A pneumonia segmentation method based on multi-connection axial graph convolution network

Abstract

The present application relates to the technical field of medical image processing, in particular to a kind of pneumonitis segmentation method based on multi-connection axial graph convolution network, comprising the following steps: obtaining chest CT image;Threshold segmentation and morphological processing are used in chest CT image, and the initial region of interest mask of lung is generated;Chest CT image and the initial region of interest mask of lung are input into multi-connection axial graph convolution network MAGNN, and the feature map for pneumonitis infection area segmentation in chest CT image is extracted and encoded;The feature map after coding is decoded and up-sampling processing, and pneumonitis infection area segmentation graph in chest CT image is obtained.The present application effectively captures the spatial and semantic information of image in graph convolution network by introducing multi-connection axial attention mechanism and multi-connection compression extraction module, and the accuracy of pneumonitis segmentation is improved by combining efficient boundary alignment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and particularly relates to a pneumonia segmentation method based on a multi-connection axial graph convolutional network. BACKGROUND

[0002] In recent years, pneumonia segmentation based on computed tomography (CT) images plays an important role in clinical diagnosis. Traditional convolutional neural networks mainly rely on local feature extraction when processing image segmentation problems, but the performance of the convolutional neural network is limited for complex boundaries and asymmetric structures in pneumonia segmentation.

[0003] Therefore, the existing segmentation method is difficult to fully capture the long-range dependence and spatial relationship of the lung infection area, thereby affecting the accuracy of segmentation. SUMMARY

[0004] The present application aims to provide a pneumonia segmentation method based on a multi-connection axial graph convolutional network to solve the technical problem that the existing technology is difficult to fully capture the long-range dependence and spatial relationship of the lung infection area, thereby affecting the accuracy of segmentation.

[0005] To solve the above technical problems, the present application specifically provides the following technical solutions:

[0006] A pneumonia segmentation method based on a multi-connection axial graph convolutional network, comprising the following steps:

[0007] Obtaining a chest CT image;

[0008] Threshold segmentation and morphological processing are used in the chest CT image to generate an initial lung region of interest mask;

[0009] The chest CT image and the initial lung region of interest mask are input into a multi-connection axial graph convolutional network (MAGNN) to extract a feature map for pneumonia infection region segmentation from the chest CT image and encode the feature map;

[0010] The encoded feature map is decoded and up-sampled to obtain a pneumonia infection region segmentation map in the chest CT image.

[0011] As a preferred scheme of the present application, the method for generating the initial lung region of interest mask comprises:

[0012] The chest CT image is binarized by using the Otsu method, and the left lung region and the right lung region in the chest CT image are retained by connected component analysis, and morphological operation Morph() is applied to ensure that the mask boundary is aligned to obtain the initial lung region of interest mask;

[0013] The expression for generating the initial lung region of interest mask is:

[0014] Seg=MorPh(OtsuThreshold(I));

[0015] Where Seg is the initial region of interest mask of the lung, I is the chest CT image, OtsuThreshold is the Otsu method function identifier, and Morph is the identifier of the morphological operation Morph() function.

[0016] As a preferred solution of the present invention, the structure of the multi-connected axial graph convolutional network MAGNN is composed of multiple layers of encoders, each layer of encoder includes a multi-connected axial graph convolutional layer MAA and a multi-region compression extraction layer MSE;

[0017] The multi-connected axial graph convolutional layer (MAA) is used to separate the feature information of the width axis and the height axis, and to determine the dependency between local features and global features by calculating the maximum difference between each node feature and its neighboring node features.

[0018] The multi-region compression extraction layer MSE is used to adaptively compress and extract features of different channels and regions of interest;

[0019] The operational expression of the multi-connected axial graph convolutional network MAGNN is:

[0020] Z (l) =MAA(MSE(Z (l―1) ,Seg (l―1) ));

[0021] Where Z (l) is the feature map output by the l-th layer encoder, Z (l―1) is the feature map output by the l-1 layer encoder, Seg (l―1) is the region of interest mask corresponding to the l-1th layer encoder, MSE is the identifier of the multi-region compression extraction layer MSE, and MAA is the identifier of the multi-connected axial graph convolution layer MAA.

