U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method

Abstract

The invention provides a brain image denoising method based on a U-shaped dynamic convolution multi-scale multi-branch network. The brain image denoising method comprises the steps that brain noise images to be denoised are input into three branch networks formed by U-netAM, DSHFN and MSDSRN in parallel; in the U-netAM branch, multi-level features are extracted through an encoder in sequence, after weighting is conducted through a channel and a space attention mechanism, the spatial resolution is recovered through a decoder, and a first feature map is obtained; in the DSHFN branch, the dynamic convolution kernel generates a corresponding convolution kernel in real time according to input image features, the image is decomposed into a low-frequency part and a high-frequency part, and the low-frequency part and the high-frequency part are subjected to weighted fusion after being processed by a low-pass filter and a high-pass filter respectively to obtain a second feature map; in the MSDSRN branch, adopting multi-scale depth separable convolution to extract multi-scale features in parallel, and obtaining a third feature map through residual connection and fusion; inputting the three feature maps into an FPB block for fusion to obtain a noise feature map; and performing pixel-by-pixel subtraction on the original brain noise image and the noise feature map, and outputting a denoised brain image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing and artificial intelligence, in particular to a U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method. BACKGROUND

[0002] Medical imaging technology obtains internal structure information of human body in a non-invasive manner, and is an important tool for disease diagnosis and scientific research. Among them, positron emission tomography (PET) uses the three-dimensional distribution of radioactive tracers in the body to reflect tissue metabolism and function, and can provide molecular level diagnostic basis for brain diseases. However, due to the limited dose of tracers and insufficient detection efficiency, random noise is easily introduced into the PET image during acquisition, which reduces the image contrast and quantitative accuracy, and affects the detection of small lesions. Therefore, denoising of brain PET images to restore the details of the original image and improve the diagnostic accuracy is a basic requirement in the field of medical image processing. Traditional image denoising techniques can be divided into two categories: transform domain and time domain. Transform domain methods often use wavelet, Fourier or discrete cosine transform to separate image signals and noise in the frequency domain and filter out high-frequency noise; time domain methods directly suppress noise in the pixel space using mean filtering, median filtering, anisotropic diffusion or improved algorithms. In recent years, with the rapid development of deep learning, convolutional neural networks (CNN) have made significant progress in medical image denoising tasks due to their end-to-end learning capabilities. In particular, the U-Net structure uses a symmetric encoder-decoder architecture and skip connections to extract high-level semantic features while preserving low-level details, and has been widely used in brain image denoising. To further improve performance, researchers have also introduced dynamic convolution, attention mechanisms and multi-scale convolution, allowing the network to adaptively adjust the convolution kernel parameters based on the input content, and fuse local and global information extracted by different receptive fields, thereby improving denoising effect and robustness. Although the above methods show certain effects in general scenarios, there are still obvious deficiencies for brain PET images: Fixed parameter convolution kernels are difficult to adapt to the randomness and non-uniformity of PET noise, resulting in blurred or missing details in the denoising results at the edges; Single-scale feature extraction cannot capture both subtle metabolic regions and overall structure information in the brain, which may cause incomplete local lesion information or global structure distortion; Traditional residual blocks use standard convolution, which has a large number of parameters and computational complexity, making it difficult to meet real-time processing requirements in clinical settings while maintaining accuracy; Existing networks lack specialized modeling of high-frequency details, which may over-smooth while removing noise, reducing the quantitative accuracy of PET images. SUMMARY

[0003] According to the technical problems proposed above, a U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method is provided. The present application studies an efficient denoising method for the noise characteristics of the PET image data set, and realizes the removal of PET random noise while completely preserving metabolic and anatomical details, so as to reduce noise while preserving key metabolic and structural information, improve imaging quality and quantitative accuracy.

