A Multi-modal Brain Tumor MRI Segmentation Method Based on Feature Decoupling

Abstract

The present invention relates to a multi-modal brain tumor MRI segmentation method based on feature decoupling. A deep supervision mechanism is introduced to extract deep features of multi-modal brain tumor MRI images. Meanwhile, a network with an attention fusion mechanism is used to extract discriminative features. An auxiliary branch network is utilized to decouple features and fully extract the features of the segmentation part. At the same time, a dynamic feature library is constructed, and an unsupervised clustering algorithm is introduced to enhance features, improving the segmentation accuracy in the case of incomplete modal data, expanding the applicability and generalization of the network model in multi-modal data. The present invention can adaptively process the correspondence relationship between different modal images of brain tumor MRI, and improve the segmentation effect of the deep model in the case of incomplete modal data.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation, and particularly relates to a multi-modal brain tumor MRI segmentation method based on feature decoupling. Background Art

[0002] The goal of medical image segmentation is to divide and extract regions such as organs, tissues, and lesions with visual homogeneity and semantic consistency in images such as CT, MRI, and X-ray; it uses image processing, computer vision, machine learning and other methods to provide theoretical and application support for the digital modeling and automated auxiliary diagnosis and treatment of diseases; as a key step in medical image intelligent analysis technology, medical image segmentation algorithms have shown broad clinical application prospects.

[0003] In recent years, with the development of artificial intelligence technology, artificial intelligence technology based on machine learning has been widely applied in fields such as computer vision and achieved significant results. Aiming at the actual clinical application scenarios and requirements, researching medical image segmentation methods with accuracy, robustness, and generality, and then developing a new generation of intelligent clinical auxiliary diagnosis systems with independent intellectual property rights has great practical significance in promoting the clinical diagnosis and treatment level and efficiency of hospitals, improving the primary diagnosis and treatment level under the multi-level diagnosis and treatment system, and enhancing the informatization strength of the medical system.

[0004] For the supervised feedback characteristics of existing deep neural networks, general natural image analysis tasks often require a large amount of training data to ensure the convergence of the model, which puts high requirements on the quantity and quality of training samples. In the field of brain tumor MRI image segmentation, due to tissue physiological differences, the influence of organ movement, and different device parameters and specifications during image acquisition, the same case will generate different modal brain tumor MRI images, and the data of various modal images often have large appearance differences; on this basis, due to the sparsity, data sensitivity, and relative closure of the disease, coupled with the fact that the labels of brain tumor MRI images need to be annotated by medical professionals, it has always been difficult to obtain a large number of high-quality training samples; due to the various reasons analyzed above, in the current brain tumor MRI image segmentation task, researchers have to face problems such as lack of samples, lack of annotation, and modal loss;

[0005] Therefore, restricted by the above problems, most existing brain tumor MRI segmentation methods are often limited to the diversity of modal data, which brings a series of drawbacks:

[0006] 1) Due to too little available data, it is difficult to establish an accurate and reliable model using the data-driven mechanism of the deep architecture;

[0007] 2) It is difficult to continuously improve the generalization ability of the learning model, and its usability and generality are restricted when the modal data is incomplete;

[0008] 3) The models for brain tumor MRI image analysis tasks of different modalities are relatively independent, which makes it difficult to effectively extract common medical knowledge. Summary of the invention

[0009] To solve the above problems, the present invention provides a multimodal brain tumor MRI segmentation method based on feature decoupling, comprising:

[0010] S1: Acquire a brain tumor MRI image dataset; wherein the brain tumor MRI image dataset includes: brain tumor MRI images of multiple modalities, modality category labels corresponding to the brain tumor MRI images, and segmentation map labels corresponding to the brain tumor MRI images;

[0011] S2: Use the encoding part of the first UNet network to downsample the brain tumor MRI image to obtain original features at multiple scales;

[0012] S3: Input the original features of multiple scales into the attention mechanism module respectively to enhance the original features and obtain the original enhanced features of multiple scales;

[0013] S4: Input the original enhanced features with the largest scale into the Resnet network to predict the modality category of the brain tumor MRI image; and supervise the Resnet network based on the prediction results of the Resnet network and the modality category label of the brain tumor MRI image;

