Medical image segmentation model based on local enhancement driving global optimization

By introducing a channel feature enhancement module and an enhanced local information extractor in the medical image segmentation model, combined with the encoder-decoder structure, the problem of insufficient global information and local detail information capture capabilities in the prior art is solved, and more accurate lesion tissue boundary identification and higher segmentation performance.

CN118887221APending Publication Date: 2024-11-01XINJIANG UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410878510.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-02
Publication Date
2024-11-01

AI Technical Summary

Technical Problem

The existing medical image segmentation algorithm has insufficient recognition capabilities when capturing global and local detailed information in the image. Especially when the lesion area is small or hybridizes with surrounding structures, it is difficult to measure the overall structure of the lesion in the feature map at the same time. Local details.

Method used

Using a medical image segmentation model based on local enhancement drive global optimization, the global context information and locality in the feature map are extracted through the channel feature enhancement module and the enhanced local information extractor, combined with the U-shaped structure and jump connection of the encoder-decoder. Detailed information, and the relationship between the feature map channel is calculated through the Spilt Attention module and the enhanced local information extractor to achieve feature fusion.

Benefits of technology

It significantly improves the ability to identify microscopic features and their boundaries of the lesion tissue, enhances the ability to identify subtle changes in the lesion boundaries, and improves the performance of medical imaging segmentation tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118887221A_ABST
    Figure CN118887221A_ABST
Patent Text Reader

Abstract

The invention discloses a medical image segmentation model based on local enhancement driving and global optimization, and belongs to the technical field of computer vision, a U-shaped structure of an encoder-decoder is adopted, and jump connection is adopted between the encoder and the decoder to relieve the problem of down-sampling information loss; the encoder comprises a channel feature enhancement module, and global context confidence and local detail information in a feature map are extracted by using an enhanced local information extractor at a jump joint; and the decoder recovers the feature map output by the encoder to the original size by using a bilinear interpolation algorithm. According to the medical image segmentation model, refined foreground information in a feature map is extracted through the channel feature enhancement module, details and the overall contour of a lesion area are concerned at the same time through the enhanced local information extractor, and the microscopic features and boundaries of lesion tissue are captured and described more accurately; and the method has great value in promoting the development of a medical image segmentation technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a medical image segmentation model based on local enhancement driven global optimization. Background Art

[0002] With the rapid development of computer vision related technologies, medical image processing provides efficient and reliable technical support for medical staff. Medical image segmentation, as an important branch of medical image processing, is very important for doctors to accurately monitor diseases, visualize injuries, and choose the right treatment plan. Medical image segmentation is a key step in medical image processing and analysis. Its goal is to identify the pixels of diseased organs from medical images and extract information features of the lesion site. Early image segmentation was mainly based on traditional methods, which all relied on manually defined rules and feature extraction, had certain limitations, and had poor adaptability to complex data sets and tasks. With the rapid development of deep learning, convolutional neural networks have been widely used in the field of medical image segmentation due to their powerful autonomous learning ability. Since various organs and tissues are adjacent to each other, especially when the lesion area is small or intertwined with surrounding structures, existing segmentation algorithms may encounter bottlenecks in recognition and performance.

[0003] In the field of medical image segmentation, CNN and Transformer have poor global modeling capabilities and local information extraction capabilities, respectively. Since local details and global contextual information are indispensable for the recognition of foreground information, researchers insert CNN and Transformer into different positions in the network to capture global information and local features in the image. Although the hybrid network as a whole can extract local detail information and global information in the feature map, local modeling and global modeling are performed independently, and it is difficult for the network to simultaneously measure the overall structure and local detail information of the lesion in the feature map, which leads to the segmentation model's inability to recognize the subtle boundaries of the lesion. Summary of the invention

[0004] In order to address the deficiencies in the above-mentioned prior art, the purpose of the present invention is to provide a medical image segmentation model based on local enhancement driven global optimization, which extracts refined foreground information in the feature map through a channel feature enhancement module, and simultaneously focuses on the details and overall contours of the lesion area through an enhanced local information extractor, thereby achieving more accurate capture and depiction of the microscopic features and boundaries of the lesion tissue, and has great value in promoting the development of medical image segmentation technology.

