CT image new coronavirus pneumonia region segmentation method and imaging method

Abstract

The application discloses a CT image new coronavirus pneumonia region segmentation method, comprising the following steps: acquiring a new coronavirus pneumonia CT image dataset with label annotation and processing to obtain a training dataset; constructing a CT image new coronavirus pneumonia region segmentation preliminary model; constructing an auxiliary classification network; training the CT image new coronavirus pneumonia region segmentation preliminary model based on the auxiliary classification network to obtain a CT image new coronavirus pneumonia region segmentation model; and using the CT image new coronavirus pneumonia region segmentation model to complete new coronavirus pneumonia region segmentation of an actual CT image. The application also discloses an imaging method comprising the CT image new coronavirus pneumonia region segmentation method. The application can capture more complete region information, continuously reduce the difference between a prediction map and a label, refine the segmentation result, and improve the accuracy of new coronavirus pneumonia region segmentation. Therefore, the application has high reliability, good accuracy and objectivity.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing, specifically relating to a method for segmenting and imaging regions of COVID-19 in CT images. Background Technology

[0002] With the development of economy and technology and the improvement of people's living standards, people are paying more and more attention to health.

[0003] Currently, with the development of artificial intelligence algorithms, AI-assisted technologies in the medical field have been widely applied. CT medical image segmentation technology can effectively assist clinicians in quickly locating relevant regions of interest, thereby improving their work efficiency. Therefore, region segmentation technology for CT images has always been one of the key research focuses for researchers.

[0004] Currently, traditional CT image segmentation methods utilize features such as texture and thresholding in CT images for segmentation, but the results are not ideal. While deep learning-based CT image segmentation methods have improved the segmentation accuracy of regions of interest to some extent, existing methods still suffer from low accuracy and poor reliability when segmenting COVID-19 regions characterized by low contrast and blurred edges. Summary of the Invention

[0005] One of the objectives of this invention is to provide a highly reliable, accurate, and objective scientific method for segmenting COVID-19 regions in CT images.

[0006] The second objective of this invention is to provide an imaging method that includes the aforementioned CT image COVID-19 region segmentation method.

[0007] The COVID-19 region segmentation method in CT images provided by this invention includes the following steps:

[0008] S1. Obtain a dataset of COVID-19 CT images with labels;

[0009] S2. Perform image processing on the dataset obtained in step S1 to obtain the training dataset;

[0010] S3. Based on a local feature fusion network, a preliminary model for COVID-19 region segmentation in CT images is constructed;

[0011] S4. Construct an auxiliary classification network to refine the segmentation results of the preliminary model for COVID-19 region segmentation in CT images;

[0012] S5. Based on the auxiliary classification network constructed in step S4, the training dataset obtained in step S2 is used to train the preliminary model for COVID-19 region segmentation in CT images constructed in step S3, so as to obtain the COVID-19 region segmentation model in CT images.

[0013] S6. Using the COVID-19 region segmentation model of the CT image obtained in step S5, complete the COVID-19 region segmentation of the actual CT image.

[0014] Step S2 involves image processing of the dataset obtained in step S1 to obtain the training dataset, specifically including the following steps:

[0015] Binarize the image to obtain grayscale image I. g ;

[0016] Set the kernel k for morphological filtering to k = getStructuringElement(1, size = [17, 17]), where getStructuringElement() is the reconstruction function;

[0017] Black hat filtering is applied to the grayscale image to obtain the lesion contour map I. e For I e =morphologyEx(I g ,MORPH_BLACKHAT,k), where morphologyEx() is the transformation function;

[0018] Outline of lesion I e Threshold repair is performed to obtain the preprocessed image;

[0019] The obtained preprocessed image is converted into a color image, thus completing the preprocessing of one image;

[0020] Each image in the acquired dataset is preprocessed to obtain the training dataset.

