Novel coronavirus lung CT image infection area segmentation method and system

Abstract

The invention relates to a novel coronavirus lung CT image infection area segmentation method and system, and the method comprises the steps: carrying out the processing of the boundary information of a CT image, obtaining a prediction boundary image, integrating the prediction boundary image into two branches outputted by an encoder, fusing the enhanced semantic information with multi-scale feature representation, and completing the segmentation of an infection area. Constructing an infection area segmentation network model, and training and optimizing the infection area segmentation network model by using a loss function to obtain an optimal infection area segmentation network model; and finally, inputting a CT image to be processed into the optimal infection area segmentation network model, and outputting a final segmentation result to complete segmentation of the novel coronavirus pulmonary infection area. Compared with the prior art, the accuracy of an infection area segmentation result can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for segmenting infected areas of novel coronavirus lung CT images. Background Art

[0002] The novel coronavirus 2019 (COVID-19) continues to spread around the world, triggering a global health crisis. Manual analysis of CT images by doctors is a time-consuming and error-prone process, and accurate diagnosis requires more experienced radiologists. Computer-aided diagnosis (CAD) can detect cases quickly and accurately, thereby reducing the pressure on radiologists. Therefore, an advanced computer-aided diagnosis method has important practical value for the early detection, diagnosis and treatment of novel coronavirus lung infection areas.

[0003] Currently, the image segmentation of computed tomography (CT) scans of the new coronavirus lungs faces many challenges. CT images often have redundant backgrounds and an imbalance between infected and healthy tissues, which can easily cause the model to be confused when distinguishing infected areas, resulting in model misjudgment. Secondly, the low contrast between infected and normal tissues makes it difficult for the model to distinguish boundaries. Thirdly, lung infections usually have variable morphologies, small and irregular shapes, and uneven distribution. This leads to incomplete and insufficient segmentation.

[0004] In the existing lung CT infection segmentation network, U-net++ reduces the semantic gap between the encoder and decoder by introducing dense jump connections, but these jump connections are mainly used for information transmission and cannot effectively improve the ability to extract feature information; PDAt-Unet adds an attention gate to reduce feature redundancy, and the combination of multi-scale convolution and hybrid attention mechanism can more effectively reduce feature redundancy; Inf-net uses implicit reverse attention and explicit edge attention to focus on the segmentation that is more suitable for the boundary of the lung infection area, but excessive noise cannot be removed during the upsampling process of the model; the BSNet model captures the infection area from the perspective of semantic relationship and boundary guidance. The common model structure will have the problem of boundary information loss when extracting feature information.

[0005] The invention patent with publication number CN114332133A discloses a method for segmenting the infected area of ​​pneumonia CT images based on improved CE-Net. This method adds an attention mechanism SE module in the encoding stage to introduce global context information, enhance the receptive field in the feature extraction stage, and increase the weight of the target-related feature channel, thereby improving the segmentation ability of small targets. The patent also introduces a feature aggregation module, which uses a bilinear interpolation method to fuse image features at different levels to obtain a more discriminative expression, further improving the segmentation accuracy of the network. However, the boundary information is not processed emphatically, and the problem of blurred boundaries still exists.

[0006] The invention patent with publication number CN117522904A discloses a dual-path image segmentation method based on contour enhancement. This method adds an encoder path based on U-Net, one of which uses the attention multi-scale feature fusion module to extract the features of the original CT image, and the other encoder path uses the contour feature extraction module to extract the contour features in the CT contour image to compensate for the boundary information lost during the downsampling process. However, the patent does not mention the use of the semantic enhancement module and the adjustment of the loss function, and there is still a problem of insufficient accuracy when segmenting small targets, irregular shapes and CT images with blurred boundaries. Summary of the invention

[0007] In order to overcome the defects of the above-mentioned prior art, the present invention provides a method and system for segmenting the infected area of ​​the new coronavirus lung CT image.

