A method and apparatus for three-dimensional medical image segmentation based on a U-shaped large kernel deep convolutional network

Abstract

This invention relates to a three-dimensional medical image segmentation method and apparatus based on a U-shaped large-kernel deep convolutional network, comprising: acquiring a three-dimensional medical image; inputting the three-dimensional medical image into an image segmentation model to obtain a segmentation result; wherein, the image segmentation model includes: a large-kernel deep convolution, a feature extraction branch, and a feature decoding branch; the feature extraction branch includes: 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 initial feature f0; the initial feature f0 is input into the four cascaded feature extraction stages for sequential feature extraction to obtain high-level features f1 to f4 with scales decreasing from large to small; the initial feature f0 and the high-level features f1 to f4 are decoded by the feature decoding branch to the same size as the input three-dimensional medical image to obtain a fusion feature F1; the three-dimensional medical image is subjected to residual convolution through a Res block to obtain a fusion feature F2; the fusion feature F1 and the fusion feature F2 are concatenated to obtain feature F3; feature F3 is input into a 1×1×1 convolutional layer with a softmax activation function to predict the final segmentation result, thereby improving the segmentation effect.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, and in particular relates to a three-dimensional medical image segmentation method and apparatus based on a U-shaped large kernel deep convolutional network. Background Technology

[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 structures of the human body, playing an indispensable role in the diagnosis and treatment of various diseases. In particular, the accurate segmentation of multiple organs in three-dimensional medical images provides doctors with precise anatomical information, helping to formulate more rational treatment plans. However, traditional manual segmentation methods rely heavily on the doctor's professional knowledge and subjective judgment, which is not only time-consuming but also makes it difficult to guarantee the objectivity and accuracy of the segmentation results.

[0003] With the continuous development of deep learning technology, deep learning-based automatic segmentation methods have gradually emerged, offering new solutions for 3D medical image segmentation due to their high efficiency, objectivity, and accuracy. These methods, trained on large amounts of data, can learn the shape, texture, and other features of organs, thus 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, many challenges remain. The diverse shapes and locations of human organs, along with the widespread differences in anatomical structures, increase the difficulty of segmentation. Furthermore, overlap between organs and decreased contrast make accurate 3D medical image segmentation exceptionally difficult. These problems often lead to issues such as missing organ boundaries and incorrect organ classification in practical applications, severely impacting the accuracy and reliability of the segmentation results. Summary of the Invention

[0005] To address the problems existing in the background art, 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 3D medical image is acquired and input into an image segmentation model to obtain the segmentation result. The image segmentation model includes a large kernel depthwise convolution, a feature extraction branch, and a feature decoding branch. The feature extraction branch includes four cascaded feature extraction stages. The large kernel depthwise convolution is used to reduce the 3D medical image to a specific resolution to obtain initial feature f0. The initial feature f0 is input into the four cascaded feature extraction stages for sequential feature extraction, obtaining high-level features f1 to f4 with scales decreasing from large to small. The initial feature f0 and high-level features f1 to f4 are decoded by the feature decoding branch to the same size as the input 3D medical image to obtain fusion feature F1. The 3D medical image is subjected to residual convolution through a Res block to obtain fusion feature F2. The fusion features F1 and F2 are concatenated to obtain feature F3, which is then input into a 1×1×1 convolutional layer with a softmax activation function for prediction to obtain the final segmentation result.

[0007] Another aspect of the present invention provides a three-dimensional medical image segmentation device based on a U-shaped large kernel depth 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 a U-shaped large kernel depth convolutional network performs the three-dimensional medical image segmentation method based on a U-shaped large kernel depth convolutional network.

