A Transformer 3D medical image segmentation method based on feature focusing

Abstract

This paper proposes a Transformer 3D medical image segmentation method based on feature focusing. For the input image to be detected, V-Net is first used to perform a first-stage coarse segmentation, and then an improved Transformer two-stage encoder based on focusing and sparse matrices is used for feature encoding. This method aims to use less computation to perform global information interaction within the 3D data space. For problems with large 3D data volumes, focused mapping is used to highlight key areas, while sparse matrices are used to reduce computational complexity. Dynamic representation position encoding is used to supplement lost position information, and a spatial feature fusion connection module is used to extract and link spatial information between the encoder and decoder. This method can retain positional and spatial information with less computational complexity, improving the ability to model spatial context.

Description

Technical Field

[0001] The present invention relates to the fields of image segmentation and computer vision technology, and in particular to a Transformer 3D medical image segmentation method based on feature focusing. Background Art

[0002] In the field of medical image segmentation, 3D medical image segmentation is an important and challenging topic. Due to the stereoscopic nature of 3D medical images, their data volume is usually N times that of 2D images, and they have complex spatial features, which enhances the sensitivity of position information. However, as a result, the variability of the position, size, and shape of the segmentation target is greater, which increases the complexity of the segmentation task. When using convolutional neural networks (CNNs) to process such large-scale 3D data, due to the huge amount of data and the complexity of the feature space, the network often cannot fully capture the contextual connections of the image, resulting in poor segmentation results. In order to meet this challenge, some current methods consider using Transformer networks that can perform global modeling. Transformer networks can effectively capture the relationship between distant pixels through the self-attention mechanism, thereby better understanding the global structure of the image. However, since the computational complexity of the self-attention mechanism is O(N 2 ), and the data volume of each sample of 3D medical images is huge, which may lead to huge overhead of computing resources. Summary of the Invention

[0003] In light of this, the present invention aims to provide a Transformer-based 3D medical image segmentation method based on feature focusing. This method reduces computational overhead through focused mapping and matrix sparsification. Secondly, a spatial feature fusion connection module is proposed to alleviate the problem of spatial position information ambiguity by fusing features from the encoder and decoder. Finally, a dynamic representation position encoding module is proposed to extract and fuse global information at different scales to enhance contextual understanding.

[0004] To achieve the above object, the present invention adopts the following technical solution: a Transformer 3D medical image segmentation method based on feature focusing, comprising the following steps:

[0005] Step S1: By training the 3D medical image dataset and applying data enhancement methods to the data, the data is fed into the V-Net backbone network for a first-stage preliminary training and feature extraction;

[0006] Step S2: The trained feature extraction network is used as the first-stage feature extraction module. The Transformer based on focus mapping and sparse matrix is ​​proposed as the basic module of the second-stage extraction network. Focus mapping is used to stimulate the change of intrinsic distribution. At the same time, sparse matrix is ​​used to reduce the amount of computation and storage space requirements, so that the model can capture key information when processing large-scale data.

[0007] Step S3: A spatial feature fusion link module SFFC is proposed to fuse the feature information in the encoder and decoder by retaining and extracting the spatial position information;

[0008] Step S4: In view of the limitations of static position coding and conventional dynamic coding, a dynamic representation position coding module DPRE is proposed. It extracts local features at different scales, integrates global information to increase context understanding, implements dynamic position coding based on content, and finally realizes the segmentation of 3D medical images.

[0009] In a preferred embodiment, step S1 specifically includes the following steps:

[0010] Step S11: Obtain a public medical image segmentation training set from the Internet and obtain relevant annotations of the training data; mainly obtain three commonly used 3D medical image datasets ADCD, Synapse, and LA, which respectively include data of multiple parts of the heart and the whole body;

[0011] Step S12: After the data is loaded, different random data augmentations are performed on the data, including random color jittering, grayscale conversion, Gaussian blurring, and random shearing, to simulate the complex changes in real medical scenarios and force the model to learn more robust feature representations.