[0005] The technical solution adopted by the present invention to solve the technical problem is:

[0006] A medical image segmentation model based on local enhancement driven global optimization is provided, which adopts a U-shaped structure of encoder-decoder, and a jump connection is used between the encoder and the decoder to alleviate the problem of downsampling information loss;

[0007] The encoder includes a channel feature enhancement module, and an enhanced local information extractor is used at the jump connection to extract global context confidence and local detail information in the feature map;

[0008] The decoder uses a bilinear interpolation algorithm to restore the feature map output by the encoder to its original size.

[0009] Furthermore, the workflow of the channel feature enhancement module is as follows:

[0010] S11: Assume that the size of the input feature map is The input feature maps are grouped according to the channel dimension to obtain k feature maps {F1, F2…F k-1 ,F k};

[0011] S12: From the channel dimension, the feature maps {F1, F2…F k-1 ,F k}Continue to group to get r feature maps, among which, feature map F1 can be obtained after grouping {F 11 ,F 12 …F 1r}, and so on for other feature maps;

[0012] S13: Convolution kernels with sizes of 1×1 and 3×3 are used to align feature maps {F i1 ,F i2 …F ir} to extract features and obtain feature maps {F i1 ′,F i2 ′…F ir '};

[0013] S14: The feature map {F i1 ′,F i2 ′…F ir ′} is input into the Spilt Attention module to calculate the relationship between the feature map channels and output the feature map T i ;

[0014] S15: Feature maps {T1, T2…T k}Concatenate and output feature map T;

[0015] S16: After the convolution operation, feature fusion is performed with the original feature map F.

[0016] Furthermore, the formula of S11 is:

[0017] {F1,F2…F k-1 ,F k}=Splite(F){F1,F2…F k-1 ,F k}∈R h×w×c / k ;

[0018] The formula for S12 is:

[0019] {F 11 ,F 12 …F 1r}=Split(F i ){F 11 ,F 12 …F 1r}∈R h×w×c / k / r ; 1≤i≤k;

[0020] The formula for S13 is:

[0021] {F i1 ′,F i2 ′…F ir ′}=Conv3(Conv1({F i1 ,F i2 …F ir})){F i1 ′,F i2 ′…F ir ′∈R h×w×c′ / k};

[0022] The formula for S14 is:

[0023] T i = SpliAttention({F i1 ′,F i2 ′…F ir ′})T i ∈R h×w×c′ / k ;

[0024] The formula for S15 is:

[0025] T=concayenate{T1,T2…T k};

[0026] The formula for S16 is:

[0027] out=conv1(T)+F.

[0028] Furthermore, in S14, the workflow of the Spilt Attention module is as follows:

[0029] (1) Assume that the feature map {F i1 ′,F i2 ′…F ir The size of '} is R h×w×c , feature map {F i1 ′,F i2 ′…F ir ′} Feature fusion is performed by adding operation to obtain feature map

[0030] (2) Feature map F i ∈R h×w×c After the global maximum pooling operation, we get F i ∈R 1×1×c ;

[0031] (3) Feature map F i ∈R 1×1×c It undergoes convolution operation, normalization operation, and activation function mapping operation in sequence;

[0032] (4) Use r parallel convolution kernels to extract feature maps F i ∈R 1×1×c′ The semantic information of r feature maps {F i1 ,F i2 …F ir};

[0033] (5) Use the Softmax function to separate the feature maps {F i1 ,F i2 …F ir}The channel dimension is calculated to obtain the corresponding channel coefficient;

[0034] (6) Channel coefficient and feature map {F i1 ′,F i2 ′…F ir ′} and then perform addition operation to output the feature map.