[0004] The technical means adopted by the present application are as follows: A U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method, comprising: S1, parallelly inputting the brain noise image to be denoised into three branch networks composed of a fusion attention mechanism U-shaped network (U-netAM), a dynamic sparse high-frequency network (DSHFN) and a multi-scale deep separable residual network (MSDSRN); S2, in the branch fusion attention mechanism U-shaped network (U-netAM), sequentially extracting multi-level features through the encoder, and then weighting through the channel attention mechanism and the spatial attention mechanism, and then recovering the spatial resolution through the decoder to obtain a first feature map; S3, in the branch dynamic sparse high-frequency network (DSHFN), using a dynamic convolution kernel to generate a corresponding convolution kernel according to the input image features to decompose the image into a low-frequency part and a high-frequency part, and then weighting and fusing after processing through a low-pass filter and a high-pass filter to obtain a second feature map; S4, in the branch multi-scale deep separable residual network (MSDSRN), multi-scale deep separable convolution is used to extract multi-scale features in parallel, and the third feature map is obtained after fusion through residual connection; S5, inputting the first feature map, the second feature map and the third feature map into a feature processing block (FPB) for fusion to obtain a noise feature map; S6, subtracting the noise feature map from the original brain noise image pixel by pixel to output a denoised brain image.

[0005] Further, the encoder of the fusion attention mechanism U-shaped network comprises at least three convolution-pooling structures, each of which is connected in series with a channel attention module and a spatial attention module in turn, and the decoder comprises an up-sampling-convolution structure corresponding to the number of layers of the encoder, and the encoder and the decoder are spliced through a jump connection.

[0006] Further, the channel attention module performs global average pooling on the input feature map according to the channel dimension, generates channel weights through two fully connected layers, and multiplies the original feature map by the channel weights; the spatial attention module performs maximum pooling and average pooling on the input feature map according to the spatial dimension, generates spatial weights through a 7x7 convolution, and multiplies the original feature map by the spatial weights pixel by pixel.

[0007] Further, the dynamic convolution kernel generation step of the dynamic sparse high-frequency network comprises: performing frequency decomposition on the input feature map to obtain low-frequency features and high-frequency features; generating low-frequency convolution kernels and high-frequency convolution kernels based on the low-frequency features and the high-frequency features, respectively; performing convolution operation on the generated convolution kernels and the corresponding feature maps; fusing the convolution results according to the preset weighting coefficients to output a second feature map.

[0008] Further, the generation of the high-frequency convolution kernel adopts a sparse mechanism, and the sparsity rate is set to 90%, so that the proportion of non-zero parameters in the high-frequency convolution kernel does not exceed 10%.

[0009] Further, the multi-scale depth separable residual network comprises a 1x1 depth convolution, a 3x3 depth convolution and a 5x5 depth convolution, wherein: the 1x1 depth convolution is used for channel feature compression; the 3x3 depth convolution is used for extracting local detail features; the 5x5 depth convolution is used for capturing global spatial information; the feature maps output by the 1x1 depth convolution, the 3x3 depth convolution and the 5x5 depth convolution are spliced in the channel dimension, then fused by 1x1 pointwise convolution, and added to the input feature map in residual.

[0010] Further, the feature processing block (FPB) performs the following operations on the first feature map, the second feature map and the third feature map from the three branches: splicing the first feature map, the second feature map and the third feature map in the channel dimension to obtain spliced features; performing 1x1 pointwise convolution, batch normalization and ReLU activation on the spliced features in sequence to output a noise feature map.

[0011] Compared with the prior art, the present application has the following advantages: 1. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method provided by the present application can more comprehensively extract different scale features in brain images by fusing the attention mechanism and the multi-branch design of dynamic convolution kernels, effectively suppresses various types of noise, and improves the image quality.

[0012] 2. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method provided by the present application introduces an U-shaped structure branch based on the attention mechanism, optimizes the information transmission between the encoder and the decoder, enhances the fusion effect of multi-level features, makes the denoising process more stable and accurate, and improves the detail restoration ability.

