A State-Space Model-Based Medical Image Segmentation Method for Abdominal Multi-Organs

Abstract

This invention belongs to the field of medical image processing technology, specifically relating to a method for abdominal multi-organ medical image segmentation based on a state-space model. Addressing the shortcomings of existing technologies such as the difficulty of CNNs in modeling long-distance dependencies, the high computational complexity of Transformers, and the large semantic gaps, incomplete segmentation contours, and easy organ encroachment issues in VM-UNet skip connections, this solution makes key improvements: It constructs an improved VM-UNet-Skip architecture, placing skip connections before downsampling to reduce the semantic gap between the small decoder features; it introduces a multi-scale global-local information aggregation module, capturing global anatomical dependencies through multi-head Mamba units and enhancing local detail representations with a convolutional gated feedforward network; and it relies on the VMamba encoder-decoder to achieve efficient feature extraction and reconstruction. The method flow includes: constructing initial feature representations through patch embedding layers, extracting multi-scale hierarchical features through the encoder, enhancing features through an information fusion module, and fusing and mapping the enhanced features to obtain the segmentation result through the decoder. This invention effectively improves the segmentation accuracy of abdominal multi-organs, solves the problems of insufficient contour integrity and organ encroachment, while reducing the number of model parameters and computational complexity, providing reliable technical support for clinical diagnosis and surgical planning.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method for segmenting abdominal multi-organ medical images based on a state-space model. Background Technology

[0002] Abdominal multi-organ medical image segmentation is a core technical step in the diagnosis and treatment of abdominal diseases. Accurate segmentation results can clearly present the anatomical structure, spatial location, and morphological characteristics of organs such as the liver, pancreas, and kidneys, providing key data support for disease screening, lesion localization, surgical planning, and efficacy evaluation. It is directly related to the accuracy of clinical diagnosis and the rationality of treatment plans, and has irreplaceable research value and practical significance in medical imaging applications.

[0003] Currently, multiple technical approaches have emerged in the field of medical image segmentation: Convolutional neural networks (such as U-Net) have become the traditional mainstream solution due to their advantages in local feature extraction, but due to the limitations of the network structure, they are difficult to effectively model long-distance anatomical dependencies between multiple organs in the abdomen; Transformer-based models (such as TransUNet) achieve global contextual information capture through self-attention mechanisms, but they suffer from the defect that computational complexity increases quadratically with image size, resulting in excessively high costs for practical applications; Visual state space models (such as VM-UNet) that have emerged in recent years model long-distance dependencies with linear complexity, which is suitable for the needs of medical image segmentation, but this architecture still has significant shortcomings—skip connections are set after downsampling operations, and there is a lack of targeted local detail enhancement mechanisms.

[0004] In existing technologies, the skip connection design of VM-UNet leads to significant semantic gaps between the corresponding layer feature maps of the encoder and decoder, which does not match the semantic consistency requirements of the Add fusion operation, resulting in an imbalance in feature fusion. Simultaneously, its weak ability to represent local structural details makes the contours prone to breakage and blurring during single-organ segmentation, and inaccurate boundary delineation of adjacent organs during multi-organ joint segmentation, resulting in mutual encroachment and severely affecting segmentation accuracy. These problems directly limit the application effect of state-space models in abdominal multi-organ segmentation and cannot meet the actual clinical needs for accurate segmentation. Therefore, an optimized segmentation scheme is urgently needed to solve these problems. Summary of the Invention

[0005] To address the problems existing in the background art, one aspect of the present invention provides a method for segmenting abdominal multi-organ medical images based on a state-space model, comprising:

[0006] S1: Obtain abdominal CT images, divide the abdominal CT images into multiple non-overlapping image blocks through a patch embedding layer, and map each image block to a high-dimensional embedding vector to construct the initial feature representation of the abdominal CT images;

[0007] S2: Input the initial feature representation of the abdominal CT image into the encoder based on the VMamba architecture, and extract multiple levels of features from large to small scale at each stage of the encoder;

[0008] S3: Flatten the features at multiple levels and concatenate them into a feature sequence; input the feature sequence into the Mamba-based information fusion module to obtain the enhanced feature sequence, and reshape the enhanced feature sequence into multiple enhanced hierarchical features with scales from large to small;

[0009] S4: Input the features encoded by the encoder into the decoder based on the VMamba architecture, and gradually fuse multiple enhanced hierarchical features in the decoder to obtain reconstructed features; map the reconstructed features through a classification layer to obtain the segmentation result.