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

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

[0010] This invention, by introducing large-kernel deep convolution, enables the network to capture broader contextual information, effectively addressing the challenges posed by the diversity of organ shapes, locations, and anatomical structures. The combination of feature extraction and feature decoding branches allows the network to gain a deeper understanding of complex features in medical images, thereby improving the accuracy and reliability of segmentation. The introduction of residual convolution enhances the model's robustness, reducing issues such as missing organ boundaries and incorrect organ classification, thus improving the overall quality of the segmentation results. The automation nature of deep learning methods significantly accelerates the segmentation process, producing results in a shorter time compared to manual segmentation, improving the efficiency of medical diagnosis. The efficient network structure and algorithm optimization further shorten processing time, making this solution more feasible in practical applications. By introducing techniques such as large-kernel deep convolution, feature extraction branches, feature decoding branches, and residual convolution, this solution significantly improves the accuracy, efficiency, objectivity, and consistency of 3D medical image segmentation, providing strong support for clinical decision-making and personalized treatment, and possessing significant clinical application value and practical significance. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the overall framework of the present invention;

[0012] Figure 2 This is a schematic diagram of the architecture of the UL block in this invention;

[0013] Figure 3 This is a schematic diagram of the architecture of the large kernel depthwise convolution module in this invention;

[0014] Figure 4 This is a schematic diagram of the architecture of the sparse multilayer perceptron layer for three-dimensional images in this invention;

[0015] Figure 5 This is a schematic diagram of the Res block structure in this invention;

[0016] Figure 6 This is a schematic diagram comparing the sparse multilayer perceptron layer of a 3D image with the ordinary multilayer perceptron layer of a 3D image in this invention. Detailed Implementation

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0018] Please see Figure 1 This invention provides a three-dimensional medical image segmentation method based on a U-shaped large-kernel deep convolutional network, comprising:

[0019] A 3D medical image is acquired and input into an image segmentation model to obtain the segmentation result. The image segmentation model includes a large kernel depthwise convolution, a feature extraction branch, and a feature decoding branch. The feature extraction branch includes four cascaded feature extraction stages. The large kernel depthwise convolution is used to reduce the 3D medical image to a specific resolution to obtain initial feature f0. The initial feature f0 is input into the four cascaded feature extraction stages for sequential feature extraction, obtaining high-level features f1 to f4 with scales decreasing from large to small. The initial feature f0 and high-level features f1 to f4 are decoded by the feature decoding branch to the same size as the input 3D medical image to obtain fusion feature F1. The 3D medical image is subjected to residual convolution through a Res block to obtain fusion feature F2. The fusion features F1 and F2 are concatenated to obtain feature F3, which is then input into a 1×1×1 convolutional layer with a softmax activation function for prediction to obtain the final segmentation result.

[0020] In this embodiment, ULD-Net was implemented using PyTorch1 and MONAI2. All experiments were performed on an NVIDIA RTX A6000 server. All ULD-Net experiments (including ablation experiments) used the AdamW optimizer with an initial learning rate of 0.0001. We used the Dice similarity coefficient as an evaluation metric to compare the overlap 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 Let i represent the true value and the predicted value of voxel i, respectively, and let I represent the number of voxels.

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

[0024] In this embodiment, due to the inherent redundancy in images, previous convolutional-based and Transformer-based methods both included a module to reduce the feature representation to a specific size. A common approach for convolutional-based methods is to use downsampling layers. Transformer-based methods employ more aggressive patch layers, using a patch size of 4 to accommodate their multi-level architecture. Combining the advantages of both methods, a design was developed that uses a 3D convolutional kernel with a kernel size of 7 and padding of 3 to reduce the resolution of the 3D image to half its original value.

[0025] Preferably, in the first feature extraction stage of the feature extraction branch, the input initial feature f0 is used to extract features to obtain high-level features F1; in the second feature extraction stage of the feature extraction branch, the input high-level features F1 are used to extract features to obtain high-level features F2; ​​in the third stage of the feature extraction branch, the input high-level features F2 are used to extract features to obtain high-level features F3; and in the fourth stage of the feature extraction branch, the input high-level features F3 are used to extract features to obtain high-level features F4.

