A coronary artery segmentation method and system based on a DUNetR model

Abstract

The application provides a coronary artery segmentation method and system based on a DUNetR model, which comprises the following steps: acquiring a three-dimensional coronary artery CTA image dataset and performing pretreatment; establishing a DUNetR model combined in parallel by a Transformer network and a UNet network, inputting the pretreated coronary artery CTA image dataset into the DUNetR model for training, inputting a coronary artery CTA image to be segmented into the trained DUNetR model for segmentation, and obtaining a preliminary segmentation result; performing smoothing processing on the preliminary segmentation result by using an image morphological algorithm to obtain a final segmentation result, and completing the segmentation of the coronary artery; the DUNetR model combined in parallel by the Transformer and UNet double networks is constructed, the attention mechanism and the full convolutional neural network are used to perform global and local detailed segmentation on the CTA, and the segmentation precision can be significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of CT blood vessel image segmentation, and more particularly to a coronary artery segmentation method and system based on a DUNetR model. BACKGROUND

[0002] At present, the segmentation of coronary artery three-dimensional CT angiography (CTA) is divided into two methods of traditional method and deep learning method. The traditional method includes image segmentation method and level set based method, which usually needs to manually place seed points or label regions for segmentation, and is basically semi-automatic. On the other hand, the segmentation method based on deep learning can be based on two-dimensional slice segmentation or three-dimensional image segmentation, however, using two-dimensional image for segmentation will cause the loss of three-dimensional space information, affecting the continuity of the segmented coronary artery image. Directly applying three-dimensional convolutional neural network for segmentation increases the computational complexity. Therefore, the existing segmentation technology has the technical limitations of insufficient spatial continuity, the need for human intervention, low segmentation precision, large consumption of computing resources, etc., which cannot be solved at the same time.

[0003] In the prior art, a fully automatic coronary artery segmentation method is disclosed, which includes the following steps: first, segmenting the heart region, which includes the coronary artery; then performing blood vessel enhancement processing on the segmented heart region to enhance the coronary artery image; then performing automatic detection of seed voxels on the image after the blood vessel enhancement processing to solve the problem of manual intervention required by the traditional region segmentation method; finally, segmenting the coronary artery through consistency judgment; however, this method needs to perform steps such as heart registration, blood vessel enhancement processing and seed point detection, compared with the method of segmenting the coronary artery by deep learning, this method needs to adjust more parameters, and is not as good as the deep learning method in feature extraction and feature interaction; in addition, in the original image, due to the CT imaging problem, the voxel value in part of the coronary artery is smaller than the surrounding voxel value, which causes the edge to be blurred, and the method of segmenting the coronary artery based on the Hessian matrix and neighborhood may be greatly affected;

[0004] A method and device for automatically detecting and segmenting aneurysm regions in intracranial CTA images are also disclosed in the prior art. The device comprises the following modules: a receiving module for receiving an intracranial CTA image to be processed; a segmentation module for performing aneurysm segmentation in a three-dimensional image block sampled from the CTA image using an aneurysm segmentation network; a resampling module for merging the aneurysm regions segmented by the segmentation module and resampling the three-dimensional image block based on connected domains; and a detection module for classifying the resampled three-dimensional image block using an aneurysm classification network to determine whether an aneurysm exists. The method in the prior art mainly detects aneurysm regions in CTA images and segments aneurysms at the patch level. However, in the area near the heart, there are disturbances other than coronary arteries, such as capillaries and similar regions, which may affect the accuracy of segmentation. SUMMARY

[0005] To overcome the defects of the prior art that full-automatic segmentation of coronary arteries cannot be achieved and the segmentation accuracy is not high, the present application provides a coronary artery segmentation method and system based on a DUNetR model, a DUNetR model combining a Transformer network and a UNet network in parallel is constructed, and a global and local detailed segmentation is performed on CTA using an attention mechanism and a fully convolutional neural network, which can significantly improve the segmentation accuracy.

[0006] To solve the above technical problems, the technical solutions of the present application are as follows:

[0007] A coronary artery segmentation method based on a DUNetR model, comprising the following steps:

[0008] S1: Obtain a three-dimensional coronary artery CTA image dataset and perform preprocessing;

[0009] S2: Establish a DUNetR model combining a Transformer network and a UNet network in parallel, input the preprocessed coronary artery CTA image dataset into the DUNetR model for training, and obtain the trained DUNetR model;

[0010] S3: Obtain a coronary artery CTA image to be segmented and input it into the trained DUNetR model for segmentation, and obtain a preliminary segmentation result;

[0011] S4: Smooth the preliminary segmentation result using an image morphological algorithm to obtain a final segmentation result, and complete the segmentation of the coronary artery.

