Light-weight brain tumor segmentation method based on MSLQA-Net

Abstract

The invention relates to a lightweight brain tumor segmentation method based on MSLQA-Net. The method is applied to tumor segmentation of brain tumor magnetic resonance imaging (MRI). The method comprises the following steps: S1, preprocessing brain MRI data, and dividing the data into a training set, a verification set and a test set; s2, constructing a multi-scale decomposition residual convolution module (MSDRC) to perform feature extraction, and forming a multi-level secondary feature map; s3, constructing a cross-layer feature aggregation module (CLFA), and performing efficient aggregation on the multi-level secondary feature map to obtain an aggregated feature map; s4, constructing a quadruple-dimension attention module (QDA), and further refining the aggregated feature map to obtain a refined feature map; s5, fusing multi-scale decomposition residual convolution, cross-layer feature aggregation and a quadruple-dimension attention module, and constructing an MSLQA-Net segmentation model; and S6, performing training, verification optimization and testing on the segmentation model by using the training set, the verification set and the test set. According to the method, in a brain tumor segmentation task, a high-precision result can be still kept while the reasoning speed is increased.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a lightweight brain tumor segmentation method based on MSLQA-Net. Background Art

[0002] Brain tumors are abnormal cell groups growing in the brain, which seriously threaten the life and health of patients. Magnetic resonance imaging (MRI) can accurately display soft tissue structures and has become a key technology for clinicians to diagnose and treat gliomas. However, the complexity and diversity of gliomas make it a difficult task that consumes a lot of time and energy for experienced radiologists.

[0003] Traditional brain tumor segmentation methods, including machine learning-based algorithms, rely on manually designed features and are unable to fully utilize the rich contextual information in MRI images. With the rapid development of deep learning (DL) technology, especially the U-Net architecture based on convolutional neural networks (CNNs), it has shown excellent performance in image segmentation tasks. However, early CNN-based brain tumor segmentation methods usually use two-dimensional (2D) models, which require three-dimensional (3D) brain MRI images to be sliced ​​into 2D images for processing. This method lacks continuous information between slices and is difficult to capture the global context of the tumor. Although some studies have improved the model by introducing three-dimensional (3D) convolution kernels to solve the above problems, many improved methods have improved the segmentation performance to a certain extent, but there are still significant limitations.

[0004] At present, although advanced deep learning models have made significant progress in the field of brain tumor segmentation, they generally have problems such as large parameter scale, low training and inference efficiency, and it is difficult to meet the demand for real-time segmentation results in clinical scenarios. To meet this challenge, many lightweight network architectures have emerged in recent years, aiming to reduce model complexity and computational overhead while maintaining high segmentation accuracy. However, due to the shallow network depth and limited number of channels, these lightweight models often lack the ability to extract features and are difficult to fully capture complex semantic information. This deficiency leads to inaccurate segmentation or blurred boundaries when the model is dealing with small target areas, detailed features, and complex boundaries. In addition, lightweight models lack sufficient adaptability and generalization ability when dealing with the heterogeneity and diversity between different MRI modalities, which further limits their stability and reliability in actual clinical applications. Therefore, it is urgent to develop an automated brain tumor segmentation method that can not only ensure high segmentation accuracy, but also have fast response capabilities and efficient computing performance to meet the clinical requirements for real-time and accuracy. This can not only significantly reduce the workload of radiologists and improve diagnostic efficiency, but also provide timely and accurate segmentation results in emergency situations, thereby improving patients' treatment outcomes and survival rates. Summary of the invention

[0005] This paper mainly aims at the problem that the 3D U-shaped architecture network model has a complex model structure and a large number of parameters, which leads to slow model training in processing 3D brain tumor image segmentation tasks. A lightweight brain tumor segmentation method based on MSLQA-Net is proposed. By optimizing the 3D U-shaped architecture network model and using a lightweight multi-scale decomposition residual convolution module to lightweight improve the traditional architecture, the parameters can be significantly reduced while effectively improving the segmentation performance. The cross-layer feature aggregation module is used on the jump connection to efficiently fuse the multi-level feature maps, and then the fused features are further refined using the quadruple dimensional attention module, thereby improving the accuracy of brain tumor segmentation.

