Two-dimensional multi-view brain tumor medical image segmentation method and system based on Vision Mama time sequence model

Abstract

The invention relates to a two-dimensional multi-view brain tumor medical image segmentation method and a two-dimensional multi-view brain tumor medical image segmentation system based on a Vision Mama time sequence model, which utilize a novel visual representation model to complete focus segmentation of a brain tumor two-dimensional medical image. The method is used for solving the problems that when an existing two-dimensional segmentation method is used for processing two-dimensional brain medical images, space depth information cannot be fully utilized, and three-dimensional structure features are difficult to accurately capture. The method comprises the following steps: 1, acquiring and preprocessing data; 2, constructing a multi-view brain tumor segmentation network based on edge feature fusion and a spatial state model; 3, constructing a combined loss function of weighted cross entropy and weighted Dess loss, and meanwhile, storing an optimal model weight in training for prediction; and 4, predicting a brain tumor medical image by using the trained optimal model, calculating evaluation indexes and performing result comparison. Through the combination of the methods, the boundary feature extraction quality of the fuzzy edge of the complex focus and the two-dimensional segmentation precision of the model on the brain tumor are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a two-dimensional multi-view brain tumor medical image segmentation method and system based on the Vision Mamba temporal model. Background Technology

[0002] Brain tumors are a common and heterogeneous type of intracranial tumor, with a high incidence and extremely high mortality rate. Brain tumors cause increased intracranial pressure, and the tumor tissue compresses surrounding brain tissue, leading to severe central nervous system damage. Early detection before the tumor becomes malignant can significantly improve patient survival and reduce treatment costs. In the past, the main treatment options for brain tumors were radiotherapy and surgical resection. Both radiotherapy and surgery required neurosurgeons to analyze the tumor's nature based on brain lesion images and manually draw the radiotherapy or surgical area. With advancements in medical imaging technology, the number of medical image slices obtained from a single MRI scan has increased dramatically. This manual drawing process is not only time-consuming but also susceptible to errors due to subjective factors.

[0003] With the rapid development of modern artificial intelligence technology, tumor diagnosis and preoperative intervention assessment methods are constantly moving towards intelligence. Traditional segmentation methods that rely on manually designed features are significantly insufficient in dealing with the complexity of lesion structures. They not only depend on visually perceptible image features but also place high demands on the professional experience and technical skills of physicians. Although semi-automatic segmentation technology introduces some human intervention, this human operation is often subjective and prone to judgment bias or even errors.

[0004] In recent years, computer vision technology has been increasingly used in medical image segmentation to improve diagnostic efficiency, simplify treatment processes, optimize disease monitoring, and help surgeons perform tumor resections more accurately. While traditional Convolutional Neural Networks (CNNs) have made significant progress in brain tumor medical image segmentation, their method of modeling global information by stacking local receptive fields layer by layer suffers from insufficient long-range dependency capture, making it difficult to achieve accurate segmentation results in regions with blurred boundaries or complex structures.

[0005] The Vision Mamba architecture relies on a state-space model, enabling efficient global modeling with linear complexity. This makes it more suitable for fine-grained boundary delineation of brain tumor MRI images, thereby improving segmentation accuracy and robustness. However, during MRI imaging, due to limitations in the imaging mechanism, T2 and FLAIR modal 3D images typically contain only a limited number of axial slices, resulting in significantly lower resolution in the coronal and sagittal planes compared to the axial plane, leading to resolution anisotropy. While 3D imaging can obtain complete volumetric data, the acquisition time is approximately three times that of 2D imaging, limiting its clinical application. To address the shortcomings of existing 2D segmentation methods that rely solely on a single axial slice and struggle to utilize the contextual information of adjacent slices, this invention proposes a multi-view segmentation method that integrates information from three orthogonal planes (axial, coronal, and sagittal) of medical images. This method constructs 2D segmentation sub-networks based on the Vision Mamba architecture in each of the three directions to extract global features and fine-grained boundary information. In the output stage, weighted fusion and attention-based fusion are used to fuse the segmentation results from the three directions, ultimately generating high-precision 3D segmentation results.

[0006] Through the above technical solution, this invention can fully utilize the complementary information from multiple perspectives of medical images and the global modeling capabilities of Vision Mamba while ensuring computational efficiency. This effectively alleviates the challenges brought about by resolution anisotropy, greatly promotes the automation process of medical image analysis, and improves the segmentation accuracy of complex heterogeneous tumors and fuzzy boundary regions. It has good clinical application value and promotion prospects. Summary of the Invention

[0007] To address the technical problems of existing two-dimensional segmentation methods that rely solely on axial slices, lack contextual information from adjacent slices, and have insufficient segmentation accuracy in regions with blurred boundaries and complex structures, this invention proposes a two-dimensional multi-view brain tumor medical image segmentation method and system based on the Vision Mamba temporal model, aiming to improve the boundary awareness and global context modeling capabilities of medical image segmentation.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model, comprising the following steps:

[0009] (1) Obtain brain tumor medical image dataset and preprocess it. Slice the three-dimensional image data in the dataset into three orthogonal planes in the axial, coronal and sagittal directions to obtain three sets of two-dimensional slice image data from different perspectives.

[0010] (2) Construct three independent two-dimensional brain tumor segmentation sub-network models based on edge feature fusion and visual spatial state equations respectively, and process slice images in the axial, coronal and sagittal directions respectively; then fuse the segmentation results output by the three sub-network models to obtain the final output; the two-dimensional brain tumor segmentation sub-network model based on edge feature fusion and visual spatial state equations includes: a backbone network based on temporalization and visual spatial state equations, an edge feature extraction network based on convolution and a bottleneck layer based on a bidirectional edge-semantic feature fusion module;

[0011] (3) Construct a combined loss function of weighted cross-entropy and weighted Descein loss. Use the three sets of two-dimensional slice image data from different perspectives obtained in step (1) as training datasets to train the two-dimensional brain tumor segmentation sub-network model based on edge feature fusion and visual space state equation. During the training process, use mini-batch gradient descent combined with adaptive optimization algorithm to update parameters. In each iteration, backpropagate and gradient correction are performed on the network weights to gradually approach the optimal solution. Save the optimal weight parameters based on the minimum loss of the validation set during the training process. After training, use the model for subsequent prediction and evaluation. After the training phase ends, load the saved optimal weight model, perform segmentation prediction on the test set, and compare the prediction results with the real labels. Use multiple evaluation indicators to evaluate the performance to verify the accuracy, robustness and edge detail prediction ability of the model in medical image segmentation tasks.

