A medical image segmentation method based on multi-domain semi-supervision
By adopting a multi-field semi-supervised method in the field of medical image segmentation, using teacher-student networks and contrast learning technology, the problem of lack of samples and labeling is solved, segmentation accuracy and generalization ability are improved, and medical image segmentation effect is achieved in multiple fields.
Patent Information
- Application Number
- CN202211130790.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-16
- Publication Date
- 2025-04-08
- Estimated Expiration
- 2042-09-16
AI Technical Summary
Due to the lack of samples and labeling, existing medical image segmentation methods are difficult to achieve accurate and general segmentation effects on multiple data modes, multiple organs and multiple diseases, resulting in insufficient generalization capabilities of model and difficult to extract common medical knowledge.
Using a multi-domain semi-supervised medical image segmentation method, a semi-supervised teacher student segmentation model is constructed, and a small amount of labeled data and a large amount of labelless data is used to extract the advanced semantic features of multi-domain medical images through interaction and contrast learning of teacher-student networks, and robust learning is performed through domain feature similarity screening and multi-domain advanced semantic comparison loss function.
It improves the accuracy and generalization ability of medical image segmentation, expands the applicability of multiple fields, realizes robust segmentation effects on unsupervised medical images in multiple fields, and promotes the improvement of the effect of depth models on image segmentation.
Smart Images
Figure CN115861164B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision and medical image fusion, and specifically relates to a medical image segmentation method based on multi-domain semi-supervision. Background Art
[0002] The goal of medical image segmentation is to divide and extract organs, tissues, lesions and other areas with visual homogeneity and semantic consistency in CT, MRI, X-ray and other images; it uses image processing, computer vision, machine learning and other methods to provide theoretical and application support for digital modeling and automated auxiliary diagnosis and treatment of diseases; as a key step in medical image intelligent analysis technology, medical image segmentation algorithms have shown broad clinical application prospects, and many scientific research institutions at home and abroad have carried out a lot of research in this field; in 2017, Stanford University developed a deep learning algorithm for diagnosing skin cancer; in 2018, the Google Brain team proposed a method for predicting the risk of cardiovascular and cerebrovascular diseases through retinal images; in recent years, the upper gastrointestinal tumor endoscopic auxiliary diagnosis system developed by my country's scientific research team has achieved great success; therefore, in view of actual clinical application scenarios and needs, research on medical image segmentation methods with accuracy, robustness and versatility, and then develop a new generation of intelligent clinical auxiliary diagnosis systems with independent intellectual property rights, which has very important practical significance in promoting the level and efficiency of clinical diagnosis and treatment in hospitals, improving the level of primary diagnosis and treatment under the multi-level diagnosis and treatment system, and enhancing the informationization strength of my country's medical system;
[0003] With the development of deep learning methods and the continuous research in related fields, deep network architectures with data-driven mechanisms have shown excellent feature learning and knowledge mining capabilities. In this wave of development, the field of medical image segmentation has also made great progress. However, due to the supervised feedback characteristics of deep neural networks, general natural image analysis tasks often require a large amount of training data to ensure the convergence of the model, which places high demands on the quantity and quality of training samples.
[0004] In contrast, in the field of medical image segmentation, due to differences in tissue physiology, the influence of organ movement, and different equipment parameters and specifications during image acquisition, various medical image data often have large appearance differences. On this basis, due to the sparsity of symptoms, data sensitivity, and relative closedness, and the fact that medical image labels require medical professionals to label them, it has always been difficult to obtain a large number of high-quality training samples. Now, in the field of natural image analysis, sample sets with a large amount of data and annotations have been obtained, and a large number of deep networks with extremely large parameter quantities have emerged. In the field of medical image analysis, it is precisely because of the various reasons analyzed above that researchers have to face the "small sample" problem of lack of samples and lack of annotations in various medical image analysis tasks at this stage.
