Cross-domain medical image segmentation method based on feature decoupling and enhancement

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a cross-domain medical image segmentation method based on feature decoupling and enhancement, and the method comprises the steps: inputting a plurality of collected medical images of different centers into a trained medical image segmentation model which comprises a feature decoupling module and an inter-domain cooperative reasoning network; after an input image is processed in the feature decoupling module, domain irrelevant features and domain specific features are output, and then the inter-domain collaborative reasoning network predicts and outputs a segmentation result after inter-domain information flow guidance and pseudo boundary perception enhancement are carried out on the domain irrelevant features and the domain specific features; an inter-domain information flow guiding module based on bidirectional information flow is adopted for inter-domain information flow guiding, the boundary resolution of domain-independent features is improved through imaging details of domain specific features, and meanwhile domain noise of the domain specific features is restrained through anatomical structures of the domain-independent features. According to the method, the domain irrelevant features are sharper at the anatomical boundary, the domain specific features are purer in imaging details, and a solid foundation is laid for subsequent segmentation.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and particularly relates to a cross-domain medical image segmentation method based on feature decoupling and enhancement. Background Art

[0002] Medical image segmentation is a core technology for computer-aided diagnosis and treatment planning, and is of great value in applications such as tumor detection, organ contour extraction, and lesion localization. However, due to the diversity of imaging modalities (e.g., CT, MRI, X-ray) and acquisition protocols, cross-domain medical image segmentation faces significant domain migration challenges. These inter-domain differences primarily stem from the deep coupling between domain-invariant features of anatomical structures (such as organ geometry and topology) and domain-specific features of the imaging process (e.g., texture, noise level, and contrast). For example, MRI images may exhibit high soft tissue contrast, while CT images are more adept at displaying skeletal structures. This modality difference significantly reduces the generalization ability of traditional deep learning models in unseen domains. In recent years, domain adaptation methods and feature decoupling techniques have been widely explored in an attempt to alleviate this problem by separating domain-independent and domain-specific information. However, existing methods often lack targeted enhancement strategies after feature separation, resulting in a struggle to balance the sharpness of anatomical boundaries with the purity of imaging details. These methods perform poorly in scenarios with small sample sizes or high noise levels. In addition, although there has been progress in the application of pseudo-label generation and boundary perception modules, how to efficiently utilize disentangled features and avoid information loss remains an open challenge.

[0003] A specific analysis of the prior art shows the following defects:

[0004] 1. Existing technologies still oversimplify domain differences: Existing technologies are all focused on domain adaptation, but most of the methods they use (such as boundary contrast, GANs, and image translation) still regard domain differences as "noise" that needs to be eliminated, without fully exploring and utilizing the valuable domain-specific information that domain differences may contain. For example, boundary contrast methods may focus on aligning boundary features, but may ignore the differences and potential value of other domain-specific features such as texture and contrast in images of different modalities.

[0005] 2. Existing technologies have limitations in global feature alignment: Existing technologies still rely to some extent on the concept of global feature alignment. Global alignment can distort or lose important domain-specific information, especially in cross-modal medical image segmentation, where different modalities inherently carry crucial diagnostic information. For example, contrastive learning aligns feature spaces by bringing positive samples closer together and negative samples further apart, while GANs use adversarial training to align the feature distribution of generated images with that of the target domain. These methods lack fine-grained distinction between features, easily leading to confusion between domain-invariant and domain-specific features.

[0006] 3. Existing technologies ignore the potential value of domain-specific information: Existing technologies do not fully consider the potential value of domain-specific information in improving segmentation performance. For example, in cross-modal medical image segmentation, although CT and MRI images are different modalities, their respective imaging characteristics (such as CT's sensitivity to bone tissue and MRI's sensitivity to soft tissue) can complement each other and jointly improve segmentation accuracy. If only pursuing domain invariance while ignoring this modality-specific information, the performance ceiling of the model will be limited. Summary of the Invention

[0007] The purpose of the present invention is to provide a cross-domain medical image segmentation method based on feature decoupling and enhancement, which is used to solve the technical problem that even if domain adaptive migration and feature decoupling are performed in the existing technology in cross-domain medical image segmentation, the sharpness of the anatomical boundaries and the purity of the imaging details in the segmentation results are difficult to balance due to the lack of targeted enhancement strategies, and there is also a technical problem that information is lost due to the decoupling of features.

[0008] The cross-domain medical image segmentation method based on feature decoupling and enhancement inputs medical images collected from multiple different centers into a trained medical image segmentation model. The medical image segmentation model includes a feature decoupling module and an inter-domain collaborative reasoning network. The input image is processed in the feature decoupling module to output domain-independent features and domain-specific features. Then, the inter-domain collaborative reasoning network predicts and outputs segmentation results for the domain-independent features and domain-specific features through inter-domain information flow guidance and pseudo-boundary perception enhancement; the inter-domain information flow guidance adopts an inter-domain information flow guidance module based on bidirectional information flow, which uses the imaging details of domain-specific features to enhance the boundary resolution of domain-independent features, and at the same time uses the anatomical structure of domain-independent features to constrain the domain noise of domain-specific features; the pseudo-boundary perception enhancement uses the structural information of the boundary enhancement feature and the texture information of the noise suppression feature to collaboratively generate a boundary weight map, and refines the result through pseudo-supervision.

