Medical image segmentation device and method based on mixed state space architecture

Abstract

The invention discloses a medical image segmentation device and method based on a mixed state space architecture, the device is provided with an input module, a mixed visual state space coding module, an axial cross-information bridging module, a decoding and supervision module and an output module, and the method comprises the following steps: 1) obtaining a medical image to be processed; the method comprises the steps of (1) image segmentation, (2) preprocessing, (3) coding feature obtaining, (4) feature reconstruction and (5) medical image segmentation device.According to the method, structure information and detail information in the image are reserved at the same time, high-precision low-delay automatic segmentation can be achieved, and the method is suitable for medical image auxiliary diagnosis, preoperative planning and other scenes.

Description

Technical Field

[0001] This invention relates to image processing technology, specifically a medical image segmentation device and method based on a hybrid state-space architecture. Background Technology

[0002] Medical image segmentation plays a crucial role in the early detection and diagnosis of diseases. In the past, doctors manually segmented and labeled target regions, but this traditional method was highly subjective and inefficient. Efficient computational methods utilizing deep learning have provided stronger guidance for improving the efficiency and accuracy of medical image segmentation. Currently, a large amount of research is dedicated to developing various medical image segmentation systems and automated segmentation methods, and significant results have been achieved.

[0003] Currently, medical image segmentation methods are mainly divided into three architectures: those based on U-Net improvements, those based on Transformer improvements, and those based on state-space models. U-Net-based architectures primarily employ a U-shaped encoder-decoder structure with various skip connections to fuse semantic information at different scales. These architectures leverage the translational equivariance and local receptive fields of convolutional operations to effectively capture local spatial patterns and fine-grained anatomical details. However, using fixed-size convolutional kernels, they cannot extract all meaningful features from medical image datasets with varying spatial backgrounds. Furthermore, the inherent locality of convolutional operations limits their ability to model long-range dependencies, making it difficult to effectively capture spatial relationships across organs and scales, thus limiting segmentation accuracy. Transformer-based architectures model global dependencies through self-attention mechanisms, effectively modeling long-range dependencies and addressing the shortcomings of convolutional neural networks in this area. However, self-attention mechanisms focus more on global information and are less effective at preserving detailed information in images, making it difficult to effectively capture fine-grained details. Boundary features are a concern. Secondly, the computational complexity of the self-attention mechanism is proportional to the square of the image resolution. When processing high-resolution medical images, the computational and memory consumption increases dramatically, leading to a significant decrease in training and inference speed, making it difficult to use in actual clinical diagnosis and treatment. Network architectures based on state-space models capture long-distance dependencies through selective state-space mechanisms while maintaining linear computational complexity relative to sequence length, solving the computational complexity problem of the Transformer architecture. However, these architectures mainly focus on global sequence modeling and lack explicit extraction mechanisms for multi-scale local features. They are difficult to simultaneously consider global context understanding and local detail preservation, resulting in insufficient segmentation accuracy for small lesions or boundary regions. In addition, the skip connections in these architectures use simple feature concatenation or addition, which cannot fully utilize the semantic differences and spatial anisotropy between encoder and decoder features, limiting the preservation of fine-grained details during feature transfer.

[0004] Currently, numerous network architectures are used in medical image segmentation research, but few can simultaneously solve the aforementioned problems. Therefore, extracting all meaningful feature information from medical image datasets with different spatial backgrounds while preserving structural and detailed information within the images, under conditions of low computational and storage overhead, remains a challenge for accurate medical image recognition and segmentation. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a medical image segmentation device and method based on a hybrid state-space architecture. This method simultaneously preserves structural and detailed information in the image, enabling high-precision, low-latency automatic segmentation, and is suitable for scenarios such as medical image-assisted diagnosis and preoperative planning.

