Cross-modal medical image segmentation method and system based on improved domain adaptation framework of Mamba network

Abstract

The present invention discloses a cross-modal medical image segmentation method and system based on a Mamba network improved domain adaptation framework, which relates to the field of image processing, wherein the method comprises: S1. Preprocessing a pre-acquired cross-modal medical image dataset, and forming a source domain dataset and a target domain dataset; S2. Building a domain adaptation framework improved by using a Mamba network, the improved domain adaptation framework comprises: a generation network and a segmentation network; the generation network comprises a generator and a first discriminator, which is used to generate a translation image based on a labeled source domain dataset; the segmentation network comprises a segmentor and a second discriminator; S3. End-to-end training of the improved domain adaptation framework obtains a segmentation network; S4. Using the segmentation network to predict the image segmentation result of the target domain dataset. The present invention can better adapt the source domain and the target domain, realize domain adaptation, and improve the test accuracy on unlabeled target domain images.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a cross-modal medical image segmentation method and system based on a Mamba network improved domain adaptation framework. Background Art

[0002] Neural networks have achieved remarkable success in medical image segmentation tasks, especially when a large amount of labeled data is available. However, well-trained models often encounter difficulties when deployed in clinical settings. This performance degradation is because data in real-world applications often come from different acquisition parameters or imaging techniques, resulting in significant differences in appearance and depth distribution. While these cross-domain variations are easily discernible by the human eye, they can be challenging for neural networks. For example, when a segmentation network trained on magnetic resonance imaging (MRI) images is applied to computed tomography (CT) images, evaluation metrics such as Dice similarity coefficient (Dice), average surface distance (ASD), etc. often drop significantly.

[0003] In order to alleviate the performance degradation caused by cross-domain differences, cross-modal domain adaptation research has emerged. The goal of domain adaptation is to leverage the knowledge of the source domain and effectively transfer it to the target domain. One of the most direct and effective methods is to fine-tune the model trained on the source domain data using labeled data from the target domain. However, in the medical field, obtaining labels can be challenging due to privacy issues or other restrictions. Therefore, unsupervised domain adaptation (UDA) methods that do not require target domain labels are becoming increasingly popular. The core idea of ​​unsupervised domain adaptation is to build a shared model between the source and target domains, and use the information of the two domains to train a model that performs well on the target domain.

[0004] Some studies have attempted to alleviate the distribution differences between the source and target domains through feature alignment to enhance the ability of modal transfer. However, due to the inherent differences in their distribution, it is impossible to completely eliminate domain differences. At the same time, due to the excellent performance of generative adversarial networks (GANs) in the field of image translation, other studies have used GAN networks and adversarial learning to achieve modal conversion from the source domain to the target domain, thereby trying to solve the problem of structural consistency. However, GAN-based methods are often affected by mode collapse and unstable training, resulting in the inability of translated images to maintain structural consistency. Subsequently, some methods have integrated feature alignment and image translation methods into a unified framework to improve adaptability and segmentation accuracy.

[0005] The above methods are mainly based on traditional convolutional neural networks (CNNs). However, convolutional neural networks have limitations in modeling long-range dependencies, which is particularly problematic in unsupervised domain adaptation tasks that require considering global style information for each image sample. Although Transformers can capture global context information, they are hampered by the trade-off between quadratic computational complexity and pixel-level local modeling. Recently, state-space models (SSMs) represented by the Mamba network model have become a promising approach. The Mamba network model is not only good at utilizing global context information for modeling, but also maintains linear computational complexity. Currently, the Mamba network model has made initial progress in the field of medical image segmentation. The hybrid method of the Mamba network and CNN has surpassed the CNN or Transformer methods on multiple segmentation tasks. However, there is currently no application of the Mamba network in cross-domain unsupervised segmentation scenarios. Summary of the invention

[0006] To solve the technical problems in the above background, the present invention aims to effectively utilize labeled source domain data and unlabeled target domain data, and improve the segmentation performance of the network during target domain testing by reducing the distribution difference between the source domain and the target domain.

