Residual error mixed attention denoising method, system and device for LDCT (Level Discrete Cosine Transform) and medium

Abstract

The invention relates to the technical field of computer image processing, and discloses a residual mixed attention denoising method, system and device for LDCT and a medium, and the method comprises the steps: inputting a low-dose computed tomography slice into a binary parallel convolution module for feature extraction, and generating a first feature map; performing feature extraction on the first feature map through a first convolution module to obtain a second feature map; extracting features of the second feature map by using a residual mixed attention group, the residual mixed attention group comprising a plurality of residual mixed attention modules, and generating a third feature map after processing; inputting the third feature map into a second convolution module for image reconstruction to obtain a fourth feature map; and adding the low-dose computed tomography slice and the fourth feature map to obtain a final de-noised image. Experiments prove that the method shows excellent denoising performance on a plurality of data sets, more reliable image basis is provided for clinical diagnosis, and the radiation risk of patients is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the field of computer image processing technology, and in particular to a residual mixed attention denoising method, system, device and medium for LDCT. Background Art

[0002] Computed tomography (CT) technology, as an important means of modern medical imaging diagnosis, provides high-resolution cross-sectional anatomical images through X-ray imaging, playing an irreplaceable role in clinical applications such as disease diagnosis, tumor screening, and surgical planning. However, the ionizing radiation associated with CT examinations may pose a potential risk of biological damage to patients, especially those with chronic diseases and children who require multiple follow-up examinations. To reduce radiation risks, low-dose CT (LDCT) technology has emerged, which reduces radiation exposure by reducing the X-ray dose. However, the dose reduction inevitably leads to a significant increase in noise and artifacts in the image, seriously affecting image quality and diagnostic accuracy.

[0003] Traditional LDCT image processing methods fall into three main categories: algorithms based on iterative reconstruction, sinusoidal filtering, and image post-processing techniques. While these methods can improve image quality in specific scenarios, due to the non-uniform distribution and non-stationary nature of LDCT noise, as well as its high coupling with anatomical structures, traditional methods often suffer from problems such as oversmoothing leading to loss of detail or incomplete denoising, making them difficult to meet the image quality requirements of clinical diagnosis.

[0004] In recent years, deep learning technology has shown great potential in the field of medical image processing. Methods based on convolutional neural networks have made significant progress in LDCT denoising tasks by extracting features through local receptive fields. Architectures such as residual networks, densely connected networks, and U-Net have effectively alleviated the gradient vanishing problem of deep networks by introducing mechanisms such as skip connections and dense connections. However, the inherent local limitations of convolutional neural networks make it difficult to model long-range image dependencies, and their performance is limited when dealing with large-scale structured noise. The Transformer architecture has advantages in global feature modeling due to its self-attention mechanism, but its direct application in medical image processing faces challenges such as high computational complexity and insufficient preservation of local details.

[0005] Current research trends focus on the fusion architecture of CNN and Transformer, aiming to balance local feature extraction and global context modeling. However, existing hybrid methods still have the following key issues: (1) the multi-scale feature fusion mechanism is not efficient enough, making it difficult to balance the importance of features at different resolutions; (2) the computational overhead of the attention mechanism is high, making it difficult to achieve efficient inference while maintaining performance; and (3) the noise-structure coupling characteristics unique to LDCT images are not adequately modeled. These issues limit the generalization ability of existing methods in complex clinical scenarios, making it difficult to achieve an ideal balance between maintaining anatomical integrity and removing noise.

[0006] In response to the above technical bottlenecks, it is urgent to develop new LDCT image denoising methods. By innovatively combining residual learning, hybrid attention mechanism and multi-scale feature fusion technologies, more accurate noise modeling and more efficient feature utilization can be achieved, thereby significantly reducing the radiation dose while obtaining high-quality CT images that meet clinical diagnostic needs. Summary of the Invention

[0007] In view of the above existing problems, the present invention is proposed.