[0022] As a preferred solution of the present invention, the decoding and upsampling process is implemented by a multi-layer decoder, and the decoding and upsampling process is expressed as:

[0023]

[0024] Where, is the feature map output by the l-th layer decoder, is the feature map output by the l+1 layer decoder, Z (l) is the feature map output by the l-th layer encoder, Seg (l) is the region of interest mask corresponding to the l-th layer encoder, Decoder (l)Identifier of the first layer encoder.

[0025] As a preferred scheme of the present application, the method for adaptive compression and extraction of features of different channels and regions of interest by the multi-region compression and extraction layer MSE comprises:

[0026] Firstly, channel compression and activation: the spatial information of the input feature map of the multi-region compression and extraction layer MSE is extracted in each channel, and the activation value of the channel is adjusted by an activation function to dynamically enhance the features related to the specific channel;

[0027] Wherein, the channel compression operation F c-squeeze extracts spatial information from the input feature map;

[0028] cs=F c―squeeze (Z);

[0029] In the formula, Z is the input feature map of the multi-region compression and extraction layer MSE, cs is the descriptor of the feature map after channel compression, F c-squeeze is the identifier of the channel compression operation;

[0030] The channel activation operation F c-excite uses the trainable weight W cs in the channel activation operation to activate cs to obtain the channel scaling coefficient;

[0031] c=F c―excite (cs,W cs );

[0032] In the formula, ce is the channel scaling coefficient, F c-excite is the identifier of the channel activation operation, and W cs is the trainable weight in the channel activation operation;

[0033] Secondly, region compression and activation: the input feature map of the multi-region compression and extraction layer MSE is globally averaged and pooled to extract the global spatial information of each region of interest, and an activation function is used to enhance the feature representation of each region;

[0034] Wherein, the region compression operation F r-squeeze globally averages and pools each region of interest region m of the input feature map to extract the global information at the region level;

[0035]

[0036] In the formula, Z is the input feature map, |region m | is the number of spatial points in the mth region of interest region m ; and F represents the feature value of the input feature map Z at the kth channel of the spatial position (i, j), rs m is a descriptor of the region compressed feature map, F r-squeeze is a region compression operation identifier;

[0037] Region activation operation F r-excite Utilize the trainable weight W in the region activation operation rs rs m , get the region scaling coefficient; activation process figure 1 , get the region scaling coefficient;

[0038]

[0039] In the formula, re is the region scaling coefficient, sigma is the Sigmoid activation function, delta is the ReLU activation function, F r-excite is a region activation operation identifier, W rs is a trainable weight in the region activation operation, including and

[0040] Third step, re-scaling operation: according to the channel scaling coefficient and the region scaling coefficient, respectively, scale the channel and region of the input feature map;

[0041] Where, on the basis of the channel scaling coefficient ce and the region scaling coefficient re, the input feature map is adjusted by using the re-scaling function F scale ;

[0042] F scale (Z, re, ce) = ce o ∑ m (re m · Z · 1 ((i, j) e region m );

[0043] In the formula, F scale (Z, re, ce) is a re-scaling function, which re-scales the feature Z by using the obtained re and ce vectors, re m is the region scaling coefficient in region m , 1 ((i, j) e region m )) is an indicator function, and o represents element-level multiplication.

[0044] As a preferred scheme of the present application, the method for separating feature information of width and height axes of the multi-connection axial graph convolution layer MAA comprises:

[0045] In a first step, a separation processing of the width axis and the height axis is performed, in which the input feature map of the multi-connection axial graph convolution layer MAA is divided along the width axis and the height axis to extract the feature components in the width direction and the feature components in the height direction, respectively.

[0046] In a second step, local features and global features are captured, in which the attention mechanism is used to model the dependence of the feature components in the width axis and the height axis, the local features are captured through the attention mechanism, the short-distance spatial detail correlation is emphasized, the global features are captured, and the long-distance spatial context relationship is modeled.

[0047] In a third step, the information of the width axis and the height axis is fused, in which the features processed in the width axis and the height axis are recombined to form a global feature representation through weighted summation or a specific fusion mechanism, so that the information of the spatial dimension can be optimized cooperatively, and the segmentation performance is improved.