[0007] To achieve the above object, the present invention provides a cross-modal medical image segmentation method based on a Mamba network improved domain adaptation framework, the steps comprising:

[0008] S1. Preprocess the pre-acquired cross-modal medical image dataset and form a source domain dataset and a target domain dataset;

[0009] S2. Building a domain adaptation framework improved by using the Mamba network, the improved domain adaptation framework comprising: a generation network and a segmentation network; the generation network comprising a generator and a first discriminator, for generating a translation image based on the labeled source domain dataset; the segmentation network comprising a segmentor and a second discriminator;

[0010] S3. training the improved domain adaptation framework end-to-end to obtain a segmentation network;

[0011] S4. Use the segmentation network to predict the image segmentation result of the target domain dataset.

[0012] Preferably, in the generation network, the generator comprises: a first encoder, a bottleneck layer and a first decoder;

[0013] The first encoder consists of input processing and two downsampling layers to extract high-level feature representations;

[0014] The bottleneck layer is a Mamba hybrid bottleneck module and a residual block alternatingly;

[0015] The first decoder consists of two upsampling layers and output processing.

[0016] Preferably, in the first encoder, the input processing includes a reflection padding layer, a convolution layer, an instance normalization layer and a ReLU activation function; the two downsampling layers have the same structure, both consisting of a convolution layer, an instance normalization layer and a ReLU activation function;

[0017] In the first decoder, the output processing includes a reflection padding layer, a convolution layer and a Tanh activation function; the two upsampling layers have the same structure, both consisting of a transposed convolution layer, an instance normalization layer and a ReLU activation function;

[0018] In the bottleneck layer, the features are first linearly processed in the Mamba hybrid bottleneck module and then divided into two branches; in the first branch, the features are sequentially subjected to a depth-separable convolutional layer, an activation layer, and a state equation modeling layer; in the second branch, the features are processed by an activation function SiLU; finally, the features from the two branches are combined using the Hadamard product, and then the obtained features are combined with the input features through a residual connection; the Mamba hybrid bottleneck module also includes a channel compression module, which is used to further reduce the number of channels and maintain key information in the features.

[0019] Preferably, the first discriminator adopts an adversarial learning approach, using the source domain label and the source domain dataset as joint inputs to serve as additional structural constraints to assist the generation process.

[0020] Preferably, in the segmentation network, the segmenter is composed of a second encoder, a style adjustment and semantic injection jump connection module and a second decoder; before the second encoder, the image is divided into non-overlapping blocks of size 4×4 by patch embedding, and the dimension is mapped to C; the second encoder includes four stages, each stage using N1, N2, N3, N4 visual state equation blocks.

[0021] Preferably, a domain-specific layer normalization module is used to replace the initial normalization module in the visual state equation block. The domain-specific layer normalization module contains two independent layer normalizations. Each layer normalization processes features from a specific domain to estimate inter-layer statistics and uses learnable affine parameters to capture specific information of each domain. The domain-specific layer normalization DSLN is defined as:

[0022] ,

[0023] Where d is the domain label, z is the feature from domain d, and represents the mean and variance, and is a learnable affine parameter and ε is a very small constant used to avoid instability when the denominator is zero.

[0024] Preferably, the style adjustment and semantic injection module is used to process the multi-scale feature map from the second encoder, firstly, the spatial and channel attention is calculated through the convolutional attention module; then, 1×1 convolution maps all features to the same dimension and performs feature fusion; finally, the features are downsampled or upsampled to each scale through adaptive average pooling or bilinear interpolation and passed to the second decoder part; each level of the second decoder accepts the fused features from the style adjustment and semantic injection module, uses three sets of transposed convolutions to upsample the underlying feature information, and fuses the underlying features with the previous level high-level features, and finally completes the resolution reconstruction and segmentation tasks.

[0025] The present invention also provides a cross-modal medical image segmentation system based on a Mamba network improved domain adaptation framework, the system is used to implement the above method, including: an acquisition module, a construction module, a training module and a segmentation module;

[0026] The acquisition module is used to preprocess the pre-acquired cross-modal medical image dataset and form a source domain dataset and a target domain dataset;

[0027] The building module is used to build a domain adaptation framework improved by using a Mamba network, wherein the improved domain adaptation framework includes: a generation network and a segmentation network; the generation network includes a generator and a first discriminator, and is used to generate a translation image based on the labeled source domain dataset; the segmentation network includes a segmentor and a second discriminator;

[0028] The training module is used to train the improved domain adaptation framework end-to-end to obtain a segmentation network;

