Image segmentation method and device
By introducing global-local feature interaction and multi-scale feature learning in image segmentation neural networks, the problem of low accuracy of gap segmentation around blood vessels in magnetic resonance images is solved, and the segmentation effect of high precision and detail retention is achieved.
Patent Information
- Application Number
- CN202410978456.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-07-19
AI Technical Summary
The existing image segmentation method is difficult to accurately segment the gap around the blood vessel in magnetic resonance images, especially because the characteristics of the tiny tubular structure are easily lost, resulting in low segmentation accuracy.
An image segmentation method is adopted, and the segmentation result graph is generated through multi-scale feature learning and global-local feature interaction using the segmentation neural network model, including input module, encoder module, global-local feature interaction. The model contains multiple encoding layers and Transformer layers, using the global-local feature interaction module and the multi-head cross-attention interaction module to enhance feature representation and restore spatial resolution.
It significantly improves segmentation accuracy, achieves high Dice values, retains detailed information, and is particularly outstanding in segmentation of small tubular structures.
Smart Images

Figure CN119091133A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of image processing, and in particular relates to an image segmentation method and a device thereof. Background Art
[0002] Perivascular Spaces (PVS) refer to a structure present in the brain and spinal cord, which is usually filled with cerebrospinal fluid or interstitial fluid. They are located in the space around blood vessels and are an important part of the brain's lymphatic system. The morphology and changes of PVS are closely related to cerebral small vessel disease, multiple sclerosis and neurodegenerative diseases.
[0003] There are various algorithms that automatically mark PVS in magnetic resonance (MR) images. However, the existing algorithms rely on filter-based image processing operations. However, PVS appears as widely distributed small tubular structures in MR images, and the average diameter of a single resolvable PVS is less than 5 mm. The segmentation algorithm is constrained by manually tuned parameters and performs poorly in actual implementation. Summary of the invention
[0004] The purpose of the present invention is to provide an image segmentation method, aiming to solve the problem that the existing segmentation method is prone to losing features and has low image accuracy:
[0005] The first aspect of the present invention provides an image segmentation method, comprising: obtaining an image to be processed; inputting the image to be processed into the segmentation neural network model to obtain a segmentation result map; the segmentation neural network model comprises: an input module, an encoder module, a plurality of global-local feature interaction sub-networks, a decoder module and an output module; the input module is used to receive the image to be processed; the encoder module comprises a plurality of encoding layers Enc i , i is a natural number from 1 to N, N is the number of coding layers; and multiple Transformer layers TL j , j is a natural number from 1 to M, M is the number of Transformer layers; the encoding layer includes a multi-scale feature learning module; the output local feature of each encoding layer The output local features of each Transformer layer are The global-local feature interaction subnetwork is used to receive the local features output by each encoding layer and extract the global features F based on the image to be processed. g , and obtain interactive features through the interaction of global features and local features The decoder module includes multiple decoding layers, namely Dec N ,Dec N-1..., Dec2, Dec1; the output local feature F of the Transformer layer lM and interaction features Input the decoding layer Dec N Receive, get output Dec m Receive decoding layer Dec m+1 Output and interaction features Get Output m is a natural number from 1 to N-1; the decoding layer restores the spatial resolution through upsampling and convolution operations; the output module includes a 1×1 convolution layer for receiving the output of the last decoding layer Get the segmentation result image.
[0006] Optionally, the image segmentation neural network model is obtained by training based on a magnetic resonance image training set, and the magnetic resonance image training set includes magnetic resonance images in which perivascular spaces have been annotated.
[0007] Optionally, the magnetic resonance (MR) images included in the training set are MR images acquired on a 3T MRI (3-Tesla magnetic resonance imaging) scanner, and the images to be processed are MR images acquired on a 3T MRI scanner.
[0008] Optionally, the global-local feature interaction subnetwork consists of a multi-scale Transformer-based global feature extraction network and a global-local feature interaction module.
[0009] Optionally, the multi-scale Transformer-based global feature extraction network (Global Feature Extractor, GFE) includes: receiving an input image x of size H×W: performing a convolution operation on the input image x using a convolution kernel to generate a feature map F of size H×W×C x ; Use three convolution kernels of different sizes to generate three feature maps of the same size of H'×W'×C' by adjusting the padding and stride {F a ,F b ,F c};The feature map {F a ,F b ,F c}Reshape to size N f ×C' embedding matrix, where N f =H'×W' represents the number of tags, C' is the number of channels for each feature, and the learnable absolute position embedding matrix is used To save the position information, and add the feature map to the embedding matrix to obtain the position embedding feature map {Ta ,T b ,T c}; embed the position into the feature map {T a ,T b ,T c} Input into the Transformer encoder and reshape the output to the same size as the original feature map and the original feature; the reshaped feature map is convolved along the channel dimension and the final global feature F is generated through convolution operation g .
[0010] Optionally, the global-local feature interaction module (GLFI) includes: obtaining the F g and the local features The global feature F g and the local features The features are extracted through convolution operation and then reshaped to obtain the global features F' g and the local features Among them, the global feature F' g and the local features The same size as F'; the one-dimensional learnable absolute position embedding matrix is combined with the global feature F' g Add to get the global feature T containing position information g ; Embed the one-dimensional learnable absolute position matrix with local features Add to get the local features containing location information The local features are analyzed through the multi-head cross attention interaction module and the global feature T g Interact to get interactive features Use continuous deconvolution operations to transform the interaction features Mapped to a feature map of the same size as the corresponding local feature right and Concatenate along the channel dimension and use convolution to generate the final global-local interaction features
[0011] Optionally, the convolution operation includes: using continuous convolution operations and reshaping operations to transform the global feature F g Convert to Global Features Where N g represents the number of markers, C g Indicates the number of channels.
[0012] Optionally, the multi-head cross attention interaction module includes: As the input of the query generator, the global feature T gAs input to the key generator and value generator, LN represents layer normalization, and the formula of the multi-head cross attention interaction module is as follows:
[0013]
[0014] in, and is a learnable parameter, C l ,h represent the embedding dimension and the number of heads respectively.