[0048] As a preferred scheme of the present application, the method for determining the dependence relationship between the local features and the global features by calculating the maximum difference between the feature of each node and the feature of the neighboring node of the multi-connection axial graph convolution layer MAA comprises:

[0049] The maximum difference between the feature of each node and the feature of the neighboring node is calculated to obtain the dependence relationship between the local features:

[0050] The expression for calculating the maximum difference between the feature of each node position (i, j) of the input feature map of the multi-connection axial graph convolution layer MAA and the feature of the neighboring node position is:

[0051]

[0052] In the formula, Δz p,(i,h) is the maximum difference between the feature of the node position (i, j) and the feature of the neighboring node position, N(i,j) is the neighborhood of the node position (i, j), z (i,j) is the feature value at the node position (i, j), and z (k,l) is the feature of the neighboring node position (k, l).

[0053] As a preferred scheme of the present application, the loss function of the multi-connection axial graph convolution network MAGNN is a weighted combination of the Dice loss and the cross-entropy loss,

[0054] The evaluation indexes of the multi-connection axial graph convolution network MAGNN include:

[0055] The Dice coefficient (Dice Coefficient) is used to evaluate the overlap degree of the segmentation result and the true label, and the value range is 0 to 1, and the larger the value is, the closer the segmentation result is to the true region.

[0056] Intersection over Union (IoU): the proportion of the intersection of the segmentation region and the true region to the union, used to quantify the accuracy of the segmentation result, the larger the value is the better.

[0057] Hausdorff Distance 95th Percentile (HD95): measures the maximum distance between the segmentation boundary and the true boundary, taking the 95th percentile to reduce the influence of outliers, the smaller the value is the better, indicating that the boundary alignment is more accurate.

[0058] Compared with the prior art, the present application has the following beneficial effects:

[0059] The present application effectively captures the spatial and semantic information of the image in the graph convolution network by introducing the multi-connection axial attention mechanism and the multi-connection compression extraction module, and combines efficient boundary alignment to improve the accuracy of pneumonia segmentation. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other embodiments can be obtained without creative labor on the basis of the provided drawings.

[0061] figure 2 The flowchart of the pneumonia segmentation method based on the multi-connection axial graph convolution network provided by the embodiment of the present application;

[0062] figure 3 The lung region of interest mask provided by the embodiment of the present application;

[0063] figure 4 The multi-region compression extraction layer MSE processing flowchart provided by the embodiment of the present application;

[0064] figure 5 The multi-connection axial graph convolution layer MAA processing flowchart provided by the embodiment of the present application;

[0065] figure 6 The structure diagram of the pneumonia segmentation model based on the multi-connection axial graph convolution network provided by the embodiment of the present application;

[0066] figure 7 The model effect evaluation diagram on the MosMedData data set provided by the embodiment of the present application;

[0067] figure 1The model effect evaluation diagram on the COVID-19 CT Lung and Infection Segmentation dataset provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0069] As shown in figure 5 and figure 2 The present application provides a pneumonia segmentation method based on a multi-connection axial graph convolutional network, comprising the following steps:

[0070] obtaining a chest CT image;

[0071] generating an initial lung region of interest mask by threshold segmentation and morphological processing in the chest CT image;

[0072] inputting the chest CT image and the initial lung region of interest mask into a multi-connection axial graph convolutional network MAGNN, extracting a feature map for pneumonia infection region segmentation in the chest CT image and encoding the feature map;

[0073] decoding and up-sampling the encoded feature map to obtain a pneumonia infection region segmentation map in the chest CT image.

[0074] The present application realizes long-range dependence modeling of the infection region by adding a multi-connection axial attention mechanism (MAA) in the graph convolutional network (MAGNN), realizes channel compression and spatial compression by combining a multi-region layer (MSE) design, captures global and local features, improves the expression ability of the features, and effectively reduces the calculation amount.

[0075] Specifically, the multi-region compression extraction layer MSE in the present application compresses and extracts the features of different channels and regions of interest, ensures that the model can fuse rich context information at different scales, and improves the segmentation result.

[0076] The multi-connection axial graph convolutional layer MAA in the present application models the features of the width axis and the height axis of the image respectively, thereby capturing the local and global dependence relationship. This separation axis information processing method can effectively reduce the calculation complexity, while ensuring that the features in each direction can be fully transmitted and captured.

