalization performance in the unknown domain. However, performing Fourier transforms at the image level and using linear operations on the low-frequency spectrum for data augmentation limits the diversity of augmented data styles. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a semi-supervised domain generalized medical image segmentation method and system. The invention introduces a style diversification enhancement method, performing Fourier transform on features at each layer of the style decoder and adding style noise and style mixing to the channel-level statistics of its low-frequency spectrum. This enhances the style of the data while preserving semantic information, thereby introducing more cross-domain information perturbations into the teacher-student framework. Furthermore, a style enhancement decoder is used to execute the style enhancement process, improving its interpretability. Simultaneously, this method proposes a confidence-aware regularization strategy based on a dual-teacher model. By selecting the more confident prediction from the dual-teacher model, the consistency regularization process is guided, significantly improving the model's generalization ability.

[0009] According to some embodiments, the first aspect of the present invention provides a semi-supervised domain generalized medical image segmentation method, which adopts the following technical solution:

[0010] A semi-supervised domain generalized medical image segmentation method includes:

[0011] The original medical image is acquired and preprocessed to obtain the preprocessed medical image;

[0012] Based on pre-processed medical images, image segmentation is performed using a pre-trained semi-supervised domain generalized medical image segmentation model.

[0013] The training process of the semi-supervised domain generalized medical image segmentation model is as follows:

[0014] The original medical images are acquired and preprocessed to obtain a source domain labeled dataset and a source domain unlabeled dataset.

[0015] The source domain labeled dataset is input into the student model for image segmentation prediction, and the supervision loss and reconstruction loss are calculated using the student image segmentation prediction results and the real labels.

[0016] Weakly augment the unlabeled dataset in the source domain and then input it into the dual-teacher model for segmentation prediction to obtain pseudo-labeled samples;

[0017] The source domain unlabeled image features extracted by the encoders in the two teacher models are integrated, and the style enhancement decoder is used to enhance the style of the integrated features to obtain a style-enhanced image.

[0018] The style-enhanced image is input into the student model for image segmentation prediction, and the style-enhanced image segmentation prediction result is obtained. The confidence-perceived consistency loss is calculated based on the style-enhanced image segmentation prediction result and the pseudo-label sample.

[0019] Given the number of training iterations, calculate the sum of the above loss functions, update the network parameters through gradient descent of the objective function, and finally obtain the trained semi-supervised generalized medical image segmentation model.

[0020] Furthermore, the calculation of supervised loss and reconstruction loss using student image segmentation prediction results and ground truth labels specifically involves:

[0021] The supervised training loss is calculated based on the similarity between the student's image segmentation prediction results and the ground truth labels.

[0022] The features extracted by the encoder in the student model are reconstructed using a style enhancement decoder, and the reconstruction loss is calculated based on the reconstruction results and the source domain labeled dataset.

[0023] Furthermore, the style enhancement decoder is used to perform style diversification enhancement on the integrated features to obtain a style-enhanced image, specifically as follows:

[0024] Based on the integrated features, the input is fed into the style enhancement decoder;

[0025] The features z extracted from each layer of the style enhancement decoder by upsampling convolution i Perform a Fourier transform, decompose the Fourier transform result, and obtain the amplitude spectrum and phase diagram;

[0026] The channel-level statistics of the amplitude spectrum calculation are further subjected to style mixing and style noise operations to obtain the style-enhanced low-frequency spectrum;

[0027] Based on the style-enhanced low-frequency spectrum and amplitude spectrum, the original high-frequency spectrum is merged to obtain the style-enhanced amplitude spectrum.

[0028] The amplitude spectrum of the style-enhanced model is transformed from the frequency domain to the feature space by inverse Fourier transform, thus obtaining the style-enhanced features.

[0029] The above style enhancement process is applied to each layer of the style enhancement decoder. After the entire style enhancement decoder upsampling process, the style-enhanced image is obtained.