[0015] Optional, output of the multi-head cross-attention interaction module It can be expressed as:
[0016]
[0017] Among them, LN(·) represents layer normalization and MLP(·) represents a fully connected layer.
[0018] Optionally, the multi-scale feature learning module includes: generating three feature maps {F1, F2, F3} by using a convolution operation of three convolution kernels; splicing the feature maps {F1, F2, F3} along the channel dimension to form a multi-channel feature map F in ; Use the squeeze and excitation module to process the multi-channel feature map F in , to increase the focus on highly correlated features and reduce the focus on weakly correlated features.
[0019] Optionally, the segmentation neural network model is obtained by training a basic neural network model based on a training set; the method for training the basic neural network model based on the training set includes: using a combination of cross entropy loss and Dice loss as a loss function, and using an Adam optimizer to train the model.
[0020] Optionally, the method for training the basic neural network model based on the training set also includes: (1) performing data enhancement on the image to be processed, wherein the data enhancement includes cropping, rotation and translation; the cropping includes: randomly cropping the image to be processed into a plurality of first image blocks, performing center cropping on the plurality of first image blocks to obtain a second image block; the first image block is larger than the second image block; then rotating and translating the first image block and the second image block to obtain a first image block after data enhancement and a second image block after data enhancement; (2) inputting the first image block after data enhancement and the corresponding second image block after data enhancement into the segmentation neural network model to obtain a first probability feature map F respectively. L and the second probability feature map F S ; For the first probability feature map F L Perform center cropping to obtain the first probability feature subgraph F' L ; F'L With F S Same size; (3) Select F' L and F S The larger value of each pixel in is used as the segmentation probability feature map of the final image block.
[0021] Another aspect of the present invention provides an image segmentation device, comprising: an acquisition unit, for acquiring an image to be processed; a segmentation unit, for inputting the image to be processed into the segmentation neural network model to obtain a segmentation result map; the segmentation neural network model comprises: an input module, an encoder module, a plurality of global-local feature interaction sub-networks, a decoder module and an output module; the input module is used to receive the image to be processed; the encoder module comprises a plurality of encoding layers Enc i , i is a natural number from 1 to N, N is the number of coding layers; and multiple Transformer layers TL j , j is a natural number from 1 to M, M is the number of Transformer layers; the encoding layer includes a multi-scale feature learning module; the output local feature of each encoding layer The output local features of each Transformer layer are The global-local feature interaction subnetwork is used to receive the local features output by each encoding layer and extract the global features F based on the image to be processed. g , and obtain interactive features through the interaction of global features and local features The decoder module includes multiple decoding layers, namely Dec N ,Dec N-1 ..., Dec2, Dec1; the output local features of the Transformer layer and interaction features Input the decoding layer Dec N Receive, get output Dec m Receive decoding layer Dec m+1 Output and interaction features Get Output m is a natural number from 1 to N-1; the decoding layer restores the spatial resolution through upsampling and convolution operations; the output module includes a 1×1 convolution layer for receiving the output of the last decoding layer Get the segmentation result image.
[0022] The segmentation neural network model is obtained by training based on an MR image training set, wherein the MR image training set includes MR images in which perivascular spaces have been annotated.
[0023] Yet another aspect of the present invention provides a computer-readable storage medium, comprising instructions, which, when executed on a computer, enable the computer to execute the image segmentation method as described in any one of the first aspects of the embodiments of the present invention.
[0024] Technical Effects
[0025] The present invention has the following advantages: the global-local feature interaction subnetwork used in the present invention can better capture feature information of different scales, and enhance feature representation through the interaction of global and local features, significantly improve segmentation accuracy, and achieve a higher Dice value; it retains detail information, which is particularly helpful for the segmentation of small tubular structures. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without creative work. It is understood that the drawings provided in this section are only used to better understand the present solution and do not constitute a limitation of the present invention.
[0027] Figure 1 A schematic diagram of a flow chart of an embodiment of the image segmentation method provided by the present invention;
[0028] Figure 2 A schematic diagram of a processing flow of an image segmentation neural network model for an input image to be processed;
[0029] Figure 3 A schematic diagram of the processing flow of the global feature extraction network based on multi-scale Transformer provided by the present invention;
[0030] Figure 4 A schematic diagram of a workflow of the global-local feature interaction module provided by the present invention
[0031] Figure 5 A schematic diagram of the convolution operation of the multi-scale feature learning module provided by the present invention
[0032] Figure 6 The figure is a schematic diagram of a processing flow of a first image block and a second image block.
[0033] Figure 7 It is a schematic diagram comparing the processing results of other segmentation methods and the image segmentation method provided by the present invention.
[0034] Figure 8 A structural schematic diagram of the image segmentation device provided by the present invention;
[0035] Fig. 9 This is another structural schematic diagram of the image segmentation device provided by the present invention. DETAILED DESCRIPTION
[0036] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only embodiments of a part of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention. At the same time, for the sake of clarity and simplicity, the description of well-known functions and structures is omitted in the following description.
[0037] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0038] Before introducing the content of the present invention, it is necessary to introduce the terms involved in the present invention:
[0039] TransUNet network, TransUNet network is a deep learning model that combines Transformer and U-Net architectures, designed for medical image segmentation tasks. TransUNet combines two architectures, convolutional neural network (CNN) and Transformer. It adopts the encoder-decoder structure of U-Net, in which the encoder part combines CNN and Transformer.
[0040] See also Figure 1 The image segmentation method provided by the present invention includes: step 101-step 102.
[0041] 101. Obtain an image to be processed.
[0042] Acquiring the image to be processed refers to acquiring the image to be processed and analyzed from the magnetic resonance imaging device. In this specific scheme, the image to be processed refers to a magnetic resonance (MR) image containing the perivascular space (PVS). The step of acquiring the image to be processed is the starting point of the entire processing flow and is the prerequisite for passing the image to be processed to the neural network model. That is, the patient's brain is scanned with an MR scanner to obtain an MR image containing the perivascular space. The acquired MR image is subjected to necessary preprocessing operations such as denoising, standardization, and alignment as the image to be processed.