[0077] The method for generating the initial lung region of interest mask comprises:

[0078] The chest CT image is binarized by using the Otsu method, and the left lung region and the right lung region in the chest CT image are reserved through connected domain analysis, and morphological operation Morph() is applied to ensure that the mask boundary is aligned, to obtain an initial lung region of interest mask;

[0079] The generation expression of the initial lung region of interest mask is:

[0080] Seg=MorPh(OtsuThreshold(I));

[0081] In the formula, Seg is the initial lung region of interest mask, I is the chest CT image, OtsuThreshold is the function identifier of the Otsu method, and Morph is the identifier of the morphological operation Morph() function.

[0082] The threshold segmentation and morphological processing flow is as follows:

[0083] The lung region (left lung region and right lung region) is extracted based on the set threshold using the pixel value range of the CT image.

[0084] The morphological operation (such as dilation and erosion) is used to remove noise and small artifacts, and an initial mask Seg (such as shown in figure 3 ) is generated.

[0085] The structure of the multi-connection axial graph convolutional network MAGNN is composed of multiple layers of encoders, and each layer of encoder includes a multi-connection axial graph convolutional layer MAA and a multi-region compression extraction layer MSE.

[0086] The multi-connection axial graph convolutional layer MAA is used to separate the feature information of the width axis and the height axis, and determine the dependency relationship between the local features and the global features by calculating the maximum difference between the feature of each node and the feature of its neighborhood nodes.

[0087] The multi-region compression extraction layer MSE is used for adaptive compression and extraction of features of different channels and regions of interest.

[0088] The operation expression of the multi-connection axial graph convolutional network MAGNN is:

[0089] Z (l) =MAA(MSE(Z (l―1) ,Seg (l―1) ));

[0090] In the formula, Z (l) is the feature map output by the l-th layer encoder, Z (l―1) is the feature map output by the (l-1)-th layer encoder, and Seg (l―1)is the mask of the region of interest corresponding to the l-1th layer encoder, MSE is the identifier of the multi-region compression extraction layer MSE, and MAA is the identifier of the multi-connected axial graph convolution layer MAA.

[0091] Seg (l―1) is the mask of the region of interest corresponding to the l-1th layer encoder, and is obtained by Seg (l―2) is obtained by performing an interpolation operation on the feature map obtained after the size is reduced by half.

[0092] The decoding and up-sampling processing is implemented by a multi-layer decoder, and the decoding and up-sampling processing expression is:

[0093]

[0094] In the formula, is the feature map output by the lth layer decoder, is the feature map output by the l+1th layer decoder, and Z (l) is the feature map output by the lth layer encoder, and Seg (l) is the mask of the region of interest corresponding to the lth layer encoder, Decoder (l) is the identifier of the lth layer encoder.

[0095] The Squeeze-and-Excitation (SE) layer adjusts the importance of each feature map adaptively, mainly recalibrates the features at the channel level, and improves the feature representation capability. However, in medical images, especially the distribution of lung lesions often spans multiple regions of interest (ROI), and these regions have different features. The SE layer adjusts at the channel level and cannot effectively capture the spatial relationship and discontinuous feature distribution between these regions. The present application expands the function of the SE layer by using a multi-region compression extraction layer MSE, which not only compresses and recalibrates the channels, but also recalibrates the regions at the region level. The MSE layer combines the channel compression and region compression operations, as follows:

[0096] As shown in activation process , the method for adaptively compressing and extracting features of different channels and regions of interest by the multi-region compression extraction layer MSE includes:

[0097] First, channel compression and activation: the spatial information of the input feature map of the multi-region compression extraction layer MSE is extracted in each channel, and the activation value of the channel is adjusted through an activation function to dynamically enhance the features related to a specific channel;

[0098] wherein the channel compression operation F c-squeeze extracts spatial information from the input feature map;

[0099] cs=F c―squeeze (Z);

[0100] where Z is the input feature map of the multi-region compression extraction layer MSE, cs is the descriptor of the channel compressed feature map, F c-squeeze is the channel compression operation identifier;

