Coronary artery segmentation method based on multi-scale multi-frequency channel attention neural network

Abstract

A coronary artery segmentation method based on a multi-scale and multi-frequency channel attention neural network is characterized in that the multi-scale and multi-frequency channel attention neural network is in a semantic segmentation model of a U-shaped structure; and connecting each layer of encoder of the down-sampling path and each layer of decoder of the up-sampling path by using a multi-scale multi-frequency channel attention module, wherein the multi-scale multi-frequency channel attention module is obtained by applying the multi-frequency channel attention module to a multi-scale network.

Description

Technical Field

[0001] The invention relates to the field of computer medical image intelligent recognition, in particular to a segmentation method for coronary angiography images. Background Art

[0002] Coronary artery disease (CAD) is a cardiovascular disease and one of the leading causes of death. Currently, there are various examination methods for coronary artery disease, such as electrocardiogram, myocardial enzyme measurement, coronary angiography, etc. Among them, invasive coronary angiography (ICA) is considered to be the gold standard for evaluating the anatomical structure and stenosis of coronary arteries.

[0003] In recent decades, the automatic segmentation of ICA vessels has received extensive attention. Compared with manual segmentation, automatic segmentation methods can often provide more objective analysis of vascular structures, and deep learning methods play a leading role in automatic segmentation by providing an end-to-end approach. With the emergence of the U-Net, a U-shaped structure semantic segmentation model based on a fully convolutional neural network (FCN) proposed in 2015, the image segmentation effect has been greatly improved by obtaining feature maps of different sizes at the spatial level and realizing the effective fusion of coarse-grained features and fine-grained features.

[0004] The U-Net network consists of a downsampling encoding path, an upsampling decoding path, and a skip connection path in the middle. The encoder path performs 4 consecutive downsampling operations, thereby reducing the image size while doubling the number of feature map channels, and obtaining shallow feature information of the feature image. The decoder path is symmetrical with it, performing 4 consecutive upsampling operations, thereby increasing the image size while halving the number of feature map channels, and obtaining deep feature information of the feature image. At the same time, a skip connection is set up between the encoder and decoder of each layer, so that the model can splice the image feature maps of the same size in the same layer, and fuse the deep feature information of the image with the shallow feature information.

[0005] Since simple skip connections are set up between each layer of encoders in the downsampling stage and each layer of decoders in the upsampling stage in U-Net, but there is a large semantic difference between the shallow features extracted in the downsampling stage and the deep image feature information in the upsampling stage, simple skip connections directly merge the two, affecting the segmentation effect. Therefore, in recent years, many improved networks based on U-Net have emerged one after another. For example, the U-Net++ model reduces the semantic difference between the feature maps of the two paths by adding nested densely connected subnetworks in the encoding path and the decoding path. The U-Net3+ model uses full-scale skip connections and deep supervision to extract information from full-scale image features. The Attention U-Net model proposes an attention-based semantic segmentation network, which adds attention gates in the skip connection path to optimize the downsampling results, and then connects them with the upsampling results to eliminate noise and irrelevant information in the skip connection process, and reduce the semantic difference between the two feature maps. The Res-UNet model is inspired by ResNet and incorporates it into UNet. It adopts an attention mechanism to extract the region of interest (ROI) and background, so that the model can focus on the target ROI area and ignore irrelevant noise background. The latest improved algorithm based on UNet is Morphi-UNet proposed in 2024. By combining multi-scale morphological modules, morphological features of different scales can be extracted to improve segmentation performance. However, most of the above segmentation algorithms operate in the spatial domain, which cannot capture the global information of the image well and are more sensitive to noise.

[0006] At the same time, from the perspective of frequency domain, segmentation models that use various frequency conversion methods to enhance image feature extraction capabilities are also emerging. For example, SENet uses SE (Squeeze-and-Excitation) blocks, which is one of the classic channel attention mechanism models. It uses global average pooling (GAP) to complete feature mapping to achieve compressed channels. Since the two-dimensional discrete cosine transform (2D DCT) transform has good spectrum perception capabilities and can convert input images from the spatial domain to the frequency domain, the FcaNet network proposes a channel attention mechanism that uses 2D DCT to compress channels. Mathematically, GAP is only equivalent to the lowest frequency component of DCT. However, these methods only extract image feature information from the image frequency domain level, which is usually global and mainly reflects the edge structural information of the image. It ignores the specific spatial position of pixels in the image and is difficult to capture local feature changes in the image.

