An Improved Scale Fusion-Based Brain Tumor Segmentation Method

Abstract

This invention claims protection for an improved scale fusion-based brain tumor segmentation method, aiming to improve scale feature fusion by aggregating scale information within and between encoder layers, thereby achieving accurate brain tumor image segmentation. This method belongs to the field of computer vision. The method includes the following steps: Step 1. This invention proposes a global-local feature mixer, which achieves global-local scale information mixing by capturing local detail features and global context information in parallel, to solve the problem of uncertain brain tumor location. Step 2. This invention designs a multi-scale expansion initialization module, which can effectively capture multi-scale shape features, aiming to improve the network's ability to adapt to different brain tumor shapes. Step 3. This invention designs a multi-scale feature aggregation module, which can effectively aggregate multi-scale features from different encoder layers, helping to alleviate the problem of uncertain brain tumor size.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and particularly relates to a brain tumor segmentation method based on improved scale fusion. BACKGROUND

[0002] In modern medical diagnosis, accurate segmentation of brain tumors is a key step to improve treatment effectiveness and patient prognosis. Magnetic resonance imaging (MRI) and computed tomography (CT) as the main brain imaging means provide rich brain structure information for doctors. However, the segmentation of brain tumors faces uncertainty in position, shape and size, which seriously affects the accuracy of segmentation and the effectiveness of clinical application.

[0003] Early brain tumor segmentation methods mainly rely on manual segmentation by imaging experts. Although this method is accurate, it is time-consuming and susceptible to subjective factors. With the development of technology, researchers began to explore threshold-based automatic segmentation methods, which distinguish tumor tissue from normal tissue by setting specific thresholds. However, these methods are highly dependent on image quality and are difficult to adapt to the diversity and complexity of tumors. In order to improve the automation and accuracy of segmentation, researchers have proposed image feature-based methods. These methods extract features such as texture, shape and intensity distribution from images, and combine machine learning algorithms to segment tumors. Support vector machines (SVM), random forests and neural networks have been widely used in this field. Although these methods have improved the accuracy of segmentation to some extent, they usually require a lot of feature engineering and have limited adaptability to morphological changes and size differences of tumors.

[0004] In recent years, deep learning technology has made revolutionary progress in image segmentation. Convolutional neural networks (CNN) have become the mainstream method for medical image segmentation due to their strong feature extraction capabilities. In particular, for brain tumor segmentation, deep learning models such as U-Net and V-Net have received widespread attention due to their excellent performance. These models can automatically learn hierarchical structural features in images, effectively improving the accuracy and robustness of tumor segmentation.

[0005] In the field of brain tumor segmentation, previous studies have explored various methods to address the uncertainty of tumor location, the diversity of tumor shape, and the variability of tumor size. These methods typically include expanding the target volume, introducing attention mechanisms, and feature fusion updates to achieve precise local positioning of tumor regions. Meanwhile, using visual Transformers for long-range dependency modeling helps capture global contextual information in images, thereby improving segmentation accuracy. However, these methods often have limitations in handling local and global features. Specifically, they usually treat the modeling process of local and global features as two independent steps rather than simultaneous modeling in the same initial state. This independent modeling approach can interfere with the quality of the extracted features, as the extraction of a single feature is influenced by the prior of another feature. Additionally, previous studies often focus on local modeling when extracting multi-scale features, lacking comprehensive perception of multi-scale features in the global range.

[0006] Furthermore, although visual Transformer-based brain tumor segmentation networks perform well in capturing global contextual information, they often come with high computational costs, requiring strict hardware resources. This makes it difficult for these networks to achieve efficient operation while maintaining high performance in practical applications.

[0007] Therefore, how to effectively integrate local and global multi-scale contextual information to achieve more accurate brain tumor segmentation and balance the performance and efficiency of the network remains a challenge. Future research needs to explore new methods and technologies to overcome the limitations of existing methods and further develop brain tumor segmentation technology.