[0008] The present invention can be implemented by the following technical solutions:

[0009] The present invention provides a method for segmenting the infected area of ​​a new coronavirus lung CT image, which processes the boundary information of the CT image to obtain a predicted boundary map, integrates the predicted boundary map into two branches of the encoder output, and uses enhanced semantic information to fuse multi-scale feature representation to jointly complete the segmentation of the infected area. The method includes the following steps:

[0010] S1, preprocessing the CT image to obtain a data set;

[0011] S2, build the infected area segmentation network model;

[0012] S3, using the data set and combining the loss function to train and optimize the segmentation model to obtain an optimized network model;

[0013] S4, input the CT image to be processed into the optimized segmentation network model, output the final segmentation result, and complete the segmentation of the new coronavirus lung infection area. As a preferred technical solution, the specific operation of preprocessing the CT image in S1 is: processing the CT image size to 352×352 pixels, normalizing the CT image, and enhancing the image data by randomly rotating 45 degrees to obtain the preprocessed data.

[0014] As a preferred technical solution, the segmentation network model (Semantic Enhancement-Boundary Guidance-Net, SEBG-Net) constructed in S2 includes a backbone network, an image boundary guidance module (Image Boundary Guidance, IBG), a multi-scale feature selection module (Multi-scale Feature Selection, MFS) and a semantic enhancement module (Semantic Enhancement, SE);

[0015] The backbone network includes five layers of encoders Conv1-5 and four layers of decoders Conv6-9, wherein the encoder is used to extract semantic features, and the decoder uses bilinear difference to gradually upsample the multi-scale feature representation and semantic information to fuse, and finally output the segmentation result;

[0016] The image boundary guidance module is used to extract the boundary feature information of the CT image and output the predicted boundary map, and includes two branches: the first branch is used to extract the boundary feature information of the image, which is a cascade of a 5×5 convolution, a 3×3 convolution and two 1×1 convolutions; the second branch is used for spatial compression, which is a cascade of a 1×1 convolution, a Relu activation function and a Sigmoid activation function;

[0017] The multi-scale feature selection module extracts and fuses feature information through a hybrid attention mechanism, outputs a multi-scale feature representation, and includes four branches, the first branch is two 3×3 convolution cascades, the second branch is a 5×5 convolution and a 3×3 convolution cascade, the third branch is a 7×7 convolution and a 3×3 convolution cascade, and the fourth branch uses a residual connection;

[0018] The semantic enhancement module is used to enhance the semantic features extracted by the encoder, construct the relationship between the complex background and the infected area and output it as semantic information, including two branches, the first branch is two 3×3 convolution cascades, and the second branch is average pooling, Sigmoid function and a 1×1 convolution cascade;

[0019] The outputs of Conv1-4 encoders in the backbone network are connected to the input of the multi-scale feature selection module, the outputs of Conv3-5 are connected to the input of the semantic enhancement module, the output of the image boundary guidance module is connected to the outputs of Conv3 and Conv4 encoders, and the output of the multi-scale feature selection module and the output of the semantic enhancement module are connected to the inputs of Conv6-9 decoders.

[0020] As a preferred technical solution, the structure of each encoder and decoder layer in the backbone network is a cascade of two 3×3 convolutions, a batch normalization module and a Relu activation function.

[0021] As a preferred technical solution, residual connections are introduced between 3×3 convolutions in the backbone network. The specific mathematical expressions of the encoder and decoder in the backbone network are:

[0022] Encode=Conv 3×3 (Conv 3×3 (I in ))+Conv 1×1 (I in )

[0023] Decode=Conv 3×3 (Conv 3×3 (I in ))+Conv 1×1 (I in )

[0024] Where Encode is the encoder convolution block, Decode is the decoder convolution block, I in ∈R C×H×W For the input image, Conv 1×1 is a 1×1 convolution kernel, Conv 3×3 is a 3×3 convolution kernel.

[0025] As a preferred technical solution, the mathematical expression of the image boundary guidance module is:

[0026]

[0027] In the formula, f IBG is the image boundary feature, I in ∈R C×H×W is the input image, Relu is the activation function, Sig is the Sigmoid activation function, Conv i×j Represents i×j convolution kernel (such as Conv 1×1 is a 1×1 convolution kernel), It is an element-wise multiplication operation.