[0007] In summary, previous deep learning segmentation methods often only focus on the spatial domain or the frequency domain, ignoring the potential benefits of combining the two, and not fully considering capturing the global and local information of the image, which limits the image segmentation accuracy. Therefore, an improved neural network structure is provided, which can fully capture the global and local semantic information of the image, thereby improving the coronary artery image segmentation effect, which has become a problem to be solved urgently in the existing technology. Summary of the invention

[0008] In view of the shortcomings of the prior art, the technical problem to be solved by the present invention is to solve the problem of insufficient image semantic feature extraction in the structural improvement of the traditional U-shaped network, improve the segmentation accuracy of invasive coronary angiography images, and propose a multi-scale multi-frequency channel attention neural network M 2 CA-Net combines image spatial domain information and frequency domain information to obtain richer semantic information. To achieve this goal, the technical solution provided by the present invention is:

[0009] A coronary artery segmentation method based on a multi-scale multi-frequency channel attention neural network, the multi-scale multi-frequency channel attention neural network (M 2 CA-Net is a U-shaped semantic segmentation model (U-Net) that uses a multi-scale multi-frequency channel attention module (M-MCM) between each encoder layer of the downsampling path and each decoder layer of the upsampling path. 2 The multi-scale multi-frequency channel attention module is connected to the multi-scale network by applying the multi-frequency channel attention module (MCA) to the multi-scale network.

[0010] A coronary artery segmentation method based on a multi-scale and multi-frequency channel attention neural network is further developed.

[0011] The multi-frequency channel attention module (MCA) specifically includes the following steps

[0012] 1.1) Use two-dimensional discrete cosine transform (2D DCT) for coronary angiography image segmentation to expand a single frequency component to multiple frequency components.

[0013] 1.2) Concatenate the multiple frequency components obtained and calculate the channel attention weight of the entire feature map:

[0014] The multi-scale multi-frequency channel attention module is obtained by applying the multi-frequency channel attention module (MCA) to the multi-scale network network, and specifically includes the following steps:

[0015] 2.1) The input coronary angiography image is processed by multi-scale branches using convolution kernels of different sizes to obtain multi-scale feature maps, and each feature map is spliced ​​along the channel dimension:

[0016] 2.2) Using the MCA module to process the feature maps of each branch in parallel, so as to obtain the attention weight information of each channel of feature maps of different scales;

[0017] 2.3) Multiply the obtained multi-scale channel attention weight vector and the overall feature map obtained by splicing feature maps of different scales channel by channel, and finally obtain the multi-scale multi-frequency channel attention weight feature map result.

[0018] Coronary artery segmentation method based on multi-scale multi-frequency channel attention neural network, further the multi-frequency channel attention module (MCA)

[0019] Step 1.1) is as follows:

[0020] The feature maps of convolutions with different kernel sizes are transformed to the frequency domain for channel compression. The compression process is to divide the input features into multiple parts along the channel dimension, and then perform 2D DCT transformation on each part of the feature map and map it to the frequency domain to obtain a single frequency vector for each part. After that, the single frequency vectors of each part are concatenated to obtain the compressed vector of the entire feature map, thereby obtaining the compression result of channel attention.

[0021] The specific steps of the compression process are:

[0022] Assume that a 2D coronary angiography feature map X of size C×H×W is given as input, where C represents the number of channels, H represents the feature map height, and W represents the feature map width. First, we divide the input feature map X into n parts along the channel direction, denoted as [X 0 ,X 1 ,…,X n-1 ], at this time each part feature map X i The dimension size becomes C′×H×W, where i∈{0,1,…,n-1} and C′=C / n, and C must be divisible by n. After that, we assign the corresponding 2D DCT component to each part as the compression result of channel attention. The compression result of each part Freq i The dimension size is C′, and the specific formula is:

[0023]

[0024] In this formula, Represents the input feature map X i The 2D DCT compression operation performed, [u i ,v i ] is X i The corresponding 2D DCT components, is the basis function of 2D DCT, so we get the compressed frequency component Freq of each part i ;

[0025] Step 1.2) is as follows:

[0026] The single frequency vectors obtained from the n parts obtained in step 1.1) are connected in series to obtain the frequency compression vector Freq of the dimension size C of the entire feature map. The specific formula is:

[0027] Freq=Cat([Freq 0 ,Freq 1 ,…,Freq n-1 ])

[0028] In this formula, Cat represents concatenating each frequency vector along the channel dimension, and finally obtaining the channel attention weight vector W of the entire feature map. The specific formula is:

[0029] W = sigmoid(fc(Freq))

[0030] In this formula, fc represents the fully connected layer, which is used to realize the information interaction between high channel and low channel dimensions. After that, the sigmoid activation function is used to assign weights to the channels after information interaction, and the weight of each channel is controlled in the range of 0-1, and the sum of the weights of each channel is added to 1.

[0031] Multi-scale and multi-frequency channel attention module (M 2 CA)

[0032] Step 2.1) is specifically:

[0033] Assume that a 2D coronary angiography feature map X of size C×H×W is given as input, where C represents the number of channels, H represents the feature map height, and W represents the feature map width. First, we divide the input feature map X into S parts along the channel direction. After each part is processed by a convolution extraction operator with different convolution kernel sizes, we obtain feature maps F with the same number of channel dimensions C′ of different scales. i , where C′=C / S,i=0,1,…,S-1. The calculation formulas for feature maps of different scales are as follows:

[0034] F i =Conv(k i ×k i )(X),i=0,1,2,…,S-1

[0035] In this formula, k i represents the size of the i-th convolution kernel, k i=2×(i+1)+1. Then, the feature maps of different scales are connected in series along the channel dimension to obtain the overall multi-scale feature map F. The specific formula is as follows:

[0036]

[0037] In this formula, represents the connection operator, and the dimension of the overall multi-scale feature map F is C×H×W.

[0038] The above content is the process of performing multi-scale processing on the input feature map to generate feature maps of different scales and splicing them along the channel dimension.

[0039] Step 2.2) is specifically:

[0040] In step 2.1) obtain feature maps F of different scales i Afterwards, the MCA module is used to extract the channel attention information of each scale feature map and obtain the channel attention weight W of each channel. i The attention weight vector is shown as follows:

[0041] W i =MCA(F i ),i=0,1,2,…,S-1

[0042] Among them, W i The dimension is C′×1×1. The MCA module is used to obtain the attention weight information of each channel from the feature map. Then, in order to interact with the channel attention information of different scales, the obtained channel attention weight vector W i Connect them in series to get the multi-scale channel attention weight W. The specific formula is as follows:

[0043]

[0044] In the above formula, represents the connection operator, after which the obtained multi-scale attention weights are processed by the activation function to obtain the multi-scale multi-frequency channel attention vector mms att , the specific formula is as follows:

[0045] mms att =softmax(W)

[0046] In the above formula, mms att Represents the multi-scale and multi-frequency channel attention vector of the input feature map. Softmax represents the Softmax activation function, which is used to map the attention weights of each channel to between 0 and 1.

[0047] The above content is the process of using the MCA module to process feature maps of different scales to extract the attention weight information of each channel, and concatenate them to obtain the multi-scale channel attention weight vector.

[0048] Step 2.3) is as follows:

[0049] After obtaining the attention weight vector mms of all channels att After that, mms att Multiply the overall multi-scale feature map F obtained in step 2.1) channel by channel to obtain the final multi-scale multi-frequency channel attention weight feature map output. The specific formula is as follows:

[0050]

[0051] In the above formula, represents a channel-by-channel multiplication operation, and output represents the obtained feature map of multi-scale and multi-frequency channel attention weights.

[0052] A method for segmenting coronary arteries based on a multi-scale multi-frequency channel attention neural network. Further, a specific method for segmenting coronary arteries based on a multi-scale multi-frequency channel attention neural network includes:

[0053] M 2 The downsampling path of CA-Net includes four encoder layers. Each layer consists of two convolution operations with a kernel size of 3 and a ReLu activation function to expand the number of channels of the feature map. After each layer of encoding convolution operation, a 2×2 maximum pooling layer operation is performed to reduce the dimension size of the feature map. The output feature map result of each layer is used as M 2 The input of the CA module is used to extract the multi-scale and multi-frequency channel attention feature map of the feature map to obtain richer semantic features. After the fourth layer of downsampling operation, two convolution operations with a kernel size of 3 and a ReLu activation function are performed to form the number of channels used to expand the feature map.

