Medical image field generalization segmentation method and system

Abstract

The invention provides a generalization segmentation method and system in the medical image field, and belongs to the technical field of medical image processing. According to the method, after multi-center medical image data is preprocessed, a network architecture based on U-Net and ResNet is constructed for feature extraction and segmentation prediction. The image features are converted into frequency domain representation through a frequency domain decomposition module, and low-frequency amplitude and phase information are separated. And constructing a style representation space by using a frequency domain style projection reconstruction module, projecting an unknown domain style to a source domain style library, and disturbing and expanding feature distribution. Through a multi-stage comparison decoupling module, in combination with class prototype comparison learning and pixel-level semantic comparison learning, decoupling of domain-independent and domain-related features is realized. And finally, combining the segmentation loss with the contrast learning loss optimization model to generate a final segmentation result. By the adoption of the method, the problem of medical image multi-center data field deviation is solved through frequency domain style alignment and comparison decoupling.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a generalized segmentation method and system in the field of medical images. Background Art

[0002] In recent years, deep learning methods have made significant progress in automatic medical image segmentation. However, most existing medical image segmentation methods are based on the assumption that training samples and test samples follow the same statistical distribution. In actual medical scenarios, due to differences in imaging equipment, imaging protocols, etc., there are distribution differences between data from different centers, that is, domain shift. These multi-center data constitute a multi-center problem in the field of medical imaging. In this context, model generalization is crucial to solving the problem of unstable performance when the model is deployed to different centers, which is of great significance for clinical auxiliary diagnosis.

[0003] Medical image segmentation methods currently used for multi-source domain generalization can be divided into three categories from the perspective of approach: data-level, model-level, and feature-level generalization. Data-level generalization methods enhance the model's generalization ability by increasing the diversity and quantity of training data through explicit or learnable transformation functions, but their enhancement range is limited and may produce unrealistic images, resulting in limited generalization performance. Model-level generalization methods use meta-learning strategies or transfer learning to improve the pre-trained model framework to alleviate the domain transfer problem, but there are risks of complex optimization and overfitting meta-tasks. Feature-level generalization methods aim to improve the model's generalization performance by using domain-independent features of the input image, and are divided into feature alignment and feature decoupling. Feature alignment enhances generalization by aligning or standardizing cross-domain feature distributions, but it is easy to destroy the discriminability and semantic information of domain-related features; feature decoupling decomposes the input sample feature vector to separate domain-specific and task-specific features. Existing methods have problems such as complex decoupling process and most of them do not consider pixel-level semantic features.

[0004] In summary, existing medical image segmentation methods have many limitations when dealing with the domain shift problem of multi-center data, and a new method that can effectively solve these defects and deficiencies is urgently needed. Summary of the invention

[0005] The purpose of the present invention is to provide a generalized segmentation method and system in the field of medical images, which solves the problem of multi-center data domain offset in medical images through frequency domain style alignment and contrast decoupling, and improves the model generalization ability and segmentation accuracy.

[0006] To achieve the above object, the present invention provides a generalized segmentation method in the field of medical images, comprising the following steps: Step S1, preprocessing multi-center medical image data; Step S2: Construct an encoder-decoder network architecture based on U-Net and ResNet for feature extraction and segmentation prediction of medical images; Step S3: Convert the input image features into a frequency-domain representation through a frequency-domain decomposition module, extract the low-frequency amplitude as style information, and retain the phase as content information; Step S4: Construct a multi-source domain style representation space through a frequency-domain style projection and reconstruction module, project the low-frequency amplitude style of the unknown domain onto the source domain style library, and perform statistical perturbation to expand the feature distribution space; Step S5: Through a multi-level contrast decoupling module, combine class prototype contrast learning and pixel-level semantic contrast learning to achieve decoupling of domain-independent features and domain-related features; Step S6: Optimize the model by combining the segmentation loss and the contrast learning loss to generate the final segmentation result.

[0007] Preferably, in step S1, the preprocessing is as follows: uniformly crop the multi-center medical images to a unified size and scale the values to 0-1, and perform random rotation and random flipping operations on the data to obtain enhanced medical image data.