[0101] Channel activation operation F c-excite uses the trainable weight W cs in the channel activation operation to activate cs to obtain the channel scaling coefficient;

[0102] ce=F c―excite (cs,W cs );

[0103] where ce is the channel scaling coefficient, F c-excite is the channel activation operation identifier, and W cs is the trainable weight in the channel activation operation;

[0104] Second step, region compression and activation: the input feature map of the multi-region compression extraction layer MSE is globally averaged pooled to extract the global spatial information of each region of interest, and an activation function is used to enhance the feature representation of each region;

[0105] where the region compression operation F r-squeeze performs global average pooling on each region of interest region m of the input feature map to extract the global information at the region level;

[0106]

[0107] where Z is the input feature map, |region m | is the number of spatial points in the mth region of interest region m , and z m (i,j,k) represents the feature value of the kth channel of the input feature map Z at the spatial position (i,j). r-squeeze is the descriptor of the region compressed feature map, F r-excite is the region compression operation identifier;

[0108] Region activation operation F rs uses the trainable weight W m in the region activation operation to activate rs r-excite to obtain the region scaling coefficient; figure 4 figure 6

[0109]

[0110] where re is the region scaling coefficient, σ is the Sigmoid activation function, δ is the ReLU activation function, and F​r-excite is an identifier of a region activation operation, W rs is a trainable weight in the region activation operation, comprising and

[0111] Third step, re-calibration operation: according to the channel scaling coefficient and the region scaling coefficient, respectively, scaling the channel and region of the input feature map, to ensure that the model can adaptively strengthen the features of the key region and channel. This extraction and optimization adjustment of features helps to significantly improve the segmentation effect, thereby improving the model performance;

[0112] wherein, on the basis of the channel scaling coefficient ce and the region scaling coefficient re, the re-calibration function F scale adjusts the input feature map;

[0113] F scale (Z,re,ce)=ce⊙∑ m (re m ·Z·1((i,j)∈region m );

[0114] In the formula, F scale (Z,re,ce) is a re-calibration function, which uses the obtained re and ce vectors to re-calibrate the feature Z, re m is the region scaling coefficient in region m 1((i,j)∈region m )) is an indicator function, which ensures that only the region scaling coefficient re m is applied to the region in region m , and ⊙ represents element-level multiplication, which applies channel-level scaling.

[0115] The multi-connection axial graph convolution layer MAA captures local and global dependencies by modeling the features of the width axis and the height axis of the image respectively. This separation of axis information processing can effectively reduce the computational complexity while ensuring that the features in each direction can be fully transmitted and captured, as follows:

[0116] As shown in figure 7 , the method for separating the feature information of the width axis and the height axis of the multi-connection axial graph convolution layer MAA includes:

[0117] First step, separation processing of the width axis and the height axis: the input feature map of the multi-connection axial graph convolution layer MAA is axially divided along the width axis and the height axis, and the feature components in the width direction and the feature components in the height direction are extracted respectively;

[0118] Second step, capturing of local and global feature dependencies: attention mechanism is used to model the dependencies of feature components on the width and height axes, local features are captured through attention mechanism to emphasize the details of short-distance space, and global features are captured to model the context relationship of long-distance space;

[0119] Third step, fusion of width axis and height axis information: the features processed by the width axis and the height axis are recombined to form a global feature representation through weighted summation or a specific fusion mechanism, ensuring that the information of the spatial dimension can be optimized cooperatively to improve the segmentation performance.

[0120] In order to capture local and global feature dependencies, the MAA layer not only calculates the features of a single node, but also calculates the relationship between each node and its neighborhood nodes. By calculating the maximum difference between the features of a node and its neighborhood nodes, the model can understand the particularity of the node in the local area and its relationship with other nodes.

[0121] The method for determining the dependency relationship between local features and global features by the multi-connection axial graph convolution layer MAA includes:

[0122] Calculate the maximum difference between the features of each node and its neighborhood nodes to obtain the dependency relationship between local features:

[0123] The expression for calculating the maximum difference between the features of each node position (i, j) of the input feature map of the multi-connection axial graph convolution layer MAA and the features of the neighborhood node position is:

[0124]

[0125] In the formula, Δz p,(i,j) is the maximum difference between the features of the node position (i, j) and the features of the neighborhood node position, N(i,j) is the neighborhood of the node position (i, j), z (i,j) is the feature value at the node position (i, j), and z (k,l) is the feature of the neighborhood node position (k, l).