[0029] The segmentation module is used to predict the image segmentation result of the target domain data set using the segmentation network.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The improved conditional generative network of the Mamba network designed by the present invention can better complete the image modality translation task and maintain the structural constraints. The improved Mamba network designed by us using the domain-specific layer normalized segmentation network can better adapt to the source domain and the target domain, achieve domain adaptation, and improve the test accuracy on the unlabeled target domain images. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0033] Figure 1 is a flowchart of an embodiment of the present invention;

[0034] Figure 2 is a schematic diagram of an improved condition generator of a Mamba network according to an embodiment of the present invention;

[0035] Figure 3 It is a schematic diagram of a Mamba hybrid bottleneck module (MMN) structure in a bottleneck layer of a generator improved by a Mamba network according to an embodiment of the present invention;

[0036] Figure 4 is a schematic diagram of an improved segmenter of a Mamba network according to an embodiment of the present invention;

[0037] Figure 5 Schematic diagram of a visual state equation (VSS) block structure in a segmenter improved by a Mamba network according to an embodiment of the present invention;

[0038] Figure 6 It is a schematic diagram of the workflow of the style adjustment and semantic injection module (SASI) in the improved generator of the Mamba network in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] 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.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Embodiment 1

[0042] This embodiment provides a cross-modal medical image segmentation method based on the Mamba network improved domain adaptation framework. The overall method flow is as follows: Figure 1 As shown, the specific implementation methods of each step will be described in detail below in combination with this embodiment.

[0043] S1. Preprocess the pre-acquired cross-modal medical image dataset and construct a source domain dataset and a target domain dataset.

[0044] The target organ area is captured, the 3D image is divided into multiple 2D images, the image is modified to a uniform size, the image is standardized and normalized, and random cropping and random rotation are performed to achieve image augmentation operations to form the source domain dataset and the target domain dataset.

[0045] S2. Build an improved domain adaptation framework using Mamba network.

[0046] In this embodiment, the improved domain adaptation framework includes: a conditional generative network improved by a Mamba network (hereinafter referred to as a generative network) and a segmentation network improved by a Mamba network using domain-specific layer normalization (hereinafter referred to as a segmentation network). The generative network includes a generator and a first discriminator, and the segmentation network includes a segmentor and a second discriminator.

[0047] Specifically, in the generative network, such as Figure 2 As shown, the generator includes: a first encoder, a bottleneck layer and a first decoder.

[0048] The first encoder consists of input processing and two downsampling layers to extract high-level feature representations. The input processing includes a reflection padding layer, a convolution layer, an instance normalization layer, and a ReLU activation function; the two downsampling layers have the same structure, both consisting of a convolution layer, an instance normalization layer, and a ReLU activation function.

[0049] The first decoder consists of two upsampling layers and output processing, where the output processing includes a reflection filling layer, a convolution layer and a Tanh activation function; the two upsampling layers have the same structure, both consisting of a transposed convolution layer, an instance normalization layer and a ReLU activation function.

[0050] In the bottleneck layer, this embodiment designs the Mamba hybrid bottleneck (MMN) module and the residual block alternately to maintain the characteristic characteristics to the greatest extent. In addition, compared with the bottleneck module of the traditional pure CNN architecture, the bottleneck module improved by Mamba can extract more comprehensive and effective context information without introducing additional computational complexity. Figure 3 As shown in the figure, in the MMN module, the features are first linearly processed and then divided into two branches; in the first branch, the features are sequentially processed through a depth-wise separable convolutional layer, an activation layer, and a state equation modeling (SSM) layer; in the second branch, the features are processed by the activation function SiLU; finally, the features from the two branches are combined using the Hadamard product, and then the obtained features are combined with the input features through a residual connection.

[0051] Finally, in the channel compression module, two sets of 3×3 convolutions, batch normalization layers, and ReLU activation functions are residually connected with a set of 1×1 convolutions, batch normalization layers, and ReLU activation functions to further reduce the number of channels and maintain the key information in the features.

[0052] The first discriminator mentioned above uses a patch-based discriminator and generator to complete adversarial learning. At the same time, based on the inspiration of the conditional generative adversarial network (ConditionalGAN), the source domain label and the source domain image are used as common generator input, so that the label can serve as an additional structural constraint to assist the generation process.

[0053] The labeled source domain dataset generates translated images in the target domain style through the generator. The specific process includes:

[0054] Introduce unsupervised contrast loss PatchNCE loss to optimize the generator:

[0055] ,

[0056] Where v is the feature vector of the query block from the translated image; v + is the feature vector of the positive sample from the source domain image and is in the same position as the query block; - The negative sample feature vector comes from the source domain image and is located at different positions of the query block; τ represents the temperature parameter, and in this embodiment, τ = 0.07; n represents the current patch block; N represents the total number of patches; Represents the current negative sample patch; L nce represents the PatchNCE loss.

