A method for ultrasonic image denoising based on multi-attention fusion

Abstract

The present invention claims protection for an ultrasonic image denoising method based on multi-attention fusion, which comprises the following steps: obtaining an original ultrasonic image, inputting the original ultrasonic image into a constructed residual codec network RED‑MAM with multi-attention mechanism fusion, wherein the residual codec network RED‑MAM comprises an encoder and a multi-attention fusion decoder, wherein the encoder is used to extract features of the input image, and in addition to a basic decoding and reconstruction module, a multi-attention mechanism fusion module MAM is introduced into the multi-attention fusion decoder, wherein each encoding block or decoding block corresponds to a ReLU activation function, and a multi-attention mechanism fusion attention block is introduced after the first and third deconvolutions of the decoding block; then, the multi-attention fusion feature vector output by the multi-attention mechanism fusion attention block is activated through the ReLU activation function; then, the multi-attention fusion feature is input into the next decoding block, and the decoded ultrasonic image is obtained through decoding and reconstruction.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular to an ultrasonic image denoising method. Background Art

[0002] In recent years, with the rapid development of computers, the Internet, data storage and wireless technology, ultrasound imaging technology has also been developing. These advances have been significantly reflected in the medical field, making the diagnosis and treatment of many diseases more effective. In practical medical applications, ultrasound images are corrupted by noise during acquisition, compression and transmission. Specifically, during the acquisition process, due to the limitations of the imaging environment and equipment performance, the obtained image is very likely to be damaged by noise or artifacts, resulting in poor quality of ultrasound images and even the loss of key information. In particular, speckle noise in ultrasound images blurs relevant details and reduces the contrast of soft tissues in the image, causing serious visual interference to physicians and affecting their judgment, making the interpretation and diagnosis of the disease a difficult task. Speckle is caused by the backscattering phenomenon generated when ultrasound waves propagate in biological media, which is difficult to reduce during the image acquisition process. Ultrasound image denoising faces a major challenge from speckle noise, which depends on image data and causes quality degradation in the form of reduced contrast, blurred details and local pseudo-features. Therefore, one solution is to design a suitable post-processing step to remove speckle noise in ultrasound images without losing any important information contained in the image. Ultrasound image denoising has important clinical significance and application value. In recent years, attention mechanisms have made great progress in the field of computer vision, such as channel attention mechanisms, spatial attention mechanisms, etc. Inspired by the development of attention mechanisms, attention mechanisms have been extended to the task of ultrasound image denoising. In order to make the denoising model better adapt to the characteristics of speckle noise, this paper designs an ultrasound image denoising model based on multi-attention fusion. It focuses on the importance of texture structure information in ultrasound images from multiple domains, making the denoising model more detailed in the reconstruction process of ultrasound images, thereby improving the denoising effect of the model on ultrasound images.

[0003] CN112200750A, a method for establishing an ultrasonic image denoising model and an ultrasonic image denoising method, belongs to the field of image denoising, and includes: a data set preprocessing step: preprocessing the images in the natural image data set into image blocks of equal size, adding speckle noise, and forming training samples by the image blocks and their corresponding noise images to obtain a training data set; a model building step: using the residual block in the ResNet network as a basic unit to establish a ResUNet network, and introducing an attention mechanism to obtain an initial denoising model to be trained, which is used to suppress speckle noise in the image; a model training step: using the training data set to train the initial denoising model, and after the training is completed, an ultrasonic image denoising model is obtained. The present invention introduces residual blocks and attention mechanisms on the basis of the UNet network, and the ultrasonic image denoising model thus established can effectively improve the denoising effect of the ultrasonic image. The present invention innovates and improves on the residual codec network RED-CNN as the basic model. In the residual codec network, unlike the basic network in the comparative patent, each coding block has a corresponding residual feature output. Second, the multi-attention fusion attention block used in the present invention is a parallel combination of spatial attention and channel attention. More specifically, it can be seen that the parallel upper route is composed of a spatial attention and a channel attention, and the channel attention in the upper route and the channel attention in the parallel lower route are different channel attention mechanisms. Due to the different types of attention mechanisms, the focus of the content is also different. Therefore, the present invention adopts parallel connection, maintains the characteristics of multiple attention mechanisms and aggregates the characteristics of multiple attention mechanisms, so that the image restoration is more realistic. Summary of the invention