[0126] The capturing of local and global feature dependencies refers to the model's ability to distinguish and effectively model the features of different regions in the image. Specifically:

[0127] Local feature dependency: represents the relationship between a node and its neighborhood nodes within its local range. For example, maximum difference, relative distance between nodes, or similarity within the neighborhood. This feature is very important for handling details and local changes.

[0128] Global feature dependency: refers to the relationship between a node and other nodes in the entire image or a larger range. This part of the feature is crucial for capturing global context information and global consistency.

[0129] The local features obtained by the application are crucial for detail areas (such as the boundaries of lung lesions). MAA can more accurately capture the boundary information of the lesion area by calculating the maximum difference between each node and the neighborhood nodes, thereby improving the segmentation ability of the model in detail areas (especially complex lung lesions).

[0130] The global features obtained by the application help the model understand the relationship between different regions in the image. In medical images, the distribution of lesions is usually greatly influenced by the global context (for example, the spread of lung inflammation), so global information is very important for improving the overall segmentation effect.

[0131] Therefore, the application realizes long-range dependency modeling of infected areas by adding a multi-connection axial attention mechanism in the graph convolution network, improves the segmentation ability of the model in detail areas (especially complex lung lesions), and improves the overall segmentation effect.

[0132] The loss function of the multi-connection axial graph convolution network MAGNN is a weighted combination of Dice loss and cross-entropy loss,

[0133] The evaluation index of the multi-connection axial graph convolution network MAGNN includes:

[0134] Dice coefficient: used to evaluate the overlap between the segmentation result and the true label, the value range is 0 to 1, the larger the value, the closer the segmentation result to the true region.

[0135] Intersection over Union (IoU): the intersection of the segmented region and the true region accounts for the proportion of the union, which is used to quantify the accuracy of the segmentation result, the larger the value, the better.

[0136] Hausdorff distance 95th percentile (HD95): measures the maximum distance between the segmentation boundary and the true boundary, takes the 95th percentile to reduce the influence of outliers, the smaller the value, the better, indicating that the boundary alignment is more accurate.

[0137] Through experiments on multiple public data sets, the application proves that the MAGNN model achieves the most advanced performance in the lung inflammation CT image segmentation task and significantly reduces the parameter amount, as follows:

[0138] As ​ shown in the performance on the MosMedData dataset, the results show that the MAGNN model shows the highest segmentation performance, with a Dice coefficient of 0.8611, a Jaccard index of 0.7591, and a HD95 value of 1.5986. Compared with traditional models such as Unet and Unet++, MAGNN improves about 5% in Dice coefficient, while only requiring 3.5 million parameters. When compared with high-capacity Transformer-based models such as TransUNet (109.5 million parameters) and H2Former (33.7 million parameters), MAGNN achieves higher accuracy with fewer parameters.

[0139] It is worth noting that the performance of Transformer-based models such as TransUNet and H2Former on the MMDD dataset is lower than expected, which may be due to the limitation of the small size of the dataset. Transformer models usually require large datasets to effectively learn complex long-distance dependencies, and it is difficult to fully utilize their advantages in the case of limited data. While the design of MAGNN is optimized for small datasets, it can achieve a good balance between parameter efficiency and segmentation accuracy.

[0140] In addition, although the UNext model has only 1.5 million parameters, its lightweight convolutional architecture exhibits excellent efficiency, but its simplified design limits the ability to capture complex long-distance dependencies, so it performs poorly in accurately segmenting complex anatomical structures compared to MAGNN. The above results show that MAGNN achieves an excellent trade-off between model complexity and segmentation performance, making it suitable for application in medical image segmentation tasks with limited data.

[0141] As ​ shown in the performance on the COVID-19CT Lung and Infection Segmentation dataset, the results show that compared with the MosMedData dataset, CLISD provides a larger amount of data and high-quality annotations, which improves the performance of most models. The proposed MAGNN model performs best on the CLISD dataset, with a Dice coefficient of 0.9022, a Jaccard index of 0.8219, and an HD95 value of only 1.0.