[0014] S5: Input the original enhanced feature with the smallest scale into the spatial pyramid pooling layer for semantic enhancement to obtain the first original semantic enhanced feature; input the first original semantic enhanced feature into the softmax classifier for coarse segmentation, and store the first original semantic enhanced feature with segmentation accuracy higher than the preset value into an ordered queue to construct a feature library;

[0015] S6: using the decoding part of the first UNet network to upsample the original features of multiple scales step by step to obtain a brain tumor MRI restored image; and using the brain tumor MRI image as a label to supervise the decoding part of the first UNet network;

[0016] S7: clustering the first original semantic enhancement features in the feature library using the Kmeans algorithm to generate multiple classes and class centers, and calculating the weight of each first original semantic enhancement feature for each class according to the number of class centers; multiplying the weight of the first original semantic enhancement feature for the class to which it belongs by the class center of the class to which it belongs, to obtain the second original semantic enhancement feature;

[0017] S8: Input the original enhanced features and the second original semantic enhanced features at different scales into the decoding part of the second UNet network to gradually decode a segmentation map consistent with the scale of the brain tumor MRI image. Input the segmentation map into Softmax to output the class prediction result of the segmentation map, and perform supervised training using the segmentation map label of the brain tumor MRI image.

[0018] The present invention has at least the following beneficial effects

[0019] According to the deep supervision mechanism, the present invention uses the encoding part of the UNet network to extract original features at multiple scales. At the same time, the attention mechanism is adopted to enhance the original features to obtain the original enhanced features. The spatial pyramid pooling layer is used to further enhance the original enhanced features with the smallest scale, and rough segmentation is performed to construct a feature library. An unsupervised clustering algorithm is introduced to enhance the features, which improves the segmentation accuracy in the case of incomplete modal data, expands the applicability and generalization of the network model in multi-modal data. The present invention can adaptively process the correspondence relationship between different modalities of brain tumor MRI images and improve the segmentation effect of the deep model in the case of incomplete modal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is the flowchart of the method of the present invention;

[0021] Figure 2 is the schematic diagram of the segmentation model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention schematically. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0023] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and cannot be construed as limitations on the present invention; for better illustrating the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, and do not represent the dimensions of actual products; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0024] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the accompanying drawings. This is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and cannot be understood as a limitation of the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0025] Please refer to Figure 1 , the present invention provides a multi-modal brain tumor MRI segmentation method based on feature decoupling, including:

[0026] S1: Obtain a brain tumor MRI image dataset; wherein, the brain tumor MRI image dataset includes: brain tumor MRI images of multiple modalities, modality category labels corresponding to the brain tumor MRI images, and segmentation map labels corresponding to the brain tumor MRI images;

[0027] Preferably, an implementation manner of obtaining a brain tumor MRI image dataset includes: In this embodiment, the BraTS2018 dataset is downloaded from the network. This dataset contains 285 cases, and each case has four different modalities. By labeling each brain tumor MRI image in the BraTS2018 dataset with modality category labels and segmentation map labels, a brain tumor MRI image dataset is obtained.

[0028] Preferably, an implementation manner of obtaining a brain tumor MRI image dataset includes: In this embodiment, 285 MRI scan data are obtained from 15 medical centers. By labeling each brain tumor MRI image with modality category labels and segmentation map labels, a brain tumor MRI image dataset is obtained. Each MRI scan includes T1, T1ce, T2, and FLAIR image sequences.

[0029] Preferably, an implementation manner of labeling modality category labels for brain tumor MRI images includes:

[0030] For example, in this embodiment, brain tumor MRI images of 4 modalities are used. Therefore, the modality category labels are divided into 4 categories, where 0 represents T1, 1 represents T1ce, 2 represents T2, and 3 represents FLAIR.

[0031] Preferably, an implementation manner of labeling segmentation map labels for each brain tumor MRI image includes:

[0032] This embodiment mainly segments the complete tumor, tumor core, and enhanced tumor in brain tumor MRI images. Therefore, the segmentation map labels in this embodiment are the complete tumor segmentation image, tumor core segmentation image, and enhanced tumor segmentation image corresponding to the brain tumor MRI image.