[0009] Preferably, the process of utilizing the imaging details of domain-specific features to enhance the boundary resolution of domain-independent features includes:

[0010] First, the domain-specific feature F spec Perform the following attention mechanism operation, namely:

[0011] A spec =sigmoid(Conv 1×1 (BN(F spec )))

[0012] Among them, BN is the batch normalization layer, Conv 1×1 is a 1×1 convolutional layer, A spec is the output of the attention mechanism;

[0013] The domain-independent features F are then enhanced by attention weighting and residual convolutioninv The boundary details are as follows:

[0014]

[0015] Among them, LN is the layer normalization layer, ⊙ represents element-by-element multiplication, Conv 3×3 is a 3×3 convolutional layer, δ is a small weight factor, is the domain-independent feature F inv Boundary enhancement features are obtained after boundary detail enhancement.

[0016] Preferably, the process of constraining the domain noise of domain-specific features using the anatomical structure of domain-independent features includes:

[0017] Domain-independent feature F inv Extracting pseudo boundary prior P of anatomical structure through multi-scale gradient operator edge , the calculation formula is as follows:

[0018]

[0019] in, Represents the Sobel operator, which is used to calculate the spatial gradient of the feature map, where k represents the scale of different convolution kernels; P edge Pseudo boundary prior P edge As a pseudo-supervisory signal, combined with depthwise separable convolution to generate a smooth constraint graph C inv , the corresponding calculation formula is as follows:

[0020] C inv =DSConv 3×3 (sigmoid(P edge ))

[0021] Among them, DSConv 3×3 represents a 3×3 depth-wise separable convolution, and sigmoid represents a sigmoid function;

[0022] The final suppression domain-specific feature F spec The noise in is calculated as follows:

[0023]

[0024] Where α represents a hyperparameter that is adaptively adjusted by gradient descent to ensure that the constraint is not too strong, and tanh represents the hyperbolic tangent function; Represents domain-specific features F spec Noise suppression features after noise suppression.

[0025] Preferably, the specific process of enhancing pseudo-boundary perception includes:

[0026] Enhance the boundary feature and noise suppression features Fusion, get the fusion feature F fuse , fusion feature F fuse Then generate the initial segmentation prediction P through a 3×3 convolution layer init ;

[0027] Pseudo-supervised refinement involves a lightweight pseudo-supervised iterative process that transforms the initial segmentation prediction P init As pseudo labels, combined with the gradient information of the original input image, the model's ability to perceive boundary areas is optimized through several iterations.

[0028] Preferably, the optimization process of each iteration of pseudo-supervised refinement includes:

[0029] 1) Boundary gradient extraction: For the input image I, calculate its gradient map G and compare the gradient map G with the initial prediction P init Perform element-level weighted fusion to obtain the boundary enhancement prediction feature F edge , the calculation formula is: F edge =P init (1 + α G), where α is a weighting coefficient used to control the strength of the gradient information;

[0030] 2) Pseudo-label generation and update: The prediction of the current round is used as the pseudo-label to supervise the training of the next round of iterations; the update rule of the pseudo-label is:

[0031]

[0032] in, represents the updated prediction after the t-1th iteration, represents the prediction of the tth round used as pseudo label, Represents the boundary enhancement prediction feature F corresponding to the t-th round of iterative prediction edge , Conv represents the convolution operation, Softmax is the Softmax function, represents the updated prediction after the tth iteration;

[0033] After several iterations of pseudo-supervision iteration, the final boundary refinement prediction P is obtained. refine .

[0034] Preferably, the inter-domain collaborative reasoning network also includes a post-processing step to implement cross-domain consistency constraints, using a fully connected conditional random field to refine the boundary prediction P refine After further optimization, the energy function of the fully connected conditional random field is:

[0035]

[0036] Among them, θ i (x i )=-logP refine (x i ), is a unary potential function, θ ij (x i , x j ) is a bilateral potential function based on the color and spatial distance between image tables, E(x) represents the output of the energy function, and the inference process of the fully connected conditional random field is completed through the mean field approximation algorithm to output the final segmentation result.