[0013] 3. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method provided by the application realizes adaptive modeling of high-frequency details of the brain image by combining a sparse mechanism with a dynamic convolution kernel, improves the adaptability and generalization capability of the model to different noise distributions, and enhances the robustness of denoising.

[0014] 4. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method provided by the application adopts deep separable convolution in the multi-scale residual module, effectively reduces the computational complexity, improves the network operation efficiency, and at the same time, strengthens the ability to capture local details, thereby meeting the dual requirements of performance and efficiency in actual applications.

[0015] 5. In the actual application of brain image denoising, the multi-branch network of the application can improve the noise suppression effect and detail preservation ability, and is superior to other technologies in multiple objective indicators, and has strong engineering promotion and clinical application value.

[0016] In summary, the multi-branch network based on U-shaped dynamic convolution multi-scale feature extraction provided by the application effectively fuses multi-scale features, improves the denoising performance by using dynamic convolution and attention mechanism, and reduces the computational complexity by deep separable convolution. Experimental results show that the network has excellent noise suppression and detail recovery capability in brain image denoising, and has good practicability and promotion value.

[0017] Based on the above reasons, the application can be widely promoted in the fields of medical image processing and artificial intelligence. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0019] Figure 1 The method flowchart of the application.

[0020] Figure 2 The UDCMMN network model diagram of the application.

[0021] Figure 3 The U-shaped network diagram of the application fused with the attention mechanism.

[0022] Figure 4 The dynamic convolution diagram of the application.

[0023] Figure 5 The multi-scale dilated convolution residual block diagram of the application.

[0024] Figure 6 Visualization comparison of brain image denoising by different methods provided by the embodiments of the present application at a noise level of 25.

[0025] Figure 7 Visualization comparison of different methods provided by the embodiments of the present application on the RNI15 dataset.

[0026] Figure 8 Visualization comparison of different methods provided by the embodiments of the present application on the SIDD dataset. DETAILED DESCRIPTION

[0027] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0028] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] As shown in Figure 1 The present application provides a U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method, which comprises: S1, parallelly inputting a brain noise image to be denoised into three branch networks composed of a fusion attention mechanism U-shaped network (U-netAM), a dynamic sparse high-frequency network (DSHFN) and a multi-scale depth separable residual network (MSDSRN); S2, in the branch fusion attention mechanism U-shaped network (U-netAM), sequentially extracting multi-level features through an encoder, and then weighting through a channel attention mechanism and a spatial attention mechanism, and then recovering the spatial resolution through a decoder to obtain a first feature map; S3, in the branch dynamic sparse high-frequency network (DSHFN), the corresponding convolution kernel is generated according to the input image features in real time by using a dynamic convolution kernel, the image is decomposed into a low-frequency part and a high-frequency part, and after being processed by a low-pass filter and a high-pass filter respectively, the low-frequency part and the high-frequency part are fused to obtain a second feature map; S4, in the branch multi-scale depthwise separable residual network (MSDSRN), multi-scale features are extracted by using a multi-scale depthwise separable convolution in parallel, and a third feature map is obtained after fusion through a residual connection; S5, the first feature map, the second feature map and the third feature map are input into a feature processing block (FPB) for fusion to obtain a noise feature map; S6, the original brain noise image is subtracted from the noise feature map pixel by pixel, and a denoised brain image is output.