[0012] Furthermore, when performing multi-view fusion for two-dimensional brain tumor image segmentation, three independent two-dimensional brain tumor segmentation sub-network models based on edge feature fusion and visual space state equations are used in the axial, coronal, and sagittal directions to process the images in each orthogonal plane. Then, the prediction outputs of these three networks are integrated through a fusion strategy to generate the final segmentation result. The fusion strategy includes mid-term attention mechanism fusion, cross-view multi-head attention mechanism fusion, and terminal voxel fusion to maximize the use of complementary information from the three perspectives.

[0013] Furthermore, the fusion strategy is as follows: first, alignment and enhancement are performed at the representation layer, then cross-view interaction is performed, and finally, robust convergence is achieved using logits attention.

[0014] 2D perspective Where Ax represents the axial direction, Co represents the coronal direction, Sa represents the sagittal direction, and the image height is... , width is The label mask is Number of channels Batch processing size Features after alignment of the same voxel ,in As a scale;

[0015] The intermediate attention mechanism is fused as follows: in By integrating channel and spatial attention mechanisms at the scale, a cross-perspective fusion representation is obtained. ,include:

[0016] 1) Channel attention: First, perform global average pooling on the features of each viewpoint. Then, a small multilayer sensor is used to obtain the channel gating value. ,in These are the parameters for a small multilayer sensor. For the Sigmoid function, For the SiLU function; finally, Expand back to the original space size and scale each channel individually. ;

[0017] 2) Spatial weighting: The position of each pixel in each viewpoint is... The weighted features of the three channels are then spliced ​​together in the channel dimension and fed into a... Pixel multilayer perceptron Output the prediction probability scores for the three viewpoints and normalize them. deal with:

[0018]

[0019] Among them, the probability scores of the three perspectives ,and ;

[0020] 3) Blending Features and Projection: First, combine pixels one by one to obtain the blending features. ;

[0021] To preserve the residual information from each viewpoint, Features from three perspectives To splice, use projection Reverting to the original standard number of channels yields the fused backbone characteristics: ;

[0022] The cross-view multi-head attention fusion is as follows: In At scale, a single-layer multi-head cross-view attention mechanism is used to unfold the planar image features. : After unfolding, there are a total of Each token; for each perspective Using its own features as the sequence Query, it interacts with the feature sequences Key and Value from the other two paths respectively, resulting in the following linear mapping:

[0023] , ,

[0024] in, From the perspective Feature sequence Query , For the other two paths, the feature sequences Key and Value are... Here are the parameter matrices for the three attention sequences; In addition to The stitching of two additional perspectives outside the main viewpoint; let the number of heads be h, and the dimension of a single head be... The sum of attention outputs for the i-th head is obtained from the above linear mapping: Then, the output results of each head are summed and concatenated. and normalize After processing the residuals, we obtain: ;in The parameter matrix for residual processing;

[0025] Then connect to the feedforward network layer. And add residual processing: Then Transform back to the size of the original feature map Finally, the updated features from the three perspectives are gated and fused. A cross-perspective fusion representation was obtained. ,in These are the backbone features derived from mid-term attention-based fusion. Hyperparameters for gated fusion;

[0026] The final voxel fusion is defined as: the prediction score of the corresponding feature of the same voxel in each direction. After splicing, it passes through a small multilayer sensor. Obtain attention weights in each direction. :

[0027] ,

[0028] in, It is an exponential function with the natural logarithm e as its base. This is the result of late-stage voxel fusion.

[0029] Furthermore, in the backbone network based on temporalization and visual space state equations, after temporalization improves the network, the model supports a set of consecutive slices of one case as input at a time, and can learn the segmentation information of each consecutive slice and share features; the encoder of the backbone network consists of the visual state space module (VSS Block) and block merging in Vision Mamba, where the visual state space module is used to extract low-level image detail features and high-level semantic information; the block merging operation is to divide the feature map into several small blocks and then perform downsampling operation; the decoder of the backbone network consists of the visual state space module and block expansion, where block expansion is the inverse operation of block merging, responsible for enlarging the spatial resolution of the feature map to gradually restore it to the size of the input image for reconstructing segmentation details;

[0030] The temporal implementation method is as follows: a two-dimensional image is divided into image blocks. Postscript ; Divide the image into blocks at each scale s The corresponding slice image features are obtained after passing through the encoder VSS Block. , is represented as: Image features at each scale Subsequently, a temporal state-space model was introduced, and a one-dimensional selective scan was performed on the slice at each spatial location. Flatten the spatial grid Output ;in It applies the VSS Block to the dimension T of the continuous slice.

[0031] Furthermore, the backbone network based on temporalization and the visual space state equation, the Vision Mamba Visual State Space Block (VSS Block), relies on the 2D selective scanning SS2D module. The visual state space block contains two branches: a local branch and a state space model (SSM) branch. After the first image segmentation, the feature map size is... After block merging, the number of channels ,and , Features after the s-th encoding stage: ; denote the continuous slice of input The local branch structure is: linear projection layer. +Normalization ;

[0032] The SSM branch uses a selective state-space model to perform a linear-time scan on the flattened one-dimensional vector, using horizontal and vertical scans to adapt to the two-dimensional model; the SSM branch structure is: linear projection layer. +Multi-scale depth-separable convolutional layers +Edge enhancement layer +SiLU function + SS2D+ Normalization The formulas are as follows:

[0033] , ,

[0034] , ,

[0035] Then, gated fusion and linear mixing of the two branches are performed, along with the residuals, using the following formula:

[0036] , ,

[0037] Then connect to the feedforward network FFN: .