[0012] Step S13: Initialize the network weights and parameters of the image to be segmented in the medical image segmentation dataset using a pre-trained V-Net feature extraction network.

[0013] In a preferred embodiment, step S2 specifically includes the following steps:

[0014] Step S21: First, perform linear transformation on the query matrix Q and the key matrix K, and use the similarity calculation method to calculate the similarity between them, thereby obtaining the self-attention matrix A. In the self-attention matrix A, each element represents the similarity score between different positions in the sequence; then, perform sparse processing on the self-attention matrix A to obtain a sparse self-attention matrix A'; most elements in the sparse self-attention matrix are zero, and only some important similarity scores are retained. The specific calculation process is as follows:

[0015] Q=xW Q , K=xW K , V=xWV

[0016]

[0017] Among them, x represents the input feature, K T represents the transpose of the key matrix, d represents the dimension of the value matrix, Sim uses the common Sotfmax function, and after sparsifying the self-attention matrix A, we get A′ ij , where k is a hyperparameter of the sparsification matrix, which is also a hyperparameter of the focus map; A′ ij Represents the sparse self-attention matrix; A ij represents the self-attention matrix;

[0018] Step S22: In the V processing of the value matrix, a focus mapping function is proposed. The focus mapping function is carefully designed and applied to the value matrix. The specific calculation formula is as follows:

[0019]

[0020] Where V′ represents the value matrix after stimulating intrinsic characteristics, and the RELU function ensures the non-negativity of the input and the validity of the denominator;

[0021] Step S23: A local rank restoration module is proposed to restore the order of the attention matrix by applying additional depthwise separable convolution (DWC) to supplement the lost details in the model. The final complete and specific calculation process is as follows:

[0022]

[0023] DWC is separable convolution, where DWC(V) represents the operation of separable convolution on the V matrix.

[0024] In a preferred embodiment, step S3 specifically includes the following steps:

[0025] Step S31: improving the skip connection module;

[0026] First, obtain the low-level feature A from the encoder and adopt the maximum pooling and average pooling strategies. Through these two pooling operations, two feature weight matrices with different characteristics are obtained, representing the salient features and global features of A respectively. The two pooling matrices are added together to fuse the salient features and global features of feature A, which can better capture the correlation and importance between features. The weighted features are processed by the sigmoid activation function to limit the weighted features between 0 and 1. Then, the weighted weight matrix is ​​multiplied by the original feature matrix A to obtain a new feature matrix A′. The purpose of this step is to make A′ pay more attention to the important spatial position information in the encoder, as well as the salient and global features through the weighted multiplication operation. The specific calculation is shown in the formula:

[0027] A′=(MLP(Max(A))+MLP(Mean(A)))×A

[0028] Among them, MLP is a multi-layer perceptron, Max is a 3D maximum pooling layer, and Mean is a 3D average pooling layer;

[0029] Step S32: Upsample the high-level features B from the decoder to make them of the same size as the feature matrix A′. After upsampling, a separable convolution operation is performed to further extract the feature representation of B while maintaining the consistency of the spatial position information. After the separable convolution operation, a decoder feature B′ with twice the spatial position information is obtained. The specific calculation process is as follows:

[0030] B′=DWC(Upsample(B))

[0031] Step S33: Finally, we concatenate the encoder feature matrix A′ and the decoder feature matrix B′ to obtain a comprehensive feature representation that contains important features from the encoder and decoder as well as spatial position information. The complete calculation process is shown in the formula:

[0032]

[0033] MLP is a multi-layer perceptron, a basic structure of a neural network; Upsample is to upsample the feature;

[0034] Step S34: Through a series of detailed operations, including pooling, weighted multiplication, upsampling, separable convolution and feature splicing, the feature information in the encoder and decoder are effectively integrated and the spatial position information is retained.