[0004] The present invention aims to solve the above problems of the prior art. A method for ultrasonic image denoising based on multi-attention fusion is proposed. The technical solution of the present invention is as follows:

[0005] A method for ultrasonic image denoising based on multi-attention fusion, comprising the following steps:

[0006] The original ultrasound image is obtained and input into the constructed residual codec network RED-MAM with multi-attention mechanism fusion. The residual codec network RED-MAM includes an encoder and a multi-attention fusion decoder. The encoder is used to extract the features of the input image. In the decoder, in addition to the basic decoding and reconstruction module, the basic decoding and reconstruction network is composed of 5 deconvolutions, and the multi-attention fusion decoder is composed of deconvolution and multi-attention fusion attention blocks. A multi-attention mechanism fusion module MAM is also introduced. The multi-attention mechanism fusion module MAM includes channel attention mechanism and spatial attention mechanism, including average pooling layer, maximum pooling layer, global average pooling layer, fully connected layer and shared multi-layer perceptron. What features does the channel attention focus on that are meaningful for image restoration. Spatial attention focuses on which feature information is more meaningful; in order to effectively calculate the channel attention, the spatial dimension of the input feature map needs to be compressed, and then the spatial information is aggregated by average pooling; the maximum pooling operation can search for the most unique features and can more carefully infer the importance of the information on the channel; global average pooling generates channel statistics, thereby compressing the global spatial information into the channel; global average pooling is also the simplest aggregation technology; each encoding block or decoding block corresponds to a ReLU activation function, and after the first and third deconvolutions of the decoding block, a multi-attention mechanism fusion attention block is introduced; then the multi-attention fusion feature vector output by the multi-attention mechanism fusion attention block is activated through the ReLU activation function; then, the multi-attention fusion feature enters the next decoding block, and the decoded ultrasound image is obtained by decoding.

[0007] Furthermore, in the decoder, the weight distribution is learned by utilizing the relevant feature maps in the spatial domain and the channel domain, and the learned weights are then applied to the original feature maps, and finally a weighted sum is performed, so that attention can be paid to both the information in the spatial domain and the information in the channel domain, and the image features in each channel are treated equally, avoiding the spatial domain attention ignoring the information in the channel domain.

[0008] Furthermore, in the multi-attention mechanism fusion module MAM, it is assumed that the output from the last encoding block is a d-dimensional vector r∈[0,1] d , the first layer of deconvolution is a mapping function f with parameter θ, which can be expressed as f θ (·), therefore, the input of the MAM module is f θ(r), simplified as F; in the multi-attention fusion attention block, in order to simultaneously play the role of channel attention and spatial attention modules to further improve the performance of the denoising model, the feature M1 is output from the conventional channel attention block, and a residual block is output at the same time. The residual block is the feature output by the channel attention and directly jumps to the specified position without passing through the next link. It does not pass through each figure in sequence. The place where the residual block is input is where the arrow points, and then passes through a conventional spatial attention module again to output the feature vector M2; in the parallel branch, the feature vector passes through a channel attention, and the output feature vector and the residual block output by the skip connection are element-wise multiplied.

[0009] Furthermore, the calculation of feature F by MAM is described as follows:

[0010] M1=r+δ(MLP(AP(F))+MLP(MP(F)))

[0011] M2=f1+δ(φ 7×7 (MP(AP(f1))))

[0012] M3=r+δ(MLP(GAP(F)))

[0013] M=M2+M3

[0014] Among them, δ represents the Sigmoid activation function, M1,M2,M3,M∈r d , M is the multi-attention fusion feature vector output from the MAM block, AP represents the average pooling operation, MP represents the maximum pooling operation, GAP represents the global average pooling operation, φ 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0015] Furthermore, the basic encoder is composed of 5 convolutions, and the conventional decoder module is just composed of simple convolution blocks.