[0006] Specifically, the method comprises the following steps:

[0007] S1: Preprocess brain MRI data and divide the data into training set, validation set and test set;

[0008] S2: Construct a multi-scale decomposition residual convolution module (MSDRC) for feature extraction to form a multi-level feature map;

[0009] S3: Construct a cross-layer feature aggregation module (CLFA) to efficiently aggregate multi-level feature maps to obtain aggregated feature maps;

[0010] S4: Construct a quadruple dimensional attention module (QDA) to further refine the aggregated feature map to obtain a refined feature map;

[0011] S5: Integrate multi-scale decomposition residual convolution, cross-layer feature aggregation and four-dimensional attention module to build the MSLQA-Net segmentation model;

[0012] S6: Use the training set, validation set and test set to train, validate, optimize and test the segmentation model.

[0013] Preferably, S1 comprises the following steps:

[0014] S1.1: Preprocess the input data by standardizing, cropping, and data augmentation;

[0015] S1.2: Divide the preprocessed data into training set, validation set and test set.

[0016] Preferably, S2 comprises the following steps:

[0017] S2.1: Three feature extraction branches are constructed using three sizes of convolution kernels;

[0018] S2.2: Decompose and refine features on each branch;

[0019] S2.3: The features of the three branches are integrated to construct a multi-scale decomposition residual convolution module.

[0020] Preferably, S2.2 comprises the following steps:

[0021] A traditional 3D convolution kernel is decomposed into two pseudo 3D convolution kernel connections. A residual connection is used between the two decomposed convolution blocks to prevent information loss during feature extraction. Finally, a 1×1×1 convolution is used to further refine the features. The feature extraction process can be described as:

[0022] F i =Conv 1×1×1 (Conv 1×i×i (Conv i×i×1 (F))+Conv i×i×1 (F)

[0023] Preferably, S2.3 comprises the following steps:

[0024] The features of the three branches are added and fused pixel by pixel. The operation can be described as:

[0025] F out =F i1 +F i2 +F i3

[0026] F out is the output of the module.

[0027] Preferably, S3 comprises the following steps:

[0028] S3.1: Aggregate shallow and middle layer features to generate preliminary aggregated features;

[0029] S3.2: Further aggregate the middle-level, preliminary aggregation, and deep-level features to build a cross-layer feature aggregation module.

[0030] Preferably, S3.1 comprises the following steps:

[0031] For shallow features F s and the middle-level features F c Use convolution operation to map shallow features to F c The feature map of the same size is then c Add pixel by pixel, integrate features and reduce the number of channels through convolution operation, calculate image attention using Sigmoid operation, and finally distribute the attention map to F by element-by-element multiplication. c Graphically, this operation can be described as:

[0032]

[0033] Among them, σ1 represents the Relu operation, σ2 represents the Sigmoid operation, Here, i represents the convolution kernel size, m represents the number of input channels, n represents the number of output channels, and F c ' is the fusion feature.

[0034] Preferably, S3.2 comprises the following steps:

[0035] F c ', F c and F d Using the convolution operation, F c ' and F c Mapped to F d The feature map of the same size, then F c ', F c and F d Add pixel by pixel, integrate features and reduce the number of channels through convolution operation, transpose convolution on the feature map to ensure that the output and input feature maps are of the same size, use Sigmoid operation to calculate image attention, and finally distribute the attention map to F c 'On the image, the calculation process is as follows:

[0036]

[0037] Among them, Conv Trans is the transposed convolution, F c " is the final fusion result.

[0038] Preferably, S4 comprises the following steps:

[0039] S4.1: Transpose the input features in different directions to form four branches;

[0040] S4.2: Attention modeling is performed on four dimensions on four branches to generate enhanced features;

[0041] S4.3: Fuse the enhanced features of the four branches to construct a four-dimensional attention module.