[0035] In a preferred embodiment, step S4 specifically includes the following steps:

[0036] Step S41: The dynamic representation position encoding module DPRE extracts local features from different scales through separable convolution, and then performs global attention operation to obtain global information to increase context understanding, thereby achieving dynamic position encoding based on content;

[0037] Step S42: First, the input features are put into two 3D separable convolutional layers with different kernel sizes for feature extraction; by extracting local features at different scales, two features F3 and F5 with different sizes are obtained. The specific calculation process of F3 and F5 is as follows:

[0038] F i =DWC i×i (F input )

[0039] Where DWC stands for depthwise separable convolution, F i It is represented as the feature after depth-wise separable convolution, where the values ​​of i are 3 and 5, and F input Represented as input features;

[0040] Step S43: Next, the self-attention mechanism is used to establish links between patch features to capture the dependencies between positions at different scales in the input sequence;

[0041] Step S44: Use a fully connected layer to further adjust the dynamic position encoding; by learning the weight and bias parameters of the position encoding, the position information is effectively integrated with the image features, and finally the feature F with dynamic position encoding is obtained. out , the specific calculation process is shown in the formula:

[0042] F pos =LN(F5)×Softmax(Reshape(LN(F3))×LN(F5))

[0043] F out =F input +Reshape(F pos )

[0044] F pos is the feature after learning the position feature; LN is the linear layer, F5 is the feature after the separable convolution of size 5 mentioned above, F3 is the feature after the separable convolution of size 3 mentioned above, Softmax is the normalization function, and Reshape is the shape transformation to facilitate the multiplication of the previous and next features in the correct rank;

[0045] Step S45: Test the trained medical image segmentation model on the validation set to obtain the final detection accuracy.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. Different from common 3D medical image segmentation, considering the problem of excessive amount of 3D medical image data, sparse matrix and focusing function are introduced on the basis of Transformer to make it more suitable for actual application scenarios.

[0048] 2. A spatial feature fusion connection module is proposed to alleviate the problem of spatial position information ambiguity by fusing the features of the encoder and decoder.

[0049] 3. A dynamic representation position encoding module is proposed to extract and fuse global information at different scales to increase contextual understanding. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flowchart of a method implementation according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0052] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0053] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form, and it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0054] like Figure 1 As shown, the present invention provides a Transformer3D medical image segmentation method based on feature focusing, comprising the following steps:

[0055] Step S1: training with 3D medical image datasets, applying data augmentation methods to the data, and feeding the data into the V-Net backbone network for a first-stage preliminary training and feature extraction;

[0056] Step S2: The trained feature extraction network is used as the first-stage feature extraction module. The Transformer based on focused mapping and sparse matrix (F-Transformer) is proposed as the basic module of the second-stage extraction network. Through focused mapping, the intrinsic distribution change is stimulated. At the same time, the use of sparse matrix reduces the amount of computation and storage space requirements, making the model more efficient in capturing key information when processing large-scale data.

[0057] Step S3: To address the limitations of traditional skip connection methods in terms of information utilization and spatial position information retention, a spatial feature fusion connection module (SFFC) is proposed. By retaining and extracting spatial position information, it can effectively fuse the feature information in the encoder and decoder;

[0058] Step S4: To address the limitations of static position coding and conventional dynamic coding, a dynamic representation position coding module (DPRE) is proposed. By extracting local features at different scales and fusing global information to increase contextual understanding, dynamic position coding can be better implemented based on content, and finally the segmentation of 3D medical images can be achieved.

[0059] Step S1 specifically includes the following steps:

[0060] Step S11: Obtain a public medical image segmentation training set from the Internet and obtain relevant annotations of the training data; mainly obtain three commonly used 3D medical image datasets (ADCD, Synapse, and LA), which respectively include data on the heart and multiple parts of the torso;

[0061] Step S12: After the data is loaded, various random data augmentations are performed on the data. These include random color jittering, grayscale conversion, Gaussian blurring, and random cropping to simulate the complex changes in real medical scenarios and force the model to learn more robust feature representations.