[0030] Furthermore, the calculation of the confidence-perceived consistency loss based on the style-enhanced image segmentation prediction results and pseudo-label samples is specifically as follows:

[0031] Calculate the entropy of the prediction results of the two teacher models for segmenting the unlabeled data in the source domain;

[0032] Calculate the corresponding segmentation weights based on the entropy of the prediction results of the two teacher models;

[0033] The dual-teacher prediction ensemble result is obtained by dividing the sum of the products of the prediction results of the two teacher models and the segmentation weights by 2.

[0034] Calculate the entropy of the dual-teacher prediction ensemble result, and calculate the dual-teacher segmentation weights based on the entropy of the dual-teacher prediction ensemble result;

[0035] Based on the style-enhanced image segmentation prediction results and the cross-entropy of pseudo-labeled samples, as well as the dual-teacher segmentation weights, the confidence-perceived consistency loss is determined.

[0036] Furthermore, the confidence-perceived consistency loss is determined based on the style-enhanced image segmentation prediction results, the cross-entropy of pseudo-labeled samples, and the dual-teacher segmentation weights, specifically as follows:

[0037]

[0038] In the formula, y t These are pseudo-labels predicted by the dual-teacher model for weakly enhanced unlabeled images. L ce Let cross-entropy be the loss function, which is defined as:

[0039]

[0040] In the formula, y i These are pseudo-labels predicted by the dual-teacher model for unlabeled data from an unknown distribution. This is used to predict segmentation of unlabeled data by the student model after style diversification enhancement.

[0041] Furthermore, given the number of training iterations, the sum of the above loss functions is calculated, and the network parameters are updated through gradient descent of the objective function, ultimately obtaining the trained semi-supervised generalized medical image segmentation model, specifically:

[0042] Given the number of training iterations;

[0043] The sum of supervised training loss, reconstruction loss, and confidence-perceived consistency loss is used as the total loss of the semi-supervised domain generalized medical image segmentation model.

[0044] The parameters of the student model are continuously updated using gradient descent of the objective function;

[0045] The dual-teacher model is updated using an exponential moving average of the student model parameters;

[0046] Only one teacher model is updated in each round;

[0047] After iteration, a well-trained semi-supervised generalized medical image segmentation model is finally obtained.

[0048] Furthermore, the total loss of the semi-supervised domain generalized medical image segmentation model is specifically as follows:

[0049] L = L sup +α*L con +L rec ;

[0050] Where α is the weight that controls the balance between the supervised training loss and the confidence-perceived consistency loss, and L is the supervised training loss. sup Reconstruction loss L rec Confidence-perceived consistency loss L con .

[0051] According to some embodiments, the second aspect of the present invention provides a semi-supervised domain generalized medical image segmentation system, which adopts the following technical solution:

[0052] A semi-supervised domain generalized medical image segmentation system includes:

[0053] The image acquisition and processing module is configured to acquire raw medical images and perform preprocessing to obtain preprocessed medical images.

[0054] The image prediction and segmentation module is configured to perform image segmentation based on pre-processed medical images using a pre-trained semi-supervised domain generalized medical image segmentation model.

[0055] The training process of the semi-supervised domain generalized medical image segmentation model is as follows:

[0056] The original medical images are acquired and preprocessed to obtain a source domain labeled dataset and a source domain unlabeled dataset.

[0057] The source domain labeled dataset is input into the student model for image segmentation prediction, and the supervision loss and reconstruction loss are calculated using the student image segmentation prediction results and the real labels.

[0058] Weakly augment the unlabeled dataset in the source domain and then input it into the dual-teacher model for segmentation prediction to obtain pseudo-labeled samples;

[0059] The source domain unlabeled image features extracted by the encoders in the two teacher models are integrated, and the style enhancement decoder is used to enhance the style of the integrated features to obtain a style-enhanced image.

[0060] The style-enhanced image is input into the student model for image segmentation prediction, and the style-enhanced image segmentation prediction result is obtained. The confidence-perceived consistency loss is calculated based on the style-enhanced image segmentation prediction result and the pseudo-label sample.

[0061] Given the number of training iterations, calculate the sum of the above loss functions, update the network parameters through gradient descent of the objective function, and finally obtain the trained semi-supervised generalized medical image segmentation model.