[0035] Furthermore, the formula of step (1) is:

[0036] F i =F i1 ′+F i2 ′+…+F ir ′F i ∈R h×w×c ;

[0037] The formula for step (2) is:

[0038] F i =Globalpooling(F i )Fi ∈R 1×1×c ;

[0039] The formula for step (3) is:

[0040] F i =Relu(BN(Conv1(F i )))F i ∈R 1×1×c′ ;

[0041] The formula for step (4) is:

[0042] {F i1 ,F i2 …F ir}=Conv1(F i ){F i1 ,F i2 …F ir}∈R 1×1×c ;

[0043] The formula for step (5) is:

[0044]

[0045] The formula for step (6) is:

[0046]

[0047] Furthermore, the workflow of the enhanced local information extractor is as follows:

[0048] S21: Define the size of the input feature map X as R h×w×c , set c = 1, h = w = 4, and use the Softmax activation function on the input feature map to obtain two attention coefficient matrices and

[0049] S22: Multiply the coefficient matrix by the input feature map and scale the output feature map to obtain two feature maps Q∈R 1×16×1 and K∈R 1×1×16 ;

[0050] S23: Perform matrix multiplication on the feature map Q and the feature map K to obtain an element matrix M1;

[0051] S24: Calculate the Euclidean distance between each element in the element matrix and the diagonal element to obtain a similarity matrix, and then use the Sigmoid derivative to convert the similarity matrix into a self-attention coefficient matrix;

[0052] S25: Self-attention coefficient matrix and scale-transformed feature map V∈R 1×16×1 Perform matrix multiplication to output the feature map Y.

[0053] Furthermore, in S21, the specific calculation formulas of α1 and β1 are as follows:

[0054]

[0055] The formula for S22 is:

[0056] Q = Reshape(α×X)

[0057] K = Reshape(β×X);

[0058] The formula for S23 is:

[0059]

[0060] The formula for S24 is:

[0061]

[0062] Among them, λ i,j represents the self-attention coefficient of a certain spatial position, and Q i K j -Q i K i It represents the specific calculation formula for converting a point element in the element matrix into a similarity value using Euclidean distance;

[0063] The formula for S25 is:

[0064] Y i =λ i,1 ×v1+λ i,2 ×v2+…+λ i,15 ×v 15 +λ i,16 ×v 16 .

[0065] Furthermore, the model was evaluated using the Synapse multi-organ segmentation dataset, the automatic cardiac diagnosis challenge dataset, and the myeloma plasma cell segmentation dataset.

[0066] Furthermore, the model uses a mixture of Dice Similarity Coefficient and Hausdorff Distance as evaluation indicators, where the DSC calculation formula is:

[0067]

[0068] The formula for calculating HD is:

[0069] HD(A,B)=max{(h(A,B),h(B,A))};

[0070] Among them, A represents the predicted segmentation result, B represents the actual segmentation result; h(A,B) represents the maximum value of the shortest distance from each point in A to B, that is:

[0071] h(A,B)=max[min(d(a,b))]

[0072] Among them, d(a,b) represents the Euclidean distance between a and b.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] In the medical image segmentation model based on local enhancement driven global optimization exemplified in the present invention, the channel feature enhancement module can measure the importance of each channel feature map, which effectively highlights the importance of the channel feature map with a large proportion of foreground information and weakens the negative impact of the channel feature map with a large proportion of background information; in addition, the enhanced local information extractor deeply analyzes the complex relationships between rows and columns in the feature map to generate a coefficient matrix describing the internal spatial correlation of the features; next, the model fuses the obtained coefficient matrix with the original feature map to produce a deep feature response rich in local contextual information; in addition, in order to enhance the model's ability to grasp local details, it processes these feature maps through dimensional conversion and matrix multiplication operations to form an element matrix that can deeply show the interaction between the original features. This process further explores the interaction between the features, so that the deep features can comprehensively characterize the complex local information of the original image data; then, the extracted deep features are analyzed for similarity and weighted by calculating the Euclidean distance method between the elements, in order to refine and optimize the feature representation to ensure that the model can accurately reflect the subtle changes at the lesion boundary. The network model developed by the present invention has shown significant performance improvements in a variety of medical image segmentation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0076] Figure 1 This is the overall network architecture diagram;

[0077] Figure 2 It is a specific flow chart of the local information extraction module;

[0078] Figure 3 This is the Spilt Attention workflow diagram;

[0079] Figure 4 This is the specific flow chart of the enhanced local information extractor. DETAILED DESCRIPTION

[0080] The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0081] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0082] like Figure 1 As shown, this embodiment provides a medical image segmentation model based on local enhancement driven global optimization, adopts a U-shaped structure of encoder-decoder, and in order to maximize the use of foreground information, a jump connection is used between the encoder and the decoder to alleviate the problem of downsampling information loss.