[0030] In the embodiment, as shown in Figure 2 , the UDCMMN brain image denoising model adopts a hierarchical structure, which is composed of four key parts of a fusion attention mechanism U-net (U-net of convergent attention mechanisms, U-netAM), a dynamic sparse high-frequency network (Dynamic Sparse High-Frequency Network, DSHFN), a multi-scale depthwise separable residual network (Multi-Scale Depthwise Separable Residual Network, MSDSRN) and a feature processing block (Feature Processing Block, FPB). In Figure 2 , the brain image with noise is first input into the network as the network input, and is input into the three branch networks respectively. The brain noise image is input into the fusion attention mechanism U-net (U-netAM), multi-scale features are extracted by adaptively focusing on key areas, the recognition ability of important structures in the brain image is enhanced, and the information flow and feature fusion are optimized at multiple levels. The dynamic sparse high-frequency network (DSHFN) and the dynamic sparse high-frequency network (DSHFN) extract the global and local features of the brain image. The different features are fused by the feature processing block (FPB), and the first extracted features are further processed to obtain the final brain image noise feature map. Then, by subtracting the input image from the second output noise feature map, a clearer brain image is obtained.

[0031] In specific implementation, as a preferred embodiment of the present application, the encoder of the fusion attention mechanism U-net comprises at least three convolution-pooling structures, each layer is sequentially connected with a channel attention module and a spatial attention module, the decoder comprises an up-sampling-convolution structure corresponding to the number of layers of the encoder, and the encoder and the decoder are spliced through a jump connection.

[0032] In this embodiment, as shown in Figure 3 The U-net of convergent attention mechanisms (U-netAM) is based on the traditional U-Net, and the ability of feature extraction and reconstruction is enhanced by introducing channel attention mechanism and spatial attention mechanism. The encoder part of the network gradually extracts multi-level features of the input image, and in this process, the channel dimension of the features is adjusted by the channel attention mechanism, highlighting important channel information. At the same time, the spatial attention mechanism focuses on the spatial information of the key regions in the image, and adjusts the spatial dimension of the feature map. The outputs of the two attention mechanisms are fused through a learnable weighting coefficient, so as to realize accurate extraction and optimization of multi-dimensional feature information. The decoder part restores the spatial resolution of the image through step-by-step upsampling, and uses the skip connection to splice the low-level features in the encoder with the high-level features in the decoder. In this way, not only the detail information is preserved, but also the global features are effectively integrated. Finally, the channel number is reduced through the convolution layer, and the output image is restored to the original resolution of the input image by using the bilinear interpolation method.

[0033] In specific implementation, as a preferred embodiment of the present application, the channel attention module performs global average pooling on the input feature map according to the channel dimension, generates channel weights through two fully connected layers, and multiplies the original feature map by channel. The spatial attention module performs maximum pooling and average pooling on the input feature map according to the spatial dimension, generates spatial weights through 7x7 convolution, and multiplies the original feature map by pixel.

[0034] In specific implementation, as a preferred embodiment of the present application, the dynamic convolution kernel generation step of the branch dynamic sparse high-frequency network includes: Frequency decomposition is performed on the input feature map to obtain low-frequency features and high-frequency features; Low-frequency convolution kernels and high-frequency convolution kernels are generated based on the low-frequency features and the high-frequency features, respectively; The generated convolution kernels are convolved with the corresponding feature maps; The convolution results are fused according to the preset weighting coefficients, and the second feature map is output.

[0035] In this embodiment, as shown in Figure 4As shown, the dynamic sparse high-frequency network (DSHFN) first processes the input brain image through a dynamic convolution module. According to the characteristics and noise patterns of the image, the network can automatically adjust the convolution kernel at different levels to generate a convolution kernel that matches the characteristics of the current input image. This process is dynamic, and the generation of the convolution kernel is adjusted in real time according to the content of the image, thereby effectively preserving important high-frequency details while removing noise. Specifically, the dynamic sparse high-frequency network decomposes the image into a low-frequency part and a high-frequency part for processing. The low-frequency part mainly contains the overall structural information of the image, while the high-frequency part focuses on the details and edge information in the image. The low-frequency information is extracted through a low-pass filter, while the high-frequency information is enhanced through a high-pass filter. The low-frequency information and the high-frequency information are combined through a specific weighted fusion strategy. Through this weighted fusion method, the dynamic sparse high-frequency network can recover high-frequency details in the denoising process, avoid over-smoothing, and remove noise in the image, ultimately significantly improving the quality of the brain image. The dynamic sparse high-frequency network can effectively extract and recover high-frequency details in brain image denoising, preserve edge and structural information, and is crucial for diagnostic analysis. The network combines dynamic convolution kernels and sparse mechanisms, where the dynamic convolution kernel decomposes the feature map through frequency selectivity and adaptively generates a convolution kernel, which is convolved with the original feature map to accurately extract and recover key information.