[0028] As a preferred technical solution, the first three branches of the multi-scale feature selection module extract and fuse feature information through a hybrid attention mechanism. Then the fused feature information is element-wise added with the fourth branch to generate new feature information as a decoder feature supplement. The specific mathematical expression is:

[0029] f AT1 =AT(Conv 3×3 (Conv 3×3 (I in )),Conv 3×3 (Conv 5×5 (I in )))

[0030] f AT2 =AT(f AT1 ,Conv 3×3 (Conv 7×7 (I in )))

[0031] f MFS =f AT2 +I in

[0032] In the formula, f AT1 、f AT2 For MFS is a multi-scale feature representation, I in ∈R C×H×W is the input image, and AT is the hybrid attention mechanism.

[0033] As a preferred technical solution, the specific mathematical expression of the semantic enhancement module is:

[0034] f Cat =Cat(UP(Conv3),UP(Conv4),UP(Conv5))

[0035]

[0036] In the formula, f Cat is the fusion feature, f SE is a semantic feature, Cat is a fusion operation, UP is a bilinear difference operation, is an element-wise multiplication operation, Sig is the Sigmoid activation function, Conv i×j Represents i×j convolution kernel (such as Conv 3×3 Represented as a 3×3 convolution kernel).

[0037] As a preferred technical solution, the loss function in S3 is the weighted sum of the binary cross entropy loss function and the dice similarity coefficient loss function, and its specific mathematical expression is:

[0038]

[0039] In the formula, is the loss function, is the binary cross entropy loss function, is the dice similarity coefficient loss function, α and β are weight coefficients, G i is the gold standard value, P i is the predicted value, and N is the number of samples.

[0040] According to another aspect of the present invention, a system for segmenting infected areas of a CT image of a new coronavirus lung is provided. The system works by applying the method for segmenting infected areas of a CT image of a new coronavirus lung as described above. The system includes an image preprocessing module, a model building module, and an image segmentation processing module.

[0041] The image preprocessing module is used to preprocess the CT image to obtain preprocessed data for subsequent use; the model building module is used to construct an infected area segmentation network model, and use the data set and the loss function to train and optimize the parameters of the segmentation network model to obtain the optimal segmentation network model; the segmentation processing module is used to input the CT image to be processed into the optimized segmentation network model, output the segmentation result, and complete the segmentation task of the new coronavirus lung infection area.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention proposes a semantically enhanced novel coronavirus lung infection area segmentation scheme based on boundary guidance. By constructing an infection area segmentation network model, the boundary information of the CT image is processed by the model to obtain a predicted boundary map, and the predicted boundary map is integrated into the two branches of the encoder output. The infected area segmentation is completed by using enhanced semantic information fusion multi-scale feature representation, thereby significantly improving the boundary detail processing, feature extraction capability and segmentation accuracy, effectively improving the recognition rate of small infection areas, reducing mis-segmentation, and overall improving the accuracy and reliability of the segmentation results.

[0044] 2. The method of the present invention constructs and uses an infected area segmentation network model, which incorporates an image boundary guidance module for extracting image boundary feature information of CT images and outputting a boundary prediction map. The module includes two branches. The first branch is used to extract image boundary feature information, which is a cascade of a 5×5 convolution, a 3×3 convolution, and two 1×1 convolutions. The second branch is used for spatial compression, which is a cascade of a 1×1 convolution, a Relu activation function, and a Sigmoid activation function. The use of this module helps to obtain rich boundary information to improve segmentation accuracy, obtain image boundary feature information from CT images and integrate it into the network, and enhance the ability to identify the boundaries of infected areas.

[0045] 3. The present invention constructs and uses an infected area segmentation network model, which incorporates a semantic enhancement module for enhancing the semantic features extracted by the encoder, constructing the relationship between the complex background and the infected area and outputting corresponding semantic information. The module comprises two branches: the first branch is a cascade of two 3×3 convolutions, and the second branch is a cascade of average pooling, a Sigmoid function and a 1×1 convolution. Through the design of the semantic enhancement module, spatial feature information and position feature information are extracted and fused to construct the connection between different infected areas, and the distinction between the complex background and the infected area can also be strengthened. The connection between the semantic information is used to help locate the infected area, enhance the distinction between the infected area and the background, and improve the recognition ability of the segmented infected area.