[0083] Channel feature enhancement module:

[0084] like Figure 2 As shown, the workflow of the channel feature enhancement module is:

[0085] S11: Assume that the size of the input feature map is The input feature maps are grouped according to the channel dimension to obtain k feature maps {F1, F2…F k-1 ,F k}:

[0086] {F1,F2…F k-1 ,F k}=Splite(F){F1,F2…F k-1 ,F k}∈R h×w×c / k ;

[0087] S12: From the channel dimension, the feature maps {F1, F2…F k-1 ,F k}Continue to group to get r feature maps, among which, feature map F1 can be obtained after grouping {F 11 ,F 12 …F 1r}, and so on for other feature maps:

[0088] {F 11 ,F 12 …F 1r}=Split(F i ){F11 ,F 12 …F 1r}∈R h×w×c / k / r ; 1≤i≤k;

[0089] S13: Convolution kernels with sizes of 1×1 and 3×3 are used to align feature maps {F i1 ,F i2 …F ir} to extract features and obtain feature maps {F i1 ′,F i2 ′…F ir ′}:

[0090] {F i1 ′,F i2 ′…F ir ′}=Conv3(Conv1({F i1 ,F i2 …F ir})){F i1 ′,F i2 ′…F ir ′∈R h×w×c′ / k};

[0091] S14: The feature map {F i1 ′,F i2 ′…F ir ′} is input into the Spilt Attention module to calculate the relationship between the feature map channels and output the feature map T i :

[0092] T i = SpliAttention({F i1 ′,F i2 ′…F ir ′})T i ∈R h×w×c′ / k ;

[0093] S15: Feature maps {T1, T2…T k} concatenate and output feature map T:

[0094] T=concayenate{T1,T2…T k};

[0095] S16: After the convolution operation, feature fusion is performed with the original feature map F:

[0096] out=conv1(T)+F.

[0097] Specifically, Figure 3 As shown, the workflow of the Spilt Attention module is:

[0098] (1) Assume that the feature map {F i1 ′,F i2 ′…F ir The size of '} is R h×w×c , feature map {F i1 ′,F i2 ′…F ir ′} Feature fusion is performed by adding operation to obtain feature map

[0099] F i =F i1 ′+F i2 ′+…+F ir ′F i ∈R h×w×c ;

[0100] (2) Feature map F i ∈R h×w×c After the global maximum pooling operation, we get F i ∈R 1×1×c :

[0101] F i =Globalpooling(F i )F i ∈R 1×1×c ;

[0102] (3) Feature map F i ∈R 1×1×c After the convolution operation (convolution kernel of size 1×1), normalization operation (BatchNorm2d), and activation function mapping (Rule) operation:

[0103] F i =Relu(BN(Conv1(F i )))F i ∈R 1×1×c′ ;

[0104] (4) Use r parallel convolution kernels (convolution kernels of size 1×1) to extract feature maps F i ∈R 1×1×c′ The semantic information of r feature maps {F i1 ,F i2 …F ir}:

[0105] {F i1 ,F i2 …F ir}=Conv1(F i ){Fi1 ,F i2 …F ir}∈R 1×1×c ;

[0106] (5) Use the Softmax function to separate the feature maps {F i1 ,F i2 …F ir}The channel dimension is calculated to obtain the corresponding channel coefficient:

[0107]

[0108] (6) Channel coefficient and feature map {F i1 ′,F i2 ′…F ir ′} after multiplication, the addition operation is performed to output the feature map:

[0109]

[0110] In summary, the feature map is first divided into several groups twice along the channel dimension, and then information is extracted from the feature map. In this way, the channel feature enhancement module can extract refined foreground information in the feature map. Subsequently, the feature map is input into the Spilt Attention module to calculate the channel attention coefficient and multiply the channel attention coefficient with the feature map to highlight the importance of a certain channel feature map. In this way, the channel feature enhancement module network highlights the importance of the channel feature map with a large proportion of foreground information and weakens the negative impact of the channel feature map with a large proportion of background information. Compared with simply using ResNet50 as a feature extractor in the downsampling process, this method is more conducive to improving the network's ability to discriminate foreground information.