[0057] Afterwards, the first discriminator is used to distinguish fake data from real data in the target domain and complete adversarial learning on the generator. The learning objectives are:

[0058] ,

[0059] In the formula, E represents the expected value, which represents the average operation of the relevant variables; x represents the image (the corresponding y represents the label) t~T represents the target domain distribution; D g represents the first discriminator; G represents the generator; x t represents the target domain image; x s~S Represents the source domain distribution.

[0060] In the segmentation network, Figure 4As shown in the figure, the segmenter consists of a second encoder, a style adjustment and semantic injection (SASI) module, and a second decoder. Before the second encoder, the image is divided into non-overlapping blocks of size 4×4 by patch embedding, and the dimension is mapped to C, which defaults to 96. The second encoder contains four stages, each of which uses [N1, N2, N3, N4] visual state equation (VSS) blocks. The VSS block originated from Visual Mamba (V Mamba) and is the core module of the encoder. The features in this module are first layer-normalized and then divided into two branches: in the first branch, the input features are once passed through a linear layer, a depth-separable convolution, an activation function SiLU, and a 2D selective scanning module for high-level feature extraction: in the other branch, the features are passed through a linear layer and an activation function SiLU and then multiplied element-by-element with the features of the first branch. Subsequently, the input features are residually connected with the original features to obtain the final VSS block output features.

[0061] like Figure 5 As shown, this implementation replaces the initial normalization module in the basic structure of the VSS block with a newly designed domain-specific layer normalization module; since different modalities exhibit different feature distributions, conventional layer normalization modules cannot accurately identify and process feature information from different modalities, and cannot learn general feature representations. Therefore, a specific layer normalization is configured for the features of each modality in the encoder to separately process its modality-related features, which will reduce the performance degradation caused by modality differences. The domain-specific layer normalization module contains two independent layer normalizations, each of which processes features from a specific domain. Its purpose is to estimate inter-layer statistics (mean and variance) and use learnable affine parameters to capture specific information for each domain. Domain-specific layer normalization DSLN can be defined as:

[0062] ,

[0063] Where d is the domain label, z is the feature from domain d belonging to [source domain s, target domain t], and represents the mean and variance, and is a learnable affine parameter and ε is a very small constant used to avoid instability when the denominator is zero.

[0064] In this design, the domain-specific layer normalization module is used to capture domain-specific affine parameters to adapt domain-specific features, while the 2D selective scanning parameters are shared between different domains to learn general anatomical representations. The first three stages use patch merging after the VSS block to reduce the feature size and increase the number of channels. The number of channels in the four stages is set to [C, 2C, 4C, 8C]. This implementation sets N1, N2, N4 = 2, N3 = 5 through experiments.

[0065] like Figure 6 As shown, the SASI module is used to process the multi-scale feature maps from the second encoder, which can effectively integrate feature information from different modalities and different scales.

[0066] First, the spatial and channel attention is calculated through the convolutional block attention module (CBAM); then, a 1×1 convolution maps all features to the same dimension and performs feature fusion. Finally, the features are downsampled or upsampled to various scales through adaptive average pooling or bilinear interpolation and passed to the second decoder part. Each level of the second decoder accepts the fused features from the SASI module. On the other hand, three sets of transposed convolutions are used to upsample the underlying feature information and fuse the underlying features with the previous high-level features to finally complete the resolution reconstruction and segmentation tasks.

[0067] The source domain image, target domain image, and translated image are segmented by the segmenter to produce their respective segmentation results ps, ps-t, and pt. The source domain label is used to supervise ps and ps-t:

[0068] ,

[0069] Where, L sup represents the supervision loss, L ce represents the cross entropy loss function, L dice represents the Dice loss function; p s represents the source domain segmentation result; p s-t represents the translation image segmentation result; p t represents the target domain image segmentation result; y s Represents the source domain label.