[0036] In specific implementation, as a preferred embodiment of the present application, the dynamic convolution schematic diagram is as shown in Figure 5 As shown, the generation of the high-frequency convolution kernel adopts a sparse mechanism, with a sparsity rate of 90%, so that the proportion of non-zero parameters in the high-frequency convolution kernel does not exceed 10%.

[0037] In specific implementation, as a preferred embodiment of the present application, the multi-scale depth separable residual network includes a multi-scale depth separable convolution structure, which includes 1x1 depth convolution, 3x3 depth convolution and 5x5 depth convolution, wherein: The 1x1 depth convolution is used for channel feature compression; The 3x3 depth convolution is used for extracting local detail features; The 5x5 depth convolution is used for capturing global spatial information; The feature maps output by the 1x1 depth convolution, the 3x3 depth convolution and the 5x5 depth convolution are spliced in the channel dimension, then fused through 1x1 pointwise convolution, and added to the input feature map in residual.

[0038] In this embodiment, the multi-scale deep separable residual network is an efficient network structure that integrates multi-scale feature extraction and deep separable convolution, aiming to optimize the image feature extraction process and solve the gradient vanishing problem in deep networks by introducing residual structure. Specifically as follows: Firstly, the network adopts a multi-scale convolution mechanism, which uses different sizes of convolution kernels (such as 1x1, 3x3 and 5x5) to extract features from input images. Multi-scale convolution can capture different scale feature information in images from local to global. Specifically, 1x1 convolution mainly acts on the feature compression in the channel dimension, 3x3 convolution can effectively extract the detailed features of local regions, and 5x5 convolution can capture more spatial information. By concatenating the output features of different scale convolutions, the network can comprehensively represent the image from multiple dimensions, enhancing the perception ability of diversified information.

[0039] Secondly, in order to improve the calculation efficiency and reduce the parameter amount, the network adopts deep separable convolution. Deep separable convolution decomposes the standard convolution operation into two steps. The first step is depth convolution, which independently convolves each input channel. The second step is pointwise convolution, which uses 1x1 convolution to map the output channels of depth convolution. This can significantly reduce the amount of calculation and parameter, while preserving important information in the image, improving the efficiency and scalability of the model.

[0040] In addition, the residual structure in the network effectively alleviates the gradient vanishing problem in the training of deep networks by directly adding the input features and the output of the convolution layer. The introduction of residual blocks makes information propagate more smoothly in the network, improving the training stability of the network and the trainability of deep networks. Each residual module contains multi-scale convolution, effectively promoting the fusion of feature information and enhancing the expression ability of the network. The flow chart of multi-scale dilated convolution residual block is shown in Figure 6 .

[0041] Embodiment The peak signal-to-noise ratio (PSNR), structural similarity (SSIM), normalized mean squared error (NMSE), multi-scale structural similarity (MS-SSIM), and learned perceptual image patch similarity (LPIPS) are used for quantitative evaluation, and the evaluation values of different methods are compared quantitatively. Table 1 lists the value range and evaluation effect of each objective evaluation index.

[0042] Table 1 Comparison of different model results in brain image test set indicators

[0043] The ECNDNet, DnCNN, DudeNet, and UDRN in the following table are denoising models based on convolutional neural networks, and the UDCMMN is a convolutional neural network denoising model proposed by the present application. The bold data represents the best performance.