[0012] Preferably, the specific method of preprocessing in step S1 includes any one or more of scale transformation, cropping, rotation, affine transformation and normalization operation.

[0013] Preferably, the DUNetR model established in the step S2 is specifically as follows:

[0014] The DUNetR model comprises a Transformer network and a UNet network arranged in parallel;

[0015] The Transformer network comprises a first Transformer module, a second Transformer module, a third Transformer module and a fourth Transformer module connected in sequence; each Transformer module is identical in structure and comprises three Transformer layers connected in sequence;

[0016] The UNet network comprises an input layer, an encoder, a decoder and an output layer;

[0017] The encoder comprises a first encoding block, a second encoding block, a third encoding block, a fourth encoding block and a fifth encoding block connected in sequence, and a first superposition block, a second superposition block, a third superposition block and a fourth superposition block arranged side by side;

[0018] The decoder comprises a first decoding block, a second decoding block, a third decoding block and a fourth decoding block connected in sequence;

[0019] The output of the first encoding block is further connected with the input of the output layer, the output of the second encoding block is further connected with the input of the first superposition block, the output of the third encoding block is further connected with the input of the second superposition block, the output of the fourth encoding block is further connected with the input of the third superposition block, and the output of the fifth encoding block is further connected with the input of the fourth superposition block;

[0020] The output of the first superposition block is connected with the input of the first decoding block, the output of the second superposition block is connected with the input of the second decoding block, the output of the third superposition block is connected with the input of the third decoding block, and the output of the fourth superposition block is connected with the input of the fourth decoding block; the output of the first decoding block is connected with the input of the output layer, and the output of the output layer is taken as the output of the DUNetR model;

[0021] The output of the first Transformer module is further connected with the input of the first superposition block, the output of the second Transformer module is further connected with the input of the second superposition block, the output of the third Transformer module is further connected with the input of the third superposition block, and the output of the fourth Transformer module is further connected with the input of the fourth superposition block.

[0022] Preferably, the DUNetR model is provided with a loss function L based on a similarity Dice coefficient dice , which is specifically as follows:

[0023]

[0024] Wherein, P represents the prediction result of the DUNetR model, and R represents the true label of the preprocessed coronary CTA image data.

[0025] Preferably, in the step S2, the specific method for obtaining the trained DUNetR model is as follows:

[0026] The preprocessed coronary CTA image dataset is input into the DUNetR model for training, and the loss function L dice When the value of the loss function L dice is greater than a preset threshold, the training is completed, and the trained DUNetR model is obtained.

[0027] Preferably, in the step S4, the specific method for smoothing the preliminary segmentation result by using the image morphological algorithm to obtain the final segmentation result is as follows:

[0028] The coronary artery label in the preliminary segmentation result is morphologically processed, specifically as follows:

[0029] The coronary artery label in the preliminary segmentation result is dilated by using a spherical structure with a radius R according to the shape of the coronary artery vessel surface, and the connected domain in the dilated coronary artery label is extracted by a connected domain analysis method, so that the smoothing processing is completed, and the final segmentation result is obtained.

[0030] Preferably, the step S4 further comprises: mapping the final segmentation result to a medical image space, and storing the mapped data in an nii format.

[0031] The application also provides a coronary artery segmentation system based on a DUNetR model, which is based on the above coronary artery segmentation method and comprises:

[0032] A data preprocessing module is configured to obtain a three-dimensional coronary CTA image dataset and perform preprocessing thereon;

[0033] A model establishing and training module is configured to establish a DUNetR model combined in parallel by a Transformer network and a UNet network, input the preprocessed coronary CTA image dataset into the DUNetR model for training, and obtain a trained DUNetR model;

[0034] A preliminary segmentation module is configured to obtain a coronary CTA image to be segmented and input the image into the trained DUNetR model for segmentation, and obtain a preliminary segmentation result;

[0035] A post-processing module is configured to smooth the preliminary segmentation result by using an image morphology algorithm to obtain a final segmentation result, thereby completing the segmentation of the coronary artery.

[0036] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps in the method.

