Brain tumor automatic segmentation method

Abstract

The invention discloses a brain tumor automatic segmentation method, which is applied to the field of medical image processing, and aims to solve the problem of low accuracy of the existing brain tumor segmentation, the whole 3D image is compressed from the angle of three axes through a three-axis strip-shaped pooling method in combination with three learnable weight vectors, and the segmentation accuracy of the brain tumor is improved. And finally, the information of the three axes is used for mutual complementation, so that the spatial information is reserved to the greatest extent, and the aim of reducing the calculation amount and the internal memory burden is fulfilled. An information entropy attention module is designed, and an area containing the most information amount is obtained by comparing information entropy, so that enhancement is performed. Noise interference of an intracranial complex environment is greatly reduced, so that the segmentation performance of the model is further improved; in addition, curvature filtering attention is designed, and the tumor edge shape is extracted through learnable convolution, so that in final segmentation, the edge of the model is depicted clearest.

Description

Technical Field

[0001] The invention belongs to the field of medical image processing, and in particular relates to a brain tumor segmentation technology. Background Art

[0002] Brain tumors are a common type of intracranial tumor in adults, including both benign and malignant types. Its incidence has gradually increased in recent years, which has had 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 are diverse and complex. According to the classification of the World Health Organization (WHO), brain tumors are graded from grade I (benign) to grade IV (malignant). Brain tumors may cause a series of symptoms such as headaches, seizures, cognitive impairment, or loss of motor function by directly compressing brain tissue or obstructing cerebrospinal fluid circulation.

[0003] In clinical practice, tumors that cause symptoms and continue to grow are usually surgically removed. Given the diversity of brain tumors in size and location, rapid and accurate positioning is crucial in clinical practice, which requires doctors to accurately judge the location and size of brain tumors. In the past, preoperative segmentation of brain tumors was usually performed manually, with experienced doctors performing single-point annotation. However, this manual segmentation is not only time-consuming, but also highly dependent on the doctor's professional level. Therefore, it is prone to errors, which is not conducive to the patient's treatment and recovery, and easily consumes the doctor's energy, while also prolonging the diagnosis and treatment cycle.

[0004] Manually segmenting brain tumors is not only time-consuming, but also highly dependent on the professional level of doctors. This not only requires doctors to have solid knowledge of dermatology, but also involves a lot of manpower costs. The most critical thing is that 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 needs of patients in a timely manner. Therefore, the automatic segmentation method of brain tumors was born. Magnetic resonance imaging is the main data for brain tumor segmentation. It is a 3D medical image. If this image is used for segmentation as a whole, it may lead to high computational costs, which in turn leads to low computational efficiency.

[0005] In view of the above shortcomings, an efficient and fast method for automatic brain tumor segmentation has become the current mainstream segmentation method. Many previous studies have shown that the U-shaped architecture has good performance by using an encoder-decoder structure to extract image features and classify pixels in 3D medical images. Over the years, more 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. As a result, 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 can be extracted and pixels in 3D medical images can be classified by using an encoder-decoder U-shaped structure. 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. As a result, the device requirements become higher, placing a heavy burden on memory resources, and these model structures often lack effective plug-and-play capabilities, limiting their transferability.

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

[0008] In order to address the challenge of high computational costs, many researchers have adopted the strategy of cutting 3D brain tumor images into a series of 2D slices. This innovative method has achieved remarkable results in practice. By simplifying complex three-dimensional data into a more manageable two-dimensional form, it 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, while this strategy brings convenience, it is also accompanied by a problem that cannot be ignored: some critical three-dimensional information may inevitably be lost during the slicing process. Specifically, key features such as the three-dimensional morphology of the tumor and its fine spatial relationship with surrounding tissues are difficult to fully retain after conversion into 2D slices. This loss of spatial information may, to a certain extent, affect the doctor's accurate understanding of the overall picture of the tumor, as well as 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. When using 2D slices, the spatial information is lost, which affects the final segmentation accuracy;

[0011] 2. The computational cost required when using 3D attention is too high, resulting in low computational efficiency. Summary of the invention

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

[0013] The technical solution adopted by the present invention is: a method for automatic segmentation of brain tumors, comprising:

[0014] S1, obtaining a magnetic resonance imaging image with segmentation labels;

[0015] S2. Construct a segmentation model, which adopts a layered encoder and decoder structure; the output result 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, and the output of the novel lightweight attention mechanism module based on information entropy-curvature basic attention is used as the input of the encoder of the next layer and the decoder of the corresponding layer after downsampling;

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

[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 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;

[0018] S3, using the magnetic resonance imaging image with segmentation labels obtained in step S1 to train the segmentation model in step S2;

[0019] S4. Inputting the magnetic resonance imaging image of the object to be segmented without segmentation labels into the trained segmentation model to obtain a brain tumor segmentation result.