[0016] Furthermore, the expressions of the conventional channel attention module and the conventional spatial attention module are respectively:

[0017] Channel attention expression M c =MLP(Avgpool(f))

[0018] Spatial attention expression M S =δ(φ 7×7 (MP(AP(f))))

[0019] Among them, f is the input feature, Avgpool is the average pooling operation, MLP is a multi-layer perceptron composed of fully connected layers, AP represents the average pooling operation, MP represents the maximum pooling operation, GAP represents the global average pooling operation, φ 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0020] The advantages and beneficial effects of the present invention are as follows:

[0021] The main innovations of the present invention are the steps of claims 1-4.

[0022] It is not easy to think of aggregating multiple attention mechanism features through parallel connection. Its advantage is that it can enrich the feature information of the image and restore the image more completely, thereby improving the denoising performance of the denoising model.

[0023] In steps 1-4, multiple attention mechanisms are used in parallel to make up for the limitation of single attention mechanism. The importance of information collected is not comprehensive. Not only does the channel pay attention to what is important information, but also We use spatial attention to focus on which information is more important, and we also use three pooling methods to further enrich Rich feature information is conducive to the restoration of image texture details. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of a structure diagram of an ultrasound image denoising model based on multi-attention fusion according to a preferred embodiment of the present invention;

[0025] Figure 2 is a schematic diagram of the encoder structure;

[0026] Figure 3 This is a schematic diagram of the MAM multi-attention mechanism fusion attention module;

[0027] Figure 4 These are the denoising results of several methods on three datasets;

[0028] Figure 5 It is the denoising results of various methods on the FH dataset;

[0029] Figure 6 This is an enlarged view of the denoising results of various methods on the FH dataset;

[0030] Figure 7 It is the absolute difference map between the denoising results of various methods and the noisy image on the FH dataset;

[0031] Figure 8 It is the absolute difference image between the denoising results of various methods and the clean image on the FH dataset. DETAILED DESCRIPTION

[0032] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0033] The technical solution of the present invention to solve the above technical problems is:

[0034] Aiming at the problem that the channel-adaptive ultrasound image denoising algorithm based on residual codec only pays attention to the importance of the channel domain and ignores the importance of information in other domains, a denoising model based on multi-attention mechanism fusion is proposed, called the residual codec network with multi-attention mechanism fusion (RED-MAM). The RED-MAM ultrasound image denoising model can overcome the problem of inter-class noise interference and the inability of existing denoising models to pay attention to the feature information of multiple domains of ultrasound images, which affects the image reconstruction ability of the model. Similarly, the encoder extracts the features of the input image. In the decoder, in addition to the basic decoding and reconstruction module, a multi-attention mechanism fusion module (MAM) is introduced to pay attention to the importance of information in multiple feature domains of the image, so that the model can better restore the detailed information of the image feature information. In the multi-attention module, the limitations of the attention mechanism are mainly considered, so that the denoising model has the ability to pay attention to the importance of information in multiple feature domains. The structural diagram of the image denoising model (RED-MAM network) based on multi-attention mechanism fusion is shown in the figure. Figure 1 shown.

[0035] This method consists of 5 encoding blocks and 5 decoding blocks based on multi-attention mechanism fusion. Each encoding block or decoding block corresponds to a ReLU activation function, such as Figure 2 As shown in the figure. After the first and third deconvolutions in the decoding block, this chapter introduces a multi-attention mechanism fusion attention block. The multi-attention fusion feature vector output by the multi-attention mechanism fusion attention block is then activated through the ReLU activation function. Then, the multi-attention fusion feature is fed into the next decoding block.

[0036] In order to speed up the training process and retain more details, skip connections and deconvolution are used. In addition, in order to pay attention to the feature information of multiple domains during the reconstruction process and make the image restoration more detailed, this chapter constructs a multi-attention mechanism fusion attention module (MAM), such as Figure 3As shown in the figure, a multi-attention mechanism is introduced in the decoder to fuse the attention block. In the decoder of the RED-MAM denoising model, the weight distribution is learned by using the relevant feature maps in multiple domains, and the learned weights are then applied to the original feature maps, and finally the weighted sum is performed. This allows the model to pay attention to both the information in the spatial domain and the information in the channel domain, and treat the image features in each channel equally, avoiding the spatial domain attention ignoring the information in the channel domain.

