Three-dimensional medical image segmentation method and device based on U-shaped large-kernel deep convolutional network

Abstract

The invention relates to a three-dimensional medical image segmentation method and device based on a U-shaped large-kernel deep convolutional network, and the method comprises the steps: obtaining a three-dimensional medical image, inputting the three-dimensional medical image into an image segmentation model, and obtaining a segmentation result; wherein the image segmentation model comprises a large-kernel deep convolution, a feature extraction branch and a feature decoding branch; the feature extraction branch comprises four cascaded feature extraction stages; the large-kernel deep convolution is used for reducing the three-dimensional medical image to a specific resolution to obtain an initial feature f0; the initial feature f0 is input into four cascaded feature extraction stages for feature extraction in sequence, and advanced features f1-f4 with the scales from large to small are obtained; decoding the initial feature f0 and the advanced features f1-f4 to the same size as the input three-dimensional medical image through a feature decoding branch to obtain a fusion feature F1; performing residual convolution on the three-dimensional medical image through a Res block to obtain a fusion feature F2; and the fusion feature F1 and the fusion feature F2 are spliced to obtain a feature F3, the feature F3 is input into a 1 * 1 * 1 convolutional layer with a softmax activation function for prediction to obtain a final segmentation result, and the segmentation effect is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation, and in particular relates to a three-dimensional medical image segmentation method and device based on a U-shaped large-kernel deep convolutional network. Background Art

[0002] In modern medical diagnosis, computed tomography (CT) and magnetic resonance imaging (MRI) are two major imaging technologies that provide doctors with detailed views of the internal structure of the human body, playing an indispensable auxiliary role in the diagnosis and treatment of a variety of diseases. In particular, the accurate segmentation of multiple organs in three-dimensional medical images provides doctors with precise anatomical information, which helps to formulate more reasonable treatment plans. However, traditional manual segmentation methods are highly dependent on the doctor's professional knowledge and subjective judgment. Not only is it time-consuming, but the objectivity and accuracy of the segmentation results are difficult to guarantee.

[0003] With the continuous development of deep learning technology, automatic segmentation methods based on deep learning have gradually emerged, bringing new solutions to 3D medical image segmentation with their high efficiency, objectivity and accuracy. Such methods can learn the shape, texture and other characteristics of organs by training a large amount of data, thereby achieving accurate segmentation of multiple organs in a short time.

[0004] However, despite the significant progress made by deep learning in 3D medical image segmentation, it still faces many challenges. The shapes and positions of human organs are diverse, and there are general differences in anatomical structures, which increases the difficulty of segmentation. In addition, the overlap between organs and the decrease in contrast make accurate 3D medical image segmentation extremely difficult. These problems often lead to the lack of organ boundaries and incorrect organ types in many current segmentation methods in practical applications, which seriously affects the accuracy and reliability of the segmentation results. Summary of the invention

[0005] In order to solve the problems existing in the background technology, one aspect of the present invention provides a three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network, comprising:

[0006] A three-dimensional medical image is acquired, and the three-dimensional medical image is input into an image segmentation model to obtain a segmentation result; wherein the image segmentation model comprises: a large kernel deep convolution, a feature extraction branch and a feature decoding branch; the feature extraction branch comprises: four cascaded feature extraction stages; the large kernel deep convolution is used to reduce the three-dimensional medical image to a specific resolution to obtain an initial feature f0; the initial feature f0 is input into the four cascaded feature extraction stages for feature extraction in sequence to obtain high-level features f1 to f4 with scales from large to small; the initial feature f0 and the high-level features f1 to f4 are decoded into the same size as the input three-dimensional medical image through a feature decoding branch to obtain a fused feature F1; the three-dimensional medical image is subjected to residual convolution through a Res block to obtain a fused feature F2; feature F3 is obtained by splicing the fused feature F1 and the fused feature F2, and feature F3 is input into a 1×1×1 convolution layer with a softmax activation function for prediction to obtain a final segmentation result.

[0007] Another aspect of the present invention provides a three-dimensional medical image segmentation device based on a U-shaped large-core deep convolutional network, characterized in that it includes a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the three-dimensional medical image segmentation device based on the U-shaped large-core deep convolutional network performs the three-dimensional medical image segmentation method based on the U-shaped large-core deep convolutional network.

[0008] A computer-readable storage medium stores a program, characterized in that when the program is executed by a processor, a three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network is implemented.