[0038] Furthermore, the structure of the SS2D module is: Scan Expanding + S6 unit + Scan Merging; firstly, the two-dimensional features are expanded into four sequences along the four diagonal directions. Where d represents the directions of the four diagonals. , TL represents the top left, BR the bottom right, TR the top right, and BL the bottom left;

[0039] The S6 element is the core element in the Vision Mamba model based on the discretized spatial state equation; for the continuous spatial state equation, we have: , ;

[0040] in, , , These are, respectively, the state value reflecting the current system, the value representing external influences introduced into the system, and the value we hope to observe from the system. Let A represent the current system output state value; where A, B, C, and D are the state matrix, control matrix, output matrix, and command matrix, respectively; the continuous state equation is discretized as follows:

[0041] ,

[0042] in, The state matrix is ​​a discrete-time state matrix. It is a continuous-time state matrix. For discrete time steps, The value of the input control matrix at step t;

[0043] Selective parameters , , Input-driven: , , ;in For learnable parameter matrix, The input features at step t , For bias terms, Soft positive function , , Represented as a learnable mapping function, This is represented as converting a vector into a diagonal matrix; a one-way recursive approach is used: , For each direction d, the final output is: ;

[0044] Scan Merging restores the output from the four directions to two dimensions and aggregates them to obtain four tensors. Then, by using weighted attention summation, the four tensors are aggregated and restored to their original size: , .

[0045] Furthermore, an edge extraction network based on convolution and spatial attention mechanisms is constructed. The edge feature extraction network consists of a two-dimensional convolution module, a linear projection, and a feature mapping layer, which is used to achieve multi-level feature extraction. It extracts edge features from the shallow layer of the image to form an edge feature map, and finally outputs the edge feature map to the edge-semantic feature fusion module.

[0046] Furthermore, the bidirectional edge-semantic feature fusion module inputs the edge feature map and the segmentation image feature map into the module, fuses the two feature maps, enhances the model's recognition of blurred boundary features, and realizes boundary-guided segmentation and segmentation correction boundaries; the implementation method of the bidirectional edge-semantic feature fusion module is as follows:

[0047] (1) Receive the boundary feature map output by the edge feature extraction network and the semantic segmentation feature map output by the backbone network encoder respectively, and denote the boundary feature map output by the edge extraction network as... Semantic segmentation features output by the backbone network encoder After passing through the same sampling layer Align to the same size r is the variation ratio, and it combines max pooling and average pooling. After suppressing noise, we obtain the aligned boundary feature map and segmentation feature map:

[0048] , ,

[0049] (2) Input the aligned two-way features into the interactive attention layer, and set up three mutually coupled attention enhancement branches in the interactive attention layer to complete deep feature selection and fusion; including:

[0050] ① Channel Attention Branch: Based on compression and activation attention mechanisms, the upsampled features of the encoder are channel-weighted to learn channel-level importance coefficients. Effective segmentation channel features are learned through adaptive reinforcement.

[0051] , ,

[0052] , ,

[0053] in, Features of the boundary passage Semantic channel features, For global average pooling, For the Sigomid function, The SiLU activation function is used. These are the parameters for a small multilayer sensor. The boundary features are those after channel attention weighting. Semantic segmentation features weighted by channel attention Represented as the Hadema product of element-wise multiplication;

[0054] ② Spatial attention branch: utilizing max pooling With average pooling Generate and stitch together spatial attention maps , Enhance the network's spatial awareness of key lesion areas and boundary locations:

[0055] ,

[0056] ,

[0057] in Calculate for k convolutional layers, Boundary features after spatial attention weighting, These are semantic segmentation features weighted by spatial attention;

[0058] ③ Bilinear Modeling Branch: Perform bilinear pooling on channel attention and spatial attention features to model higher-order semantic relationships, and fuse it with the outputs of the first two branches; in the same spatial location Take the outer product of the two channel vectors And compressed to The fused feature vector is obtained from the dimension. ,in Indicates the first Boundary features at each location, No. Semantic segmentation features at each location; then, the fused feature vectors from all locations are collected and arranged. For the new feature map ;

[0059] (3) The features fused through the interactive attention layer are then processed... convolution Implement gate inter-gate control to achieve bidirectional guidance of boundary features and semantic segmentation features:

[0060] ,

[0061] in These are semantic segmentation features after bidirectional guidance. These are the boundary features after bidirectional guidance;

[0062] Finally, the above three branches and the bidirectional interaction results are concatenated at the channel dimension and residual processing is performed. Its expression is and introduce To alleviate overfitting:

[0063] ,

[0064] Where F is the concatenated feature map and Y is the final output of the bidirectional edge-semantic feature fusion module; the output fused feature map Y is then used by the decoder to reconstruct the high-resolution lesion segmentation result; in the skip connection path, bidirectional feature fusion replaces the traditional method of only concatenating, and the spatial attention map generated by max pooling and average pooling is used to emphasize key regions and suppress redundant background, thereby improving boundary consistency and segmentation accuracy.

[0065] Furthermore, the method for constructing the combined loss function of weighted cross-entropy and weighted Dessian loss is as follows:

[0066] A weighted approach is used for the loss, determining the weight of each category of pixels in the loss function based on their proportion in the segmentation image. Specifically, the total loss is a combination of the weighted cross-entropy (WCE) loss function and the weighted Dessell (WD) loss function. (WCE loss) WD loss and combined loss function They are represented as follows:

[0067]

[0068]

[0069]

[0070] in, Indicates input data, This represents the corresponding actual segmentation label. Represents network parameters, This represents the weight of class C in the WCE loss. Represents the true segmentation mask. This represents the predicted segmentation probability mask. This represents the balancing weight of class c in the WD loss, set to... .

[0071] The present invention also provides a two-dimensional multi-view brain tumor medical image segmentation system based on the Vision Mamba time series model, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it can implement the above-mentioned method.