[0033] Preferably, an implementation manner of labeling each brain tumor MRI image with a segmentation map label includes:

[0034] This embodiment mainly segments low-grade glioma (LGG), high-grade glioma, and cerebrospinal fluid-secreting tumor (HGG) in brain tumor MRI images. Therefore, the segmentation map labels in this embodiment are the images corresponding to low-grade glioma (LGG), high-grade glioma, and cerebrospinal fluid-secreting tumor (HGG) of the brain tumor MRI image.

[0035] Preferably, a manner of preprocessing brain tumor MRI image data includes:

[0036] Performing the same cropping and data augmentation on the brain tumor MRI image;

[0037] Preferably, a manner of preprocessing brain tumor MRI image data includes:

[0038] Randomly cropping the brain tumor MRI image to increase the number of tumor MRI images.

[0039] Preferably, a manner of preprocessing brain tumor MRI image data includes:

[0040] Using the Gaussian filtering algorithm to filter the noise in the brain tumor MRI image to improve the accuracy of image classification.

[0041] Preferably, a manner of preprocessing brain tumor MRI image data includes:

[0042] Stretching the brain tumor MRI images in the brain tumor MRI image dataset to a scale of 572*572. The scale of 572*572 is generally used as the input scale of the UNet network.

[0043] Please refer to Figure 2 , S2: Using the encoding part of the first UNet network to perform downsampling on the brain tumor MRI image to obtain original features at multiple scales;

[0044] Preferably, the encoding part of the first UNet network includes: four downsampling modules. Among them, each downsampling module includes a convolutional layer, a batch normalization BN layer, a ReLU activation function layer, and a pooling layer connected in sequence. The output of each downsampling module is used as the input of the next downsampling module, and the output of each downsampling module is used as the original feature to obtain original features at multiple scales.

[0045] Preferably, an implementation of the first UNet network encoding part includes: the convolution kernel size of the convolutional layer is 3, the downsampling scale of the pooling layer is 2, and the intermediate features output by each downsampling module in the first UNet network encoding part are used as the original semantic features at multiple scales, and are sorted by scale as the feature map of 80×80×80, the feature map of 40×40×40, the feature map of 20×20×20, and the feature map of 10×10×10.

[0046] S3: Input the original features at multiple scales into the attention mechanism module respectively to enhance the features of the original features to obtain the original enhanced features at multiple scales;

[0047] Preferably, the feature enhancement of the original features includes:

[0048] Channelattention = X * Sigmoid(MLP(Avgpool(X) + Maxpool(X)))

[0049] where Channelattention represents the original enhanced feature, X represents the original feature, Sigmoid represents the activation function, MLP represents the multi-layer perceptron, Avgpool represents global average pooling, and Maxpool represents global max pooling.

[0050] S4: Input the original enhanced feature with the largest scale into the Resnet network to predict the modal category of the brain tumor MRI image; and supervise and train the Resnet network according to the prediction result of the Resnet network and the modal category label of the brain tumor MRI image;

[0051] Preferably, the supervision and training of the Resnet network includes:

[0052] Construct a cross-entropy loss function according to the prediction result of the Resnet network and the modal category label of the brain tumor MRI image, and use the adam optimizer to optimize the parameters of the Resnet network, where the cross-entropy loss function is as follows:

[0053]

[0054] where, represents the cross-entropy loss function, y i represents the modal category label of the brain tumor MRI image, P(x i ) represents the prediction result of the Resnet network for the brain tumor MRI image, and N represents the number of samples.

[0055] S5: Input the original enhanced feature with the smallest scale into the spatial pyramid pooling layer for semantic enhancement to obtain the first original semantic enhanced feature; input the first original semantic enhanced feature into the softmax classifier for rough segmentation, and store the first original semantic enhanced feature with a segmentation accuracy higher than the preset value into an ordered queue to construct a feature library;

[0056] Preferably, the calculation method of the segmentation accuracy includes:

[0057]

[0058]

[0059] where Loss represents the segmentation accuracy, x i represents the first original semantic enhanced feature, P(x i ) represents the predicted value of the Softmax classifier for the first original semantic enhanced feature, y ij represents the label of the first original semantic enhanced feature, M represents the number of labels, and there are M segmentation maps of the brain tumor MRI image. For example, when the segmentation maps are complete tumor, tumor core, and enhanced tumor, M is 3.