[0037] The basic idea of ​​the attention mechanism in computer vision is to let the model learn to focus, focus on important information and ignore unimportant information. The spatial domain attention mechanism performs corresponding spatial transformation on the spatial domain information in the image, thereby extracting the key information. The channel domain attention mechanism adds a weight to the signal on each channel to represent the relevance of the key information in the channel domain. The larger the weight, the higher the relevance. However, the multi-attention fusion attention block proposed in this chapter is a mechanism that enables the network to adaptively strengthen key information according to different tasks. It pays attention to important feature information in both the spatial domain and the channel domain, and is mainly used to enhance the denoising performance and image restoration capabilities of the denoising model. Considering multiple feature domains, the image can be reconstructed more finely, so that the image details and textures can be restored more comprehensively. The multi-attention fusion attention module is specifically described as follows:

[0038] During the exploration, it was found that both single channel attention and single spatial attention can significantly improve the denoising performance of the model, and the effect of channel attention is more significant than that of spatial attention. Assume that the output from the last encoding block is a d-dimensional vector r∈[0,1] d , the first layer of deconvolution is a mapping function f with parameter θ, which can be expressed as f θ (·). Therefore, the input of the MAM module is f θ (r), simplified to F. In the multi-attention fusion attention block, in order to simultaneously play the role of channel attention and spatial attention modules to further improve the performance of the denoising model, the feature M1 is output from the conventional channel attention block, and a residual block M1 is output at the same time, and then it passes through a conventional spatial attention again to output the feature vector M2. In the parallel branch, the feature vector passes through a channel attention, and the output feature vector and the residual block output by the jump connection are element-wise multiplied. This MAM module not only increases the scale of the feature, but also adds richer feature vectors to the reconstruction network for recovery, which is the basis for a more comprehensive and detailed reconstruction of ultrasound images. After reasonable analysis, the structure of the multi-attention fusion attention block was finally determined, as shown below: Figure 3 The calculation of feature F by MAM is described as follows:

[0039] M1=r+δ(MLP(AP(F))+MLP(MP(F))) (5.1)

[0040] M2=f1+δ(φ 7×7 (MP(AP(f1)))) (5.2)

[0041] M3=r+δ(MLP(GAP(F))) (5.3)

[0042] M=M2+M3 (5.4)

[0043] Among them, δ represents the Sigmoid activation function, M1,M2,M3,M∈r d , M is the multi-attention fusion feature vector output from the MAM block, AP represents the average pooling operation, MP represents the maximum pooling operation, GAP represents the global average pooling operation, φ 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0044] Experiments were conducted on three ultrasound image datasets: FH, GS, and CAMUS. The three evaluation indicators of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Root Mean Square Error (RMSE) were used to compare the performance of the RED-MAM method with other denoising methods. This study is based on two public ultrasound datasets and one private dataset to verify the effectiveness of the RED-MAM algorithm on three ultrasound image datasets: FH, GS, and CAMUS. The datasets used are briefly shown in Table 1:

[0045] Table 1 Description of ultrasound image dataset

[0046]

[0047] The original ultrasound dataset is an image set collected by physical correction means, which contains a small amount of real noise in ultrasound images. Taking it as experimental data, the practicability of the model is guaranteed to a certain extent. According to the characteristics of speckle noise in ultrasound images, by adding additive Gaussian white noise to the original ultrasound images, the noise in ultrasound images is partially simulated, which is conducive to controlling the noise level of the image set for multiple groups of experiments to evaluate the robustness and stability of the proposed algorithm. Therefore, by adding additive Gaussian white noise to the above three ultrasound image datasets to simulate the additive noise in ultrasound images, not only can part of the noise in ultrasound images be simulated, but also the degree of noise in ultrasound images can be distinguished by controlling the amount of noise added, so that the proposed model can verify whether it is robust to different degrees. In addition, according to the holdout method, the training set and test set of the above three datasets are randomly divided in the ratio of 7:3 from the images with simulated speckle noise.

[0048] The denoising model and parameter configuration designed in this paper are shown in Table 2.