[0037] The application further provides an electronic device, which comprises a processor and a memory, and the memory stores computer readable instructions, and the computer readable instructions are executed by the processor to implement the steps in the method.

[0038] Compared with the prior art, the technical scheme of the application has the beneficial effects that:

[0039] The application provides a coronary artery segmentation method and system based on a DUNetR model, which comprises the following steps: firstly, acquiring a three-dimensional coronary artery CTA image dataset and preprocessing the same; establishing a DUNetR model combined in parallel by a Transformer network and a UNet network, inputting the preprocessed coronary artery CTA image dataset into the DUNetR model for training, and obtaining a trained DUNetR model; acquiring a coronary artery CTA image to be segmented and inputting the same into the trained DUNetR model for segmentation, and obtaining a preliminary segmentation result; and smoothing the preliminary segmentation result by using an image morphology algorithm to obtain a final segmentation result, thereby completing the segmentation of the coronary artery.

[0040] The application enhances and expands the original image by using digital image processing methods such as image scaling, rotation and image enhancement, so that the network model has better learning efficiency; meanwhile, the application proposes a novel coronary artery segmentation model combined in parallel by a global context segmentation network and a local detail segmentation network, which has higher segmentation accuracy and efficiency; the application introduces a Transformer model, uses the attention mechanism to acquire long-distance features, and can more effectively extract feature information between the upper and lower parts of a 3D image to realize accurate global and local segmentation; in addition, the application fuses the global segmentation feature information extracted by the Transformer and the local segmentation detail features extracted by the UNet network by feature splicing, so that the global and local segmentation results are combined together, thereby obtaining more accurate segmentation results. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A coronary artery segmentation method flowchart based on the DUNetR model provided in embodiment 1.

[0042] Figure 2 A DUNetR model structure diagram provided in embodiment 2.

[0043] Figure 3 A DUNetR model data flow diagram provided for example 2.

[0044] Figure 4 A coronary artery segmentation result diagram provided for example 2.

[0045] Figure 5 A TransUNet network structure diagram provided for example 2.

[0046] Figure 6 A MISSU network structure diagram provided for example 2.

[0047] Figure 7 A UNETR network structure diagram provided for example 2.

[0048] Figure 8 A coronary artery segmentation system structure diagram based on a DUNetR model provided for example 3. DETAILED DESCRIPTION

[0049] The accompanying drawings are only for illustrative purposes and should not be construed as limiting the patent;

[0050] In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;

[0051] It is understandable that some well-known structures and their descriptions in the drawings may be omitted for those skilled in the art.

[0052] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0053] Example 1

[0054] As shown in Figure 1 The present application provides a coronary artery segmentation method based on a DUNetR model, comprising the following steps:

[0055] S1: Obtain a three-dimensional coronary CTA image dataset and pre-process it;

[0056] S2: Establish a DUNetR model combined in parallel by a Transformer network and a UNet network, input the pre-processed coronary CTA image dataset into the DUNetR model for training, and obtain the trained DUNetR model;

[0057] S3: Obtain a coronary CTA image to be segmented and input it into the trained DUNetR model for segmentation, and obtain a preliminary segmentation result;

[0058] S4: smoothing the preliminary segmentation result by using an image morphology algorithm to obtain a final segmentation result, and completing the segmentation of the coronary artery.

[0059] In the specific implementation process, first, a three-dimensional coronary CTA image dataset is acquired and preprocessed; a DUNetR model combined in parallel by a Transformer network and a UNet network is established, the preprocessed coronary CTA image dataset is input into the DUNetR model for training, and a trained DUNetR model is acquired; a coronary CTA image to be segmented is acquired and input into the trained DUNetR model for segmentation, and a preliminary segmentation result is acquired; the preliminary segmentation result is smoothed by using an image morphology algorithm to obtain a final segmentation result, and the segmentation of the coronary artery is completed;

[0060] The embodiment proposes a coronary artery segmentation method combined in parallel by a Transformer network and a UNet network, uses an attention mechanism and a fully convolutional neural network to perform a global and local detailed segmentation on a CTA, and compared with a traditional coronary artery segmentation method, the method has achieved significant improvement in feature extraction and segmentation precision; compared with other deep learning methods, the scheme can more effectively extract and fuse global information and local information, thereby realizing an accurate full-automatic segmentation model on a CTA image, the improvement improves the prediction efficiency and precision, and provides beneficial help for subsequent clinical diagnosis and treatment.