[0005] Therefore, under the constraints of the small sample problem, most existing medical image segmentation methods are often limited to a single data modality, a single organ, a fixed disease, and a few verification data sets, which brings a series of disadvantages:
[0006] 1) Due to the lack of available data, it is difficult to use the data-driven mechanism of deep architecture to build accurate and reliable models;
[0007] 2) The generalization ability of the learning model is difficult to continuously improve, and its usability and versatility are restricted;
[0008] 3) The models for different medical image analysis tasks are relatively independent, which makes it difficult to effectively extract common medical knowledge. Summary of the invention
[0009] In order to solve the above technical problems, the present invention proposes a medical image segmentation method based on multi-domain semi-supervision, comprising:
[0010] Construct and train a semi-supervised teacher-student segmentation model, input the alienated domain data to be segmented into the trained segmentation model, and obtain the segmentation result;
[0011] The training process of the semi-supervised teacher-student segmentation model includes:
[0012] S1: Acquire source domain medical images, where the source domain medical images consist of a small amount of labeled data and a large amount of unlabeled data;
[0013] S2: Input the labeled data in the source domain medical image into the student segmentation network, extract the high-level tissue semantic features of the labeled data through the encoder, use the decoder to restore the feature channels of the high-level tissue semantic features, use the segmentation head MLP to predict the pixel-by-pixel prediction of each channel after restoration, binarize the predicted pixels through the Softmax function, and obtain the segmentation prediction confidence matrix. Perform supervised training based on the segmentation prediction confidence matrix and the labeled data in the source domain medical image;
[0014] S3: Use the exponential moving average algorithm to update the network parameters of the student network for supervised training to obtain the teacher network. The unlabeled data is input into the teacher network to obtain the segmentation prediction confidence matrix, which is processed by the Detach function to become the teacher pseudo label.
[0015] S4: Unlabeled data extracts high-level semantic features of unlabeled data through the student network encoder θ (X s ), f θ (X s ) is decomposed into three groups of independent semantic features Q, K, and V through group convolution operations, and high-level semantic features are extracted through one-dimensional convolution.θ (X s ), use Gaussian whitening function to make two groups of independent semantic features Q and K independent of each other, and embed them into binary morphological features M through transposed dot product. V groups of independent semantic features are operated with gate point G to form unary domain features D;
[0016] S5: Store the unary domain feature D into the queue and construct the domain memory storage unit Meta Bank;
[0017] S6: In Meta Bank, the nearest neighbor algorithm is used to extract the univariate domain feature D* that is similar to the current univariate domain feature D. D and D* are regarded as the current positive domain features, and the remaining features in Meta Bank are regarded as negative domain features. Contrastive learning is used to bring the positive domain features and the domain features closest to them in the domain memory storage unit Meta Bank closer to each other, and to isolate the positive and negative domain features from each other.
[0018] S7: Select the unary domain feature D* as the new domain feature, add it to the binary morphological feature M, and obtain the new high-level semantic feature F*;
[0019] S8: F* passes through the student segmentation network decoder and is re-decoded to obtain a feature map of the same size as the original input image. The feature map is classified by the segmentation head MLP to obtain the student segmentation confidence prediction, and pseudo-supervised training is performed with the teacher pseudo-label;
[0020] S9: Start the gradient back-propagation mechanism, optimize the model's loss function, and update the network parameters according to the exponential moving average algorithm. When the model converges or reaches the set number of epochs, the model training is completed and the model parameters are fixed.
[0021] Preferably, the predicted pixels are binarized by the Softmax function to obtain a segmentation prediction confidence matrix, which is expressed as:
[0022]
[0023] Among them, Y represents the segmentation prediction confidence matrix obtained by the student segmentation network, represents the student network encoder-decoder and segmentation head processing operations inside the student segmentation network, X represents the labeled data in the input source domain medical image, σ(·) represents the Softmax function, and τ represents the parameter that controls the differentiation of the segmentation result.
[0024] Preferably, supervised training is performed based on the segmentation prediction confidence matrix and labeled data in the source domain medical image, expressed as:
[0025]
[0026] in, represents the loss function of supervised training, T represents the total pixel size of labeled data in the source domain medical image, X represents the labeled data in the input source domain medical image, GT(X) represents the label of the labeled data in the source domain medical image, and Y represents the segmentation prediction confidence matrix obtained by the student segmentation network.
[0027] Preferably, the unlabeled data is input into the teacher network to obtain the segmentation prediction confidence matrix, and is processed by the Detach function to become the teacher pseudo label, which is expressed as:
[0028]
[0029] Among them, Y * represents the teacher pseudo label, represents the teacher network encoder-decoder and segmentation head processing operations inside the teacher network, Detach and Softmax function operations, X represents the labeled data in the input source domain medical image, and τ represents the parameter that controls the differentiation of the segmentation results.