[0006] The technical solution to achieve the objective of this invention is: A medical image segmentation device based on a hybrid state-space architecture, the medical image segmentation device comprising: Input module: The input module is used to acquire the medical image to be segmented, and to perform normalization, enhancement and patch embedding on the medical image in sequence, and output a multi-scale feature medical image; Hybrid visual state space coding module: The hybrid visual state space coding module has parallel global state space branches and multi-scale local branches, which are used to fuse the global features output by the global state space branches and the local features output by the multi-scale local branches according to smooth progressive weights to obtain the coded features; Axial cross-information bridging module: The axial cross-information bridging module is equipped with a channel alignment unit for unifying the encoded features and corresponding decoded features, a context modulation unit for combining the channel-aligned features with global context information, a horizontal pooling unit and a vertical pooling unit for performing average pooling along the horizontal and vertical directions respectively, a directional attention unit for generating horizontal and vertical attention coefficients based on the pooling results and multiplying them element-wise with the modulation features, and a residual fusion unit for performing residual fusion of the output of the directional attention unit with the corresponding decoded features to obtain refined features. The axial cross-information bridging module is used to perform channel alignment, context modulation, directional attention generation, and residual fusion on the encoded features and decoded features, and output refined features. Decoding and Supervision Module: The decoding and supervision module is used to perform multi-level decoding on the refined features and apply progressively weighted depth supervision at multiple decoding levels to generate segmentation features; Output module: The output module is used to convert the segmentation features into medical image segmentation results and output them.

[0007] The global state space branch in the hybrid vision state space coding module performs sequential scanning along four orthogonal directions: horizontal, vertical, and their opposite directions, and outputs selective state responses; the multi-scale local branches in the hybrid vision state space coding module are sequentially equipped with depth-separable convolutions and channel attention modules with dilation rates of 1, 3, and 5.