[0060] Preferably, the labels of the first original semantic enhanced feature include:

[0061] Downsample the segmentation map label of the brain tumor MRI image to make it consistent with the scale of the first original semantic enhanced feature to obtain the label of the first original semantic enhanced feature.

[0062] S6: Use the decoding part of the first UNet network to upsample the original features of multiple scales step by step to obtain the restored brain tumor MRI image; and use the brain tumor MRI image as a label to supervise the training of the decoding part of the first UNet network;

[0063] Preferably, the supervision and training of the decoding part of the first UNet network include:

[0064] Create a loss function based on the restored brain tumor MRI image and the brain tumor MRI image, and use the adam optimizer to update the parameters of the decoding part of the first UNet network. The loss function is as follows:

[0065]

[0066] where, represents the loss function of the decoding part of the first UNet network, f(x i ) represents the restored brain tumor MRI image, x i represents the brain tumor MRI image, MSE represents the mean square error function, and M represents the number of samples.

[0067] Preferably, the decoding part of the first UNet network is mainly composed of four upsampling modules, wherein each upsampling module includes a pooling layer, a convolution layer, a batch normalization BN layer and an activation function ReLU layer connected in sequence, and the output of each upsampling module serves as the input of the next upsampling module.

[0068] Preferably, the decoding parts of the first UNet and the second UNet network each include: four upsampling modules, wherein each upsampling module includes a pooling layer, a convolution layer, a batch normalization BN layer and an activation function ReLU layer connected in sequence, and the output of each upsampling module serves as the input of the next upsampling module.

[0069] Preferably, an implementation of the first UNet network decoding part includes: the convolution kernel size of the convolution layer is 3, the upsampling scale of the pooling layer is 2, the output of the first UNet network decoding part is set to a single-channel image output, and the multi-scale original features are passed through the decoder to obtain a reconstructed prediction image consistent with the scale of the brain tumor MRI image, that is, a brain tumor MRI restored image.

[0070] Preferably, upsampling the original features at multiple scales to obtain a brain tumor MRI restored image comprises:

[0071] The original feature with the smallest scale is input into the first upsampling module of the decoding part of the first UNet network to output the first sub-intermediate feature; the first sub-intermediate feature is spliced ​​with the original feature of the same scale on the feature dimension and then input into the next upsampling module to obtain the second sub-intermediate feature; the second sub-intermediate feature is spliced ​​with the original feature of the same scale on the feature dimension and then input into the next upsampling module to obtain the third sub-intermediate feature; the third sub-intermediate feature is spliced ​​with the original feature of the same scale on the feature dimension and then input into the next upsampling module to obtain the fourth sub-intermediate feature; similarly, the fourth sub-intermediate feature is spliced ​​with the original feature of the same scale on the feature dimension and then input into the last upsampling module to output a feature map, which is the brain tumor MRI restored image.

[0072] S7: clustering the first original semantic enhancement features in the feature library using the Kmeans algorithm to generate multiple classes and class centers, and calculating the weight of each first original semantic enhancement feature for each class according to the number of class centers; multiplying the weight of the first original semantic enhancement feature for the class to which it belongs by the class center of the class to which it belongs, to obtain the second original semantic enhancement feature;

[0073] The Kmeans algorithm belongs to unsupervised learning, and K-means clustering is the most basic and commonly used clustering algorithm. Its basic idea is to find a partitioning scheme of K clusters (Clusters) through iteration, so that the loss function corresponding to the clustering result is minimized. Among them, the loss function can be defined as the sum of the squared errors of each sample from the center point of its belonging cluster.

[0074] Preferably, the second original semantic enhancement feature includes:

[0075]

[0076] Among them, F′ represents the second original semantic enhancement feature, and F represents the first original semantic enhancement feature. represents the dot product operation, and C kmeans (F bank ) represents the class center point.

[0077] S8: Input the original enhancement features and the second original semantic enhancement features of different scales into the decoding part of the second UNet network to decode level by level to obtain a segmentation map consistent with the scale of the brain tumor MRI image. Input the segmentation map into Softmax to output the class prediction result of the segmentation map, and use the segmentation map label of the brain tumor MRI image for supervised training.