[0043] 102. Input the image to be processed into the segmentation neural network model to obtain a segmentation result image.
[0044] Specifically, the MR images that have been acquired and preprocessed in the previous step are used as input data, and these images to be processed are transmitted to the input module of the segmentation neural network model.
[0045] The segmentation neural network model involved in the present invention is explained in detail below:
[0046] The segmentation neural network model includes: an input module, an encoder module, multiple global-local feature interaction sub-networks, a decoder module and an output module; specifically, the processing flow of the segmentation neural network model can be referred to Figure 2 , Figure 2 This is a schematic diagram of the processing flow of the segmentation neural network model on the input image.
[0047] 1. The input module is used to receive the image to be processed, that is, to receive the image to be processed formed in step 101.
[0048] 2. Encoder module, including multiple encoding layers Enc i , i is a natural number from 1 to N, N is the number of encoding layers; in this embodiment, there are four encoding layers, N=4, namely Enc1, Enc2, Enc3 and Enc4. The encoding layer includes a multi-scale feature learning module; the output local feature of each encoding layer In this embodiment, the output local features of each encoding layer are and Among them, the output local features of the encoding layer Enc1 are The output local features of the encoding layer Enc2 are The output local features of the encoding layer Enc3 are The output local features of the encoding layer Enc4 are
[0049] and multiple Transformer layers TL j, j is a natural number from 1 to M, M is the number of Transformer layers; the output local features of each Transformer layer are In this embodiment, there are two Transformer layers, M=2, namely TL1 and TL2. The output local features of each Transformer layer are TL1 and TL2 respectively. Among them, the output local feature of Transformer layer TL1 is TL1; the output local feature of Transformer layer TL2 is TL2.
[0050] The encoder module is used to output multi-scale features; these features have different spatial resolutions and semantic information. The multi-scale feature learning module aims to extract features of different scales and different semantic levels from the input image for use in segmentation tasks.
[0051] The multi-scale feature learning module includes convolutional layers and pooling layers; it can also include attention mechanisms, such as channel attention and spatial attention. It can effectively improve the performance of the segmentation network and enhance the network's ability to perceive structures of different sizes and complexity in images.
[0052] Specifically, three feature maps {F1, F2, F3} are generated by using the convolution operation of three convolution kernels; the feature maps {F1, F2, F3} are concatenated along the channel dimension to form a multi-channel feature map F in ; Use the squeeze and excitation module to process the multi-channel feature map F in , to increase the focus on highly correlated features and reduce the focus on weakly correlated features.
[0053] More specifically, three feature maps {F1, F2, F3} are generated by using three convolution operations with kernel sizes of 3×3, 5×5, and 7×7. In the implementation process, cascaded 3×3 convolution operations are used to extract features of different scales. In this way, the receptive field can be expanded and the network parameters can be reduced. Specifically, two cascaded 3×3 convolution operations are used to create a receptive field of approximately 5×5 size, and three cascaded 3×3 convolution operations are used to generate a receptive field of approximately 7×7 size. The specific processing process can be referred to Figure 5 , Figure 5 This is a schematic diagram of the convolution operation of the multi-scale feature learning module provided by the present invention.
[0054] Concatenate the feature maps {F1, F2, F3} along the channel dimension to form a multi-channel feature map F in ;
[0055] Use the squeeze and excitation module to process the multi-channel feature map F in , to increase the focus on highly correlated features and reduce the focus on weakly correlated features.
[0056] The formula used by the squeeze and excitation module is as follows:
[0057] F1=Conv 3×3 (x),F2=Conv 3×3 (F1),F3=Conv 3×3 (F2)
[0058] F out =F SE (Concate(F1,F2,F3))
[0059] Among them, x represents the input feature map, Conv 3×3 represents a 3×3 convolution operation, Concate represents concatenation along the channel dimension, and F SE Represents the squeeze operation and excitation operation of the SE module.
[0060] The multi-scale feature learning module generates multiple feature maps through convolution operations of different scales, and processes these feature maps through the squeezing and excitation modules to enhance the model's perception of highly correlated features while reducing attention to weakly correlated features, thereby improving the expression and distinction capabilities of features.
[0061] 3. The global-local feature interaction sub-network consists of a multi-scale Transformer-based global feature extraction network and a global-local feature interaction module.
[0062] 1. A global feature extraction network based on a multi-scale Transformer, including:
[0063] (1) Receive an input image x of size H×W:
[0064] (2) Use a convolution kernel (preferably a 3×3 convolution kernel) to perform a convolution operation on the input image x to generate a feature map F of size H×W×C x ;
[0065] (3) Use three convolution kernels of different sizes (3×3, 5×5, and 7×7) to generate three feature maps of the same size H'×W'×C' by adjusting the padding and stride {F a ,F b ,F c};
[0066] (4) The feature map {F a ,F b ,F c}Reshape to size N f ×C' embedding matrix, where N f=H'×W' represents the number of tags, C' is the number of channels for each feature, and the learnable absolute position embedding matrix is used To save the position information, and add the feature map to the embedding matrix to obtain the position embedding feature map {T a ,T b ,T c};
[0067] (5) Embed the position into the feature map {T a ,T b ,T c} Input into the Transformer encoder and reshape the output to the same size as the original feature map and original features;
[0068] (6) The reshaped feature map is convolved along the channel dimension and the final global feature F is generated through a convolution operation (a 1×1 convolution operation). g .
[0069] The processing flow of the global feature extraction network based on multi-scale Transformer can be referred to Figure 3 As shown, Figure 3 This is a schematic diagram of the processing flow of the Transformer-based global feature extraction network provided by the present invention. Based on the above processing method, by using multiple convolution kernels of different sizes, the multi-scale features of the input image can be effectively captured. This method enables the network to extract richer and more detailed feature information at different scales, which is beneficial to the segmentation task of structural objects of different sizes and shapes.
[0070] 2. Global-local feature interaction module
[0071] The workflow of the global-local feature interaction module can be referred to Figure 4 , Figure 4 That is a schematic diagram of the workflow of the global-local feature interaction module provided by the present invention.