[0008] A medical image segmentation method based on a hybrid state-space architecture, employing the aforementioned medical image segmentation device based on a hybrid state-space architecture, includes the following steps: 1) Acquire the medical images to be processed: The images can be dermoscopy images, endoscopy images, multi-organ CT images, MRI images or other medical images. The medical images are from public datasets, hospital PACS systems or local storage devices. Medical images include dermoscopy images, endoscopy images, multi-organ CT images and MRI images. The medical images are unified into single-channel or three-channel format and adjusted to the preset spatial resolution of 224×224 or 256×256 pixels according to the network input requirements. 2) Preprocessing: The medical images are normalized, enhanced, and patched sequentially to obtain multi-scale feature medical images. Normalization, resizing, and standardization operations are performed on the original images to map pixel values ​​to the range of [0,1], reducing the distribution differences caused by different acquisition devices and imaging parameters. During the training phase, random rotation, horizontal / vertical flipping, random cropping, brightness and contrast adjustment data augmentation strategies are also used to expand the training samples to improve the robustness of the model to scale changes, deformation, and illumination changes. The preprocessed medical images are then processed by a shallow convolutional feature extraction module and downsampling operations to construct multi-scale initial feature representations, providing input for the subsequent hybrid state space coding network. 3) Obtaining Encoded Features: Multi-scale features of medical images are input into a hybrid visual state-space coding network. Global and local features are fused using smooth progressive weights to obtain encoded features. The smooth progressive weights employ cosine scheduling. In the first training phase, the global branch weights are set to be greater than the local branch weights. In the second training phase, the local branch weights are set to be greater than the global branch weights. Specifically: The encoding network uses Hybrid Visual State Space (HVST) modules as basic units at each encoding stage. Each HVST module includes a Visual State Space (VSS) branch and a multi-scale local enhancement branch, and combines features through a smooth progressive fusion mechanism. The VSS branch is implemented based on the VMamba visual state space mechanism. After linear projection and depthwise separable convolution enhancement of the input features, the branch rearranges the two-dimensional features into a sequence according to the horizontal, vertical and their opposite directions. The sequence is scanned using a selective state space recursive formula to generate state responses in four directions, which are then aggregated along the channel dimension to form global context features with long-distance dependencies. , , in, Indicates position The state vector at that point, and These represent the input features and the output features, respectively. These are learnable parameters; The multi-scale local enhancement branch employs three sets of depthwise separable convolutions with different dilation rates to perform convolution operations on the input features. The dilation rates are 1, 2, and 4, corresponding to effective receptive fields of 3×3, 5×5, and 9×9, respectively. This jointly captures fine-grained boundary, small target lesion, and large-scale structural information. The outputs of the three convolutions are concatenated along the channel dimension and then input into the Squeeze-and-Excitation channel attention module. The channel attention module generates channel weights through global average pooling and two fully connected transformations, reweighting the channels to enhance the channel responses related to lesions and important anatomical structures. The convolution outputs are normalized using GroupNorm to ensure numerical stability during small-batch training. , in, These are features resulting from multi-scale local enhancement. Indicates the expansion rate The convolution operator acts on feature F. This represents the k-th channel attention module; The HVST module sets a learnable local branch weight parameter and a maximum weight upper limit that changes with training progress, forming a smooth and progressive fusion mechanism. During training, the training progress is calculated based on the current epoch, and the corresponding maximum weight value is generated through a piecewise cosine curve, so that the maximum fusion weight of the local branch smoothly increases from 0 to about 0.6. The actual weight of the local branch is limited to not exceeding the current maximum weight after the sigmoid function is applied to the learnable parameter. The global branch weight is equal to 1 minus the local branch weight. The VSS branch output and the multi-scale local branch output are weighted and summed according to their corresponding weights, and then added to the module input through a residual connection to obtain the fused encoded features. Global semantic information and local details are preserved simultaneously within the same feature path. , in, and These represent the features of the global state space branch and the multi-scale local branch, respectively. and The corresponding fusion weights; 4) Feature Reconstruction: The encoded features are input into the axial cross-information bridging module and collaborate with the decoding network to complete context modulation and axial refinement, thus completing feature reconstruction. Specifically, the axial cross-information bridging module sequentially performs channel alignment, context modulation, horizontal average pooling, vertical average pooling, directional attention generation, and residual fusion. It then performs element-wise multiplication of the modulated features based on the horizontal and vertical attention coefficients to obtain the refined features. The axial cross-information bridging module is positioned between each layer of the encoder and the corresponding decoding layer. First, it performs convolutional transformations and size alignment on the features from the encoder and decoder to obtain feature representations with consistent channel count and spatial resolution. Then, it performs element-wise summation on the aligned encoded and decoded features, followed by ReLU activation and convolution to generate a modulation factor α. This α is then constrained to the range [0,1] by Sigmoid activation. This modulation factor is used as a weight coefficient and multiplied with the encoded features to form enhanced jump features. To introduce axial structural information, the axial cross-information bridging module performs average pooling on the features along the horizontal and vertical axes in the downstream branch, obtaining axial features in the X and Y directions. The two axial features are concatenated along the channel dimension and input into the convolutional layer. The convolutional output is then passed through a batch normalization layer and a ReLU linear rectified function. The nonlinear activation process yields fused features that simultaneously encode horizontal and vertical structural information. These features are then processed by a convolutional layer to generate an axial weight map. The axial weight map is then multiplied element-wise with the aforementioned enhanced jump features to achieve feature refinement that considers both scale and axis. The output is used as the input features for the decoding stage. Through this process, the decoding network gains stronger boundary awareness and structural consistency when restoring spatial resolution. The decoding network employs a multi-level upsampling structure. The features of the lowest resolution layer are first restored to the spatial size of the previous scale by the upsampling module and then fused with the encoded features corresponding to that scale by the axial cross-information bridging module. The features are then further refined through convolution and nonlinear activation. This process is repeated at each scale to restore the feature map resolution step by step, ultimately resulting in a high-resolution feature map consistent with the input image resolution, providing support for the output of fine segmentation results. 5) Outputting Medical Image Segmentation Results: Progressively weighted depth supervision is applied to multiple decoding layers of the decoding network to output medical image segmentation results. The progressively weighted depth supervision uses an exponential normalization function to assign weights to the auxiliary losses of each decoding layer. Shallow layer weights decrease with each training epoch, while deep layer weights increase. Specifically, the decoding network sets up depth supervision heads on feature maps at multiple scales. Each supervision head consists of a convolutional layer and an upsampling operation, outputting segmentation predictions with the same resolution as the original image. During training, Dice loss, cross-entropy loss, and boundary-aware loss are calculated for each scale prediction, and time-varying weights are assigned to the losses at different scales. Shallow supervision heads have higher weights in the early stages of training, while deep supervision heads have lower weights. As training progresses, the weights of shallow supervision heads gradually decrease, while the weights of deep supervision heads gradually increase, enabling the network to converge quickly in the early stages and rely more on semantically rich high-level features for optimization in the later stages. The sum of the weighted losses at each scale is used as the total loss to guide network parameter updates.