[0061] Embodiment 2

[0062] The embodiment provides a coronary artery segmentation method based on a DUNetR model, including the following steps:

[0063] S1: acquiring a three-dimensional coronary CTA image dataset and preprocessing;

[0064] S2: establishing a DUNetR model combined in parallel by a Transformer network and a UNet network, inputting the preprocessed coronary CTA image dataset into the DUNetR model for training, and acquiring a trained DUNetR model;

[0065] S3: acquiring a coronary CTA image to be segmented and inputting the image into the trained DUNetR model for segmentation, and acquiring a preliminary segmentation result;

[0066] S4: smoothing the preliminary segmentation result by using an image morphology algorithm to obtain a final segmentation result, and completing the segmentation of the coronary artery.

[0067] The specific method of preprocessing in the step S1 includes any one or more of a scale transformation, a cropping, a rotation, an affine transformation and a normalization operation.

[0068] The DUNetR model established in the step S2 is specifically as follows:

[0069] As shown in Figure 2 The DUNetR model includes a Transformer network and a UNet network arranged in parallel;

[0070] The DUNetR model includes a Transformer network and a UNet network arranged in parallel;

[0071] The Transformer network includes a first Transformer module, a second Transformer module, a third Transformer module and a fourth Transformer module connected in sequence; each Transformer module has the same structure and includes three Transformer layers connected in sequence;

[0072] The UNet network includes an input layer, an encoder, a decoder and an output layer;

[0073] The encoder includes a first encoding block, a second encoding block, a third encoding block, a fourth encoding block and a fifth encoding block connected in sequence, and a first superposition block, a second superposition block, a third superposition block and a fourth superposition block arranged side by side;

[0074] The decoder includes a first decoding block, a second decoding block, a third decoding block and a fourth decoding block connected in sequence;

[0075] The output of the first encoding block is further connected with the input of the output layer, the output of the second encoding block is further connected with the input of the first superposition block, the output of the third encoding block is further connected with the input of the second superposition block, the output of the fourth encoding block is further connected with the input of the third superposition block, and the output of the fifth encoding block is further connected with the input of the fourth superposition block;

[0076] The output of the first superposition block is connected with the input of the first decoding block, the output of the second superposition block is connected with the input of the second decoding block, the output of the third superposition block is connected with the input of the third decoding block, and the output of the fourth superposition block is connected with the input of the fourth decoding block; the output of the first decoding block is connected with the input of the output layer, and the output of the output layer is taken as the output of the DUNetR model;

[0077] The output of the first Transformer module is further connected with the input of the first superposition block, the output of the second Transformer module is further connected with the input of the second superposition block, the output of the third Transformer module is further connected with the input of the third superposition block, and the output of the fourth Transformer module is further connected with the input of the fourth superposition block;

[0078] The DUNetR model is provided with a loss function L based on a similarity Dice coefficient dice , specifically:

[0079]

[0080] Wherein, P represents the prediction result of the DUNetR model, and R represents the true label of the preprocessed coronary CTA image data;

[0081] In the step S2, the specific method for obtaining the trained DUNetR model is:

[0082] The preprocessed coronary CTA image data set is input into the DUNetR model for training, and the loss function L is calculated after each training dice When the value of the loss function L dice is greater than a preset threshold, the training is completed, and the trained DUNetR model is obtained;

[0083] In the step S4, the specific method for smoothing the preliminary segmentation result using an image morphological algorithm to obtain the final segmentation result is:

[0084] The coronary artery label in the preliminary segmentation result is morphologically processed, specifically:

[0085] The coronary artery label in the preliminary segmentation result is dilated according to the shape of the coronary artery vessel surface using a spherical structure with a radius R, and the connected domain in the dilated coronary artery label is extracted by a connected domain analysis method, the smoothing processing is completed, and the final segmentation result is obtained.

[0086] The step S4 further includes: mapping the final segmentation result to a medical image space, and storing the mapped data in nii format.