[0009] The present invention has at least the following beneficial effects

[0010] By introducing large kernel deep convolution, the present invention can capture a wider range of contextual information and effectively cope with the challenges brought by organ shape, position diversity and anatomical structure differences; the combination of feature extraction branch and feature decoding branch enables the network to understand the complex features in medical images more deeply, thereby improving the accuracy and reliability of segmentation; the introduction of residual convolution enhances the robustness of the model, reduces problems such as missing organ boundaries and organ type errors, and improves the overall quality of segmentation results; the automation characteristics of deep learning methods greatly accelerate the segmentation process, and can produce segmentation results in a shorter time than manual segmentation, thereby improving the efficiency of medical diagnosis. The efficient network structure and algorithm optimization further shorten the processing time, making the present solution more feasible in practical applications; by introducing large kernel deep convolution, feature extraction branch, feature decoding branch and residual convolution and other technologies, the present solution significantly improves the accuracy, efficiency, objectivity and consistency of three-dimensional medical image segmentation, provides strong support for clinical decision-making and personalized treatment, and has important clinical application value and practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is the overall framework flow chart of the present invention;

[0012] Figure 2 Schematic diagram of the architecture of the UL block in the present invention;

[0013] Figure 3 Schematic diagram of the architecture of the large-kernel deep convolution module in the present invention;

[0014] Figure 4 It is a schematic diagram of the architecture of the sparse multi-layer perception machine layer of a three-dimensional image in the present invention;

[0015] Figure 5 It is a structural schematic diagram of the Res block in the present invention;

[0016] Figure 6 It is a schematic diagram comparing the sparse multi-layer perceptron layer of three-dimensional images and the common multi-layer perceptron layer of three-dimensional images in the present invention. DETAILED DESCRIPTION

[0017] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0018] See also Figure 1 The present invention provides a three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network, comprising:

[0019] A three-dimensional medical image is acquired, and the three-dimensional medical image is input into an image segmentation model to obtain a segmentation result; wherein the image segmentation model comprises: a large kernel deep convolution, a feature extraction branch and a feature decoding branch; the feature extraction branch comprises: four cascaded feature extraction stages; the large kernel deep convolution is used to reduce the three-dimensional medical image to a specific resolution to obtain an initial feature f0; the initial feature f0 is input into the four cascaded feature extraction stages for feature extraction in sequence to obtain high-level features f1 to f4 with scales from large to small; the initial feature f0 and the high-level features f1 to f4 are decoded into the same size as the input three-dimensional medical image through a feature decoding branch to obtain a fused feature F1; the three-dimensional medical image is subjected to residual convolution through a Res block to obtain a fused feature F2; feature F3 is obtained by splicing the fused feature F1 and the fused feature F2, and feature F3 is input into a 1×1×1 convolution layer with a softmax activation function for prediction to obtain a final segmentation result.

[0020] In this example, ULD-Net is implemented using Pytorch1 and MONAI2. All experiments of this method are performed on an NVIDIA RTX A6000 server. All experiments of ULD-Net (including ablation experiments) use the AdamW optimizer with an initial learning rate of 0.0001. We use the Dice similarity coefficient as an evaluation metric to compare the overlap area between the predicted and ground-truth labels, i.e., the accuracy of the segmentation. The Dice metric is defined as:

[0021]

[0022] Among them, G i and P i They represent the true value and predicted value of voxel i respectively, and I represents the number of voxels.

[0023] Preferably, the large kernel depth convolution uses a three-dimensional convolution with a size of 7×7×7 and a padding of 3 to reduce the resolution of the input three-dimensional medical image to half of the original resolution.

[0024] In this embodiment, due to the inherent redundancy in the image, both the previous convolution-based method and the Transformer-based method have a module for reducing the feature representation to a specific size. A common method based on convolution is to use a downsampling layer. The Transformer-based method uses a more aggressive patch layer, using a patch size of 4 to adapt to its multi-level architecture design. Combining the advantages of the two methods, a three-dimensional convolution kernel with a convolution kernel size of 7 and a padding of 3 is designed to reduce the resolution of the three-dimensional image to half of the original.