[0142] Notably, although Attention Unet introduces an attention mechanism, its performance on the CLISD dataset is significantly worse, with a Dice coefficient of only 0.8181 and an HD95 value as high as 2.9984. This performance difference is mainly attributed to the implementation of its attention mechanism. Attention Unet employs localized attention, which, while enhancing features in specific regions, limits the model's overall performance due to its spatial limitations. In the CLISD dataset, capturing extensive spatial relationships is crucial for achieving accurate segmentation.

[0143] In contrast, MAGNN employs axial attention operations that can model both local and global dependencies simultaneously across multiple regions of interest (ROIs). This mechanism enables MAGNN to better capture the complex spatial relationships of infected lung regions, resulting in superior segmentation performance on the CLISD dataset. These results demonstrate that MAGNN outperforms traditional attention mechanism models in terms of accuracy and generalization, particularly in medical image segmentation tasks that require global dependencies.

[0144] The present application effectively captures spatial and semantic information of images in graph convolution networks by introducing a multi-connection axial attention mechanism and a multi-connection compression extraction module, and improves the accuracy of pneumonia segmentation by combining efficient boundary alignment.

[0145] The above examples are only exemplary embodiments of the present application and are not intended to limit the present application, and the protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements shall also be considered to fall within the protection scope of the present application.

Claims

1. A pneumonia segmentation method based on a multi-connected axial graph convolutional network, characterized in that: The following steps are involved: Obtain chest CT images; Threshold segmentation and morphological processing are used in chest CT images to generate the initial region of interest mask of the lungs; The chest CT image and the initial region of interest mask of the lungs are input into the multi-connected axial graph convolutional network MAGNN to extract and encode the feature map for pneumonia infection area segmentation in the chest CT image; Decoding and upsampling the encoded feature map to obtain a segmentation map of the pneumonia infection area in the chest CT image; The structure of the multi-connected axial graph convolutional network MAGNN consists of multiple layers of encoders, each of which includes a multi-connected axial graph convolution layer MAA and a multi-region compression extraction layer MSE; Among them, the multi-connected axial graph convolution layer MAA is used to separate the feature information of the width axis and the height axis, and determine the dependency between local features and global features by calculating the maximum difference between each node feature and its neighboring node features; The multi-region compression extraction layer MSE is used to adaptively compress and extract features of different channels and regions of interest; The operational expression of the multi-connected axial graph convolutional network MAGNN is: ； Where, is the feature map output by the l-th layer encoder, is the feature map output by the l-1 layer encoder, is the region of interest mask corresponding to the l-1th layer encoder, MSE is the identifier of the multi-region compression extraction layer MSE, and MAA is the identifier of the multi-connected axial graph convolution layer MAA.

2. The pneumonia segmentation method based on a multi-connected axial graph convolutional network according to claim 1, characterized in that: The initial region of interest mask of the lungs is generated by: The chest CT image is binarized using the Otsu method, and the left and right lung regions in the chest CT image are retained through connected domain analysis. The morphological operation Morph() is applied to ensure that the mask boundaries are aligned, and the initial region of interest mask of the lung is obtained. The generation expression of the initial region of interest mask of the lung is: Seg=MorPh(OtsuThreshold(I)); Where Seg is the initial region of interest mask of the lung, I is the chest CT image, OtsuThreshold is the Otsu method function identifier, and Morph is the identifier of the morphological operation Morph() function.

3. The pneumonia segmentation method based on a multi-connected axial graph convolutional network according to claim 1, characterized in that: The decoding and up-sampling process is implemented by a multi-layer decoder, and the decoding and up-sampling process expression is: ； Where, is the feature map output by the l-th layer decoder, is the feature map output by the l+1 layer decoder, is the feature map output by the l-th layer encoder, is the region of interest mask corresponding to the l-th layer encoder, is the identifier of the encoder at layer l.