[0026] Please see Figure 2 Preferably, in each feature extraction stage of the feature extraction branch, a cascaded UL module and a downsampling layer are used to extract the input features, and the resolution of the input features is reduced to half of the original in the downsampling layer; wherein, the UL module includes: two cascaded feature extraction modules; each feature extraction module includes a cascaded first normalization layer, a large kernel depth convolution module, a first Gaussian error linear unit activation function, a second normalization layer, a 3D image sparse multilayer perceptron layer, and a second Gaussian error linear unit activation function; 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 kernel depth convolution module extracts the input feature map by using a depth 3D convolution kernel with kernels of 9, 7, 5, and 3 to reduce the resolution of the 3D image to half of its original value, resulting in four different feature outputs. The four different feature outputs are concatenated instead of added. The concatenated result of four channels is reduced to a single channel through a 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, which is smoother than the rectified linear unit activation function, thus solving 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 a factor of 2 each time, features are extracted progressively, and the results for each depth are saved.

[0029]

[0030] Among them, BLDC and UMLP are respectively a cascaded large-kernel deep convolutional module and a 3D image sparse multilayer perceptron layer; and The outputs of large-kernel depthwise convolutional blocks with cascaded structures of different depths; z l and z l+1 This represents the output of a 3D large-kernel depthwise convolutional sparse multilayer perceptron layer at different depths. The result for each depth is saved and used as one of the inputs for merging the results.

[0031] Preferably, the large kernel depthwise convolution module includes: a 9×9×9 three-dimensional convolution, a 7×7×7 three-dimensional convolution, a 5×5×5 three-dimensional convolution, a 3×3×3 three-dimensional convolution, and a linear layer; the 9×9×9, 7×7×7, 5×5×5, and 3×3×3 three-dimensional convolutions are used to extract features from the input feature map, and each three-dimensional convolution kernel reduces the resolution of the input feature map to half of its original value; the output features of the 9×9×9, 7×7×7, 5×5×5, and 3×3×3 three-dimensional convolutions are concatenated to obtain feature Y0, and the linear layer is used to linearly map feature Y0 to convert the multi-channel input feature Y0 into a single-channel output feature.

[0032] Please see Figure 3In this embodiment, a cascaded large-kernel deep convolution module (3D convolutions of size 9×9×9, 7×7×7, 5×5×5, and 3×3×3) is used to extract features from the input image. Functionally, each input channel in the deep convolution has a corresponding convolution kernel for computation, ensuring that the input and output features have the same dimension. Large-kernel deep convolutions can accommodate a considerable area during computation, meaning they can provide a large receptive field. In terms of performance, it is generally believed that the computational cost and number of parameters of large-kernel convolutions increase quadratically with the kernel size; deep convolutions, by virtue of their inherent characteristics, can greatly overcome this problem. Simultaneously, while deep convolutions have relatively low performance on modern parallel computing devices like GPUs due to the lower computational and memory access costs they introduce, large convolution kernels reduce the number of times memory is accessed, which can improve the efficiency of deep convolutions to some extent. Therefore, large-kernel deep convolutions are highly compatible with 3D medical image segmentation tasks in terms of both functionality and performance. Large kernel depthwise convolutions rely on the self-attention of a large receptive field and the inherent properties of convolution to extract features globally. However, limited by a fixed kernel size, large kernel depthwise convolutions also have a significant drawback. In medical volumetric image datasets such as abdominal multi-organ CT images, there are localized positional and morphological connections between different organs. These connections are often very subtle due to differences in anatomical structures and organ overlap. If the distance between these connected organs exceeds the range of the large kernel depthwise convolution, it becomes difficult to detect these connections. Therefore, inspired by the concept of reparameterization, large kernel depthwise convolution kernels of different sizes are arranged using a reparameterization method. Since the model does not need to share parameters and weights during the testing phase, a structure reparameterization structure is used directly. This structure can take into account and balance the feature representations of large kernel depthwise convolutions of different sizes, improving model performance compared to methods that use large kernel depthwise convolutions of different sizes alone. The improved structure is called a branched LK depthwise convolution block.