[0046] 4. The present invention constructs and uses an infected area segmentation network model, which incorporates a multi-scale feature selection module. The model extracts and fuses feature information through a hybrid attention mechanism, and outputs a multi-scale feature representation. The module includes four branches: the first branch is a cascade of two 3×3 convolutions; the second branch is a cascade of a 5×5 convolution and a 3×3 convolution; the third branch is a cascade of a 7×7 convolution and a 3×3 convolution; the fourth branch uses a residual connection. Through the design of the multi-scale feature selection module, the connection between features is strengthened, and the hybrid attention mechanism guides the network to pay attention to spatial information and location information, and perform information fusion. In addition, the residual connection reduces information loss, thereby improving the effectiveness of information fusion.

[0047] 5. The present invention uses a loss function to train and optimize the infected area segmentation network, and the loss function used is composed of a binary cross entropy loss function and a dice similarity coefficient loss function. Among them, the binary cross entropy loss function enables the model to accurately predict the probability that each pixel belongs to the lesion area or the normal area; the dice similarity coefficient loss function uses pixel-level accuracy to focus on the overall contour of the lesion area. Reasonably allocate the focus of the model, and by weighting the weights of the two, the segmentation network model pays more attention to the edge features of the infected area, so that the segmentation of the lesion infected area is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of the steps of a method for segmenting the infected area of ​​a new coronavirus lung CT image in the present invention;

[0049] Figure 2 It is a schematic diagram of the SEBG-Net structure in the present invention;

[0050] Figure 3 This is a flow chart of a method for segmenting infected areas of a novel coronavirus lung CT image in the present invention;

[0051] Figure 4 Schematic diagram of the structure of the image boundary guidance module in the present invention;

[0052] Figure 5 is a structural schematic diagram of the semantic enhancement module in the present invention;

[0053] Figure 6 is a schematic diagram of the structure of the multi-scale feature selection module in the present invention;

[0054] Figure 7 1 is a comparison diagram of the segmentation results of different models in the embodiment and the example of this method. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0056] Example

[0057] This embodiment applies a method for segmenting the infected area of ​​a new coronavirus lung CT image, the steps are as follows: Figure 1 As shown, including:

[0058] S1, preprocessing the CT image to obtain a preprocessed data set;

[0059] S2. Build and use the infected area segmentation network model, input the data set into the model to output the segmentation result;

[0060] S3, using the loss function to train and optimize the segmentation network model to obtain the optimal segmentation network model;

[0061] S4. Input the CT image to be processed into the optimized segmentation network model, output the segmentation result, and complete the segmentation of the new coronavirus lung infection area.

[0062] Among them, the segmentation network model SEBG-Net is constructed and used, and its structure is as follows Figure 2 As shown in Figure 1, it includes a backbone network, an image boundary guidance module, a multi-scale feature selection module, and a semantic enhancement module. After adding the specific application of the infection segmentation network model, the flowchart of the new coronavirus lung infection area segmentation method is as follows: Figure 3 As shown, the process is:

[0063] Input CT images and perform preprocessing, build and use the infected area segmentation network model SEBG-Net, and initialize its network parameters. Subsequently, extract the image boundary feature information to obtain the boundary prediction map; then input the data set into the encoder, fuse the boundary prediction map with the encoder's processing results and input it into the multi-scale feature selection module to obtain a multi-scale feature representation; then use the semantic enhancement module to extract semantic information from the encoder's processing results; finally, in the decoder, the multi-scale feature representation obtained in S3 and the semantic information obtained in S4 are fused through step-by-step upsampling to obtain the segmentation result.

[0064] The segmentation model is optimized through the loss function to determine whether the current number of iterations is less than the predetermined number of iterations. If so, the system returns to the initialization parameter step until the optimal segmentation result is obtained through iteration. The segmented image is output to complete the segmentation of the new coronavirus lung infection area.

[0065] In this segmentation network model, the outputs of Conv1~4 encoders in the backbone network are connected to the input of the multi-scale feature selection module, the outputs of Conv3~5 are connected to the input of the semantic enhancement module, the output of the image boundary guidance module is connected to the outputs of Conv3 and Conv4 encoders, and the outputs of the multi-scale feature selection module and the outputs of the semantic enhancement module are connected to the inputs of Conv6~9 decoders.