[0072] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a two-dimensional multi-view brain tumor medical image segmentation method and system based on the VisionMamba temporal model. It can not only be used for academic experimental verification (comparing with the results of existing segmentation models) to verify the effectiveness of the method of the present invention, but also be deployed in subsequent clinical auxiliary diagnosis, lesion identification and other application scenarios, with good scalability and practical value. Attached Figure Description

[0073] Figure 1 The flowchart illustrates a two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model, provided in an embodiment of the present invention.

[0074] Figure 2 This is a structural diagram of the two-dimensional brain tumor image segmentation method with multi-view fusion in an embodiment of the present invention.

[0075] Figure 3This is a network structure diagram of a two-dimensional brain tumor medical image segmentation model based on temporal Vision Mamba and edge feature extraction in an embodiment of the present invention.

[0076] Figure 4 This is a structural diagram of the Visual State Space Block (VSS Block) in an embodiment of the present invention.

[0077] Figure 5 This is a structural diagram of the bidirectional edge-semantic feature fusion module in an embodiment of the present invention.

[0078] Figure 6 This is a dataset and brain tumor segmentation result diagram of the method of the present invention in an embodiment of the present invention. Detailed Implementation

[0079] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0080] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0081] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0082] like Figure 1 As shown in the figure, this embodiment provides a two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model, and its implementation steps are as follows.

[0083] (1) Obtain brain tumor medical image dataset and preprocess it. Slice the three-dimensional image data in the dataset into three orthogonal planes in the axial, coronal and sagittal directions to obtain three sets of two-dimensional slice image data from different perspectives.

[0084] (2) Construct three independent two-dimensional brain tumor segmentation sub-network models based on edge feature fusion and Vision State Space (VSS) to process slice images in the axial, coronal and sagittal directions respectively; then fuse the segmentation results output by the three sub-network models to obtain the final output; the two-dimensional brain tumor segmentation sub-network model based on edge feature fusion and Vision State Space includes: a backbone network based on temporalization and Vision State Space, an edge feature extraction network based on convolution and a bottleneck layer based on a bidirectional edge-semantic feature fusion module.

[0085] (3) Construct a combined loss function of weighted cross-entropy and weighted Dice loss. Use the three sets of two-dimensional slice image data from different perspectives obtained in step (1) as training datasets to train the two-dimensional brain tumor segmentation sub-network model based on edge feature fusion and visual space state equation. During training, use mini-batch gradient descent combined with adaptive optimization algorithm (Adam optimizer) to update parameters. In each iteration, backpropagate and gradient correction are performed on the network weights to gradually approach the optimal solution. Save the optimal weight parameters based on the minimum loss of the validation set during training. After training, use the model for subsequent prediction and evaluation. After the training phase, load the saved optimal weight model, perform segmentation prediction on the test set, and compare the prediction results with the real labels. Use various evaluation metrics (such as Dice coefficient, IoU, Hausdorff distance, etc.) to evaluate the performance and verify the accuracy, robustness and edge detail prediction ability of the model in medical image segmentation tasks.

[0086] The original dataset for brain tumors came from BraTS2021, a dataset provided in the International Multimodal Brain Tumor Segmentation Challenge. This dataset, provided by multiple international institutions, contains four modalities (T1, T1CE, T2, and FLAIR) of 3D medical images per sample. Each modality has a file size of 240×240×155. The segmentation labels were manually annotated by professional physicians, containing four subregions: background (0); necrosis and non-enhanced tumor (1); edema (2); and enhanced tumor (4). The four modalities of brain tumor MRI images from one patient in the original dataset were sliced. The 3D brain glioma medical image data were sliced ​​in three orthogonal planes (axial, coronal, and sagittal) to obtain three sets of 2D data. After slicing the data from the three modalities, the original single-channel 2D images of the four modalities were fused into a single 4-channel multimodal 2D image. Data augmentation included the following operations:

[0087] ① Random cropping: Randomly offset and crop the image within the range of [-10, 10] in the x and y directions to simulate changes in image position; ② Flip transformation: Perform horizontal or vertical flipping with a certain probability to improve the model's adaptability to changes in direction; ③ Gaussian noise perturbation: Add Gaussian noise with a mean of 0 and a variance randomly selected between [0.01, 0.09] to the image to simulate possible interference during the imaging process.

[0088] The augmented dataset containing brain tumor case images was divided into training, validation, and test sets according to a preset ratio (8:1:1). Normalization or standardization algorithms were used to standardize the intensity of the image data of each modality to reduce the numerical differences between different modalities.

[0089] Figure 2 This is a structural diagram of a multi-view fusion-based 2D brain tumor image segmentation method. To achieve multi-view fusion-based 2D brain tumor image segmentation, three independent 2D brain tumor segmentation sub-network models based on edge feature fusion and visual space state equations are used to process the image in each orthogonal plane in the axial, coronal, and sagittal directions. The predicted outputs of these three networks are then integrated through a fusion strategy to generate the final segmentation result. The fusion strategy includes mid-stage attention mechanism fusion, cross-view multi-head attention mechanism fusion, and terminal voxel fusion to maximize the use of complementary information from the three perspectives. The fusion strategy's approach is as follows: first, alignment and enhancement are performed at the representation layer; then, cross-view interaction is performed; and finally, robust convergence is achieved using logits attention.

[0090] 2D perspective Where Ax represents the axial direction, Co represents the coronal direction, Sa represents the sagittal direction, and the image height is... , width is The label mask is Number of channels Batch processing size Features after alignment of the same voxel ,in For the scale.

[0091] The intermediate attention mechanism is fused as follows: in By integrating channel and spatial attention mechanisms at the scale, a cross-perspective fusion representation is obtained. ,include:

[0092] 1) Channel attention: First, perform global average pooling on the features of each viewpoint. Then, a small multilayer sensor is used to obtain the channel gating value. ,in These are the parameters for a small multilayer sensor. For the Sigmoid function, For the SiLU function; finally, Expand back to the original space size and scale each channel individually. .