[0062] Step S13: Initialize the network weights and parameters of the image to be segmented in the medical image segmentation dataset using a pre-trained V-Net feature extraction network.

[0063] Step S2 specifically includes the following steps:

[0064] Step S21: First, we perform a linear transformation on the query matrix Q and the key matrix K, and use the similarity calculation method to calculate the similarity between them, thereby obtaining the self-attention matrix A. In the self-attention matrix A, each element represents the similarity score between different positions in the sequence. Then, in order to reduce the computational complexity, we perform a sparse processing on the self-attention matrix A to obtain a sparse self-attention matrix A'. Most of the elements in the sparse self-attention matrix are zero, and only some important similarity scores are retained, which significantly reduces the amount of computation and storage space requirements. This sparse processing method can effectively improve the computational efficiency of the self-attention mechanism. The specific calculation process is as follows:

[0065] Q=xW Q , K=xW K , V=xW V

[0066]

[0067] Among them, x represents the input feature, K T represents the transpose of the key matrix, d represents the dimension of the value matrix, Sim uses the common Sotfmax function, and after sparsifying the self-attention matrix A, we get A′ ij , where k is a hyperparameter of the sparsification matrix, which is also a hyperparameter of the focus map. ij Represents the sparse self-attention matrix; A ij Represents the self-attention matrix.

[0068] Step S22: In the V processing of the value matrix, an innovative focus mapping function is proposed. This function is carefully designed and applied to the value matrix. Its purpose is to stimulate the distribution changes within the value matrix through a specific mapping method, making the features corresponding to different positions more prominent and significant. The design of this mapping function takes into account the intrinsic characteristics of the value matrix and the application scenarios in the self-attention mechanism, in order to achieve more accurate and effective feature extraction and representation. At the same time, it enhances the corresponding features and alleviates the feature smoothing caused by the sparse matrix. The specific calculation formula is as follows:

[0069]

[0070] Where V′ represents the value matrix after stimulating intrinsic characteristics, the RELU function ensures the non-negativity of the input and the validity of the denominator, and k is a hyperparameter of the sparsification matrix, which is also a hyperparameter of the focus map.

[0071] Step S23: A local rank restoration module is proposed to restore the order of the attention matrix by applying additional depthwise separable convolution (DWC) to supplement the lost details in the model. The final complete and specific calculation process is as follows:

[0072]

[0073] Where Sparse() is the sparsification function, O is the final self-attention output; DWC is the separable convolution, where DWC(V) represents the operation of the separable convolution on the V matrix.

[0074] The proposed Transformer based on focus mapping and sparse matrix reduces the subsequent computational complexity through the sparse self-attention matrix A, and at the same time performs focus mapping on the value matrix V to stimulate intrinsic distribution changes and enhance feature contrast. Finally, depthwise separable convolution is used to maintain the matrix rank lost due to the sparse matrix, so that the model can capture key information more efficiently when processing large-scale data and maintain the diversity of output features.

[0075] Step S3 specifically includes the following steps:

[0076] Step S31: In order to enhance the features generated by the model, better utilize the information in the feature map, and retain and utilize the spatial location information to further improve the performance and generalization ability of the model, we improve the skip connection module;

[0077] First, we obtain the low-level feature A from the encoder. In order to retain its spatial position information, the maximum pooling and average pooling strategies are adopted. Through these two pooling operations, we obtain two feature weight matrices with different characteristics, representing the salient features and global features of A respectively. By adding these two pooling matrices, the salient features and global features of feature A are integrated, which can better capture the correlation and importance between features. The weighted features are processed by the sigmoid activation function to limit the weighted features between 0 and 1. Next, we multiply the weighted weight matrix with the original feature matrix A to obtain a new feature matrix A′. The purpose of this step is to make A′ pay more attention to the important spatial position information in the encoder, as well as the salient and global features through the weighted multiplication operation. The specific calculation is shown in the formula:

[0078] A′=(MLP(Max(A))+MLP(Mean(A)))×A

[0079] Among them, MLP is a multi-layer perceptron, Max is a 3D maximum pooling layer, and Mean is a 3D average pooling layer.