[0037] Preferably, the feature decoupling module preprocesses the medical image data of multiple different centers and obtains the feature F through an encoder. Then, a dynamic soft mask strategy is introduced to achieve fine feature separation, and a feature response map M is generated through a learnable convolutional layer:

[0038] M=σ(W m *F+b m )

[0039] Where σ is the sigmoid activation function, ensuring that the mask value range is between [0, 1], W m and b m They represent the weights and bias terms of the learnable convolutional layer, respectively, and the response map M is the dynamic soft mask. The feature map is processed using a dual-channel adaptive feature routing mechanism to decompose the input features into domain-independent features and domain-specific features. The expression of the dual-channel adaptive feature routing mechanism is as follows:

[0040] F inv =F⊙(1-M)+ResBlock inv (F⊙(1-M))

[0041] F spec =F⊙M+ResBlock spec (F⊙M)

[0042] Among them, ⊙ represents element-by-element multiplication, ResBlock inv and ResBlock spec Two dedicated residual blocks adopt multi-scale receptive field structures to enhance the expression capabilities of invariant features and specific features respectively; inv represents domain-independent features, F spec Represents domain-specific features.

[0043] Preferably, the feature decoupling module also includes a bidirectional attention enhancement mechanism based on mutual information minimization, which first uses the domain-independent feature F inv and domain-specific features F specCalculate the feature channel correlation matrix A, and then feed the correlation matrix A back to each feature branch. The expression is as follows:

[0044] F inv =F inv +γ inv ChannelAttention(F inv , A)

[0045] F spec =F spec +γ spec SpatialAttention(F spec , A)

[0046] Among them, γ inv and γ spec It is a learnable balance parameter used to dynamically adjust the intensity of attention influence. ChannelAttention represents channel attention, and SpatialAttention represents spatial attention.

[0047] Preferably, the feature decoupling module introduces a gradient modulation mechanism based on mutual information minimization, which is expressed as follows:

[0048]

[0049] Among them, I(F inv , F spec ) represents the mutual information between domain-independent features and domain-specific features; is the mutual information gradient of domain-independent features, is the mutual information gradient of domain-specific features; sign(·) is the sign function, which means the modulation strength of mutual information in the gradient direction is retained, λ MI represents the strength of the control mutual information constraint; |·| 1+∈ Indicates adaptive adjustment of the gradient through the exponent 1+∈; and denote the initial gradient and modulated gradient of the domain-independent feature, and represent the initial gradient and modulated gradient of domain-specific features, respectively.

[0050] Preferably, the mutual information I(F inv , F spec ) is estimated using a neural estimator based on noise contrast estimation, which is expressed as:

[0051]

[0052] Among them, T θ It is a parameterized neural network with feature pairs as input and scalar scores as output; the joint distribution expectation Represents the feature pairs extracted from the same sample (F inv and F spec )Find the expected marginal distribution expectation; Represents the feature pairs extracted from different samples (F inv and F spec ) Find the expectation maximization operation; represents the expression that is maximized by optimizing the network parameters θ.

[0053] The present invention has the following advantages:

[0054] 1. This paper proposes an inter-domain information flow guidance module. Based on bidirectional information flow guidance, it achieves mutual enhancement of domain-independent and domain-specific features during the inference process. Specifically, it uses the detailed information of domain-specific features to enhance the boundary resolution of domain-independent features, and uses the structural information of domain-independent features to constrain the domain noise of domain-specific features. This achieves the coordinated utilization and complementary enhancement of domain-independent and domain-specific features, avoiding the blindness of simple fusion. It makes domain-independent features sharper at anatomical boundaries and domain-specific features purer in imaging details, laying a solid foundation for subsequent segmentation and resolving technical problems existing in the prior art.

[0055] 2. This paper proposes a pseudo-boundary perception enhancement module that improves segmentation accuracy through dynamic boundary prediction and iterative optimization. Building on the inter-domain information flow guidance module, this module leverages the structural information of domain-independent features and the texture information of domain-specific features to collaboratively generate a boundary weight map and iteratively refine the segmentation results through pseudo-supervision. This significantly improves the boundary accuracy of the segmentation results, especially in small sample scenarios. Through iterative optimization, the model can more accurately perceive and segment boundary areas, improving overall segmentation performance.

[0056] 3. This paper abandons the traditional approach of treating domain differences as simple "noise" and innovatively proposes a dynamic soft masking strategy to adaptively decouple input image features into domain-independent and domain-specific features. Unlike hard threshold segmentation, soft masking allows for smooth feature transitions, more consistent with the continuous distribution of natural image features. Combined with innovative improvements to the inter-domain collaborative inference network, this approach avoids feature confusion and information loss that can result from global feature alignment, laying the foundation for the effective utilization of subsequent domain-specific information and achieving a more natural and refined feature separation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a basic flow chart of the cross-domain medical image segmentation method based on feature decoupling and enhancement of the present invention.

[0058] Figure 2 It is a schematic diagram of the feature decoupling module in the present invention.

[0059] Figure 3 It is a schematic diagram of the inter-domain collaborative reasoning network in the present invention.

[0060] Figure 4 It is a flow chart of the structural constraints of the inter-domain information flow guidance module from domain-independent features to domain-specific features in the present invention.