[0072] The global-local feature interaction module includes:
[0073] (1) Obtaining the global feature F g and local features
[0074] In this embodiment, there are four local features, namely and The local features are obtained through the encoder module in step (2). The global features F are obtained through the global feature extraction network of the multi-scale Transformer. g .
[0075] (2) The global feature Fg and local features The features are extracted through convolution operation and then reshaped to obtain the global features F' g and local features Among them, the global feature F' g and local features The same size;
[0076] The convolution operation involves using continuous convolution operations and reshaping operations to transform the global feature F g Convert to Global Features Where N g represents the number of markers, C g Indicates the number of channels.
[0077] Specifically, in this embodiment, continuous 3×3 convolution operations with a stride of 2 and reshaping operations are used to transform the global feature F g Convert to global feature F' g ; where N g =20×20 represents the number of markers, C g =512 indicates the number of channels.
[0078] (3) The one-dimensional learnable absolute position embedding matrix and the global feature F' g Add to get the global feature T containing position information g ; Embed the one-dimensional learnable absolute position matrix with local features Add to get the local features containing location information
[0079] (4) Local features are analyzed through multi-head cross attention interaction modules and the global feature T g Interact to get interactive features
[0080] Global features and local features interact through a multi-head cross-attention mechanism to enhance the richness and accuracy of features and retain detail information; by retaining feature information at different levels, it helps to improve the precision and accuracy of the segmentation network.
[0081] The multi-head cross attention interaction module includes: As the input of the query generator, the global feature T g As input to the key generator and value generator, LN represents layer normalization, and the formula of the multi-head cross attention interaction module is as follows:
[0082]
[0083]
[0084] in, and is a learnable parameter, C l ,h represent the embedding dimension and the number of heads respectively;
[0085] Output of the multi-head cross-attention interaction module It can be expressed as:
[0086]
[0087] Among them, LN(·) represents layer normalization and MLP(·) represents a fully connected layer.
[0088] (5) Use continuous deconvolution operations to transform the interactive features Mapped to a feature map of the same size as the corresponding local feature right and Concatenate along the channel dimension and use convolution (preferably 3×3) to generate the final global-local interaction features
[0089] The global-local feature interaction module enhances the expressiveness of features through the multi-head cross-attention mechanism of Transformer; at the same time, through deconvolution and convolution operations, it ensures that the final generated global-local interaction features have good spatial resolution and semantic information, which is suitable for subsequent perivascular space segmentation tasks.
[0090] 4. Decoder Module
[0091] The decoder module includes multiple decoding layers, namely Dec N ,Dec N-1 ..., Dec2, Dec1. In this embodiment, there are four decoding layers, namely Dec4, Dec3, Dec2 and Dec1.
[0092] The output local features F of the Transformer layer lM and interaction features Input the decoding layer Dec N Receive, get output Dec m Receive decoding layer Dec m+1 Output and interaction features Get Output m is a natural number from 1 to N-1; the decoding layer restores the spatial resolution through upsampling and convolution operations. That is, the output local features of the Transformer layer and interaction features The input decoding layer Dec4 receives and obtains the output Dec3 receives the output of the decoding layer Dec4 and interaction features Get Output Dec2 receives the output of the decoding layer Dec3 and interaction features Get Output Dec1 receives the output of the decoding layer Dec2 and interaction features Get Output
[0093] The above operations combine global and local feature information to more accurately predict the position and shape of the target.
[0094] 5. Output Module
[0095] The output module includes a convolutional layer that receives the output of the last decoding layer. Specifically, a 1×1 convolutional layer is used to receive the output of the last decoding layer to generate the final segmentation result map.
[0096] Model Training
[0097] The segmentation neural network model is trained based on a magnetic resonance image training set; the magnetic resonance image training set includes magnetic resonance images in which perivascular spaces have been annotated.
[0098] The training process of the segmentation neural network model begins with the acquisition and preprocessing of a large number of MR image datasets, each of which contains information about the annotated perivascular spaces. First, these MR images undergo preprocessing steps such as denoising, grayscale, and contrast adjustment to meet the training requirements of the neural network model. Then, the entire dataset is divided into a training set, a validation set, and a test set. The training set is used to optimize model parameters, the validation set is used to adjust hyperparameters and monitor model performance, and the test set is used to ultimately evaluate the generalization ability of the model. An initial neural network model consistent with the structure of the segmentation neural network model is set, and an appropriate learning rate is set. After the model training is completed, the validation set is used to evaluate the performance of the model on unseen data. The evaluation indicators may include Dice coefficient, accuracy, and recall. Finally, the generalization ability of the model is verified by the test set to ensure that the model can accurately extract the perivascular region of interest from new MR images, providing reliable support for medical image analysis and diagnosis.
[0099] The method for training the basic neural network model based on the training set also includes:
[0100] (1) performing data enhancement on the image to be processed, wherein the data enhancement includes cropping, rotation and translation;
[0101] The cropping includes: randomly cropping the image to be processed into a plurality of first image blocks, and performing center cropping on the plurality of first image blocks to obtain a second image block; the first image block is larger than the second image block;
[0102] Then, the first image block and the second image block are rotated and translated to obtain a data-enhanced first image block and a data-enhanced second image block;
[0103] (2) Inputting the first image block after data enhancement and the corresponding second image block after data enhancement into the segmentation neural network model, respectively obtaining a first probability feature map F L and the second probability feature map F S ; For the first probability feature map F L Perform center cropping to obtain the first probability feature subgraph F' L ; F' L With F S Same size;
[0104] (3) Select F' L and F S The larger value of each pixel in is used as the segmentation probability feature map of the final image block.
[0105] Specifically, the process of data enhancement for the processed image can be referred to Figure 6 , Figure 6 That is a schematic diagram of a processing flow for the first image block and the second image block. Figure 6 The GLoMS-Net shown in the figure is the segmentation neural network model provided by the present invention. This method uses random cropping to enhance the data set, so that the model can process image blocks of different sizes, and uses cropping and pixel-level maximum operations to merge the segmentation results of image blocks of different sizes, thereby improving the accuracy and robustness of perivascular space segmentation.