[0008] After searching, the application publication number CN115424103A, an improved U-Net brain tumor segmentation method based on attention mechanism and multi-scale feature fusion, belongs to the field of semantic segmentation, including the following steps: S1, data acquisition and data preprocessing; S2, constructing an improved U-Net brain tumor segmentation model based on attention mechanism and multi-scale feature fusion; S3, constructing a hybrid loss function and training the improved U-Net brain tumor segmentation model, saving the optimal model; S4, using the optimal model for prediction, saving the prediction results, online verification, obtaining evaluation indicators, and finally comparing the results; After adding a residual module to the U-Net model, using deep over-parameterized convolution instead of ordinary convolution, adding a multi-scale feature fusion module and an attention mechanism module, the segmentation effect is quantitatively evaluated using the Dice similarity coefficient. The experimental results show that the improved U-Net network can effectively improve the segmentation accuracy of MRI brain tumor images and has good segmentation performance.

[0009] In a patent based on attention mechanism and multi-scale feature fusion improved U-Net brain tumor segmentation method, the attention mechanism and multi-scale feature are constructed to solve the problems of brain tumor position area not fixed and tumor shape size different. However, it is worth noting that when extracting multi-scale features, the research focuses on local modeling, but lacks comprehensive perception of multi-scale features in the global range, and the feature fusion efficiency is low. How to effectively fuse local and global multi-scale context information is crucial to realize accurate brain tumor segmentation. Based on this, the present application constructs a global-local feature mixer. The module captures local detail features and global context information in parallel, designs a feature fusion enhancer, and realizes effective fusion of global-local scale information. Through multi-scale channel attention and multi-scale feature aggregation module, the multi-scale features from different levels of encoder are effectively aggregated. The local and global multi-scale context information is effectively fused. SUMMARY

[0010] The present application aims to solve the problems of the above prior art. An improved scale fusion brain tumor segmentation method is proposed. The technical scheme of the present application is as follows:

[0011] An improved scale fusion brain tumor segmentation method, comprising the following steps:

[0012] Step 1, input 4 multi-modal MR brain images containing tumor lesions into the network for brain tumor segmentation, and draw three nested lesion sub-regions ET (peritumoral edema region), WT (whole tumor region) and TC (tumor core region) of brain tumor;

[0013] Step 2, construct a global-local feature mixer GLFM (spatial mixing Transformer, input dependent dynamic convolution, feature fusion enhancer), capture local detail features and global context information in parallel, and realize information mixing of global-local scale;

[0014] Step 3, construct a multi-scale expansion initial module MSDI (dilated convolution + multi-scale channel attention), realize effective capture of multi-scale deformation features through dilated convolution with different expansion rates, Inception's unique structure and multi-scale channel attention module MS-CAM (global average pooling + point convolution + activation function);

[0015] Step 4, construct a multi-scale feature aggregation module MFSA (downsampling, channel attention, convolution block + spatial attention, regular linear attention), integrate multi-level multi-resolution features into long-range dependent modeling, and realize multi-scale global perception ability.

[0016] Further, the step 1 inputs 4 multi-modal MR brain images containing tumor lesions into the network for brain tumor segmentation, and outlines three nested lesion sub-regions ET, WT and TC of the brain tumor, specifically:

[0017] The training images are input into the brain tumor image segmentation network, which takes a 3D U-Net structure as the basic network structure, encodes and extracts features from the input images to obtain layer features F1, F2, F3, F4, the first three layers of features pass through the multi-scale expansion initial module MSDI, and the fourth layer of features passes through the multi-scale feature aggregation module MFSA, then the corresponding decoder is used for feature splicing to restore the details lost by downsampling, and the decoder gradually restores the feature mapping to the size of the original input image through progressive upsampling and convolution operation, thereby realizing end-to-end brain tumor image segmentation.

[0018] Further, the step 2 constructs a global-local feature mixer GLFM, which effectively captures global-local layer representations to integrate local and global context information, and realizes information mixing of global-local scales by capturing local detail features and global context information in parallel, specifically:

[0019] A, the input feature x is first uniformly divided into two sub-feature mappings along the channel dimension to obtain corresponding feature mappings {x1, X2, and the calculation process can be formulated as:

[0020] x1, x2 = split(x) (1)

[0021] Wherein, split represents uniform division of the input feature along the channel dimension.