[0010] Another aspect of the present invention provides a medical image segmentation system for abdominal multi-organs based on a state-space model. The system includes a memory and a processor. The memory is used to store an application program. The processor is used to run the application program and execute the aforementioned medical image segmentation method for abdominal multi-organs based on a state-space model.

[0011] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for segmenting abdominal multi-organ medical images based on a state-space model.

[0012] The present invention has at least the following beneficial effects

[0013] This invention addresses the challenges of existing medical image segmentation techniques, such as the difficulty of CNNs in modeling long-distance dependencies, the high computational complexity of Transformers, and the large semantic discrepancies in skip connections, insufficient fidelity of single organ contours, and mutual encroachment between multiple organ segments in VM-UNet. By placing skip connections before the downsampling operation, the semantic differences in the encoder-decoder feature maps are effectively reduced, matching the semantic consistency requirements of the Add fusion operation. Utilizing a Mamba-based information fusion module, global anatomical structural dependencies are efficiently captured through multi-head Mamba units. This is combined with a convolutional gated feedforward network to enhance local detail feature representation. Simultaneously, the VMamba architecture's encoder-decoder achieves linear complexity feature extraction and reconstruction. Ultimately, this significantly improves the accuracy of abdominal multi-organ segmentation, resolves issues of incomplete organ contours and mutual encroachment between organs, and reduces the number of model parameters and computational complexity, providing more accurate and efficient technical support for clinical diagnosis and surgical planning. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0015] Figure 2 This is a schematic diagram of the method flow framework of the present invention;

[0016] Figure 3 This is a schematic diagram of the encoder and decoder based on the VMamba architecture of this invention;

[0017] Figure 4 This is a schematic diagram of the information fusion module based on Mamba of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 One aspect of the present invention provides a method for segmenting abdominal multi-organ medical images based on a state-space model, comprising:

[0020] S1: Obtain abdominal CT images, divide the abdominal CT images into multiple non-overlapping image blocks through a patch embedding layer, and map each image block to a high-dimensional embedding vector to construct the initial feature representation of the abdominal CT images;

[0021] Preferably, the initial feature representation of the abdominal CT image is constructed as follows: the acquired abdominal CT image has a dimension of H×W×3; the patch embedding layer uses convolution to segment the abdominal CT image into non-overlapping k×k image blocks; each image block is mapped to a feature dimension C, thus constructing the initial feature representation of the abdominal CT image. , where H and W represent the height and width of the abdominal CT image, and k represents the kernel size.

[0022] In this embodiment, the Synapse dataset was used for training. The Synapse dataset is a publicly available abdominal CT dataset designed for automatic multi-organ segmentation tasks, comprising 30 clinical abdominal CT cases. Each CT volume contains 85 to 198 slices of 512×512 pixels, totaling 3,779 axial slices. The dataset includes annotations for eight abdominal organs: aorta, gallbladder, spleen, left kidney, right kidney, liver, pancreas, and stomach. The images in the Synapse dataset were first uniformly scaled to a resolution of 224×224. To mitigate the risk of overfitting, data augmentation strategies, including random flipping and random rotation, were introduced. Subsequently, a patch embedding layer was used to transform the features of the preprocessed abdominal CT image. This layer employed a convolutional kernel of size k=4 to segment the 224×224×3 dimensional CT image into 56×56 non-overlapping 4×4 image patches. Each image patch was then mapped to a high-dimensional embedding vector of dimension C=96 through convolution operations, ultimately constructing an initial feature representation of the abdominal CT image with a size of 56×56×96. .

