A brain tumor automatic segmentation method

Abstract

The application discloses an automatic brain tumor segmentation method applied to the field of medical image processing and aiming at the problem of low accuracy of existing brain tumor segmentation. The three-axis strip pooling method is used to compress the whole 3D image from the perspective of three axes in cooperation with three learnable weight vectors, and the three axis information is utilized to complement each other at the end, so that the spatial information is retained to the maximum extent, and the purpose of reducing the calculation amount and the memory burden is achieved. An information entropy attention module is designed, the region with the largest information amount is obtained by comparing the information entropy, and then the region is enhanced. The noise and interference of the intracranial complex environment are greatly reduced, and the segmentation performance of the model is further improved. A curvature filtering attention is also designed, and a learnable convolution is used to extract the tumor edge shape, so that the edge of the model of the application is most clearly depicted in the final segmentation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical image processing, and particularly relates to a brain tumor segmentation technology. BACKGROUND

[0002] Brain tumors are common intracranial tumor types in adults, including both benign and malignant. Its incidence has gradually increased in recent years, causing a significant impact on patients' health and quality of life. Brain tumors can originate from brain tissue, meninges, neural tissue or other intracranial structures, and have diversity and complexity. According to the classification of the World Health Organization (WHO), the classification of brain tumors ranges from grade I (benign) to grade IV (malignant). Brain tumors can cause a series of symptoms such as headache, seizures, cognitive impairment or loss of motor function by directly compressing brain tissue or obstructing cerebrospinal fluid circulation.

[0003] Clinical practice usually surgically removes tumors that cause symptoms and continue to grow. Given the diversity of brain tumor size and location, it is crucial to achieve rapid and accurate positioning in clinical practice, which requires doctors to accurately judge the location and size of brain tumors. In the past, preoperative segmentation of brain tumors usually used manual segmentation, with single-point labeling by experienced doctors, but this manual segmentation not only consumes time, but also is highly dependent on the professional level of doctors, so it is prone to errors, which is not conducive to the treatment and recovery of patients, and consumes the energy of doctors while lengthening the diagnosis and treatment cycle.

[0004] Manual segmentation of brain tumors not only consumes time, but also is highly dependent on the professional level of doctors, which not only requires doctors to have solid dermatological knowledge, but also involves a large amount of human cost. Most importantly, the professional ability of doctors directly affects the diagnosis effect of patients, and the number of excellent doctors is limited, which cannot meet the increasing demand of patients in time. Therefore, brain tumor automatic segmentation methods have emerged. Magnetic resonance imaging is the main data for brain tumor segmentation, which is a 3D medical image. If the entire image is put into segmentation, it may result in high computing cost, and thus low computing efficiency.

[0005] In view of the above shortcomings, an efficient and rapid brain tumor automatic segmentation method has become the mainstream segmentation method. Many previous studies have shown that by using an encoder-decoder structure to extract image features and classify pixels in 3D medical images, the U-shaped architecture has good performance. Over the years, people have proposed more U-shaped model structures. However, these studies mainly focus on improving the accuracy of tumor segmentation, while ignoring the complexity of the model. Therefore, the equipment requirements become higher, which puts a heavy burden on memory resources, and these model structures often lack effective plug-and-play functions, limiting their transferability.

[0006] Many previous studies have shown that image features are extracted by using an encoder-decoder U-shaped structure and classifying pixels in 3D medical images. Over the years, numerous U-shaped model structures have been proposed. However, these studies mainly focus on improving the accuracy of tumor segmentation, while ignoring the complexity of the model. Therefore, the device requirements become higher, which puts a heavy burden on memory resources, and these model structures often lack effective plug-and-play functionality, limiting their transferability.

[0007] In order to enhance the feature extraction ability and improve the segmentation accuracy, various attention methods have been proposed in recent years and widely applied in 3D medical image segmentation tasks. Although these attention methods improve the segmentation accuracy, they still require a large amount of memory and exhibit high computational complexity. This is because each voxel is processed individually when performing 3D medical image segmentation tasks, which results in high computational cost.

[0008] In order to solve the challenge of high computational cost, many researchers have adopted the strategy of cutting 3D brain tumor images into a series of 2D slices, which has achieved remarkable results in practice. By simplifying complex three-dimensional data into more easily processed two-dimensional form, not only greatly reduces the computational burden, but also makes the analysis process more efficient and intuitive, providing strong support for the diagnosis and treatment of brain tumors. However, this strategy brings convenience, but also brings a problem that cannot be ignored: some important three-dimensional information may be inevitably lost during the slicing process. Specifically, the three-dimensional morphology of the tumor and its fine spatial relationship with the surrounding tissue are difficult to be completely preserved after being transformed into 2D slices. This loss of spatial information may affect the doctor's accurate understanding of the overall picture of the tumor and the precise planning of the surgical path and resection range.