[0030] Preferably, the two independent semantic features Q and K are made independent of each other using a Gaussian whitening function, and are embedded into a binary morphological feature M through a transposed dot product, which is expressed as:
[0031]
[0032] Among them, M represents the binary morphological feature, represents the independent semantic features after the transposed Gaussian whitening operation, Q i represents the i-th pixel in Q, K j represents the jth pixel in K, T represents the transposition operation, μ Q represents the mean of independent semantic features Q, μ K represents the mean of the independent semantic features K, Ω represents the set of all pixels in the feature map, t represents a specific threshold pixel, σ(·) represents the Softmax function, G represents the gating point of the high-level semantic features of unlabeled data, and exp() represents the exponential function.
[0033] Preferably, the V group of independent semantic features is subjected to a dot product operation with the gate point G to form a unary domain feature D, which is expressed as:
[0034]
[0035] in, Represents the dot product operation, V represents a set of independent semantic features decomposed by the grouped convolution operation, and G represents the gating point of the high-level semantic features of unlabeled data.
[0036] Preferably, the nearest neighbor algorithm is used to extract the unary term domain feature D* similar to the current unary term domain feature D, which is expressed as:
[0037] D*=L d (D) Meta
[0038] Among them, L d (·) Meta represents the shortest distance operation between metrics D and D*.
[0039] Preferably, contrastive learning is used to bring the positive domain features and the domain features closest to them in the domain memory storage unit Meta Bank closer to each other, and to isolate the positive and negative domain features from each other, which can be expressed as:
[0040]
[0041] in, represents the contrast loss function between the i-th pixel of the positive and negative domain features, exp(D i ·D* i / τ) represents the feature closeness term between the positive domain and the nearest neighbor domain, exp(·) represents the exponential function, and D i represents the i-th pixel in the positive domain feature, D* i represents the i-th pixel in the nearest neighbor domain feature, τ represents the differentiation parameter controlling the segmentation result, ∑ D`∈Meta Bank exp(D i ·D` i / τ) represents the positive and negative domain feature isolation term, D` represents the negative domain feature, and F` i is the i-th pixel of the negative domain feature.
[0042] Preferably, the student segmentation confidence prediction is pseudo-supervised with the teacher pseudo-label, expressed as:
[0043]
[0044] in, represents the pseudo-supervisory loss function, N represents the total pixel size of the unlabeled data in the source domain medical image, and Y b represents the prediction result of the student network, Represents the prediction result of the teacher network.
[0045] Preferably, the loss function of the model is expressed as:
[0046]
[0047] in, represents the overall loss function of the teacher-student network, represents the loss function of supervised training, represents the contrast loss function between the i-th pixel of the positive and negative domain features, is the pseudo-supervisory loss function.
[0048] The beneficial effects of the present invention are as follows: the present invention mines a large number of high-level semantic features of unlabeled multi-domain medical images through a teacher-student network, adopts a network with a self-attention disentanglement mechanism to extract domain features and segmentation part features, uses a domain feature similarity screening mechanism and a multi-domain high-level semantic contrast loss function for robust learning, introduces an exponential sliding average algorithm to heterogeneously transform the student model into a teacher model, and implements a pixel-level errata scheme under the constraint of teacher-student consistency. Finally, the segmentation area is obtained based on the student output judgment result, which improves the accuracy of segmentation and expands the applicability and generalization of multiple fields. The present invention can achieve robust segmentation effects in multi-domain unsupervised medical images, improve the effect of deep models on image segmentation, and promote the development of related technical fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 This is a structural diagram of the medical image segmentation method based on multi-domain semi-supervision of the present invention;
[0050] Figure 2 This is a segmentation result diagram of an embodiment of the present invention. DETAILED DESCRIPTION
[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0052] A medical image segmentation method based on multi-domain semi-supervised Figure 1 As shown, including:
[0053] Construct and train a semi-supervised teacher-student segmentation model, input the alienated domain data to be segmented into the trained segmentation model, and obtain the segmentation result;
[0054] The training process of the semi-supervised teacher-student segmentation model includes:
[0055] S1: Acquire source domain medical images, where the source domain medical images consist of a small amount of labeled data and a large amount of unlabeled data;
[0056] S2: Input the labeled data in the source domain medical image into the student segmentation network, extract the high-level tissue semantic features of the labeled data through the encoder, use the decoder to restore the feature channels of the high-level tissue semantic features, use the segmentation head MLP to predict the pixel-by-pixel prediction of each channel after restoration, binarize the predicted pixels through the Softmax function, and obtain the segmentation prediction confidence matrix. Perform supervised training based on the segmentation prediction confidence matrix and the labeled data in the source domain medical image;
[0057] S3: Use the exponential moving average algorithm to update the network parameters of the student network for supervised training to obtain the teacher network. The unlabeled data is input into the teacher network to obtain the segmentation prediction confidence matrix, which is processed by the Detach function to become the teacher pseudo label.