4. The pneumonia segmentation method based on a multi-connected axial graph convolutional network according to claim 3, characterized in that: The method for adaptively compressing and extracting features of different channels and regions of interest by the multi-region compression extraction layer MSE includes: The first step is channel compression and activation: the spatial information of each channel of the input feature map of the multi-region compression extraction layer MSE is extracted, and the activation value of the channel is adjusted through the activation function to dynamically enhance the features related to the specific channel; Among them, the channel compression operation F c-squeeze Extract spatial information from the input feature map; ； Where Z is the input feature map of the multi-region compression extraction layer MSE, cs is the descriptor of the feature map after channel compression, and F c-squeeze The identifier of the channel compression operation; Channel activation operation F c-excite Utilize the trainable weights W in the channel activation operation cs Activate cs to obtain the channel scaling coefficient; ； Where ce is the channel scaling factor, F c-excite is the channel activation operation identifier, W cs Trainable weights in channel activation operations; The second step is region compression and activation: global average pooling is performed on the input feature map of the multi-region compression extraction layer MSE to extract the global spatial information of each region of interest, and the feature representation of each region is enhanced through the activation function; Among them, the region compression operation F r-squeeze For each region of interest in the input feature map m Perform global average pooling to extract global information at the regional level; ； Where Z is the input feature map, |region m ∣ is the mth region of interest m The number of spatial points in , Indicates the spatial position of the input feature map Z The eigenvalue of the kth channel, rs m is the descriptor of the feature map after region compression, F r-squeeze The identifier of the region compression operation; Region activation operation F r-excite Utilize the trainable weights W in the region activation operation rs tors m Perform activation processing to obtain the regional scaling coefficient; ； Where re is the region scaling coefficient, σ is the Sigmoid activation function, δ is the ReLU activation function, and F r-excite is the region activation operation identifier, W rs It is the trainable weight in the region activation operation, including and ; The third step is recalibration operation: according to the channel scaling factor and the area scaling factor, the channel and area of ​​the input feature map are scaled respectively; Among them, based on the channel scaling factor ce and the regional scaling factor re, the recalibration function F is used scale Adjust the input feature map; ； Where, is the recalibration function, which uses the obtained re and ce vectors to recalibrate the feature Z. For region m The regional scaling factor in , 1((i,j)∈region m )) is the indicator function and ⊙ represents element-wise multiplication.

5. The pneumonia segmentation method based on a multi-connected axial graph convolutional network according to claim 4, characterized in that: The method of separating feature information of the width axis and the height axis by the multi-connected axial graph convolution layer MAA includes: The first step is the separation of the width axis and the height axis: the input feature map of the multi-connected axial graph convolution layer MAA is axially split along the width axis and the height axis to extract the feature components in the width direction and the feature components in the height direction respectively; The second step is capturing local and global features: Attention mechanisms are used to model dependencies between feature components along both the width and height axes. This mechanism captures local features, emphasizing detail relevance in short-range spaces, while simultaneously capturing global features and modeling contextual relationships in long-range spaces. The third step is the fusion of width axis and height axis information: Recombining the features processed by width axis and height axis to form a global feature representation through weighted summation or specific fusion mechanism.

6. The pneumonia segmentation method based on a multi-connected axial graph convolutional network according to claim 5, characterized in that: The method of the multi-connected axial graph convolution layer MAA determining the dependency relationship between local features and global features by calculating the maximum difference between each node feature and its neighboring node features includes: Calculate the maximum difference between each node feature and its neighboring node features to obtain the dependency relationship between local features: The expression for calculating the maximum difference between the feature at each node position (i, j) and the feature at the neighboring node position in the input feature graph of the multi-connected axial graph convolution layer MAA is: ； Where, is the maximum difference between the feature at the node position (i, j) and the feature at the neighboring node position, N(i, j) is the neighborhood of the node position (i, j), z (i,j) is the eigenvalue at the node position (i, j), z (k,l) is the feature of the neighborhood node position (k,l).

7. The pneumonia segmentation method based on a multi-connected axial graph convolutional network according to claim 1, characterized in that: The loss function of the multi-connected axial graph convolutional network MAGNN is a weighted combination of Dice loss and cross entropy loss. The evaluation metrics of the multi-connected axial graph convolutional network MAGNN include: Dice coefficient: used to evaluate the degree of overlap between the segmentation result and the true label; Intersection-over-Union: used to quantify the accuracy of segmentation results; Hausdorff distance 95%: measures the maximum distance between the segmentation boundary and the true boundary.

Images