[0087] In the specific implementation process, first, a three-dimensional coronary CTA image data set is obtained and preprocessed;

[0088] In the 400 coronary CTA data sets collected in this embodiment, five-fold cross-validation is adopted, the 400 data sets are divided into five equal parts, one of which (80 examples) is taken as test data, and the remaining four (320 examples) are taken as training data sets, a total of five rounds, so that each data set can be used as test set;

[0089] All training data sets are subjected to image technology processing such as scaling, cropping, rotation, affine transformation and normalization, to enhance and expand the CTA data set;

[0090] A DUNetR model combined in parallel by a Transformer network and a UNet network is established, and a data processing flow of the DUNetR model is as shown in Figure 3 The preprocessed coronary CTA image dataset is input into the DUNetR model for training, and a trained DUNetR model is obtained, Figure 3 In the formula, Tf represents a Transformer module, C represents an encoding block, F represents a stacking block, and U represents a decoding block.

[0091] The attention mechanism in the Transformer can model the features between the upper and lower images of the coronary image, and realize the construction of long-distance feature dependence; the UNet network is responsible for extracting the local feature information of the coronary image, and the combination of the features extracted by the two realizes the fusion of the global information and the local information of the coronary artery, thereby effectively improving the segmentation accuracy.

[0092] The dual-parallel network structure of the Transformer and the Unet effectively realizes the combination of long-distance dependence and local information, the splicing and fusion operation of shallow features and deep features, receives features with the same resolution, and helps to generate labels with the same size. At the same time, through the introduction of deep features, the receptive field is increased to generate high-precision segmentation labels.

[0093] In the embodiment, the DUNetR model is provided with a loss function L dice based on the similarity Dice coefficient, and specifically:

[0094]

[0095] In the formula, P represents the prediction result of the DUNetR model, and R represents the true label of the preprocessed coronary CTA image data. The closer the predicted image is to the true label image, the higher the loss function value is, and vice versa. When the value of the loss function L dice is greater than a preset threshold value, the training is completed, and the trained DUNetR model is obtained.

[0096] The coronary CTA image to be segmented is obtained and input into the trained DUNetR model for segmentation to obtain a preliminary segmentation result.

[0097] The image morphological algorithm is used to smooth the preliminary segmentation result to smooth and complete the burrs and small broken parts caused by the segmentation of the blood vessels.

[0098] In order to capture the shape of the blood vessel surface of the coronary artery, a spherical structure with a radius R is used for inflation operation, in order to obtain local information around the coronary artery and reduce redundancy, according to medical prior knowledge, under the premise of existence of left and right coronary arteries, the largest two connected domains are extracted by connected domain analysis, and the coronary artery is smoothed, so that the final segmentation result is more smooth, and the breaking and splitting situation is reduced;

[0099] In the morphological processing module, the coronary artery label is subjected to morphological processing, and a spherical structure with a radius R is used for inflation according to the shape of the blood vessel surface, and the segmented coronary artery is used as the skeleton main body for smoothing processing, and some disconnected parts are connected through connected domain, and finally the mapping storage module maps the coronary artery result after segmentation and morphological processing back to the original image space, and stores the three-dimensional matrix data into medical image format to obtain the final segmentation result, completes the segmentation of the coronary artery, and the final segmentation result prediction label is as shown in Figure 4

[0100] The final segmentation result is mapped to the medical image space, and the mapped data is stored in nii format with spatial information;

[0101] The results of five-fold cross-validation are shown in Table 1, and it can be seen that the DUNetR network model based on the parallel combination of Transformer and Unet is much better than the ordinary UNet network structure and the network model combined with Transformer and Unet without parallel combination, and the DUNetR network combined with Transformer and Unet is more stable than other segmentation effects in different folds, and the mean square error is the smallest; and the final average dice coefficient of the method is 0.8376, which reaches a high segmentation level;

[0102] Table 1: Dice index table under five folds

[0103] Name Fold_1 (%) Fold_2 (%) Fold_3 (%) Fold_4 (%) Fold_5 (%) ACC (%) MISSU 75.34±3.16 74.06±3.13 74.21±3.63 73.80±3.22 74.87±3.19 74.46±3.32 UNETR 73.81±3.25 73.62±2.93 73.67±2.71 73.89±2.47 74.37±2.46 73.87±2.78 DUNETR 77.12±2.53 76.96±2.80 77.26±2.63 76.20±3.04 77.26±2.69 76.96±2.76

[0104] In this embodiment, the structure of the segmentation model is improved based on the CTA image, and the DUNetR network framework is constructed, and the 3D coronary artery segmentation is defined as a one-dimensional sequence to sequence prediction problem, the left Transformer encoder converts the image to a one-dimensional sequence, and the attention mechanism is used to learn the context information, the extracted information is connected with the features extracted by the CNN encoder under multi-scale, and then combined with the CNN decoder, and finally the segmentation prediction result is obtained;