[0009] In summary, the existing brain tumor segmentation methods mainly have the following problems:

[0010] 1. The use of 2D slices leads to the loss of spatial information, affecting the final segmentation accuracy;

[0011] 2. The use of 3D attention requires a high computational cost, resulting in low computational efficiency. SUMMARY

[0012] To solve the above technical problems, the present application proposes an automatic brain tumor segmentation method, which considers accuracy and reasoning efficiency, and designs a series of novel compression and feature extraction modules to reduce computational complexity while improving segmentation accuracy.

[0013] The technical scheme adopted by the present application is as follows: an automatic brain tumor segmentation method, comprising:

[0014] S1, acquire a magnetic resonance scan image with a labeled segmentation label;

[0015] S2, construct a segmentation model, the segmentation model adopts a hierarchical encoder and decoder structure; the output result of each layer of the encoder is taken as the input of a novel lightweight attention mechanism module based on information entropy-curvature basic attention of the corresponding layer of the encoder, and the output of the novel lightweight attention mechanism module based on information entropy-curvature basic attention is taken as the input of the next layer of the encoder and the decoder of the corresponding layer after down-sampling;

[0016] The input of each layer of the decoder also includes the result of the output of the lower layer of the decoder after up-sampling;

[0017] The novel lightweight attention mechanism module based on information entropy-curvature basic attention includes a three-axis strip pooling unit, an information entropy attention unit and a curvature filtering attention unit; the three-axis strip pooling unit is used for compressing the input image, the information entropy attention unit is used for eliminating noise interference in the compressed image, and the curvature filtering attention unit is used for extracting local features of the image processed by the information entropy attention unit;

[0018] S3, the magnetic resonance scan image with the labeled segmentation label acquired in step S1 is used to train the segmentation model in step S2;

[0019] S4, input the magnetic resonance scan image of the object to be segmented without a labeled segmentation label into the trained segmentation model to obtain a brain tumor segmentation result.

[0020] The present application has the following advantages: the three-axis strip pooling method is used to compress the entire 3D image from the perspective of three axes, and the three-axis information is used to complement each other at the end, so that the spatial information is retained to the maximum extent, and the purpose of reducing the calculation amount and the memory burden is achieved; the information entropy attention module is designed, the region with the largest information amount is obtained by comparing the information entropy, and then the region is enhanced, so that the noise interference of the intracranial complex environment is greatly reduced, and the segmentation performance of the model is further improved; the curvature filtering attention is also designed, the tumor edge shape is extracted through the learnable convolution, so that the edge of the model is most clearly described in the final segmentation. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 Different brain tumor image size shape position examples are shown;

[0022] Figure 2 The overall architecture of the present application is shown;

[0023] Figure 3 The overall design of the model is shown;

[0024] Figure 4 is a three-axis strip pooling structure diagram;

[0025] Figure 5 is a curvature filtering attention structure diagram;

[0026] Figure 6 is a visualization result on the BraTS2021 dataset;

[0027] Figure 7 is a heat map visualization result of the method of the present application. DETAILED DESCRIPTION

[0028] In order to facilitate those skilled in the art to understand the technical content of the present application, the content of the present application is further explained below in conjunction with the drawings.

[0029] Segmentation of brain tumors is an important reference and powerful guide for doctors before and during surgery, and automatic brain tumor segmentation based on deep learning brings great convenience to doctors during surgery and can maximize the prevention of incomplete segmentation and misjudgment of brain tumors. However, in balancing segmentation accuracy and computational cost, it is difficult to achieve satisfactory results according to existing methods. At the same time, due to the following characteristics of brain tumor images:

[0030] 1. Brain tumor images are 3D medical images, and the required memory burden and computational cost are high when training automatic tumor segmentation, which affects the inference efficiency of the entire model;

[0031] 2. The size, location and shape of brain tumors are different and may differ greatly, which makes it difficult to segment;

[0032] 3. The growth environment of brain tumors is extremely complex, and tumors are often closely adjacent to surrounding brain tissue, which makes it necessary to accurately distinguish the boundary between tumor tissue and normal brain tissue during segmentation, which may cause noise interference to automatic segmentation;

[0033] Due to the above characteristics, when using 3D brain tumor image data for segmentation, the situation of incomplete segmentation and segmentation error may occur, resulting in poor segmentation effect. Figure 1 Different position, size, shape of brain tumor images are shown. As shown in Figure 1 the segmentation result is divided into three regions, which are whole tumor (WT), tumor core (TC), and enhanced tumor (ET).