[0023] S2: Input the initial feature representation of the abdominal CT image into the encoder based on the VMamba architecture, and extract multiple levels of features from large to small scale at each stage of the encoder;

[0024] Preferably, the encoder based on the VMamba architecture includes four cascaded feature extraction stages; wherein the first, second, and third feature extraction stages each include two cascaded VSS modules and one downsampling module; the fourth feature extraction stage includes one VSS module; and the input features of the first feature extraction stage are used as hierarchical features. The input features from the downsampling modules in the first, second, and third feature extraction stages are used as hierarchical features. , and The output features of the fourth feature extraction stage are used as the features encoded by the encoder.

[0025] In this embodiment, the initial feature representation is first... The input is a VMamba-based encoder (containing four cascaded feature extraction stages). The first to third feature extraction stages each consist of two cascaded VSS modules (Visual State Space modules, used to refine features and capture local structure and long-range dependencies) and a block merging downsampling module. The fourth feature extraction stage contains only one VSS module. Initial feature representation... First, the raw input features are directly used as hierarchical features in the first feature extraction stage. After the feature representation is gradually optimized by two VSS modules, the input block merging downsampling module achieves scale compression by merging adjacent 2×2 feature blocks (with a stride of 2 and keeping the number of channels C=96 unchanged). The input features of this downsampling module are denoted as hierarchical features. Features after the first stage of block merging and downsampling (dimension becomes) (The number of channels is doubled to 2C=192) and then fed into the second feature extraction stage. The features are first optimized by two VSS modules, and the optimized features are used as hierarchical features. The block merging downsampling module input to this stage compresses the scale again through 2×2 block merging (step size 2). (The number of channels increases to 4C=384); the features after the second stage of downsampling enter the third feature extraction stage. After being processed by two VSS modules, the processed features are used as hierarchical features. Input the corresponding block merging downsampling module, and through 2×2 block merging (step size 2), compress the height and width from 14×14 to 7×7, doubling the number of channels from 384 to 768. The downsampling output feature dimension is... Finally, the features downsampled in the third stage are fed into the fourth feature extraction stage. After further refinement by the VSS module, the output dimension is... The features are the features encoded by the encoder, thus completing the extraction of features at multiple levels from large to small scale (56×56×96→56×56×96→28×28×192→14×14×384).

[0026] S3: Flatten the features at multiple levels and concatenate them into a feature sequence; input the feature sequence into the Mamba-based information fusion module to obtain the enhanced feature sequence, and reshape the enhanced feature sequence into multiple enhanced hierarchical features with scales from large to small;

[0027] Preferably, step S3 includes: representing multiple hierarchical features in descending order of scale as follows: , , and Hierarchical features , , and Features were obtained after flattening. , , and ; Features , , and Along the channel dimension The feature sequence is obtained by concatenation. ; The feature sequence is processed through the Mamba-based information fusion module. The feature sequence is obtained by performing forward and backward scans respectively. and ; feature sequence and The corresponding multi-head Mamba modules are input separately for processing to obtain feature sequences. and ; feature sequence and The feature sequence is obtained by adding the features together. , feature sequence The initial enhanced hierarchical features are reshaped into multiple scales through reshaping, namely... , , and The scales of the multiple initial augmentation hierarchical features correspond one-to-one with the scales of the multiple hierarchical features. Each of the initial augmentation hierarchical features is processed through its corresponding convolutional gated feedforward network to obtain multiple augmentation hierarchical features with scales decreasing from large to small. , , and .

[0028] In this embodiment, the encoder based on the VMamba architecture extracts hierarchical features from large to small scale through four feature extraction stages. , , and First, flatten these four hierarchical features in descending order of scale. and After flattening, they are respectively and , and Flattened and Then, these four flattened features are concatenated along the channel dimension to obtain the feature sequence. ;Then the feature sequence Inputting the Mamba-based information fusion module, performing forward and backward scans respectively yields... and The two sequences are respectively input into a multi-head Mamba module containing h parallel Mamba modules (h is set to 4, so that the input features are evenly divided into 4 sub-spatial features, processed by each Mamba module and then concatenated to output the feature sequence). and The feature sequence is obtained by adding the two elements one by one. Then, reshape it. Reconstructed into initial enhanced hierarchical features that correspond one-to-one with the original hierarchical feature scale. , , and Finally, these four initial enhanced hierarchical features are input into the corresponding convolutional gated feedforward networks (after being processed sequentially by sub-layer normalization, linear layers, DW convolutional layers, GeLu activation functions, and linear layers, and then added element-wise with the input features), ultimately yielding enhanced hierarchical features of decreasing scale. , , and .