[0080] Step S32: Next, we upsample the high-level features B from the decoder to make them of the same size as the feature matrix A'. After upsampling, we use a separable convolution operation to further extract the feature representation of B while maintaining the consistency of the spatial position information. Separable convolution is an effective convolution operation that can reduce the number of parameters while maintaining the effectiveness of the model. After the separable convolution process, we obtain a decoder feature B' that is magnified twice and has spatial position information. The specific calculation process is as follows:

[0081] B′=DWC(Upsample(B))

[0082] Step S33: Finally, we concatenate the encoder feature matrix A′ and the decoder feature matrix B′ to obtain a comprehensive feature representation that contains important features from both the encoder and decoder as well as spatial location information. This operation process aims to fully utilize the low-level and high-level features in the encoder and decoder, and retain and extract spatial location information through weighted fusion and convolution operations, thereby providing a more accurate and effective feature representation for medical image segmentation tasks. The complete calculation process is shown in the formula:

[0083]

[0084] MLP is a multi-layer perceptron, a basic structure of a neural network; Upsample is to upsample the feature;

[0085] Step S34: Through a series of detailed operations, including pooling, weighted multiplication, upsampling, separable convolution and feature splicing, the feature information in the encoder and decoder are effectively fused and the spatial position information is preserved.

[0086] Step S4 specifically includes the following steps:

[0087] Step S41: There are many methods for dynamic position coding, among which the current mainstream methods include dynamic position coding based on convolution and dynamic position coding based on global information. The convolution-based method utilizes the local receptive field characteristics of convolutional neural networks and can better capture local features of the image. However, due to its strong locality, it is always unable to effectively consider situations involving global information, resulting in the inability to fully utilize the overall semantic information of the image, which limits the performance of the model. Another method based on global information adopts a strategy of global analysis of the entire image, and generates position coding by comprehensively considering the information of all pixels in the image. This method can better capture the global features and semantic information of the image, increase contextual understanding, and thus improve the model's ability to understand the overall structure of the image. A novel dynamic representation position coding module (DPRE) is designed, which extracts local features from different scales through separable convolution, and then performs attention operations on the global to obtain global information to increase contextual understanding, so that dynamic position coding can be better implemented according to content;

[0088] Step S42: First, the input features are put into two 3D separable convolution layers with different kernel sizes for feature extraction. The advantage of this design is that it can quickly extract local features within each patch image. By extracting local features at different scales, two features F3 and F5 with different sizes are obtained. The model can better capture the details in the image, thereby enhancing the understanding of local structures. At the same time, the use of separable convolution makes the model more computationally efficient, which can reduce the number of parameters and reduce computational complexity. The specific calculation process of F3 and F5 is as follows:

[0089] F i =DWC i×i (F input )

[0090] Where DWC stands for depthwise separable convolution, F i It is represented as the feature after depth-wise separable convolution, where the values ​​of i are 3 and 5, and F input is represented as input features.

[0091] Step S43: Next, a self-attention mechanism is used to link patch features to capture dependencies between positions at different scales in the input sequence. By focusing globally on the entire sequence, the model can better understand the overall structure and semantic information of the input data, thereby improving its ability to capture global features and enhancing its understanding of the overall semantics of the image.