[0021] Step S3, which involves constructing a preliminary model for COVID-19 region segmentation in CT images based on a local feature fusion network, specifically includes the following steps:

[0022] The model includes an encoding layer and a decoding layer;

[0023] Dilated convolutions are used in the encoding layer to expand the receptive field of the network, and depthwise separable convolutions are used to reduce the training parameters of the network. A triple hybrid domain attention mechanism is used in the encoding and decoding layers to promote the expression of lesion information and suppress the expression of non-lesion information.

[0024] The decoding layer of the network consists of a shape network and deconvolution, where the shape network uses gate convolution and residual convolution to determine the complete lesion region by capturing the contour information of the lesion.

[0025] The specific process of shape network to extract contour features is as follows: First, residual convolution is used to obtain the global features of the lesion, and then gated convolution is used to extract the edge contour information of the lesion and filter out non-contour information.

[0026] In the upsampling process, firstly, a 1*1 convolution is performed on the four layers of features to unify the number of channels in the upsampled feature map, preparing for subsequent feature fusion. Then, the sampled features of the last layer of the model are input into the residual block, and the output of the residual block and the first layer upsampling are simultaneously fed into the gated convolution. Finally, the feature map output from the gated convolution is processed again by the residual convolution, and the result is added to the second layer upsampling of the gated convolution. This process is repeated until the last layer of the upsampled feature map is reached.

[0027] The last layer of features in the upsampling process in the model is subjected to a final gated convolution with the previous layer feature map to obtain the contour features of the lesion; the contour features of the lesion extracted by the shape network are fused with the lesion features to obtain a more accurate segmentation result than a single lesion feature.

[0028] The aforementioned triple hybrid domain attention mechanism specifically includes the following steps:

[0029] The input image is sequentially subjected to dilated convolution and depthwise separable convolution to extract the lesion region features F;

[0030] The lesion region feature F is processed by 3*3 average pooling to obtain the average pooled feature F1 as F1 = Avg_Pool(F, pool size =[3,3]). Avg_Pool(·) represents the average pooling operation, pool size =[3,3] indicates that pooling uses a 3x3 convolution kernel;

[0031] In the local feature fusion network, a channel attention map is generated to obtain the channel attention feature F2, F2 = MLP(F1), where MLP(·) is the operation function of the shared network;

[0032] Perform a nonlinear processing on the channel attention feature F2 to obtain the activation region feature F. c F c =σ(F2), where σ() is the Sigmoid function, which is a non-linear activation function;

[0033] Activate region feature F c Perform 7x7 average pooling and 7x7 max pooling respectively to obtain the average pooling activation region feature F. avg and the maximum pooling activation region feature F max ;

[0034] The average pooling activation region feature F avgand the maximum pooling activation region feature F max Channel splicing is performed in the third dimension to obtain the spliced ​​and fused region feature F. f ;

[0035] Features F of the splicing and fusion region f The region feature F is obtained by performing void space pyramid pooling and nonlinear processing. s For F s =σ(ASPP(F) f )), where ASPP() is the hollow space pyramid pooling process, and σ() is the Sigmoid nonlinear activation function;

[0036] The region feature F is input into the position attention module, and three 3*3 convolutions are performed followed by a reshape operation to obtain the first region feature F3, the second region feature F4, and the third region feature F5.

[0037] Perform matrix multiplication on the first region feature F3 and the second region feature F4 to obtain the multiplicative region feature F6.

[0038] Perform matrix multiplication again on the third region feature F5 and the multiplication region feature F6 to obtain the position attention region feature F. p ;

[0039] Activate region feature F c Regional characteristics F s and location attention region features F p The merging and fusion process is performed to obtain the output lesion region features F. out For F out =Stack[F c ,F s ,F p Stack() is an element-wise stacking and merging operation for several features.