[0025] Preferably, in the first feature extraction stage of the feature extraction branch, feature extraction is performed on the input initial feature f0 to obtain the high-level feature F1; in the second feature extraction stage of the feature extraction branch, feature extraction is performed on the input high-level feature F1 to obtain the high-level feature F2; in the third stage of the feature extraction branch, feature extraction is performed on the input high-level feature F2 to obtain the high-level feature F3; in the fourth stage of the feature extraction branch, feature extraction is performed on the input high-level feature F3 to obtain the high-level feature F4.

[0026] See also Figure 2 Preferably, in each feature extraction stage of the feature extraction branch, a series of UL modules and downsampling layers are used to extract input features, and the resolution of the input features is reduced to 1 / 2 of the original in the downsampling layer; wherein the UL module includes: two cascaded feature extraction modules; each feature extraction module includes a first normalization layer, a large kernel deep convolution module, a first Gaussian error linear unit activation function, a second normalization layer, a three-dimensional image sparse multi-layer perceptron layer and a second Gaussian error linear unit activation function connected in series; the input features of the first normalization layer and the output features of the first Gaussian error linear unit activation function are added to obtain the input features of the second normalization layer; the input features of the second normalization layer and the output features of the second Gaussian error linear unit activation function are added to obtain the output features of the feature extraction module.

[0027] In this embodiment, the large-core deep convolution module extracts the input feature map and reduces the three-dimensional image resolution to half of the original through the deep three-dimensional convolution kernel with convolution kernels of 9, 7, 5, and 3, to obtain four different feature outputs; the four different feature outputs are spliced ​​by connection rather than addition; the splicing result of the four channels is reduced to a single channel through the fully connected layer; layer normalization is used to accelerate convergence and reduce overfitting; a variant of the rectified linear unit activation function, the Gaussian error linear unit activation function, is used to solve the problem that the rectified linear unit activation function suddenly becomes 0 in the negative input range.

[0028] In the UL module, the image resolution is reduced by 2 times each time, features are extracted step by step, and the results of each depth are saved;

[0029]

[0030] Among them, BLDC and UMLP are respectively the large-core deep convolution module and the 3D image sparse multi-layer perception machine layer of the cascade structure; and is the output of the large kernel depth convolution block with cascade structure of different depths; l and z l+1 The output of the 3D large kernel deep convolution sparse multilayer perceptron layer of different depths. The result of each depth will be saved as one of the inputs for the result merging.

[0031] Preferably, the large-core deep convolution module includes: three-dimensional convolution of size 9×9×9, three-dimensional convolution of size 7×7×7, three-dimensional convolution of size 5×5×5, three-dimensional convolution of size 3×3×3 and a linear layer; the three-dimensional convolution of size 9×9×9, the three-dimensional convolution of size 7×7×7, the three-dimensional convolution of size 5×5×5 and the three-dimensional convolution of size 3×3×3 are respectively used to extract features from the input feature map, and each three-dimensional convolution kernel reduces the resolution of the input feature map to 1 / 2 of the original; the output features of the three-dimensional convolution of size 9×9×9, the three-dimensional convolution of size 7×7×7, the three-dimensional convolution of size 5×5×5 and the three-dimensional convolution of size 3×3×3 are concatenated to obtain feature Y0, and the linear layer is used to linearly map the feature Y0 to map the multi-channel of the input feature Y0 to a single channel to obtain the output feature.

[0032] See also Figure 3In this embodiment, a cascaded large-core deep convolution module (three-dimensional convolution of size 9×9×9, three-dimensional convolution of size 7×7×7, three-dimensional convolution of size 5×5×5, and three-dimensional convolution of size 3×3×3) is used to extract the features of the input image. Functionally, each channel of the input in the deep convolution has a corresponding convolution kernel for calculation, so that the input features and the output features have the same dimension. The large-core deep convolution can accommodate a considerable area when calculating. This means that they can provide a larger receptive field. In terms of performance, it is generally believed that the amount of calculation and the number of parameters of large-core convolution will increase with the square of the kernel size, and deep convolution can greatly overcome this problem by relying on its inherent characteristics. At the same time, deep convolution has relatively low performance on modern parallel computing devices such as GPUs due to the introduction of lower computing and memory access costs, but the large convolution kernel reduces the number of times the video memory is accessed, which can improve the efficiency of deep convolution to a certain extent. Therefore, deep convolution with a large kernel size has a high degree of matching with the three-dimensional medical image segmentation task in terms of function and performance. Large kernel deep convolution relies on the inherent characteristics of large receptive field self-attention and convolution to extract features in the global range. However, due to the fixed convolution kernel size, large kernel deep convolution also has a defect that is difficult to ignore. In medical volume image datasets such as abdominal multi-organ CT images, there are local position and morphological connections between different organs. These connections are often very hidden due to anatomical differences and organ overlap. If the distance between these connected organs exceeds the range of large kernel deep convolution, it will be difficult to find these connections. Therefore, inspired by the concept of reparameterization, large kernel deep convolution kernels of different sizes are arranged using a reparameterization method. Since the model does not need to share parameters and weights in the test phase, the structure of structural reparameterization is directly used. This structure can take into account and balance the feature representation of large kernel deep convolutions of different sizes, and can improve the performance of the model compared to the method of using large kernel deep convolutions of different sizes alone. The improved structure is called the branched LK deep convolution block.