[0008] Preferably, in step S2, the encoder adopts a four-stage residual structure of ResNet-50, and the decoder adopts a transposed convolution architecture, with each stage containing a transposed convolution upsampling layer and a skip connection fusion unit.

[0009] Preferably, in step S3, the frequency-domain decomposition module specifically includes: Decompose the image features into amplitude and phase components using the Fourier transform; Extract the low-frequency amplitude as global style information and retain the phase as domain-independent global edge texture information.

[0010] Preferably, in step S4, the frequency-domain style projection and reconstruction module specifically includes: Initialize the style bases of each source domain, including mean and variance parameters; Calculate the similarity between the input feature style and the style bases based on the Wasserstein distance to generate a normalized similarity coefficient; Generate style affine parameters through weighted style bases to perform instance normalization and reconstruction on the low-frequency amplitude features; Introduce noise perturbation based on variance estimation into the statistics of the reconstructed style to simulate the style distribution of the unknown domain.

[0011] Preferably, in step S5, the multi-level contrast decoupling module specifically includes: Through class prototype contrast learning, calculate the class prototypes before and after style reconstruction for each domain and use the moving average strategy to update the historical class prototypes to promote feature separation between different domain classes; Through pixel-level semantic contrast learning, anchor points, positive samples, and negative samples are constructed for easily confused pixel features to optimize semantic consistency in the feature space.

[0012] Preferably, class prototype contrast learning is achieved through the following formula: ; where respectively represent the height and width of the feature map, represents the pixel feature 's one-hot label, that is, when belongs to class, that is , represents the index of the current pixel feature; The class prototype is updated using a moving average strategy as follows: ; where respectively represent the update of the class and the historical class prototype, represents the momentum, set to 0.999.

[0013] Preferably, in step S6, the segmentation loss function is: ; where represents the segmentation loss, represents the binary cross-entropy loss, represents the Dice loss; ; where represents the total number of pixel points, represents the true class of the th pixel, represents the th pixel's predicted probability value; .

[0014] Preferably, the contrast learning loss function is: ; where represents the contrast loss, respectively represent the embedding vectors of the anchor point, positive sample, and negative sample, represents the temperature hyperparameter, used to adjust the smoothness of the distribution, represents the negative sample set.

[0015] The present invention also provides a generalization segmentation system in the field of medical images, including: A data preprocessing module for preprocessing multi-center medical image data; A network construction module for constructing an encoder-decoder network architecture based on U-Net and ResNet, where the encoder adopts a four-level residual structure of ResNet-50, and the decoder adopts a transposed convolution architecture, with each level containing a transposed convolution upsampling layer and a skip connection fusion unit; A frequency domain decomposition module for converting the input image features into a frequency domain representation, separating the low-frequency amplitude as style information and the phase as content information; A frequency domain style projection and reconstruction module for constructing a multi-source domain style representation space, projecting the low-frequency amplitude style of the unknown domain onto the source domain style library, and performing statistical perturbation to expand the feature distribution space; A multi-level contrast decoupling module for combining class prototype contrast learning and pixel-level semantic contrast learning to achieve decoupling of domain-independent features and domain-related features; A model optimization module for jointly optimizing the model with segmentation loss and contrast learning loss to generate the final segmentation result.