[0040] Step S4, which involves constructing an auxiliary classification network to refine the segmentation results of the preliminary model for COVID-19 region segmentation in CT images, specifically includes the following steps:

[0041] The auxiliary classification network is constructed using the following steps:

[0042] Perform a 3x3 convolution operation on the input image to obtain the lesion features F';

[0043] The lesion feature F' is processed using a channel attention module, and then 3*3 average pooling is performed on the lesion feature F' to obtain the average pooled lesion feature F1'. Then, a channel attention map is generated in the shared network to obtain the channel attention feature F2'. Nonlinear processing is performed on the channel attention feature F2' to obtain the lesion activation feature F.c ', F c '=σ(F2'), where σ() is the Sigmoid nonlinear activation function;

[0044] The lesion feature F' is input into the position attention module, and three 3*3 convolutions are performed followed by a reshape operation to obtain the first lesion feature F3', the second lesion feature F4', and the third lesion feature F5'. The first lesion feature F3' and the second lesion feature F4' are then multiplied together to obtain the multiplicative lesion feature F6'. The multiplicative lesion feature F6' and the third lesion feature F5' are then multiplied together to obtain the position attention lesion feature F. p ';

[0045] Activation feature of lesions F c 'and location attention lesion features F p Perform a 3-dimensional multiplication operation and apply an activation function to obtain the final output lesion feature F. out ;

[0046] The classification loss is calculated by using the discrimination results predicted by the iterative auxiliary classification network, and the network parameters of the discriminant network and the preliminary model for COVID-19 region segmentation in CT images are updated by back gradient propagation; specifically, the following steps are included:

[0047] The prediction result P and label L of the preliminary model for COVID-19 region segmentation in CT images are input into the auxiliary classification network to calculate the cross-entropy loss value R of the auxiliary classification network. The parameters of the auxiliary classification network are updated by calculating the discrimination result output by the auxiliary classification network and the classification loss of the label. The output result of the auxiliary classification network is fed back to the preliminary model for COVID-19 region segmentation in CT images, thereby updating the parameters of the preliminary model for COVID-19 region segmentation in CT images in reverse.

[0048] The training described in step S5 specifically includes the following steps:

[0049] The cross-entropy loss, regularization loss, and edge loss between the prediction results of the preliminary model for COVID-19 region segmentation in CT images and the labels are calculated. Then, the loss function of the preliminary model for COVID-19 region segmentation in CT images is obtained by weighting the three types of losses.

[0050] In the training process of the COVID-19 region segmentation model for CT images, the difference between the segmented image generated by L1 distance minimization and the manually annotated segmented image is incorporated. The calculation formula is as follows: Where x is the data input to the local feature fusion network G for training, and P data(x) Let x be the distribution of x, and y be the segmented image generated by G(x,y); The probability distribution of the generated segmented image; the loss function of the local feature fusion network G is expressed as:

[0051]

[0052] In the formula L CGAN (D,G) represents the cross-entropy loss of the local feature fusion network. λ is the regularization loss of the local feature fusion network, EPE(G) is the edge loss of the local feature fusion network, and λ and η are set parameters.

[0053] The loss function of the auxiliary classification network D is:

[0054]

[0055] In the formula y i p represents the label corresponding to image i. i This represents the probability that image i is predicted to be of the positive class.

[0056] This invention also discloses an imaging method that includes the aforementioned CT image COVID-19 region segmentation method, specifically comprising the following steps:

[0057] A. Obtain actual CT images of the lung region;

[0058] B. Using the aforementioned CT image COVID-19 region segmentation method, the COVID-19 region is segmented in the CT image obtained in step S1;

[0059] C. The region segmentation results obtained in step B are annotated and re-imagined on the CT image obtained in step A to obtain the final CT image of the lung region with the COVID-19 region segmentation results.

[0060] The COVID-19 region segmentation and imaging method for CT images provided by this invention designs a local feature fusion network, which extracts richer regional contour information, helping the network to capture more complete regional information. The constructed auxiliary classification network can continuously reduce the difference between the prediction image and the label, further refine the segmentation results, and improve the accuracy of COVID-19 region segmentation. Therefore, this invention has high reliability, good accuracy, and is objective and scientific. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the segmentation method of the present invention.