[0061] Figure 5 This is an effect diagram of processing fundus image segmentation using the present invention. DETAILED DESCRIPTION

[0062] The following is a further detailed description of the specific implementation methods of the present invention through the description of the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0063] like Figures 1-4 As shown, the present invention provides a cross-domain medical image segmentation method based on feature decoupling and enhancement. Medical images collected from multiple different centers are input into a trained medical image segmentation model. The medical image segmentation model includes a feature decoupling module (FDM) and a domain-specific collaborative inference network (DSIN). The input image is processed in the feature decoupling module and outputs domain-independent features and domain-specific features. The domain-specific collaborative inference network then predicts and outputs segmentation results based on the domain-independent features and domain-specific features through inter-domain information flow guidance and pseudo-boundary perception enhancement. This effectively improves the accuracy and robustness of the segmentation results.

[0064] 1. Feature Decoupling Module (FDM): Traditional domain adaptation methods (such as adversarial training) attempt to eliminate domain differences through global feature alignment, and regard domain differences as simple "noise", ignoring the valuable information that may be contained therein. To this end, this solution improves the feature decoupling module so that it can achieve natural decoupling through adaptive learning rather than artificial rules, and sequentially performs dynamic soft mask generation, feature decomposition, bidirectional attention enhancement, and optimization constraints. The feature decoupling module first uses dynamic soft masks and a dual-channel adaptive feature routing mechanism to decompose the input features into domain-independent features and domain-specific features, and optimizes the decoupling quality through mutual information minimization and bidirectional attention enhancement. Finally, the input image features are decomposed into domain-independent features F inv and domain-specific features F spec .

[0065] The specific process of image processing by the feature decoupling module is as follows:

[0066] First, the feature decoupling module preprocesses medical image data from multiple different centers and then processes them through an encoder to obtain features F. Then, a dynamic soft masking strategy is introduced to achieve fine feature separation, and a learnable convolutional layer is used to generate a feature response map M:

[0067] M=σ(W m *F+b m )

[0068] Where σ is the sigmoid activation function, ensuring that the mask value range is between [0,1], W m and b m where represents the weights and biases of the learnable convolutional layer, respectively. This response map M essentially captures the sensitivity distribution of each position in the feature map F to domain information, providing the basis for feature decoupling, namely the dynamic soft mask required for subsequent processing. Unlike traditional hard threshold segmentation methods, soft masking allows for smooth feature transitions, which is more consistent with the continuous distribution characteristics of natural image features.

[0069] Based on the generated dynamic soft mask M, the feature decoupling module further uses the dual-channel adaptive feature routing mechanism to process the feature map and decompose the input features into domain-independent features and domain-specific features. The expression of the dual-channel adaptive feature routing mechanism is as follows:

[0070] F inv =F⊙(1-M)+ResBlock inv (F⊙(1-M))

[0071] F spec =F⊙M+ResBlock spec (F⊙M)

[0072] Among them, ⊙ represents element-by-element multiplication, ResBlock inv and ResBlock spec Two dedicated residual blocks adopt multi-scale receptive field structures to enhance the expression capabilities of invariant features and specific features respectively; inv represents domain-independent features, F spec Represents domain-specific features. Among them, the multi-scale residual block ResBlock inv It contains three parallel branches, using 1×1, 3×3 and 5×5 three hole convolutions to collect information of different scales, and enhance the ability to capture cross-domain common features (i.e., domain-invariant features); and the densely connected residual block ResBlock spec A densely connected structure is used to preserve domain-specific details (i.e., domain-specific features). The asymmetric design of the dual-channel adaptive feature routing mechanism reflects the different properties of domain-invariant features and domain-specific features.

[0073] To further improve the decoupling quality, the feature decoupling module also includes a bidirectional attention enhancement mechanism based on mutual information minimization. First, the feature channel correlation matrix is calculated as follows:

[0074]

[0075] Among them, d is the feature dimension, softmax is the softmax function, That is, domain-independent feature F inv The transpose of F spec Represents domain-specific features, and the correlation matrix A captures the implicit dependency between the two features. Afterwards, the correlation matrix A is fed back to each feature branch, and the expression is as follows:

[0076] F inv =F inv +γ inv ChannelAttention(F inv , A)

[0077] F spec =F spec +γ spec SpatialAttention(F spec , A)

[0078] Among them, γ inv and γ spec ChannelAttention is a learnable balancing parameter used to dynamically adjust the strength of attention. It represents channel attention and SpatialAttention represents spatial attention. Here, channel attention enhances domain-invariant features by reconstructing inter-channel relationships, while spatial attention strengthens the local sensitivity of domain-specific features.