[0016] Therefore, the present invention adopts the above-mentioned generalization segmentation method and system in the field of medical images, and the beneficial technical effects are as follows: (1) Through the visualization analysis of the low-frequency amplitude style transfer and its amplitude and phase of paired images, it is found that the low-frequency amplitude of the image contains the main style-related information, and the phase contains more content information, and the two can be well separated. Therefore, the low-frequency amplitude is regarded as style information and introduced into the feature level. Then, the target domain style is projected onto the source domain statistical style library for style alignment, while retaining the discriminability of domain-related features when aligning features; in addition, statistical transformation perturbation is considered for the low-frequency amplitude to simulate the feature distributions of more low-frequency styles, better adapting to the unknown target domain, and at the same time helping the model to emphasize and retain the high-frequency components of global detail information (2) Combining the use of class-level semantic prototypes and pixel-level semantic features for contrast learning, the class prototype feature contrast can promote the model to globally separate and identify different categories of source images and style reconstruction images, achieving preliminary class separation and domain decoupling. The pixel-level semantic features further refine the recognition ability of pixel features, decouple different domain features and easily confused features, and promote the semantic consistency between different domain features, further improving the prediction performance and generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of a generalization segmentation method for medical images according to the present invention; Figure 2 It is a frequency domain style projection reconstruction and low-frequency perturbation module; Figure 3 It is a multi-level contrast decoupling module; Figure 4 For visual comparison of fundus data segmentation; Figure 5 For visual comparison of prostate MRI segmentation. Detailed implementation manners

[0018] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs.

[0020] Embodiment 1 As Figures 1 - 3 shown, it is a flowchart of a generalization segmentation method in the field of medical images according to the present invention, which specifically includes the following steps: Step S1: Preprocess the multi-center medical image data.

[0021] Step S11: In this embodiment, the multi-center medical image data includes 3D prostate MRI data and 2D retinal fundus data sets. The 3D prostate MRI data is sliced, and the 2D slice results are saved in the pkl format; Step S12: When loading the data, randomly perform data augmentation processing on the data and the corresponding labels, including operations such as random rotation and random flipping, and set that each source domain data is randomly included in a batch.

[0022] Step S2: Construct an encoder-decoder network architecture based on U-Net and ResNet for feature extraction and segmentation prediction of medical images.

[0023] The encoder is composed of four convolutional stages of the ResNet-50 network. Each stage contains multiple residual units with 3×3 convolutional kernels, and its mathematical expression is: ; Among them, represents the network-generated feature, represents the input data, represents the convolution operation with output channels, represents the LeakyReLU activation function, represents the batch normalization layer.

[0024] The feature dimension reduction mechanism is implemented through cascaded max pooling layers. A 2×2 max pooling layer with a stride of 2 is set at the end of each encoding unit, and its dimension reduction formula is: ; Among them, represents the dimensionality-reduced feature, represents the max pooling operation, represents the original feature, is a 1×1 convolution for channel adjustment; The decoder adopts a transposed convolution architecture, and each level contains a transposed convolution upsampling layer and a skip connection fusion unit. Its feature reconstruction process is expressed as: ; where, represents the feature map of the th level of the decoder, represents the convolution operation, represents the feature map of the th level of the decoder, represents bilinear interpolation upsampling, represents the channel concatenation operation, represents the th level of the corresponding encoding stage of the feature map.

[0025] Step S3: Convert the input image features into a frequency domain representation through a frequency domain decomposition module, and separate the low-frequency amplitude as the style information and the phase as the content information.

[0026] Step S31: Use the Fourier transform to decompose the input feature representation into two parts, amplitude and phase; The low-frequency part of the feature amplitude mainly contains global style information, while the phase information contains domain-independent global structural information; Step S32: Use the low-frequency amplitude statistical information of the target domain image to perform style transfer and alignment on the source domain image. It can achieve alignment of different domain styles while retaining the global structural information. The specific formula is as follows: ; ; where, represents the representation of the original two-dimensional signal in the spatial domain, respectively represent the real and imaginary parts of the Fourier transform, and respectively represent the amplitude and phase, represents the complex value corresponding to the frequency in the frequency domain, represents the width of the image, represents the height of the image, represents the pixel index in the horizontal direction, represents the pixel index in the vertical direction, is the imaginary unit, represents the index of the horizontal frequency component, Indicates the index of the vertical frequency component.

[0027] After moving the low-frequency component of the amplitude to the center of the spectrum, a binary mask is introduced , whose value is zero except in the central region: ; Among them, Indicates the ratio of control Size, used to distinguish high-frequency and low-frequency components, , Respectively represent the height and width of the frequency.

[0028] Low-frequency amplitude And high-frequency amplitude : ; ; Among them, Indicates element-wise multiplication.