[0062] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0063] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a semi-supervised domain generalized medical image segmentation method as described in the first aspect above.

[0064] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0065] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of a semi-supervised domain generalized medical image segmentation method as described in the first aspect above.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] This invention proposes a novel style enhancement method that performs style mixing and style perturbation operations on low-frequency feature information, diversifying image style enhancement while preserving semantic information. Simultaneously, a style enhancement decoder is introduced to generate style-enhanced images, thereby improving the interpretability of the style enhancement process. Furthermore, from the perspective of consistency regularization, a confidence-perceived consistency method based on dual teachers is proposed. This method uses confidence-perceived consistency loss to select pseudo-labels that the teacher model is more confident in regarding unlabeled data from unknown domains, thus better guiding the consistency regularization process and improving the model's generalization performance in unknown domains. Attached Figure Description

[0068] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0069] Figure 1 This is a training flowchart of a semi-supervised domain generalized medical image segmentation method according to an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the training process of the semi-supervised domain generalized medical image segmentation model in an embodiment of the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0072] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0073] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0074] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0075] Example 1

[0076] This embodiment provides a semi-supervised domain generalized medical image segmentation method. This embodiment uses the application of this method to a server as an example for illustration. It is understood that this method can also be applied to terminals, and can also be applied to systems including terminals, servers, and other components, and can be implemented through interaction between the terminal and the server. The server can be a standalone physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. In this embodiment, the method includes the following steps:

[0077] The original medical image is acquired and preprocessed to obtain the preprocessed medical image;

[0078] Based on pre-processed medical images, image segmentation is performed using a pre-trained semi-supervised domain generalized medical image segmentation model.

[0079] The training process of the semi-supervised domain generalized medical image segmentation model is as follows:

[0080] The original medical images are acquired and preprocessed to obtain a source domain labeled dataset and a source domain unlabeled dataset.

[0081] The source domain labeled dataset is input into the student model for image segmentation prediction, and the supervision loss and reconstruction loss are calculated using the student image segmentation prediction results and the real labels.

[0082] Weakly augment the unlabeled dataset in the source domain and then input it into the dual-teacher model for segmentation prediction to obtain pseudo-labeled samples;

[0083] The source domain unlabeled image features extracted by the encoders in the two teacher models are integrated, and the style enhancement decoder is used to enhance the style of the integrated features to obtain a style-enhanced image.

[0084] The style-enhanced image is input into the student model for image segmentation prediction, and the style-enhanced image segmentation prediction result is obtained. The confidence-perceived consistency loss is calculated based on the style-enhanced image segmentation prediction result and the pseudo-label sample.

[0085] Given the number of training iterations, calculate the sum of the above loss functions, update the network parameters through gradient descent of the objective function, and finally obtain the trained semi-supervised generalized medical image segmentation model.

[0086] like Figure 1 and Figure 2 As shown, in this embodiment, the training process of the semi-supervised domain generalized medical image segmentation method includes the following steps:

[0087] Step 1: Supervised training with labeled data in the source domain. Train the student model and style enhancement decoder using labeled data, and calculate the supervised loss and reconstruction loss.

[0088] Step 2: Unsupervised training on unlabeled data from the source domain. The dual-teacher model performs segmentation prediction on the weakly augmented unlabeled data.

[0089] Step 3: Style diversification enhancement. The extracted image features are processed using a style enhancement decoder to obtain a style-enhanced image.

[0090] Step 4: Confidence-Aware Regularization Based on Dual Teachers. We further calculate the confidence-aware consistency loss between the predictions of the dual-teacher model and the student model, and provide more accurate guidance for consistency regularization learning by selecting the more confident predictions from the dual-teacher model.

[0091] Step 5: Given the number of training iterations, calculate the loss function mentioned above, update the network parameters through gradient descent of the objective function, and finally obtain the best medical image segmentation model.

[0092] Step 1, the specific implementation of supervised training with labeled data in the source domain, is as follows:

[0093] Given multiple source domains D S ={D1,D2,...,D k}, each source domain D k have Labeled data and Unlabeled data in k represents the total number of source domains.