[0062] Figure 2 This is a schematic diagram of the imaging method of the present invention. Detailed Implementation

[0063] like Figure 1The diagram shown is a flowchart of the segmentation method of the present invention: The COVID-19 region segmentation method in CT images provided by the present invention includes the following steps:

[0064] S1. Obtain a dataset of COVID-19 CT images with labels;

[0065] S2. Perform image processing on the dataset obtained in step S1 to obtain the training dataset; specifically, this includes the following steps:

[0066] Binarize the image to obtain grayscale image I. g ;

[0067] Set the kernel k for morphological filtering to k = getStructuringElement(1, size = [17, 17]), where getStructuringElement() is the reconstruction function;

[0068] Black hat filtering is applied to the grayscale image to obtain the lesion contour map I. e For I e =morphologyEx(I g ,MORPH_BLACKHAT,k), where morphologyEx() is the transformation function;

[0069] Outline of lesion I e Threshold repair is performed to obtain the preprocessed image;

[0070] The obtained preprocessed image is converted into a color image, thus completing the preprocessing of an image;

[0071] Each image in the acquired dataset is preprocessed to obtain the training dataset;

[0072] S3. Based on a local feature fusion network, construct a preliminary model for COVID-19 region segmentation in CT images; specifically including the following steps:

[0073] The model includes an encoding layer and a decoding layer;

[0074] Dilated convolutions are used in the encoding layer to expand the receptive field of the network, and depthwise separable convolutions are used to reduce the training parameters of the network. A triple hybrid domain attention mechanism is used in the encoding and decoding layers to promote the expression of lesion information and suppress the expression of non-lesion information.

[0075] The decoding layer of the network consists of a shape network and deconvolution, where the shape network uses gate convolution and residual convolution to determine the complete lesion region by capturing the contour information of the lesion.

[0076] The specific process of shape network to extract contour features is as follows: First, residual convolution is used to obtain the global features of the lesion, and then gated convolution is used to extract the edge contour information of the lesion and filter out non-contour information.

[0077] In the upsampling process, firstly, a 1*1 convolution is performed on the four layers of features to unify the number of channels in the upsampled feature map, preparing for subsequent feature fusion. Then, the sampled features from the last layer of the model are input into the residual block, and the output of the residual block and the first layer upsampling are simultaneously fed into the gated convolution. Finally, the feature map output from the gated convolution is processed again by the residual convolution, and the result is added to the second layer upsampling of the gated convolution. This process is repeated until the last layer of the upsampled feature map is reached.

[0078] The last layer of features in the upsampling process in the model is subjected to a final gated convolution with the previous layer feature map to obtain the contour features of the lesion; the contour features of the lesion extracted by the shape network are fused with the lesion features to obtain a more accurate segmentation result than a single lesion feature.

[0079] In practice, the triple hybrid domain attention mechanism includes the following steps:

[0080] The input image is sequentially subjected to dilated convolution and depthwise separable convolution to extract the lesion region features F;

[0081] The lesion region feature F is processed by 3*3 average pooling to obtain the average pooled feature F1 as F1 = Avg_Pool(F, pool size =[3,3]). Avg_Pool(·) represents the average pooling operation, pool size =[3,3] indicates that pooling uses a 3x3 convolution kernel;

[0082] In the local feature fusion network, a channel attention map is generated, and the channel attention feature F2 is obtained as F2 = MLP(F1), where MLP() is the operation function of the shared network;

[0083] Perform a nonlinear processing on the channel attention feature F2 to obtain the activation region feature F. c F c =σ(F2), where σ() is the Sigmoid nonlinear activation function;

[0084] Activate region feature F c Perform 7x7 average pooling and 7x7 max pooling respectively to obtain the average pooling activation region feature F. avg and the maximum pooling activation region feature F max ;