[0029] Step S4, construct a multi-source domain style representation space through the frequency-domain style projection reconstruction module, project the low-frequency amplitude style of the unknown domain to the source domain style library, and perform statistical perturbation to expand the feature distribution space.

[0030] Step S41, initialization of the style basis.

[0031] This module randomly initializes a corresponding set of style bases for each processing domain . Among them Represents the mean parameter of the style basis of the th domain, Represents the standard deviation parameter of the style basis of the th domain.

[0032] During the initialization process, Is initialized with zero mean, Is initialized with unit standard deviation.

[0033] Step S42, low-frequency amplitude style similarity measurement.

[0034] For the low-frequency amplitude style statistics of the input features, estimate the sample distance for the style basis of each source domain, and use it as the similarity coefficient for weighting. Specifically, calculate the mean And standard deviation Of the low-frequency amplitude of the input features along the channel dimension, and then use the Wasserstein distance to measure the style distribution difference between the low-frequency amplitude style of the input features and the th style basis, as shown in the following formula: ; Among them, and respectively represent the mean and variance obtained by calculating the characteristics along the channel dimension. represents the distance from the th style basis. represents the mean parameter of the style basis of the th domain. represents the Euclidean distance squared operation. represents the th domain's style basis standard deviation parameter; Step S43: To characterize the similarity between the input feature style and the style basis, normalize the reciprocals of all style bases, and use the Softmax operation to make the sum of each item equal to 1 as the similarity index. The process is as follows: ; Among them, represents the style similarity of the th domain, represents the total number of source domains.

[0035] Step S44: Generation of style affine parameters.

[0036] Based on the normalized similarity coefficients, weight each style basis to obtain the final style affine parameter , and the calculation formula is: ; Among them, represents the mean of the reconstructed style representation, represents the standard deviation value of the reconstructed style representation; Step S45: Low-frequency amplitude perturbation.

[0037] Assume that the distribution of each statistic of the reconstructed style follows a Gaussian distribution, and the standard deviation of these statistics is calculated as follows: ; ; Among them, represents the variance of the mean within the current batch, represents the variance of the standard deviation within the current batch, represents the batch size, represents the current batch, represents calculating the mean A Gaussian distribution of low-frequency spectrum probability statistics is established, and new means and standard deviations are randomly sampled: ; ; Among them, represents the mean after perturbation, represents the standard deviation after perturbation, represents the random noise sampled from the normal distribution added to the mean, represents the random noise sampled from the normal distribution added to the standard deviation, represents the intensity of the perturbation; Step S46, Style Projection and Alignment Reconstruction.

[0038] Perform instance normalization on the low-frequency amplitude features and convert them to the standard normal distribution to achieve the effect of removing the original style information. Denote the normalized features as , and the normalization formula is: ; Among them, represents the original amplitude feature, represents a small positive number to avoid division by zero; Subsequently, perform style projection reconstruction on through the style affine parameters after perturbation to obtain the amplitude features aligned with the style, that is: ; Step S47, Momentum Update of Style Information.

[0039] In the training stage, in order to maintain the accuracy and stability of the style information of each domain, this module adopts a momentum update strategy to update the style bases of each domain. Let the momentum coefficient be , and the update formula is as follows: ; ; Among them, is the momentum coefficient, which is set to 0.9 here. This ensures that the style bases can smoothly track the changes in the style information of different domains during the network training process, and further improves the adaptability of the system to the generalized style.

[0040] Step S5, Through the multi-level contrast decoupling module, combine class prototype contrast learning and pixel-level semantic contrast learning to achieve the decoupling of domain-independent features and domain-related features.

[0041] Step S51, Class Prototype Contrast Learning. It is intended to perform contrast learning on the class prototype features of different domains to promote the separation of the feature spaces of different domains and different categories, so as to enhance the recognition of domain-independent features. First, map the features to a unified high-dimensional representation space through a non-linear projection head P.