[0094] The student model makes predictions on labeled data:

[0095]

[0096] Then, the Dice loss is calculated for the obtained predictions and the true label y. The Dice loss is defined as follows:

[0097] L dice =1 - DiceCoefficient (2);

[0098] DiceCoefficient is used to measure the similarity between two sets, and is defined as follows:

[0099]

[0100] In the formula: x and y represent the segmentation result predicted by the model and the true label, respectively. |x∩y| represents the overlapping region between x and y, and |x|+|y| represents the joint region.

[0101] Therefore, the supervised training loss is defined as:

[0102]

[0103] Next, the style enhancement decoder D A Features Z extracted from the student encoder S The reconstruction process yields the reconstructed image X. rec And calculate the reconstruction loss:

[0104]

[0105] L mse The definition is as follows:

[0106]

[0107] Among them, y i It's a real label. It is the result of the model prediction.

[0108] Step 2, the unsupervised training steps for unlabeled data in the source domain, includes:

[0109] (1) Input unlabeled data from the multi-source domain Then, weak augmentation is performed on the unlabeled data, which includes random rotation and random cropping.

[0110] (2) The dual-teacher model predicts the unlabeled data after weak enhancement.

[0111]

[0112] Step 3, the style diversity enhancement step, includes:

[0113] (1) Integrate the features extracted from the weakly enhanced unlabeled data by the dual-teacher encoder in step two to obtain the integrated feature Z. t The features z extracted by the upsampling convolution at each layer of the style enhancement decoder are then processed. i Perform Fourier transform:

[0114]

[0115] Then, solve F(z) i The amplitude spectrum A is obtained.i Phase diagram P i Among them, amplitude spectrum A i The phase map P contains local features of a specific domain, i.e., style information. i This includes the high-level semantics of the original signal. Therefore, for the amplitude spectrum A... i The calculated channel-level statistics are further processed with style mixing and style noise operations to explore as much of the data in the unknown domain as possible.

[0116]

[0117] in, This is the low-frequency spectrum after style enhancement. Style mixing parameters can be represented as:

[0118] γ mix =λ mix σ(A j ⊙Μ)+(1-λ mix )σ(A i ⊙Μ) (11);

[0119] β mix =λ mix μ(A j ⊙Μ)+(1-λ mix )μ(A i ⊙Μ) (12);

[0120] Where, λ mix It is a coefficient that controls the interpolation level, and is randomly sampled from [0,1].

[0121] Style noise is constrained to a reasonable range using ∑γ and ∑β, where

[0122] ∑γ=σ 2 ({σ(A j ⊙Μ)}) j＝1...j,...B (13);

[0123] ∑β=μ 2 ({μ(A j ⊙Μ)}) j＝1...j,...B (14);

[0124] Style noise parameter ε γ , M is a binary mask used to control the spatial range of the amplitude spectrum to be swapped, setting M to the central region of the amplitude spectrum containing low-frequency information.

[0125] Then, the original high-frequency spectra are merged to obtain the enhanced amplitude spectrum.

[0126]

[0127] Next, through the inverse Fourier transform F -1 The style-enhanced data is transformed from the frequency domain to the feature space to obtain style-enhanced features.

[0128]

[0129] The style enhancement process described above is applied to each layer of the style enhancement decoder. After the entire style enhancement decoder upsampling process, the final style-enhanced image is obtained.

[0130] Step 4, the confidence-aware regularization step based on dual teachers, includes:

[0131] (1) The student model enhances the style diversity of the image obtained in step three. Make a prediction:

[0132]

[0133] (2) Calculate the prediction results of the two teacher models for unlabeled data from an unknown distribution. Entropy:

[0134]

[0135] Based on the entropy of the obtained segmentation prediction results, the corresponding weights are further calculated:

[0136]

[0137] The predictions from the two teachers are integrated based on the obtained weights:

[0138]

[0139] Further calculation of the dual-teacher prediction ensemble result P t The entropy is then calculated, and its corresponding weights are obtained:

[0140]

[0141] If the prediction entropy is small, the calculated weight is large, indicating that the segmentation prediction quality is high; and vice versa.