[0085] The average pooling activation region feature F avgand the maximum pooling activation region feature F max Channel splicing is performed in the third dimension to obtain the spliced ​​and fused region feature F. f ;

[0086] Features F of the splicing and fusion region f The region feature F is obtained by performing void space pyramid pooling and nonlinear processing. s For F s =σ(ASPP(F) f )), where ASPP() is the hollow space pyramid pooling process, and σ() is the Sigmoid nonlinear activation function;

[0087] The region feature F is input into the position attention module, and three 3*3 convolutions are performed followed by a reshape operation to obtain the first region feature F3, the second region feature F4, and the third region feature F5.

[0088] Perform matrix multiplication on the first region feature F3 and the second region feature F4 to obtain the multiplicative region feature F6.

[0089] Perform matrix multiplication again on the third region feature F5 and the multiplication region feature F6 to obtain the position attention region feature F. p ;

[0090] Activate region feature F c Normalized regional features F s and location attention region features F p The merging and fusion process is performed to obtain the output lesion region features F. out For F out =Stack[F c ,F s ,F p ], where Stack() is an element-wise stacking and merging operation for several features;

[0091] S4. Construct an auxiliary classification network to refine the segmentation results of the preliminary model for COVID-19 region segmentation in CT images; specifically including the following steps:

[0092] The auxiliary classification network is constructed using the following steps:

[0093] Perform a 3x3 convolution operation on the input image to obtain the lesion features F';

[0094] The lesion feature F' is processed using a channel attention module, and then 3*3 average pooling is performed on the lesion feature F' to obtain the average pooled lesion feature F1'. Then, a channel attention map is generated in the shared network to obtain the channel attention feature F2'. Nonlinear processing is performed on the channel attention feature F2' to obtain the lesion activation feature F.c ', F c '=σ(F2'), where σ() is the Sigmoid nonlinear activation function;

[0095] The lesion feature F' is input into the position attention module, and three 3*3 convolutions are performed followed by a reshape operation to obtain the first lesion feature F3', the second lesion feature F4', and the third lesion feature F5'. The first lesion feature F3' and the second lesion feature F4' are then multiplied together to obtain the multiplicative lesion feature F6'. The multiplicative lesion feature F6' and the third lesion feature F5' are then multiplied together to obtain the position attention lesion feature F. p ';

[0096] Activation feature of lesions F c 'and location attention lesion features F p Perform a 3-dimensional multiplication operation and apply an activation function to obtain the final output lesion feature F. out ;

[0097] The classification loss is calculated by using the discrimination results predicted by the iterative auxiliary classification network, and the network parameters of the discriminant network and the preliminary model for COVID-19 region segmentation in CT images are updated by back gradient propagation; specifically, the following steps are included:

[0098] The prediction result P and label L of the preliminary model for COVID-19 region segmentation in CT images are input into the auxiliary classification network to calculate the cross-entropy loss value R of the auxiliary classification network. The parameters of the auxiliary classification network are updated by calculating the discrimination result output by the auxiliary classification network and the classification loss of the label. The output result of the auxiliary classification network is fed back to the preliminary model for COVID-19 region segmentation in CT images to update the parameters of the preliminary model for COVID-19 region segmentation in CT images in reverse.

[0099] S5. Based on the auxiliary classification network constructed in step S4, the training dataset obtained in step S2 is used to train the preliminary model for COVID-19 region segmentation in CT images constructed in step S3, so as to obtain the COVID-19 region segmentation model in CT images.

[0100] In practice, the training process includes the following steps:

[0101] The cross-entropy loss, regularization loss, and edge loss between the prediction results of the preliminary model for COVID-19 region segmentation in CT images and the labels are calculated. Then, the loss function of the preliminary model for COVID-19 region segmentation in CT images is obtained by weighting the three types of losses.