[0029] Preferably, the multi-head Mamba module includes h parallel Mamba modules. The input features of the multi-head Mamba module are evenly divided into h sub-spatial features. The h sub-spatial features are input into the h Mamba modules for processing to obtain h intermediate features. The h intermediate features are then concatenated to obtain the output feature sequence of the multi-head Mamba module.

[0030] This embodiment clarifies the core structure and working mechanism of the multi-head Mamba module, which consists of h parallel-deployed Mamba modules. During operation, the input feature sequence is first uniformly divided into h non-overlapping sub-space features according to the feature dimension, ensuring that each sub-space feature has the same dimension and completely retains the key information of the original features. Then, these h sub-space features are input into the corresponding h Mamba modules. Each Mamba module independently refines the received sub-space features through internal normalization, linear transformation, DW convolution, SSM module processing, and residual connections, capturing local details and long-range dependencies within the sub-space, and outputting h optimized intermediate features. Finally, these h intermediate features from different sub-spaces are spliced ​​and fused along the channel dimension, integrating the feature representation advantages of multiple sub-spaces to form an output feature sequence that combines global dependency capture and local detail characterization capabilities.

[0031] Preferably, the Mamba module includes: a first normalization layer, a first linear layer, a second linear layer, a DW convolutional layer, a first SiLU activation function, a second SiLU activation function, an SSM module, a second normalization module, and a third linear layer; the input features of the Mamba module are used as the input features of the first normalization layer; the output features of the first normalization layer are used as the input features of the first and second linear layers; the output features of the first linear layer are processed sequentially through the DW convolutional layer, the first SiLU activation function, the SSM module, and the second normalization module; the output features of the second linear layer are processed through the second SiLU activation function; the output features of the second normalization module and the output features of the second SiLU activation function are multiplied element-wise and used as the input features of the third linear layer; the output features of the third linear layer and the input features of the Mamba module are added element-wise and used as the output features of the Mamba module.

[0032] This embodiment clarifies the detailed structure and feature processing flow of the Mamba module. This module consists of a first normalization layer, a first linear layer, a second linear layer, a DW convolutional layer, a first SiLU activation function, a second SiLU activation function, an SSM module, a second normalization module, and a third linear layer. During operation, the input features are first fed into the first normalization layer for data standardization. The output features of the first normalization layer are simultaneously input into the first and second linear layers via two paths. The first path, after mapping by the first linear layer, sequentially passes through the DW convolutional layer (extracting local features) and the first SiLU activation function (extracting local features). The feature optimization is completed by the nonlinear input module, the SSM module (capturing long-range dependencies), and the second normalization module (quadratic standardization). The second path is mapped through the second linear layer and then subjected to a nonlinear transformation through the second SiLU activation function. Subsequently, the output features of the second normalization module and the output features of the second SiLU activation function are multiplied element-wise to obtain the fused features, which are then input into the third linear layer for dimensionality adjustment. Finally, the output features of the third linear layer are added element-wise to the original input features of the Mamba module (residual connection) to obtain the output features of the Mamba module that contain both local details and long-range dependency information.

[0033] Preferably, the convolutionally gated feedforward network includes: a first sub-layer normalization layer, a first sub-linear layer, a first sub-DW convolutional layer, a GeLu activation function, and a second sub-linear layer; the input features of the convolutionally gated feedforward network are processed sequentially through the first sub-layer normalization layer, the first sub-linear layer, the first sub-DW convolutional layer, the GeLu activation function, and the second sub-linear layer; the input features of the convolutionally gated feedforward network and the output features of the second sub-linear layer are summed element-wise to obtain the output features of the convolutionally gated feedforward network.