[0033] See also Figure 4 Preferably, the three-dimensional image sparse multi-layer perceptron layer comprises:

[0034] X W =Linea rW→W (HDC×W)

[0035] X H =Linear H→H (WDC×H)

[0036] X D =Linear D→D (HWC×D)

[0037] X out =Linear4→1 (concat(X W , X H , X D , X))

[0038] Among them, X out represents the output features of the sparse multi-layer perceptron layer of the three-dimensional image, X w , X H and X D They represent the output results of the mixed information of the three-dimensional image sparse multi-layer perceptron layer in three dimensions; Linear w→W (HDC×W) means using the linear layer function provided by the torch nn package to calculate the linear relationship between the input features in dimension H, dimension D, and number of channels C, and taking dimension W as the weight without calculating it; Linear H→H (WDC×H) means using the linear layer function provided by torch's nn package to calculate the linear relationship between the input features in dimension W, dimension D, and number of channels C, and taking dimension H as the weight without calculating it; Linear D→D (HWC×D) means using the linear layer function provided by torch's nn package to calculate the linear relationship between the input features in dimension H, dimension W, and number of channels C, and taking dimension D as the weight without calculating it; Linear 4→1 (concat(X W , X H , X D , X)) represents the concatenation of the output result of the mixed information of the three-dimensional image sparse multilayer perceptron layer and the input features, and then reducing the four channels to one channel through the linear layer function; X represents the input features of the three-dimensional image sparse multilayer perceptron layer.

[0039] The multi-layer perceptron layer uses more parameters and higher computational complexity to achieve feature scaling. Therefore, the present invention hopes to reduce the number of parameters and computational complexity of the multi-layer perceptron layer as much as possible without reducing the model segmentation performance. Therefore, the sparse multi-layer perceptron layer with sparsity and weight sharing is selected for application in a two-dimensional network. These characteristics enable it to reduce computational complexity while modeling global dependencies. The comparison between the sparse multi-layer perceptron layer of a three-dimensional image and the ordinary multi-layer perceptron layer of a three-dimensional image is shown in FIG. Figure 6 shown.

[0040] Preferably, decoding the initial feature f00 and the advanced features f1-f4 to the same size as the input three-dimensional medical image through the feature decoding branch to obtain the fused feature F1 comprises:

[0041] The initial feature f0 and high-level features f1~f4 are respectively subjected to residual convolution through a Res block to obtain features f01, f11, f21, f31 and f41;

[0042] After upsampling feature f41 to double its original resolution, it is concatenated with feature f31 and then subjected to residual convolution through a Res block to obtain feature f341;

[0043] After upsampling feature f341 to double its original resolution, it is concatenated with feature f21 and then subjected to residual convolution through a Res block to obtain feature f231;

[0044] After upsampling feature f231 to double its original resolution, it is concatenated with feature f11 and then subjected to residual convolution through a Res block to obtain feature f121;

[0045] After upsampling feature f121 to double its original resolution, it is concatenated with feature f01 and then subjected to residual convolution through a Res block to obtain feature f011;

[0046] The feature f011 is upsampled to make its resolution twice as high as the original one to obtain the fused feature F1.

[0047] See also Figure 5 Preferably, the Res block includes two 3×3×3 three-dimensional convolutions connected in series; the input features of the first 3×3×3 three-dimensional convolution and the output features of the second 3×3×3 three-dimensional convolution are added to obtain the output features of the Res block, wherein the input features of the first 3×3×3 three-dimensional convolution are the input features of the Res block.