[0020] Beneficial effects of the present invention: The present invention compresses the entire 3D image from the perspective of three axes through a three-axis strip pooling method, in conjunction with three learnable weight vectors, and finally uses the three-axis information to complement each other, thereby retaining spatial information to the greatest extent and achieving the purpose of reducing the amount of calculation and memory burden; an information entropy attention module is designed to obtain the area containing the most information by comparing the size of the information entropy, thereby enhancing it, greatly reducing the noise and interference of the complex intracranial environment, and further improving the segmentation performance of the model; a curvature filtering attention is also designed to extract the tumor edge shape through learnable convolution, so that in the final segmentation, the edge of the model of the present invention is most clearly portrayed. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Examples of the size, shape and location of different brain tumor images;

[0022] Figure 2 It is the overall architecture diagram of the present invention;

[0023] Figure 3 Provide an overall design diagram for the model;

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

[0025] Figure 5 This is the curvature filter attention structure diagram;

[0026] Figure 6 Visualization results on the BraTS2021 dataset;

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

[0028] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.

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

[0030] 1. Brain tumor images are 3D medical images. When training for automatic tumor segmentation, the memory burden and computing cost are high, which in turn affects the reasoning efficiency of the entire model.

[0031] 2. Brain tumors vary in size, location, and morphology, and can vary greatly, which makes segmentation difficult;

[0032] 3. The growth environment of brain tumors is extremely complex. Tumors are often closely adjacent to surrounding brain tissues. This requires precise distinction of the boundary between tumor tissues and normal brain tissues 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, incomplete segmentation and segmentation errors may occur, resulting in poor final segmentation results. Figure 1 The images show brain tumors of different locations, sizes, and shapes. Figure 1 As shown, the segmentation results are divided into three regions, namely the whole tumor (WT), tumor core (TC), and enhancing tumor (ET).

[0034] This paper 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). This attention mechanism proposes a pooling method and two spatial attention modules based on the characteristics of brain tumors themselves, which effectively reduces the computational complexity while capturing spatial feature information, and improves parameter efficiency and inference accuracy. The mechanism includes three components: Tri-axial Strip Pooling (TSP) and two cascaded 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 features after TSP, while CFA aims to effectively extract local features from images.

[0035] This paper proposes a complete brain tumor automatic segmentation method, which includes four parts: 1. data preparation, 2. model design, 3. model training, and 4. model testing. The overall structural relationship diagram is shown in the figure Figure 2 As shown:

[0036] The following is a detailed description of the implementation process of each part:

[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 model training to make the data meet the requirements of model training. After that, the collected brain tumor images will be divided into training sets and test sets according to a certain ratio. The training set is used to train the model, and the test set is used to test the final performance of the model after training. It includes the following steps:

[0039] 11. Get data and labels

[0040] The present invention used the BraTS2021 dataset, and a total of 1237 cases were experimentally verified. Each case includes four modalities: T1, T1ce, T2, and FLAIR. These modalities provide different imaging information: T1-weighted imaging is used to display anatomical structures, T1ce is a contrast enhancement of T1, which can highlight the tumor enhancement area, T2-weighted imaging shows the range of edema in and around the tumor, and FLAIR imaging is used to more clearly display abnormal tissue areas. The present invention superimposes these four modalities into a 4-channel image input, thereby achieving accurate segmentation and analysis of brain tumors, thereby verifying the effectiveness and robustness of the model.

[0041] 12. Data preprocessing

[0042] Data preprocessing mainly involves preprocessing the image data in the data set before training the model so that it can meet the requirements of model training. In the present invention, the data in the BraTS2021 data set are all in 5-dimensional format (batch size × image depth × image length × image width), and are adjusted to 2×4×256×256×256 before being put into the model, where the batch size is 2, representing two samples processed each time forward propagation and reverse propagation, and the number of channels is 4, representing four modes, and the image depth, length, and width are all 256. Such preprocessing ensures the uniform format of the data and lays the foundation for efficient training of the model.

[0043] 13. Dataset Division

[0044] Dataset partitioning aims to divide the data set into a training set and a test set according to a certain ratio, which are used for training models and testing models respectively. In the present invention, the present invention divides the BraTS2021 data set, 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 present invention and also the key protected part of the present invention, including the specific design of three-axis strip pooling, information entropy attention, and curvature filtering attention.

[0047] The present invention 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 present invention processes two 3×3×3 convolutions with a stride of 1 as one convolution layer processing operation.

[0049] Each layer of the encoder is processed by a convolutional layer and then input into the ECA, and then down-sampled by 3D max pooling, thereby doubling the number of channels and halving the resolution of the image feature map. In the first encoding layer, the input image of size 2×4×256×256 is directly used as the input of the first encoding layer.

