Brain tumor segmentation method and device based on multi-view feature dynamic interactive fusion

Abstract

The invention relates to a brain tumor segmentation method and device based on multi-view feature dynamic interactive fusion, and the method comprises the following steps: S1, collecting a multi-modal MRI brain tumor image, constructing a data set, carrying out the preprocessing, and highlighting the details and features of a tumor; s2, constructing a multi-dimensional visual angle dynamic interaction fusion module, and realizing multi-visual angle feature dynamic fusion through three orthogonal visual angles in combination with a local detail and global space relationship; s3, constructing a hierarchical gating adaptive enhancement fusion module, carrying out dynamic weight adjustment, and carrying out selective fusion on cross-scale local details and global semantic information; s4, network parameters are updated through a self-adaptive structure sensing joint loss function, and a network model is trained and adjusted to optimize segmentation performance; and S5, performing MRI brain tumor segmentation according to the trained network model. According to the method, through multi-view information cooperative interaction and enhancement of cross-scale feature fusion, the recognition capability of the brain tumor area is effectively improved, and accurate segmentation of the brain tumor is realized.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a brain tumor segmentation method and device based on dynamic interactive fusion of multi-view features. Background Technology

[0002] In the field of medical image analysis, MRI, as an imaging technique with high contrast and no radiation damage, is widely used in the detection, evaluation, and auxiliary diagnosis of brain tumors. Multimodal MRI images can provide multi-dimensional information such as the internal structure of the tumor, the extent of invasion, and changes in surrounding tissues from different imaging sequences, serving as an important data foundation for automated brain tumor segmentation. However, due to factors such as the complex morphology of tumors, blurred tissue boundaries, and significant differences in imaging characteristics between different modalities, existing segmentation methods still have significant limitations in feature expression and fusion.

[0003] Existing brain tumor segmentation methods still have shortcomings in spatial structure understanding and multimodal information utilization, making it difficult to fully extract and integrate tumor morphological features from different perspectives. Brain tumor tissue exhibits significant structural differences across different sections, and traditional convolutional methods often rely on fixed directions or sequential operations, failing to effectively capture complementary spatial relationships between different perspectives, resulting in insufficient overall structural representation. Tumor regions often contain details and macroscopic morphology with large scale spans; a fixed receptive field cannot simultaneously capture boundary characterization and overall structure, making the model insufficiently adaptable to irregular or significantly sized tumor regions. Common skip connections simply perform direct splicing or superposition, lacking the ability to select the importance of features, easily introducing background noise and weakening the structural reconstruction effect in the decoding stage. Furthermore, the imaging differences between multimodal brain tumors are significant; without dynamic adjustment of the contribution of different modalities, the fusion effect is often limited.

[0004] To address the aforementioned issues, this invention proposes a brain tumor segmentation approach that leverages multi-perspective, multi-scale, and multi-sequence information. By employing cross-perspective dynamic interaction and orthogonal multi-scale feature extraction, it enhances the ability to express structural differences and the correlation between local details and overall morphology. Furthermore, it utilizes dynamic sequence fusion analysis to strengthen the global spatial semantics between different sections. Simultaneously, it introduces a dynamic fusion concept with hierarchical relationships and adaptive adjustment capabilities, enabling selective enhancement of local and global features based on content. Combined with dynamic gating information fusion, it improves the stability of structural reconstruction. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a brain tumor segmentation method and apparatus based on dynamic interactive fusion of multi-view features, which can achieve accurate segmentation of tumor regions in brain tumor MRI images.

[0006] Specifically, the method includes the following steps: S1: Acquire multimodal MRI brain tumor images, construct a dataset and perform preprocessing to highlight tumor details and features; S2: Construct a multi-dimensional perspective dynamic interaction fusion module, which combines local details and global spatial relationships through three orthogonal perspectives to achieve dynamic fusion of multi-perspective features; S3: Construct a hierarchical gated adaptive enhancement fusion module to perform dynamic weight adjustment and selectively fuse cross-scale local details and global semantic information; S4: Update network parameters using an adaptive structure-aware joint loss function, train and tune the network model to optimize segmentation performance; S5: Perform MRI brain tumor segmentation based on the trained network model.

[0007] Preferably, S1 includes the following steps: S1.1: Acquire multimodal brain tumor MRI images and brain tumor label data; S1.2: Denoising and enhancement operations are performed on multimodal brain tumor MRI images to obtain preprocessed image data; S1.3: Divide the preprocessed brain tumor MRI image data into datasets.