[0034] The present application re-evaluates the challenge of improving inference efficiency while maintaining accuracy, and proposes a new lightweight 3D attention mechanism based on entropy-curvature based attention (ECA), which proposes a pooling method and two spatial attention modules according to the characteristics of brain tumors themselves, effectively reduces the computational complexity while capturing spatial feature information, and improves the parameter efficiency and inference accuracy. The mechanism includes three components: a tri-axial strip pooling method (TSP) and two cascading designed attention modules, namely entropy guided attention (EGA) and curvature filter attention (CFA). TSP is used to compress 3D medical images to reduce computational load and memory requirements. EGA aims to eliminate noise interference in the features after TSP, while CFA aims to effectively extract local features from the image.

[0035] The present application proposes a complete automatic brain tumor segmentation method, which includes four parts in total: 1. Data preparation, 2. Model design, 3. Model training, and 4. Model testing. The overall structure relationship diagram is as shown in Figure 2

[0036] The implementation process of each part is described in detail as follows:

[0037] 1. Data preparation:

[0038] This part is mainly responsible for collecting brain tumor images required for training the model, and pre-processing the data before starting the model training to make the data meet the requirements of the model training. After that, the collected brain tumor images will be divided into training set and test set according to a certain proportion, wherein the training set is used for training the model, and the test set is used for testing the final performance of the trained model. Specifically, the following steps are included:

[0039] 11. Obtain data and labels

[0040] The present application uses the BraTS2021 dataset, a total of 1237 cases for experimental verification. Each case includes four modalities: T1, T1ce, T2 and FLAIR. These modalities provide different image information: T1 weighted imaging is used to display anatomical structure, T1ce is the contrast enhancement of T1, which can highlight the enhanced area of the tumor, T2 weighted imaging shows the edema range of the tumor and the surrounding area, and FLAIR imaging is used to more clearly display the abnormal tissue area. The present application combines the four modalities to form a 4-channel image input, and then realizes accurate segmentation and analysis of brain tumors, thereby verifying the effectiveness and robustness of the model.​

[0041] 12、Data preprocessing

[0042] Data preprocessing is mainly to preprocess the image data in the data set before training the model, so that it can meet the requirements of model training. In this application, the data in the BraTS2021 data set is in 5-dimensional format (batch size x picture depth x picture length x picture width), which is adjusted to 2 x 4 x 256 x 256 x 256 before being put into the model, where the batch size is 2, representing processing two samples at a time of forward propagation and back propagation, the channel number is 4, representing four modalities, and the picture depth, length and width are all 256. Such preprocessing ensures the uniformity of the data format and lays a foundation for efficient training of the model.

[0043] 13、Dataset division

[0044] Dataset division is to divide the dataset into training set and test set according to a certain proportion, which are used for training model and testing model respectively. In this application, the BraTS2021 dataset is divided, the training set contains 700 cases, the validation set contains 299 cases, and the test set contains 252 cases.

[0045] 2、Model design:

[0046] This part is the core content of the application and is the key protection part of the application, including the specific design of three-axis strip pooling, information entropy attention and curvature filter attention.

[0047] The application proposes a segmentation architecture, and embeds the designed attention ECA into the architecture, which includes an encoder with four encoding layers and a decoder with four decoding layers, and ECA is used in each encoding layer.

[0048] The application processes two 3x3x3 convolution operations with a stride of 1 as a convolution layer operation.

[0049] Each layer of the encoder is input into ECA after being processed by a convolution layer, and then down-sampling operation is completed by 3D max pooling, so as to realize doubling of the channel number and halving of the image feature map resolution. In the first encoding layer, the input image with a size of 2x4x256x256 is directly used as the input of the first encoding layer.

[0050] In the decoder of each layer, a convolutional layer is used to process the combination of image features from the skip connection and the lower layer transmission, and then 3D transposed convolution is used to complete the upsampling operation, so that the resolution of the image features output by the layer decoder is doubled to be consistent with the resolution of the image feature map transmitted by the skip connection of the upper layer. After combining the image with the image of the skip connection, the image is processed by a convolutional layer and then continues to be upsampled, so that the original resolution is gradually restored. As shown in Figure 3 After the last upsampling is completed, dimension reduction is performed through 1x1x1 convolution processing to obtain a segmentation image result of 2x4x256x256. Figure 3 In the formula, Input represents input, Output represents output, Skip-connection represents skip connection, and concat is the abbreviation of concatenate, which means merging and integrating.