[0050] In the decoder of each layer, a convolution layer is used to process the combination of image features from the skip connection and the lower layer, and then the upsampling operation is completed through 3D transposed convolution, which doubles the resolution of the image feature map output by the decoder of this layer to make it 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 upsampling operation is continued after processing through the convolution layer, so as to gradually restore it to the original resolution. Figure 3 As shown in FIG. 1 , after the final upsampling, the dimension is reduced by a 1×1×1 convolution process to obtain a 2×4×256×256 segmentation image result. Figure 3 Here, Input means input, Output means output, Skip-connection is skip connection, and concat is the abbreviation of concatenate, which means merge and integrate.

[0051] The encoding output result of the last layer is processed by ECA, down-sampled by 3D maximum pooling, and then processed by a convolution layer. Up-sampling is completed by 3D transposed convolution, and then integrated with the result of the encoding output result of the last layer after ECA processing as the input of the last decoding layer.

[0052] Each ECA module consists of three submodules: TSP, EGA and CFA, which are arranged in series. The model design diagram of the whole method is shown in Figure 3 shown.

[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, maximum pooling only focuses on significant information, which may lead to the loss of detail information; average pooling emphasizes global information, but may lose important image features; spatial pyramid pooling can handle inputs of different sizes, but it will increase the computational complexity.

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

[0056] In order to maintain the spatial information in three dimensions, the data from the sagittal, coronal and transverse planes are compressed separately. TSP applies weights through independent learnable vectors to emphasize important features in each plane, making it possible to ignore unimportant areas and highlight tumor areas. For example, the pooling operation on the transverse plane can be expressed by the following formula:

[0057]

[0058] Among them, Y DWrepresents the result of pooling along the cross section, H represents the height of the plane, W H is the learnable weight matrix on the H axis, A i is the component at height i corresponding to the input feature map. Through the above formula, pooling operations are performed on these 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 being processed by information entropy attention and curvature filter attention, the three feature maps are merged after expansion to complement each other's spatial information.

[0059] 22. Information Entropy Attention

[0060] As a key indicator for quantifying information uncertainty, information entropy has irreplaceable advantages in segmentation tasks. On the one hand, information entropy has a certain robustness to noise. In images with large noise, information entropy can still effectively reflect the grayscale 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 grayscale values ​​for information filtering. Due to the wide distribution of grayscale values ​​in MRI images, it is not feasible to directly apply this method to the entire image, and determining the grayscale threshold is also very challenging. Therefore, the direct application of information entropy in image segmentation has certain limitations.

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

[0062]

[0063] Among them, E is the entropy of the entire small window, n is the range of grayscale values, and 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, which means the more complex the image. Next, the present invention selects the top k regions with the highest entropy values. In the present invention, it is found through experiments that setting k to the small window of the first 40% is the best, and this value is verified in subsequent experiments. These regions usually contain more useful information. In order to enhance the feature representation of these regions, the present 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] Among them, 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 the image segmentation task, both local features and global features are very important. In order to efficiently fuse global information, TSP is used to extract feature information from three planes, and the global information is successfully compressed into three two-dimensional feature maps. Subsequently, EGA is used to further filter these feature maps. However, in order to achieve accurate edge extraction, the present invention proposes an efficient local information extraction module.

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

[0069] However, due to the lack of specific curvature parameters, the wide application of curvature is limited. To this end, the present invention makes a series of improvements to curvature and designs a CFA module to better extract local features from medical images. This method can more effectively highlight the edges of objects and provides the possibility of using curvature to enhance edge features. In order to solve the problem of curvature parameter selection, the present invention sets a 3×3 learnable matrix as the initial curvature value, where the initial value of the curvature matrix is ​​set to like Figure 5 As shown, the 3×3 learnable matrix is ​​represented as:

[0070]

[0071] In the learning process, the present invention embeds curvature into the convolution operation through weighted convolution. Through learning, the matrix gradually adapts to the general shape of the tumor. In the learning process, the learnable parameters are updated through back propagation. The update formula is as follows:

[0072]

[0073] Among them, α t represents the learnable parameter α of the tth iteration, α t+1 Denotes the learnable parameters α, β at the t+1th iteration t represents the learnable parameters β at the tth iteration, β t+1 represents the learnable parameter β of the t+1th iteration, η is the learning rate, is the gradient of the loss function L with respect to α, is the gradient of the loss function L with respect to β. and It is calculated by nn.parameter. Those skilled in the art should know that nn.parameter is a method for defining learnable parameters in pytorch, which is used to derive the gradient of parameters in the learnable matrix.