[0008] Traditional convolutional neural network methods are limited by their local receptive fields and have difficulty modeling global dependencies. Transformer methods have high computational complexity and insufficient preservation of local details. Furthermore, existing hybrid architectures still lack efficiency in multi-scale feature fusion and noise-structure decoupling. Therefore, the present invention provides a residual hybrid attention denoising method, system, device, and medium for LDCT to address the problems of low contrast, high noise, and low resolution in low-dose computed tomography images.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] In a first aspect, the present invention provides a residual hybrid attention denoising method for LDCT, comprising:

[0011] The low-dose computed tomography slices are input into a binary parallel convolution module for feature extraction to generate a first feature map;

[0012] Performing feature extraction on the first feature map through a first convolution module to obtain a second feature map;

[0013] Extracting features of the second feature map using a residual mixed attention group, wherein the residual mixed attention group includes a plurality of residual mixed attention modules, and generating a third feature map after processing;

[0014] Inputting the third feature map into the second convolution module for image reconstruction to obtain a fourth feature map;

[0015] The low-dose computed tomography slice is added to the fourth feature map to obtain a final denoised image.

[0016] As a preferred solution of the residual hybrid attention denoising method for LDCT according to the present invention, the generating of the first feature map includes:

[0017] Low-dose computed tomography slices are input into a binary parallel convolution module and extracted in parallel using multi-scale filters with different convolution kernel sizes to obtain the output of each layer.

[0018] The output result of each layer is added to the low-dose computed tomography slice to obtain a first feature map.

[0019] As a preferred solution of the residual hybrid attention denoising method for LDCT described in the present invention, wherein: the residual hybrid attention module includes two multi-scale feature extraction modules, a convolution module with a LeakyReLU activation function and a residual connection, wherein the multi-scale feature extraction module consists of a first convolution activation module, a global-local feature extraction block, an adaptive feature fusion module and a first pooling module.

[0020] As a preferred solution of the residual hybrid attention denoising method for LDCT according to the present invention, the processing of the multi-scale feature extraction module includes:

[0021] The first layer data flow includes a first convolution activation module, a global-local feature extraction block, and an adaptive feature fusion module. The first convolution activation module extracts shallow information of the second feature map. The global-local feature extraction block uses a multi-head attention strategy to extract global features and local features. The adaptive feature fusion module fuses the current layer features with the next layer features and outputs an adjusted multi-scale feature map.

[0022] The second layer of data flow includes a first pooling module, a global-local feature extraction block, and an adaptive feature fusion module. The first pooling module reduces the size of the feature map through a pooling operation;

[0023] The third layer of data flow includes the first pooling module and the global-local feature extraction block, which further reduces the size of the feature map to capture more global feature information;

[0024] The output feature maps of the three data streams are fused through residual connections to generate the final output feature map of the multi-scale feature extraction module.

[0025] As a preferred solution of the residual hybrid attention denoising method for LDCT described in the present invention, the processing of the global-local feature extraction block includes:

[0026] The input feature x is adjusted by convolution and activation function to obtain the feature map z. l-1 ;

[0027] The feature map z l-1 Through layer normalization, window multi-head self-attention module and residual connection processing, the feature map is obtained The window multi-head self-attention module extracts local features of the CT image by calculating self-attention within the window; the feature map The feature map z is obtained by layer normalization, multi-layer perceptron and residual connection processing l ;

[0028] The feature map z l After layer normalization, variable window multi-head self-attention module and residual connection processing, the feature map z is obtained by layer normalization, multi-layer perceptron and residual connection processing. l+1 , wherein the variable window multi-head self-attention module extracts the global features of the CT image through a variable window operation;

[0029] By l-1 , feature map z l+1 Apply residual connection to the input feature x to obtain the final output feature.

[0030] As a preferred solution of the residual hybrid attention denoising method for LDCT described in the present invention, the processing of the window multi-head self-attention module includes:

[0031] The input feature map is divided into four regular and non-overlapping windows, and self-attention is calculated independently in each window. The formula defines how the attention weight is calculated for each divided window;

[0032] The softmax function is used to normalize the dot product between the features after the linear transformation of the feature map to obtain the filtered local features.