[0042] The construction of class prototypes calculates the prototypes by averaging the features in each class region: ; where respectively represent the height and width of the feature map. is the pixel feature 's one - hot label, that is, when belongs to class, that is .

[0043] To obtain the class prototypes in the entire training dataset, use the moving average strategy to update the class prototypes: ; where is the updated and historical class prototype of is the th class prototype calculated in the current training batch, is the momentum set to 0.999.

[0044] Suppose the feature corresponding to the predicted class prototypes of the original image for classes 1, 2, and 3 are , , ; the predicted class prototype of the style - reconstructed image for class 1 is the feature . Then: positive pair: ; negative pairs: , , .

[0045] Step S52, Confusing pixel contrast learning. Using the traditional cross - entropy as the task loss only independently penalizes pixel - level predictions and ignores the semantic relationships between pixels. To further refine the semantic classification of features and reduce the impact of the domain, contrast learning of features of confusing classes in different domains is introduced to promote the separation of features that are difficult to classify due to style influence. The above - mentioned style - reconstructed samples are called domain - enhanced samples. First, anchor the sampling features that are confusing with the domain - enhanced samples, and the corresponding predicted label class is , while the corresponding true - label class is ( ), the positive sample is the feature sampled from the original image, and its predicted class label is ( ), and the negative sample is the feature sampled from the enhanced image, and its predicted class label is ( ).

[0046] Suppose the correct predicted class of the original image The corresponding feature is ; the corresponding position of the domain-enhanced image is incorrectly predicted as a category The corresponding feature is , and the others are correctly predicted as categories The feature is . Then: Positive pair: , ; Negative pair: , .

[0047] Step S6, optimize the model by combining the segmentation loss and the contrastive learning loss to generate the final segmentation result.

[0048] To simultaneously achieve medical image segmentation and cross-domain feature alignment, the overall loss function is decomposed into two parts: the segmentation loss and the contrastive learning loss. Among them, the segmentation loss is used to guide the segmentation network to accurately segment the target area at the pixel level, while the contrastive learning loss is used to further align different domain features, so that the extracted features not only have excellent segmentation performance but also have cross-domain generalization ability. The expression of the overall training objective function is as follows: ; Among them, represents the total loss, represents the segmentation loss, represents the contrastive loss, represents the hyperparameter that controls the weight; The segmentation loss function is: ; Among them, represents the binary cross-entropy loss, represents the Dice loss; ; Among them, represents the total number of pixel points, represents the true category of the th pixel, represents the predicted probability value of the th pixel; .

[0049] The contrastive learning loss function is: ; Among them, represent the embedding vectors of the anchor point, positive sample, and negative sample respectively, represents the temperature hyperparameter, which is used to adjust the smoothness of the distribution, represents the negative sample set.

[0050] Segmentation performance metric calculation module: The metrics are Dice coefficient, Hausdorff distance HD, and average surface distance ASD respectively.

[0051] The following is a further illustration of the present invention through specific examples.

[0052] Training and testing on the retinal fundus dataset and the prostate MRI dataset fully demonstrate the superiority of the method proposed in the present invention. The specific dataset characteristics and processing procedures are as follows.

[0053] The multi - center retinal fundus image dataset is used to evaluate the segmentation performance of the optic cup (OC) and optic disc (OD) to measure the model generalization ability. This dataset integrates four public fundus image datasets, covering images obtained from different locations and different scanners. Center 1 contains 50 training images and 51 test images; Center 2 contains 99 training images and 60 test images; Center 3 and Center 4 each contain 320 training images and 80 test images. During pre - processing, all images are unified into 2D three - channel color images with a size of 256×256 and normalized to ensure data consistency and stability.

[0054] The prostate MRI dataset is a publicly available multi - center dataset for prostate segmentation tasks in T2 - weighted MRI. The data comes from lower abdominal MRI images of 6 different institutions, and the specific distribution is as follows: Center 1 and Center 2 are from the NCI - ISBI13 dataset, each containing 30 patient images; Center 3 is from the I2CVB dataset, containing 19 patient images; Center 4, Center 5, and Center 6 are from the PROMISE12 dataset, containing 13, 12, and 12 patient images respectively. The pre - processing steps are: first, adjust the images to a size of 384×384 in the axial plane, then normalize the pixel intensity to conform to a distribution with a mean of 0 and a standard deviation of 1, and finally crop the images to retain the slices containing the prostate region, removing irrelevant information and focusing on the key region.