[0105] ​The output of the Transformer as the encoder of the network is a multi-scale output. Currently, the traditional Transformer and Unet combined network directly outputs the encoding layer of the Transformer as the encoder, rather than performing multi-scale output, such as TransUNet, MISSU network structure, etc., as shown in Figure 5 , Figure 6 ,

[0106] In addition, although UNETR also performs multi-scale output of the Transformer, as shown in Figure 7 , DUNetR performs double parallel multi-scale output, has an encoder with two branches, and the right side also has a CNN encoder for supplementary detailed feature extraction. This is the reason why DUNetR has higher accuracy and smaller error than UNETR, MISSU and other network structures in 3D coronary artery segmentation, as shown in Table 1 of the experimental data;

[0107] The embodiment proposes a coronary artery segmentation method combining a Transformer and a UNet double network in parallel. The method uses an attention mechanism and a fully convolutional neural network to perform global and local detailed segmentation on CTA. Compared with the traditional coronary artery segmentation method, the method has achieved significant improvement in feature extraction and segmentation accuracy. Compared with other deep learning methods, the scheme can more effectively extract and fuse global information and local information, thereby realizing an accurate fully automatic segmentation model on CTA images. This improvement improves the efficiency and accuracy of prediction and provides beneficial help for subsequent clinical diagnosis and treatment.

[0108] Embodiment 3

[0109] As shown in Figure 8 , the embodiment provides a coronary artery segmentation system based on a DUNetR model. Based on the coronary artery segmentation method described in embodiment 1 or 2, the system includes:

[0110] A data preprocessing module 301 is configured to acquire a three-dimensional coronary artery CTA image dataset and perform preprocessing.

[0111] A model establishment and training module 302 is configured to establish a DUNetR model combining a Transformer network and a UNet network in parallel, input the preprocessed coronary artery CTA image dataset into the DUNetR model for training, and acquire a trained DUNetR model.

[0112] A preliminary segmentation module 303 is configured to acquire a coronary artery CTA image to be segmented and input the image into the trained DUNetR model for segmentation, and acquire a preliminary segmentation result.

[0113] The post-processing module 304 is used for smoothing the preliminary segmentation result by using an image morphology algorithm to obtain a final segmentation result, and completing the segmentation of the coronary artery.

[0114] In the specific implementation process, firstly, the data preprocessing module 301 acquires a three-dimensional coronary CTA image data set and performs preprocessing; the model establishment and training module 302 establishes a DUNetR model combined in parallel by a Transformer network and a UNet network, inputs the preprocessed coronary CTA image data set into the DUNetR model for training, and acquires the trained DUNetR model; the preliminary segmentation module 303 acquires the coronary CTA image to be segmented and inputs it into the trained DUNetR model for segmentation, and acquires a preliminary segmentation result; finally, the post-processing module 304 smoothes the preliminary segmentation result by using an image morphology algorithm to obtain a final segmentation result, and completes the segmentation of the coronary artery.

[0115] The embodiment proposes a coronary artery segmentation system combined in parallel by a Transformer network and a UNet network, uses an attention mechanism and a full convolutional neural network to perform a global and local detailed segmentation method on CTA, and compared with a traditional coronary artery segmentation scheme, the system has achieved significant improvement in feature extraction and segmentation precision; in addition, the system can more effectively extract and fuse global information and local information, thereby realizing an accurate full-automatic segmentation model on a CTA image and improving the prediction efficiency and precision.

[0116] The same or similar reference signs correspond to the same or similar components;

[0117] The positional relationship described in the drawings is only used for illustrative description, and cannot be understood as a limitation on the patent;

[0118] Obviously, the above embodiments of the application are only examples for clearly illustrating the application, and are not intended to limit the implementation modes of the application. Based on the above description, other different forms of changes or modifications can be made by those skilled in the art. Here, all the implementation modes need not and cannot be exhausted. Any modification, equivalent replacement and improvement made within the spirit and principle of the application should be included in the protection scope of the claims of the application.