[0093] 2) Spatial weighting: The position of each pixel in each viewpoint is... The weighted features of the three channels are then spliced ​​together in the channel dimension and fed into a... Pixel multilayer perceptron Output the prediction probability scores for the three viewpoints and normalize them. deal with:

[0094]

[0095] Among them, the probability scores of the three perspectives ,and .

[0096] 3) Blending Features and Projection: First, combine pixels one by one to obtain the blending features. .

[0097] To preserve the residual information from each viewpoint, Features from three perspectives To splice, use projection Reverting to the original standard number of channels yields the fused backbone characteristics: .

[0098] The cross-view multi-head attention fusion is as follows: In At scale, a single-layer multi-head cross-view attention mechanism is used to unfold the planar image features. : After unfolding, there are a total of Each token; for each perspective Using its own features as the sequence Query, it interacts with the feature sequences Key and Value from the other two paths respectively, resulting in the following linear mapping:

[0099] , ,

[0100] in, From the perspective Feature sequence Query , For the other two paths, the feature sequences Key and Value are... Here are the parameter matrices for the three attention sequences; In addition to The stitching of two additional perspectives outside the main viewpoint; let the number of heads be h, and the dimension of a single head be... The sum of attention outputs for the i-th head is obtained from the above linear mapping: Then, the output results of each head are summed and concatenated. and normalize After processing the residuals, we obtain: ;in This is the parameter matrix for residual processing.

[0101] Then connect to the feedforward network layer. And add residual processing: Then Transform back to the size of the original feature map Finally, the updated features from the three perspectives are gated and fused. A cross-perspective fusion representation was obtained. ,in These are the backbone features derived from mid-term attention-based fusion. The hyperparameters for gating fusion.

[0102] The final voxel fusion is defined as: the prediction score of the corresponding feature of the same voxel in each direction. After splicing, it passes through a small multilayer sensor. Obtain attention weights in each direction. :

[0103] ,

[0104] in, It is an exponential function with the natural logarithm e as its base. This is the result of late-stage voxel fusion.

[0105] Figure 3 This is a network structure diagram of a two-dimensional brain tumor medical image segmentation model based on temporal Vision Mamba and edge feature extraction. In the backbone network based on temporalization and the visual-space state equation, after temporalization improvement, the model supports inputting a set of consecutive slices from one case at a time, and can learn the segmentation information of each consecutive slice, sharing features.

[0106] The backbone network's encoder consists of the Visual State Space Block (VSS Block) and Patch Merging from Vision Mamba. The VSS Block is used to extract low-level image detail features and high-level semantic information. The Patch Merging operation is a downsampling operation performed after dividing the feature map into several small blocks. The backbone network's decoder consists of the VSS Block and Patch Expanding. Patch Expanding is the inverse operation of Patch Merging and is responsible for enlarging the spatial resolution of the feature map to gradually restore it to the size of the input image for reconstructing segmentation details.

[0107] The temporal implementation method is as follows: a two-dimensional image is divided into image blocks. Postscript ; Divide the image into blocks at each scale s The corresponding slice image features are obtained after passing through the encoder VSS Block. , is represented as: Image features at each scale Subsequently, a Time Vision State Space (TVSS) model was introduced, and a one-dimensional selective scan was performed on the slice at each spatial location. Flatten the spatial grid Output ;in It applies the VSS Block to the dimension T of the continuous slice.

[0108] Figure 4 This is a structural diagram of the Vision State Space Block (VSSBlock) in Vision Mamba. The backbone network based on temporalization and the vision space state equation, the Vision State Space Block (VSS Block) in Vision Mamba, has a core two-dimensional selective scan (SS2D) module. The VSS Block contains two branches: a local branch and a state space model (SSM) branch; after the first image block segmentation, the feature map size is... After block merging, the number of channels ,and , Features after the s-th encoding stage: ; denote the continuous slice of input The local branch structure is: linear projection layer. +Normalization .

[0109] The SSM branch uses a selective state-space model to perform a linear-time scan on the flattened one-dimensional vector, using horizontal and vertical scans to adapt to the two-dimensional model; the SSM branch structure is: linear projection layer. +Multi-scale depth-separable convolutional layers +Edge enhancement layer +SiLU function + SS2D+ Normalization The formulas are as follows:

[0110] , ,

[0111] , ,

[0112] Then, gated fusion and linear mixing of the two branches are performed, along with the residuals, using the following formula:

[0113] , ,

[0114] Then connect to a feed-forward network (FFN): .

[0115] The SS2D module has the following structure: Scan Expanding + S6 unit + Scan Merging; firstly, the two-dimensional features are expanded into four sequences along the four diagonal directions. Where d represents the directions of the four diagonals. , TL represents the top left, BR the bottom right, TR the top right, and BL the bottom left.

[0116] The S6 element is the core element in the Vision Mamba model based on the discretized spatial state equation; for the continuous spatial state equation, we have: , .

[0117] in, , , These are, respectively, the state value reflecting the current system, the value representing external influences introduced into the system, and the value we hope to observe from the system. The output state values ​​of the current system are represented by A, B, C, and D, respectively, which are the state matrix, control matrix, output matrix, and command matrix. The continuous state equations are discretized (zero-order hold, step size...). ):

[0118] ,

[0119] in, The state matrix is ​​a discrete-time state matrix. It is a continuous-time state matrix. For discrete time steps, The value of the input control matrix at step t.

[0120] Selective parameters , , Input-driven (data-dependent): , , ;in For learnable parameter matrix, The input features at step t , For bias terms, Soft positive function , , Represented as a learnable mapping function, This is represented as converting a vector into a diagonal matrix; a one-way recursive approach is used: , For each direction d, the final output is: .

[0121] Scan Merging restores the output from the four directions to two dimensions and aggregates them to obtain four tensors. Then, by using weighted attention summation, the four tensors are aggregated and restored to their original size: , .