[0042] Preferably, S4.2 comprises the following steps:

[0043] Model attention in four dimensions: axial, coronal, sagittal, and channel. Taking the first branch as an example, the attention modeling in the axial direction is performed. The input feature is F A , the features are pooled by average pooling and standard deviation pooling to obtain two pooling results F avg and F std , and adaptively aggregate them to generate a new feature tensor F A '. The operation can be described as:

[0044] F avg =AvgPool(F A ),F std =StdPool(F A )

[0045]

[0046] Where α and β are trainable parameters in the range of (0, 1). Convolution is used to capture the interaction between features in the dimension, and then the Sigmoid activation function is used to generate the attention weight A in the axial direction. A , and finally, the attention weight A A By element-wise multiplication in F A ', thus obtaining the enhanced feature map on the axis The operation can be described as:

[0047] F A =Conv(F A ')

[0048] A A =σ(F A ”)

[0049]

[0050] σ represents the Sigmoid operation, and ⊙ represents pixel-by-pixel multiplication. The operations of the other three branches are the same as those in the axial direction. Attention modeling is performed in the coronal, sagittal, and channel dimensions, and finally the attention map is applied to the corresponding feature map to generate enhanced features.

[0051] Preferably, S4.3 comprises the following steps:

[0052] The four branches are fused through an average summation operation to obtain the output of the module. This operation can be described as:

[0053]

[0054] in, and They are represented as the features after attention enhancement in the coronal, sagittal and channel directions, respectively, and F' represents the output of the entire module.

[0055] Preferably, S5 comprises the following steps:

[0056] Taking the 3D U-shaped network as the basic architecture, the multi-scale decomposition residual convolution module is embedded in the encoder and decoder to extract features, a cross-layer feature aggregation module is added to the jump connection to aggregate multi-layer features, and a quadruple-dimensional attention module is used in the decoder to enhance the aggregated features to form the MSLQA-Net model.

[0057] Preferably, S6 comprises the following steps:

[0058] Use the training set and validation set to train and optimize the model to obtain the optimal model weight, and then use the test set to test the model.

[0059] The beneficial effects of the method of the present invention are as follows: first, in the data preprocessing stage, by standardizing, cropping, data enhancement and other operations on the collected brain tumor MR images, the data quality is improved, the interference of irrelevant information is reduced, and the data is reasonably divided into training set, validation set and test set, laying a good foundation for subsequent model training and testing. In the process of model construction, a multi-scale decomposition residual convolution is innovatively designed, which not only significantly reduces the amount of model parameters, but also provides richer feature information. Then, a cross-layer feature aggregation module is proposed, which realizes the efficient fusion of robust semantic information in low-resolution feature maps and rich spatial information in high-resolution feature maps while introducing fewer parameters. Subsequently, a four-dimensional attention module is proposed to capture the correlation between space and channels from the four dimensions of MR images. Finally, the MSLQA-Net model is constructed based on the 3D-UNet network. These innovative designs enable the model to maintain a high level of segmentation accuracy while significantly reducing the amount of parameters and accelerating the inference speed, providing an efficient and reliable solution for brain tumor segmentation tasks, showing good clinical application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the prior art and the drawings required for use in the embodiments. The following drawings are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0061] Figure 1 It is a flowchart of a lightweight brain tumor segmentation method based on MSLQA-Net of the present invention;

[0062] Figure 2 This is a model architecture diagram of a lightweight brain tumor segmentation method based on MSLQA-Net of the present invention;

[0063] Figure 3It is a schematic diagram of a multi-scale decomposition residual convolution structure of a lightweight brain tumor segmentation method based on MSLQA-Net of the present invention;

[0064] Figure 4 It is a schematic diagram of a cross-layer feature aggregation structure of a lightweight brain tumor segmentation method based on MSLQA-Net of the present invention;

[0065] Figure 5 It is a schematic diagram of a four-dimensional attention structure of a lightweight brain tumor segmentation method based on MSLQA-Net of the present invention;

[0066] Figure 6 This is a comparison chart of the segmentation results of a specific embodiment of the present invention on a brain tumor dataset with other lightweight deep learning networks. Specific implementation plan

[0067] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all other embodiments obtained by ordinary technicians in this field without creative work based on the embodiments of the present invention are within the scope of protection of the present invention.

[0068] The embodiment of the present invention provides a lightweight brain tumor segmentation method based on MSLQA-Net, which is used to achieve efficient segmentation of brain tumors, assist clinical diagnosis and treatment decision-making, and improve the accuracy and efficiency of medical image processing.