[0070] At the same time, adversarial learning is also performed in the prediction space to complete domain adaptation. Specifically, the second discriminator and the segmenter will perform adversarial learning. In this setting, the segmenter is trained to generate similar output space distributions for source domain images and target domain images to deceive the discriminator, which recognizes the target domain segmentation structure as a fake sample:

[0071] ,

[0072] In the formula, E represents the expected value, which represents the average operation of the relevant variables; x represents the image (the corresponding y represents the label) t~T represents the target domain distribution; D S represents the second discriminator; p t represents the target domain image segmentation result; x s~S Represents the source domain distribution.

[0073] S3. Train the improved domain adaptation framework end-to-end to obtain a segmentation network.

[0074] Use the target domain dataset, load the trained model, select the domain-specific layer normalization module, test the dataset, and get the final segmentation prediction result.

[0075] S4. Use the segmentation network to predict the image segmentation result of the target domain dataset.

[0076] Embodiment 2

[0077] The following will describe in detail how the present invention solves the technical problems in medical image processing in conjunction with the present embodiment.

[0078] In this embodiment, the acquired data is a cross-modal heart segmentation dataset from MMWHS2017. It contains 20 computed tomography (CT) modality data and 20 magnetic resonance imaging (MRI) modality data. We select four substructures for segmentation: ascending aorta, left atrial blood cavity, left ventricular blood cavity, and left ventricular myocardium. We select 80% of them for training and 20% for testing. All data are normalized to zero mean and unit variance. In order to train our 2D network for volume data segmentation, we use coronal view image slices of the heart volume, and the slices are resampled to 128*128 size.

[0079] In this embodiment, the generation network and the segmentation network are trained end-to-end, and the parameters are updated in the order of generator → segmenter → second discriminator & first discriminator. The Adam optimizer parameters β1=0.5, β2=0.999 are used. Taking into account the limitation of video memory, the training batch is set to 8, the number of training rounds is set to 200, and the learning rate is set to 0.0002. Table 1 is a comparison between this embodiment and the prior art in the direction of MRI→CT of the cardiac multimodal dataset; Table 2 is a comparison between the present invention and the prior art in the direction of CT→MRI of the cardiac multimodal dataset. It can be seen that the present invention is significantly superior to other algorithms, and is close to the upper limit of full supervision in some indicators.

[0080] Table 1

[0081] ,

[0082] Table 2

[0083] .

[0084] Embodiment 3

[0085] In this embodiment, the acquired data is a cross-modal abdominal multi-organ segmentation dataset from . It contains 20 computed tomography (CT) modality data from the 2019 CHAOS Challenge and 30 magnetic resonance imaging (MRI) modality data from the BTCV abdominal organ segmentation dataset. This embodiment selects four organs for segmentation: liver, right kidney, left kidney, and spleen. This embodiment selects 80% of them for training and 20% for testing. All data are normalized to zero mean and unit variance. In order to train the 2D network of this embodiment for volumetric data segmentation, this embodiment uses axial view image slices of the abdominal volume, and the slices are resampled to 256*256 size.

[0086] In this embodiment, the generator network and the segmentation network are trained end-to-end, and the parameters are updated in the order of generator → segmentation → second discriminator & first discriminator. The Adam optimizer parameters β1=0.5, β2=0.999 are used. Considering the memory limitation, the training batch is set to 4, the number of training rounds is set to 200, and the learning rate is set to 0.0002.

[0087] Table 3 shows the comparison between the present invention and the prior art in the MRI→CT direction of the abdominal multimodal dataset; Table 4 shows the comparison between the present invention and the prior art in the CT→MRI direction of the abdominal multimodal dataset. It can be seen that the present invention is significantly superior to other algorithms and is close to the upper limit of full supervision in some indicators.

[0088] Table 3

[0089] ,

[0090] Table 4

[0091] .

[0092] Embodiment 4

[0093] This embodiment also provides a cross-modal medical image segmentation system based on an improved domain adaptation framework of a Mamba network, including: an acquisition module, a construction module, a training module and a segmentation module; the acquisition module is used to preprocess a pre-acquired cross-modal medical image dataset and form a source domain dataset and a target domain dataset; the construction module is used to build an improved domain adaptation framework using a Mamba network, and the improved domain adaptation framework includes: a generation network and a segmentation network; the generation network includes a generator, a channel compression module and a first discriminator, and is used to generate a translation image based on a labeled source domain dataset; the segmentation network includes a segmentor and a second discriminator; the training module is used to train the improved domain adaptation framework end-to-end to obtain a segmentation network; the segmentation module is used to use the segmentation network to predict the image segmentation result of the target domain dataset.