[0009] The beneficial effects of this technical solution are: This approach employs a state-space model to capture global contextual information, mitigating the shortcomings of traditional convolutional neural networks in long-range dependency modeling while maintaining linear computational complexity. The multi-scale local branches in this solution effectively capture local features at different scales, significantly improving segmentation performance, particularly for small lesions and complex boundary regions. Compared to Transformer-based architectures, this approach significantly reduces computational complexity and memory consumption while maintaining high performance, making it more suitable for processing high-resolution medical images. The use of an axial cross-information bridging module and a multi-scale feature fusion mechanism effectively preserves detailed information and structural features in the image, improving the accuracy of segmentation boundaries. Furthermore, the progressively weighted deep supervision strategy enables end-to-end training, simplifying the model training process and improving training efficiency and model performance.

[0010] This method preserves both structural and detail information in the image, enabling high-precision, low-latency automatic segmentation, and is suitable for scenarios such as medical image-assisted diagnosis and preoperative planning. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the method used in the embodiment. Figure 2 This is a schematic diagram of the hybrid vision state space module structure in the embodiment; Figure 3 This is a schematic diagram of the axial cross-information bridging module structure in the embodiment; Figure 4 This is a comparison chart showing the performance of the example method and existing methods on medical image segmentation tasks. Detailed Implementation

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0013] Example: A medical image segmentation device based on a hybrid state-space architecture, the medical image segmentation device comprising: Input module: The input module is used to acquire the medical image to be segmented, and to perform normalization, enhancement and patch embedding on the medical image in sequence, and output a multi-scale feature medical image; Hybrid visual state space coding module: The hybrid visual state space coding module has parallel global state space branches and multi-scale local branches, which are used to fuse the global features output by the global state space branches and the local features output by the multi-scale local branches according to smooth progressive weights to obtain the coded features; Axial cross-information bridging module: The axial cross-information bridging module is equipped with a channel alignment unit for unifying the encoded features and corresponding decoded features, a context modulation unit for combining the channel-aligned features with global context information, a horizontal pooling unit and a vertical pooling unit for performing average pooling along the horizontal and vertical directions respectively, a directional attention unit for generating horizontal and vertical attention coefficients based on the pooling results and multiplying them element-wise with the modulation features, and a residual fusion unit for performing residual fusion of the output of the directional attention unit with the corresponding decoded features to obtain refined features. The axial cross-information bridging module is used to perform channel alignment, context modulation, directional attention generation, and residual fusion on the encoded features and decoded features, and output refined features. Decoding and Supervision Module: The decoding and supervision module is used to perform multi-level decoding on the refined features and apply progressively weighted depth supervision at multiple decoding levels to generate segmentation features; Output module: The output module is used to convert the segmentation features into medical image segmentation results and output them.

[0014] The global state space branch in the hybrid vision state space coding module performs sequential scanning along four orthogonal directions: horizontal, vertical, and their opposite directions, and outputs selective state responses; the multi-scale local branches in the hybrid vision state space coding module are sequentially equipped with depth-separable convolutions and channel attention modules with dilation rates of 1, 3, and 5.