[0106] The training process uses a segmentation neural network model to train a basic neural network model based on a training set; the method for training the basic neural network model based on the training set includes: using a combination of cross entropy loss and Dice loss as a loss function, and using an Adam optimizer to train the model.
[0107] The cross entropy loss is used to evaluate the difference between the predicted result and the true label. For each pixel, the cross entropy loss calculates the logarithmic probability difference between the predicted category probability and the actual category. Dice loss is used to evaluate the overlap in the image segmentation task, which is defined as the ratio of the intersection of the predicted segmentation and the true segmentation to their union. Dice loss pays more attention to the overall shape and area of the segmentation result, especially in the segmentation of small objects (such as PVS), and can improve the model's sensitivity to these fine structures. By combining these two loss functions, the classification accuracy and the overlap of the segmented area can be taken into account at the same time, so that the model can learn all aspects of the segmentation task more comprehensively during the training process, thereby improving the accuracy and robustness of the final segmentation result. The initial learning rate of the model is set to 1e-3, and the initial learning rate of the deconvolution layer in the decoder is set to 1e-4. The cosine annealing scheduling strategy is used to update the learning rate, and the Adam optimizer is used to update the parameters of the network. The segmentation model obtained with the above structure and training process has a high accuracy.
[0108] The MR images included in the training set are MR images acquired on a 3T MRI scanner, and the images to be processed are MR images acquired on a 3T MRI scanner.
[0109] 3T MRI refers to the magnetic field strength of the Magnetic Resonance Imaging (MRI) device. In MRI technology, the magnetic field strength is expressed in Tesla (T for short). T MRI refers to an MRI device with a magnetic field strength of 3 Tesla. Previous studies on automatic PVS segmentation have mainly focused on 7T MR images, which have not been widely used in clinical practice. In order to improve clinical applicability. Using 3T MRI images for model training can undoubtedly better meet clinical needs.
[0110] More specifically, they include:
[0111] (1) Obtain the original MR images and perform preprocessing to construct the training set, validation set, and test set data.
[0112] (1-a) MR images were acquired on a 3T MRI scanner.
[0113] (1-b) Neuroradiologists used ITK-Snap software to annotate PVS on T2-weighted MR images. For each subject, the slice with the largest number of PVS in the basal ganglia, centrum semiovale, and body of the lateral ventricle was selected for slice annotation (groundtruth).
[0114] (1-c) The original data is randomly divided into training set, validation set and test set in the ratio of 7:1:2 at the subject level. The labels are the same, and there is no overlap between the three groups.
[0115] (2) Perform data enhancement on the training set and the validation set to obtain a first image block after data enhancement and a second image block after data enhancement.
[0116] (2-a) Perform data augmentation on the processed images, including cropping, rotation, and translation, to expand the number of training samples.
[0117] (2-b) In the cropping of step (2-a), the size of the first image block is set to 96×96, and then these first image blocks are center cropped to generate second image blocks with a size of 80×80.
[0118] (3) The first image block after data enhancement and the corresponding second image block after data enhancement are input into the segmentation neural network model to obtain the first probability feature map F L and the second probability feature map F S ; For the first probability feature map F L Perform center cropping to obtain the first probability feature subgraph F' L ; F' L With F S Same size;
[0119] (3-1) The segmentation neural network model includes: an input module, an encoder module, multiple global-local feature interaction sub-networks, a decoder module and an output module;
[0120] (3-1-1) The input module is used to receive the image to be processed; that is, to receive the first image block after data enhancement and the second image block after corresponding data enhancement.
[0121] (3-1-2) Encoder module, including multiple encoding layers Enc i , i is a natural number from 1 to N, N is the number of encoding layers; in this embodiment, there are four encoding layers, N=4, namely Enc1, Enc2, Enc3 and Enc4. The encoding layer includes a multi-scale feature learning module; the output local feature of each encoding layer In this embodiment, the output local features of each encoding layer are and Among them, the output local features of the encoding layer Enc1 are The output local features of the encoding layer Enc2 are The output local features of the encoding layer Enc3 are The output local features of the encoding layer Enc4 are and multiple Transformer layers TL j, j is a natural number from 1 to M, M is the number of Transformer layers; the output local features of each Transformer layer are In this embodiment, there are two Transformer layers, M=2, namely TL1 and TL2. The output local features of each Transformer layer are TL1 and TL2 respectively. Among them, the output local feature of Transformer layer TL1 is TL1; the output local feature of Transformer layer TL2 is TL2.
[0122] (3-1-3) The multi-scale feature learning module includes a convolution layer and a pooling layer; it may also include an attention mechanism. By using a convolution operation with three convolution kernels, three feature maps {F1, F2, F3} are generated; specifically, by using three convolution operations with convolution kernels of size 3×3, 5×5, and 7×7, three feature maps {F1, F2, F3} are generated; in the implementation process, cascaded 3×3 convolution operations are used to extract features of different scales. In this way, the receptive field can be expanded and the network parameters can be reduced. Specifically, two cascaded 3×3 convolution operations are used to create a receptive field of approximately 5×5 size, and three cascaded 3×3 convolution operations are used to generate a receptive field of approximately 7×7 size.
[0123] Concatenate the feature maps {F1, F2, F3} along the channel dimension to form a multi-channel feature map F in ; Use the squeeze and excitation module to process the multi-channel feature map F in , to increase the focus on highly correlated features and reduce the focus on weakly correlated features.
[0124] (3-1-4) The global-local feature interaction subnetwork consists of a multi-scale Transformer-based global feature extraction network and a global-local feature interaction module.