[0122] An edge extraction network based on convolution and spatial attention mechanisms is constructed. The edge feature extraction network consists of a two-dimensional convolution module, a linear projection, and a feature mapping layer. It is used to achieve multi-level feature extraction, extract edge features from the shallow layer of the image, form an edge feature map, and finally output the edge feature map to the edge-semantic feature fusion module.

[0123] Figure 5 This is a structural diagram of the bidirectional edge-semantic feature fusion module. The bidirectional edge-semantic feature fusion module inputs edge feature maps and segmentation image feature maps, fuses the two feature maps, enhances the model's recognition of blurred boundary features, and achieves "boundary-guided segmentation" and "segmentation-corrected boundary." The implementation method of the bidirectional edge-semantic feature fusion module is as follows:

[0124] (1) Receive the boundary feature map output by the edge feature extraction network and the semantic segmentation feature map output by the backbone network encoder respectively, and denote the boundary feature map output by the edge extraction network as... Semantic segmentation features output by the backbone network encoder After passing through the same sampling layer Align to the same size r is the variation ratio, and it combines max pooling and average pooling. After suppressing noise, we obtain the aligned boundary feature map and segmentation feature map:

[0125] , ,

[0126] (2) Input the aligned two-way features into the interactive attention layer, and set up three mutually coupled attention enhancement branches in the interactive attention layer to complete deep feature selection and fusion; including:

[0127] ① Channel Attention Branch: Based on compression and activation attention mechanisms, the upsampled features of the encoder are channel-weighted to learn channel-level importance coefficients. Effective segmentation channel features are learned through adaptive reinforcement.

[0128] , ,

[0129] , ,

[0130] in, Features of the boundary passage Semantic channel features, For global average pooling, For the Sigomid function, The SiLU activation function is used. These are the parameters for a small multilayer sensor. The boundary features are those after channel attention weighting. Semantic segmentation features weighted by channel attention It is represented as the Hadema product of element-wise multiplication.

[0131] ② Spatial attention branch: utilizing max pooling With average pooling Generate and stitch together spatial attention maps , Enhance the network's spatial awareness of key lesion areas and boundary locations:

[0132] ,

[0133] ,

[0134] in Calculate for k convolutional layers, Boundary features after spatial attention weighting, These are semantic segmentation features weighted by spatial attention.

[0135] ③ Bilinear Modeling Branch: Perform bilinear pooling on channel attention and spatial attention features to model higher-order semantic relationships, and fuse it with the outputs of the first two branches; in the same spatial location Take the outer product of the two channel vectors And compressed to The fused feature vector is obtained from the dimension. ,in Indicates the first Boundary features at each location, No. Semantic segmentation features at each location; then, the fused feature vectors from all locations are collected and arranged. For the new feature map .

[0136] (3) The features fused through the interactive attention layer are then processed... convolution Implement gate inter-gate control to achieve bidirectional guidance of boundary features and semantic segmentation features:

[0137] ,

[0138] in These are semantic segmentation features after bidirectional guidance. These are the boundary features after bidirectional guidance.

[0139] Finally, the above three branches and the bidirectional interaction results are concatenated at the channel dimension and residual processing is performed. Its expression is and introduce To alleviate overfitting:

[0140] ,

[0141] Where F is the concatenated feature map, and Y is the final output of the bidirectional edge-semantic feature fusion module; finally, the output fused feature map Y is used by the decoder to reconstruct the high-resolution lesion segmentation result. In the skip connection path, bidirectional feature fusion replaces the traditional method of only concatenating, and the spatial attention map generated by max pooling and average pooling is used to emphasize key regions and suppress redundant background, thereby improving boundary consistency and segmentation accuracy.

[0142] The method for constructing the combined loss function of weighted cross-entropy and weighted Descein loss is as follows:

[0143] In glioma segmentation, the number of samples for different categories (such as background, tumor core, necrotic areas, etc.) varies greatly. During training, the model tends to predict larger categories (background), neglecting smaller categories, leading to class imbalance. To avoid this, a weighted loss scheme is used, determining the weight of each category in the loss function based on its proportion in the segmentation image. Specifically, a combination of Weighted Cross Entropy (WCE) and Weighted DICE (WD) loss is used as the total loss. WCE loss aims to improve image-level segmentation and aids model convergence. Weighting the cross entropy loss function increases the weight of deep regions with high imbalance (i.e., fewer target pixels compared to the background). To address severe class imbalance, WD loss weights are set based on the proportion of each category. Training a network with WD loss in its loss function helps detect small lesions and reduces false negatives. This combined loss function effectively reduces the uncertainty of category prediction and improves the model's classification accuracy in multi-class segmentation tasks.

[0144] WCE loss WD loss and combined loss function They are represented as follows:

[0145]

[0146]

[0147]

[0148] in, Indicates input data, This represents the corresponding actual segmentation label. Represents network parameters, This represents the weight of class C in the WCE loss. Represents the true segmentation mask. This represents the predicted segmentation probability mask. This represents the balancing weight of class c in the WD loss, set to... .

[0149] Figure 6 These are partial slice images of brain tumors in the BraTS2021 dataset from the method of this invention, as well as the results of a two-dimensional multi-view brain tumor medical image segmentation model based on the VisionMamba time series model.

[0150] This embodiment also provides a two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba time series model, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it can implement the above-mentioned method.