[0111] Enhanced Local Information Extractor:

[0112] like Figure 4 As shown, the workflow of the enhanced local information extractor is:

[0113] S21: Define the size of the input feature map X as R h×w×c , set c = 1, h = w = 4, and use the Softmax activation function on the input feature map to obtain two attention coefficient matrices and Among them, α i (1≤i≤16) is obtained by calculating the correlation between column elements, β i (1≤i≤16) is obtained by calculating the correlation between row elements. The specific calculation formulas of α1 and β1 are as follows:

[0114]

[0115]

[0116] It can be seen that α1 in the coefficient matrix α is obtained by calculating the correlation between the elements X1, X2, X3 and X4 in the input feature map. Similarly, α2, α3 and α4 are also obtained by calculating the correlation between X1, X2, X3 and X4, and the other coefficients are deduced in the same way. At this time, α in the coefficient matrix α is obtained by calculating the correlation between the column elements in the input feature map. i .

[0117] β1, β5, β9 and β in the coefficient matrix β 13 is calculated by calculating the elements X1, X5, X9 and X 13 The correlation between them is obtained. At this time, the β in the coefficient matrix β i They are all obtained by calculating the correlation between row elements in the input feature map.

[0118] S22: Multiply the coefficient matrix by the input feature map and scale the output feature map to obtain two feature maps Q∈R 1×16×1 and K∈R 1×1×16 :

[0119] Q = Reshape(α×X)

[0120] K = Reshape(β×X);

[0121] S23: Perform matrix multiplication on the feature map Q and the feature map K to obtain an element matrix M1. At this time, each element in the element matrix M1 is obtained by using the correlation between the row and column elements in the input feature map. For example, the element Q1K1 in the element matrix is ​​obtained by using the elements X1, X2, X3, X4, X5, X9 and X1 in the input feature map. 13 The correlation between the obtained:

[0122]

[0123] S24: Calculate the Euclidean distance between each element in the element matrix and the diagonal element to obtain the similarity matrix, and then use the Sigmoid derivative to convert the similarity matrix into a self-attention coefficient matrix:

[0124]

[0125] Among them, λ i,j represents the self-attention coefficient of a certain spatial position, and Q i K j -Qi K i It represents the specific calculation formula for converting a point element in the element matrix into a similarity value using Euclidean distance;

[0126] S25: Self-attention coefficient matrix and scale-transformed feature map V∈R 1×16×1 Perform matrix multiplication to output feature map Y:

[0127] Y i =λ i,1 ×v1+λ i,2 ×v2+…+λ i,15 ×v 15 +λ i,16 ×v 16 .

[0128] Each element Y in the output feature map i are generated by using all elements of the feature map V to participate in the calculation. In addition, the coefficient λ i,j It is obtained by calculating the correlation between row and column elements in the input feature map. In summary, the enhanced local information extractor simultaneously performs local modeling capabilities and global modeling during the feature extraction process. In this way, the enhanced local information extractor can focus on the details of the lesion area and the overall contour information at the same time.

[0129] Dataset:

[0130] In order to evaluate the method proposed in this embodiment, three popular data sets with different numbers of categories and different imaging methods were selected for verification under the same experimental parameters.

[0131] Synapse multi-organ segmentation dataset (Synapse): This dataset is collected from 30 cases and consists of 3779 clinical CT images. Each image contains 8 abdominal organs, including aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, stomach, etc. Among them, 2211 images are used as training sets and the rest are used as test sets.

[0132] Automatic Cardiac Diagnosis Challenge Dataset (ACDC): This dataset was collected from 100 patients using MRI scanners and contains a total of 1,579 images, each of which is labeled with the left ventricle, right ventricle, and myocardium. Among them, 1,539 images are used as training sets and the remaining 40 images are used as test sets.