[0125] (3-1-4a) A global feature extraction network based on a multi-scale Transformer, including:
[0126] Taking an input image x of size H×W:
[0127] Use a 3×3 convolution kernel to perform a convolution operation on the input image x to generate a feature map F of size H×W×C x ;
[0128] Three convolution kernels of different sizes (3×3, 5×5, and 7×7) are used to generate three feature maps of the same size H'×W'×C' by adjusting the padding and stride {F a ,F b ,F c};
[0129] The feature map {F a ,F b ,F c}Reshape to size N f ×C' embedding matrix, where N f =H'×W' represents the number of tags, C' is the number of channels for each feature, and the learnable absolute position embedding matrix is used To save the position information, and add the feature map to the embedding matrix to obtain the position embedding feature map {T a ,T b ,T c};
[0130] The position is embedded into the feature map {T a ,T b ,T c} Input into the Transformer encoder and reshape the output to the same size as the original feature map and original features;
[0131] The reshaped feature map is then transformed along the channel dimension and a 1×1 convolution operation is performed to generate the final global feature F g .
[0132] (3-1-4b) The global-local feature interaction module includes:
[0133] Get the global feature Fg 和 Local features
[0134] The local features are obtained through the encoder module; the four local features are and The global feature F is obtained through the global feature extraction network of the multi-scale Transformer g .
[0135] The global feature F g and local features The features are extracted through convolution operation and then reshaped to obtain the global features F' g and local features Among them, the global feature F' g and local features The same size;
[0136] The convolution operation involves using continuous convolution operations and reshaping operations to transform the global feature F g Convert to Global Features Where N g represents the number of markers, C gIndicates the number of channels. Specifically, the global feature F is transformed into g Convert to global feature F' g ; where N g =20×20 represents the number of markers, C g =512 indicates the number of channels.
[0137] Combine the one-dimensional learnable absolute position embedding matrix with the global feature F' g Add to get the global feature T containing position information g ; Embed the one-dimensional learnable absolute position matrix with local features Add to get the local features containing location information
[0138] The local features are analyzed through the multi-head cross attention interaction module and the global feature T g Interact to get interactive features
[0139] The multi-head cross attention interaction module includes: As the input of the query generator, the global feature T g As input to the key generator and value generator, LN represents layer normalization, and the formula of the multi-head cross attention interaction module is as follows:
[0140]
[0141] in, and is a learnable parameter, C l ,h represent the embedding dimension and the number of heads respectively;
[0142] Output of the multi-head cross-attention interaction module It can be expressed as:
[0143]
[0144] Among them, LN(·) represents layer normalization and MLP(·) represents a fully connected layer.
[0145] Use continuous deconvolution operations to transform the interaction features Mapped to a feature map of the same size as the corresponding local feature right and Concatenate along the channel dimension and use 3×3 convolution kernel to generate the final global-local interaction feature
[0146] (3-1-5) The decoder module includes multiple decoding layers, namely Dec N ,Dec N-1 ..., Dec2, Dec1. Specifically, there are four decoding layers, namely Dec4, Dec3, Dec2 and Dec1.
[0147] The output local features F of the Transformer layer lM and interaction features Input the decoding layer Dec N Receive, get output Dec m Receive decoding layer Dec m+1 Output and interaction features Get Output m is a natural number from 1 to N-1; the decoding layer restores the spatial resolution through upsampling and convolution operations. That is, the output local feature F of the Transformer layer l4 and interaction features The input decoding layer Dec4 receives and obtains the output Dec3 receives the output of the decoding layer Dec4 and interaction features Get Output Dec2 receives the output of the decoding layer Dec3 and interaction features Get Output Dec1 receives the output of the decoding layer Dec2 and interaction features Get Output
[0148] (3-1-6) The output module includes a convolutional layer to receive the output of the last decoding layer. Specifically, a 1×1 convolutional layer is used to receive the output of the last decoding layer to generate the final segmentation result map.
[0149] (3-1-7) Output two corresponding first probability feature maps F for PVS segmentation L and the second probability feature map F S In order to integrate F L and F S , crop F L So that the trimmed F' L With F S Same size.
[0150] (4) Select F' L and F SThe larger value of each pixel in is used as the segmentation probability feature map of the final image block.
[0151] Then calculate F' at the pixel level L and F S The maximum probability of the two is used as the segmentation probability feature map of the second image block.
[0152] (5) Use the training set data and the validation set data to train the segmentation neural network model to obtain the optimal parameter model.
[0153] The proposed segmentation network is implemented on the pytorch framework and trained on an NVIDIA A100 GPU for 1000 epochs. The batch size is set to 16. A combination of cross entropy loss and Dice loss is used as the loss function. The initial learning rate of the entire model is set to 1e-3, and the initial learning rate of the deconvolution layer in the decoder is set to 1e-4. The cosine annealing scheduling strategy is used to update the learning rate, and the Adam optimizer is used to update the parameters of the network.
[0154] (6) Use the trained network to test the test set data and automatically obtain the perivascular space segmentation results.
[0155] The image segmentation method of the present invention is compared with five other segmentation methods in terms of segmentation performance. The PVS segmentation performance of these methods is shown in Table 1. Figure 7 , Figure 7 That is a schematic diagram comparing the processing results of other segmentation methods and the image segmentation method provided by the present invention.
[0156] Compared with other methods, the segmentation method provided by the present invention achieves the highest Dice value. At the same time, the segmentation results of the test set are visualized to intuitively compare the performance of different methods. It can be seen that the segmentation method of this solution is more accurate in segmenting the perivascular space.
[0157] Table 1
[0158]
[0159] U-Net was proposed by Ronneberger et al. in 2015. It is a convolutional neural network architecture composed of an encoder and a decoder for biomedical image segmentation. The encoder gradually reduces the spatial resolution of the feature map through convolutional layers and pooling layers, and the decoder gradually restores the spatial resolution through convolution and upsampling layers. At the same time, the feature map in the encoder is directly passed to the corresponding layer in the decoder through multiple jump connections to help restore high-resolution features (U-net: Convolutional networks for biomedical image segmentation). UNet++ is based on U-Net by Zhou et al., who used a series of densely connected modules to replace direct jump connections in the jump connection part to improve the performance of the network (Unet++: A nested u-net architecture for medical image segmentation). AttU-Net is based on U-Net by Oktay et al., who introduced an attention mechanism to adaptively select important features and suppress irrelevant features (Attention u-net: Learning where to look for the pancreas). nnU-Net is an adaptive deep learning framework proposed by Isensee et al., which aims to automatically configure the optimal U-Net model to adapt to different medical image segmentation tasks (nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation); M 2 EDN is a CNN model for PVS segmentation with end-to-end multi-channel input and multi-scale feature extraction proposed by Lian et al. (Multi-channel multi-scale fully convolutional network for 3D perivascular spaces segmentation in 7T MR images); GLoMS-Net is the image segmentation method of the present invention.