[0066] In this segmentation network model, the backbone network consists of five layers of encoders Conv1~5 and four layers of decoders Conv6~9; the structure of each layer of encoder and decoder is a cascade of two 3×3 convolutions, batch normalization modules and Relu activation functions.

[0067] In the backbone network, residual connection is introduced in the 3×3 convolutional layer, and its mathematical expression is:

[0068] Encode=Conv 3×3 (Conv 3×3 (I in ))+Conv 1×1 (I in )

[0069] Decode=Conv 3×3 (Conv3×3 (I in ))+Conv 1×1 (I in )

[0070] Where Encode is the encoder convolution block, Decode is the encoder convolution block, I in ∈R C×H×W For the input image, Conv i×j Represents i×j convolution kernel (such as Conv 3×3 is a 3×3 convolution kernel).

[0071] In this segmentation network model, the structure of the image boundary guidance module is as follows: Figure 4 As shown in the figure, it contains two branches. The first branch is used to extract the image boundary feature information, which is a cascade of a 5×5 convolution, a 3×3 convolution and two 1×1 convolutions. The second branch is used for spatial compression, which is a cascade of a 1×1 convolution, a Relu activation function and a Sigmoid activation function. The mathematical expression is:

[0072]

[0073] In the formula, f IBG is the image boundary feature, I in ∈R C×H×W is the input image, Relu and Sigmoid are activation functions, Conv i×j Represented as i×j convolution kernel (such as Conv 1×1 is a 1×1 convolution kernel), It is an element-wise multiplication operation.

[0074] In this segmentation network model, the structure of the semantic enhancement module is as follows: Figure 5 As shown in the figure, it includes two branches. The first branch is two 3×3 convolution cascades, and the second branch is average pooling, Sigmoid function and a 1×1 convolution cascade. Their mathematical expressions are:

[0075] f Cat =Cat(UP(Conv3),UP(Conv4),UP(Conv5))

[0076]

[0077] In the formula, f Cat represents feature fusion, f SE Indicates semantic features, Cat is a fusion operation, UP is a bilinear difference operation, is an element-wise multiplication operation, Sig is the Sigmoid activation function, Conv 3×3 is a 3×3 convolution kernel, Conv 1×1is a 1×1 convolution kernel.

[0078] In this segmentation network model, the structure of the multi-scale feature selection module is as follows: Figure 6 As shown in the figure, it contains four branches: the first branch is two 3×3 convolution cascades; the second branch is a 5×5 convolution and a 3×3 convolution cascade; the third branch is a 7×7 convolution and a 3×3 convolution cascade; the fourth branch uses residual connection. Among them, the first three branches extract and fuse feature information through the hybrid attention mechanism, and then the fused feature information is element-wise added with the fourth branch to generate new feature information as the decoder feature supplement. Its mathematical expression is:

[0079] f AT1 =AT(Conv 3×3 (Conv 3×3 (I in )),Conv 3×3 (Conv 5×5 (I in )))

[0080] f AT2 =AT(f AT1 ,Conv 3×3 (Conv 7×7 (I in )))

[0081] f MFS =f AT2 +I in

[0082] In the formula, f AT1 、f AT2 For MFS is a multi-scale feature representation, I in ∈R C×H×W is the input image, AT is the hybrid attention mechanism, Conv i×j Represents i×j convolution kernel (such as Conv 1×1 is a 1×1 convolution kernel).

[0083] This embodiment applies the above solution, and is specifically implemented as follows:

[0084] The novel coronavirus lung CT images were input for preprocessing. All images were resized to 352×352 pixels and normalized to (0,1) before training. Image enhancement was then performed and the image data was enhanced by random rotation of 45 degrees to obtain the preprocessed data set, which was then divided into training set, test set, and validation set in a ratio of 8:1:1.

[0085] Construct a segmentation network SEBG-Net, build an encoder and decoder structure, and add jump connections to realize the information flow between the encoder and decoder.

[0086] Construct an image boundary guidance module IBG, input the image, extract the image boundary feature information through IBG, and then obtain the boundary prediction map S e , S e They are introduced into Conv3 and Conv4 branches respectively to enhance the network's attention to the boundaries of the infected area.