[0008] Preferably, S2 includes the following steps: S2.1: Construct a multi-dimensional view dynamic interaction fusion module (MVDIA) to achieve local and global feature integration through dual-branch collaborative processing; S2.2: Multi-Scale Orthogonal Directional Integration (MSODI) is used to construct multi-scale convolutions along three orthogonal directions, and combined with channel detail extraction to obtain local spatial feature representations; S2.3: The Tri-perspective Sequential Dynamic Combination (TPSDC) module is used to perform dynamic sequence modeling and feature interaction on the three perspectives to obtain global semantic features; S2.4: Combining the multi-scale orthogonal direction integration module and the three-view sequence dynamic fusion module, the output features of the two branches are integrated to obtain the fused features; Preferably, S2.2 includes the following steps: S2.2.1: The input features are processed in multiple directions and at multiple scales using four convolutional branches. The depthwise separable convolution is used to extract fine-grained texture and boundary details, while the three orthogonal perceptual convolutions adjust the convolution kernels according to different orthogonal planes, focusing on local structural changes in three directions.

[0009]

[0010] In the above formula, This represents the input MRI image data. This represents data from one of the three perspectives: coronal, axial, and sagittal. DWConv This represents depthwise separable convolution. DAConv This represents orientation-aware convolution. This represents the output of a depthwise separable convolution. This represents the output of the orientation-aware convolution operation; S2.2.2: The outputs of the four convolutional branches are concatenated and linearly integrated to combine convolutions at multiple directional levels, resulting in enhanced local features:

[0011] In the above formula, Conv This represents the convolution operation. Concat This indicates a splicing operation. This represents the local spatial features output by the multi-scale orthogonal direction integration module; Preferably, S2.3 includes the following steps: S2.3.1: Through three-view sequence construction and bidirectional state processing, context capture is performed on sequences along three orthogonal directions to learn the long-range dependencies of tumor regions from different perspectives, and direction-sensitive global sequence features are generated. The most effective information flow direction is further selected through direction weight adjustment to improve the expressive power of the sequence.

[0012] In the above formula, This represents the adaptive fusion sequence features output in one of the axial, coronal, or sagittal planes. Reshape This means expanding the data into a sequence along orthogonal dimensions. FWMamba Indicates a forward Mamba operation. RVMamba This indicates a reverse Mamba operation. , Indicates adaptive weights; S2.3.2: Through attentional interaction, global structural information from three directions is complementary and fused to obtain tumor features with better semantic consistency and global spatial understanding.

[0013]

[0014] In the above formula, , , These represent the results of attention calculations performed on sequences in the coronal, axial, and sagittal directions, respectively. This represents the calculation result in one direction. This represents the global spatial features output by the three-view sequence dynamic fusion module. Q, K, V It is a tensor generated from input features through convolution operations. softmax This represents the activation function. This indicates point-by-point multiplication. This indicates addition point by point; Preferably, S2.4 includes the following steps: The output features of the two branches are concatenated along the channel dimension, and the concatenated features are compressed and linearly fused using convolution. The result is then input into an MLP to further extract cross-scale and cross-view fusion relationships. Finally, residual connections are performed using the original input.

[0015]

[0016]

[0017] In the above formula, This indicates the output of the concatenation operation. These represent the local and global features obtained through the two branches, respectively. This represents the initial input data. Indicates the characteristics after fusion. This represents the output characteristics of the multi-dimensional perspective dynamic interaction fusion module. LN Representation layer normalization, MLP This represents a multilayer perceptron.

[0018] Preferably, S3 includes the following steps: S3.1: Construct a hierarchical gated adaptive enhancement fusion module (HGAEF) to achieve cross-scale feature reconstruction; S3.2: Adaptively adjust and selectively fuse the jump features based on dynamically generated weights, and perform feature dimension concatenation with the upsampled input to fuse cross-scale information; S3.3: Perform multi-branch convolution operation on the concatenated fused features, combine gated fusion for feature selection, and then introduce the initial features through residual connection and further fuse them to obtain the enhanced feature output; Preferably, S3.2 includes the following steps: Local detail features and global structural features at different semantic levels in skip connections are processed. The importance of the two types of features is adaptively adjusted based on cross-scale information, thereby forming a fusion representation with hierarchical dependencies. This fusion result is then integrated with the decoder upsampled features to ensure that the input has a hierarchical structure that simultaneously covers shallow details and deep semantics.