[0069] Reference Figure 1 , the method comprises the following steps:

[0070] S1: Preprocess brain MRI data and divide the data into training set, validation set and test set;

[0071] S2: Construct a multi-scale decomposition residual convolution module (MSDRC) for feature extraction to form a multi-level feature map;

[0072] S3: Construct a cross-layer feature aggregation module (CLFA) to efficiently aggregate multi-level feature maps to obtain aggregated feature maps;

[0073] S4: Construct a quadruple dimensional attention module (QDA) to further refine the aggregated feature map to obtain a refined feature map;

[0074] S5: Integrate multi-scale decomposition residual convolution, cross-layer feature aggregation and four-dimensional attention module to build the MSLQA-Net model;

[0075] S6: Use the training set, validation set, and test set to train, validate, optimize, and test the model.

[0076] Further, S1 comprises the following steps:

[0077] S1.1: Preprocess the input data by standardizing, cropping, and data augmentation;

[0078] Furthermore, the step of preprocessing the input data by standardizing, center cropping, randomly flipping and randomly rotating specifically includes:

[0079] The acquired data is a public dataset. Each case image completely covers the entire area of ​​the brain. The MRI data is cropped to 128×128×128 size after removing irrelevant areas. The Z-score method is then used to standardize each image. The images are then randomly flipped and rotated to enrich the diversity of the images.

[0080] S1.2: Divide the preprocessed data into training set, validation set and test set.

[0081] Furthermore, the step of dividing the preprocessed data into a training set, a validation set, and a test set specifically includes:

[0082] The MRI data were divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0083] Further, refer to Figure 3 , S2 includes the following steps:

[0084] S2.1: Three feature extraction branches are constructed using three sizes of convolution kernels;

[0085] S2.2: Decompose and refine features on each branch;

[0086] S2.3: The features of the three branches are integrated to construct a multi-scale decomposition residual convolution module.

[0087] Further, S2.2 includes the following steps:

[0088] A traditional 3D convolution kernel is decomposed into the connection of two pseudo 3D convolution kernels, that is, the 3×3×3 convolution kernel is decomposed into 3×3×1 and 1×3×3 convolution kernels, and the 5×5×5 convolution kernel is decomposed into 5×5×1 and 1×5×5 convolution kernels. Between the two decomposed convolution blocks, residual connection is used to prevent information loss during feature extraction. Finally, a 1×1×1 convolution is used to further refine the features. The feature extraction of the three branches can be described as:

[0089] F1=Conv1×1×1 (F)

[0090] F2=Conv 1×1×1 (Conv 1×3×3 (Conv 3×3×1 (F))+Conv 3×3×1 (F)

[0091] F3=Conv 1×1×1 (Conv 1×5×5 (Conv 5×5×1 (F))+Conv 5×5×1 (F)) Further, S2.3 includes the following steps:

[0092] The features of the three branches are added and fused pixel by pixel. The operation can be described as:

[0093] F out =F1+F2+F3

[0094] F out is the output of the module.

[0095] Further, refer to Figure 4 , S3 includes the following steps:

[0096] S3.1: Aggregate shallow and middle layer features to generate preliminary aggregated features;

[0097] S3.2: Further aggregate the middle-level, preliminary aggregation, and deep-level features to build a cross-layer feature aggregation module.

[0098] Further, S3.1 includes the following steps:

[0099] For shallow features F s and the middle-level features F c Use convolution operation to map shallow features to F c The feature map of the same size is then c Add pixel by pixel, integrate features and reduce the number of channels through convolution operation, calculate image attention using Sigmoid operation, and finally distribute the attention map to F by element-by-element multiplication. c Graphically, this operation can be described as:

[0100]

[0101] Among them, σ1 represents the Relu operation, σ2 represents the Sigmoid operation, Here, i represents the convolution kernel size, m represents the number of input channels, n represents the number of output channels, and F c ' is the fusion feature.

[0102] Further, S3.2 includes the following steps:

[0103] F c ', F c and F d Using the convolution operation, F c ' and F c Mapped to F d The feature map of the same size, then F c ', F c and F d Add pixel by pixel, integrate features and reduce the number of channels through convolution operation, transpose convolution on the feature map to ensure that the output and input feature maps are of the same size, use Sigmoid operation to calculate image attention, and finally distribute the attention map to F c 'On the image, the calculation process is as follows:

[0104]

[0105] Among them, Conv Trans is a transposed convolution with a kernel size of 2 and a stride of 2, F c " is the final fusion result.