[0022] B, the feature mappings x1, x2 are modeled in parallel to obtain global feature f g and local feature f | , and the calculation process is formulated as follows:

[0023] f0(x1) = x1 + Norm(WMSA / shuffle WMSA(x1)) (2)

[0024] f g (x1) = f0(x1) + Norm(MLP(f0(x1))) (3)

[0025] A' = Conv(Conv(AdaptivePool(x2))) (4)

[0026] A = Softmax(Reshape(A')) (5)

[0027]

[0028] f l (x2) = Softmax(BN(W * x2)) (7)

[0029] where x1 and x2 are the feature maps {x1, x2} generated in step A, WMSA and shuffle WMSA in equation (2) represent window-based multi-head attention and shuffle window multi-head self-attention respectively, Norm represents the normalization operation, MLP represents the multi-layer perception, f0(x1) represents the feature generated after attention and normalization operation, f g is the final generated global feature; Adaptivepool in equation (4) represents adaptive average pooling, Softmax is an activation function, and A' represents an intermediate generated attention map; A in equation (5) is an attention weight, G in equation (6) represents an output dimension, P represents a set of learnable parameters, and W represents a convolution kernel dependent on the input; BN in equation (7) represents a normalization operation, and f l is the final local feature.

[0030] C, fuse the global feature and the local feature to obtain a fused mixed feature Y, and then pass the mixed feature Y through a feature fusion enhancer (TFE) to obtain a final mixed feature, and the calculation process is as follows:

[0031]

[0032] where f g and f l are the global feature and the local feature generated in step C, the feature fusion enhancer (TFE) includes a 3x3x3 deep convolution and a 1x1x1 point convolution, an extended 1x1x1 convolution, and a residual connection, which are used to enhance the local-global information fusion.

[0033] Further, the step 3 constructs a multi-scale expansion initial module MSDI, which realizes effective capture of multi-scale deformation features through expansion convolutions with different expansion rates, Inception-specific structures, and a multi-scale channel attention module MS-CAM, and specifically:

[0034] A, taking the first three layers of the encoder feature y as the multi-scale feature input, respectively passing through four convolution operations, the first two layers are expansion convolutions with different expansion rates, and the last two layers are 1x1x1 convolution and 1x1x1 convolution with pool operation, splicing the feature maps obtained by the first two layers of convolution to obtain y1, and splicing the feature maps obtained by the last two layers of convolution to obtain y2, and specifically:

[0035] y1 = DC(DC(Conv(y))) + Cnov(DC(y)) (8)

[0036] y2 = Pool(Conv(y)) + Conv(y) (9)

[0037] wherein, in formula (8), DC represents dilated convolution, the first two DCs have a dilation rate of 1, and the last DC has a dilation rate of 2, and Conv represents 1x1x1 convolution; in formula (9), Pool represents a pooling operation;

[0038] B, send y1 and y2 into the MS-CAM module to obtain the final feature map y, specifically:

[0039] y = MS-CAM(y1, y2) (10)

[0040] wherein, y1 and y2 are the feature maps obtained in step A1, and MS-CAM is a multi-scale channel attention.

[0041] Further, the step 4 constructs a multi-scale feature aggregation module MFSA to integrate multi-level multi-resolution features into long-range dependency modeling, and realizes multi-scale global perception capability, specifically:

[0042] A, taking the output features (x 1 , x 2 , x 3 , x 4 ) of each layer of the encoder as input, down-sampling them to the same resolution size through cross-step convolution operation with different step lengths, then performing feature stacking along the channel direction, i.e. Concatenation operation, and then performing channel self-adaptation on the spliced features, i.e. using channel attention SE to assign different weights to each channel, thereby obtaining the fused inter-layer multi-scale features, and the calculation process can be formulated as follows:

[0043]

[0044] wherein, DownSampling represents a down-sampling operation, which is realized by cross-step convolution in this network, which is conducive to learning better sampling points, Concat represents stacking along the channel direction, and SE represents channel attention operation, and X fuse is the fused multi-scale feature;

[0045] B, local modeling is performed on the fused feature X fuse obtained in step A, first passing through two consecutive convolution blocks CB (convolution + normalization + nonlinear activation) to obtain local feature mapping, and then generating local constraint weights through spatial attention SA, and the calculation formula is as follows:

[0046] ω local = SA(CB(X fuse )) (12)

[0047] wherein ω local is the generated locality constraint weight.