[0015] like Figure 1 As shown, a medical image segmentation method based on a hybrid state-space architecture employs the aforementioned medical image segmentation device based on a hybrid state-space architecture. The method includes the following steps: 1) Acquire the medical images to be processed: The images are dermoscopy images, endoscopy images, multi-organ CT images, MRI images or other medical images. The medical images are from public datasets, hospital PACS systems or local storage devices. The medical images include dermoscopy images, endoscopy images, multi-organ CT images and MRI images. In this example, the medical images are unified into single-channel or three-channel format and adjusted to the preset spatial resolution of 224×224 or 256×256 pixels according to the network input requirements. 2) Preprocessing: The medical images are normalized, enhanced, and patched sequentially to obtain multi-scale feature medical images. Normalization, resizing, and standardization operations are performed on the original images to map pixel values ​​to the range of [0,1], reducing the distribution differences caused by different acquisition devices and imaging parameters. During the training phase, random rotation, horizontal / vertical flipping, random cropping, brightness and contrast adjustment data augmentation strategies are also used to expand the training samples to improve the robustness of the model to scale changes, deformation, and illumination changes. The preprocessed medical images are then processed by a shallow convolutional feature extraction module and downsampling operations to construct multi-scale initial feature representations, providing input for the subsequent hybrid state space coding network. 3) Obtaining Encoded Features: Multi-scale features of medical images are input into a hybrid visual state-space coding network. Global and local features are fused using smooth progressive weights to obtain encoded features. The smooth progressive weights employ cosine scheduling. In the first training phase, the global branch weights are set to be greater than the local branch weights. In the second training phase, the local branch weights are set to be greater than the global branch weights. Specifically: like Figure 2 As shown, the encoding network uses Hybrid Visual State Space (HVST) modules as basic units at each encoding stage. Each HVST module includes a Visual State Space (VSS) branch and a multi-scale local enhancement branch, and combines features through a smooth progressive fusion mechanism. The VSS branch is implemented based on the VMamba visual state space mechanism. After linear projection and depthwise separable convolution enhancement of the input features, the branch rearranges the two-dimensional features into a sequence according to the horizontal, vertical, and their opposite directions. The sequence is scanned using a selective state space recursive formula to generate state responses in four directions, which are then aggregated along the channel dimension to form global context features with long-distance dependencies. , , in, Indicates position The state vector at that point, and These represent the input features and the output features, respectively. These are learnable parameters; The multi-scale local enhancement branch employs three sets of depthwise separable convolutions with different dilation rates to perform convolution operations on the input features. The dilation rates are 1, 2, and 4, corresponding to effective receptive fields of 3×3, 5×5, and 9×9, respectively. This jointly captures fine-grained boundary, small target lesion, and large-scale structural information. The outputs of the three convolutions are concatenated along the channel dimension and then input into the Squeeze-and-Excitation channel attention module. The channel attention module generates channel weights through global average pooling and two fully connected transformations, reweighting the channels to enhance the channel responses related to lesions and important anatomical structures. The convolution outputs are normalized using GroupNorm to ensure numerical stability during small-batch training. , in, These are features resulting from multi-scale local enhancement. Indicates the expansion rate The convolution operator acts on feature F. This represents the k-th channel attention module; The HVST module sets a learnable local branch weight parameter and a maximum weight upper limit that changes with training progress, forming a smooth and progressive fusion mechanism. During training, the training progress is calculated based on the current epoch, and the corresponding maximum weight value is generated through a piecewise cosine curve, so that the maximum fusion weight of the local branch smoothly increases from 0 to about 0.6. The actual weight of the local branch is limited to not exceeding the current maximum weight after the sigmoid function is applied to the learnable parameter. The global branch weight is equal to 1 minus the local branch weight. The VSS branch output and the multi-scale local branch output are weighted and summed according to their corresponding weights, and then added to the module input through a residual connection to obtain the fused encoded features. Global semantic information and local details are preserved simultaneously within the same feature path. , in, and These represent the features of the global state space branch and the multi-scale local branch, respectively. and The corresponding fusion weights; 4) Feature Reconstruction: The encoded features are input into the axial cross-information bridging module and collaborate with the decoding network to complete context modulation and axial refinement, thus completing feature reconstruction. Specifically, the axial cross-information bridging module sequentially performs channel alignment, context modulation, horizontal average pooling, vertical average pooling, directional attention generation, and residual fusion. It then performs element-wise multiplication of the modulated features based on the horizontal and vertical attention coefficients to obtain the refined features. like Figure 3As shown, the axial cross-information bridging module is positioned between each layer of the encoder and the corresponding decoding layer. First, it performs convolutional transformation and size alignment on the features from the encoder and decoder to obtain feature representations with consistent channel count and spatial resolution. Then, it performs element-wise summation on the aligned encoded and decoded features, followed by ReLU activation and convolution to generate a modulation factor α. This α is then constrained to the range [0,1] by Sigmoid activation. This modulation factor is used as a weight coefficient and multiplied with the encoded features to form enhanced jump features. To introduce axial structural information, the axial cross-information bridging module performs average pooling on the features along the horizontal and vertical axes in the downstream branch to obtain axial features in the X and Y directions. The two axial features are concatenated along the channel dimension and input into the convolutional layer. The convolutional output sequentially passes through a batch normalization layer and a ReLU linear rectified function. The nonlinear activation process yields fused features that simultaneously encode horizontal and vertical structural information. These features are then processed by a convolutional layer to generate an axial weight map. The axial weight map is then multiplied element-wise with the aforementioned enhanced jump features to achieve feature refinement that considers both scale and axis. The output is used as the input features for the decoding stage. Through this process, the decoding network gains stronger boundary awareness and structural consistency when restoring spatial resolution. The decoding network employs a multi-level upsampling structure. The features of the lowest resolution layer are first restored to the spatial size of the previous scale by the upsampling module and then fused with the encoded features corresponding to that scale by the axial cross-information bridging module. The features are then further refined through convolution and nonlinear activation. This process is repeated at each scale to restore the feature map resolution step by step, ultimately resulting in a high-resolution feature map consistent with the input image resolution, providing support for the output of fine segmentation results. 5) Outputting Medical Image Segmentation Results: Progressively weighted depth supervision is applied to multiple decoding layers of the decoding network to output medical image segmentation results. The progressively weighted depth supervision uses an exponential normalization function to assign weights to the auxiliary losses of each decoding layer. Shallow layer weights decrease with each training epoch, while deep layer weights increase. Specifically, the decoding network sets up depth supervision heads on feature maps at multiple scales. Each supervision head consists of a convolutional layer and an upsampling operation, outputting segmentation predictions with the same resolution as the original image. During training, Dice loss, cross-entropy loss, and boundary-aware loss are calculated for each scale prediction, and time-varying weights are assigned to the losses at different scales. Shallow supervision heads have higher weights in the early stages of training, while deep supervision heads have lower weights. As training progresses, the weights of shallow supervision heads gradually decrease, while the weights of deep supervision heads gradually increase, enabling the network to converge quickly in the early stages and rely more on semantically rich high-level features for optimization in the later stages. The sum of the weighted losses at each scale is used as the total loss to guide network parameter updates.

