Image segmentation method based on boundary structure decoupling optimization of double-branch complementary attention

Abstract

The present application relates to the technical field of computer vision and image processing, and discloses a boundary structure decoupling optimization image segmentation method based on double-branch complementary attention, which establishes a multi-organ image segmentation system based on Mamba-U type architecture, and internally has an image acquisition module, a data preprocessing module, a double-branch encoder module, a double-branch complementary attention module, a boundary perception auxiliary learning module, a mask generation decoder module, an anatomical structure coherence reinforcement module and a result output module, realizes efficient complementary fusion of global semantics and local boundary features through a channel attention and space-frequency attention collaborative enhancement mechanism, improves the perception ability of the model to organ regions and small structures, and through training-pushing decoupling design, strengthens boundary feature learning in the training stage, corrects topological errors in the inference stage, realizes boundary refinement and anatomical structure rationality of the segmentation result, and finally achieves the beneficial effects of collaborative optimization of geometric precision and topological coherence.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to an image segmentation method based on boundary structure decoupling optimization using bi-branch complementary attention. Background Technology

[0002] Image segmentation is one of the core tasks in computer vision and image processing, widely used in medical image analysis, industrial defect detection, and remote sensing image interpretation. Existing segmentation methods are mainly based on encoder-decoder architectures or introduce attention mechanisms to enhance feature extraction and context modeling capabilities. However, these methods generally suffer from the following shortcomings: a lack of adaptive selection mechanisms for the importance of different feature channels and spatial regions, making it difficult to accurately focus on the target region while suppressing background redundancy; insufficient sensitivity to high-frequency boundary information, leading to blurred segmentation boundaries and loss of fine structures; and a lack of effective constraints on the topological and structural coherence of the prediction results, easily resulting in isolated misclassified regions or structurally discontinuous predictions. Recently, while efficient Mamba-based models (such as EM-Net) have made significant progress in global context modeling and computational efficiency, their limitations in feature selection lead to insufficient focusing of target organs, a lack of boundary detail perception resulting in blurred organ boundaries, and a failure to introduce structural connectivity constraints during training, thus easily leading to topological errors. How to integrate multidimensional attention mechanisms into an efficient U-shaped network to enhance feature representation, and simultaneously improve the geometric accuracy and anatomical coherence of segmentation through boundary supervision and topological constraints, has become a key problem that urgently needs to be solved in current multi-organ segmentation tasks. Summary of the Invention

[0003] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an image segmentation method based on dual-branch complementary attention with decoupled boundary structure. This method has the advantages of accurate feature representation, clear boundary details, coherent topological structure, and consistency with training and inference objectives. It solves the problems of insufficient adaptive feature selection, weak perception of high-frequency boundary details, and lack of topological constraints in prediction results in existing Mamba-based segmentation models.

[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: an image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention, comprising the following steps: Step 1: Establish a multi-organ image segmentation system based on the U-net architecture. The system has built-in image acquisition module, data preprocessing module, dual-branch encoder module, dual-branch complementary attention module, boundary-aware assisted learning module, decoder module, anatomical structure coherence enhancement module, and result output module. Step 2: The image acquisition module is responsible for collecting multimodal raw image data of multiple human organs, as well as the corresponding pixel-level gold standard annotation, organ boundary annotation, and anatomical topology annotation data. Step 3: The data preprocessing module receives the raw images and labeled data output by the image acquisition module and performs standardized preprocessing. Step 4: The dual-branch encoder module receives the preprocessed standardized image data and uses a dual-branch parallel structure to perform multi-scale downsampling encoding to achieve multi-scale feature map extraction. Step 5: The dual-branch complementary attention module is connected to the multi-scale feature map output by the dual-branch encoder, and branch-specific attention enhancement is implemented. The Mamba branch is enhanced through channel attention, and the spectral branch is enhanced through spatial attention. The enhanced feature map is output through cross-branch complementary feature fusion. Step Six: The decoder module adopts a U-net decoding structure symmetrical to the dual-branch encoder. The enhanced feature map output by the dual-branch complementary attention module is restored to the original image size through deconvolution and upsampling operations. At the same time, feature information of the corresponding scale in the encoding stage is fused to achieve deep fusion of global and local features. In the training stage, jump to Step Seven; in the inference stage, jump to Step Eight. Step 7: During the model training phase, the boundary-aware auxiliary learning module receives the feature map output by the decoder, combines it with the organ boundary contour information in the gold standard annotation, constructs the boundary-aware loss function, and trains the boundary discrimination ability of the model feature layer. At the same time, the segmentation head also receives the feature map output by the decoder, processes it, and outputs the segmentation mask; jump to step 9. Step 8: In the model inference stage, the segmentation head also receives the feature map output by the decoder and processes it. Then, it inputs the initial segmentation mask into the anatomical structure coherence enhancement module. Based on the prior rules of human multi-organ anatomy, it builds a lightweight topological constraint and correction strategy and outputs the final segmentation mask. Step 9: The result output module receives the final segmentation mask output by the mask generation decoder module, completes the tasks of visualization, data archiving and quantitative evaluation, and achieves the goals of intuitive presentation of segmentation results, traceable and iterative data, and continuous optimization of model performance.