[0033] As a preferred solution of the residual hybrid attention denoising method for LDCT described in the present invention, the processing of the adaptive feature fusion module includes:

[0034] Add the input feature maps of the two resolution data streams pixel by pixel to generate the output feature map;

[0035] To aggregate the feature maps in each channel, the output feature maps are processed by applying maximum pooling and average pooling at the same time to obtain channel statistics s1 and s2;

[0036] The channel statistics s1 and s2 are added together and then passed through a 1×1 convolution and activated using the LeakyReLU activation function. Then two parallel channel upsampling convolution layers and the softmax function are applied to obtain the attention activation value. and

[0037] The attention activation value and Multiply the original input feature map channel by channel to obtain two channel attention feature maps, and then add the two channel attention feature maps to obtain the final output feature map.

[0038] In a second aspect, the present invention provides a residual hybrid attention denoising system for LDCT, comprising:

[0039] a first extraction unit, configured to input the low-dose computed tomography slice into a binary parallel convolution module for feature extraction to generate a first feature map;

[0040] A second extraction unit, configured to perform feature extraction on the first feature map through a first convolution module to obtain a second feature map;

[0041] a third extraction unit, configured to extract features of the second feature map using a residual mixed attention group, wherein the residual mixed attention group includes a plurality of residual mixed attention modules, and generates a third feature map after processing;

[0042] a fourth extraction unit, configured to input the third feature map into the second convolution module for image reconstruction to obtain a fourth feature map;

[0043] A denoising output unit is configured to add the low-dose computed tomography slice to the fourth feature map to obtain a final denoised image.

[0044] In a third aspect, the present invention provides an electronic device comprising a memory and a processor; the memory is used to store computer-executable instructions, and the processor implements the steps of a residual mixed attention denoising method for LDCT when executing the computer-executable instructions.

[0045] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of a residual hybrid attention denoising method for LDCT.

[0046] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides a residual hybrid attention denoising method, system, device and medium for LDCT, which effectively solves the key technical difficulties in low-dose CT image denoising through an innovative multi-scale feature extraction mechanism and adaptive feature fusion strategy. After extracting shallow features using a binary parallel convolution module, the present invention uses the multi-scale feature extraction module in the residual hybrid attention group to coordinate the global-local feature extraction block and the adaptive feature fusion module to achieve adaptive fusion of multi-scale complementary features while retaining the original high-resolution features. The global-local feature extraction block achieves the continuity of local feature extraction and contextual information through an attention mechanism based on continuous window and variable window technology, thereby compensating for the limitation of convolutional neural networks that only extract local features. The adaptive feature fusion module uses a self-attention mechanism to adaptively select and fuse key information from information streams of different resolutions. Experiments have shown that the method provided by the present invention exhibits excellent denoising performance on multiple datasets such as Mayo, LDCT-and-Projection-data and Piglet, not only significantly improving image quality but also having excellent generalization ability. This technological breakthrough not only provides a more reliable imaging basis for clinical diagnosis, but also effectively reduces the radiation risk of patients (radiation dose can be reduced by 50-80%). It is of great significance to promote the development of precision medicine. At the same time, its modular design facilitates integration into existing medical imaging systems, and has important clinical application value and industrial prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 This is a logic diagram of the overall process of the residual hybrid attention denoising method for LDCT according to an embodiment of the present invention.

[0049] Figure 2 This is a diagram of the overall network architecture of the residual hybrid attention denoising method for LDCT described in one embodiment of the present invention.

[0050] Figure 3 Schematic diagram of a binary parallel convolution module of a residual hybrid attention denoising method for LDCT according to an embodiment of the present invention.

[0051] Figure 4Schematic diagram of a multi-scale feature extraction module (MSFE) of a residual hybrid attention denoising method for LDCT according to an embodiment of the present invention.

[0052] Figure 5 This is a structural diagram of the global-local feature extraction block (GLB) of the residual hybrid attention denoising method for LDCT described in one embodiment of the present invention.

[0053] Figure 6 Schematic diagram of the adaptive feature fusion module (AFF) of the residual hybrid attention denoising method for LDCT according to an embodiment of the present invention.

[0054] Figure 7 This is a denoised image of various denoising algorithms of the residual mixed attention denoising method for LDCT described in an embodiment of the present invention, where LDCT represents a noisy computed tomography image, NDCT represents a normal-dose CT image, and the rest are CT images denoised by various denoising algorithms.