[0094] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A cross-modal medical image segmentation method based on an improved domain adaptation framework of a Mamba network, characterized in that the steps include: S1. Preprocess the pre-acquired cross-modal medical image dataset and form a source domain dataset and a target domain dataset; S2. Building a domain adaptation framework improved by using the Mamba network, the improved domain adaptation framework comprising: a generation network and a segmentation network; the generation network comprising a generator and a first discriminator, for generating a translation image based on the labeled source domain dataset; the segmentation network comprising a segmentor and a second discriminator; S3. training the improved domain adaptation framework end-to-end to obtain a segmentation network; S4. using the segmentation network to predict the image segmentation result of the target domain data set; Wherein, in the generation network, the generator includes: a first encoder, a bottleneck layer and a first decoder; The first encoder consists of input processing and two downsampling layers to extract high-level feature representations; The bottleneck layer is a Mamba hybrid bottleneck module and a residual block alternately performed; the Mamba hybrid bottleneck module includes a channel compression module, and the channel compression module is used to further reduce the number of channels and maintain key information in the features; The first decoder is composed of two upsampling layers and output processing; In the first encoder, the input processing includes a reflection padding layer, a convolution layer, an instance normalization layer and a ReLU activation function; the two downsampling layers have the same structure, both consisting of a convolution layer, an instance normalization layer and a ReLU activation function; In the first decoder, the output processing includes a reflection padding layer, a convolution layer and a Tanh activation function; the two upsampling layers have the same structure, both consisting of a transposed convolution layer, an instance normalization layer and a ReLU activation function; In the bottleneck layer, the features are first linearly processed in the Mamba hybrid bottleneck module and then divided into two branches; in the first branch, the features are sequentially processed by a depthwise separable convolutional layer, an activation layer, and a state equation modeling layer; in the second branch, the features are processed by an activation function SiLU; finally, after combining the features from the two branches using the Hadamard product, the obtained features are combined with the input features through a residual connection; In the segmentation network, the segmenter is composed of a second encoder, a style adjustment and semantic injection jump connection module, and a second decoder; before the second encoder, the image is divided into non-overlapping blocks of size 4×4 by patch embedding, and the dimension mapping is C; the second encoder includes four stages, each of which uses N1, N2, N3, and N4 visual state equation blocks; the style adjustment and semantic injection module is used to process the multi-scale feature map from the second encoder, and first calculates the spatial and channel attention through the convolutional attention module; then, 1×1 convolution maps all features to the same dimension and performs feature fusion; finally, the features are downsampled or upsampled to each scale and passed to the second decoder part through adaptive average pooling or bilinear difference; each level of the second decoder accepts the fused features from the style adjustment and semantic injection module, uses three sets of transposed convolutions to upsample the underlying feature information, and fuses the underlying features with the previous level high-level features, and finally completes the resolution reconstruction and segmentation tasks; The initial normalization module in the visual state equation block is replaced by a domain-specific layer normalization module. The domain-specific layer normalization module contains two independent layer normalizations. Each layer normalization processes features from a specific domain to estimate inter-layer statistics and uses learnable affine parameters to capture specific information of each domain. The domain-specific layer normalization DSLN is defined as: ， Where d is the domain label, z is the feature from domain d, and represents the mean and variance, and is a learnable affine parameter and ε is a very small constant used to avoid instability when the denominator is zero.

2. The cross-modal medical image segmentation method based on the Mamba network improved domain adaptation framework according to claim 1, characterized in that: The first discriminator adopts an adversarial learning approach, using the source domain label and the source domain dataset as joint input to serve as additional structural constraints to assist the generation process.

3. A cross-modal medical image segmentation system based on a Mamba network improved domain adaptation framework, the system being used to implement the method according to any one of claims 1-2, characterized in that: include: Acquisition module, construction module, training module and segmentation module; The acquisition module is used to preprocess the pre-acquired cross-modal medical image dataset and form a source domain dataset and a target domain dataset; The building module is used to build a domain adaptation framework improved by using a Mamba network, wherein the improved domain adaptation framework includes: a generation network and a segmentation network; the generation network includes a generator and a first discriminator, and is used to generate a translation image based on the labeled source domain dataset; the segmentation network includes a segmentor and a second discriminator; The training module is used to train the improved domain adaptation framework end-to-end to obtain a segmentation network; The segmentation module is used to predict the image segmentation result of the target domain data set using the segmentation network.

Images