[0106] Further, refer to Figure 5 , S4 comprises the following steps:

[0107] S4.1: Transpose the input features in different directions to form four branches;

[0108] S4.2: Attention modeling is performed on four dimensions on four branches to generate enhanced features;

[0109] S4.3: Fuse the enhanced features of the four branches to construct a four-dimensional attention module.

[0110] Further, S4.2 includes the following steps:

[0111] Model attention in four dimensions: axial, coronal, sagittal, and channel. Taking the first branch as an example, the attention modeling in the axial direction is performed. The input feature is F A , the features are pooled by average pooling and standard deviation pooling to obtain two pooling results F avg and F std , and adaptively aggregate them to generate a new feature tensor F A '. The operation can be described as:

[0112] F avg =AvgPool(F A ),F std =StdPool(F A)

[0113]

[0114] Where α and β are trainable parameters in the range of (0, 1). A convolution kernel of size 1×1×3 is used to capture the interaction between features in the dimension, and then a Sigmoid activation function is used to generate the attention weight A in the axial direction. A , and finally, the attention weight A A By element-wise multiplication in F A ', thus obtaining the enhanced feature map on the axis The operation can be described as:

[0115] F A =Conv 1×1×3 (F A ')

[0116] A A =σ(F A ”)

[0117]

[0118] σ represents the Sigmoid operation, and ⊙ represents pixel-by-pixel multiplication. The operations of the other three branches are the same as those in the axial direction. Attention modeling is performed in the coronal, sagittal, and channel dimensions, and finally the attention map is applied to the corresponding feature map to generate enhanced features.

[0119] Further, S4.3 includes the following steps:

[0120] The four branches are fused through an average summation operation to obtain the output of the module. This operation can be described as:

[0121]

[0122] in, and They are represented as the features after attention enhancement in the coronal, sagittal and channel directions, respectively, and F' represents the output of the entire module.

[0123] Further, refer to Figure 2 , S5 comprises the following steps:

[0124] Taking the 3D U-shaped network as the basic architecture, the multi-scale decomposition residual convolution module is embedded in the encoder and decoder to extract features, a cross-layer feature aggregation module is added to the jump connection to aggregate multi-layer features, and a quadruple-dimensional attention module is used in the decoder to enhance the aggregated features to form the MSLQA-Net model.

[0125] Further, S6 comprises the following steps:

[0126] Use the training set and validation set to train and optimize the model to obtain the optimal model weights, then import the optimal model weights into the test code and use the test set to test the model.

[0127] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make other equivalent modifications or substitutions without violating the spirit of the invention, and these equivalent modifications or substitutions are included in the scope defined by the application claims.

Claims

1. A lightweight brain tumor segmentation method based on MSLQA-Net, characterized in that: The method comprises the following steps: S1: Preprocess brain MRI data and divide the data into training set, validation set and test set; S2: Construct a multi-scale decomposition residual convolution module (MSDRC) for feature extraction to form a multi-level feature map; S3: Construct a cross-layer feature aggregation module (CLFA) to efficiently aggregate multi-level feature maps to obtain aggregated feature maps; S4: Construct a quadruple dimensional attention module (QDA) to further refine the aggregated feature map to obtain a refined feature map; S5: Integrate multi-scale decomposition residual convolution, cross-layer feature aggregation and four-dimensional attention module to build the MSLQA-Net segmentation model; S6: Use the training set, validation set and test set to train, validate, optimize and test the segmentation model.

2. According to claim 1, a lightweight brain tumor segmentation method based on MSLQA-Net is characterized in that: S1 includes the following steps: S1.1: Preprocess the input data by standardizing, cropping, and data augmentation; S1.2: Divide the preprocessed data into training set, validation set and test set.

3. The lightweight brain tumor segmentation method based on MSLQA-Net according to claim 1, characterized in that: S2 includes the following steps: S2.1: Three feature extraction branches are constructed using three sizes of convolution kernels; S2.2: Decompose and refine features on each branch; S2.3: The features of the three branches are integrated to construct a multi-scale decomposition residual convolution module. Wherein, S2.2 comprises the following steps: A traditional 3D convolution kernel is decomposed into two pseudo 3D convolution kernel connections. A residual connection is used between the two decomposed convolution blocks to prevent information loss during feature extraction. Finally, a 1×1×1 convolution is used to further refine the features. The feature extraction process can be described as: F i =Conv 1×1×1 (Conv 1×i×i (Conv i×i×1 (F))+Conv i×i×1 (F)) Among them, S2.3 includes the following steps: The features of the three branches are added and fused pixel by pixel. The operation can be described as: F out =F i1 +F i2 +F i3 F out is the output of the module.