[0019]

[0020]

[0021] In the above formula, This indicates that the weights are generated dynamically. These represent local details and global semantic features, respectively. This represents the hierarchical features after fusion. It is an upsampled input. Feature output after splicing and fusion AvgPool Indicates average pooling; Preferably, S3.3 includes the following steps: S3.3.1: Construct a dual-branch convolutional structure. Dynamically dilated convolution is used to extract cross-scale information with variable receptive fields, while regular convolution is used to generate corresponding gating signals. Combining the two branches enables gating enhancement selection, achieving enhancement of cross-scale features and noise suppression.

[0022]

[0023] In the above formula, d Indicates the continuous dynamic void ratio. As a void ratio regulating factor, The feature input after splicing and fusion, DDConv Represents dynamic dilated convolution. This is for gating and enhancing feature output; S3.3.2: Linear integration is performed through channel mapping convolution, and residual paths are used to supplement the information of the gating results. Then, the fusion is further completed through a multilayer perceptron module to obtain more expressive feature output.

[0024] In the above formula, To enhance feature output, The feature input after splicing and fusion, This is for gating and enhancing feature output.

[0025] Preferably, S4 includes the following steps: S4.1: Calculate the basic segmentation loss and ensure the stability of the model's classification of the tumor region by using pixel classification constraints that combine Dice and cross-entropy; S4.2: Calculate multi-directional shape loss to constrain the continuity and stability of the morphology and boundaries of the tumor region from different perspectives; S4.3: Construct an adaptive structure-aware joint loss function (ASAJ Loss) and train the model. Update the parameters based on backpropagation to improve the model's segmentation performance. Preferably, S4.1 includes the following steps: Using Dice and cross-entropy loss, the overlap between the predicted region and the actual tumor region, as well as the pixel-level classification error, are calculated:

[0026]

[0027]

[0028] In the above formula, , Let Dice loss and cross-entropy loss represent the losses respectively. Indicates the basic segmentation loss. Represents the set of image pixels. Represents pixels i The predicted probability of it being a tumor. Represents pixels i The true label; Preferably, S4.2 includes the following steps: The morphological continuity and boundary stability of the segmentation results in different spatial directions are constrained by multi-directional shape loss, and the prediction probability difference between adjacent pixels in each view direction is calculated respectively:

[0029] In the above formula, Indicates multi-directional shape loss. Indicates the predicted probability. This represents the predicted probability of adjacent pixels. Represents the set of image pixels. Indicates spatial direction d Up pixel i The neighborhood set, These correspond to three orthogonal directions respectively.| | express d The number of pixels in the directional neighborhood. This represents the directional weighting parameter used to balance the strength of shape constraints in different spatial directions; Preferably, S4.3 includes the following steps: By combining the basic segmentation loss and multi-directional shape loss, a synergistic optimization of segmentation accuracy and shape structure consistency is achieved. The Adaptive Structure-Aware Joint Loss (ASAJLoss) function is constructed as follows:

[0030] In the above formula, This represents the adaptive structure-aware joint loss function. This represents the equilibrium parameters.

[0031] Preferably, S5 includes the following steps: S5.1: Input the brain tumor MRI image into the trained segmentation network model and output a feature fusion map containing semantic information of the tumor region; S5.2: Process the feature fusion map to obtain a segmentation probability map of the brain tumor region; S5.3: Binarize the probability map to obtain the final brain tumor segmentation result.

[0032] The beneficial effects of this invention are as follows: This invention, through the design of a systematic feature extraction and fusion process, enables multimodal brain tumor MRI data to be progressively enhanced during segmentation. The multi-dimensional dynamic interaction module MVDIA acquires the 3D structural differences and spatial correlations between different perspectives. Specifically, the multi-scale orthogonal direction integration module MSODI combines channel feature processing and multi-scale convolution in three orthogonal perspective directions to extract different detailed local features. The three-view sequence dynamic fusion module TPSDC captures global semantics through long-sequence dynamic analysis and attention enhancement mechanisms. The hierarchical gated adaptive enhancement fusion module HGAEF, through dynamic weight adjustment, gated selective enhancement, and dynamic dilated convolution, enables adaptive fusion of local and global information at different levels in skip features, enhancing structural reconstruction capabilities and suppressing noise interference in image data. The adaptive structure-aware joint loss function ASAJ Loss, by calculating Dice loss and multi-directional shape consistency loss, jointly supervises the model from three levels: region overlap, pixel-level classification, and spatial structural shape consistency constraints from different perspectives, improving the accuracy of tumor boundary characterization, the ability to maintain internal structural continuity, and the overall segmentation accuracy in complex scenes. Attached Figure Description