[0044] Table 2 Comparison of different model results in brain image test set indicators

[0045] Table 3 Comparison of different model results in brain image training set indicators

[0046] Table 4 Comparison of different model results in CBSD68 indicators

[0047] Figure 5 Comparison of different models in Set15 indicators

[0048] Table 6 Running time and parameters of different models in single training

[0049] The experimental conditions and settings are as follows: The Adam algorithm is used to train and optimize the network, and the network is trained based on the Pytorch framework. After multiple experiments, the hyperparameters of the first-order and second-order momentum estimation decay rates are both 0.9.

[0050] Before training and testing, all images are converted to tensors and normalized to The learning rate determines the magnitude of model parameter updates, and a high learning rate accelerates convergence but can cause oscillation, so a strategy of gradually reducing the learning rate is usually adopted to balance training speed and stability.

[0051] This embodiment dynamically adjusts the learning rate according to different stages during the training process. Specifically, when , the learning rate is set to When , the learning rate is set to When , the learning rate is set to .

[0052] This embodiment selects the AANLIB medical image dataset to evaluate the denoising performance of the UDCMMN model on brain noise images, and introduces the CBSD68 and Set15 datasets to verify its generalization ability on other color images. To test the performance of the model in a real noise environment, the SIDD and RNI15 real noise datasets are used. By integrating multiple datasets, the scientificity and wide applicability of the experiment are ensured, and reliable evaluation of the denoising performance under different noise conditions is provided.

[0053] This embodiment selects 40 brain images from the dataset, of which 20 brain images are used for pre-training and 20 brain images are used for testing. The pre-training images are cropped by a sliding window with a size of and a step size of 10, and normalized to In order to expand the diversity of the training dataset, the following eight transformations are performed on each cropped image: horizontal flip; clockwise rotation; clockwise rotation; clockwise rotation; original cropped image; horizontal flip after clockwise rotation; horizontal flip after clockwise rotation; horizontal flip after clockwise rotation. Finally, a total of 3872 brain images of are obtained and used as the real training dataset. All methods are performed on the same training and test sets, and the same number of training iterations are maintained to ensure the fairness of the experiment. As shown in Figure 6 , 7 , 8, respectively, are the visual comparison of brain image denoising at a noise level of 25 for different methods, the visual comparison of different methods on the RNI15 dataset, and the visual comparison of different methods on the SIDD dataset.

[0054] In this embodiment, the loss function uses a combined function of mean absolute error and mean square error, and by using the denoising model of the application, the following effects can be obtained: The image denoising quality is improved: the model effectively enhances the feature extraction and reconstruction ability by fusing the attention mechanism U-shaped network, provides richer and more accurate image information for the denoising task, and is superior to other methods in quantitative indicators such as PSNR, SSIM, etc.

[0055] High-frequency detail and structure information is efficiently preserved: the dynamic sparse high-frequency net optimizes the high-frequency detail recovery, highlights the image edges and key structures, and effectively avoids the problem of high-frequency information loss in conventional denoising methods.

[0056] The multi-scale feature expression ability of the model is enhanced: the multi-scale depth separable residual network combines multi-scale feature extraction and deep residual connection, which not only improves the global perception ability of the model, but also overcomes the problems of gradient disappearance and other deep network problems.

[0057] The feature fusion efficiency and denoising precision are improved: the feature processing block integrates the information output by multiple branches, realizes effective fusion of features, and further improves the overall denoising performance.

[0058] It performs well on multiple benchmark datasets: the comprehensive qualitative and quantitative experimental results show that UDCMMN exhibits good robustness and image restoration ability under different noise levels.

[0059] It shows good clinical application potential: the model achieves a better balance between noise suppression and image detail preservation, providing clearer and more reliable input for subsequent medical image analysis.

[0060] In summary, the brain image denoising model UDCMMN proposed in the application achieves excellent denoising performance on multiple benchmark datasets by fusing various structures and attention mechanisms, and shows good high-frequency detail preservation ability and clinical application potential.