[0054] M 2 The upsampling path of CA-Net includes four layers of decoder operations. The first feature map of each layer is the feature map of the lower layer after the 2×2 up convolution operation and the output result of the corresponding layer of the downsampling path. 2 The multi-scale and multi-frequency channel attention feature maps obtained after processing by the CA module are spliced, and then two convolution operations with a convolution kernel size of 3 and a ReLu activation function are performed to reduce the number of channels of the feature map corresponding to the encoder; after the convolution operation and ReLu activation operation in the last layer, a 1×1 convolution operation is performed to change the channel dimension of the feature map to 1, and the segmentation result of the input coronary angiography image is obtained.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] The present invention provides an image feature extraction method combining image spatial domain and frequency domain information, and proposes a multi-scale multi-frequency channel attention neural network M in combination with the traditional U-shaped network structure. 2 CA-Net. This method extracts image features in both the spatial domain and the frequency domain, improving the limitations of previous models that only focus on the image spatial domain or the frequency domain, and ignores the benefits of combining the image spatial domain and the frequency domain.

[0057] The present invention combines multi-scale spatial features with multi-frequency features, and combines multi-scale features with the MCA module to propose an MCA 2 The CA module is used to extract feature information of feature maps of different scales, and integrate the multi-scale spatial information and multi-frequency channel attention information of the image. In this way, local and global channel attention is integrated to achieve better information interaction, solving the limitation of traditional attention modules that only rely on single spatial information or frequency information. Multi-scale features and the encoding and decoding paths of the U-shaped network extract the global and local features of the image well, reduce the receptive field, and extract features of different scales in the channel direction to improve the segmentation performance, solving the problem of vascular rupture in the segmentation of coronary angiography images by traditional models; at the same time, 2D DCT is introduced to convert the image from the spatial domain to the frequency domain, and the global features of the image are obtained by extracting the frequency feature information of the image, expanding the receptive field of the image feature map, and expanding the single frequency to multiple frequencies, extracting richer image channel attention features, solving the problem of incorrect background prediction of coronary artery images caused by traditional models.

[0058] The example of the present invention applies this method to two coronary artery data sets for segmentation verification, and selects the currently more advanced image segmentation model as a comparison model. The comparative experimental results are used to prove the effectiveness of the present invention in the field of coronary artery image segmentation, and also prove that the multi-frequency and multi-scale settings improve the segmentation performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The overall flow chart of the coronary artery segmentation method based on multi-scale multi-frequency channel attention neural network provided by the present invention;

[0060] Figure 2 A structural diagram of a multi-frequency channel attention module proposed for a coronary artery segmentation method based on a multi-scale multi-frequency channel attention neural network provided in an embodiment of the present invention;

[0061] Figure 3A multi-scale multi-frequency channel attention module diagram proposed for the coronary artery segmentation method based on a multi-scale multi-frequency channel attention neural network provided in an embodiment of the present invention;

[0062] Figure 4 A comparison chart of the segmentation effects of the coronary artery segmentation method based on the multi-scale and multi-frequency channel attention neural network provided in an embodiment of the present invention and other models. DETAILED DESCRIPTION

[0063] In order to make the purpose and advantages of the present invention more obvious and understandable, the technical scheme in the specific example of the present invention will be fully described below in combination with three drawings in the example of the present invention. The specific embodiment is only used to further explain the present invention in detail and does not limit the protection scope of the claims of this application.

[0064] Example

[0065] This embodiment provides a coronary artery segmentation method based on a multi-scale and multi-frequency channel attention neural network.

[0066] See also Figure 1 As shown, in this embodiment, a multi-scale multi-frequency channel attention neural network (M 2 The overall process of segmenting coronary artery images using CA-Net) includes the following steps.

[0067] S1: Divide the coronary artery dataset to be segmented into a training dataset and a test dataset.

[0068] S2: Data augmentation was performed on the training dataset, including random horizontal and vertical flipping, followed by random rotation between 0° and 180°, and all coronary artery dataset images were uniformly adjusted to 256×256 pixels.