[0049] Table 2 Parameter configuration of RED-MAM denoising network structure

[0050]

[0051]

[0052] First, on three different ultrasound datasets, RED-MAM has the highest evaluation indexes in almost all noise levels among the comparison methods. In Table 5, when the noise level is 15, RED-SENet and RDN10 are slightly higher than the method in this chapter. In the other noise levels, RED-MAM is higher than other comparison methods. When the noise level is 10, RED-MAM is 2.0%, 0.02% and 0.8% higher than the second method in the three evaluation indicators of PSNR, SSIM and RMSE. When the noise level is 25, RED-MAM improves by 8.3%, 0.3% and 4.7% respectively in the three evaluation indicators compared with the second method. When the noise level is 30, the algorithm proposed in this chapter improves by 0.6%, 0.5% and 1.0% in the evaluation indicators compared with the second method. When the noise level is 50, the method proposed in this chapter improves by 4.4%, 0.1% and 3.8% in the evaluation indicators compared with the second method. This shows that the model introduced in this chapter has a certain degree of stability. In most cases, the objective evaluation index obtained after denoising by this model is slightly improved compared with the second index method. RED-MAM has a greater improvement in the index than RED-CNN.

[0053] Table 5 FH dataset

[0054]

[0055] Secondly, the results of the experiment on the HC18 ultrasound dataset are shown in Table 6. As can be seen from Table 6, compared with the RED-CNN method, the PSNR of RED-MAM is at least 1.79 higher and at most 3.41 higher at different noise levels. On the HC18 ultrasound dataset, except for the three evaluation indicators being suboptimal when the noise level is 50, the other noise levels all achieved the highest evaluation indicators. This shows that the RED-MAM model proposed in this chapter has good robustness and can perform well at different noise levels.

[0056] Table 6 HC18 dataset

[0057]

[0058]

[0059] Finally, in order to verify the robustness and stability of the model again, experiments were conducted on the CAMUS adult cardiac ultrasound dataset, and the experimental results are shown in Table 7. It can be observed from Table 7 that although RED-MAM has only a slight improvement compared with other methods, it still achieves the highest evaluation index in most noise levels. When the noise level is 30, although the optimal evaluation index is not achieved, it is second only to the optimal. In summary, the existing work in this chapter has achieved relatively good results in 3 ultrasound datasets and 5 different noise levels, which further illustrates that the RED-MAM method has strong robustness and stability.

[0060] Table 7 CAMUS dataset

[0061]

[0062]

[0063] This section will analyze the denoising effect of the model proposed in this chapter from the perspective of subjective evaluation, and visualize the denoising results of the above datasets. Figure 4 As shown, the denoising effects of various methods can be seen more clearly from the figure.

[0064] exist Figure 4In the figure, 1, 2, and 3 represent the FH dataset, HC18 dataset, and CAMUS dataset, respectively. (a) is a noise-free image, (b) is a noisy image, (c) is the denoising result of BM3D, (d) is the denoising result of K-SVD, (e) is the denoising result of CNN10, (f) is the denoising result of RDN10, (g) is the denoising result of RED-CNN, (h) is the denoising result of RED-SENet, (i) is the denoising result of RED-CBAM, and (j) is the denoising result of RED-MAM. Figure 4 From the denoising results of various methods on the three ultrasound datasets, it can be observed that the images after denoising by methods (c) and (d) are significantly smoother. Compared with methods (c) and (d), the denoising results of methods (e) to (j) do not over-denoise the images but focus more on restoring texture details. The following is a visual analysis of the denoising results of various methods on the fetal heart dataset. Figure 5 shown.

[0065] observe Figure 5 In the green frame area, compare the positions indicated by the black arrow and the white arrow. Figure 5 In (b), the presence of noise is clearly visible. In (c), the noise on the image is significantly reduced, and the texture structure of the image becomes too smooth. In (d), a lot of noise is removed, and the texture details are slightly smoother than (a), but the texture smoothness is lower than that in (c). By carefully observing the two arrows and comparing them with Figure 5 (a) By comparison, we found Figure 5 The texture structure in (j) is restored most finely.