[0133] Myeloma plasma cell segmentation dataset (SegPC-2021): This dataset has a total of 497 images, each of which contains several myeloma plasma cells. The purpose is to segment the cytoplasm and nucleus. Among them, 398 images are used as training sets and the remaining 99 images are used as test sets.

[0134] Evaluation indicators:

[0135] The model uses Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) as evaluation indicators. The DSC calculation formula is:

[0136]

[0137] The formula for calculating HD is:

[0138] HD(A,B)=max{(h(A,B),h(B,A))}

[0139] Among them, A represents the predicted segmentation result, B represents the actual segmentation result; h(A,B) represents the maximum value of the shortest distance from each point in A to B, that is:

[0140] h(A,B)=max[min(d(a,b))]

[0141] Among them, d(a,b) represents the Euclidean distance between a and b.

[0142] Experimental analysis:

[0143] The medical image analysis model of this embodiment is compared with other network models in abdominal multi-organ segmentation and myeloma plasma cell segmentation. The comparison results are shown in Tables 1 and 2.

[0144] Table 1 Comparative experimental results of abdominal multi-organ segmentation dataset

[0145]

[0146] Table 2 Comparative experimental results of myeloma plasma cell segmentation dataset

[0147]

[0148] After extensive experimental verification, the network model developed in this paper has shown significant performance improvements in a variety of medical image segmentation tasks. This result not only confirms the excellent performance of the method of this embodiment in depicting the boundary of lesion tissue, but also highlights the great value of enhanced local information extractors in promoting the development of medical image segmentation technology.

[0149] Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A medical image segmentation model based on local enhancement driven global optimization, It is characterized in that A U-shaped structure of encoder-decoder is adopted, and a skip connection is adopted between the encoder and decoder to alleviate the problem of downsampling information loss; The encoder includes a channel feature enhancement module, and an enhanced local information extractor is used at the jump connection to extract global context confidence and local detail information in the feature map; The decoder uses a bilinear interpolation algorithm to restore the feature map output by the encoder to its original size.

2. The medical image segmentation model based on local enhancement driven global optimization according to claim 1, It is characterized in that The workflow of the channel feature enhancement module is as follows: S11: Assume that the size of the input feature map is The input feature maps are grouped from the channel dimension to obtain k feature maps {F 1 ,F 2 …F k-1 ,F k }; S12: From the channel dimension, the feature map {F 1 ,F 2 …F k-1 ,F k }Continue to group to get r feature maps, among which the feature map F 1 After grouping, we can get {F 11 ,F 12 …F 1r }, and so on for other feature maps; S13: Convolution kernels with sizes of 1×1 and 3×3 are used to align feature maps {F i1 ,F i2 …F ir } to extract features and obtain feature maps {F i1 ′,F i2 ′…F ir '}; S14: The feature map {F i1 ′,F i2 ′…F ir ′} is input into the Spilt Attention module to calculate the relationship between the feature map channels and output the feature map T i ; S15: Feature map {T 1 ,T 2 …T k }Concatenate and output feature map T; S16: After the convolution operation, feature fusion is performed with the original feature map F.

3. The medical image segmentation model based on local enhancement driven global optimization according to claim 2, It is characterized in that The formula for S11 is: {F 1 ,F 2 …F k-1 ,F k }=Splite(F){F 1 ,F 2 …F k-1 ,F k }∈R h×w×c / k ; The formula for S12 is: {F 11 ,F 12 …F 1r }=Split(F i ){F 11 ,F 12 …F 1r }∈R h×w×c / k / r ;1≤i≤k; The formula for S13 is: {F i1 ′,F i2 ′…F ir ′}=Conv3(Conv1({F i1 ,F i2 …F ir })){F i1 ′,F i2 ′…F ir ′∈R h×w×c′ / k }; The formula for S14 is: T i =SpliAttention({F i1 ′,F i2 ′…F ir ′}) T i ∈R h×w×c′ / k ; The formula for S15 is: T=concayenate{T 1 ,T 2 …T k }; The formula for S16 is: out=conv1(T)+F.