[0058] S4: Unlabeled data extracts high-level semantic features of unlabeled data through the student network encoder θ (X s ), f θ (X s ) is decomposed into three groups of independent semantic features Q, K, and V through group convolution operations, and high-level semantic features are extracted through one-dimensional convolution. θ (X s ), use Gaussian whitening function to make two groups of independent semantic features Q and K independent of each other, and embed them into binary morphological features M through transposed dot product. V groups of independent semantic features are operated with gate point G to form unary domain features D;
[0059] S5: Store the unary domain feature D into the queue and construct the domain memory storage unit Meta Bank;
[0060] S6: In Meta Bank, the nearest neighbor algorithm is used to extract the univariate domain feature D* that is similar to the current univariate domain feature D. D and D* are regarded as the current positive domain features, and the remaining features in Meta Bank are regarded as negative domain features. Contrastive learning is used to bring the positive domain features and the domain features closest to them in the domain memory storage unit Meta Bank closer to each other, and to isolate the positive and negative domain features from each other.
[0061] S7: Select the unary domain feature D* as the new domain feature, add it to the binary morphological feature M, and obtain the new high-level semantic feature F*;
[0062] S8: F* passes through the student segmentation network decoder and is re-decoded to obtain a feature map of the same size as the original input image. The feature map is classified by the segmentation head MLP to obtain the student segmentation confidence prediction, and pseudo-supervised training is performed with the teacher pseudo-label;
[0063] S9: Start the gradient back-propagation mechanism, optimize the model's loss function, and update the network parameters according to the exponential moving average algorithm. When the model converges or reaches the set number of epochs, the model training is completed and the model parameters are fixed.
[0064] The acquired multi-domain medical imaging data are cropped, and in this embodiment, the size is set to 255×255 pixels; different data enhancements are performed on them and input into the teacher and student networks. The specific enhancement methods include (Mixup, Cutmix, Gaussian white noise, etc.).
[0065] The predicted pixels are binarized through the Softmax function to obtain the segmentation prediction confidence matrix, which is expressed as:
[0066]
[0067] Among them, Y represents the segmentation prediction confidence matrix obtained by the student segmentation network, represents the student network encoder-decoder and segmentation head processing operations inside the student segmentation network, X represents the labeled data in the input source domain medical image, σ(·) represents the Softmax function, and τ represents the parameter that controls the differentiation of the segmentation result.
[0068] The network parameters of the student network for supervised training are updated using the exponential moving average algorithm, expressed as:
[0069] X′ t =αX′ t-1 +(1-α)X t
[0070] Among them, X′ t represents the teacher network parameters at the tth round, θ t The parameters of the student network at the tth round, α represents the weighting coefficient.
[0071] Supervised training is performed based on the segmentation prediction confidence matrix and labeled data in the source domain medical images, expressed as:
[0072]
[0073] in, represents the loss function of supervised training, T represents the total pixel size of labeled data in the source domain medical image, X represents the labeled data in the input source domain medical image, GT(X) represents the label of the labeled data in the source domain medical image, and Y represents the segmentation prediction confidence matrix obtained by the student segmentation network.
[0074] The unlabeled data is input into the teacher network to obtain the segmentation prediction confidence matrix, which is processed by the Detach function to become the teacher pseudo label, expressed as:
[0075]
[0076] Among them, Y * represents the teacher pseudo label, represents the teacher network encoder-decoder and segmentation head processing operations inside the teacher network, Detach and Softmax function operations, X represents the labeled data in the input source domain medical image, and τ represents the parameter that controls the differentiation of the segmentation results.
[0077] The Gaussian whitening function is used to make the two sets of independent semantic features Q and K independent of each other, and they are embedded into binary morphological features M through transposed dot products, which can be expressed as:
[0078]
[0079] Among them, M represents the binary morphological feature, represents the independent semantic features after the transposed Gaussian whitening operation, Q i represents the i-th pixel in Q, K j represents the jth pixel in K, T represents the transposition operation, μ Q represents the mean of independent semantic features Q, represents the mean of the independent semantic features K, Ω represents the set of all pixels in the feature map, t represents a specific threshold pixel, σ(·) represents the Softmax function, G represents the gating point of the high-level semantic features of unlabeled data, and exp() represents the exponential function.