[0102] In the training process of the COVID-19 region segmentation model for CT images, the difference between the segmented image generated by L1 distance minimization and the manually annotated segmented image is incorporated. The calculation formula is as follows: Where x is the data input to the local feature fusion network G for training, and P data(x) Let x be the distribution of x, and y be the segmented image generated by G(x,y); The probability distribution of the generated segmented image; the loss function of the local feature fusion network G is expressed as:

[0103]

[0104] In the formula L CGAN (D,G) represents the cross-entropy loss of the local feature fusion network, and L L1 (G) is the regularization loss of the local feature fusion network, EPE(G) is the edge loss of the local feature fusion network, and λ and η are set parameters;

[0105] The loss function of the auxiliary classification network D is:

[0106]

[0107] In the formula y i p represents the label corresponding to image i. i This represents the probability that image i is predicted to be of the positive class;

[0108] S6. Using the COVID-19 region segmentation model of the CT image obtained in step S5, complete the COVID-19 region segmentation of the actual CT image.

[0109] The segmentation results of the present invention will be described below with reference to an embodiment:

[0110] This method was experimentally tested on the COVID-19 and COVID-19SemiSeg lesion segmentation datasets. The experiments used the PyTorch framework, with MELoss as the loss function for LF-Net and cross-entropy loss as the loss function for the discriminant network. Stochastic Gradient Descent (SGD) optimizer was used for network weight optimization, with a learning rate of 0.001, momentum of 0.9, image size of 96×96, batch size of 2, and epochs of 100. The proposed method and classic segmentation networks were trained on COVID-19SemiSeg and COVID-19 datasets, respectively. After training, the networks were compared experimentally on the test set. Table 1 shows the experimental results of different classic networks on COVID-19SemiSeg.

[0111] Table 1. Schematic diagram of experimental results of different classical networks tested on COVID19 SemiSeg.

[0112] R2U-Net 0.665 0.551 0.788 DeepLabV3+ 0.568 0.440 0.672 Attention U-Net 0.677 0.557 0.823 R2AU-Net 0.602 0.485 0.760 U-Net++ 0.671 0.557 0.811 This invention 0.717 0.608 0.823

[0113] Table 2 shows a comparison of the test results of different classic networks on the COVID-19 dataset:

[0114] Table 2. Comparison of test results of different classic networks on the COVID-19 dataset.

[0115] R2U-Net 0.627 0.521 0.750 DeepLabV3+ 0.522 0.413 0.651 Attention U-Net 0.614 0.513 0.737 R2AU-Net 0.613 0.511 0.733 U-Net++ 0.625 0.519 0.750 This invention 0.652 0.550 0.760

[0116] As shown in Tables 1 and 2, the experimental results demonstrate that the proposed method achieves good segmentation performance on both lesion segmentation datasets. The data in Tables 1 and 2 also show that, compared to classic segmentation networks, the proposed method achieves the highest values ​​in Dsc, Iou, and Sen. accuracy, proving that the present invention can capture richer lesion contour information to determine lesion regions, capture more complete lesion features, and use an auxiliary classification network to improve the network's confidence in the labels, further narrowing the difference between the prediction results and the labels, and refining the prediction results.

[0117] like Figure 2 The diagram shown is a flowchart of the imaging method of the present invention: The imaging method disclosed in this invention, which includes the COVID-19 region segmentation method in CT images, specifically includes the following steps:

[0118] A. Obtain actual CT images of the lung region;

[0119] B. Using the aforementioned CT image COVID-19 region segmentation method, the COVID-19 region is segmented in the CT image obtained in step S1;

[0120] C. The region segmentation results obtained in step B are annotated and re-imagined on the CT image obtained in step A to obtain the final CT image of the lung region with the COVID-19 region segmentation results.

[0121] The imaging method provided by this invention can be applied to existing lung imaging CT scanners. After application, the lung imaging CT scanner acquires the original CT image of the actual lung region. After performing COVID-19 region segmentation and secondary imaging using the method of this invention, the final output is a lung region CT image with COVID-19 region segmentation results. This enables the application of lung imaging CT scanners to be more widespread and can directly provide auxiliary data to help doctors in subsequent work.