Claims

1. A coronary artery segmentation method based on the DUNetR model, characterized in that, Includes the following steps: S1: Acquire three-dimensional coronary CTA image data and perform preprocessing; S2: Establish a DUNetR model that combines Transformer network and UNet network in parallel. Input the preprocessed coronary CTA image dataset into the DUNetR model for training and obtain the trained DUNetR model. S3: Obtain the coronary artery CTA image to be segmented and input it into the trained DUNetR model for segmentation to obtain preliminary segmentation results; S4: Use image morphology algorithms to smooth the preliminary segmentation results to obtain the final segmentation results and complete the segmentation of the coronary arteries; The specific preprocessing methods in step S1 include any one or more of the following: scaling transformation, cropping, rotation, affine transformation, and normalization operations; The DUNetR model established in step S2 is specifically as follows: The DUNetR model includes a Transformer network and a UNet network configured in parallel. The Transformer network includes a first Transformer module, a second Transformer module, a third Transformer module, and a fourth Transformer module connected in sequence; each Transformer module has the same structure, including three Transformer layers connected in sequence. The UNet network includes an input layer, an encoder, a decoder, and an output layer; The encoder includes a first encoding block, a second encoding block, a third encoding block, a fourth encoding block, and a fifth encoding block connected in sequence, as well as a first superimposed block, a second superimposed block, a third superimposed block, and a fourth superimposed block arranged in parallel; The decoder includes a first decoding block, a second decoding block, a third decoding block, and a fourth decoding block connected in sequence; The output of the first coding block is also connected to the input of the output layer, the output of the second coding block is also connected to the input of the first stacking block, the output of the third coding block is also connected to the input of the second stacking block, the output of the fourth coding block is also connected to the input of the third stacking block, and the output of the fifth coding block is also connected to the input of the fourth stacking block. The output of the first stacking block is connected to the input of the first decoding block, the output of the second stacking block is connected to the input of the second decoding block, the output of the third stacking block is connected to the input of the third decoding block, and the output of the fourth stacking block is connected to the input of the fourth decoding block; the output of the first decoding block is connected to the input of the output layer, and the output of the output layer is used as the output of the DUNetR model. The output of the first Transformer module is also connected to the input of the first overlay block, the output of the second Transformer module is also connected to the input of the second overlay block, the output of the third Transformer module is also connected to the input of the third overlay block, and the output of the fourth Transformer module is also connected to the input of the fourth overlay block.

2. The coronary artery segmentation method based on the DUNetR model according to claim 1, characterized in that, The DUNetR model is configured with a loss function based on the similarity Dice coefficient. Specifically: Where P represents the prediction result of the DUNetR model, and R represents the true label of the preprocessed coronary CTA image data.

3. The coronary artery segmentation method based on the DUNetR model according to claim 2, characterized in that, In step S2, the specific method for obtaining the trained DUNetR model is as follows: The preprocessed coronary CTA image dataset was input into the DUNetR model for training, and the loss function was calculated after each training iteration. When the loss function When the value is greater than the preset threshold, training is completed and the trained DUNetR model is obtained.

4. The coronary artery segmentation method based on the DUNetR model according to claim 3, characterized in that, In step S4, the method for smoothing the preliminary segmentation result using an image morphology algorithm to obtain the final segmentation result is as follows: Morphological processing was performed on the coronary artery tags in the preliminary segmentation results, specifically as follows: Based on the shape of the coronary artery surface, a spherical structure with radius R is used to expand the coronary artery labels in the preliminary segmentation results. The connected components are then extracted from the expanded coronary artery labels using connected component analysis, and smoothing is performed to obtain the final segmentation results.

5. A coronary artery segmentation method based on the DUNetR model according to claim 1 or 4, characterized in that, Step S4 is followed by mapping the final segmentation result to a medical image space and storing the mapped data in nii format.

6. A coronary artery segmentation system based on the DUNetR model, based on the coronary artery segmentation method according to any one of claims 1 to 5, characterized in that, include: Data preprocessing module: used to acquire and preprocess 3D coronary CTA image datasets; Model building and training module: used to build a DUNetR model that combines Transformer network and UNet network in parallel. The preprocessed coronary CTA image dataset is input into the DUNetR model for training, and the trained DUNetR model is obtained. Preliminary segmentation module: Used to acquire the coronary artery CTA image to be segmented and input it into the trained DUNetR model for segmentation to obtain preliminary segmentation results; Post-processing module: Used to smooth the preliminary segmentation results using image morphology algorithms to obtain the final segmentation result and complete the segmentation of the coronary arteries.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 5.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer-readable instructions, which, when executed by the processor, perform the steps of the method as described in any one of claims 1 to 5.

Images