[0160] Compared with the above methods, the segmentation algorithm of the present invention adds a Transformer-based global-local feature interaction subnetwork to compensate for the information loss caused by downsampling in the encoder, and uses a multi-scale feature learning module and a multi-scale fusion strategy, which can ultimately capture more accurate PVS detail information and thus improve segmentation accuracy. As can be seen from Table 1, the method of the present invention has the best segmentation performance, with a Dice value of 0.678, which is higher than the existing method model. Compared with U-Net, UNet++, AttU-Net, nnU-Net and M 2 For EDN, the segmentation accuracy of the segmentation algorithm of the present invention is improved by 12.6%, 8.0%, 10.1%, 26.7% and 9.9% respectively. Figure 7 Four segmentation cases randomly selected from the test set are shown. Each row in the figure represents an independent case. The first and second columns represent the processed images and the true annotation results, and the other columns represent the results of six different methods. Figure 7 (Details pointed by circles and arrows) It can be seen that compared with the other five methods, the segmentation algorithm of the present invention can segment more PVS details, which is closer to the real annotation. It can be seen that the segmentation method of the present invention can effectively segment these tiny PVS detail structures, thereby improving the segmentation performance (Dice value) compared with other segmentation algorithms.
[0161] The above content describes the image segmentation method provided by the present invention. To support the above embodiment, the present invention also provides an image segmentation device. Figure 8 , an embodiment of the image segmentation device provided by the present invention includes:
[0162] An acquisition unit, used for acquiring an image to be processed;
[0163] A segmentation unit, used for inputting the image to be processed into the segmentation neural network model to obtain a segmentation result map;
[0164] The segmentation neural network model includes: an input module, an encoder module, a plurality of global-local feature interaction sub-networks, a decoder module and an output module;
[0165] The input module is used to receive the image to be processed; the encoder module includes multiple encoding layers Enc i , i is a natural number from 1 to N, N is the number of coding layers; and multiple Transformer layers TL j , j is a natural number from 1 to M, M is the number of Transformer layers; the encoding layer includes a multi-scale feature learning module; the output local feature of each encoding layer The output local features of each Transformer layer are F lj;
[0166] The global-local feature interaction subnetwork is used to receive the local features output by each encoding layer and extract the global features F based on the image to be processed. g , and obtain interactive features through the interaction of global features and local features
[0167] The decoder module includes multiple decoding layers, namely DEc N ,DEc N-1 ...,Dec2,Dec1;
[0168] The output local feature F of the Transformer layer lM and interaction features Input the decoding layer Decm N Receive, get output
[0169] Dec m Receive decoding layer Dec m+1 Output and interaction features Get Output m is a natural number from 1 to N-1;
[0170] The decoding layer restores the spatial resolution through upsampling and convolution operations;
[0171] The output module includes a 1×1 convolutional layer for receiving the output of the last decoding layer. Get the segmentation result image.
[0172] The segmentation neural network model is obtained by training based on an MR image training set, wherein the MR image training set includes MR images in which perivascular spaces have been annotated.
[0173] In this embodiment, the process performed by each unit is the same as the above Figure 1 The method flow described in the corresponding embodiment is similar, and the segmentation neural network model used by the present device may include the aforementioned Figures 2 to 6 The structure corresponding to any embodiment will not be described in detail here.
[0174] Fig. 9 It is a structural diagram of an image segmentation device provided by an embodiment of the present invention. The image segmentation device 900 may include one or more central processing units (CPU) 901 and a memory 905. The memory 905 stores one or more application programs or data.
[0175] In this embodiment, the specific functional module division in the central processing unit 901 can be the same as the above Figure 8 The functional module division method of each unit described in is similar and will not be repeated here.
[0176] The memory 905 may be a volatile storage or a persistent storage. The program stored in the memory 905 may include one or more modules, and each module may include a series of instruction operations on the server. Furthermore, the central processor 901 may be configured to communicate with the memory 905 and execute a series of instruction operations in the memory 905 on the perivascular space segmentation device 900.
[0177] The image segmentation device 900 may further include one or more power supplies 902 , one or more wired or wireless network interfaces 903 , and one or more input and output interfaces 904 .
[0178] The CPU 901 can execute the aforementioned Figure 1 The operations performed by the image segmentation method in the illustrated embodiment will not be described in detail here.
[0179] An embodiment of the present invention further provides a computer storage medium, which is used to store computer software instructions used for the above-mentioned image segmentation method, and includes a program designed for executing the image segmentation method.
[0180] The image segmentation method can be as described above Figure 1 The image segmentation method described in .
[0181] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the equivalent transformation of the circuit, the division of the unit, is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.
[0182] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.
[0183] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0184] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent substitution or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. An image segmentation method, characterized in that: include: Get the image to be processed; Inputting the image to be processed into the segmentation neural network model to obtain a segmentation result graph; The segmentation neural network model includes: an input module, an encoder module, a plurality of global-local feature interaction sub-networks, a decoder module and an output module; The input module is used to receive the image to be processed; the encoder module includes multiple encoding layers Ennc i , i is a natural number from 1 to N, N is the number of coding layers; and multiple Transformer layers TL j , j is a natural number from 1 to M, M is the number of Transformer layers; the encoding layer includes a multi-scale feature learning module; the output local feature of each encoding layer The output local features of each Transformer layer are The global-local feature interaction subnetwork is used to receive the local features output by each encoding layer and extract the global features F based on the image to be processed. g , and obtain interactive features through the interaction of global features and local features The decoder module includes multiple decoding layers, namely Dec N ,Dec N-1 ..., Dec2, Dec1; The output local features of the Transformer layer and interaction features Input the decoding layer Dec N Receive, get output Dec m Receive decoding layer Dec m+1 Output and interaction features Get Output m is a natural number from 1 to N-1; The decoding layer restores the spatial resolution through upsampling and convolution operations; The output module includes a convolutional layer for receiving the output of the last decoding layer. Get the segmentation result image.