[0033] To more clearly illustrate the technical methods in the embodiments of the present invention, the accompanying drawings used in the prior art and embodiments are incorporated in and constitute a part of this specification. The following drawings are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart illustrating a brain tumor segmentation method based on dynamic interactive fusion of multi-view features according to the present invention. Figure 2 This is a schematic diagram of the multi-dimensional perspective dynamic interactive fusion module structure of a brain tumor segmentation method based on multi-perspective feature dynamic interactive fusion according to the present invention. Figure 3 This is a schematic diagram of the hierarchical gated adaptive enhancement fusion module structure of a brain tumor segmentation method based on dynamic interactive fusion of multi-view features according to the present invention. Figure 4 This is a schematic diagram of the network model structure of a brain tumor segmentation method based on dynamic interactive fusion of multi-view features according to the present invention. Detailed Implementation Plan To make the objectives, features, and advantages of this invention more apparent and understandable, the technical methods of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0035] This invention provides a brain tumor segmentation method and apparatus based on multi-view feature dynamic interactive fusion. By fusing the spatial differences of multimodal brain tumor images under different views, a cross-view dynamic interaction is constructed to enhance the expression of three-dimensional structure. Multi-scale feature extraction in orthogonal directions is used to strengthen the association between local details and overall semantics. Furthermore, a hierarchical dynamic gating enhancement mechanism is combined to adaptively adjust the importance of local and global features, selectively enhance key regions, and achieve accurate segmentation of brain tumor regions.

[0036] A typical embodiment of the present invention is described below. Figure 1 , Figure 4 The method includes the following steps: S1: Acquire multimodal MRI brain tumor images, construct a dataset and perform preprocessing to highlight tumor details and features; S2: Construct a multi-dimensional perspective dynamic interaction fusion module, which combines local details and global spatial relationships through three orthogonal perspectives to achieve dynamic fusion of multi-perspective features; S3: Construct a hierarchical gated adaptive enhancement fusion module to perform dynamic weight adjustment and selectively fuse cross-scale local details and global semantic information; S4: Update network parameters using an adaptive structure-aware joint loss function, train and tune the network model to optimize segmentation performance; S5: Segment MRI brain tumors based on the trained network model; The following is a detailed description of each step.

[0037] Furthermore, as a preferred embodiment of this method, S1 includes the following steps: S1.1: Acquire multimodal brain tumor MRI images and brain tumor label data; Furthermore, S1.1 includes the following steps: S1.1.1: Acquire multimodal brain tumor MRI images and corresponding labeled data. The images include T1-weighted imaging, T1-weighted enhanced imaging, T2-weighted imaging and the liquid attenuated inversion recovery sequence FLAIR, ensuring that the acquired data can fully cover the morphological and structural features of the tumor tissue. S1.1.2: Perform preliminary screening of the data, remove abnormal samples, and retain valid MRI data and corresponding labels that can reflect the morphological characteristics of the tumor; S1.2: Denoising and enhancement operations are performed on multimodal brain tumor MRI images to obtain preprocessed image data; Furthermore, S1.2 includes the following steps: S1.2.1: Resample and unify the spatial resolution between different modalities, normalize the grayscale range of multimodal images, and reduce intensity deviations caused by different devices and scanning conditions; S1.2.2: Denoising algorithms such as nonlocal mean or wavelet transform are used to suppress noise in the image to improve the clarity of the brain tumor region. Image enhancement methods such as random flipping and local contrast enhancement are combined to highlight the tumor edge and internal structure, thereby improving the stability of subsequent feature extraction. S1.3: Divide the preprocessed brain tumor MRI image data into datasets; Furthermore, S1.3 includes the following steps: S1.3.1: Associate the preprocessed multimodal brain tumor MRI images with the corresponding labels to ensure that the data input matches the supervision information; S1.3.2: Divide the associated dataset into 80% training set, 10% validation set, and 10% test set.