[0092] Step S44: At the same time, it only needs to perform self-attention on the features between patches, which can greatly reduce the computational complexity of training and inference, and is particularly effective for large-scale images. Finally, a fully connected layer is used to further adjust the dynamic position encoding. By learning the weights and bias parameters of the position encoding, the position information is effectively integrated with the image features, and finally the feature F with dynamic position encoding is obtained. out , the specific calculation process is shown in the formula:

[0093] F pos =LN(F5)×Softmax(Reshape(LN(F3))×LN(F5))

[0094] F out =F input +Reshape(F pos )

[0095] F pos is the feature after learning the position feature; LN is the linear layer, F5 is the feature after the separable convolution of size 5 mentioned above, F3 is the feature after the separable convolution of size 3 mentioned above, Softmax is the normalization function, and Reshape is the shape transformation to facilitate the multiplication of the previous and next features in the correct rank;

[0096] Step S45: Test the trained medical image segmentation model on the validation set to obtain the final detection accuracy.

[0097] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0098] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0099] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0100] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

[0102] This patent is not limited to the above-mentioned optimal implementation mode. Anyone can derive various other forms of Transformer3D medical image segmentation methods based on feature focusing under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of this invention should be covered by this patent.

Claims

1. A Transformer 3D medical image segmentation method based on feature focusing, characterized in that: The following steps are involved: Step S1: By training the 3D medical image dataset and applying data enhancement methods to the data, the data is fed into the V-Net backbone network for a first-stage preliminary training and feature extraction; Step S2: The trained feature extraction network is used as the first-stage feature extraction module. The Transformer based on focus mapping and sparse matrix is ​​proposed as the basic module of the second-stage extraction network. Focus mapping is used to stimulate the change of intrinsic distribution. At the same time, sparse matrix is ​​used to reduce the amount of computation and storage space requirements, so that the model can capture key information when processing large-scale data. Step S3: A spatial feature fusion link module SFFC is proposed to fuse the feature information in the encoder and decoder by retaining and extracting the spatial position information; Step S4: To address the limitations of static position coding and conventional dynamic coding, a dynamic representation position coding module (DPRE) is proposed. It extracts local features at different scales, integrates global information to enhance contextual understanding, implements dynamic position coding based on content, and finally achieves segmentation of 3D medical images. Step S3 specifically includes the following steps: Step S31: improving the skip connection module; First, obtain the low-level feature A from the encoder and adopt the maximum pooling and average pooling strategies. Through these two pooling operations, two feature weight matrices with different characteristics are obtained, representing the salient features and global features of A respectively. The two pooling matrices are added together to fuse the salient features and global features of feature A, which can better capture the correlation and importance between features. The weighted features are processed by the sigmoid activation function to limit the weighted features between 0 and 1. Then, the weighted weight matrix is ​​multiplied by the original feature matrix A to obtain a new feature matrix A'. The purpose of this step is to make A' pay more attention to the important spatial position information in the encoder, as well as the salient and global features through the weighted multiplication operation. The specific calculation is shown in the formula: A′=(MLP(Max(A))+MLP(Mean(A)))×A Among them, MLP is a multi-layer perceptron, Max is a 3D maximum pooling layer, and Mean is a 3D average pooling layer; Step S32: Upsample the high-level features B from the decoder to make them of the same size as the feature matrix A'. After upsampling, a separable convolution operation is performed to further extract the feature representation of B while maintaining the consistency of the spatial position information. After the separable convolution operation, a decoder feature B' with twice the spatial position information is obtained. The specific calculation process is as follows: B′=DWC(Upsample(B)) Step S33: Finally, we concatenate the encoder feature matrix A' and the decoder feature matrix B' to obtain a comprehensive feature representation that contains important features from the encoder and decoder as well as spatial position information. The complete calculation process is shown in the formula: MLP is a multi-layer perceptron, a basic structure of a neural network; Upsample is to upsample the feature; Step S34: Through a series of detailed operations, including pooling, weighted multiplication, upsampling, separable convolution and feature splicing, the feature information in the encoder and decoder are effectively integrated and the spatial location information is preserved; Step S4 specifically includes the following steps: Step S41: The dynamic representation position encoding module DPRE extracts local features from different scales through separable convolution, and then performs global attention operation to obtain global information to increase context understanding, thereby achieving dynamic position encoding based on content; Step S42: First, the input features are put into two 3D separable convolutional layers with different kernel sizes for feature extraction; by extracting local features at different scales, two features F3 and F5 with different sizes are obtained. The specific calculation process of F3 and F5 is as follows: F i =DWC i×i (F input ) Where DWC stands for depthwise separable convolution, F i It is represented as the feature after depth-wise separable convolution, where the values ​​of i are 3 and 5, and F input Represented as input features; Step S43: Next, the self-attention mechanism is used to establish links between patch features to capture the dependencies between positions at different scales in the input sequence; Step S44: Use a fully connected layer to further adjust the dynamic position encoding; by learning the weight and bias parameters of the position encoding, the position information is effectively integrated with the image features, and finally the feature F with dynamic position encoding is obtained. out , the specific calculation process is shown in the formula: F pos =LN(F5)×Softmax(Reshape(LN(F3))×LN(F5)) F out =F input +Reshape(F pos ) F pos is the feature after learning the position feature; LN is the linear layer, F5 is the feature after the separable convolution of size 5, F3 is the feature after the separable convolution of size 3, Softmax is the normalization function, and Reshape is the shape transformation to facilitate the multiplication of the previous and next features in the correct rank; Step S45: Test the trained medical image segmentation model on the validation set to obtain the final detection accuracy.