[0079] The entire decoupling process of the feature decoupling module is subject to orthogonal constraints and feature enhancement: To ensure the effective separation of the two types of features, the feature decoupling module also introduces an orthogonal constraint based on the Wasserstein distance. The corresponding loss function is as follows:

[0080]

[0081] in, represents the Wassertein-2 distance, p(F inv ) represents the distribution of domain-independent features, p(F spec ) represents the distribution of domain-specific features, so the first term It is to promote the separation of feature distribution in Wasserstein space; Represents KL divergence, which is used to measure p(F inv ) and p(F spec ), λ div Indicates that KL divergence is in loss The weight coefficient in , therefore, the second term It increases the distribution difference through KL divergence.

[0082] The feature decoupling module further introduces a gradient modulation mechanism based on mutual information minimization to ensure that the two types of features are decoupled at the information theory level. The expression is as follows:

[0083]

[0084] Among them, I(F inv , F spec ) represents the mutual information between domain-independent features and domain-specific features; is the mutual information gradient of the domain-independent feature, which represents the gradient of the mutual information to the domain-independent feature. is the mutual information gradient of the domain-specific feature, which indicates the gradient of the mutual information with respect to the domain-specific feature. The two are used to indicate how to adjust the feature to reduce the mutual information; sign(·) is the sign function, which indicates the modulation strength of the mutual information in the gradient direction. MI represents the strength of the control mutual information constraint; |·| 1+∈ It means that the exponent 1+∈ is used to amplify large gradients and suppress small gradients, thus achieving adaptive adjustment of gradients. and denote the initial gradient and modulated gradient of the domain-independent feature, and represent the initial gradient and modulated gradient of domain-specific features, respectively.

[0085] Mutual Information I(F inv , F spec ) is estimated using a neural estimator based on noise contrast estimation, which is expressed as:

[0086]

[0087] Among them, T θ is a parameterized neural network with the input being a feature pair (F inv and F spec ), the output is a scalar score; the joint distribution expectation Represents the feature pairs extracted from the same sample (F inv and F spec )Find the expected marginal distribution expectation; Represents the feature pairs extracted from different samples (F inv and F spec )Find the expectation-maximizing operation. It represents the expression that is maximized by optimizing the network parameters θ. The feature decoupling module performs approximate mutual information estimation through the above algorithm.

[0088] The training objective function of the entire feature decoupling module is:

[0089]

[0090] in, Denotes the adversarial loss of the feature decoupling module, which is optimized to ensure the domain-invariant feature F inv The distribution in the source and target domains is consistent, while the domain-specific features F spec The mutual information constraint of domain difference information is retained; I(F inv , F spec ) represents minimizing the information overlap of the two feature subspaces, λ mii is the corresponding weight parameter; is the loss of orthogonality constraint, λ orth To correspond to the weight parameters, optimizing this loss is used to promote the separation of features in the representation space; is the reconstruction loss, λ recon To correspond to the weight parameters, the loss is optimized to ensure that the feature decoupling process does not lose useful information.

[0091] 2. Inter-domain Coordinated Inference Network (DSIN): The inter-domain collaborative inference network processes the domain-independent features and domain-specific features output by the feature decoupling module. Through multi-level feature interaction, it enhances pseudo-boundary perception and cross-domain consistency constraints, and directly transforms domain-independent features and domain-specific features into high-precision and robust segmentation results. The inter-domain coordinated inference network not only makes full use of F inv and F spec In addition to the complementarity of the proposed method, a dynamic reasoning mechanism and pseudo-supervision strategy are introduced, which enables it to perform well in image segmentation tasks in the form of small samples, noisy data, and multimodal (such as CT, MRI) cross-domain.

[0092] The innovative inspiration of the inter-domain coordinated reasoning network comes from the collaborative processing of structure and details in human cognition. It abandons the simple splicing or weighted fusion of features in traditional methods, and instead builds an efficient feature utilization framework through three modules: guiding inter-domain information flow, enhancing pseudo-boundary perception, and implementing cross-domain consistency constraint post-processing.

[0093] The specific process of inter-domain coordinated reasoning network processing is as follows:

[0094] Inter-domain Information Flow Guidance (DIF): In order to make the domain-independent feature F inv and domain-specific features F spec Mutual enhancement in reasoning, the inter-domain coordinated reasoning network includes a domain-specific information flow guidance (DIF) module based on bidirectional information flow, which dynamically adjusts the expressive power of the two features through conditional dependencies. This mechanism avoids the blindness of traditional fusion and makes full use of domain-specific features F spec Imaging details enhance domain-independent features F inv The boundary resolution is achieved by using the domain-independent feature F invThe anatomical structure constrains domain-specific features F spec domain noise.