[0005] Preferably, the image acquisition module has built-in a diversity acquisition unit, a boundary blur acquisition unit, a complex topology acquisition unit, and a cross-center acquisition unit; The data preprocessing module preprocesses the following steps: S1.1 First, remove image noise, artifacts and redundant interference information through Gaussian filtering and nonlocal mean denoising algorithm; S1.2 Then, perform pixel value normalization and grayscale value standardization on the images of different modalities to unify the data distribution range; S1.3 Continue to carry out data augmentation operations such as random flipping, rotation, scaling, and elastic deformation to expand the training sample size; S1.4 Finally, ensure the annotation data format is unified, registration and alignment are performed, and output high-quality preprocessed data.

[0006] Preferably, the diversity acquisition unit is responsible for acquiring original image data of multiple human organs under different modalities and different devices, as well as corresponding pixel-level gold standard annotation data; the boundary blur acquisition unit is responsible for acquiring images of difficult organs with blurred organ boundaries, low tissue contrast, and unclear edge contours, and focuses on capturing image data of difficult-to-segment targets such as the adhesion edges of adjacent organs, small organs, and thin-walled organs.

[0007] Preferably, the complex topology acquisition unit acquires image data of organs with complex physiological adhesions, pathological deformations, and anatomical structural variations; the cross-center acquisition unit collects standardized desensitized image data from multiple medical centers of different regions, levels, and equipment configurations.

[0008] Preferably, the dual-branch encoder module employs a dual-branch parallel structure to perform multi-scale downsampling encoding, enabling simultaneous extraction of global semantics and local details from the preprocessed standardized image data. Specifically: (1) A method that relies on Mamba blocks to efficiently capture long-range global semantic dependencies and extract deep global organ features; (2) Another branch extracts shallow local detail features through the spectrum module, taking into account both global semantic information and local detail information, and outputs multi-scale and multi-level encoded feature maps.

[0009] Preferably, the dual-branch complementary attention module implements branch-specific attention enhancement, specifically as follows: (1) The Mamba branch embedding channel attention mechanism focuses on information-rich feature channels, suppresses redundant channel interference, and strengthens global feature expression and long-range dependency modeling. (2) The spectral branch fusion spatial attention and frequency domain filtering operation accurately extracts high-frequency boundary information, refines local spatial positioning, and enhances the ability to perceive organ edges.

[0010] Preferably, the dual-branch complementary attention module integrates dual-branch features using a weighted fusion formula, outputting an enhanced feature map that combines global semantics and local details. The dual-branch feature fusion calculation formula is as follows: In the formula, This represents the fused enhanced features output by the dual-branch complementary attention module. This represents the global features output by the channel attention branch. denoted by , represents the local boundary features of the spatial-frequency domain attention branch output, and µ represents the feature fusion weight coefficient.

[0011] Preferably, the boundary-aware assisted learning module is enabled during the training phase and incorporates an auxiliary boundary loss, with the total loss function... From the main segmentation loss and auxiliary boundary loss The weighted composition is calculated using the following formula: In the formula, and These are the principal segmentation prediction and the ground truth, respectively. To obtain from morphological gradient operators The derived boundary truth labels use Dice Loss as the boundary loss function. , The balance coefficient is minimized. The resulting gradients are backpropagated to the shared encoder-decoder backbone, thereby correcting the underlying feature representations and making them more sensitive to high-frequency boundaries while focusing on semantic regions.