[0038] Furthermore, refer to Figure 2 S2 includes the following steps: S2.1: Construct a multi-dimensional view dynamic interaction fusion module (MVDIA) to achieve local and global feature integration through dual-branch collaborative processing; S2.2: Multi-Scale Orthogonal Directional Integration (MSODI) is used to construct multi-scale convolutions along three orthogonal directions, and combined with channel detail extraction to obtain local spatial feature representations; Furthermore, S2.2 includes the following steps: S2.2.1: The input features are processed in multiple directions and at multiple scales based on four convolutional branches, whereby... Depthwise separable convolutions are used to extract fine-grained textures and boundary details. Three orthogonal direction-aware convolutions adjust the convolution kernels according to different orthogonal planes. , and Focus on local structural changes in different directions:

[0039]

[0040] In the above formula, This represents the input MRI image data. This represents data from one of the three perspectives: coronal, axial, and sagittal. DWConv This represents depthwise separable convolution. DAConv This represents orientation-aware convolution. This represents the output of a depthwise separable convolution. This represents the output of the orientation-aware convolution operation; S2.2.2: Concatenate the outputs of the four convolution branches. Linear convolutional integration combines the outputs of convolutions at multiple directions and levels to obtain enhanced local features:

[0041] In the above formula, Conv This represents the convolution operation. Concat This indicates a splicing operation. This represents the local spatial features output by the multi-scale orthogonal direction integration module; S2.3: The Tri-perspective Sequential Dynamic Combination (TPSDC) module is used to perform dynamic sequence modeling and feature interaction on the three perspectives to obtain global semantic features; Furthermore, S2.3 includes the following steps: S2.3.1: Through three-view sequence construction and bidirectional state processing, context capture is performed on sequences along three orthogonal directions to learn the long-range dependencies of tumor regions from different perspectives, and direction-sensitive global sequence features are generated. The most effective information flow direction is further selected through direction weight adjustment to improve the expressive power of the sequence.

[0042] In the above formula, This represents the adaptive fusion sequence features output in one of the axial, coronal, or sagittal planes. Reshape This means expanding the data into a sequence along orthogonal dimensions. FWMamba Indicates a forward Mamba operation. RVMamba This indicates a reverse Mamba operation. , Indicates adaptive weights; S2.3.2: Through attentional interaction, global structural information from three directions is complementary and fused to obtain tumor features with better semantic consistency and global spatial understanding.

[0043]

[0044] In the above formula, , , These represent the results of attention calculations performed on sequences in the coronal, axial, and sagittal directions, respectively. This represents the calculation result in one direction. This represents the global spatial features output by the three-view sequence dynamic fusion module. Q, K, V It is a tensor generated from input features through convolution operations. softmax This represents the activation function. This indicates point-by-point multiplication. This indicates addition point by point; S2.4: Combining the multi-scale orthogonal direction integration module and the three-view sequence dynamic fusion module, the output features of the two branches are integrated to obtain the fused features; Furthermore, S2.4 includes the following steps: The output features of the two branches are concatenated along the channel dimension and then... Convolution performs channel compression and linear fusion on the concatenated features, then inputs the normalized result into the MLP to further extract cross-scale and cross-view fusion relationships, and finally performs residual connections with the original input.

[0045]

[0046]

[0047] In the above formula, This indicates the output of the concatenation operation. These represent the local and global features obtained through the two branches, respectively. This represents the initial input data. Indicates the characteristics after fusion. This represents the output characteristics of the multi-dimensional perspective dynamic interaction fusion module. LN Representation layer normalization, MLP This represents a multilayer perceptron.

[0048] Furthermore, refer to Figure 3 S3 includes the following steps: S3.1: Construct a hierarchical gated adaptive enhancement fusion module (HGAEF) to achieve cross-scale feature reconstruction; S3.2: Adaptively adjust and selectively fuse the jump features based on dynamically generated weights, and perform feature dimension concatenation with the upsampled input to fuse cross-scale information; Furthermore, S3.2 includes the following steps: Local detail features and global structural features at different semantic levels in skip connections are processed, and dynamic weights are generated based on cross-scale information to adaptively adjust the importance of the two types of features, thereby forming a fusion representation with hierarchical dependencies. This fusion result is then integrated with the decoder upsampled features to ensure that the input has a hierarchical structure that simultaneously covers shallow details and deep semantics.

[0049]

[0050]

[0051] In the above formula, This indicates that the weights are generated dynamically. These represent local details and global semantic features, respectively. This represents the hierarchical features after fusion. It is an upsampled input. Feature output after splicing and fusion AvgPool Indicates average pooling; S3.3: Perform multi-branch convolution operation on the concatenated fused features, combine gated fusion for feature selection, and then introduce the initial features through residual connection and further fuse them to obtain the enhanced feature output; Furthermore, S3.3 includes the following steps: S3.3.1: Construct a dual-branch convolutional structure. Dynamically dilated convolution is used to extract cross-scale information with variable receptive fields. Convolution is used to generate corresponding gated signals. Combining two branches enables gated enhancement selection, achieving enhancement of cross-scale features and noise suppression:

[0052]

[0053] In the above formula, d Indicates the continuous dynamic void ratio. As a void ratio regulating factor, The feature input after splicing and fusion, DDConv Represents dynamic dilated convolution. This is for gating and enhancing feature output; S3.3.2: Through channel mapping Convolutional processing is performed linearly, and residual paths are used to supplement the gating results with information. Then, the fusion is further completed through a multilayer perceptron module to obtain more expressive feature outputs.