[0151] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0155] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A two-dimensional multi-view brain tumor medical image segmentation method based on a Vision Mamba time sequence model, characterized in that, Includes the following steps: (1) Obtain brain tumor medical image dataset and preprocess it. Slice the three-dimensional image data in the dataset into three orthogonal planes in the axial, coronal and sagittal directions to obtain three sets of two-dimensional slice image data from different perspectives. (2) Construct three independent two-dimensional brain tumor segmentation sub-network models based on edge feature fusion and visual spatial state equations to process slice images in the axial, coronal and sagittal directions respectively; The segmentation results from the three sub-network models are then fused to obtain the final output. The two-dimensional brain tumor segmentation sub-network model based on edge feature fusion and visual spatial state equation includes: a backbone network based on temporalization and visual spatial state equation, an edge feature extraction network based on convolution, and a bottleneck layer based on a bidirectional edge-semantic feature fusion module. (3) Construct a combined loss function of weighted cross-entropy and weighted Descein loss. Use the three sets of two-dimensional slice image data from different perspectives obtained in step (1) as training datasets to train the two-dimensional brain tumor segmentation sub-network model based on edge feature fusion and visual space state equation. During the training process, use mini-batch gradient descent combined with adaptive optimization algorithm to update parameters. In each iteration, backpropagate and gradient correction are performed on the network weights to gradually approach the optimal solution. Save the optimal weight parameters based on the minimum loss of the validation set during the training process. After training, use the model for subsequent prediction and evaluation. After the training phase ends, load the saved optimal weight model, perform segmentation prediction on the test set, and compare the prediction results with the real labels. Use multiple evaluation indicators to evaluate the performance to verify the accuracy, robustness and edge detail prediction ability of the model in medical image segmentation tasks.

2. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba time series model according to claim 1, characterized in that, To achieve multi-view fusion for 2D brain tumor image segmentation, three independent 2D brain tumor segmentation sub-network models based on edge feature fusion and visual space state equations are used to process the images in each orthogonal plane in the axial, coronal, and sagittal directions. Then, the prediction outputs of these three networks are integrated through a fusion strategy to generate the final segmentation result. The fusion strategy includes mid-term attention mechanism fusion, cross-view multi-head attention mechanism fusion, and terminal voxel fusion to maximize the use of complementary information from the three perspectives.

3. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba time series model according to claim 2, characterized in that, The fusion strategy is as follows: first, alignment and enhancement are performed at the representation layer, then cross-view interaction is performed, and finally, logits attention is used for robust convergence. recording 2D view where Ax represents axial, Co represents coronal, Sa represents sagittal, height of image is , width is , label mask is , number of channels , batch size ; same voxel aligned features where is scale; The intermediate attention mechanism is fused as follows: in By integrating channel and spatial attention mechanisms at the scale, a cross-perspective fusion representation is obtained. ,include: 1) Channel attention: First, perform global average pooling on the features of each viewpoint. Then, a small multilayer sensor is used to obtain the channel gating value. ,in These are the parameters for a small multilayer perceptron. For the Sigmoid function, For the SiLU function; finally, Expand back to the original space size and scale each channel individually. ; 2) Spatial weighting: The position of each pixel in each viewpoint is... The weighted features of the three channels are then spliced ​​together in the channel dimension and fed into a... Pixel multilayer perceptron Output the prediction probability scores for the three viewpoints and normalize them. deal with: Among them, the probability scores of the three perspectives ,and ; 3) Blending Features and Projection: First, combine pixels one by one to obtain the blending features. ; To preserve the residual information from each viewpoint, Features from three perspectives To splice, use projection Reverting to the original standard number of channels yields the fused backbone characteristics: ; The cross-view multi-head attention fusion is as follows: In At scale, a single-layer multi-head cross-view attention mechanism is used to unfold the planar image features. : After unfolding, there are a total of Each token; for each perspective Using its own features as the sequence Query, it interacts with the feature sequences Key and Value from the other two paths respectively, resulting in the following linear mapping: ， ， in, From the perspective Feature sequence Query , For the other two paths, the feature sequences Key and Value are... Here are the parameter matrices for the three attention sequences; In addition to The stitching of two additional perspectives outside the main viewpoint; let the number of heads be h, and the dimension of a single head be... The sum of attention outputs for the i-th head is obtained from the above linear mapping: Then, the output results of each head are summed and concatenated. and normalize After processing the residuals, we obtain: ;in The parameter matrix for residual processing; Then connect to the feedforward network layer. And add residual processing: Then Transform back to the size of the original feature map Finally, the updated features from the three perspectives are gated and fused. A cross-perspective fusion representation was obtained. ,in These are the backbone features derived from mid-term attention-based fusion. Hyperparameters for gated fusion; The final voxel fusion is defined as: the prediction score of the corresponding feature of the same voxel in each direction. After splicing, it passes through a small multilayer sensor. Obtain attention weights in each direction. : ， in, It is an exponential function with the natural logarithm e as its base. This is the result of late-stage voxel fusion.

4. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model according to claim 2, characterized in that, In the backbone network based on temporalization and visual space state equations, after the temporalization improvement network, the model supports a set of continuous slices of one case as input at a time, and can learn the segmentation information of each continuous slice and share features; the encoder of the backbone network is composed of the visual state space module VSS Block and block merging in Vision Mamba, where the visual state space module is used to extract low-level image detail features and high-level semantic information. The block merging operation is a downsampling operation performed after dividing the feature map into several small blocks; the decoder of the backbone network consists of a visual state space module and block expansion. Block expansion is the inverse operation of block merging and is responsible for increasing the spatial resolution of the feature map to gradually restore it to the size of the input image for reconstructing segmentation details. The temporal implementation method is as follows: a two-dimensional image is divided into image blocks. Postscript ; Divide the image into blocks at each scale s The corresponding slice image features are obtained after passing through the encoder VSS Block. , represented as: Image features at each scale Subsequently, a temporal state-space model was introduced, and a one-dimensional selective scan was performed on the slice at each spatial location. Flatten the spatial grid Output ;in It applies the VSS Block to the dimension T of the continuous slice.

5. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model according to claim 4, characterized in that, The backbone network based on temporalization and visual space state equations, the Vision Mamba Visual State Space Block (VSS Block), relies on the two-dimensional selective scanning SS2D module. The Visual State Space Block contains two branches: a local branch and a State Space Model (SSM) branch. After the first image segmentation, the feature map size is: After block merging, the number of channels ,and , Features after the s-th encoding stage: ; denote the continuous slice of input ; The local branching structure is: linear projection layer +Normalization ; The SSM branch uses a selective state-space model to perform a linear-time scan on the flattened one-dimensional vector, using horizontal and vertical scans to adapt to the two-dimensional model; the SSM branch structure is: linear projection layer. +Multi-scale depth-separable convolutional layers +Edge enhancement layer +SiLU function + SS2D+ Normalization The formulas are as follows: ， ， ， ， Then, gated fusion and linear mixing of the two branches are performed, along with the residuals, using the following formula: ， ， Then connect to the feedforward network FFN: .

6. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model according to claim 5, characterized in that, The SS2D module has the following structure: Scan Expanding + S6 unit + Scan Merging; firstly, the two-dimensional features are expanded into four sequences along the four diagonal directions. Where d represents the directions of the four diagonals. , TL represents the top left, BR the bottom right, TR the top right, and BL the bottom left; The S6 element is the core element in the Vision Mamba model based on the discretized spatial state equation; for the continuous spatial state equation, we have: , ; in, , , These are, respectively, the state value reflecting the current system, the value representing external influences introduced into the system, and the value we hope to observe from the system. Let A represent the current system output state value; where A, B, C, and D are the state matrix, control matrix, output matrix, and command matrix, respectively; the continuous state equation is discretized as follows: ， in, The state matrix is ​​a discrete-time state matrix. It is a continuous-time state matrix. For discrete time steps, The value of the input control matrix at step t; Selective parameters , , Input-driven: , , ;in For learnable parameter matrix, The input features at step t , For bias terms, Soft positive function , , Represented as a learnable mapping function, This is represented as converting a vector into a diagonal matrix; a one-way recursive approach is used: , For each direction d, the final output is: ; Scan Merging restores the output from the four directions to two dimensions and aggregates them to obtain four tensors. Then, by using weighted attention summation, the four tensors are aggregated and restored to their original size: , .

7. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model according to claim 2, characterized in that, An edge extraction network based on convolution and spatial attention mechanisms is constructed. The edge feature extraction network consists of a two-dimensional convolution module, a linear projection, and a feature mapping layer. It is used to achieve multi-level feature extraction, extract edge features from the shallow layer of the image, form an edge feature map, and finally output the edge feature map to the edge-semantic feature fusion module.

8. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model according to claim 1, characterized in that, The bidirectional edge-semantic feature fusion module inputs edge feature maps and segmentation image feature maps into the module, fuses the two feature maps, enhances the model's recognition of blurred boundary features, and realizes boundary-guided segmentation and segmentation boundary correction; the implementation method of the bidirectional edge-semantic feature fusion module is as follows: (1) Receive the boundary feature map output by the edge feature extraction network and the semantic segmentation feature map output by the backbone network encoder respectively, and denote the boundary feature map output by the edge extraction network as... Semantic segmentation features output by the backbone network encoder After passing through the same sampling layer Align to the same size r is the variation ratio, and it combines max pooling and average pooling. After suppressing noise, we obtain the aligned boundary feature map and segmentation feature map: ， ， (2) Input the aligned two-way features into the interactive attention layer, and set up three mutually coupled attention enhancement branches in the interactive attention layer to complete deep feature selection and fusion; including: ① Channel Attention Branch: Based on compression and activation attention mechanisms, the upsampled features of the encoder are channel-weighted to learn channel-level importance coefficients. Effective segmentation channel features are learned through adaptive reinforcement. ， ， ， ， in, Features of the boundary passage Semantic channel features, For global average pooling, For the Sigomid function, The SiLU activation function is used. These are the parameters for a small multilayer perceptron. The boundary features are those after channel attention weighting. Semantic segmentation features weighted by channel attention Represented as the Hadema product of element-wise multiplication; ② Spatial attention branch: utilizing max pooling With average pooling Generate and stitch together spatial attention maps , Enhance the network's spatial awareness of key lesion areas and boundary locations: ， ， in Calculate for k convolutional layers, The boundary features are spatial attention-weighted. These are semantic segmentation features weighted by spatial attention; ③ Bilinear Modeling Branch: Perform bilinear pooling on channel attention and spatial attention features to model higher-order semantic relationships, and fuse it with the outputs of the first two branches; in the same spatial location Take the outer product of the two channel vectors And compressed to The fused feature vector is obtained from the dimension. ,in Indicates the first Boundary features at each location, No. Semantic segmentation features at each location; then, the fused feature vectors from all locations are collected and arranged. For the new feature map ; (3) The features fused through the interactive attention layer are then processed... convolution Implement gate inter-gate control to achieve bidirectional guidance of boundary features and semantic segmentation features: ， in These are semantic segmentation features after bidirectional guidance. These are the boundary features after bidirectional guidance; Finally, the above three branches and the bidirectional interaction results are concatenated at the channel dimension and residual processing is performed. Its expression is and introduce To alleviate overfitting: ， Where F is the concatenated feature map and Y is the final output of the bidirectional edge-semantic feature fusion module; the output fused feature map Y is then used by the decoder to reconstruct the high-resolution lesion segmentation result; in the skip connection path, bidirectional feature fusion replaces the traditional method of only concatenating, and the spatial attention map generated by max pooling and average pooling is used to emphasize key regions and suppress redundant background, thereby improving boundary consistency and segmentation accuracy.

9. The two-dimensional multi-view brain tumor medical image segmentation method based on the Vision Mamba temporal model according to claim 1, characterized in that, The method for constructing the combined loss function of weighted cross-entropy and weighted Descein loss is as follows: A weighted approach is used for the loss, determining the weight of each category of pixels in the loss function based on their proportion in the segmentation image. Specifically, the total loss is a combination of the weighted cross-entropy (WCE) loss function and the weighted Dessell (WD) loss function. (WCE loss) WD loss and combined loss function They are represented as follows: in, Indicates input data, This represents the corresponding actual segmentation label. Represents network parameters, This represents the weight of class C in the WCE loss. Represents the true segmentation mask. This represents the predicted segmentation probability mask. This represents the balancing weight of class c in the WD loss, set to... .

10. A two-dimensional multi-view brain tumor medical image segmentation system based on the Vision Mamba temporal model, characterized in that, It includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the method as described in any one of claims 1-9.

Images