[0051] The output result of the last encoding layer is processed by ECA, 3D maximum pooling is used to complete the downsampling operation, and then a convolutional layer is used for processing, and 3D transposed convolution is used to complete the upsampling. After integrating the result of the last encoding output result processed by ECA, the result is used as the input of the last decoding layer.

[0052] Each ECA module is composed of three sub-modules: TSP, EGA and CFA, and these sub-modules are arranged in series. The model design diagram of the whole method is as shown in Figure 3 .

[0053] 21. Three-axis strip pooling

[0054] At present, the commonly used pooling methods in neural networks include maximum pooling, average pooling and spatial pyramid pooling. However, the maximum pooling only pays attention to significant information, which may lead to the loss of detailed information; the average pooling emphasizes global information, but may lose important image features; the spatial pyramid pooling can process inputs of different sizes, but increases the computational complexity.

[0055] In order to fundamentally reduce the memory consumption in the process of 3D medical image segmentation, the present application proposes three-axis strip pooling, as shown in Figure 4 .

[0056] In order to maintain the spatial information of three dimensions, the data from the sagittal plane, the coronal plane and the transverse plane are compressed respectively. TSP applies weights through independent learnable vectors to emphasize important features in each plane, so that unimportant areas can be ignored and tumor areas can be highlighted. For example, the pooling operation on the transverse plane can be represented by the following formula:

[0057]

[0058] Y = max (0, x + a) (1) DWdenotes the result after pooling along the cross-section, H denotes the height of the plane, and W denotes the width of the plane H is a learnable weight matrix on the H axis, A i is a component on the height i corresponding to the input feature map. Through the above formula, the pooling operation is performed on the three planes to obtain three two-dimensional feature maps, which capture three-axis information. These feature maps effectively retain global information from different perspectives. After information entropy attention and curvature filtering attention processing, the three feature maps are merged after swelling to complement each other's spatial information.

[0059] 22. Information entropy attention

[0060] Information entropy, as a key indicator of quantifying information uncertainty, has irreplaceable advantages in segmentation tasks. On the one hand, information entropy has a certain robustness to noise. In an image with high noise, information entropy can still effectively reflect the gray distribution characteristics, thereby guiding the segmentation process. On the other hand, due to its sensitivity to local features, information entropy can effectively identify structures and boundaries in images, thereby improving the accuracy of segmentation. However, the current application of information entropy mainly relies on the maximum entropy threshold method based on gray value for information filtering. Since the gray value distribution in MRI images is wide, it is not feasible to directly apply this method to the entire image, and it is also very challenging to determine the gray threshold. Therefore, the direct application of information entropy in image segmentation has certain limitations.

[0061] Based on this, the present application has made a series of improvements in this aspect. Considering the complexity and uncertainty of setting a threshold, the present application does not use thresholding methods, but divides the image feature map into a number of small windows with a size of 3x3. Due to the different resolutions of image features at each encoding layer, the number of small windows divided is also different. In each window, the present application calculates the entropy of the gray value to quantify the amount of information in the region, and then uses the top-k algorithm to select the region with the most information. The formula for calculating the entropy is as follows:

[0062]

[0063] where E is the entropy of the entire small window, n is the range of gray values, x i is the gray value of the image, p(x i ) is the gray value x iThe higher the entropy value, the greater the amount of information, meaning the more complex the image. Next, the invention selects the top k regions with the highest entropy values. In the invention, it is found through experiments that setting k to take the top 40% of small windows is the best, and this value is verified in subsequent experiments. These regions usually contain more useful information. To enhance the feature representation of these regions, the invention merges these selected high-entropy regions with the original image, thereby improving the accuracy of subsequent processing, as shown below:

[0064] Y = Conv(concat(X, topk(E))

[0065] where Y represents the enhanced result, X represents the feature map before processing, and E represents the entropy value of the window.

[0066] 23. Curvature filter attention

[0067] In image segmentation tasks, both local features and global features are very important. To efficiently fuse global information, TSP is used to extract feature information from three planes, and global information is successfully compressed into three two-dimensional feature maps. Subsequently, EGA is used to further filter these feature maps. However, to achieve accurate edge extraction, the invention proposes an efficient local information extraction module.