[0055] To verify the effectiveness of the method of the present invention, comparative experiments are conducted with various effective methods in recent years. The experiments mainly evaluate the model performance based on two segmentation metrics, Dice and ASD. In addition, for 3D prostate data, the predicted results of all 2D slices are restored to 3D predictions and then compared with mainstream methods. The experimental results are listed in detail in Tables 1, 2, 3, and 4, demonstrating the strong competitiveness and excellent generalization performance of the method of the present invention in the field of medical image segmentation.

[0056] Table 1 Dice metrics (%) for multi - center data segmentation of prostate MRI ;

[0057] Table 2 ASD metrics (mm) for multi - center data segmentation of prostate MRI ;

[0058] Table 3 Dice Index (%) of Fundus Multicenter Data Segmentation ;

[0059] Table 4 ASD Index (mm) of Fundus Multicenter Data Segmentation ;

[0060] The experimental results were qualitatively compared with other methods, and the visualization results of the specific segmentation effects are shown in Figure 4 、 Figure 5 . The segmentation effects of these methods on the first three domains of these two datasets were compared respectively. For the retinal fundus data segmentation task, the segmentation results of the optic cup and optic disc obtained by the present invention are closer to the ground truth labels. Compared with the baseline and other domain generalization methods, there are more over-segmented and under-segmented regions, especially the blurred segmentation at the junction of the optic cup and optic disc. For the prostate MRI segmentation task, the segmentation of the target region by the present invention is more refined and smooth, while other methods are prone to positioning errors in some low-brightness images compared with the comparison methods.

[0061] Example 2 A domain generalization segmentation system in the field of medical images, comprising: A data preprocessing module for preprocessing multi-center medical image data; A network construction module for constructing an encoder-decoder network architecture based on U-Net and ResNet, wherein the encoder adopts a four-stage residual structure of ResNet-50, the decoder adopts a transposed convolution architecture, and each stage contains a transposed convolution upsampling layer and a skip connection fusion unit; A frequency domain decomposition module for converting the input image features into a frequency domain representation, separating the low-frequency amplitude as the style information and the phase as the content information; A frequency domain style projection reconstruction module for constructing a multi-source domain style representation space, projecting the low-frequency amplitude style of the unknown domain to the source domain style library, and performing statistical perturbation to expand the feature distribution space; A multi-level contrast decoupling module for combining class prototype contrast learning and pixel-level semantic contrast learning to achieve decoupling of domain-independent features and domain-related features; A model optimization module for jointly optimizing the model with the segmentation loss and the contrast learning loss to generate the final segmentation result.

[0062] It should be noted that the content not elaborated in detail in the present invention is all prior art and is well known to those skilled in the art.

[0063] Therefore, the present invention adopts the above-mentioned method and system for generalized segmentation in the field of medical images, and solves the problem of domain shift in multi-center medical image data through frequency-domain style alignment and contrast decoupling, improving the model generalization ability and segmentation accuracy.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A generalization segmentation method in the field of medical images, characterized in that, It includes the following steps: Step S1: Preprocess the multi-center medical image data; Step S2: Construct an encoder-decoder network architecture based on U-Net and ResNet for feature extraction and segmentation prediction of medical images; Step S3: Convert the input image features into frequency-domain representations through a frequency-domain decomposition module, extract the low-frequency amplitude as style information, and retain the phase as content information; Step S4: Construct a multi-source domain style representation space through a frequency-domain style projection and reconstruction module, project the low-frequency amplitude style of the unknown domain onto the source domain style library, and perform statistical perturbation to expand the feature distribution space; Step S5: Through a multi-level contrast decoupling module, combine class prototype contrast learning and pixel-level semantic contrast learning to achieve decoupling of domain-independent features and domain-related features; Step S6: Optimize the model by combining the segmentation loss and the contrast learning loss to generate the final segmentation result; In Step S4, the frequency-domain style projection and reconstruction module specifically includes: Initialize the style bases of each source domain, including mean and variance parameters; Calculate the similarity between the input feature style and the style bases based on the Wasserstein distance to generate a normalized similarity coefficient; Generate style affine parameters through weighted style bases to perform instance normalization and reconstruction on the low-frequency amplitude features; Introduce noise perturbation based on variance estimation into the statistics of the reconstructed style to simulate the style distribution of the unknown domain.