[0069] S3: Distinguish the source images and labels of the data set after data augmentation and use them as M 2 The input of CA-Net is used to compare the segmentation results of the source image with the label during the training process to calculate the loss. The model outputs the segmentation results of the training data set, and the predicted segmentation results and the data set labels are used to calculate the Dice coefficient to measure the similarity between the two. At the same time, the binary cross entropy (BCE) loss function is combined to minimize the loss to calculate the model error and optimize the model parameters. After the training is completed, the final model is output, and the test data set is used as the input of the final model to output the final segmentation result.

[0070] M 2 CA-Net adopts the traditional U-type encoding and decoding structure combined with M-type fusion of multi-scale information and multi-frequency channel attention. 2 CA block, M 2The downsampling path of CA-Net includes four encoder layers. Each layer consists of two convolution operations with a kernel size of 3 and a ReLu activation function to expand the number of channels of the feature map. After each layer of encoding convolution operation, a 2×2 maximum pooling layer operation is performed to reduce the dimension size of the feature map. The output feature map result of each layer is used as M 2 The input of the CA module is used to extract the multi-scale and multi-frequency channel attention feature map of the feature map to obtain richer semantic features. After the fourth layer of downsampling operation, two convolution operations with a kernel size of 3 and a ReLu activation function are performed to form the number of channels used to expand the feature map.

[0071] M 2 The upsampling path of CA-Net includes four layers of decoder operations. The first feature map of each layer is the feature map of the lower layer after the 2×2 up convolution operation and the output result of the corresponding layer of the downsampling path. 2 The multi-scale and multi-frequency channel attention feature maps obtained after processing by the CA module are spliced, and then two convolution operations with a convolution kernel size of 3 and a ReLu activation function are performed to reduce the number of channels of the feature map corresponding to the encoder; after the convolution operation and ReLu activation operation in the last layer, a 1×1 convolution operation is performed to change the channel dimension of the feature map to 1, and the segmentation result of the input coronary angiography image is obtained.

[0072] This embodiment provides a multi-scale multi-frequency channel attention module (M 2 The structure diagram of CA is as follows Figure 2 shown.

[0073] Figure 2 Showing M 2 The overall structure of the CA module, the M 2 The CA module is obtained by applying the MCA module to the multi-scale network, which includes the following steps:

[0074] The input coronary angiography image is processed by multi-scale branches using convolution kernels of different sizes to obtain multi-scale feature maps, and each feature map is spliced ​​along the channel dimension: the specific steps are:

[0075] Assume that a 2D coronary angiography feature map X of size C×H×W is given as input, where C represents the number of channels, H represents the feature map height, and W represents the feature map width. First, the input feature map X is evenly divided into S parts along the channel direction. After each part is processed by a convolution extraction operator with different convolution kernel sizes, feature maps F with the same number of channel dimensions C′ with different scales are obtained. i , where C′=C / S,i=0,1,…,S-1. The calculation formulas for feature maps of different scales are as follows:

[0076] F i =Conv(k i ×k i )(X),i=0,1,2,…,S-1

[0077] In this formula, k i represents the size of the i-th convolution kernel, k i =2×(i+1)+1. Then, the feature maps of different scales are connected in series along the channel dimension to obtain the overall multi-scale feature map F. The specific formula is as follows:

[0078]

[0079] In this formula, represents the connection operator, and the dimension of the overall multi-scale feature map F is C×H×W.

[0080] The above content is the process of performing multi-scale processing on the input feature map to generate feature maps of different scales and splicing them along the channel dimension.

[0081] The feature maps of each branch are processed in parallel using the MCA module to obtain the attention weight information of each channel of feature maps of different scales; specifically:

[0082] After obtaining feature maps F of different scales i Afterwards, the MCA module is used to extract the channel attention information of each scale feature map and obtain the channel attention weight W of each channel. i The attention weight vector is shown as follows:

[0083] W i =MCA(F i ),i=0,1,2,…,S-1

[0084] Among them, W i The dimension is C′×1×1. The MCA module is used to obtain the attention weight information of each channel from the feature map. Then, in order to interact with the channel attention information of different scales, the obtained channel attention weight vector W i Connect them in series to get the multi-scale channel attention weight W. The specific formula is as follows:

[0085]

[0086] In the above formula, represents the connection operator, after which the obtained multi-scale attention weights are processed by the activation function to obtain the multi-scale multi-frequency channel attention vector mms att , the specific formula is as follows:

[0087] mms att =softmax(W)

[0088] In the above formula, mms att Represents the multi-scale and multi-frequency channel attention vector of the input feature map. Softmax represents the Softmax activation function, which is used to map the attention weights of each channel to between 0 and 1.