[0016] In this example, during the model training phase, the AdamW optimizer is used with an initial learning rate of 0.001. A cosine annealing learning rate scheduling strategy is employed, with a minimum learning rate of 1 / 10 of the initial learning rate (0.001 in this example). The batch size is set to 16 or 32 depending on the available GPU memory. An early stopping strategy is used, terminating training prematurely if the validation set performance does not improve within 20 consecutive epochs. During the inference phase, the medical image to be segmented is input, preprocessed, and then fed into the trained model to obtain preliminary segmentation results. Post-processing of the segmentation results includes thresholding of the probability map, connected component analysis, and small region removal, outputting the final medical image segmentation mask. Figure 4 As shown, the proposed method is compared with typical methods such as U-Net, TransUNet, MALUNet, and VM-UNet on several publicly available medical image segmentation datasets for skin lesion segmentation, polyp segmentation, and multi-organ segmentation. On these datasets, the proposed method achieves higher segmentation accuracy in evaluation metrics such as Dice coefficient and IoU, and performs outstandingly in the localization of small targets and complex boundaries. At the same time, relying on the linear computational complexity of the state-space model and the lightweight multi-scale local branch design, the proposed method reduces the number of parameters and computational load while maintaining segmentation accuracy, and improves inference speed, which can meet the comprehensive requirements of clinical applications for real-time performance and resource consumption.