[0034] This embodiment clarifies the specific composition and feature processing logic of the convolutional gated feedforward network. The network consists of a first sub-layer normalization layer, a first sub-linear layer, a first sub-DW convolutional layer, a GeLu activation function, and a second sub-linear layer. Its core function is to enhance the local detail representation of the initial enhanced hierarchical features. During operation, the input initial enhanced hierarchical features are first fed into the first sub-layer normalization layer for data standardization, eliminating interference from feature distribution differences. Then, the first sub-linear layer completes the feature dimension adaptation mapping. Next, the first sub-DW convolutional layer (depthseparable convolution) accurately extracts local structural detail information. Finally, the GeLu activation function is used to introduce a non-linear transformation, enhancing the network's ability to handle complex features. The fitting ability is then improved by adjusting the feature dimensions in reverse through the second sub-linear layer to match the subsequent fusion requirements. Finally, the output features of the second sub-linear layer are added element-wise with the original input features of the network (residual connection), which effectively preserves the original feature information and alleviates the gradient vanishing problem. The final output is an enhanced feature with both local detail clarity and feature integrity, laying a high-quality foundation for the fusion with the decoded features in the decoder. In this embodiment, the descriptions of the first and second layers are only to distinguish different blocks. The descriptions of the first sub-layer normalization layer, the first sub-linear layer, and the first sub-DW convolutional layer are also only to distinguish the special processing done for different modules. In essence, they are still the layer normalization layer, linear layer, etc.

[0035] S4: Input the features encoded by the encoder into the decoder based on the VMamba architecture, and gradually fuse multiple enhanced hierarchical features in the decoder to obtain reconstructed features; map the reconstructed features through a classification layer to obtain the segmentation result.

[0036] Preferably, the VMamba-based decoder includes four cascaded feature decoding stages; wherein, the first feature decoding stage includes a VSS module; the second, third, and fourth feature decoding stages each include an upsampling module and two cascaded VSS modules; the input features of the first feature decoding stage are the features encoded by the encoder; the output features of the upsampling modules in the second, third, and fourth feature decoding stages are respectively represented as... , and The output features of the fourth feature decoding stage are represented as follows: Among them, features Enhanced hierarchical features of the same scale After element-wise addition, the features are used as input features for the first VSS module in the second feature decoding stage; features Enhanced hierarchical features of the same scale After element-wise addition, the features are used as input features for the first VSS module in the third feature decoding stage; features Enhanced hierarchical features of the same scale After element-wise addition, the features are used as input features for the first VSS module in the fourth feature decoding stage; features Enhanced hierarchical features of the same scale The elements are summed one by one and then used as the input features for the classification layer.

[0037] In this embodiment, the features encoded by the encoder (dimension 1) The input is a VMamba-based decoder, which consists of four cascaded feature decoding stages. The first feature decoding stage contains only one VSS module (for feature refinement and dependency capture), while the second, third, and fourth feature decoding stages each consist of one upsampling module and two cascaded VSS modules. The features encoded by the encoder are first fed into the first feature decoding stage, where they are optimized by the VSS module and used as the output of that stage. Then, they enter the second feature decoding stage, where the upsampling module (using interpolation or deconvolution) enlarges the feature height and width from 7×7 to 14×14 and halves the number of channels from 768 to 384, resulting in the feature... ;Will Enhanced hierarchical features with the same scale Element-wise addition is performed, and the fused features are input into two VSS modules at this stage for further optimization. Then, the features are passed to the third feature decoding stage, where the upsampling module enlarges the feature height and width from 14×14 to 28×28 and halves the number of channels to 192, resulting in... , and the corresponding enhanced hierarchical features After element-wise addition, the features are processed by two VSS modules; then, the features enter the fourth feature decoding stage, where the upsampling module enlarges the feature height and width from 28×28 to 56×56, and halves the number of channels to 96, resulting in... ,and After element-wise addition and optimization by two VSS modules, the output features are determined. Then and Element-wise addition yields a reconstructed feature that combines the advantages of decoding features and enhanced hierarchical features. This reconstructed feature is then input into a classification layer, which includes a linear layer and a Softmax activation function. The linear layer maps the dimensions of the reconstructed feature (converting the number of channels from 96 to the number of abdominal organ categories, such as the 8 organ categories in the Synapse dataset), and then restores it to a 224×224×8 dimension output consistent with the original abdominal CT image size through a linear transformation. The output value of each pixel on the 8 channels is activated by Softmax to obtain the probability that the pixel belongs to each organ category. Based on the probability that each pixel belongs to each organ, the corresponding organs are segmented from the abdominal CT image, ultimately obtaining the segmentation result of multiple abdominal organs.