[0089] The above content is the process of using the MCA module to process feature maps of different scales to extract the attention weight information of each channel, and concatenate them to obtain the multi-scale channel attention weight vector.

[0090] The obtained multi-scale channel attention weight vector is multiplied channel by channel with the overall feature map obtained by concatenating feature maps of different scales, and finally the multi-scale multi-frequency channel attention weight feature map result is obtained, specifically:

[0091] After obtaining the attention weight vector mms of all channels att After that, mms att Multiply the overall multi-scale feature map F obtained in step 2.1) channel by channel to obtain the final multi-scale multi-frequency channel attention weight feature map output. The specific formula is as follows:

[0092]

[0093] In the above formula, represents a channel-by-channel multiplication operation, and output represents the obtained feature map of multi-scale and multi-frequency channel attention weights.

[0094] This embodiment provides a multi-frequency channel attention module structure diagram as shown in Figure 4 shown.

[0095] The multi-frequency channel attention module includes

[0096] Two-dimensional discrete cosine transform (2D DCT) is used for coronary angiography image segmentation to expand a single frequency component into multiple frequency components.

[0097] The two-dimensional discrete cosine transform (2D DCT) is used for coronary angiography image segmentation. The feature maps of convolution with different kernel sizes are transformed into the frequency domain and channel compression is performed. Given a 2D coronary angiography feature map F with a size of C×H×W i As input, named X, where C represents the number of channels, H represents the feature map height, and W represents the feature map width. First, the input feature map X is evenly divided into n parts along the channel direction, denoted as [X 0 ,X 1 ,…,X n-1 ], at this time each part feature map Xi The dimension size becomes C′×H×W, where i∈{0,1,…,n-1} and C′=C / n, and it is necessary to ensure that C is divisible by n. Then, the corresponding 2D DCT component is assigned to each part as the compression result of the channel attention. The compression result of each part Freq i The dimension size is C′, and the specific formula is:

[0098]

[0099] In this formula, Represents the input feature map X i The 2D DCT compression operation performed, [u i ,v i ] is X i The corresponding 2D DCT components, is the basis function of 2D DCT, In this way, the compressed frequency component Freq of each part is obtained i .

[0100] Then the multiple frequency components obtained are concatenated, and the channel attention weight of the entire feature map is calculated, specifically:

[0101] Then the compression frequency component Freq of each part i By concatenating them, we can get the frequency compression vector Freq of the dimension C of the entire feature map. The specific formula is:

[0102] Freq=Cat([Freq 0 ,Freq 1 ,…,Freq n-1 ])

[0103] In this formula, Cat represents concatenating each frequency vector along the channel dimension, and finally obtaining the channel attention weight vector W of the entire feature map. The specific formula is:

[0104] W = sigmoid(fc(Freq))

[0105] In this formula, fc represents the fully connected layer, which is used to realize the information interaction between high channel and low channel dimensions. After that, the sigmoid activation function is used to assign weights to the channels after information interaction, and the weight of each channel is controlled in the range of 0-1, and the sum of the weights of each channel is added to 1.

[0106] See also Figure 4As shown in FIG. 1 , the essence of the present invention is to design a multi-scale multi-frequency channel attention neural network to improve the accuracy of coronary artery segmentation. Taking the public data set DCA1 as an example, the model of the present invention and the existing model are used to segment coronary artery images, and the Dice similarity coefficient (DSC), sensitivity (SN), specificity (SP) and accuracy (ACC) are used as evaluation indicators, among which DSC is considered to be the most authoritative indicator. The experimental results of various models are shown in Table 1, and the effects of each model on segmenting the image are compared. Figure 4 shown.