[0054] In the above formula, To enhance feature output, The feature input after splicing and fusion, This is for gating and enhancing feature output.

[0055] Furthermore, S4 includes the following steps: S4.1: Calculate the basic segmentation loss and ensure the stability of the model's classification of the tumor region by using pixel classification constraints that combine Dice and cross-entropy; Furthermore, S4.1 includes the following steps: Using Dice and cross-entropy loss, the overlap between the predicted region and the actual tumor region, as well as the pixel-level classification error, are calculated:

[0056]

[0057]

[0058] In the above formula, , Let Dice loss and cross-entropy loss represent the losses respectively. Indicates the basic segmentation loss. Represents the set of image pixels. Represents pixels i The predicted probability of it being a tumor. Represents pixels i The true label; S4.2: Calculate multi-directional shape loss to constrain the continuity and stability of the morphology and boundaries of the tumor region from different perspectives; Furthermore, S4.2 includes the following steps: The morphological continuity and boundary stability of the segmentation results in different spatial directions are constrained by multi-directional shape loss, and the prediction probability difference between adjacent pixels in each view direction is calculated respectively:

[0059] In the above formula, Indicates multi-directional shape loss. Indicates the predicted probability. This represents the predicted probability of adjacent pixels. Represents the set of image pixels. Indicates spatial direction d Up pixel i The neighborhood set, These correspond to three orthogonal directions respectively. | | express d The number of pixels in the directional neighborhood. This represents the directional weighting parameter used to balance the strength of shape constraints in different spatial directions; S4.3: Construct an adaptive structure-aware joint loss function (ASAJ Loss) and train the model. Update the parameters based on backpropagation to improve the model's segmentation performance. Furthermore, S4.3 includes the following steps: By combining the basic segmentation loss and multi-directional shape loss, a synergistic optimization of segmentation accuracy and shape structure consistency is achieved. The Adaptive Structure-Aware Joint Loss (ASAJLoss) function is constructed as follows:

[0060] In the above formula, This represents the adaptive structure-aware joint loss function. This represents the equilibrium parameters.

[0061] Furthermore, S5 includes the following steps: S5.1: Input the brain tumor MRI image into the trained segmentation network model and output a feature fusion map containing semantic information of the tumor region; S5.2: Process the feature fusion map to obtain a segmentation probability map of the brain tumor region; S5.3: Binarize the probability map to obtain the final brain tumor segmentation result.

[0062] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make other equivalent modifications or substitutions without departing from the spirit of the invention. These equivalent modifications or substitutions are included within the scope defined by the claims.

Claims

1. A brain tumor segmentation method based on dynamic interactive fusion of multi-view features, characterized in that, Includes the following steps: S1: Acquire multimodal MRI brain tumor images, construct a dataset and perform preprocessing to highlight tumor details and features; S1 includes the following steps: S1.1: Acquire multimodal brain tumor MRI images and brain tumor label data; S1.2: Denoising and enhancement operations are performed on multimodal brain tumor MRI images to obtain preprocessed image data; S1.3: Divide the preprocessed brain tumor MRI image data into datasets; S2: Construct a multi-dimensional perspective dynamic interaction fusion module, which combines local details and global spatial relationships through three orthogonal perspectives to achieve dynamic fusion of multi-perspective features; S2 includes the following steps: S2.1: Construct a multi-dimensional view dynamic interaction fusion module (MVDIA) to achieve local and global feature integration through dual-branch collaborative processing; S2.2: Multi-Scale Orthogonal Directional Integration (MSODI) is used to construct multi-scale convolutions along three orthogonal directions, and combined with channel detail extraction to obtain local spatial feature representations; S2.3: The Tri-perspective Sequential Dynamic Combination (TPSDC) module is used to perform dynamic sequence modeling and feature interaction on the three perspectives to obtain global semantic features; S2.4: Combining the multi-scale orthogonal direction integration module and the three-view sequence dynamic fusion module, the output features of the two branches are integrated to obtain the fused features; S3: Construct a hierarchical gated adaptive enhancement fusion module to perform dynamic weight adjustment and selectively fuse cross-scale local details and global semantic information; S3 includes the following steps: S3.1: Construct a hierarchical gated adaptive enhancement fusion module (HGAEF) to achieve cross-scale feature reconstruction; S3.2: Adaptively adjust and selectively fuse the jump features based on dynamically generated weights, and perform feature dimension concatenation with the upsampled input to fuse cross-scale information; S3.3: Perform multi-branch convolution operation on the concatenated fused features, combine gated fusion for feature selection, and then introduce the initial features through residual connection and further fuse them to obtain the enhanced feature output; S4: Update network parameters using an adaptive structure-aware joint loss function, train and tune the network model to optimize segmentation performance; S4 includes the following steps: S4.1: Calculate the basic segmentation loss and ensure the stability of the model's classification of the tumor region by using pixel classification constraints that combine Dice and cross-entropy; S4.2: Calculate multi-directional shape loss to constrain the continuity and stability of the morphology and boundaries of the tumor region from different perspectives; S4.3: Construct an adaptive structure-aware joint loss function (ASAJ Loss) and train the model. Update the parameters based on backpropagation to improve the model's segmentation performance. S5: Perform MRI brain tumor segmentation based on the trained network model.