[0142] Furthermore, a confidence-aware consistency loss is proposed to better guide the learning process of the student model, thereby optimizing network training:

[0143]

[0144] In the formula: y tThese are pseudo-labels predicted by the dual-teacher model for weakly enhanced unlabeled images. L ce Let cross-entropy be the loss function, which is defined as:

[0145]

[0146] Among them, y i These are pseudo-labels predicted by the dual-teacher model for unlabeled data from an unknown distribution. This is used to predict segmentation of unlabeled data by the student model after style diversification enhancement.

[0147] In step 5, given the number of training iterations, the parameters of the student model are continuously updated using the gradient descent of the objective function, in conjunction with the aforementioned loss function.

[0148] L = L sup +α*L con +L rec (27);

[0149] Here, α is the weight that balances the control supervision loss and the confidence perceived consistency loss.

[0150] The dual-teacher model is updated using an exponential moving average of the student model parameters:

[0151]

[0152] in, This controls the weight of the transfer parameters in each round, updating only one teacher in each round.

[0153] During the reasoning process, the student model's predictions serve as the final segmentation result.

[0154] Experimental content

[0155] Experimental platform: All experiments were conducted on servers with NVIDIA A800 GPUs, using the Python programming language in PyCharm software, with PyTorch as the deep learning framework.

[0156] Dataset Introduction: The Multicenter, Multi-Supplier, and Multi-Disease Cardiac Image Segmentation Dataset (M&Ms) includes 320 subjects scanned using equipment from four different MRI scanner vendors (Siemens, Philips, GE, and Canon). Subjects from different scanners were treated as different domains. Therefore, the M&Ms dataset has four domains: Domain A has 95 topics, Domain B has 125 topics, and Domains C and D each contain 50 topics. For each subject, only end-systolic and end-diastolic phases were annotated. Voxel resolution ranges from 0.85 × 0.85 × 10 mm to 1.45 × 1.45 × 9.9 mm.

[0157] Experimental Results: As shown in the table below, under the conditions of 2% and 5% labeled data in the M&Ms dataset, our method outperforms other methods in all target domains. Under the 2% labeled data condition in the M&Ms dataset, the average Dice score (%) is improved by 6.44%; under the 5% labeled data condition, the average Dice score (%) is improved by 1.75%. Therefore, the experimental results demonstrate the effectiveness of our method.

[0158] Table 1 Comparison of experimental results

[0159]

[0160] This embodiment proposes a novel style diversification enhancement method. While preserving domain-invariant global structural information, it performs diverse style enhancements on the data to introduce cross-domain data perturbations into the teacher-student framework and improve the accuracy of the teacher module's predictions of unlabeled data from unknown distributions. This, in turn, better integrates with consistency regularization to improve the model's robustness in unknown domains. Furthermore, we propose a confidence-aware regularization strategy between predictions of weakly enhanced and style-enhanced images, thereby better regularizing the teacher-student framework. Experiments show that our proposed method achieves improved generalization performance on the M&Ms dataset.

[0161] Example 2

[0162] This embodiment provides a semi-supervised domain generalized medical image segmentation system, including:

[0163] The image acquisition and processing module is configured to acquire raw medical images and perform preprocessing to obtain preprocessed medical images.

[0164] The image prediction and segmentation module is configured to perform image segmentation based on pre-processed medical images using a pre-trained semi-supervised domain generalized medical image segmentation model.

[0165] The training process of the semi-supervised domain generalized medical image segmentation model is as follows:

[0166] The original medical images are acquired and preprocessed to obtain a source domain labeled dataset and a source domain unlabeled dataset.

[0167] The source domain labeled dataset is input into the student model for image segmentation prediction, and the supervision loss and reconstruction loss are calculated using the student image segmentation prediction results and the real labels.

[0168] Weakly augment the unlabeled dataset in the source domain and then input it into the dual-teacher model for segmentation prediction to obtain pseudo-labeled samples;

[0169] The source domain unlabeled image features extracted by the encoders in the two teacher models are integrated, and the style enhancement decoder is used to enhance the style of the integrated features to obtain a style-enhanced image.