[0087] The dataset is input into the encoder, a multi-scale feature selection module is constructed, the boundary prediction map is fused with the processing result of the encoder, and then input into the multi-scale feature selection module to obtain a multi-scale feature representation. The multi-scale feature selection module MFS guides the network to learn multi-scale semantic features and extract more useful features; this module uses a hybrid attention mechanism to focus on key feature information, guides the network to focus on spatial information and location information and perform information fusion, and strengthens the connection between features. In addition, the residual connection reduces information loss, so that the model can perform information fusion more effectively.

[0088] Construct a semantic enhancement module. The CT image contains spatial feature information and position feature information. The semantic enhancement module extracts semantic information from the processing results of the encoder, and extracts feature information for fusion. Construct the connection between different infected areas, strengthen the distinction between complex background and infected areas, and use the connection between semantic information to help locate the infected area. Strengthening the distinction between infected areas and background will help network recognition and segmentation.

[0089] The multi-scale feature selection module MFS combines the semantic feature information obtained in multiple branches with the decoder Conv6~9. The semantic enhancement module SE provides the enhanced semantic information to the decoder. The decoder uses the bilinear difference method to fuse the feature information of different sizes in the encoding process, thereby achieving the purpose of reconstructing the infected area. In the decoder, the multi-scale feature representation obtained in S3 and the semantic information obtained in S4 are gradually upsampled and fused using the bilinear difference method to obtain the segmentation result. The segmentation result is evaluated by the loss function to obtain the optimized segmentation result, thereby completing the segmentation of the infected area.

[0090] The loss function used in this embodiment is composed of a binary cross entropy loss function and a dice similarity coefficient loss function. The binary cross entropy loss function is used to predict the probability that each pixel belongs to a lesion area or a normal area to improve the accuracy of image segmentation; the dice similarity coefficient loss function is used to focus on the boundary information of the lesion area and weight the two. The mathematical expression of the loss function is:

[0091]

[0092] In the formula, is the loss function, is the binary cross entropy loss function, is the dice similarity coefficient loss function, α and β are weight coefficients, G i is the gold standard value, P i is the predicted value, and N is the number of samples.

[0093] To verify the effectiveness of this solution, this example performs infection area segmentation of the novel coronavirus lung CT image, and evaluates the performance and generalization ability of the segmentation model SEBG-Net of this solution using three datasets, including the MedSeg Covid Dataset 1 dataset, the COVID-19-CT-Seg dataset, and the MedSeg CovidDataset 2 dataset:

[0094] The MedSeg Covid Dataset 1 dataset contains 100 CT images of 40 cases, classified by radiologists using three labels: ground glass, consolidation, and pleural effusion.

[0095] The COVID-19-CT-Seg dataset contains 20 cases and a total of 3520 CT images. The gold standard is that radiologists use ITK-SNAP to manually annotate axial images layer by layer.

[0096] The MedSeg Covid Dataset 2 dataset contains 9 cases and 829 CT images in total. Among them, 373 images were assessed as positive by radiologists. There are three labels in the image, which set ground glass and lung consolidation to 1 and pleural effusion to 0.

[0097] In this embodiment, the main evaluation indicators are sensitivity (Sensitivity, Sens), specificity (Spec), precision (Precision, Prec) and Dice similarity coefficient (Dice), and the expressions are:

[0098]

[0099] Where TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

[0100] The data set used in this embodiment is composed of the above three data sets, with a total of 69 cases and 4449 CT images, and the performance and generalization ability of SEBG-Net are evaluated and verified. In addition, under the same experimental conditions and the same experimental data set, the segmentation network model of this scheme is also compared with the current mainstream models such as U-net, U-net++, Inf-Net, BSNet and PAtt-Unet for segmentation performance, as shown in Table 1. It can be seen from the table that the sensitivity, specificity, accuracy and Dice similarity coefficient of this segmentation network reached 67.52%, 99.38%, 65.15% and 64.86% respectively. Compared with the U-net model, the sensitivity, accuracy and Dice similarity coefficient were improved by 9.8%, 0.49% and 5.74% respectively.

[0101] The experimental results show that this scheme is significantly better than other models in terms of sensitivity, accuracy and Dice similarity coefficient, and is basically consistent with other models in terms of specificity. The comparative experiment shows that the segmentation network model SEBG-Net of this scheme can effectively segment the new coronavirus lung infection area.