[0080] V groups of independent semantic features are operated by dot product with the gate point G to form a unary domain feature D, which is expressed as:
[0081]
[0082] in, Represents the dot product operation, V represents a set of independent semantic features decomposed by the grouped convolution operation, and G represents the gating point of the high-level semantic features of unlabeled data.
[0083] The nearest neighbor algorithm is used to extract the univariate domain feature D* similar to the current univariate domain feature D, which is expressed as:
[0084] D*=L d (D) Meta
[0085] Among them, L d (·) Meta represents the shortest distance operation between metrics D and D*.
[0086] Use contrastive learning to bring the positive domain features closer to the domain features closest to them in the domain memory storage unit Meta Bank, and isolate the positive and negative domain features from each other, expressed as:
[0087]
[0088] in, represents the contrast loss function between the i-th pixel of the positive and negative domain features, exp(D i ·D* i / τ) represents the feature pull-up term between the positive domain and the nearest neighbor domain. Its purpose is to compactly represent similar domains in the domain library so that the network can perform clustering modeling on the same domain well. Exp(·) is an exponential level metric function, and D i is the i-th pixel in the positive domain feature, D* i is the i-th pixel in the nearest neighbor feature, τ represents the differentiation parameter controlling the segmentation result; D`∈Meta Bank exp(D i ·D` i / τ) is the positive and negative domain feature isolation item, which aims to structure the different domains in the domain library so that the network can robustly learn cross-domain perturbations. MetaBank is the domain memory storage unit Meta Bank, D` is all the features of non-positive domain features D and nearest neighbor domain features D* in Meta Bank, also called negative domain features, D` i is the i-th pixel of the negative domain feature.
[0089] The student segmentation confidence prediction is used together with the teacher pseudo-label for pseudo-supervised training, expressed as:
[0090]
[0091] in, is a pseudo-supervisory loss function, N represents the total pixel size of the unlabeled data in the source domain medical image, and Y b Predict results for the student network, Prediction results for the teacher network, represents the correct conditional entropy for prediction, It represents the prediction error conditional entropy. The sum of the above two conditional entropies is the pseudo-supervision loss.
[0092] The loss function of the model is expressed as:
[0093]
[0094] in, represents the overall loss function of the teacher-student network, represents the loss function of supervised training, represents the contrast loss function between the i-th pixel of the positive and negative domain features, is a pseudo-supervisory loss function. The above three sets of loss functions are added together to obtain the overall loss function of the current teacher-student network.
[0095] In S9, the network parameters are updated according to the exponential moving average algorithm, which is expressed as:
[0096] θ′ t =αθ′ t-1 +(1-α)θ t
[0097] Among them, θ′ t represents the network parameters updated by the model at the tth round, θ′ t-1 represents the parameters of the teacher network at the tth round, θ t represents the parameters of the student network at the tth round, and α represents the weighting coefficient.
[0098] A medical image segmentation system based on multi-domain semi-supervision, comprising: a data collection module, a downsampling module, a feature exchange module, an MLP module, a pooling module, a splicing module, a confidence calculation module, a Gaussian feature whitening module, a domain memory storage module, an attention module, a nearest neighbor candidate feature module, an exponential sliding average module, a matching module, and a consistency constraint module;
[0099] The data collection module is used to obtain multi-domain labeled medical images and multi-domain unlabeled medical images input into the model;
[0100] The downsampling module is used to filter out high-level semantic features;
[0101] The feature exchange module is used to extract the unary field features and the nearest neighbor unary field features, and then exchange the feature information of the two groups of fields;
[0102] The MLP module is used to extract the deep features of the original high-level semantics and the nearest neighbor high-level semantics based on the output of the feature segmentation module using multiple MLPs;
[0103] The pooling module is used to perform maximum pooling on the deep features to obtain the global features of the source domain and the global features of the target domain;
[0104] The splicing module is used to fuse the feature information of features, splicing the univariate domain features and the binary morphological features with the deep features of each point;
[0105] The confidence calculation module is used to input the splicing features of each pixel into the fully connected layer to obtain the confidence of each pixel;
[0106] The Gaussian feature whitening module is used to independently decouple multiple high-level semantic features;
[0107] The domain memory storage module is used to store unary domain features, and its internal data structure conforms to the queue storage mode;
[0108] The attention module is used to obtain the correlation information between features and enhance the feature representation. That is, the high-level semantic feature representation is processed according to the self-attention mechanism and the cross-attention mechanism to obtain four sets of decoupled independent feature representations.