[0033] Please see Figure 4 Preferably, the 3D image sparse multilayer 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 X represents the output features of a sparse multilayer perceptron layer in a 3D image. w X H and X D These represent the output results of a sparse multilayer perceptron layer in a 3D image, where information is mixed across three dimensions; Linear w→W (HDC×W) indicates that the linear layer function provided by the nn package in torch is used to calculate the linear relationship between the input features in dimensions H, D, and the number of channels C, while dimension W is used as a weight and not processed in the calculation; Linear H→H (WDC×H) indicates that the linear layer function provided by the nn package in torch is used to calculate the linear relationship between the input features in dimension W, dimension D, and number of channels C, while dimension H is used as a weight and is not processed in the calculation; Linear D→D (HWC×D) indicates that the linear layer function provided by the nn package in torch is used to calculate the linear relationship between the input features in dimensions H, W, and C, while dimension D is used as a weight and not processed. 4→1 (concat(X W X H X D X)) represents the concatenation of the output and input features of the sparse multilayer perceptron layer of the 3D image in three dimensions, and then the reduction of 4 channels to 1 channel through a linear layer function; X represents the input features of the sparse multilayer perceptron layer of the 3D image.

[0039] Multilayer perceptrons (MPS) achieve feature scaling with a large number of parameters and high computational complexity. Therefore, this invention aims to reduce the number of parameters and computational complexity of MPS layers as much as possible without compromising model segmentation performance. Thus, a sparse MPS layer, possessing sparsity and weight sharing in 2D networks, is chosen. These characteristics enable it to reduce computational complexity while modeling global dependencies. A comparison is provided between a sparse MPS layer for 3D images and a standard MPS layer for 3D images. Figure 6 As shown.

[0040] Preferably, the step of decoding the initial feature f00 and high-level features f1 to f4 through a feature decoding branch to obtain the fused feature F1 of the same size as the input three-dimensional medical image includes:

[0041] The initial feature f0 and the high-level features f1 to 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 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 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 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 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 double its resolution to obtain the fused feature F1.

[0047] Please see Figure 5 Preferably, the Res block comprises two cascaded 3×3×3 three-dimensional convolutions; 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 summed 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 results from each depth are fed into a residual block consisting of two 3×3×3 convolutional layers, and then into instance normalization to ensure stable feature extraction. These feature maps are then deconvolutionally processed, doubling their resolution and concatenating the input from the previous depth with themselves. The concatenated features are then 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, calculated as follows:

[0049]

[0050] Where I is the number of voxels; J is the number of classes; Y i,j and G I,j Let be 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 kernel depth 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 a U-shaped large kernel depth convolutional network performs the three-dimensional medical image segmentation method based on a U-shaped large kernel depth convolutional network.

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

[0053] In summary, by introducing large-kernel deep convolution, this invention can capture a wider range of contextual information, effectively addressing the challenges posed by the diversity of organ shapes, locations, and anatomical structures. The combination of feature extraction and feature decoding branches enables the network to gain a deeper understanding of complex features in medical images, thereby improving the accuracy and reliability of segmentation. The introduction of residual convolution enhances the robustness of the model, reducing issues such as missing organ boundaries and incorrect organ classification, thus improving the overall quality of the segmentation results. The automation nature of deep learning methods significantly accelerates the segmentation process, producing results in a shorter time compared to manual segmentation, improving the efficiency of medical diagnosis. The efficient network structure and algorithm optimization further shorten the processing time, making this solution more feasible in practical applications. By introducing techniques such as large-kernel deep convolution, feature extraction branches, feature decoding branches, and residual convolution, this solution significantly improves the accuracy, efficiency, objectivity, and consistency of 3D medical image segmentation, providing strong support for clinical decision-making and personalized treatment, and possessing significant clinical application value and practical significance.