[0012] Preferably, the mask generation decoder module generates a pixel-level initial segmentation mask through activation function mapping, and its calculation formula is as follows: In the formula, Represents image coordinates Pixel segmentation probability at that location, This represents the feature map value output by the decoder.

[0013] Preferably, the result output module receives the final segmentation mask output by the mask generation decoder module and performs the following tasks: (1) The segmentation mask is superimposed on the original image for visualization, and the regions and boundaries of each organ are marked for easy clinical viewing; (2) Classify, store, and archive the original images, preprocessed data, initial segmentation results, and final segmentation results, and support data traceability and model iteration; (3) The segmentation accuracy, boundary quality and structural rationality are evaluated by using Dice similarity coefficient, IoU intersection-union ratio, boundary F1 score and topology consistency index. The evaluation results are fed back to the training stage to help the model parameters be continuously optimized.

[0014] Compared with existing technologies, this invention provides an image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention, which has the following beneficial effects: 1. This invention achieves efficient complementary fusion of global semantics and local boundary features through the channel attention and spatial-frequency domain attention collaborative enhancement mechanism of the dual-branch complementary attention module (DCAM), thereby significantly improving the model's ability to perceive organ regions and fine structures, and ultimately achieving the beneficial effects of accurate segmentation and clear boundary.

[0015] 2. This invention employs a decoupled training-inference design between the Boundary Awareness Assisted Learning Module (ABAS) and the Anatomical Structure Coherence Enhancement Module (ACES). This design enhances boundary feature learning during the training phase and corrects topological errors during the inference phase, thereby refining the boundary of the segmentation results and ensuring the rationality of the anatomical structure. This achieves the beneficial effect of synergistic optimization of geometric accuracy and topological coherence.

[0016] 3. This invention comprehensively ensures the generalization ability and robustness of the model by acquiring and preprocessing images in multiple modes, multiple scenarios and across centers, combined with quantitative evaluation of results and feedback loop, thus achieving the beneficial effect of stable application and continuous optimization of the model in diverse clinical scenarios. Attached Figure Description

[0017] Figure 1 This is a system flowchart 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 An image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention includes the following steps: Step 1: Establish a multi-organ image segmentation system based on the U-net architecture. The system has built-in image acquisition module, data preprocessing module, dual-branch encoder module, dual-branch complementary attention module (DCAM), boundary-aware assisted learning module (ABAS), mask generation decoder module, anatomical coherence enhancement module (ACES), and result output module. The system adopts a training-inference decoupled design. During the training phase, the boundary-aware assisted learning module (ABAS) is enabled for boundary supervision, and during the inference phase, the anatomical coherence enhancement module (ACES) is enabled for topology correction. Step 2: The image acquisition module is responsible for collecting multimodal, multi-scene, and cross-center raw image data of multiple human organs, as well as supporting pixel-level gold standard annotation, organ boundary annotation, and anatomical topology annotation data. After the collected dataset is divided into training set, validation set, and test set in a 7:2:1 ratio, it is uniformly transmitted to the data preprocessing module. The rich data dimensions ensure the generalization ability and robustness of the model. Step 3: The data preprocessing module receives the raw image and annotation data output by the image acquisition module and transmits them to the dual-branch encoder module; Step 4: The dual-branch encoder module receives the preprocessed standardized image data and, based on the Mamba-U architecture, uses a dual-branch parallel structure to perform multi-scale downsampling encoding, thereby achieving simultaneous extraction of global semantics and local details, providing a high-quality feature base for the subsequent attention enhancement module. Step 5: The Dual-Branch Complementary Attention Module (DCAM) interfaces with the multi-scale feature map output by the dual-branch encoder. Following the principles of representation specialization and complementary optimization, it implements branch-specific attention enhancement and outputs an enhanced feature map that combines global semantics and local details through cross-branch complementary feature fusion. This enhanced feature map is then simultaneously transmitted to the mask generation decoder module and the boundary awareness assisted learning module. Step Six: The decoder module adopts a U-net decoding structure symmetrical to the dual-branch encoder. The enhanced feature map output by the dual-branch complementary attention module is restored to the original image size through deconvolution and upsampling operations. At the same time, feature information of the corresponding scale in the encoding stage is fused to achieve deep fusion of global and local features. In the training stage, jump to Step Seven; in the inference stage, jump to Step Eight. Step 7: During the model training phase, the boundary-aware auxiliary learning module receives the feature map output by the decoder, combines it with the organ boundary contour information in the gold standard annotation, constructs a boundary-aware loss function, and trains the boundary discrimination ability of the model feature layer. At the same time, the segmentation head also receives the feature map output by the decoder, processes it, and outputs a segmentation mask; jump to step 9; Step 8: In the model inference stage, the segmentation head also receives the feature map output by the decoder and processes it. Then, it inputs the initial segmentation mask into the anatomical structure coherence enhancement module. Based on the prior rules of human multi-organ anatomy, it builds a lightweight topological constraint and correction strategy and outputs the final segmentation mask. Step 9: The results output module receives the final segmentation mask output by the mask generation decoder module, completes the tasks of visualization, data archiving, and quantitative evaluation, and achieves the goals of intuitive presentation of segmentation results, traceable and iterative data, and continuous optimization of model performance. The tasks include: First, visualizing the segmentation mask overlaid with the original image, clearly marking the regions and boundaries of each organ for easy clinical review; Second, classifying, storing, and archiving the original image, preprocessed data, initial segmentation results, and final segmentation results to support data traceability and model iteration; Third, quantitatively evaluating segmentation accuracy, boundary quality, and structural rationality through indicators such as Dice similarity coefficient, IoU (Intersection over Union), boundary F1 score, and topological consistency index, and feeding the evaluation results back to the training phase to help continuously optimize model parameters.