[0102] Table 1 Performance comparison of different models

[0103]

[0104] In addition, 6 CT images are randomly selected from the test set and the segmentation effect is demonstrated, such as Figure 7 shown. Figure 7 In the figure, (a) is the original CT image of the new coronavirus lung, (b) is the gold standard image, and (c) to (h) are the segmentation results of the U-net, U-net++, Inf-Net, BSNet, PAtt-Unet, and SEBG-Net models of this scheme. Figure 7 It can be seen that the segmentation network model of this scheme can effectively segment small infected areas, which is closest to the gold standard. The segmentation results of the other five models are incomplete, the boundaries are blurred, or the segmentation errors are wrong. This network can completely segment the boundaries of the infected area, while the other five models are incomplete and the boundaries are blurred.

[0105] In summary, this scheme proposes a semantically enhanced novel coronavirus lung CT image infection segmentation method based on boundary guidance. The image boundary guidance module uses the boundary features of the image and introduces it between the encoder and the decoder, thereby enhancing the network's ability to recognize the boundaries of the infected area. The semantic enhancement module uses the feature information from the encoder to establish the relationship between local information and global information, highlighting the infected area in a complex background. The multi-scale feature selection module uses convolution kernels of different scales to flexibly select features, and combines the hybrid attention mechanism to guide the network to recognize and fuse feature information. The experimental results show that the model proposed in this scheme can effectively extract the boundary features of the infected area, is particularly sensitive to small lesion areas, can effectively segment small infected areas, and achieves good segmentation results.

[0106] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for segmenting infected areas of new coronavirus lung CT images, characterized in that: The method processes the boundary information of the CT image to obtain a predicted boundary map, then integrates the predicted boundary map into the two branches of the encoder output, and uses enhanced semantic information to fuse multi-scale feature representation to jointly complete the infected area segmentation. The method includes the following steps: S1. preprocessing the CT image to obtain a preprocessed data set; S2, build the infected area segmentation network model; S3. Using the data set and combining the loss function to train and optimize the infected area segmentation network model, so as to obtain the optimal infected area segmentation network model; S4. Input the CT image to be processed into the optimal infected area segmentation network model, output the final segmentation result, and complete the segmentation of the new coronavirus lung infection area.

2. According to claim 1, a method for segmenting infected areas of a new coronavirus lung CT image is characterized in that: The specific operation of preprocessing the CT image in S1 is: processing the CT image size to 352×352 pixels, then normalizing the CT image, and enhancing the image data by randomly rotating it by 45 degrees to obtain a preprocessed data set.

3. According to claim 1, a method for segmenting infected areas of a new coronavirus lung CT image is characterized in that: The infected area segmentation network model constructed in S2 includes a backbone network, an image boundary guidance module, a multi-scale feature selection module and a semantic enhancement module; The backbone network includes five layers of encoders Conv1 to 5 and four layers of decoders Conv6 to 9, wherein the encoder is used to extract semantic features, and the decoder uses bilinear difference to gradually upsample and fuse multi-scale feature representation and semantic information, and outputs the segmentation result; The image boundary guidance module is used to extract the image boundary feature information of the CT image and output the boundary prediction map, and includes two branches: the first branch is used to extract the image boundary feature information, and a 5×5 convolution, a 3×3 convolution and two 1×1 convolutions are cascaded; the second branch is used for spatial compression, and a 1×1 convolution, a Relu activation function and a Sigmoid activation function are cascaded; The multi-scale feature selection module extracts and fuses feature information through a hybrid attention mechanism, outputs a multi-scale feature representation, and includes four branches. The first branch is a cascade of two 3×3 convolutions, the second branch uses a 5×5 convolution and a 3×3 convolution cascade, and the third branch uses a 7×7 convolution and a 3×3 convolution cascade; the fourth branch uses a residual connection; The semantic enhancement module is used to enhance the semantic features extracted by the encoder, construct the relationship between the complex background and the infected area and output it as semantic information, including two branches: the first branch is two 3×3 convolution cascades; the second branch is average pooling, Sigmoid function and a 1×1 convolution cascade; The outputs of Conv1-4 encoders in the backbone network are connected to the input of the multi-scale feature selection module, the outputs of Conv3-5 are connected to the input of the semantic enhancement module, the output of the image boundary guidance module is connected to the outputs of Conv3 and Conv4 encoders, and the output of the multi-scale feature selection module and the output of the semantic enhancement module are connected to the inputs of Conv6-9 decoders.