[0095] The inter-domain information flow guidance module is used to realize the spec To domain-independent feature F inv The details of the feature F are first enhanced by spec Perform the following attention mechanism operation, namely:

[0096] A spec =sigmoid(Conv 1×1 (BN(F spec )))

[0097] Among them, BN is the batch normalization layer, Conv 1×1 is a 1×1 convolutional layer, A spec is the output of the attention mechanism. The attention mechanism is used to extract domain-specific features F spec The domain-specific information related to the segmentation task (such as texture details) is used to enhance the domain-independent features F inv The boundary of is then used to provide guidance. The domain-independent feature F is then enhanced through attention weighting and residual convolution. inv The boundary details are as follows:

[0098]

[0099] Among them, LN is the layer normalization layer, ⊙ represents element-by-element multiplication, Conv 3×3 It is a 3×3 convolutional layer, δ is a small weight factor (the default value is 0.1 in this example), is the domain-independent feature F inv The boundary enhancement feature is obtained after the boundary details are enhanced. The above boundary detail enhancement method introduces a small amount of domain-specific features F spec Direct contribution to enhance cross-domain adaptability.

[0100] In order to achieve the domain-independent feature F inv To the domain-specific feature F spec The structural constraints, domain-independent features F inv Extracting pseudo boundary prior P of anatomical structure through multi-scale gradient operator edge , the calculation formula is as follows:

[0101]

[0102] in, Represents the Sobel operator, which is used to calculate the spatial gradient of the feature map, where k∈{1, 3} represents the scale of different convolution kernels of 1×1 and 3×3; Represents the domain-independent feature F invAfter applying the Sobel operator, the absolute value is taken to obtain the gradient magnitude of a single scale to represent the edge strength; It means adding the gradient amplitudes of different scales 1×1 and 3×3, thereby fusing multi-scale boundary information to form a more robust pseudo boundary prior P edge .

[0103] Pseudo-boundary prior P edge Captures domain-independent features F inv The anatomical boundary information in is used as a pseudo-supervisory signal to constrain the domain-specific features F spec Noise in the pseudo-boundary prior P edge Combined with depthwise separable convolution to generate a smooth constraint graph C inv , the corresponding calculation formula is as follows:

[0104] C inv =DSConv 3×3 (sigmoid(P edge ))

[0105] Among them, DSConv 3×3 represents a 3×3 depthwise separable convolution, and sigmoid represents the sigmoid function.

[0106] The final suppression domain-specific feature F spec The noise in is calculated as follows:

[0107]

[0108] Among them, α represents the hyperparameter adaptively adjusted by gradient descent, which ensures that the constraint is not too strong, tanh represents the hyperbolic tangent function; ⊙ represents element-by-element multiplication, Represents domain-specific features F spec Noise suppression features after noise suppression.

[0109] This bidirectional guidance mechanism enables boundary enhancement features Sharper at anatomical boundaries and let noise suppress features The imaging details are purer, laying a solid foundation for subsequent segmentation.

[0110] Pseudo boundary perception enhancement: The inter-domain coordinated reasoning network also includes a pseudo boundary perception module (PBM), which improves segmentation accuracy through dynamic boundary prediction and iterative optimization. The core of the pseudo boundary perception module is to use boundary enhancement features. Structural information and noise suppression features The texture information of the image is used to collaboratively generate a boundary weight map and the result is refined through pseudo-supervision.

[0111] The implementation of pseudo boundary perception enhancement relies on a lightweight pseudo supervision iterative process, the specific process includes: first, the boundary enhancement feature and noise suppression features Fusion, get the fusion feature F fuse , the corresponding formula is: Concat represents the connection operation, fusion feature F fuse Then generate the initial segmentation prediction P through a 3×3 convolution layer init This initial segmentation predicts P init Typically, images are rough, especially at the edges, where blur or offset may occur. Therefore, this method uses these as pseudo-labels and combines them with the gradient information of the original input image to further optimize the model's ability to perceive these edges. The optimization process requires two to three iterations, each of which includes the following steps.

[0112] 1) Boundary gradient extraction: For the input image I, calculate its gradient map Extract the high-frequency information of the edge area. Combine the gradient map G with the initial prediction P init Perform element-level weighted fusion to obtain the boundary enhancement prediction feature F edge , the calculation formula is: F edge =P init ·(1+α·G), where α is a weighting coefficient, which is 0.3 in the embodiment and is used to control the intensity of the gradient information.

[0113] 2) Pseudo-label generation and update: The prediction of the current round is used as the pseudo-label to supervise the training of the next round of iteration. The update rule of the pseudo-label is:

[0114]

[0115] in, represents the updated prediction after the t-1th round (previous round), represents the prediction of round t (the current round) used as pseudo-label, Represents the boundary enhancement prediction feature F corresponding to the t-th round of iterative prediction edge , Conv represents the convolution operation, Softmax is the Softmax function, represents the updated prediction after the tth round (the current round) iteration.

[0116] 3) Loss function: To ensure the effectiveness of pseudo-supervision, a composite loss function is used to optimize the model in each iteration. The loss function is:

[0117]

[0118] Where CE is the cross entropy loss, Dice is the Dice coefficient loss, and β is the weight of the Dice loss, which is 0.7 in this embodiment. By fusing features F fuse Generate directly, and then use the updated prediction obtained after the previous iteration

[0119] After 2 to 3 iterations, the pseudo boundary perception module (PBM) significantly improves the boundary accuracy, especially in small sample scenarios. Experiments show that after more than 3 iterations, the performance improvement tends to saturate, so t = 2 is selected as the default setting in this embodiment to balance efficiency and accuracy. After completing the pseudo supervision iteration, the final boundary refinement prediction P is obtained. refine .