[0020] The image acquisition module incorporates a variety acquisition unit, a boundary blur acquisition unit, a complex topology acquisition unit, and a cross-center acquisition unit; Data preprocessing module preprocessing process: S1.1 First, remove image noise, artifacts and redundant interference information through Gaussian filtering and nonlocal mean denoising algorithm; S1.2 Then, perform pixel value normalization and grayscale value standardization on the images of different modalities to unify the data distribution range; S1.3 Continue to carry out data augmentation operations such as random flipping, rotation, scaling, and elastic deformation to expand the training sample size; S1.4 Finally, the annotation data format is unified, registration and alignment are completed, and high-quality preprocessed data is output and transmitted to the dual-branch encoder module.

[0021] The advantages are: through a four-step progressive processing of joint noise reduction by Gaussian filtering and nonlocal mean denoising, pixel-level normalization and grayscale standardization, data augmentation by random flipping, rotation, scaling and elastic deformation, and unified annotation format and registration alignment, noise interference is systematically eliminated, data distribution is unified, sample size is expanded and annotation information is aligned, so as to achieve stable output of high-quality standardized training data.

[0022] The diverse acquisition unit adapts to the interface protocols and data reading programs of different imaging devices, covering multiple clinically commonly used medical imaging modalities such as CT, MRI, and ultrasound, as well as various brands and models of medical imaging acquisition equipment. It acquires raw image data of multiple human organs under different modalities and devices, along with corresponding pixel-level gold standard labeled data, covering organs in multiple locations such as the abdomen, chest, and brain, while also taking into account images of normal and diseased organs. This comprehensively ensures the diversity of data sources, modalities, and scenarios, overcoming the problem of limited model generalization ability caused by single data. The boundary blur acquisition unit uses a combination of image feature filtering algorithms and manual verification to specifically acquire images of difficult organs with blurred organ boundaries, low tissue contrast, and unclear edge contours. It focuses on capturing image data of difficult-to-segment targets such as the adhesion edges of adjacent organs, small organs, and thin-walled organs, which are used for subsequent targeted training of the model. This enhances the model's ability to perceive and capture high-frequency boundary details and overcomes the technical defects of existing models such as blurred boundaries and loss of fine structures.

[0023] The complex topology acquisition unit collects image data of organs with physiological adhesions, pathological deformations, and complex anatomical variations through anatomical structure feature retrieval and clinical case screening mechanisms. It includes various complex scene samples such as organ displacement, overlap, and irregular shape, thereby expanding the model's complex structure learning samples, enhancing the model's ability to model the coherence and topological rationality of organ structures, and reducing problems such as isolated misjudged regions and topological disorder during segmentation. The cross-center acquisition unit collects standardized de-identified image data from multiple medical centers of different regions, levels, and equipment configurations through compliant data sharing interfaces and de-identification processing procedures. It takes into account the differences in imaging parameters, image quality, and population characteristics among different medical centers. By fusing multi-center data, it reduces the domain bias problem caused by single-center data and improves the adaptability and stability of the model in different clinical scenarios.