[0054] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within 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 3D medical image is acquired and input into an image segmentation model to obtain the segmentation result. The image segmentation model includes a large-kernel depthwise convolution, a feature extraction branch, and a feature decoding branch. The feature extraction branch includes four cascaded feature extraction stages. The large-kernel depthwise convolution is used to reduce the 3D medical image to a specific resolution to obtain initial feature f0. The initial feature f0 is input into the four cascaded feature extraction stages for sequential feature extraction, obtaining high-level features f1~f4 with scales decreasing from large to small. The initial feature f0 and high-level features f1~f4 are decoded by the feature decoding branch to the same size as the input 3D medical image to obtain fusion feature F1. The 3D medical image is subjected to residual convolution through a Res block to obtain fusion feature F2. The fusion features F1 and F2 are concatenated to obtain feature F3, which is then input with a softmax activation function. The convolutional layer predicts the final segmentation result; the large kernel depth convolution uses a 7×7×7 three-dimensional convolution with padding of 3, which reduces the resolution of the input three-dimensional medical image to half of its original value; in each feature extraction stage of the feature extraction branch, a cascaded UL module and a downsampling layer are used to extract the input features, and the downsampling layer reduces the resolution of the input features to 1 / 2 of its original value; wherein, the UL module includes: two cascaded feature extraction modules; each feature extraction module includes a cascaded first normalization layer, a large kernel depth convolution module, a first Gaussian error linear unit activation function, a second normalization layer, a three-dimensional image sparse multilayer perceptron layer, and a second Gaussian error linear unit activation function; 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 second normalization layer... The input features of the sub-layer and the output features of the second Gaussian error linear unit activation function are summed to obtain the output features of the feature extraction module. The large kernel depth convolution module includes: 9×9×9 three-dimensional convolutions, 7×7×7 three-dimensional convolutions, 5×5×5 three-dimensional convolutions, 3×3×3 three-dimensional convolutions and linear layers. The 9×9×9, 7×7×7, 5×5×5 and 3×3×3 three-dimensional convolutions are 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 its original value. The output features of the 9×9×9, 7×7×7, 5×5×5 and 3×3×3 three-dimensional convolutions are concatenated to obtain feature Y0. The linear layer is used to linearly map feature Y0 to convert the multi-channel input feature Y0 into a single channel to obtain the output feature.

2. The three-dimensional medical image segmentation method based on a U-shaped large-kernel depthwise convolutional network according to claim 1, characterized in that, In the first feature extraction stage of the feature extraction branch, the initial input feature f0 is extracted to obtain the high-level feature F1. In the second feature extraction stage of the feature extraction branch, the high-level input feature F1 is extracted 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.

3. The three-dimensional medical image segmentation method based on a U-shaped large-kernel depthwise convolutional network according to claim 1, characterized in that, The 3D image sparse multilayer perceptron layer includes: in, This represents the output features of a sparse multilayer perceptron layer in a 3D image. , and These represent the output results of a sparse multilayer perceptron layer in a 3D image, where information is mixed across three dimensions. This 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 D and number of channels C, while treating dimension W as a weight and not processing it; This means using the linear layer function provided by the nn package in torch to calculate the linear relationship between the input features in dimension W, dimension D and number of channels C, while using dimension H as a weight without performing any calculation; This 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, while treating dimension D as a weight and not processing it; This method involves concatenating the output of a sparse multilayer perceptron layer of a 3D image with the input features, and then reducing the 4 channels to 1 channel using a linear layer function. This represents the input features of a sparse multilayer perceptron layer in a 3D image.

4. The three-dimensional medical image segmentation method based on a U-shaped large-kernel depthwise convolutional network according to claim 1, characterized in that, The step of decoding the initial feature f0 and high-level features f1~f4 into a fused feature F1 of the same size as the input 3D medical image through the feature decoding branch includes: The initial feature f0 and the 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 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 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 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 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 double its resolution to obtain the fused feature F1.

5. The three-dimensional medical image segmentation method based on a U-shaped large-kernel depthwise convolutional network according to claim 1, characterized in that, The Res block comprises two cascaded 3×3×3 three-dimensional convolutions; 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 together 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.

6. A three-dimensional medical image segmentation device based on a U-shaped large-kernel depthwise convolutional network, characterized in that, It includes a processor and a memory; the memory is used to store computer programs; the processor is connected to the memory and is used to execute the computer programs stored in the memory, so that the three-dimensional medical image segmentation device based on a U-shaped large kernel deep convolutional network performs the 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 5.

7. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the three-dimensional medical image segmentation method based on a U-shaped large kernel deep convolutional network as described in any one of claims 1-5.

Images