[0048] C, the fusion feature X obtained in step A is subjected to global modeling, specifically including: X fuse is subjected to global modeling, specifically including: X fuse First, it is fed into the rule linear attention for the capture of global context information, to obtain a feature containing global dependency, and then the locality weight generated in step B is subjected to matrix multiplication operation with it, to give it locality constraint, to avoid excessive attention of deep features to globality, and the calculation formula is as follows:

[0049] Z = ω local · CB(Xfus e )(13)

[0050] wherein CB represents the proposed rule linear attention, including CBAM and 1x1 convolution, ω local is the locality constraint weight, and Z is the output feature after the aggregation of the multi-scale context information.

[0051] An electronic device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the improved scale fusion brain tumor segmentation method according to any one of claims 1 to 5 when executing the program.

[0052] A non-transitory computer readable storage medium, having a computer program stored thereon, wherein the computer program is executable by a processor to implement the improved scale fusion brain tumor segmentation method according to any one of claims 1 to 5.

[0053] The advantages and beneficial effects of the present application are as follows:

[0054] The present application mainly aims at the problem that the traditional convolutional neural network has poor segmentation accuracy when facing brain tumors with variable shape, size and position, and designs an improved scale fusion brain tumor segmentation method, the core of which is to realize multi-context learning through the fusion of multi-scale features within and between layers, so as to more effectively adapt to tumors with different shapes and positions.

[0055] Firstly, a global-local feature mixer (GLFM) is constructed, which effectively captures global-local layer representation to integrate local and global context information, and realizes information mixing of global-local scales by capturing local detail features and global context information in parallel. Among them, the designed feature fusion enhancer enhances the feature fusion without increasing the calculation overhead; the designed local branch (input-dependent dynamic convolution) generates a spatially varying attention map for each attention group, which better supports local feature coding.

[0056] Secondly, a multi-scale dilation initial module (MSDI) is designed to effectively capture multi-scale deformation features through dilation convolution with different dilation rates, the unique structure of Inception and a multi-scale channel attention module (MS-CAM). The combination of dilation convolution with different dilation rates and Inception V2 extracts more extensive multi-scale features without increasing the computational complexity, and the introduction of multi-scale channel attention effectively aggregates multi-scale semantic information.

[0057] Finally, a multi-scale feature aggregation module (MFSA) is proposed to integrate multi-level multi-resolution features into long-range dependency modeling to achieve multi-scale global perception. The rule linear attention (CBAM+1x1x1 convolution) is ingeniously used to replace the Transformer for long-range dependency modeling, saving the computational overhead, and the designed auxiliary local branch (convolution block+spatial attention) also effectively guides the long-range modeling, avoiding the excessive preference of high-level features for global context.

[0058] The experimental results on two public and well-known brain tumor segmentation image datasets BraTS2019 and BraTS2020 show that the proposed method has comparable or even more advanced performance than other state-of-the-art methods.

[0059] Unlike most previous convolutional neural networks and visual Transformer hybrid structures, the parallel hybrid structure is adopted in the method to effectively realize the synchronous perception of local details and global dependencies, thereby decoupling the local and global feature capture and avoiding mutual crosstalk caused by the serial connection. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is the overall network model structure schematic diagram of the preferred embodiment provided by the application.

[0061] Figure 2 is the global-local feature mixer structure schematic diagram of the application.

[0062] Figure 3 is the multi-scale dilation initial module structure schematic diagram of the application.

[0063] Figure 4 is the multi-scale feature aggregation module structure schematic diagram of the application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the application will be described in detail below with reference to the drawings of the embodiments of the application. The described embodiments are only a part of the embodiments of the application.