[0055] Figure 8 The residual hybrid attention denoising method for LDCT according to an embodiment of the present invention is Figure 5 Schematic diagram of the corresponding ROI3 in .

[0056] Figure 9 This is a visualization of the residual hybrid attention denoising method for LDCT described in one embodiment of the present invention, on the LDCT-and-Projection-data dataset, including low-dose computed tomography images of the head, low-dose computed tomography images denoised by different methods, and normal-dose computed tomography (NDCT) images.

[0057] Figure 10 Schematic diagram of denoising results of low-dose computed tomography images of the Piglet dataset using different methods of the residual hybrid attention denoising method for LDCT described in one embodiment of the present invention (the selected area is magnified). DETAILED DESCRIPTION

[0058] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0059] Example 1, with reference to Figures 1-6As an embodiment of the present invention, a residual hybrid attention denoising method for LDCT is provided, such as Figure 1 The specific steps shown include:

[0060] S100: Inputting the low-dose computed tomography slice into a binary parallel convolution module for feature extraction to generate a first feature map;

[0061] S200: performing feature extraction on the first feature map through a first convolution module to obtain a second feature map;

[0062] S300: extracting features of the second feature map using a residual mixed attention group, where the residual mixed attention group includes multiple residual mixed attention modules, and generating a third feature map after processing;

[0063] S400: Inputting the third feature map into the second convolution module for image reconstruction to obtain a fourth feature map;

[0064] S500: Adding the low-dose computed tomography slice to the fourth feature map to obtain a final denoised image.

[0065] It should be noted that the key technical difficulties in low-dose CT image denoising are effectively solved through the innovative multi-scale feature extraction mechanism and adaptive feature fusion strategy. After the present invention uses a binary parallel convolution module to extract shallow features, the multi-scale feature extraction module in the residual mixed attention group cooperates with the global-local feature extraction block and the adaptive feature fusion module to achieve adaptive fusion of multi-scale complementary features while retaining the original high-resolution features. The global-local feature extraction block achieves the continuity of local feature extraction and contextual information through the attention mechanism based on continuous window and variable window technology, thereby making up for the limitation of convolutional neural networks that only extract local features. The adaptive feature fusion module adaptively selects and fuses key information from information streams of different resolutions through the self-attention mechanism. Experiments have shown that the method provided by the present invention exhibits excellent denoising performance on multiple data sets such as Mayo, LDCT-and-Projection-data and Piglet. It can not only significantly improve image quality, but also has excellent generalization ability. This technological breakthrough not only provides a more reliable imaging basis for clinical diagnosis, but also effectively reduces the radiation risk of patients (radiation dose can be reduced by 50-80%). It is of great significance to promote the development of precision medicine. At the same time, its modular design facilitates integration into existing medical imaging systems, and has important clinical application value and industrial prospects.

[0066] like Figure 2 The overall network architecture diagram of the embodiment of the present invention is shown below. Figure 1 and Figure 2The specific implementation of the residual hybrid attention denoising method for LDCT provided by the present invention is introduced.

[0067] In the embodiment of the present application, the step S100 inputs the low-dose computed tomography slice into the binary parallel convolution module for feature extraction, and generates the first feature map including:

[0068] A series of continuous low-dose computed tomography slices are input into a binary parallel convolution module and extracted in parallel through multi-scale filters with different convolution kernel sizes to obtain the output results of each layer;

[0069] The output of each layer is added to the low-dose computed tomography slice to obtain the first feature map, which is the output feature map of the binary parallel convolution module.

[0070] It should be noted that the convolution kernel sizes of the multi-scale filter are set to 1, 3, 5, 7, and 9, respectively. Smaller convolution kernels are used to capture local details in medical images, including tiny structures and boundaries. Furthermore, to encompass a wider range of anatomical structures, such as organs or lesion areas, this embodiment uses larger convolution kernels. These larger convolution kernels play a crucial role by expanding the receptive field and providing a wider range of contextual information.