Claims

1. A method for segmenting COVID-19 regions in CT images, comprising the following steps: S1. Obtain a dataset of labeled CT images of COVID-19; S2. Perform image processing on the dataset obtained in step S1 to obtain the training dataset; S3. Construct a preliminary model for COVID-19 region segmentation in CT images based on a local feature fusion network; specifically including the following steps: The model includes an encoding layer and a decoding layer; Dilated convolutions are used in the encoding layer to expand the receptive field of the network, and depthwise separable convolutions are used to reduce the training parameters of the network. A triple hybrid domain attention mechanism is used in the encoding and decoding layers to promote the expression of lesion information and suppress the expression of non-lesion information. The decoding layer of the network consists of a shape network and deconvolution, where the shape network uses gated convolution and residual convolution to determine the complete lesion region by capturing the contour information of the lesion. The specific process of shape network to extract contour features is as follows: First, residual convolution is used to obtain the global features of the lesion, and then gated convolution is used to extract the edge contour information of the lesion and filter out non-contour information. In the upsampling process, firstly, a 1*1 convolution is performed on the four layers of features to unify the number of channels in the upsampled feature map, preparing for subsequent feature fusion. Then, the sampled features of the last layer of the model are input into the residual block, and the output of the residual block and the first layer upsampling are simultaneously fed into the gated convolution. Finally, the feature map output from the gated convolution is processed again by the residual convolution, and the result is added to the second layer upsampling of the gated convolution. This process is repeated until the last layer of the upsampled feature map is reached. The last layer of features in the upsampling process in the model is subjected to a final gated convolution with the previous layer feature map to obtain the contour features of the lesion; the contour features of the lesion extracted by the shape network are fused with the lesion features to obtain a more accurate segmentation result than a single lesion feature. S4. Construct an auxiliary classification network to refine the segmentation results of the preliminary model for COVID-19 region segmentation in CT images; specifically including the following steps: The auxiliary classification network is constructed using the following steps: Perform a 3x3 convolution operation on the input image to obtain lesion features. ; lesion characteristics The channel attention module is used for processing to identify lesion features. Perform 3x3 average pooling to obtain average pooling lesion features. Then, channel attention maps are generated in the shared network to obtain channel attention features. Channel attention features Nonlinear processing was performed to obtain lesion activation characteristics. , ,in It is the Sigmoid non-linear activation function; lesion characteristics The input location attention module uses three 3x3 convolutions followed by a reshape operation to obtain the first lesion feature. Second lesion characteristics and third lesion characteristics The characteristics of the first lesion Second lesion characteristics Perform matrix multiplication to obtain multiplicative lesion features. ; Multiplication lesion characteristics and third lesion characteristics Perform matrix multiplication to obtain the location attention lesion features. ; Activation characteristics of lesions and location attention lesion characteristics Perform a 3-dimensional multiplication operation and apply an activation function to obtain the final output lesion features. ; The classification loss is calculated by using the discrimination results predicted by the iterative auxiliary classification network, and the network parameters of the discriminant network and the preliminary model for COVID-19 region segmentation in CT images are updated by back gradient propagation; specifically, the following steps are included: The prediction result P and label L of the preliminary model for COVID-19 region segmentation in CT images are input into the auxiliary classification network to calculate the cross-entropy loss value R of the auxiliary classification network. The parameters of the auxiliary classification network are updated by calculating the discrimination result output by the auxiliary classification network and the classification loss of the label. The output result of the auxiliary classification network is fed back to the preliminary model for COVID-19 region segmentation in CT images to update the parameters of the preliminary model for COVID-19 region segmentation in CT images in reverse. S5. Based on the auxiliary classification network constructed in step S4, and using the training dataset obtained in step S2, train the preliminary model for COVID-19 region segmentation in CT images constructed in step S3 to obtain the COVID-19 region segmentation model in CT images; specifically including the following steps: The cross-entropy loss, regularization loss, and edge loss between the prediction results of the preliminary model for COVID-19 region segmentation in CT images and the labels are calculated. Then, the loss function of the preliminary model for COVID-19 region segmentation in CT images is obtained by weighting the three types of losses. In the training process of the COVID-19 region segmentation model for CT images, the difference between the segmented image generated by minimizing the L1 distance and the manually annotated segmented image is incorporated. The calculation formula is as follows: ,in Input to the local feature fusion network The training data, for The distribution for The generated segmented image; The probability distribution of the generated segmented image; local feature fusion network The loss function is expressed as In the formula For the cross-entropy loss of the local feature fusion network, For the regularization loss of the local feature fusion network, The edge loss of the local feature fusion network, and The parameters are set; The loss function of the auxiliary classification network D is: In the formula Image The corresponding tags Image The probability of predicting it as positive; S6. Using the COVID-19 region segmentation model of the CT image obtained in step S5, complete the COVID-19 region segmentation of the actual CT image.