2. The image segmentation method according to claim 1, characterized in that: The global-local feature interaction subnetwork consists of a global feature extraction network based on a multi-scale Transformer and a global-local feature interaction module.
3. The image segmentation method according to claim 2, characterized in that: The multi-scale Transformer-based global feature extraction network includes: Taking an input image x of size H×W: Use the convolution kernel to perform a convolution operation on the input image x to generate a feature map F of size H×W×C x ; Three convolution kernels of different sizes are used to generate three feature maps of the same size H′×W′×C′ by adjusting the padding and stride {F a , F b , F c }; The feature map {F a , F b , F c }Reshape to size N f ×C′ embedding matrix, where N f =H′×W′ represents the number of tags, C′ is the number of channels for each feature, and the learnable absolute position embedding matrix is used To save the position information, and add the feature map to the embedding matrix to obtain the position embedding feature map {T a , T b , T c }; The position is embedded into the feature map {T a , T b , T c } Input into the Transformer encoder and reshape the output to the same size as the original feature map and original features; The reshaped feature map is then convolved along the channel dimension to generate the final global feature F g .
4. The image segmentation method according to claim 3, characterized in that: The global-local feature interaction module includes: Get the global feature F g and the local features The global feature F g and the local features The features are extracted through convolution operation and then reshaped to obtain the global features F′ g and the local features Among them, the global feature F′ g and the local features The same size; The position embedding matrix and the global feature F′ g Add to get the global feature T containing position information g ; Embed the position matrix with local features Add to get the local features containing location information The local features are analyzed through the multi-head cross attention interaction module and the global feature T g Interact to get interactive features Use continuous deconvolution operations to transform the interaction features Mapped to a feature map of the same size as the corresponding local feature right and Concatenate along the channel dimension and use convolution to generate the final global-local interaction features 5. The image segmentation method according to claim 4, characterized in that: The convolution operation includes: using continuous convolution operations and reshaping operations to transform the global feature F g Convert to Global Features Where N g represents the number of markers, C g Indicates the number of channels; The multi-head cross-attention interaction module includes: Local features As the input of the query generator, the global feature T g As input to the key generator and value generator, LN represents layer normalization, and the formula of the multi-head cross attention interaction module is as follows: in, and is a learnable parameter, C l , h represent the embedding dimension and the number of heads respectively; Output of the multi-head cross-attention interaction module It can be expressed as: Among them, LN(·) represents layer normalization and MLP(·) represents a fully connected layer.
6. The image segmentation method according to claim 1, characterized in that: The multi-scale feature learning module includes: By using the convolution operation of three convolution kernels, three feature maps {F1, F2, F3} are generated; Concatenate the feature maps {F1, F2, F3} along the channel dimension to form a multi-channel feature map F in ; The multi-channel feature map F is processed using a squeeze and excitation module in ; to increase the focus on highly correlated features and reduce the focus on weakly correlated features.
7. The image segmentation method according to claim 1, characterized in that: The segmentation neural network model is obtained by training the basic neural network model based on the training set; The method for training the basic neural network model based on the training set includes: using a combination of cross entropy loss and Dice loss as a loss function and using an Adam optimizer to train the model.
8. The image segmentation method according to claim 7, characterized in that: The method for training the basic neural network model based on the training set also includes: (1) performing data enhancement on the image to be processed, wherein the data enhancement includes cropping, rotation and translation; The cropping includes: randomly cropping the image to be processed into a plurality of first image blocks, and performing center cropping on the plurality of first image blocks to obtain a second image block; the first image block is larger than the second image block; Then, the first image block and the second image block are rotated and translated to obtain a data-enhanced first image block and a data-enhanced second image block; (2) Inputting the first image block after data enhancement and the corresponding second image block after data enhancement into the segmentation neural network model, respectively obtaining a first probability feature map F L and the second probability feature map F S ; For the first probability feature map F L Perform center cropping to obtain the first probability feature subgraph F′ L ; F′ L With F S Same size; (3) Select F′ L and F S The larger value of each pixel in is used as the segmentation probability feature map of the final image block.
9. An image segmentation device, characterized in that: include: An acquisition unit, used for acquiring an image to be processed; A segmentation unit, used for inputting the image to be processed into the segmentation neural network model to obtain a segmentation result map; The segmentation neural network model includes: an input module, an encoder module, a plurality of global-local feature interaction sub-networks, a decoder module and an output module; The input module is used to receive the image to be processed; the encoder module includes multiple encoding layers Enc i , i is a natural number from 1 to N, N is the number of coding layers; and multiple Transformer layers TL j , j is a natural number from 1 to M, M is the number of Transformer layers; the encoding layer includes a multi-scale feature learning module; the output local feature of each encoding layer The output local features of each Transformer layer are The global-local feature interaction subnetwork is used to receive the local features output by each encoding layer and extract the global features F based on the image to be processed. g , and obtain interactive features through the interaction of global features and local features The decoder module includes multiple decoding layers, namely Dec N , Dec N-1 ,..., Dec2, Dec1; The output local features of the Transformer layer and interaction features Input the decoding layer Dec N Receive, get output Dec m Receive decoding layer Dec m+1 Output and interaction features Get Output m is a natural number from 1 to N-1; The decoding layer restores the spatial resolution through upsampling and convolution operations; The output module includes a convolutional layer for receiving the output of the last decoding layer. Get the segmentation result image.
10. A computer-readable storage medium, characterized in that: The method comprises instructions, which, when executed on a computer, enable the computer to execute the image segmentation method according to any one of claims 1 to 8.
Citation Information
Patent Citations
Multi-organ segmentation method and device for simultaneously extracting and fusing global and local features
CN116258685A
Multi-modal brain tumor image segmentation method based on self-supervised learning
WO2024108522A1