[0068] Curvature can accurately reflect the geometric shape information of image edges, which is particularly important for image segmentation. Generally, high curvature regions correspond to edges and details in images, while low curvature regions correspond to flat regions. During the filtering process, the pixels in high curvature regions change relatively more, while those in low curvature regions are relatively smooth. By accurately calculating the curvature of image edges, small changes in edges can be captured. This ability enables curvature to more accurately identify the boundaries of objects, thereby achieving accurate segmentation of different regions.

[0069] However, due to the lack of specific curvature parameters, the widespread application of curvature has been limited. To this end, the invention has made a series of improvements to curvature and designed a CFA module to better extract local features from medical images. This method can more effectively highlight the edges of objects, making it possible to enhance edge features using curvature. To solve the problem of curvature parameter selection, the invention sets a 3x3 learnable matrix as the initial curvature value, where the initial value of the curvature matrix is set to As Figure 5 shown, the 3x3 learnable matrix is represented as:

[0070]

[0071] In the learning process, the curvature is embedded into the convolution operation by weighted convolution, and through learning, the matrix gradually adapts to the approximate shape of the tumor.

[0072]

[0073] Wherein, alpha t represents the learnable parameter alpha of the tth iteration, alpha t+1 represents the learnable parameter alpha of the t+1th iteration, beta t represents the learnable parameter beta of the tth iteration, beta t+1 represents the learnable parameter beta of the t+1th iteration, and eta is the learning rate, is the gradient of the loss function L with respect to alpha, is the gradient of the loss function L with respect to beta. and It is calculated by nn.parameter, which is a method for defining learnable parameters in pytorch, and is used for parameter gradient derivation in learnable matrix.

[0074] Finally, the three feature maps are converted to the same three-dimensional space through the puffing operation, so that the features from the three dimensions can complement each other, thereby further enhancing the representation of the tumor region. Finally, in order to combine global and local information, the feature maps obtained from the three planes are integrated, and the input and output of ECA are kept the same size, realizing the plug-and-play feature, as shown in the structural diagram Figure 5 .

[0075] 3、Model training

[0076] The actual training process includes how to encode brain tumor images through the network, how to connect the three models in series to maximize the effect of the three models, and how to embed the attention into the model to make the feature extraction most efficient, thereby assisting the model to make accurate judgments.

[0077] 31、Model hyperparameter setting

[0078] The method of the application is implemented on an Nvidia GeForce RTX 3090 20GB GPU, uses a 256x256x256 size 3D image as input, trains for 400 epochs, sets the batch size to 2, the learning rate to 0.01, and the weight decay to 3e-5.

[0079] 32、Error backpropagation

[0080] The backpropagation error process describes how the model is trained. When given an input, the model generates a predicted output. By calculating the difference between this predicted output and the actual value, the invention can obtain the prediction error. Then, this error is backpropagated through the network, allowing the weights in the entire network to be adjusted. Next, the learning rate is applied to the weights at each point in the network to fine-tune these weights. In the invention, the learning rate is set to 0.01. The entire training process uses the SGD optimizer to train, with the weight decay weight set to 3e-5. The batch size during training is set to 2, including 2 brain membrane images and their corresponding labels, and the epoch is set to 200, with the training ending when the epoch iteration reaches 200.

[0081] The loss function during training is a weighted sum of the commonly used Dice loss and cross-entropy loss. The Dice loss improves the model's sensitivity to small targets, while the cross-entropy loss (CE Loss) ensures the model's ability to classify all pixels. The two have complementary properties. The loss function during training is defined as follows:

[0082]

[0083] where p i represents the predicted probability of the i-th pixel; g i represents the true label of the i-th pixel, with the label taking values of 0 or 1; represents the total number of pixels in the image; ω1, ω2 represent weight coefficients for balancing the contribution of the two losses to the total loss.

[0084] 4、Model testing

[0085] This part is mainly responsible for testing the performance of the model, including testing the model on the divided test set and calculating various performance indicators.

[0086] The model testing stage mainly uses the trained model for testing to verify the prediction performance of the model. This test uses the public BraTS2021 dataset. The meanings of the five indicators are introduced as follows:

[0087] 1) Dice coefficient: measures the overlap between the model's predicted results and the true labels, the higher the better.

[0088] 2) Hausdorff distance: measures the shape difference between the segmentation results and the true labels, the smaller the better. A smaller Hausdorff distance means that the difference between the predicted boundary and the true boundary is smaller. Float operation: detects the proportion of positive examples to all true positive examples, the higher the better.