2. The method for segmenting COVID-19 regions in CT images according to claim 1, characterized in that... Step S2 involves image processing of the dataset obtained in step S1 to obtain the training dataset, specifically including the following steps: Binarize the image to obtain a grayscale image. ; Kernel for setting morphological filtering for ,in For refactoring functions; Black hat filtering is applied to the grayscale image to obtain the lesion contour map. for ,in For conversion functions; Outline of the lesion Threshold repair is performed to obtain the preprocessed image; The obtained preprocessed image is converted into a color image, thus completing the preprocessing of an image; Each image in the acquired dataset is preprocessed to obtain the training dataset.

3. The method for segmenting COVID-19 regions in CT images according to claim 2, characterized in that... The aforementioned triple hybrid domain attention mechanism specifically includes the following steps: Dilated convolution and depthwise separable convolution are performed sequentially on the input image to extract features of the lesion region. ; Characteristics of the lesion area Perform 3x3 average pooling to obtain average pooling features. for ; This indicates the average pooling operation. This indicates that pooling uses a 3x3 convolution kernel; Channel attention maps are generated in the local feature fusion network to obtain channel attention features. , ,,in Operation functions for shared networks; Channel attention features Perform nonlinear processing to obtain the features of the activated region. , ,in Sigmoid is a non-linear activation function; Activate region features Perform 7x7 average pooling and 7x7 max pooling respectively to obtain the activation region features of the average pooling. and max pooling activation region features ; Average pooling activates region features and max pooling activation region features Channel splicing is performed in the third dimension to obtain the spliced ​​and blended region features. ; Features of splicing and fusion areas By performing void space pyramid pooling and nonlinear processing, the region features are obtained. for ,in For the pyramid pooling treatment of empty spaces, It is the Sigmoid non-linear activation function; regional characteristics The input location attention module performs three 3x3 convolutions followed by a reshape operation to obtain the features of the first region. Second region characteristics and third region features ; Features of the first region Second region features Perform matrix multiplication to obtain the features of the multiplication region. ; Features of the third region Features of the multiplication region Perform matrix multiplication again to obtain the location attention region features. ; Activate region features Regional characteristics and location attention region features By merging and fusing, the characteristics of the output lesion area are obtained. for ,in This is to perform an element-wise stacking and merging operation on several features.

4. An imaging method comprising the COVID-19 region segmentation method for CT images according to any one of claims 1 to 3, specifically comprising the following steps: A. Obtain actual CT images of the lung region; B. Using the CT image COVID-19 region segmentation method according to any one of claims 1 to 3, the COVID-19 region is segmented in the CT image obtained in step S1; C. The region segmentation results obtained in step B are annotated and re-imagined on the CT image obtained in step A to obtain the final CT image of the lung region with the COVID-19 region segmentation results.

Images