[0024] The advantages are: through the collaborative design of diverse acquisition units, boundary-blurred acquisition units, complex topology acquisition units, and cross-center acquisition units, it covers multi-modal imaging, difficult boundary scenes, complex anatomical structures, and cross-center offset, etc., and achieves a comprehensive improvement in the representativeness of training data and the generalization ability of the model.

[0025] The dual-branch encoder module receives preprocessed standardized image data and, based on a u-net architecture, employs a dual-branch parallel structure to perform multi-scale downsampling encoding, achieving simultaneous extraction of global semantics and local details: one branch relies on Mamba blocks to efficiently capture long-range global semantic dependencies and extract deep global organ features; the other branch extracts shallow local detail features through a spectrum module, taking into account both global semantic information and local detail information, and outputs multi-scale, multi-level encoded feature maps, providing a high-quality feature base for the subsequent attention enhancement module.

[0026] The advantages are: by using the parallel dual-branch architecture of long-range dependency modeling of Mamba blocks and local feature extraction of convolutional neural networks, deep global semantic information and shallow local detail information are captured simultaneously, achieving efficient collaborative extraction of global and local features and stable supply of high-quality feature base.

[0027] The dual-branch complementary attention module implements branch-specific attention enhancement, as detailed below: (1) The Mamba branch embedding channel attention mechanism focuses on information-rich feature channels, suppresses redundant channel interference, and strengthens global feature expression and long-range dependency modeling. (2) The spectral branch fusion spatial attention and frequency domain filtering operation accurately extracts high-frequency boundary information, refines local spatial positioning, and enhances the ability to perceive organ edges.

[0028] The advantages are: by focusing on information-enriching channels through channel attention mechanism, and by using a branch differentiation design that combines spatial attention and frequency domain filtering to extract high-frequency boundary information, the ability to express global features and perceive local edges is enhanced in a targeted manner, thereby achieving a dual improvement in adaptive selection of feature channels and accurate capture of boundary details.

[0029] The Dual-Branch Complementary Attention Module (DCAM) integrates dual-branch features through cross-branch complementary feature fusion using a weighted fusion formula, outputting an enhanced feature map that combines global semantics with local details. The dual-branch feature fusion calculation formula is as follows: In the formula, This represents the fused enhanced features output by the dual-branch complementary attention module. This represents the global features output by the channel attention branch. denoted by , represents the local boundary features of the spatial-frequency domain attention branch output, and µ represents the feature fusion weight coefficient, with a value ranging from 0.4 to 0.6, which is adaptively adjusted according to the organ segmentation task.

[0030] The advantage is that it outputs fused enhanced features through a dual-branch feature fusion calculation formula, using learned weight coefficients. Dynamically balancing global features With local features The contribution ratio is adaptively adapted to the feature requirements of different organ segmentation tasks, achieving an enhanced feature expression that combines global semantic consistency and local boundary accuracy.

[0031] The boundary-aware assisted learning module is enabled during the training phase and incorporates auxiliary boundary loss. The total loss function is... From the main segmentation loss and auxiliary boundary loss The weighted composition is calculated using the following formula: In the formula, and These are the principal segmentation prediction and the ground truth, respectively. To obtain from morphological gradient operators The derived boundary truth labels use Dice Loss as the boundary loss function. , The balance coefficient is minimized. The resulting gradients are backpropagated to the shared encoder-decoder backbone, thereby correcting the underlying feature representations and making them more sensitive to high-frequency boundaries while focusing on semantic regions.

[0032] The advantages are: by enabling the boundary-aware learning module and adding auxiliary boundary loss during the training phase, accurate segmentation is achieved; and by enabling the anatomical coherence enhancement module during the inference phase, anatomical consistency can be enhanced, further improving the segmentation results.

[0033] The mask generation decoder module generates a pixel-level initial segmentation mask through activation function mapping, and its calculation formula is as follows: In the formula, Represents image coordinates The pixel segmentation probability at a given location, with a value ranging from 0 to 1; This represents the feature map value output by the decoder; a probability value greater than 0.5 indicates the target organ region, while a probability value less than 0.5 indicates the background region, thus generating the initial segmentation mask.