[0038] Furthermore, in this embodiment, a composite loss function is used during model training (referring to all network structures mentioned in this document). The loss function consists of two parts: cross-entropy loss ( ) and Dice coefficient loss ( ), which is represented as:

[0039]

[0040] )

[0041]

[0042] in, and Represents the weight parameters. This represents the true label of pixel i in the sample; This represents the prediction result for pixel i in the sample; This indicates the number of pixels in the sample.

[0043] Another aspect of the present invention provides a medical image segmentation system for abdominal multi-organs based on a state-space model. The system includes a memory and a processor. The memory is used to store an application program. The processor is used to run the application program and execute the aforementioned medical image segmentation method for abdominal multi-organs based on a state-space model.

[0044] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for segmenting abdominal multi-organ medical images based on a state-space model.

[0045] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0046] In summary, this invention addresses the challenges of existing medical image segmentation techniques, such as the difficulty of CNNs in modeling long-distance dependencies, the high computational complexity of Transformers, and the large semantic discrepancies in skip connections, insufficient fidelity of single organ contours, and mutual encroachment between multiple organ segments in VM-UNet. By placing skip connections before the downsampling operation, the semantic differences in the encoder-decoder feature maps are effectively reduced, matching the semantic consistency requirements of the Add fusion operation. Utilizing a Mamba-based information fusion module, global anatomical structural dependencies are efficiently captured through multi-head Mamba units. This is combined with a convolutional gated feedforward network to enhance local detail feature representation. Simultaneously, the VMamba architecture's encoder-decoder achieves linear complexity feature extraction and reconstruction. Ultimately, this invention significantly improves the accuracy of abdominal multi-organ segmentation, resolves issues of incomplete organ contours and mutual encroachment between organs, while reducing the number of model parameters and computational complexity, providing more accurate and efficient technical support for clinical diagnosis and surgical planning.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for segmenting abdominal multi-organ medical images based on a state-space model, characterized in that, include: S1: Obtain abdominal CT images, divide the abdominal CT images into multiple non-overlapping image blocks through a patch embedding layer, and map each image block to a high-dimensional embedding vector to construct the initial feature representation of the abdominal CT images; S2: Input the initial feature representation of the abdominal CT image into the encoder based on the VMamba architecture, and extract multiple levels of features from large to small scale at each stage of the encoder; S3: Flatten the features at multiple levels and concatenate them into a feature sequence; input the feature sequence into the Mamba-based information fusion module to obtain the enhanced feature sequence, and reshape the enhanced feature sequence into multiple enhanced hierarchical features with scales from large to small; S4: Input the features encoded by the encoder into the decoder based on the VMamba architecture, and gradually fuse multiple enhanced hierarchical features in the decoder to obtain reconstructed features; The reconstructed features are mapped through a classification layer to obtain the segmentation result.

2. The method for segmenting abdominal multi-organ medical images based on a state-space model according to claim 1, characterized in that, The initial feature representation of the abdominal CT image is constructed as follows: the acquired abdominal CT image has dimensions of H×W×3; the patch embedding layer uses convolution to segment the abdominal CT image into non-overlapping k×k image patches; each image patch is mapped to a feature dimension C, thus constructing the initial feature representation of the abdominal CT image. , where H and W represent the height and width of the abdominal CT image, and k represents the kernel size.

3. The method for segmenting abdominal multi-organ medical images based on a state-space model according to claim 1, characterized in that, The encoder based on the VMamba architecture includes four cascaded feature extraction stages; the first, second, and third feature extraction stages each include two cascaded VSS modules and one downsampling module; the fourth feature extraction stage includes one VSS module; the input features from the first feature extraction stage are used as hierarchical features. The input features from the downsampling modules in the first, second, and third feature extraction stages are used as hierarchical features. , and The output features of the fourth feature extraction stage are used as the features encoded by the encoder.