[0074] Finally, the three feature maps are transformed into the same three-dimensional space through the expansion operation, so that the features from the three dimensions can complement each other, thereby further enhancing the representation of the tumor area. 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. The structure diagram is shown in the figure. Figure 5 shown.

[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 their effectiveness, and how to embed the attention into the model to maximize feature extraction efficiency, thereby assisting the model in making accurate judgments.

[0077] 31. Model hyperparameter setting

[0078] The method of the present invention is implemented on an Nvidia GeForce RTX 3090 20GB GPU, using a 3D image of size 256×256×256 as input, training for 400 epochs, with the batch size set to 2, the learning rate to 0.01, and the weight decay to 3e-5.

[0079] 32. Error Back Propagation

[0080] The back propagation 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 present invention can obtain the prediction error. This error is then back-propagated through the network so that the weights in the entire network can be adjusted. Next, the learning rate is applied to the weights of each point in the network to fine-tune these weights. In the present invention, the learning rate is set to 0.01. The entire training process is trained using the SGD optimizer, and its weight decay weight is set to 3e-5. The batch size during training is set to 2, including 2 meningeal images and their corresponding labels, and the epoch is set to 200. When the epoch iteration reaches 200, the training ends.

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

[0082]

[0083] 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; ω 1 ,ω 2 Represents the weight coefficient, which is used to balance the contribution of the two losses to the total loss.

[0084] 4. Model testing

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

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

[0087] 1) Dice coefficient: measures the degree of overlap between the model's prediction results and the true label, the higher the better.

[0088] 2) Hosdorff distance: measures the shape difference between the segmentation result and the true label, the smaller the better. A smaller Hosdorff distance means that the difference between the predicted boundary and the true boundary is smaller. Floating point operation: the ratio of detected positive examples to all true positive examples, the higher the better.

[0089] 3) Floating-point operations: It is used to measure the number of floating-point calculations performed by a model during execution. The smaller the value, the more efficient it is.

[0090] 4) Parameter quantity: It measures the complexity and learning ability of the model, expressed as the total number of learnable parameters in the model. Models with fewer parameters are 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] As for the Dice coefficient, the higher the Dice coefficient value, the higher the segmentation accuracy. The method of the present invention performs best among all the methods (WT: 83.07, TC: 87.04, ET: 87.15), which is significantly better than other models.

[0095] In addition, the present invention also conducts a visual test on the present invention, by comparing the segmentation results of different models, so as to compare the segmentation differences with other models. Figure 6 As shown, Figure 6 ECA is the method proposed by the present invention. Figure 6 In FIG. 5 , each row represents a different segmentation result of the test result of a case, and it can be seen that the segmentation performance of the present invention is superior.

[0096] The present invention performs a visual test on the model's focus area, where each sub-image represents the focus area of ​​different epochs. Figure 7 As shown, the red area represents the model's focus on the input features. It can be seen that the model proposed in the present invention can perceive the tumor more accurately and can not only locate the tumor features more accurately, but also show superior performance.

[0097] In addition, the present invention has conducted experimental verification on the number of topk regions before selecting the entropy value in EGA, as shown in Table 2. The results show that the best performance is achieved when the present invention selects the top 40%.

[0098] Table 2 Verification of the proportion of EGA selected areas on the BraTS2021 dataset

[0099]

[0100] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, the present invention may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for automatic segmentation of brain tumors, characterized in that: include: S1, obtaining a magnetic resonance imaging image with segmentation labels; S2. constructing a segmentation model, wherein the segmentation model adopts a hierarchical encoder and decoder structure; The output result of each layer of encoder is used as the input of the new lightweight attention mechanism module based on information entropy-curvature basic attention corresponding to the encoder of this layer. The output of the new lightweight attention mechanism module based on information entropy-curvature basic attention is used as the input of the encoder of the next layer and the decoder of the corresponding layer after downsampling. 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; S3, using the magnetic resonance imaging image with segmentation labels obtained in step S1 to train the segmentation model in step S2; S4. Inputting the magnetic resonance imaging 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 three-axis strip pooling unit compresses the data from the sagittal plane, coronal plane, and transverse plane respectively to obtain feature maps of the three dimensions: sagittal plane, coronal plane, and transverse plane.

3. The method for automatic brain tumor segmentation according to claim 2, 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.

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

5. The method for automatic brain tumor segmentation according to claim 3, characterized in that: The curvature filter attention unit processes the three-dimensional feature map output by the information entropy attention submodule through a convolution operation embedded in a learnable matrix; converts the three-dimensional feature map output by the convolution operation embedded in a learnable matrix to the same three-dimensional space through a puffing operation; and finally integrates the three-dimensional feature maps converted to the same three-dimensional space.

6. The method for automatic brain tumor segmentation according to claim 5, characterized in that: The loss function during training 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.

Images