[0065] The technical scheme for solving the above technical problems of the present application is:

[0066] As Figure 1 The figure shows the overall framework of the improved scale fusion brain tumor segmentation method provided by the embodiment, including but not limited to the following steps:

[0067] Step 1, using 3D U-net as baseline, the details of the composition include but are not limited to the following contents: A, input the training image into the brain tumor image segmentation network, the brain tumor image segmentation network takes the 3D U-Net structure as the basic network structure, encodes and extracts the features of the input image to obtain the layer features F1, F2, F3, F4, the first three layer features pass through the multi-scale dilation initial module (MSDI), the fourth layer feature passes through the multi-scale feature aggregation module (MFSA), then the corresponding decoder is used for feature splicing to restore the details information lost by downsampling, the decoder gradually restores the feature mapping to the size of the original input image through progressive upsampling and convolution operation, so as to realize end-to-end brain tumor image segmentation.

[0068] Step 2, build a global-local feature mixer (GLFM) to effectively capture global-local layer representation to integrate local and global context information, capture local detail features and global context information in parallel to realize information mixing of global-local scale;

[0069] Step 3, build a multi-scale dilation initial module (MSDI) to effectively capture multi-scale deformation features through dilation convolution with different dilation rates, Inception-specific structure and multi-scale channel attention module (MS-CAM).

[0070] Step 4, build a multi-scale feature aggregation module (MFSA) to integrate multi-level multi-resolution features into long-range dependency modeling to realize multi-scale global perception ability.

[0071] As Figure 2 The figure shows the global-local feature mixer (GLFM) module diagram constructed by the embodiment, and the specific contents include but are not limited to the following contents:

[0072] A, the input feature x is first evenly divided into two sub-feature mappings along the channel dimension to obtain the corresponding feature mappings {x1, X2, and the calculation process can be formulated as:

[0073] x1, x2 = split (x) (1)

[0074] Wherein, split represents evenly dividing the input feature along the channel dimension.

[0075] B. Parallelly modeling global-local features for feature maps x1, x2, obtaining global feature f g and local feature f l , whose calculation process is formulated as follows:

[0076] f0(x1) = x1 + Norm(WMSA / shuffle WMSA(x1)) (2)

[0077] f g (x1) = f0 + Norm(MLP(f0)) (3)

[0078] A' = Conv(Conv(AdaptivePool(x2))) (4)

[0079] A = Softmax(Reshape(A')) (5)

[0080]

[0081] f l (x2) = Softmax(BN(W*x2)) (7)

[0082] wherein x1 and x2 are feature maps generated in step B, WMSA and shuffle WMSA in equation (2) represent window-based multi-head attention and shuffle window multi-head self-attention respectively, Norm represents normalization operation, MLP represents multi-layer perception, f g is the final generated global feature; Adaptivepool in equation (4) represents adaptive average pooling, Softmax is an activation function, and A' represents an intermediate generated attention map; A in equation (5) is an attention weight, G in equation (6) represents an output dimension, P represents a set of learnable parameters, and W represents an input-dependent convolution kernel; BN in equation (7) represents a normalization operation, and f l is the final local feature.

[0083] C. Effectively fusing the global feature and the local feature to obtain a fused mixed feature Y, and then passing through a feature fusion enhancer (TFE) to obtain the final mixed feature, whose calculation process is formulated as follows:

[0084]

[0085] wherein f g and f l are the global feature and the local feature generated in step C, the feature fusion enhancer (TFE) includes a 3x3x3 deep convolution and a 1x1x1 point convolution, an extended 1x1x1 convolution, and a residual connection, which are used to enhance local-global information fusion.

[0086] As Figure 3 shown is a multi-scale expansion initial (MSDI) module diagram constructed by the embodiment of the application, and the specific content includes but is not limited to the following content:

[0087] A. Take the first three layers of the encoder feature y as the multi-scale feature input, respectively pass through four layers of convolution operation, the first two layers are respectively dilated convolution with different expansion rates, and the last two layers are respectively 1x1x1 convolution and 1x1x1 convolution with pool operation, the feature maps obtained by the first two layers of convolution are spliced to obtain y1, and the feature maps obtained by the last two layers of convolution are spliced to obtain y2, specifically:

[0088] y1=DC(DC(Conv(y)))+Cnov(DC(y)) (8)

[0089] y2=Pool(Conv(y))+Conv(y) (9)

[0090] Wherein, in the formula (8), DC represents dilated convolution, the expansion rates of the first two DCs are 1, and the expansion rate of the last DC is 2, and Conv represents 1x1x1 convolution; In formula (9), Pool represents the pooling operation.