2. The generalization segmentation method in the field of medical images according to claim 1, wherein In Step S1, the preprocessing is as follows: Uniformly crop the multi-center medical images and scale the values to 0-1, and perform random rotation and random flipping operations on the data to obtain enhanced medical image data.

3. A generalization segmentation method in the field of medical images according to claim 1, characterized in that, In Step S2, the encoder adopts a four-level residual structure of ResNet-50, and the decoder adopts a transposed convolution architecture, with each level containing a transposed convolution upsampling layer and a skip connection fusion unit.

4. A generalization segmentation method in the field of medical images according to claim 1, characterized in that, In Step S3, the frequency-domain decomposition module specifically includes: Decompose the image features into amplitude and phase components using the Fourier transform; Extract the low-frequency amplitude as global style information and retain the phase as domain-independent global edge texture information.

5. A generalization segmentation method in the field of medical images according to claim 1, characterized in that, In Step S5, the multi-level contrast decoupling module specifically includes: Through class prototype contrast learning, calculate the class prototypes before and after style reconstruction for each domain and use the moving average strategy to update the historical class prototypes to promote feature separation between different domain classes; Through pixel-level semantic contrast learning, construct anchors, positive samples, and negative samples for easily confused pixel features to optimize the semantic consistency in the feature space.

6. A generalization segmentation method in the field of medical images according to claim 5, characterized in that, The class prototype contrast learning is implemented through the following formula: ； Among them, respectively represent the height and width of the feature map, represents the pixel feature of the one-hot label, that is, when belongs to class, that is , represents the index of the current pixel feature; Update the class prototypes using the moving average strategy as follows: ； Among them, respectively represent the update of the class and the historical class prototype, represents momentum and is set to 0.

999.

7. A generalization segmentation method in the field of medical images according to claim 1, characterized in that In Step S6, the segmentation loss function is: ； Among them, represents the segmentation loss, represents the binary cross-entropy loss, represents the Dice loss; ； Among them, represents the total number of pixel points, represents the true category of the th pixel, represents the predicted probability value of the th pixel; 。 8. A generalization segmentation method in the field of medical images according to claim 1, characterized in that The contrast learning loss function is: ； Among them, represents the contrastive loss, respectively represent the embedding vectors of the anchor, positive sample, and negative sample, represents the temperature hyperparameter, which is used to adjust the smoothness of the distribution, represents the negative sample set.

9. A generalization segmentation system in the field of medical images, characterized in that, It includes: A data preprocessing module for preprocessing multi-center medical image data; A network construction module for constructing an encoder-decoder network architecture based on U-Net and ResNet, where the encoder adopts a four-level residual structure of ResNet-50, and the decoder adopts a transposed convolution architecture, with each level containing a transposed convolution upsampling layer and a skip connection fusion unit; A frequency-domain decomposition module for converting the input image features into frequency-domain representations, separating the low-frequency amplitude as style information and the phase as content information; Frequency-domain style projection reconstruction module, which is used to construct a multi-source domain style representation space, project the low-frequency amplitude style of the unknown domain onto the source domain style library, and perform statistical perturbation to expand the feature distribution space; Multi-level contrast decoupling module, which is used to combine class prototype contrast learning and pixel-level semantic contrast learning to achieve the decoupling of domain-independent features and domain-related features; Model optimization module, which is used to jointly optimize the model with segmentation loss and contrast learning loss to generate the final segmentation result.

Images