4. The lightweight brain tumor segmentation method based on MSLQA-Net according to claim 1, characterized in that: S3 includes the following steps: S3.1: Aggregate shallow and middle layer features to generate preliminary aggregated features; S3.2: Further aggregate the middle-level, preliminary aggregation, and deep-level features to build a cross-layer feature aggregation module. Among them, S3.1 includes the following steps: For shallow features F s and the middle-level features F c Use convolution operation to map shallow features to F c The feature map of the same size is then c Add pixel by pixel, integrate features and reduce the number of channels through convolution operation, calculate image attention using Sigmoid operation, and finally distribute the attention map to F by element-by-element multiplication. c Graphically, this operation can be described as: Among them, σ1 represents the Relu operation, σ2 represents the Sigmoid operation, Here, i represents the convolution kernel size, m represents the number of input channels, n represents the number of output channels, and F c ' is the fusion feature. Among them, S3.2 includes the following steps: F c ', F c and F d Using the convolution operation, F c ' and F c Mapped to F d The feature map of the same size, then F c ', F c and F d Add pixel by pixel, integrate features and reduce the number of channels through convolution operation, transpose convolution on the feature map to ensure that the output and input feature maps are of the same size, use Sigmoid operation to calculate image attention, and finally distribute the attention map to F c 'On the image, the calculation process is as follows: Among them, Conv Trans is the transposed convolution, F c " is the final fusion result.

5. The lightweight brain tumor segmentation method based on MSLQA-Net according to claim 1, characterized in that: S4 includes the following steps: S4.1: Transpose the input features in different directions to form four branches; S4.2: Attention modeling is performed on four dimensions on four branches to generate enhanced features; S4.3: Fuse the enhanced features of the four branches to construct a four-dimensional attention module. Among them, S4.2 includes the following steps: Model attention in four dimensions: axial, coronal, sagittal, and channel. Taking the first branch as an example, the attention modeling in the axial direction is performed. The input feature is F A , the features are pooled by average pooling and standard deviation pooling to obtain two pooling results F avg and F std , and adaptively aggregate them to generate a new feature tensor F A '. The operation can be described as: F avg =AvgPool(F A ),F std =StdPool(F A ) Where α and β are trainable parameters in the range of (0, 1). Convolution is used to capture the interaction between features in the dimension, and then the Sigmoid activation function is used to generate the attention weight A in the axial direction. A , and finally, the attention weight A A By element-wise multiplication in F A ', thus obtaining the enhanced feature map on the axis The operation can be described as: F A ”=Conv(F A ’) A A =σ(F A ") σ represents the Sigmoid operation, and ⊙ represents pixel-by-pixel multiplication. The operations of the other three branches are the same as those in the axial direction. Attention modeling is performed in the coronal, sagittal, and channel dimensions, and finally the attention map is applied to the corresponding feature map to generate enhanced features. Among them, S4.3 includes the following steps: The four branches are fused through an average summation operation to obtain the output of the module. This operation can be described as: in, and They are represented as the features after attention enhancement in the coronal, sagittal and channel directions, respectively, and F' represents the output of the entire module.

6. The lightweight brain tumor segmentation method based on MSLQA-Net according to claim 1, characterized in that: S5 includes the following steps: Taking the 3D U-shaped network as the basic architecture, the multi-scale decomposition residual convolution module is embedded in the encoder and decoder to extract features, a cross-layer feature aggregation module is added to the jump connection to aggregate multi-layer features, and a quadruple-dimensional attention module is used in the decoder to enhance the aggregated features to form the MSLQA-Net model.

7. The lightweight brain tumor segmentation method based on MSLQA-Net according to claim 1, characterized in that: S6 includes the following steps: Use the training set and validation set to train and optimize the model to obtain the optimal model weight, and then use the test set to test the model.

Images