4. The medical image segmentation model based on local enhancement driven global optimization according to claim 2, It is characterized in that In S14, the workflow of the Spilt Attention module is as follows: (1) Assume that the feature map {F i1 ′,F i2 ′…F ir The size of '} is R h ×w×c , feature map {F i1 ′,F i2 ′…F ir ′} Feature fusion is performed by adding operation to obtain feature map (2) Feature map F i ∈R h×w×c After the global maximum pooling operation, we get F i ∈R 1×1×c ; (3) Feature map F i ∈R 1×1×c It undergoes convolution operation, normalization operation, and activation function mapping operation in sequence; (4) Use r parallel convolution kernels to extract feature maps F i ∈R 1×1×c′ The semantic information of r feature maps {F i1 ,F i2 …F ir }; (5) Use the Softmax function to separate the feature maps {F i1 ,F i2 …F ir }The channel dimension is calculated to obtain the corresponding channel coefficient; (6) Channel coefficient and feature map {F i1 ′,F i2 ′…F ir ′} and then perform addition operation to output the feature map.

5. The medical image segmentation model based on local enhancement driven global optimization according to claim 4, It is characterized in that The formula for step (1) is: F i =F i1 ′+F i2 ′+…+F ir ′F i ∈R h×w×c ; The formula for step (2) is: F i =Globalpooling(F i ) F i ∈R 1×1×c ; The formula for step (3) is: F i =Relu(BN(Conv1(F i ))) F i ∈R 1×1×c′ ; The formula for step (4) is: {F i1 ,F i2 …F ir }=Conv1(F i ) {F i1 ,F i2 …F ir }∈R 1×1×c ; The formula for step (5) is: The formula for step (6) is:

6. The medical image segmentation model based on local enhancement driven global optimization according to claim 1, It is characterized in that The workflow of the enhanced local information extractor is as follows: S21: Define the size of the input feature map X as R h×w×c , set c = 1, h = w = 4, and use the Softmax activation function on the input feature map to obtain two attention coefficient matrices and S22: Multiply the coefficient matrix by the input feature map and scale the output feature map to obtain two feature maps Q∈R 1×16×1 and K∈R 1×1×16 ; S23: Perform matrix multiplication on the feature map Q and the feature map K to obtain the element matrix M 1 ; S24: Calculate the Euclidean distance between each element in the element matrix and the diagonal element to obtain a similarity matrix, and then use the Sigmoid derivative to convert the similarity matrix into a self-attention coefficient matrix; S25: Self-attention coefficient matrix and scale-transformed feature map V∈R 1×16×1 Perform matrix multiplication to output the feature map Y.

7. The medical image segmentation model based on local enhancement driven global optimization according to claim 5, It is characterized in that In S21, α 1 and β 1 The specific calculation formula is as follows: The formula for S22 is: Q = Reshape(α×X) K = Reshape(β×X); The formula for S23 is: The formula for S24 is: Among them, λ i,j represents the self-attention coefficient of a certain spatial position, and Q i K j -Q i K i It represents the specific calculation formula for converting a point element in the element matrix into a similarity value using Euclidean distance; The formula for S25 is: Y i =λ i,1 ×v 1 +λ i,2 ×v 2 +…+l i,15 ×v 15 +λ i,16 ×v 16 。 8. The medical image segmentation model based on local enhancement driven global optimization according to claim 1, It is characterized in that The model was evaluated on the Synapse multi-organ segmentation dataset, the automatic cardiac diagnosis challenge dataset, and the myeloma plasma cell segmentation dataset.

9. The medical image segmentation model based on local enhancement driven global optimization according to claim 1, It is characterized in that The model uses a mixture of Dice Similarity Coefficient and Hausdorff Distance as evaluation indicators, where the DSC calculation formula is: The formula for calculating HD is: HD(A,B)=max{(h(A,B),h(B,A))}; Among them, A represents the predicted segmentation result, B represents the actual segmentation result; h(A,B) represents the maximum value of the shortest distance from each point in A to B, that is: h(A,B)=max[min(d(a,b))] Among them, d(a,b) represents the Euclidean distance between a and b.