[0078] Preferably, the network decoding part of the second UNet includes: four upsampling modules. Among them, each upsampling module includes a pooling layer, a convolutional layer, a batch normalization BN layer, and a ReLU activation function layer connected in sequence. The output of each upsampling module is used as the input of the next upsampling module.

[0079] Preferably, the implementation method of the network decoding part of the second UNet includes: the convolutional kernel size of the convolutional layer is 3, the upsampling scale of the pooling layer is 2, the output of the network decoding part of the second UNet is set to an image output with 4 channels, and the original enhancement features and the second original semantic enhancement features of different scales are input into the decoding part of the second UNet network to obtain a segmentation map consistent with the scale of the brain tumor MRI image through decoding.

[0080] The second original semantic enhancement feature is input into the first upsampling module of the decoding part of the second UNet network to output the first intermediate feature. The first intermediate feature and the original enhancement feature with the same scale are concatenated and then input into the next upsampling module to output the second intermediate feature. The second intermediate feature and the original enhancement feature with the same scale are concatenated and then input into the next upsampling module to output the third intermediate feature. Similarly, the third intermediate feature and the original enhancement feature with the same scale are concatenated and then input into the next upsampling module to obtain the fourth intermediate feature. Finally, the fourth intermediate feature and the enhancement feature with the same scale are concatenated in the feature dimension and then input into the last upsampling module to output a segmentation map consistent with the scale of the brain tumor MRI image. The segmentation map of each channel is input into Softmax to output the class prediction result of the segmentation map, and supervised training is performed using the segmentation map label of the brain tumor MRI image.

[0081] Preferably, the supervised training using the segmentation map label of the brain tumor MRI image includes:

[0082] Construct a Dice loss function based on the segmentation map and the segmentation map label, and use the backpropagation mechanism to update the parameters of the decoding part of the second UNet network. Among them, the loss function is as follows:

[0083]

[0084] Among them, represents the loss function of the decoding part of the second UNet network, t represents the number of samples, f represents the Dice segmentation loss function, X represents the brain tumor MRI image, Y represents the segmentation map label corresponding to the brain tumor MRI image, and p(X) represents the class of the segmentation map output by Softmax.

[0085] In the embodiment, experiments are carried out using the dataset (BraTS2018). The dataset (BraTS2018) comes from the brain tumor segmentation challenge and is widely used for brain tumor segmentation tasks. This dataset contains 285 cases, and each case has four different modalities, and three tumor sites need to be segmented. The present invention mainly focuses on segmentation under the condition of incomplete modality information, and combinations of modality loss situations will be simulated in the experimental settings. Table 1 shows the experimental results obtained by testing under different modality loss situations.

[0086] Table 1 Test results of different modality combinations

[0087]

[0088] As can be seen from Table 1, by testing under different modality combinations, the segmentation accuracy (Dice coefficient) obtained by the present invention is significantly improved compared with the baseline model under the same conditions. This result shows that the present invention can effectively improve the segmentation accuracy of brain tumor MRI images in the case of modality loss based on feature decoupling. In contrast, when the modality information data is incomplete, the performance of the baseline model drops significantly. Therefore, the present invention is effective.

[0089] In this embodiment, the maximum number of epochs is set to 450, and the initial learning rate is 0.001;

[0090] The present invention extracts raw features at multiple scales using the encoding part of the UNet network according to the deep supervision mechanism, and at the same time uses the attention mechanism to enhance the raw features to obtain raw enhanced features. The spatial pyramid pooling layer is used to further enhance the raw enhanced features with the smallest scale, and a rough segmentation is performed to construct a feature library. An unsupervised clustering algorithm is introduced to enhance the features, improving the segmentation accuracy in the case of incomplete modality data, and expanding the applicability and generalization of the network model in multi-modal data. The present invention can adaptively process the corresponding relationship between different modality images of brain tumor MRI, and improve the segmentation effect of the deep model in the case of incomplete modality data.