[0061] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for part or all of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solution deviate from the scope of the technical solutions of the embodiments of the application.

Claims

1. A brain image denoising method based on U-shaped dynamic convolution multi-scale multi-branch network, characterized in that, The method comprises the following steps: S1, parallel input the brain noise image to be denoised into three branch networks composed of a fusion attention mechanism U-shaped network, a dynamic sparse high-frequency network and a multi-scale deep separable residual network; S2, in the branch fusion attention mechanism U-shaped network, multi-level features are extracted in turn through the encoder, and after being weighted through the channel attention mechanism and the spatial attention mechanism, the spatial resolution is restored through the decoder to obtain a first feature map; S3, in the branch dynamic sparse high-frequency network, a dynamic convolution kernel is used to generate a corresponding convolution kernel according to the input image features in real time, the image is decomposed into a low-frequency part and a high-frequency part, and after being processed by a low-pass filter and a high-pass filter respectively, the two are weighted and fused to obtain a second feature map; S4, in the branch multi-scale deep separable residual network, multi-scale features are extracted in parallel through multi-scale deep separable convolution, and the third feature map is obtained after fusion through residual connection; S5, the first feature map, the second feature map and the third feature map are input into a feature processing block for fusion to obtain a noise feature map; S6, the original brain noise image is subtracted from the noise feature map pixel by pixel to output a denoised brain image.

2. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method according to claim 1, characterized in that, The encoder of the fusion attention mechanism U-shaped network comprises at least three layers of convolution-pooling structure, each layer is connected in series with a channel attention module and a spatial attention module in turn, and the decoder comprises an up-sampling-convolution structure corresponding to the number of layers of the encoder, and the features of the same scale of the encoder and the decoder are spliced through a jump connection.

3. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method according to claim 2, characterized in that, The channel attention module performs global average pooling on the input feature map in the channel dimension, generates channel weights through two fully connected layers, and multiplies the original feature map in each channel; the spatial attention module performs maximum pooling and average pooling on the input feature map in the spatial dimension, generates spatial weights through 7×7 convolution, and multiplies the original feature map pixel by pixel.

4. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method according to claim 1, characterized in that, The dynamic convolution kernel generation step of the dynamic sparse high-frequency network comprises: Frequency decomposition is performed on the input feature map to obtain low-frequency features and high-frequency features; Low-frequency convolution kernels and high-frequency convolution kernels are generated based on the low-frequency features and the high-frequency features respectively; The generated convolution kernels are convolved with the corresponding feature maps; The convolution results are fused according to the preset weighting coefficients to output the second feature map.

5. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method according to claim 4, characterized in that, The generation of the high-frequency convolution kernel adopts a sparse mechanism, and the sparsity is set to 90%, so that the proportion of non-zero parameters in the high-frequency convolution kernel is not more than 10%.

6. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method according to claim 1, characterized in that, The multi-scale deep separable convolution structure of the multi-scale deep separable residual network comprises: 1×1 deep convolution, 3×3 deep convolution and 5×5 deep convolution, wherein: The 1×1 deep convolution is used for channel feature compression; The 3×3 deep convolution is used for extracting local detail features; The 5×5 deep convolution is used for capturing global spatial information; The feature maps output by the 1×1 deep convolution, the 3×3 deep convolution and the 5×5 deep convolution are spliced in the channel dimension, then fused through 1×1 point-by-point convolution, and added to the input feature map in residual.

7. The U-shaped dynamic convolution multi-scale multi-branch network brain image denoising method according to claim 1, characterized in that, The feature processing block performs the following operations on the first feature map, the second feature map and the third feature map from the three branches: Splice the first feature map, the second feature map and the third feature map in the channel dimension to obtain spliced features; The concatenation feature is sequentially subjected to 1x1 point-by-point convolution, batch normalization and ReLU activation, and a noise feature map is output.

Images