2. The brain tumor segmentation method based on dynamic interactive fusion of multi-view features according to claim 1, characterized in that, S2.2 includes the following steps: S2.2.1: The input features are processed in multiple directions and at multiple scales using four convolutional branches. The depthwise separable convolution is used to extract fine-grained texture and boundary details, while the three orthogonal perceptual convolutions adjust the convolution kernels according to different orthogonal planes, focusing on local structural changes in three directions. In the above formula, This represents the input MRI image data. This represents data from one of the three perspectives: coronal, axial, and sagittal. DWConv This represents depthwise separable convolution. DAConv This represents orientation-aware convolution. This represents the output of a depthwise separable convolution. This represents the output of the orientation-aware convolution operation; S2.2.2: The outputs of the four convolutional branches are concatenated and linearly integrated to combine convolutions at multiple directional levels, resulting in enhanced local features: In the above formula, Conv This represents the convolution operation. Concat This indicates a splicing operation. This represents the local spatial features output by the multi-scale orthogonal direction integration module; S2.3 includes the following steps: S2.3.1: Through three-view sequence construction and bidirectional state processing, context capture is performed on sequences along three orthogonal directions to learn the long-range dependencies of tumor regions from different perspectives, and direction-sensitive global sequence features are generated. The most effective information flow direction is further selected through direction weight adjustment to improve the expressive power of the sequence. In the above formula, This represents the adaptive fusion sequence features output in one of the axial, coronal, or sagittal planes. Reshape This means expanding the data into a sequence along orthogonal dimensions. FWMamba Indicates a forward Mamba operation. RVMamba This indicates a reverse Mamba operation. , Indicates adaptive weights; S2.3.2: Through attentional interaction, global structural information from three directions is complementary and fused to obtain tumor features with better semantic consistency and global spatial understanding. In the above formula, , , These represent the results of attention calculations performed on sequences in the coronal, axial, and sagittal directions, respectively. This represents the calculation result in one direction. This represents the global spatial features output by the three-view sequence dynamic fusion module. Q, K, V It is a tensor generated from input features through convolution operations. softmax This represents the activation function. This indicates point-by-point multiplication. This indicates addition point by point; S2.4 includes the following steps: The output features of the two branches are concatenated along the channel dimension, and the concatenated features are compressed and linearly fused using convolution. The result is then input into an MLP to further extract cross-scale and cross-view fusion relationships. Finally, residual connections are performed using the original input. In the above formula, This indicates the output of the concatenation operation. These represent the local and global features obtained through the two branches, respectively. This represents the initial input data. Indicates the characteristics after fusion. This represents the output characteristics of the multi-dimensional perspective dynamic interaction fusion module. LN Representation layer normalization, MLP This represents a multilayer perceptron.