[0089] 3) Float operations: is used to measure the number of floating point calculations a model performs when executing, the smaller the value represents the more efficient.

[0090] 4) Parameter quantity: measures the complexity and learning ability of the model, represented as the total number of learnable parameters in the model, the model with less parameter quantity is more efficient.

[0091] The test results on the BraTS2021 dataset are shown in Table 1:

[0092] Table 1 Test results on the BraTS2021 dataset

[0093]

[0094] In terms of Dice coefficient, the higher the Dice coefficient value represents the higher the segmentation accuracy, the method of the present application performs best among all methods (WT: 83.07, TC: 87.04, ET: 87.15), which is significantly better than other models.

[0095] In addition, the present application also carries out visual test on the present application, compares the segmentation results of different models to compare the segmentation difference with other models, and the results are shown in Figure 6 , wherein ECA is the method proposed in the present application, and Figure 6 , each row respectively represents different segmentation results of the test results of each case, and the superior performance of the segmentation of the present application can be seen. Figure 6

[0096] The present application carries out visual test on the attention area of the model, and each sub-image respectively represents the attention area of different epochs, as shown in Figure 7 , the red area represents the attention point of the model to the input feature, and it can be seen that the model proposed in the present application can more accurately perceive the tumor, not only can more accurately locate the tumor feature, and shows superior performance.

[0097] In addition, the number of topk regions selected before the entropy value in the verification EGA is verified by experiment, as shown in Table 2, and the results prove that when the present application selects the top 40%, the best performance is achieved.

[0098] Table 2 Verification of region selection ratio on BraTS2021 dataset for EGA

[0099]

[0100] ​Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and understanding of the principles of the application and should not be construed as limiting the scope of the application to such specifically enumerated embodiments. Various modifications and changes can be made to the application by those skilled in the art which will be apparent from this disclosure without departing from the spirit and principles of the application. Any modifications, equivalent substitutions, improvements, etc. made to the application should be included within the scope of the application as defined in the following claims.

Claims

1. A method for automatic brain tumor segmentation, characterized in that: include: S1, obtain the MRI scan image with the segmentation label; S2. Build a segmentation model, where the segmentation model adopts a layered encoder and decoder structure; The output of each layer of encoder is used as the input of the novel lightweight attention mechanism module based on information entropy-curvature basic attention corresponding to the encoder of this layer. The output of the novel lightweight attention mechanism module based on information entropy-curvature basic attention is downsampled and used as the input of the encoder of the next layer and the decoder of the corresponding layer. The input of each layer of decoder also includes the up-sampled output of the lower layer decoder; The novel lightweight attention mechanism module based on information entropy-curvature basic attention includes: a three-axis strip pooling unit, an information entropy attention unit, and a curvature filter attention unit; the three-axis strip pooling unit is used to compress the input image, the information entropy attention unit is used to eliminate noise interference in the compressed image, and the curvature filter attention unit is used to extract local features of the image processed by the information entropy attention unit; The three-axis strip pooling unit compresses the data from the sagittal, coronal and transverse planes respectively to obtain feature maps of the three dimensions: sagittal, coronal and transverse. The curvature filter attention unit processes the three-dimensional feature maps output by the information entropy attention submodule through a convolution operation embedded in a learnable matrix; the three-dimensional feature maps output by the convolution operation embedded in the learnable matrix are converted to the same three-dimensional space through a dilation operation; finally, the three-dimensional feature maps converted to the same three-dimensional space are integrated; S3. Using the MRI scan image with segmentation labels obtained in step S1, the segmentation model of step S2 is trained. The loss function during the training process is expressed as: Among them, p i represents the predicted probability of the i-th pixel; g i represents the true label of the i-th pixel, and the label value is 0 or 1; represents the total number of pixels in the image; ω1 and ω2 represent weight coefficients; S4. Inputting the MRI scan image of the object to be segmented without segmentation labels into the trained segmentation model to obtain a brain tumor segmentation result.

2. The method for automatic brain tumor segmentation according to claim 1, characterized in that: The information entropy attention unit divides the feature maps of the three dimensions into several small windows, calculates the entropy of the grayscale value of each small window, and merges the top k small windows with the highest entropy of the grayscale value corresponding to the feature map of the current dimension with the feature map of that dimension.

3. The method for automatic brain tumor segmentation according to claim 2, characterized in that: The value of k is 40% of the total number of small windows corresponding to the current dimension feature map.

Images