[0120] Segmentation result generation: To further improve the smoothness and consistency of the segmentation results, the inter-domain coordination reasoning network also introduces a post-processing step to implement cross-domain consistency constraints, using the fully connected conditional random field (CRF) to refine the boundary prediction P refine Further optimization is performed. The energy function of the fully connected conditional random field (CRF) is:

[0121]

[0122] Among them, θ i (x i )=-log P refine (x i ), is a unary potential function, θ ij (x i , x j ) is a bilateral potential function based on the color and spatial distance between image tables, and E(x) represents the output of the energy function. The inference process of the fully connected conditional random field (CRF) is completed by the mean field approximation algorithm, and the final segmentation result is output. If the experiment verifies that no post-processing step is required, the boundary refinement prediction P is directly used. refine as the final segmentation result.

[0123] In the training phase of the final segmentation result, this method uses the following loss function for end-to-end optimization:

[0124]

[0125] in, is the segmentation result loss, Y represents the true label, λ1 and λ2 are the cross entropy CE(P final , Y) and Dice loss (1-Dice(P final ,Y)), in this embodiment, λ1 and λ2 take values of 1.0 and 0.5 respectively.

[0126] This method can be used for multimodal image pathology collaborative analysis tasks, such as fundus image segmentation (e.g. Figure 5 As shown in the figure, cross-modal lesion localization of cardiac ultrasound and CT, fusion segmentation of multi-sequence MRI of brain tumors, and joint diagnosis of heterogeneous data of X-ray and tissue slice images.

[0127] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the inventive concept and technical solution of the present invention, or the inventive concept and technical solution are directly applied to other occasions without improvement, they are all within the scope of protection of the present invention.

Claims

1. A cross-domain medical image segmentation method based on feature decoupling and enhancement, characterized by: Medical images collected from multiple different centers are input into a trained medical image segmentation model, which includes a feature decoupling module and an inter-domain collaborative inference network. The input image is processed in the feature decoupling module to output domain-independent features and domain-specific features. The inter-domain collaborative inference network then predicts and outputs segmentation results for domain-independent features and domain-specific features through inter-domain information flow guidance and pseudo-boundary perception enhancement. The inter-domain information flow guidance adopts an inter-domain information flow guidance module based on bidirectional information flow, which uses the imaging details of domain-specific features to enhance the boundary resolution of domain-independent features, and at the same time uses the anatomical structure of domain-independent features to constrain the domain noise of domain-specific features. The pseudo-boundary perception enhancement uses the structural information of boundary enhancement features and the texture information of noise suppression features to collaboratively generate a boundary weight map, and refines the results through pseudo-supervision.

2. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 1, characterized in that: The process of using the imaging details of domain-specific features to improve the boundary resolution of domain-independent features includes: first, the domain-specific features F spec Perform the following attention mechanism operation, namely: A spec =sigmoid(Conv 1×1 (BN(F spec ))) Among them, BN is the batch normalization layer, Conv 1×1 is a 1×1 convolutional layer, A spec is the output of the attention mechanism; The domain-independent features F are then enhanced by attention weighting and residual convolution inv The boundary details are as follows: Among them, LN is the layer normalization layer, ⊙ represents element-by-element multiplication, Conv 3×3 is a 3×3 convolutional layer, δ is a small weight factor, is the domain-independent feature F inv Boundary enhancement features are obtained after boundary detail enhancement.

3. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 2, characterized in that: The process of using the anatomical structure of domain-independent features to constrain the domain noise of domain-specific features includes: Domain-independent feature F inv Extracting pseudo boundary prior P of anatomical structure through multi-scale gradient operator edge , the calculation formula is as follows: in, Represents the Sobel operator, which is used to calculate the spatial gradient of the feature map, where k represents the scale of different convolution kernels; P edge Pseudo boundary prior P edge As a pseudo-supervisory signal, combined with depthwise separable convolution to generate a smooth constraint graph C inv , the corresponding calculation formula is as follows: C inv =DSConv 3×3 (sigmoid(P edge )) Among them, DSConv 3×3 represents a 3×3 depth-wise separable convolution, and sigmoid represents a sigmoid function; The final suppression domain-specific feature F spec The noise in is calculated as follows: Where α represents a hyperparameter that is adaptively adjusted by gradient descent to ensure that the constraint is not too strong, and tanh represents the hyperbolic tangent function; Represents domain-specific features F spec Noise suppression features after noise suppression.

4. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 3, characterized in that: The specific process of enhancing pseudo-boundary perception includes: Enhance the boundary feature and noise suppression features Fusion, get the fusion feature F fuse , fusion feature F fuse Then generate the initial segmentation prediction P through a 3×3 convolution layer init ; Pseudo-supervised refinement involves a lightweight pseudo-supervised iterative process that transforms the initial segmentation prediction P init As a pseudo-label, combined with the gradient information of the original input image; through several iterations, the model's perception of the boundary area is optimized.

5. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 4, characterized in that: The optimization process for each iteration of pseudo-supervised refinement includes: 1) Boundary gradient extraction: For the input image I, calculate its gradient map G and compare the gradient map G with the initial prediction P init Perform element-level weighted fusion to obtain the boundary enhancement prediction feature F edge , the calculation formula is: F edge =P init (1 + α G), where α is a weighting coefficient used to control the strength of the gradient information; 2) Pseudo-label generation and update: The prediction of the current round is used as the pseudo-label to supervise the training of the next round of iterations; the update rule of the pseudo-label is: in, represents the updated prediction after the t-1th iteration, represents the prediction of the tth round used as pseudo label, Represents the boundary enhancement prediction feature F corresponding to the t-th round of iterative prediction edge , Conv represents the convolution operation, Softmax is the Softmax function, represents the updated prediction after the tth iteration; After several iterations of pseudo-supervision iteration, the final boundary refinement prediction P is obtained. refine .

6. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to any one of claims 1 to 5, characterized in that: The inter-domain collaborative reasoning network also includes a post-processing step to implement cross-domain consistency constraints, using fully connected conditional random fields to refine the boundary prediction P refine After further optimization, the energy function of the fully connected conditional random field is: Among them, θ i (x i )=-logP refine (x i ), is a unary potential function, θ ij (x i , x j ) is a bilateral potential function based on the color and spatial distance between image tables, E(x) represents the output of the energy function, and the inference process of the fully connected conditional random field is completed through the mean field approximation algorithm to output the final segmentation result.

7. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 1, characterized in that: The feature decoupling module preprocesses medical image data from multiple different centers and obtains features F through an encoder. It then introduces a dynamic soft mask strategy to achieve fine feature separation and generates a feature response map M through a learnable convolutional layer: M=σ(W m *F+b m ) Where σ is the sigmoid activation function, ensuring that the mask value range is between [0, 1], W m and b m They represent the weights and bias terms of the learnable convolutional layer, respectively, and the response map M is the dynamic soft mask. The feature map is processed using a dual-channel adaptive feature routing mechanism to decompose the input features into domain-independent features and domain-specific features. The expression of the dual-channel adaptive feature routing mechanism is as follows: F inv =F⊙(1-M)+ResBlock inv (F⊙(1-M)) F spec =F⊙M+ResBlock spec (F⊙M) Among them, ⊙ represents element-by-element multiplication, ResBlock inv and ResBlock spec Two dedicated residual blocks adopt multi-scale receptive field structures to enhance the expression capabilities of invariant features and specific features respectively; inv represents domain-independent features, F spec Represents domain-specific features.

8. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 7, characterized in that: The feature decoupling module also includes a bidirectional attention enhancement mechanism based on mutual information minimization, which first uses the domain-independent feature F inv and domain-specific features F spec Calculate the feature channel correlation matrix A, and then feed the correlation matrix A back to each feature branch. The expression is as follows: F inv =F inv +g inv ·ChannelAttention(F inv ,A) F spec =F spec +g spec ·SpatialAttention(F spec ,A) Among them, γ inv and γ spec It is a learnable balance parameter used to dynamically adjust the intensity of attention influence. ChannelAttention represents channel attention, and SpatialAttention represents spatial attention.

9. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 8, characterized in that: The feature decoupling module introduces a gradient modulation mechanism based on mutual information minimization, which is expressed as follows: Among them, I(F inv , F spec ) represents the mutual information between domain-independent features and domain-specific features; is the mutual information gradient of domain-independent features, is the mutual information gradient of domain-specific features; sign(·) is the sign function, which means the modulation strength of mutual information in the gradient direction is retained, λ MI represents the strength of the control mutual information constraint; |·| 1+∈ Indicates adaptive adjustment of the gradient through the exponent 1+∈; and denote the initial gradient and modulated gradient of the domain-independent feature, and represent the initial gradient and modulated gradient of domain-specific features, respectively.

10. The cross-domain medical image segmentation method based on feature decoupling and enhancement according to claim 9, characterized in that: Mutual Information I(F inv , F spec ) is estimated using a neural estimator based on noise contrast estimation, which is expressed as: Among them, T θ It is a parameterized neural network with feature pairs as input and scalar scores as output; the joint distribution expectation Represents the feature pairs extracted from the same sample (F inv and F spec )Find the expected marginal distribution expectation; Represents the feature pairs extracted from different samples (F inv and F spec ) Find the expectation maximization operation; represents the expression that is maximized by optimizing the network parameters θ.

Images