[0071] Specifically, the schematic diagram of the binary parallel convolution module (DPC module) is as follows Figure 3 As shown in Figure 1, the network consists of five parallel structures with dense connections and residual connections. To calculate the intermediate output of each layer, x is used as the initial input and passed through the five parallel structures. In the first row, two convolutions with a kernel of 1×1 are used, and the ReLU activation function is used in the middle. The formula is expressed as:

[0072] Z1=Conv 1×1 (x)

[0073] Among them, Z1 represents the feature map after the first convolution, and the feature map is activated by the ReLU function. The formula is as follows:

[0074]

[0075] Among them, Z1′ represents the feature map after the ReLU activation function, and then passes through a 1×1 to obtain the output Z1″ of this layer. The formula is as follows:

[0076] Z1″=Conv 1×1 (Z1′)

[0077] Combining the above formula, we can get:

[0078]

[0079] From this, the output of each layer can be expressed as the following formula:

[0080]

[0081] Among them, the value of i is i∈{1,2,3,4,5}, Z i ″ is the output of each layer, (2i-1)(2i-1) represents the convolution kernel size;

[0082] The output of each layer above is added to the low-dose computed tomography slice to obtain the output feature map of the module, which is expressed as:

[0083]

[0084] Where y represents the output feature map of the DPC module.

[0085] It should be noted that step S100 not only processes long-range spatial information, but also captures and fuses spatial information at finer scales, thereby expanding the receptive field and achieving better aggregation of local information. This design network can extract features at multiple scales while preserving input details, further improving denoising performance.

[0086] In the embodiment of the present application, step S200 extracts features from the first feature map using the first convolution module to obtain the second feature map, including:

[0087] The first convolution module includes a convolution (Conv) and a LeakyReLU activation function. The first feature map is input into the first convolution module to further extract features and obtain the second feature map, which is the output feature map of the first convolution module. The formula is expressed as:

[0088] Z=Conv(I)

[0089]

[0090] Among them, Z represents the feature after convolution, A represents the feature map after the LeakyReLU activation function, and α is a small positive number that represents the slope when the input is negative.

[0091] In the embodiment of the present application, the above step S300 uses the residual mixed attention group to extract the features of the second feature map. The residual mixed attention group includes multiple residual mixed attention modules. After processing, the third feature map generated includes:

[0092] The second feature map is passed through a residual hybrid attention (RHA) group. Each RHA contains two multi-scale feature extraction modules (MSFE), a convolution module with a LeakyReLU activation function, and a residual connection. The formula can be expressed as:

[0093] RHA1=A+LeakyReLU(Conv(MSFE2(MSFE1(A))))

[0094] Among them, MSFE1(A) represents the output of the first MSFE module, MSFE2(MSFE1(A)) represents the output of the second MSFE module, and RHA1 represents the feature output of the first RHA module. The RHA group consists of 4 RHA modules, so the output feature map of the RHA group can be expressed as RHA4;

[0095] It should be noted that the core module of each RHA is MSFE, which is capable of generating outputs with spatial accuracy. MSFE achieves this by retaining high-resolution feature representations while absorbing a large amount of contextual data from low resolution. MSFE consists of three parallel attention streams. In each stream, the global-local feature extraction block (GLB) adopts window-based and variable window-based attention mechanisms to extract global and local features. In addition, by introducing the self-attention mechanism, the network is able to adaptively select and fuse key information from different resolution streams, thereby enabling contextual information to be transferred from low-resolution features to high-resolution features.

[0096] Specifically, the multi-scale feature extraction module consists of a first convolutional activation module, a global-local feature extraction block, an adaptive feature fusion module and a first pooling module, wherein the first convolutional activation module is a convolution, leaky ReLU module (CLR), and the first pooling module is a convolution, leaky ReLU, pooling module (CLRP).