[0048] The result of each depth is fed into a residual block consisting of two convolutional layers with a kernel size of 3×3×3, and then fed into instance normalization to ensure stable extracted features. These feature maps have their resolution doubled after the deconvolution layer, and the input of the previous depth is concatenated with itself. Subsequently, the concatenated features are fed into the next residual block. Finally, a 1×1×1 convolutional layer with a softmax activation function is used to predict the final segmentation result. The loss function used is a combination of soft Dice loss and cross entropy loss, which is calculated as:

[0049]

[0050] Where I is the number of voxels; J is the number of classes; Y i,j and G I,j are the output probability of class j at voxel i and the one-hot encoded ground truth, respectively.

[0051] Another aspect of the present invention provides a three-dimensional medical image segmentation device based on a U-shaped large-core deep convolutional network, characterized in that it includes a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the three-dimensional medical image segmentation device based on the U-shaped large-core deep convolutional network performs the three-dimensional medical image segmentation method based on the U-shaped large-core deep convolutional network.

[0052] A computer-readable storage medium stores a program, characterized in that when the program is executed by a processor, a three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network is implemented.

[0053] In summary, by introducing large kernel deep convolution, the present invention can capture a wider range of contextual information and effectively cope with the challenges brought by organ shape, position diversity and anatomical structure differences; the combination of feature extraction branch and feature decoding branch enables the network to understand the complex features in medical images more deeply, thereby improving the accuracy and reliability of segmentation; the introduction of residual convolution enhances the robustness of the model, reduces problems such as missing organ boundaries and organ type errors, and improves the overall quality of segmentation results; the automation characteristics of deep learning methods greatly accelerate the segmentation process, and can produce segmentation results in a shorter time than manual segmentation, thereby improving the efficiency of medical diagnosis. The efficient network structure and algorithm optimization further shorten the processing time, making the present scheme more feasible in practical applications; by introducing large kernel deep convolution, feature extraction branch, feature decoding branch and residual convolution and other technologies, the present scheme significantly improves the accuracy, efficiency, objectivity and consistency of three-dimensional medical image segmentation, provides strong support for clinical decision-making and personalized treatment, and has important clinical application value and practical significance.

[0054] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network, characterized in that: include: A three-dimensional medical image is acquired, and the three-dimensional medical image is input into an image segmentation model to obtain a segmentation result; wherein the image segmentation model comprises: a large kernel deep convolution, a feature extraction branch and a feature decoding branch; the feature extraction branch comprises: four cascaded feature extraction stages; the large kernel deep convolution is used to reduce the three-dimensional medical image to a specific resolution to obtain an initial feature f0; the initial feature f0 is input into the four cascaded feature extraction stages for feature extraction in sequence to obtain high-level features f1 to f4 with scales from large to small; the initial feature f0 and the high-level features f1 to f4 are decoded into the same size as the input three-dimensional medical image through a feature decoding branch to obtain a fused feature F1; the three-dimensional medical image is subjected to residual convolution through a Res block to obtain a fused feature F2; feature F3 is obtained by splicing the fused feature F1 and the fused feature F2, and feature F3 is input into a 1×1×1 convolution layer with a softmax activation function for prediction to obtain a final segmentation result.

2. The three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network according to claim 1, characterized in that: The large kernel depth convolution uses a three-dimensional convolution with a size of 7×7×7 and a padding of 3 to reduce the resolution of the input three-dimensional medical image to half of the original resolution.

3. The three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network according to claim 1, characterized in that: In the first feature extraction stage of the feature extraction branch, feature extraction is performed on the input initial feature f0 to obtain the high-level feature F1, and in the second feature extraction stage of the feature extraction branch, feature extraction is performed on the input high-level feature F1 to obtain the high-level feature F2; In the third stage of the feature extraction branch, feature extraction is performed on the input high-level feature F2 to obtain high-level feature F3; in the fourth stage of the feature extraction branch, feature extraction is performed on the input high-level feature F3 to obtain high-level feature F4.

4. The three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network according to claim 3, characterized in that: In each feature extraction stage of the feature extraction branch, a series of UL modules and downsampling layers are used to extract input features, and the resolution of the input features is reduced to 1 / 2 of the original in the downsampling layer; wherein the UL module includes: two cascaded feature extraction modules; each feature extraction module includes a first normalization layer, a large kernel deep convolution module, a first Gaussian error linear unit activation function, a second normalization layer, a three-dimensional image sparse multi-layer perceptron layer and a second Gaussian error linear unit activation function connected in series; the input features of the first normalization layer and the output features of the first Gaussian error linear unit activation function are added to obtain the input features of the second normalization layer; the input features of the second normalization layer and the output features of the second Gaussian error linear unit activation function are added to obtain the output features of the feature extraction module.