4. The method for segmenting abdominal multi-organ medical images based on a state-space model according to claim 1, characterized in that, Step S3 includes: representing multiple hierarchical features in descending order of scale as follows: , , and ; Hierarchical features , , and Features were obtained after flattening. , , and ; Features , , and Along the channel dimension The feature sequence is obtained by concatenation. The feature sequence is processed using a Mamba-based information fusion module. The feature sequence is obtained by performing forward and backward scans respectively. and ; feature sequence and The corresponding multi-head Mamba modules are input separately for processing to obtain feature sequences. and ; feature sequence and The feature sequence is obtained by adding the features together. , feature sequence The initial enhanced hierarchical features are reshaped into multiple scales through reshaping, namely... , , and The scales of the multiple initial augmentation hierarchical features correspond one-to-one with the scales of the multiple hierarchical features. Each of the initial augmentation hierarchical features is processed through its corresponding convolutional gated feedforward network to obtain multiple augmentation hierarchical features with scales decreasing from large to small. , , and .

5. The method for segmenting abdominal multi-organ medical images based on a state-space model according to claim 4, characterized in that, The multi-head Mamba module includes h parallel Mamba modules. The input features of the multi-head Mamba module are evenly divided into h sub-spatial features. The h sub-spatial features are input into the h Mamba modules for processing to obtain h intermediate features. The h intermediate features are concatenated to obtain the output feature sequence of the multi-head Mamba module.

6. The method for segmenting abdominal multi-organ medical images based on a state-space model according to claim 5, characterized in that, The Mamba module includes: a first normalization layer, a first linear layer, a second linear layer, a DW convolutional layer, a first SiLU activation function, a second SiLU activation function, an SSM module, a second normalization module, and a third linear layer. The input features of the Mamba module serve as the input features of the first normalization layer. The output features of the first normalization layer serve as the input features of the first and second linear layers. The output features of the first linear layer are processed sequentially through the DW convolutional layer, the first SiLU activation function, the SSM module, and the second normalization module. The output features of the second linear layer are processed through the second SiLU activation function. The output features of the second normalization module and the output features of the second SiLU activation function are multiplied element-wise and then used as the input features of the third linear layer. The output features of the third linear layer and the input features of the Mamba module are added element-wise and then used as the output features of the Mamba module.

7. The method for segmenting abdominal multi-organ medical images based on a state-space model according to claim 4, characterized in that, The convolutionally gated feedforward network includes: a first sub-layer normalization layer, a first sub-linear layer, a first sub-DW convolutional layer, a GeLu activation function, and a second sub-linear layer. The input features of the convolutionally gated feedforward network are processed sequentially through the first sub-layer normalization layer, the first sub-linear layer, the first sub-DW convolutional layer, the GeLu activation function, and the second sub-linear layer. The input features of the convolutionally gated feedforward network and the output features of the second sub-linear layer are summed element-wise to obtain the output features of the convolutionally gated feedforward network.

8. The method for segmenting abdominal multi-organ medical images based on a state-space model according to claim 1, characterized in that, The VMamba-based decoder comprises four cascaded feature decoding stages. The first feature decoding stage includes a VSS module; the second, third, and fourth feature decoding stages each include an upsampling module and two cascaded VSS modules. The input features of the first feature decoding stage are the features encoded by the encoder. The output features of the upsampling modules in the second, third, and fourth feature decoding stages are respectively represented as... , and The output features of the fourth feature decoding stage are represented as follows: Among them, features Enhanced hierarchical features of the same scale After element-wise addition, the features are used as input features for the first VSS module in the second feature decoding stage; features Enhanced hierarchical features of the same scale After element-wise addition, the features are used as input features for the first VSS module in the third feature decoding stage; features Enhanced hierarchical features of the same scale After element-wise addition, the features are used as input features for the first VSS module in the fourth feature decoding stage; features Enhanced hierarchical features of the same scale The elements are summed one by one and then used as the input features for the classification layer.

9. A medical image segmentation system for multiple organs of the abdomen based on a state-space model, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the abdominal multi-organ medical image segmentation method based on a state-space model as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by the processor, implements a method for segmenting abdominal multi-organ medical images based on a state-space model as described in any one of claims 1 to 8.

Images