[0091] B. Send y1 and y2 into the MS-CAM module to obtain the final feature map y, specifically:

[0092] y=MS-CAM(y1,y2) (10)

[0093] Wherein, y1 and y2 are the feature maps obtained in step A, and MS-CAM is a multi-scale channel attention.

[0094] As Figure 4 shown is a multi-scale feature aggregation (MFSA) module diagram constructed by the embodiment of the application, and the specific content includes but is not limited to the following content:

[0095] A. Take the output features (x 1 , x 2 , x 3 , x 4 ) of each layer of the encoder as input, downsample them to the same resolution size through cross-step convolution operation with different step lengths, then stack the features along the channel direction (i.e. Concatenation operation), and then perform channel self-adaptation on the spliced features, i.e. use channel attention (SE) to assign different weights to each channel, thereby obtaining the fused inter-layer multi-scale features.

[0096] The calculation process can be formulated as follows:

[0097]

[0098] where DownSampling represents the down-sampling operation, which is realized by the strided convolution in this network, which is beneficial to learn better sampling points, Concat represents the stacking along the channel direction, and SE represents the channel attention operation, X fuse is the fused multi-scale feature.

[0099] B, the fused feature X fuse is obtained in step A is locally modeled. First, two consecutive convolution blocks CB (convolution + normalization + nonlinear activation) are passed to obtain a local feature map, and then a locality constraint weight is generated by spatial attention SA, and the calculation formula is as follows:

[0100] ω local = SA(CB M (X fuse )) (12)

[0101] where ω local is the generated locality constraint weight.

[0102] C, the fused feature X fuse obtained in step A is globally modeled. Specifically, X fuse is first fed into the regular linear attention for global context information capture to obtain a feature containing global dependency, and then the locality weight generated in step B is multiplied with it to give it locality constraint to avoid excessive attention to globality by deep features, and the calculation formula is as follows:

[0103] Z = ω local ·CB(Xfus e ) (13)

[0104] where CB represents the proposed regular linear attention, including CBAM and 1x1 convolution, ω local is the locality constraint weight, and Z is the output feature after the final multi-scale context information aggregation.

[0105] The system, device, module or unit illustrated in the above embodiments can be specifically implemented by a computer chip or entity, or by a product with certain functions.

[0106] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can implement information storage by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0107] It should also be noted that the terms "comprising", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such a process, method, article or apparatus. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus including the element.

[0108] The above embodiments should be understood as only for illustrating the present application and not for limiting the protection scope of the present application. After reading the content of the present application, the skilled in the art can make various changes or modifications to the present application, and these equivalent changes and modifications also fall within the scope defined by the claims of the present application.