2. The feature-focused Transformer 3D medical image segmentation method according to claim 1, wherein: Step S1 specifically includes the following steps: Step S11: Obtain a public medical image segmentation training set from the Internet and obtain relevant annotations of the training data; mainly obtain three commonly used 3D medical image datasets ADCD, Synapse, and LA, which respectively include data of multiple parts of the heart and the whole body; Step S12: After the data is loaded, different random data augmentations are performed on the data, including random color jittering, grayscale conversion, Gaussian blurring, and random shearing, to simulate the complex changes in real medical scenarios and force the model to learn more robust feature representations. Step S13: Initialize the network weights and parameters of the image to be segmented in the medical image segmentation dataset using a pre-trained V-Net feature extraction network.

3. The Transformer 3D medical image segmentation method based on feature focusing according to claim 1, characterized in that: Step S2 specifically includes the following steps: Step S21: First, perform linear transformation on the query matrix Q and the key matrix K, and use the similarity calculation method to calculate the similarity between them, thereby obtaining the self-attention matrix A. In the self-attention matrix A, each element represents the similarity score between different positions in the sequence; then, perform sparse processing on the self-attention matrix A to obtain a sparse self-attention matrix A'; most elements in the sparse self-attention matrix are zero, and only some important similarity scores are retained. The specific calculation process is as follows: Q=xW Q ,K=xW K ,V=xW V Among them, x represents the input feature, K T represents the transpose of the key matrix, d represents the dimension of the value matrix, Sim uses the common Softmax function, and after sparsifying the self-attention matrix A, we get A′ ij , where k is a hyperparameter of the sparsification matrix, which is also a hyperparameter of the focus map; A′ ij Represents the sparse self-attention matrix; A ij represents the self-attention matrix; Step S22: In the V processing of the value matrix, a focus mapping function is proposed. The focus mapping function is carefully designed and applied to the value matrix. The specific calculation formula is as follows: Where V′ represents the value matrix after stimulating intrinsic characteristics, and the RELU function ensures the non-negativity of the input and the validity of the denominator; Step S23: A local rank restoration module is proposed to restore the order of the attention matrix by applying additional depthwise separable convolution (DWC) to supplement the lost details in the model. The final complete and specific calculation process is as follows: Where Sparse() is the sparsification function, O is the final self-attention output; DWC is the separable convolution, where DWC(V) represents the operation of the separable convolution on the V matrix.

Images