3. The brain tumor segmentation method based on dynamic interactive fusion of multi-view features according to claim 1, characterized in that, S3.2 includes the following steps: Local detail features and global structural features at different semantic levels in skip connections are processed. The importance of the two types of features is adaptively adjusted based on cross-scale information, thereby forming a fused representation with hierarchical dependencies. This fusion result is then integrated with the decoder upsampled features to ensure that the input has a hierarchical structure that simultaneously covers shallow details and deep semantics. In the above formula, This indicates that the weights are generated dynamically. These represent local details and global semantic features, respectively. This represents the hierarchical features after fusion. It is an upsampled input. Feature output after splicing and fusion AvgPool Indicates average pooling; S3.3 includes the following steps: S3.3.1: Construct a dual-branch convolutional structure. Dynamically dilated convolution is used to extract cross-scale information with variable receptive fields, and regular convolution is used to generate corresponding gating signals. Combining the two branches enables gating enhancement selection, achieving enhancement of cross-scale features and noise suppression. In the above formula, d Indicates the continuous dynamic void ratio. As a void ratio regulating factor, The feature input after splicing and fusion, DDConv Represents dynamic dilated convolution. This is for gating and enhancing feature output; S3.3.2: Linear integration is performed through channel mapping convolution, and residual paths are used to supplement the information of the gating results. Then, the fusion is further completed through a multilayer perceptron module to obtain more expressive feature output. In the above formula, To enhance feature output, The feature input after splicing and fusion, This is for gating and enhancing feature output.

4. The brain tumor segmentation method based on dynamic interactive fusion of multi-view features according to claim 1, characterized in that, S4.1 includes the following steps: Using Dice and cross-entropy loss, the overlap between the predicted region and the actual tumor region, as well as the pixel-level classification error, are calculated: In the above formula, , Let Dice loss and cross-entropy loss represent the losses respectively. Indicates the basic segmentation loss. Represents the set of image pixels. Represents pixels i The predicted probability of it being a tumor. Represents pixels i The true label; S4.2 includes the following steps: The morphological continuity and boundary stability of the segmentation results in different spatial directions are constrained by multi-directional shape loss, and the prediction probability difference between adjacent pixels in each view direction is calculated respectively: In the above formula, Indicates multi-directional shape loss. Indicates the predicted probability. This represents the predicted probability of adjacent pixels. Represents the set of image pixels. Indicates spatial direction d Up pixel i The neighborhood set, These correspond to three orthogonal directions respectively. | | express d The number of pixels in the directional neighborhood. This represents the directional weighting parameter used to balance the strength of shape constraints in different spatial directions; S4.3 includes the following steps: By combining the basic segmentation loss and multi-directional shape loss, a synergistic optimization of segmentation accuracy and shape structure consistency is achieved. The Adaptive Structure-Aware Joint Loss (ASAJLoss) function is constructed as follows: In the above formula, This represents the adaptive structure-aware joint loss function. This represents the equilibrium parameter.

5. The brain tumor segmentation method based on dynamic interactive fusion of multi-view features according to claim 1, characterized in that, S5 include: S5.1: Input the brain tumor MRI image into the trained segmentation network model and output a feature fusion map containing semantic information of the tumor region; S5.2: Process the feature fusion map to obtain a segmentation probability map of the brain tumor region; S5.3: Binarize the probability map to obtain the final brain tumor segmentation result.

6. A brain tumor segmentation device based on dynamic interactive fusion of multi-view features, characterized in that, include: Preprocessing module: used to acquire multimodal MRI brain tumor image data and corresponding labels, construct datasets and perform a series of preprocessing steps such as denoising, enhancement, normalization and segmentation on the data, and convert them into the input format of the model; The MVDIA module is used to construct multi-angle local and global dynamic feature interactions, focusing on the key structures and semantic relationships of brain tumor regions from multiple perspectives; the multi-scale orthogonal direction integration module MSODI uses channel-independent detail feature extraction and orthogonal direction structural analysis to extract local information from the data; the three-view sequence dynamic fusion TPSDC expands the features along three perspectives and combines bidirectional information processing and attention interaction to obtain global semantics; the collaborative fusion of the two branches generates a multi-view feature representation that combines detail sensitivity and global semantic expression, realizing comprehensive processing of complex brain tumor regions; The HGAEF module is used to construct a dynamic gated fusion of hierarchical skip features and decoded features, focusing on selective enhancement of cross-scale detailed structure and global semantics. It can dynamically control the information flow according to the feature distribution, so that skip connections retain both boundary details and high-level semantics, further enhancing the model's cross-scale consistency and reconstruction stability. The adaptive structure-aware joint loss module is used to update model parameters and improve model segmentation performance. It combines pixel-level segmentation loss with shape consistency loss from different viewpoints to comprehensively constrain the classification accuracy, structural stability, and overall shape rationality of the predicted region. At the same time, it restricts tumor boundary mutations, avoids local structural breaks in the segmentation results, and improves overall stability, thereby enhancing the segmentation reliability of the network from multiple levels.

Images