Claims

1. A medical image segmentation device based on a hybrid state-space architecture, characterized in that, The medical image segmentation device is equipped with: Input module: The input module is used to acquire the medical image to be segmented, and to perform normalization, enhancement and patch embedding on the medical image in sequence, and output a multi-scale feature medical image; Hybrid visual state space coding module: The hybrid visual state space coding module has parallel global state space branches and multi-scale local branches, which are used to fuse the global features output by the global state space branches and the local features output by the multi-scale local branches according to smooth progressive weights to obtain the coded features; Axial cross-information bridging module: The axial cross-information bridging module is equipped with a channel alignment unit for unifying the encoded features and corresponding decoded features, a context modulation unit for combining the channel-aligned features with global context information, a horizontal pooling unit and a vertical pooling unit for performing average pooling along the horizontal and vertical directions respectively, a directional attention unit for generating horizontal and vertical attention coefficients based on the pooling results and multiplying them element-wise with the modulation features, and a residual fusion unit for performing residual fusion of the output of the directional attention unit with the corresponding decoded features to obtain refined features. The axial cross-information bridging module is used to perform channel alignment, context modulation, directional attention generation, and residual fusion on the encoded features and decoded features, and output refined features. Decoding and Supervision Module: The decoding and supervision module is used to perform multi-level decoding on the refined features and apply progressively weighted depth supervision at multiple decoding levels to generate segmentation features; Output module: The output module is used to convert segmentation features into medical image segmentation results and output them.

2. The medical image segmentation device based on a hybrid state-space architecture according to claim 1, characterized in that, The global state space branch in the hybrid vision state space coding module performs sequential scanning along four orthogonal directions: horizontal, vertical, and their opposite directions, and outputs selective state responses; the multi-scale local branches in the hybrid vision state space coding module are sequentially equipped with depth-separable convolutions and channel attention modules with dilation rates of 1, 3, and 5.

3. A medical image segmentation method based on a hybrid state-space architecture, employing the medical image segmentation apparatus based on a hybrid state-space architecture as described in any one of claims 1-2, characterized in that, The method includes the following steps: 1) Acquire the medical images to be processed: The images are dermoscopy images, endoscopy images, multi-organ CT images, MRI images or other medical images. The medical images are from public datasets, hospital PACS systems or local storage devices. The medical images include dermoscopy images, endoscopy images, multi-organ CT images and MRI images. The medical images are unified into single-channel or three-channel format and adjusted to the preset spatial resolution of 224×224 or 256×256 pixels according to the network input requirements. 2) Preprocessing: The medical images are normalized, enhanced, and patched sequentially to obtain multi-scale feature medical images. Normalization, resizing, and standardization operations are performed on the original images to map pixel values ​​to the range of [0,1], reducing the distribution differences caused by different acquisition devices and imaging parameters. During the training phase, random rotation, horizontal / vertical flipping, random cropping, brightness and contrast adjustment data augmentation strategies are also used to expand the training samples. The preprocessed medical images are then processed by a shallow convolution feature extraction module and downsampling operations to construct multi-scale initial feature representations, providing input for the subsequent hybrid state space coding network. 3) Obtain the encoded features: Multi-scale features of medical images are input into a hybrid visual state-space coding network. Global and local features are fused using smooth progressive weights to obtain encoded features. The smooth progressive weights employ cosine scheduling. In the first training phase, the global branch weights are set to be greater than the local branch weights, and in the second training phase, the local branch weights are set to be greater than the global branch weights. Specifically: The encoding network uses Hybrid Visual State Space (HVST) modules as basic units at each encoding stage. Each HVST module includes a Visual State Space (VSS) branch and a multi-scale local enhancement branch, and combines features through a smooth progressive fusion mechanism. The VSS branch is implemented based on the VMamba visual state space mechanism. After linear projection and depthwise separable convolution enhancement of the input features, the branch rearranges the two-dimensional features into a sequence according to the horizontal, vertical and their opposite directions. The sequence is scanned using a selective state space recursive formula to generate state responses in four directions. These responses are then aggregated along the channel dimension to form global context features with long-distance dependencies. ， ， in, Indicates position The state vector at that point, and These represent the input features and the output features, respectively. These are learnable parameters; The multi-scale local enhancement branch employs three sets of depthwise separable convolutions with different dilation rates to perform convolution operations on the input features. The dilation rates are 1, 2, and 4, corresponding to effective receptive fields of 3×3, 5×5, and 9×9, respectively. This jointly captures fine-grained boundary, small target lesion, and large-scale structural information. The outputs of the three convolutions are concatenated along the channel dimension and then input into the Squeeze-and-Excitation channel attention module. The channel attention module generates channel weights through global average pooling and two fully connected transformations, reweighting the channels to enhance the channel responses related to lesions and important anatomical structures. The convolution outputs are normalized using GroupNorm. ， in, These are features resulting from multi-scale local enhancement. Indicates the expansion rate The convolution operator acts on feature F. This represents the k-th channel attention module; The HVST module sets a learnable local branch weight parameter and a maximum weight upper limit that changes with training progress, forming a smooth and progressive fusion mechanism. During training, the training progress is calculated based on the current epoch, and the corresponding maximum weight value is generated through a piecewise cosine curve, so that the maximum fusion weight of the local branch smoothly increases from 0 to about 0.