4. A method for segmenting infected areas of a new coronavirus lung CT image according to claim 3, characterized in that: The structure of each layer of encoder and decoder in the backbone network is a cascade of two 3×3 convolutions, a batch normalization module and a Relu activation function.

5. A method for segmenting infected areas of a new coronavirus lung CT image according to claim 4, characterized in that: In the backbone network, residual connections are introduced between 3×3 convolutions. The mathematical expressions of the encoder and decoder in the backbone network are: Encode=Conv 3×3 (Conv 3×3 (I in ))+Conv 1×1 (I in ) Decode=Conv 3×3 (Conv 3×3 (I in ))+Conv 1×1 (I in ) Where Encode is the encoder convolution block, Decode is the decoder convolution block, I in ∈R C×H×W For the input image, Conv 1×1 is a 1×1 convolution kernel, Conv 3×3 is a 3×3 convolution kernel.

6. A method for segmenting infected areas of a novel coronavirus lung CT image according to claim 3, characterized in that: The mathematical expression of the image boundary guidance module is: In the formula, f IBG is the image boundary feature, I in ∈R C×H×W is the input image, Relu is the activation function, Sig is the Sigmoid activation function, Conv 1×1 is a 1×1 convolution kernel, Conv 3×3 is a 3×3 convolution kernel, Conv 5×5 is a 5×5 convolution kernel, It is an element-wise multiplication operation.

7. According to claim 3, a method for segmenting infected areas of a new coronavirus lung CT image is characterized in that: In the multi-scale feature selection module, the first three branches extract and fuse feature information through a hybrid attention mechanism, and finally the fused feature information is element-wise added with the fourth branch to generate new feature information as a decoder feature supplement. The mathematical expression is: f AT1 =AT(Conv 3×3 (Conv 3×3 (I in )),Conv 3×3 (Conv 5×5 (I in ))) f AT2 =AT(f AT1 ,Conv 3×3 (Conv 7×7 (I in ))) f MFS =f AT2 +I in In the formula, f AT1 、f AT2 For MFS is a multi-scale feature representation, I in ∈R C×H×W is the input image, AT represents the hybrid attention mechanism, Conv 3×3 is a 3×3 convolution kernel, Conv 5×5 is a 5×5 convolution kernel, Conv 7×7 is a 7×7 convolution kernel.

8. A method for segmenting infected areas of a novel coronavirus lung CT image according to claim 3, characterized in that: The specific mathematical expression of the semantic enhancement module is: f Cat =Cat(UP(Conv3),UP(Conv4),UP(Conv5)) In the formula, f Cat is the fusion feature, f SE is a semantic feature, Cat is a fusion operation, UP is a bilinear difference operation, Conv 3×3 is a 3×3 convolution kernel, Conv 1×1 is a 1×1 convolution kernel, It is an element-level multiplication operation, and Sig is the Sigmoid activation function.

9. A method for segmenting infected areas of a new coronavirus lung CT image according to claim 1, characterized in that: The loss function in S3 is the weighted sum of the binary cross entropy loss function and the dice similarity coefficient loss function, and the mathematical expression is: In the formula, is the loss function, is the binary cross entropy loss function, is the dice similarity coefficient loss function, α and β are weight coefficients, G i is the gold standard value, P i is the predicted value, and N is the number of samples.

10. A novel coronavirus lung CT image infection area segmentation system, characterized in that: The system is applied to work by a novel coronavirus lung CT image infection area segmentation method as described in any one of claims 1 to 9, the system comprising an image preprocessing module, a model building module and a segmentation processing module; The image preprocessing module is used to preprocess the CT image to obtain corresponding data for subsequent use; The model building module is used to build an infected area segmentation network model, and use the data set and the loss function to train and optimize the parameters of the infected area segmentation network model to obtain a segmentation network model; The segmentation processing module is used to input the CT image to be processed into the optimal infection area segmentation network model to output the final segmentation result and complete the segmentation task of the new coronavirus lung infection area.

Images