5. The three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network according to claim 4, characterized in that: The large-core deep convolution module includes: three-dimensional convolution of size 9×9×9, three-dimensional convolution of size 7×7×7, three-dimensional convolution of size 5×5×5, three-dimensional convolution of size 3×3×3 and linear layer; three-dimensional convolution of size 9×9×9, three-dimensional convolution of size 7×7×7, three-dimensional convolution of size 5×5×5 and three-dimensional convolution of size 3×3×3 are respectively used to extract features from the input feature map, and each three-dimensional convolution kernel reduces the resolution of the input feature map to 1 / 2 of the original; the output features of the three-dimensional convolution of size 9×9×9, the three-dimensional convolution of size 7×7×7, the three-dimensional convolution of size 5×5×5 and the three-dimensional convolution of size 3×3×3 are concatenated to obtain feature Y0, and the linear layer is used to linearly map the feature Y0 to map the multi-channel of the input feature Y0 to a single channel to obtain the output feature.

6. The three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network according to claim 4, characterized in that: The three-dimensional image sparse multi-layer perception machine layer includes: X W =Linear W→W (HDC×W) X H =Linear H→H (WDC×H) X D =Linear D→D (HWC×D) X out =Linear 4→1 (concat(X W ,X H ,X D ,X)) Among them, X out represents the output features of the sparse multi-layer perceptron layer of the three-dimensional image, X W , X H and X D They represent the output results of the mixed information of the three-dimensional image sparse multi-layer perceptron layer in three dimensions; Linear W→W (HDC×W) means using the linear layer function provided by the torch nn package to calculate the linear relationship between the input features in dimension H, dimension D, and number of channels C, and taking dimension W as the weight without calculating it; Linear H→H (WDC×H) means using the linear layer function provided by torch's nn package to calculate the linear relationship between the input features in dimension W, dimension D, and number of channels C, and taking dimension H as the weight without calculating it; Linear D→D (HWC×D) means using the linear layer function provided by torch's nn package to calculate the linear relationship between the input features in dimension H, dimension W, and number of channels C, and taking dimension D as the weight without calculating it; Linear 4→1 (concat(X W , X H , X D , X)) represents the concatenation of the output result of the mixed information of the three-dimensional sparse multilayer perceptron layer and the input features, and then reducing the four channels to one channel through the linear layer function; X represents the input features of the three-dimensional sparse multilayer perceptron layer.

7. The three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network according to claim 1, characterized in that: The initial feature f0 and the advanced features f1-f4 are decoded to the same size as the input three-dimensional medical image through the feature decoding branch to obtain the fused feature F1, which includes: The initial feature f0 and high-level features f1~f4 are respectively subjected to residual convolution through a Res block to obtain features f01, f11, f21, f31 and f41; After upsampling feature f41 to double its original resolution, it is concatenated with feature f31 and then subjected to residual convolution through a Res block to obtain feature f341; After upsampling feature f341 to double its original resolution, it is concatenated with feature f21 and then subjected to residual convolution through a Res block to obtain feature f231; After upsampling feature f231 to double its original resolution, it is concatenated with feature f11 and then subjected to residual convolution through a Res block to obtain feature f121; After upsampling feature f121 to double its original resolution, it is concatenated with feature f01 and then subjected to residual convolution through a Res block to obtain feature f011; The feature f011 is upsampled to make its resolution twice as high as the original one to obtain the fused feature F1.

8. The three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network according to claim 1, characterized in that: The Res block includes two 3×3×3 three-dimensional convolutions connected in series; the input features of the first 3×3×3 three-dimensional convolution and the output features of the second 3×3×3 three-dimensional convolution are added to obtain the output features of the Res block, wherein the input features of the first 3×3×3 three-dimensional convolution are the input features of the Res block.

9. A three-dimensional medical image segmentation device based on a U-shaped large-kernel deep convolutional network, characterized in that: It comprises a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the three-dimensional medical image segmentation device based on a U-shaped large-kernel deep convolutional network performs a three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, a three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network as described in any one of claims 1 to 8 is implemented.

Images