[0170] The style-enhanced image is input into the student model for image segmentation prediction, and the style-enhanced image segmentation prediction result is obtained. The confidence-perceived consistency loss is calculated based on the style-enhanced image segmentation prediction result and the pseudo-label sample.

[0171] Given the number of training iterations, calculate the sum of the above loss functions, update the network parameters through gradient descent of the objective function, and finally obtain the trained semi-supervised generalized medical image segmentation model.

[0172] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0173] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0174] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0175] Example 3

[0176] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a semi-supervised domain generalized medical image segmentation method as described in Embodiment 1 above.

[0177] Example 4

[0178] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the semi-supervised domain generalization medical image segmentation method described in Embodiment 1 above.

[0179] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0180] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0181] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0182] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0183] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0184] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A semi-supervised domain generalization medical image segmentation method, characterized in that, include: The original medical image is acquired and preprocessed to obtain the preprocessed medical image; Based on pre-processed medical images, image segmentation is performed using a pre-trained semi-supervised domain generalized medical image segmentation model. The training process of the semi-supervised domain generalized medical image segmentation model is as follows: The original medical images are acquired and preprocessed to obtain a source domain labeled dataset and a source domain unlabeled dataset. The source domain labeled dataset is input into the student model for image segmentation prediction, and the supervision loss and reconstruction loss are calculated using the student image segmentation prediction results and the real labels. Weakly augment the unlabeled dataset in the source domain and then input it into the dual-teacher model for segmentation prediction to obtain pseudo-labeled samples; The source domain unlabeled image features extracted by the encoders in the two teacher models are integrated, and the style enhancement decoder is used to enhance the style of the integrated features to obtain a style-enhanced image. The style-enhanced image is input into the student model for image segmentation prediction, yielding the style-enhanced image segmentation prediction result. Based on the style-enhanced image segmentation prediction result and pseudo-labeled samples, the confidence-perceived consistency loss is calculated, specifically as follows: Calculate the entropy of the prediction results of the two teacher models for segmenting the unlabeled data in the source domain; Calculate the corresponding segmentation weights based on the entropy of the prediction results of the two teacher models; The dual-teacher prediction ensemble result is obtained by dividing the sum of the products of the prediction results of the two teacher models and the segmentation weights by 2. Calculate the entropy of the dual-teacher prediction ensemble result, and calculate the dual-teacher segmentation weights based on the entropy of the dual-teacher prediction ensemble result; Based on the style-enhanced image segmentation prediction results and the cross-entropy of pseudo-labeled samples and the bi-teacher segmentation weights, the confidence-perceived consistency loss is determined. Given the number of training iterations, calculate the sum of the above loss functions, update the network parameters through gradient descent of the objective function, and finally obtain the trained semi-supervised generalized medical image segmentation model.

2. The semi-supervised domain generalized medical image segmentation method as described in claim 1, characterized in that, The calculation of supervised loss and reconstruction loss using student image segmentation prediction results and ground truth labels is as follows: The supervised training loss is calculated based on the similarity between the student's image segmentation prediction results and the ground truth labels. The features extracted by the encoder in the student model are reconstructed using a style enhancement decoder, and the reconstruction loss is calculated based on the reconstruction results and the source domain labeled dataset.

3. The semi-supervised domain generalized medical image segmentation method as described in claim 1, characterized in that, The process of using a style enhancement decoder to perform style diversification enhancement on the integrated features to obtain a style-enhanced image is as follows: Based on the integrated features, the input is fed into the style enhancement decoder; Features extracted from each layer of the style enhancement decoder by upsampling convolution Perform a Fourier transform, decompose the Fourier transform result, and obtain the amplitude spectrum and phase diagram; The channel-level statistics of the amplitude spectrum calculation are further subjected to style mixing and style noise operations to obtain the style-enhanced low-frequency spectrum; Based on the style-enhanced low-frequency spectrum and amplitude spectrum, the original high-frequency spectrum is merged to obtain the style-enhanced amplitude spectrum. The amplitude spectrum of the style-enhanced model is transformed from the frequency domain to the feature space by inverse Fourier transform, thus obtaining the style-enhanced features. The above style enhancement process is applied to each layer of the style enhancement decoder. After the entire style enhancement decoder upsampling process, the style-enhanced image is obtained.