[0091] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A multimodal brain tumor MRI segmentation method based on feature decoupling, It is characterized in that include: S1: Acquire a brain tumor MRI image dataset; wherein the brain tumor MRI image dataset includes: brain tumor MRI images of multiple modalities, modality category labels corresponding to the brain tumor MRI images, and segmentation map labels corresponding to the brain tumor MRI images; S2: Use the encoding part of the first UNet network to downsample the brain tumor MRI image to obtain original features at multiple scales; S3: Input the original features of multiple scales into the attention mechanism module respectively to enhance the original features and obtain the original enhanced features of multiple scales; S4: Input the original enhanced features with the largest scale into the Resnet network to predict the modality category of the brain tumor MRI image; and supervise the Resnet network based on the prediction results of the Resnet network and the modality category label of the brain tumor MRI image; S5: Input the original enhanced feature with the smallest scale into the spatial pyramid pooling layer for semantic enhancement to obtain the first original semantic enhanced feature; input the first original semantic enhanced feature into the softmax classifier for coarse segmentation, and store the first original semantic enhanced feature with segmentation accuracy higher than the preset value into an ordered queue to construct a feature library; S6: using the decoding part of the first UNet network to upsample the original features of multiple scales step by step to obtain a brain tumor MRI restored image; and using the brain tumor MRI image as a label to supervise the decoding part of the first UNet network; S7: clustering the first original semantic enhancement features in the feature library using the Kmeans algorithm to generate multiple classes and class centers, and calculating the weight of each first original semantic enhancement feature for each class according to the number of class centers; multiplying the weight of the first original semantic enhancement feature for the class to which it belongs by the class center of the class to which it belongs, to obtain the second original semantic enhancement feature; S8: The original enhanced features and the second original semantic enhanced features of different scales are input into the decoding part of the second UNet network to decode the segmentation map consistent with the scale of the brain tumor MRI image step by step, the segmentation map is input into the Softmax output segmentation map category prediction result, and the segmentation map label of the brain tumor MRI image is used for supervised training.

2. The multimodal brain tumor MRI segmentation method based on feature decoupling according to claim 1, It is characterized in that The encoding part of the first UNet network includes: four downsampling modules, wherein each downsampling module includes a convolutional layer, a batch normalization BN layer, an activation function ReLU layer and a pooling layer connected in sequence, and the output of each downsampling module is used as the input of the next downsampling module, and the output of each downsampling module is used as the original feature to obtain original features of multiple scales.

3. The multimodal brain tumor MRI segmentation method based on feature decoupling according to claim 1, It is characterized in that The decoding parts of the first UNet and the second UNet networks both include: four upsampling modules, where each upsampling module includes a pooling layer, a convolutional layer, a batch normalization (BN) layer, and a rectified linear unit (ReLU) layer connected in sequence, and the output of each upsampling module serves as the input of the next upsampling module.

4. A multi-modal brain tumor MRI segmentation method based on feature decoupling according to claim 1, wherein, the feature enhancement of the original features includes: Channelattention = X * Sigmoid(MLP(Avgpool(X) + Maxpool(X))) where Channelattention represents the original enhanced features, X represents the original features, Sigmoid represents the activation function, MLP represents the multi-layer perceptron, Avgpool represents the global average pooling, and Maxpool represents the global maximum pooling.

5. A multi-modal brain tumor MRI segmentation method based on feature decoupling according to claim 1, wherein, the calculation method of the segmentation accuracy includes: Among them, Loss represents the segmentation accuracy, and x i represents the first original semantic enhancement feature, and P(x i ) represents the predicted value of the softmax classifier for the first original semantic enhancement feature, and y ij represents the label of the first original semantic enhancement feature, M represents the number of labels; N represents the number of samples.

6. A multi-modal brain tumor MRI segmentation method based on feature decoupling according to claim 5, wherein, the labels of the first original semantic enhanced features include: Downsampling the segmentation map labels of the brain tumor MRI images to be consistent with the scale of the first original semantic enhanced features to obtain the labels of the first original semantic enhanced features.

7. A multi-modal brain tumor MRI segmentation method based on feature decoupling according to claim 6, wherein, the second original semantic enhanced features include: Among them, F′ represents the second original semantic enhancement feature, and F represents the first original semantic enhancement feature. represents the dot product operation, and C kmeans (F bank ) represents the class center point.

Images