6. The actual weight of the local branch is limited to not exceeding the current maximum weight after the sigmoid function is applied to the learnable parameter. The global branch weight is equal to 1 minus the local branch weight. The VSS branch output and the multi-scale local branch output are weighted and summed according to their corresponding weights, and then added to the module input through a residual connection to obtain the fused encoded features. Global semantic information and local details are preserved simultaneously within the same feature path. , in, and These represent the features of the global state space branch and the multi-scale local branch, respectively. and The corresponding fusion weights; 4) Feature Reconstruction: The encoded features are input into the axial cross-information bridging module and collaborate with the decoding network to complete context modulation and axial refinement, thus completing feature reconstruction. Specifically, the axial cross-information bridging module sequentially performs channel alignment, context modulation, horizontal average pooling, vertical average pooling, directional attention generation, and residual fusion. It then performs element-wise multiplication of the modulated features based on the horizontal and vertical attention coefficients to obtain the refined features. An axial cross-information bridging module is positioned between each layer of the encoder and its corresponding decoding layer. First, it performs convolutional transformations and size alignment on the features from the encoder and decoder to obtain feature representations with consistent channel count and spatial resolution. Then, it performs element-wise summation on the aligned encoded and decoded features, followed by ReLU activation and convolution to generate a modulation factor α, which is then constrained to [0,1] by Sigmoid activation. Within the specified range, the modulation factor is used as a weight coefficient and multiplied with the encoded features to form enhanced jump features. The axial cross-information bridging module performs average pooling on the features along the horizontal and vertical axes in the downstream branch to obtain axial features in the X and Y directions. The two axial features are concatenated along the channel dimension and input into the convolutional layer. The convolution output is processed by batch normalization layer and nonlinear activation of linear rectified function ReLU to obtain fused features that simultaneously encode horizontal and vertical structural information. Then, it is processed by another convolution to generate an axial weight map. The axial weight map is multiplied element-wise with the aforementioned enhanced jump features to achieve cross-scale and cross-axial feature refinement. The output is used as the input features in the decoding stage. The decoding network adopts a multi-level upsampling structure. The features of the lowest resolution layer are first restored to the spatial size of the previous scale by the upsampling module and fused with the encoded features corresponding to that scale by the axial cross-information bridging module. Then, the features are further refined by convolution and nonlinear activation. This process is repeated at each scale to restore the feature map resolution step by step, and finally a high-resolution feature map with the same resolution as the input image is obtained. 5) Outputting Medical Image Segmentation Results: Progressively weighted depth supervision is applied to multiple decoding layers of the decoding network to output medical image segmentation results. The progressively weighted depth supervision uses an exponential normalization function to assign weights to the auxiliary losses of each decoding layer. Shallow layer weights decrease with each training epoch, while deep layer weights increase. Specifically, the decoding network sets up depth supervision heads on feature maps at multiple scales. Each supervision head consists of a convolutional layer and an upsampling operation, outputting segmentation predictions with the same resolution as the original image. During training, Dice loss, cross-entropy loss, and boundary-aware loss are calculated for each scale prediction, and time-varying weights are assigned to the losses at different scales. Shallow supervision heads have higher weights in the early stages of training, while deep supervision heads have lower weights. As training progresses, the weights of shallow supervision heads gradually decrease, while the weights of deep supervision heads gradually increase, enabling the network to converge quickly in the early stages and rely more on semantically rich high-level features for optimization in the later stages. The sum of the weighted losses at each scale is used as the total loss to guide network parameter updates.

Images