Claims

1. A brain tumor image segmentation method based on improved scale fusion, characterized in that, The method comprises the following steps: 4 multi-modal MR brain images containing tumor lesions are input into a brain tumor image segmentation network for brain tumor image segmentation to delineate three nested sub-regions of brain tumors: an edema region ET around the tumor, a whole tumor region WT, and a tumor core region TC; The global-local feature mixer GLFM is constructed to capture local detailed features and global context information in parallel, and to realize information mixing in global-local scales. The multi-scale expansion initial module MSDI is constructed to capture multi-scale deformation features effectively through expansion convolutions with different expansion rates, the unique structure of Inception, and the multi-scale channel attention module MS-CAM. The multi-scale feature aggregation module MFSA is constructed to integrate multi-level multi-resolution features into long-range dependency modeling to realize multi-scale global perception. The 4 multi-modal MR brain images containing tumor lesions are input into the brain tumor image segmentation network for brain tumor image segmentation to delineate the three nested sub-regions ET, WT, and TC of brain tumors, specifically as follows: The training images are input into the brain tumor image segmentation network, which comprises four GLFMs, wherein the first GLFM, the second GLFM, the MFSA, the third GLFM, and the fourth GLFM are connected in sequence. The brain tumor image segmentation network takes a 3D U-Net structure as a basic network structure, encodes and extracts features of an input image to obtain features of each layer , , , The first three layers of features pass through a multi-scale expansion initial module MSDI, the fourth layer of features passes through a multi-scale feature aggregation module MFSA, and then is spliced with a corresponding decoder to restore the lost detail information of down sampling, the decoder gradually restores the feature mapping to the size of the original input image through progressive up sampling and convolution operation, so that end-to-end brain tumor image segmentation is realized. The global-local feature mixer GLFM is constructed to capture local detailed features and global context information in parallel, and to realize information mixing in global-local scales. A. The input feature x is first evenly divided into two sub-feature maps along the channel dimension to obtain corresponding feature maps {x1, x2}; B、 Parallelly model global-local features for feature map x1,x2, get global feature f g and local Feature f l ; C. The global feature and the local feature are fused to obtain a fused mixed feature Y, and then the feature fusion enhancer TFE is used to obtain the final mixed feature; The multi-scale expansion initial module MSDI is constructed to capture multi-scale deformation features effectively through expansion convolutions with different expansion rates, the unique structure of Inception, and the multi-scale channel attention module MS-CAM. A1. The features y of the first three layers of the encoder are taken as multi-scale feature inputs, which are respectively subjected to four convolution operations, the first two layers are expansion convolutions with different expansion rates, and the last two layers are 1x1x1 convolutions and 1x1x1 convolutions with pool operations, respectively. B1, send y1, y2 into MS-CAM module, get final feature map y 最终 ; The multi-scale feature aggregation module MFSA is constructed to integrate multi-level multi-resolution features into long-range dependency modeling to realize multi-scale global perception. A2, the output features (x 1 , x 2 , x 3 , x 4 ) of each layer of the encoder are taken as input, down-sampled to the same resolution size through cross-step convolution operations of different step lengths, then stacked along the channel direction, i.e. Concatenation operation, and then channel self-adaptation is performed on the spliced features, i.e. channel attention SE is used to assign different weights to each channel, thereby obtaining the fused inter-layer multi-scale features; B2, to the fusion features obtained in step A2 Local modeling is performed, first through two successive convolutional blocks CB, which comprise a convolution + normalization + non-linear activation, to obtain local feature maps, then through a spatial attention SA to generate locality constraint weights; C2. To the fusion features obtained in step A2 Global modeling is performed.

2. The improved scale fusion brain tumor image segmentation method according to claim 1, wherein (1) represents uniform division of the input features along the channel dimension.

3. The improved scale fusion brain tumor image segmentation method according to claim 2, characterized in that, (8) = Pool(Conv(y)) + Conv(y) (9) wherein, in formula (8), DC represents dilated convolution, the first two DCs have a dilated rate of 1, and the last DC has a dilated rate of 2, and Conv represents 1x1x1 convolution; in formula (9), Pool represents a pooling operation. (10) 。 4. The improved scale fusion brain tumor image segmentation method according to claim 3, characterized in that, （11） wherein, represents a down-sampling operation, which is implemented by using a stride convolution in this network, It is beneficial to learn better sampling points, represents stacking along the channel direction, represents the channel attention operation, that is, the fused multi-scale features; The fusion features obtained in step A2 The global modeling specifically includes: The global context information is fed into the rule linear attention for capturing, and the features containing global dependency are obtained. Then, the locality constraint weight generated in step B2 is multiplied with the features to give them locality constraints, so as to avoid excessive attention of deep features to the globality.

5. An electronic device, comprising: A computer program product, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the improved scale fusion brain tumor image segmentation method according to any one of claims 1 to 4 when executing the program.

6. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program product, wherein the processor implements the improved scale fusion brain tumor image segmentation method according to any one of claims 1 to 4 when executing the program.

Images