[0109] The nearest neighbor candidate feature module is used to query another univariate domain feature that is most similar to the univariate domain feature in the domain memory storage module according to the nearest neighbor algorithm;
[0110] The exponential sliding average module is used to estimate the weight parameters of the student network in the previous round and update the weight parameters of the teacher network after averaging;
[0111] The matching module is used to obtain a corresponding point set matching the first point set in the target point cloud;
[0112] The consistency constraint module is used to introduce consistency constraints in the teacher-student model to obtain the teacher-student pseudo-supervision loss.
[0113] In one embodiment, the datasets (M&Ms) and (SCGM) are used for experiments. The multi-center, multi-vendor and multi-disease cardiac image segmentation (M&Ms) dataset: The M&Ms challenge dataset contains 320 subjects, who were scanned in 6 clinical centers in 3 different countries using 4 different magnetic resonance scanner vendors (Siemens, Philips, GE and Canon), i.e., domains A, B, C and D. Each subject only has end-systolic and end-diastolic phases annotated, and the voxel resolution ranges from 0.85×0.85×10 mm to 1.45×1.45×9.9 mm. Domain A contains 95 cases and domain B contains 125 cases. Domain C and domain D both contain 50 subjects.
[0114] Spinal Cord Gray Matter Segmentation (SCGM) Dataset: The SCGM
[38] data comes from four different medical centers using different MRI systems (Philips Achieve, Siemens Triple, Siemens Skyra), namely domains 1, 2, 3, and 4, with voxel resolution ranging from 0.25×0.25×2.5mm to 0.5×0.5×5mm; there are 10 labeled subjects and 10 unlabeled subjects in each domain.
[0115] like Figure 2As shown, example images and predicted segmentation masks for each model in different cases are shown. When the baseline model is trained with less labeled data, the performance drops significantly. In contrast, our model produces satisfactory segmentation masks in every case.
[0116] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A medical image segmentation method based on multi-domain semi-supervision, characterized in that: include: Construct and train a semi-supervised teacher-student segmentation model, input the alienated domain data to be segmented into the trained segmentation model, and obtain the segmentation result; The training process of the semi-supervised teacher-student segmentation model includes: S1: Acquire a source domain medical image and perform data enhancement on the acquired medical image, wherein the source domain medical image consists of a small amount of labeled data and a large amount of unlabeled data; S2: Input the labeled data in the source domain medical image after data enhancement into the student segmentation network, extract the high-level tissue semantic features of the labeled data through the encoder, use the decoder to restore the feature channels of the high-level tissue semantic features, use the segmentation head MLP to predict the pixel-by-pixel prediction of each channel after restoration, binarize the predicted pixels through the Softmax function, and obtain the segmentation prediction confidence matrix. Perform supervised training based on the segmentation prediction confidence matrix and the labeled data in the source domain medical image; S3: Use the exponential moving average algorithm to update the network parameters of the student network for supervised training to obtain the teacher network. The unlabeled data is input into the teacher network to obtain the segmentation prediction confidence matrix, which is processed by the Detach function to become the teacher pseudo label. S4: Unlabeled data extracts high-level semantic features of unlabeled data through the student network encoder θ (X s ), f θ (X s ) is decomposed into three groups of independent semantic features Q, K, and V through group convolution operations, and high-level semantic features are extracted through one-dimensional convolution. θ (X s ), use Gaussian whitening function to make two groups of independent semantic features Q and K independent of each other, and embed them into binary morphological features M through transposed dot product. V groups of independent semantic features are operated with gate point G to form unary domain features D; S5: Store the unary domain feature D into the queue and construct the domain memory storage unit Meta Bank; S6: In Meta Bank, the nearest neighbor algorithm is used to extract the univariate domain feature D* that is similar to the current univariate domain feature D. D and D* are regarded as the current positive domain features, and the remaining features in Meta Bank are regarded as negative domain features. Contrastive learning is used to bring the positive domain features and the domain features closest to them in the domain memory storage unit Meta Bank closer to each other, and to isolate the positive and negative domain features from each other. S7: Select the unary domain feature D* as the new domain feature, add it to the binary morphological feature M, and obtain the new high-level semantic feature F*; S8: F* passes through the student segmentation network decoder and is re-decoded to obtain a feature map of the same size as the original input image. The feature map is classified by the segmentation head MLP to obtain the student segmentation confidence prediction, and pseudo-supervised training is performed with the teacher pseudo-label; S9: Start the gradient back-propagation mechanism, optimize the model's loss function, and update the network parameters according to the exponential moving average algorithm. When the model converges or reaches the set number of epochs, the model training is completed and the model parameters are fixed.
2. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: The predicted pixels are binarized through the Softmax function to obtain the segmentation prediction confidence matrix, which is expressed as: Among them, Y represents the segmentation prediction confidence matrix obtained by the student segmentation network, represents the student network encoder-decoder and segmentation head processing operations inside the student segmentation network, X represents the labeled data in the input source domain medical image, σ(·) represents the Softmax function, and τ represents the parameter that controls the differentiation of the segmentation result.
3. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: Supervised training is performed based on the segmentation prediction confidence matrix and labeled data in the source domain medical images, expressed as: in, represents the loss function of supervised training, T represents the total pixel size of labeled data in the source domain medical image, X represents the labeled data in the input source domain medical image, GT(X) represents the label of the labeled data in the source domain medical image, and Y represents the segmentation prediction confidence matrix obtained by the student segmentation network.
4. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: The unlabeled data is input into the teacher network to obtain the segmentation prediction confidence matrix, which is processed by the Detach function to become the teacher pseudo label, expressed as: Among them, Y * represents the teacher pseudo label, represents the teacher network encoder-decoder and segmentation head processing operations inside the teacher network, Detach and Softmax function operations, X represents the labeled data in the input source domain medical image, and τ represents the parameter that controls the differentiation of the segmentation results.
5. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: The Gaussian whitening function is used to make the two sets of independent semantic features Q and K independent of each other, and they are embedded into binary morphological features M through transposed dot products, which can be expressed as: Among them, M represents the binary morphological feature, represents the independent semantic features after the transposed Gaussian whitening operation, Q i represents the i-th pixel in Q, K j represents the jth pixel in K, T represents the transposition operation, μ Q represents the mean of independent semantic features Q, μ K represents the mean of the independent semantic features K, Ω represents the set of all pixels in the feature map, t represents a specific threshold pixel, σ(·) represents the Softmax function, G represents the gating point of the high-level semantic features of unlabeled data, and exp() represents the exponential function.
6. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: V groups of independent semantic features are operated by dot product with the gate point G to form a unary domain feature D, which is expressed as: in, Represents the dot product operation, V represents a set of independent semantic features decomposed by the grouped convolution operation, and G represents the gating point of the high-level semantic features of unlabeled data.
7. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: The nearest neighbor algorithm is used to extract the univariate domain feature D* similar to the current univariate domain feature D, which is expressed as: D*=L d (D) Meta Among them, L d (·) Meta represents the shortest distance operation between metrics D and D*.
8. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: Use contrastive learning to bring the positive domain features closer to the domain features closest to them in the domain memory storage unit Meta Bank, and isolate the positive and negative domain features from each other, expressed as: in, represents the contrast loss function between the i-th pixel of the positive and negative domain features, exp(D i ·D* i / τ) represents the feature closeness term between the positive domain and the nearest neighbor domain, exp(·) represents the exponential function, and D i represents the i-th pixel in the positive domain feature, D* i represents the i-th pixel in the nearest neighbor domain feature, τ represents the differentiation parameter controlling the segmentation result, ∑ D`∈MetaBank exp(D i ·D` i / τ) represents the positive and negative domain feature isolation term, D` represents the negative domain feature, and D` i is the i-th pixel of the negative domain feature.
9. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: The student segmentation confidence prediction is used together with the teacher pseudo-label for pseudo-supervised training, expressed as: in, represents the pseudo-supervisory loss function, N represents the total pixel size of the unlabeled data in the source domain medical image, and Y b represents the prediction result of the student network, Represents the prediction result of the teacher network.
10. The medical image segmentation method based on multi-domain semi-supervision according to claim 1, characterized in that: The loss function of the model is expressed as: in, represents the overall loss function of the teacher-student network, represents the loss function of supervised training, represents the contrast loss function between the i-th pixel of the positive and negative domain features, is the pseudo-supervisory loss function.