4. The semi-supervised domain generalization medical image segmentation method as described in claim 1, characterized in that, The confidence-perceived consistency loss is determined based on the cross-entropy of style-enhanced image segmentation prediction results and pseudo-labeled samples, and the dual-teacher segmentation weights. Specifically: ； In the formula, These are pseudo-labels predicted by the dual-teacher model for weakly enhanced unlabeled images. ; Let cross-entropy be the loss function, defined as follows: ； In the formula, These are pseudo-labels predicted by the dual-teacher model for unlabeled data from an unknown distribution. This is used to predict segmentation of unlabeled data by the student model after style diversification enhancement.

5. The semi-supervised domain generalization medical image segmentation method as described in claim 1, characterized in that, Given a given number of training iterations, the sum of the above loss functions is calculated, and the network parameters are updated using gradient descent of the objective function. Finally, a well-trained semi-supervised generalized medical image segmentation model is obtained, specifically: Given the number of training iterations; The sum of supervised training loss, reconstruction loss, and confidence-perceived consistency loss is used as the total loss of the semi-supervised domain generalized medical image segmentation model. The parameters of the student model are continuously updated using gradient descent of the objective function; The dual-teacher model is updated using an exponential moving average of the student model parameters; Only one teacher model is updated in each round; After iteration, a well-trained semi-supervised generalized medical image segmentation model is finally obtained.

6. The semi-supervised domain generalization medical image segmentation method as described in claim 5, characterized in that, The total loss of the semi-supervised domain generalized medical image segmentation model is as follows: ； in, It is a weight that balances the supervised training loss and the confidence-perceived consistency loss. Reconstruction losses Confidence-perceived consistency loss .

7. A semi-supervised domain generalization medical image segmentation system, characterized in that, include: The image acquisition and processing module is configured to acquire raw medical images and perform preprocessing to obtain preprocessed medical images. The image prediction and segmentation module is configured to perform image segmentation based on pre-processed medical images using a pre-trained semi-supervised domain generalized medical image segmentation model. The training process of the semi-supervised domain generalized medical image segmentation model is as follows: The original medical images are acquired and preprocessed to obtain a source domain labeled dataset and a source domain unlabeled dataset. The source domain labeled dataset is input into the student model for image segmentation prediction, and the supervision loss and reconstruction loss are calculated using the student image segmentation prediction results and the real labels. Weakly augment the unlabeled dataset in the source domain and then input it into the dual-teacher model for segmentation prediction to obtain pseudo-labeled samples; The source domain unlabeled image features extracted by the encoders in the two teacher models are integrated, and the style enhancement decoder is used to enhance the style of the integrated features to obtain a style-enhanced image. The style-enhanced image is input into the student model for image segmentation prediction, yielding the style-enhanced image segmentation prediction result. Based on the style-enhanced image segmentation prediction result and pseudo-labeled samples, the confidence-perceived consistency loss is calculated, specifically as follows: Calculate the entropy of the prediction results of the two teacher models for segmenting the unlabeled data in the source domain; Calculate the corresponding segmentation weights based on the entropy of the prediction results of the two teacher models; The dual-teacher prediction ensemble result is obtained by dividing the sum of the products of the prediction results of the two teacher models and the segmentation weights by 2. Calculate the entropy of the dual-teacher prediction ensemble result, and calculate the dual-teacher segmentation weights based on the entropy of the dual-teacher prediction ensemble result; Based on the style-enhanced image segmentation prediction results and the cross-entropy of pseudo-labeled samples and the bi-teacher segmentation weights, the confidence-perceived consistency loss is determined. Given the number of training iterations, calculate the sum of the above loss functions, update the network parameters through gradient descent of the objective function, and finally obtain the trained semi-supervised generalized medical image segmentation model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of a semi-supervised domain generalized medical image segmentation method as described in any one of claims 1-6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the semi-supervised domain generalized medical image segmentation method as described in any one of claims 1-6.

Images