[0066] To further illustrate the effectiveness of the denoising methods in this chapter, we continue to zoom in on the denoising results of various methods on the fetal heart dataset and further observe the texture details of the denoised image, such as Figure 6 Figure 6 The enlarged image of the denoising results of various methods on the FH dataset is shown. By enlarging the same area, the restoration of local texture details and the removal of noise are compared.

[0067] In order to clearly show the amount of noise removed by various methods and the distribution of noise locations, on the one hand, this chapter makes an absolute difference between the denoising result images of various methods and the noise image to obtain an absolute difference image with the noise image, from which we can observe the amount of noise removed by various methods, such as Figure 7 As shown; on the other hand, by making absolute differences between the denoising results of various methods and the clean image, the absolute difference map with the clean image is obtained, and the difference between the denoised image and the clean image can be observed, such as Figure 8 shown.

[0068] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0069] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. An ultrasonic image denoising method based on multi-attention fusion, characterized in that: The following steps are involved: The original ultrasound image is obtained, and the original ultrasound image is input into the constructed residual codec network RED-MAM with multi-attention mechanism fusion. The residual codec network RED-MAM includes an encoder and a multi-attention fusion decoder. The encoder is used to extract the features of the input image. The multi-attention fusion decoder is composed of deconvolution and multi-attention mechanism fusion modules MAM. The multi-attention mechanism fusion module MAM includes a channel attention mechanism and a spatial attention mechanism, including an average pooling layer, a maximum pooling layer, a global average pooling layer, a fully connected layer and a shared multi-layer perceptron. The channel attention needs to compress the spatial dimension of the input feature map, and then aggregate the spatial information by average pooling; global average pooling generates channel statistics, thereby compressing the global spatial information into the channel; each encoding block or decoding block corresponds to a ReLU activation function, and after the first and third deconvolutions of the decoding block, a multi-attention mechanism fusion attention block is introduced; Then, the multi-attention fusion feature vector output by the multi-attention mechanism fusion attention block is activated through the ReLU activation function; Next, the multi-attention fusion features are fed into the next decoding block, and the decoded ultrasound image is obtained; In the decoder, the weight distribution is learned by using the relevant feature maps in the spatial domain and the channel domain, and the learned weights are then applied to the original feature maps, and finally weighted summation is performed, so that both the information in the spatial domain and the information in the channel domain can be paid attention to, and the image features in each channel are treated equally; In the multi-attention mechanism fusion module MAM, the output from the last encoding block is a d-dimensional vector r∈[0,1] d , the first layer of deconvolution is a mapping function f with parameter θ, which can be expressed as f θ (·), therefore, the input of the MAM module is f θ (r), simplified as F; in the multi-attention fusion attention block, the feature M1 is output from the conventional channel attention block, and a residual block is output at the same time. The residual block is the feature output by the channel attention and directly jumps to the specified position without passing through the next link. It does not pass through each figure in sequence to draw the structure diagram. The input place of the residual block is the place pointed by the arrow, and then passes through a conventional spatial attention module again to output the feature vector M2; in the parallel branch, the feature vector passes through a channel attention, and the output feature vector and the residual block output by the jump connection are element-wise multiplied.

2. The ultrasonic image denoising method based on multi-attention fusion according to claim 1, characterized in that: The calculation of feature F through MAM is described as follows: M1=r+δ(MLP(AP(F))+MLP(MP(F))) M2=f1+δ(φ 7×7 (MP(AP(f1)))) M3=r+δ(MLP(GAP(F))) M=M2+M3 Among them, δ represents the Sigmoid activation function, M1,M2,M3,M∈r d , M is the multi-attention fusion feature vector output from the MAM block, AP represents the average pooling operation, MP represents the maximum pooling operation, GAP represents the global average pooling operation, φ 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

3. The ultrasonic image denoising method based on multi-attention fusion according to claim 1, characterized in that: The expressions of the conventional channel attention module and the conventional spatial attention module are respectively: Channel attention expression M c =MLP(Avgpool(f)) Spatial attention expression M S =δ(φ 7×7 (MP(AP(f)))) Where f is the input feature, Avgpool is the average pooling operation, MLP is a multi-layer perceptron consisting of fully connected layers, AP represents the average pooling operation, MP represents the maximum pooling operation, GAP represents the global average pooling operation, and φ 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

Images