[0034] The advantage is: through image coordinates Pixel segmentation probability at The decoder output is mapped to the probability space using the Sigmoid activation function; a foreground-background binary classification is achieved using a 0.5 threshold; multi-scale feature fusion ensures segmentation consistency, and finally, a pixel-level initial segmentation mask is stably generated, enabling accurate completion of the basic multi-organ segmentation task and reliable input supply for the subsequent topology correction module.

[0035] The result output module receives the final segmentation mask output by the mask generation and decoder module, and completes three core tasks: (1) The segmentation mask is superimposed on the original image for visualization, and the regions and boundaries of each organ are clearly marked, which is convenient for clinical viewing; (2) Classify, store, and archive the original images, preprocessed data, initial segmentation results, and final segmentation results to support data traceability and model iteration; (3) By using indicators such as Dice similarity coefficient, IoU intersection-union ratio, boundary F1 score, and topological consistency index, the segmentation accuracy, boundary quality and structural rationality are quantitatively evaluated, and the evaluation results are fed back to the training stage to help the model parameters be continuously optimized.

[0036] The advantages are: by overlaying and visualizing the segmentation mask with the original image, classifying and archiving the data throughout the process, and automatically evaluating and providing feedback on multi-dimensional quantitative indicators, it meets the clinical need for intuitive viewing, supports data traceability and model iteration, drives continuous optimization of model parameters, and achieves visualized presentation, systematic management and closed-loop optimization of segmentation results.

[0037] In summary, this invention achieves global-local feature adaptive enhancement through a dual-branch complementary attention mechanism, separates and optimizes boundary supervision and topology correction through a training-inference decoupling framework, ensures model generalization ability through multi-dimensional data acquisition and preprocessing, and balances accuracy and efficiency through lightweight module design. Finally, it forms a complete intelligent segmentation closed loop from data input to result output, achieving a comprehensive improvement in accuracy, boundary quality, structural rationality, and computational efficiency in multi-organ image segmentation, ultimately providing reliable technical support for clinical diagnosis and treatment planning.

[0038] 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. An image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention, characterized in that, Includes the following steps: Step 1: Establish a multi-organ image segmentation system based on the U-net architecture. The system has built-in image acquisition module, data preprocessing module, dual-branch encoder module, dual-branch complementary attention module, boundary-aware assisted learning module, decoder module, anatomical structure coherence enhancement module, and result output module. Step 2: The image acquisition module is responsible for collecting multimodal raw image data of multiple human organs, as well as the corresponding pixel-level gold standard annotation, organ boundary annotation, and anatomical topology annotation data. Step 3: The data preprocessing module receives the raw images and labeled data output by the image acquisition module and performs standardized preprocessing. Step 4: The dual-branch encoder module receives the preprocessed standardized image data and uses a dual-branch parallel structure to perform multi-scale downsampling encoding to achieve multi-scale feature map extraction. Step 5: The dual-branch complementary attention module is connected to the multi-scale feature map output by the dual-branch encoder, and branch-specific attention enhancement is implemented. The Mamba branch is enhanced through channel attention, and the spectral branch is enhanced through spatial attention. The enhanced feature map is output through cross-branch complementary feature fusion. Step Six: The decoder module adopts a U-net decoding structure symmetrical to the dual-branch encoder. The enhanced feature map output by the dual-branch complementary attention module is restored to the original image size through deconvolution and upsampling operations. At the same time, feature information of the corresponding scale in the encoding stage is fused to achieve deep fusion of global and local features. In the training stage, jump to Step Seven; in the inference stage, jump to Step Eight. Step 7: During the model training phase, the boundary awareness auxiliary learning module receives the feature map output by the decoder, combines it with the organ boundary contour information in the gold standard annotation, constructs the boundary awareness loss function, and trains the boundary discrimination ability of the model feature layer. At the same time, the segmentation head also receives the feature map output by the decoder, processes it, and outputs the segmentation mask. Jump to step nine; Step 8: In the model inference stage, the segmentation head also receives the feature map output by the decoder and processes it. Then, it inputs the initial segmentation mask into the anatomical structure coherence enhancement module. Based on the prior rules of human multi-organ anatomy, it builds a lightweight topological constraint and correction strategy and outputs the final segmentation mask. Step 9: The result output module receives the final segmentation mask output by the mask generation decoder module, completes the tasks of visualization, data archiving and quantitative evaluation, and achieves the goals of intuitive presentation of segmentation results, traceable and iterative data, and continuous optimization of model performance.

2. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 1, characterized in that, The image acquisition module incorporates a diversity acquisition unit, a boundary blur acquisition unit, a complex topology acquisition unit, and a cross-center acquisition unit. The data preprocessing module preprocesses the following steps: S1.1 First, remove image noise, artifacts and redundant interference information through Gaussian filtering and nonlocal mean denoising algorithm; S1.2 Then, perform pixel value normalization and grayscale value standardization on the images of different modalities to unify the data distribution range; S1.3 Continue to carry out data augmentation operations such as random flipping, rotation, scaling, and elastic deformation to expand the training sample size; S1.4 Finally, ensure the annotation data format is unified, registration and alignment are performed, and output high-quality preprocessed data.

3. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 2, characterized in that, The diversity acquisition unit is responsible for acquiring raw image data of multiple human organs under different modalities and devices, along with corresponding pixel-level gold standard annotation data. The boundary blur acquisition unit is responsible for acquiring images of difficult organs with blurred organ boundaries, low tissue contrast, and unclear edge contours, and focuses on capturing image data of difficult-to-segment targets such as the adhesion edges of adjacent organs, small organs, and thin-walled organs.

4. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 2, characterized in that, The complex topology acquisition unit collects image data of organs with complex physiological adhesions, pathological deformations, and anatomical structural variations; the cross-center acquisition unit collects standardized desensitized image data from multiple medical centers of different regions, levels, and equipment configurations.

5. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 1, characterized in that, The dual-branch encoder module employs a dual-branch parallel structure to perform multi-scale downsampling encoding, enabling simultaneous extraction of global semantics and local details from preprocessed standardized image data. Specifically: (1) A method that relies on Mamba blocks to efficiently capture long-range global semantic dependencies and extract deep global organ features; (2) Another branch extracts shallow local detail features through the spectrum module, taking into account both global semantic information and local detail information, and outputs multi-scale and multi-level encoded feature maps.

6. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 1, characterized in that, The dual-branch complementary attention module implements branch-specific attention enhancement, specifically as follows: (1) The Mamba branch embedding channel attention mechanism focuses on information-rich feature channels, suppresses redundant channel interference, and strengthens global feature expression and long-range dependency modeling. (2) The spectral branch fusion spatial attention and frequency domain filtering operation accurately extracts high-frequency boundary information, refines local spatial positioning, and enhances the ability to perceive organ edges.

7. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 6, characterized in that, The dual-branch complementary attention module integrates dual-branch features using a weighted fusion formula, outputting an enhanced feature map that combines global semantics and local details. The dual-branch feature fusion calculation formula is as follows: In the formula, This represents the fused enhanced features output by the dual-branch complementary attention module. This represents the global features output by the channel attention branch. denoted by , represents the local boundary features of the spatial-frequency domain attention branch output, and µ represents the feature fusion weight coefficient.

8. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 1, characterized in that, The boundary-aware assisted learning module is enabled during the training phase and incorporates auxiliary boundary loss, with the total loss function... From the main segmentation loss and auxiliary boundary loss The weighted composition is calculated using the following formula: In the formula, and These are the principal segmentation prediction and the ground truth, respectively. To obtain from morphological gradient operators The derived boundary truth labels use Dice Loss as the boundary loss function. , The balance coefficient is minimized. The resulting gradients are backpropagated to the shared encoder-decoder backbone, thereby correcting the underlying feature representations and making them more sensitive to high-frequency boundaries while focusing on semantic regions.

9. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 1, characterized in that, The mask generation decoder module generates a pixel-level initial segmentation mask through activation function mapping, and its calculation formula is as follows: In the formula, Represents image coordinates Pixel segmentation probability at that location, This represents the feature map value output by the decoder.

10. The image segmentation method based on boundary structure decoupling optimization using dual-branch complementary attention as described in claim 1, characterized in that, The result output module receives the final segmentation mask output by the mask generation decoder module and performs the following tasks: (1) The segmentation mask is superimposed on the original image for visualization, and the regions and boundaries of each organ are marked for easy clinical viewing; (2) Classify, store, and archive the original images, preprocessed data, initial segmentation results, and final segmentation results, and support data traceability and model iteration; (3) The segmentation accuracy, boundary quality and structural rationality are evaluated by using Dice similarity coefficient, IoU intersection-union ratio, boundary F1 score and topological consistency index. The evaluation results are fed back to the training stage to help the model parameters be continuously optimized.

Images