[0107] Table 1 Experimental results of each model on the DCA1 dataset

[0108]

[0109] As shown in Table 1, in the experiments conducted on the DCA1 dataset, the performance of the model proposed in the present invention is slightly lower than that of the FR-UNet model only in the ACC index. Compared with the Res-UNet, ResAttUNet, FR-UNet, DenseUNet, U-Net and AttUNet network models, the DSC index is improved by 6.9%, 6.6%, 4.2%, 3.2%, 4.5% and 0.7%, respectively. From the experimental results in the table, it can be seen that the experimental effects of the Res-UNet and ResAttUNet segmentation models on the DCA1 dataset are not as good as those of the U-Net. This may be due to the over-smoothing effect caused by their residual connections, which has a negative impact on fine boundary segmentation. In addition, irrelevant features may be learned during the training process, further affecting the segmentation performance. However, by comparing the experimental results of Res-UNet and ResAttUNet with U-Net and AttUNet, we can observe that the attention mechanism has a positive effect on the segmentation performance of the model. The M in the method provided by the present invention 2 The CA-Net model uses the attention mechanism to compress the channels of the image feature map using the frequency domain information of the image, calculates the weight of each channel, and extracts an effective representation of the image in combination with the multi-scale structure.

[0110] In an embodiment of the present invention application, a computer program that can be run on a processor and stored in a memory is also provided. When the computer program is executed by the processor, the steps of the multi-scale multi-frequency channel attention neural network as described above can be implemented by the terminal device.

[0111] The above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Any design that utilizes the design concept of the present invention, makes some simple changes or replaces some of the technical features thereof with equivalents should fall within the protection scope of the present invention.

Claims

1. A coronary artery segmentation method based on a multi-scale and multi-frequency channel attention neural network, characterized by The multi-scale multi-frequency channel attention neural network is a U-shaped structure semantic segmentation model in which each layer of the encoder in the downsampling path and each layer of the decoder in the upsampling path are connected by a multi-scale multi-frequency channel attention module, and the multi-scale multi-frequency channel attention module is obtained by applying the multi-frequency channel attention module to the multi-scale network network.

2. The coronary artery segmentation method based on a multi-scale multi-frequency channel attention neural network as claimed in claim 1, characterized in that The multi-frequency channel attention module specifically includes the following steps: 1.1) Use two-dimensional discrete cosine transform for coronary angiography image segmentation to expand a single frequency component to multiple frequency components. 1.2) Concatenate the multiple frequency components obtained and calculate the channel attention weight of the entire feature map: The multi-scale multi-frequency channel attention module is obtained by applying the multi-frequency channel attention module to the multi-scale network network, specifically comprising the following steps 2.1) The input coronary angiography image is processed by multi-scale branches using convolution kernels of different sizes to obtain multi-scale feature maps, and each feature map is spliced ​​along the channel dimension: 2.2) Using the MCA module to process the feature maps of each branch in parallel, so as to obtain the attention weight information of each channel of feature maps of different scales; 2.3) Multiply the obtained multi-scale channel attention weight vector and the overall feature map obtained by splicing feature maps of different scales channel by channel, and finally obtain the multi-scale multi-frequency channel attention weight feature map result.

3. The coronary artery segmentation method based on multi-scale multi-frequency channel attention neural network as claimed in claim 2, characterized in that: The specific methods of segmenting coronary arteries using multi-scale and multi-frequency channel attention neural network include: The downsampling path of the multi-scale multi-frequency channel attention neural network includes four encoder layers, each of which consists of two convolution operations with a kernel size of 3 and a ReLu activation function to expand the number of channels of the feature map. After each layer of encoding convolution operation, a 2×2 maximum pooling layer operation is performed to reduce the dimensional size of the feature map; the output feature map result of each layer is used as the input of the multi-scale multi-frequency channel attention module to extract the multi-scale multi-frequency channel attention feature map of the feature map and obtain richer semantic features; after the fourth layer of downsampling operation, two convolution operations with a kernel size of 3 and a ReLu activation function are continued to be performed to expand the number of channels of the feature map; The upsampling path of the multi-scale multi-frequency channel attention neural network includes four layers of decoder operations. The first feature map of each layer is concatenated with the multi-scale multi-frequency channel attention feature map obtained by processing the output of the corresponding layer of the downsampling path through the multi-scale multi-frequency channel attention module. Then, two convolution operations with a kernel size of 3 and a ReLu activation function are performed to reduce the number of channels of the feature map corresponding to the encoder. After the convolution operation and ReLu activation operation are performed on the last layer, a 1×1 convolution operation is performed to change the channel dimension of the feature map to 1, and the segmentation result of the input coronary angiography image is obtained.

Images