[0097] Specifically, such as Figure 4As shown in the figure, the schematic diagram of the multi-scale feature extraction module (MSFE) consists of three parallel feature extraction streams and residual connections. The input feature map size is H, W, C, where H and W represent the height and width of the feature map, respectively, and C represents the number of channels. The first layer of the data stream includes the first convolutional activation module, the global-local feature extraction block, and the adaptive feature fusion module. The first convolutional activation module extracts shallow information from the second feature map. The global-local feature extraction block uses a multi-head attention strategy to extract global and local features. The adaptive feature fusion module fuses the current layer features with the next layer features and outputs the adjusted multi-scale feature map. The second layer of the data stream includes the first pooling module, the global-local feature extraction block, and the adaptive feature fusion module. The first pooling module reduces the size of the feature map through pooling operations. The third layer of the data stream includes the first pooling module and the global-local feature extraction block, further reducing the size of the feature map to capture more global feature information. The output feature maps of the three data streams are fused through residual connections to generate the final output feature map of the multi-scale feature extraction module.

[0098] It should be noted that the network architecture design of the multi-scale feature extraction module has significant advantages: on the one hand, the multi-scale feature extraction module can simultaneously capture global contextual information and local detail features in the image, thereby achieving comprehensive learning of features at different scales; on the other hand, with the help of the residual connection mechanism, it not only alleviates the gradient vanishing problem in deep networks, but also effectively retains the input low-level feature information, such as edges, textures and other key details. This multi-level feature fusion design strategy significantly enhances the network's representational capabilities, allowing the model to effectively remove noise during the denoising process while better preserving the image's detail information, thereby achieving a significant improvement in denoising performance. In addition, this design is also highly versatile and scalable, providing a good infrastructure for subsequent network optimization and improvement.

[0099] Furthermore, in the embodiments of the present application, Figure 5 The processing of the global-local feature extraction block (GLB) shown includes:

[0100] The input feature x is adjusted by convolution and activation function to obtain the feature map z. l-1 , the formula is:

[0101] z l-1 =LR(Conv(x))

[0102] Where LR(·) represents the LeakyReLU activation function operation;

[0103] Feature map z l-1 Through layer normalization, window multi-head self-attention module and residual connection processing, the feature map is obtained Among them, the window multi-head self-attention module extracts local features of the CT image and captures detailed information by calculating self-attention within the window. The formula is expressed as:

[0104]

[0105] Where LN(·) represents the layer normalization (LayerNorm) operation, and WM represents the window multi-head self-attention operation (W-MSA);

[0106] Feature Map The feature map z is obtained by layer normalization, multi-layer perceptron and residual connection processing l , the formula is:

[0107]

[0108] Where MLP(·) represents the fully connected layer operation, and LN(·) represents the layer normalization operation;

[0109] Feature map z l After layer normalization, variable window multi-head self-attention module and residual connection processing, the feature map z is obtained by layer normalization, multi-layer perceptron and residual connection processing. l+1 , where the variable window multi-head self-attention module extracts the global features of the CT image through variable window operation. The formula is expressed as:

[0110]

[0111] Where SM(·) represents the variable window multi-head self-attention operation (SW-MSA);

[0112] By l-1 , feature map z l+1 Apply residual connection to the input feature x to obtain the final output feature, the formula is:

[0113] y=z l+1 +x+z l-1

[0114] Where y represents the feature map output by the global-local feature extraction block.

[0115] It should be noted that the attention mechanisms based on windowed multi-head self-attention (W-MSA) and variable windowed multi-head self-attention (SW-MSA) are core components of the global-local feature extraction block (GLB). W-MSA efficiently extracts local features of CT images and captures detailed information by calculating self-attention within the window. SW-MSA, on the other hand, breaks the limitations of fixed windows through variable window operations and enhances the feature connections between different windows, thus playing an important role in extracting global features. This design, which combines local and global feature extraction, enables the global-local feature extraction block to simultaneously capture both detailed information and the overall structure of the image, significantly improving the model's ability to understand CT images.

[0116] Furthermore, the processing of the window multi-head self-attention module includes:

[0117] The input feature map is divided into four regular and non-overlapping windows (a, b, c, and d), and self-attention is calculated independently in each window. The formula defines how the attention weight of each divided window is calculated, specifically:

[0118]

[0119] Among them, q, k represent the features of the feature map after linear transformation. T The dot product of can be used as an indicator of their similarity. However, this similarity lacks normalization. Therefore, the softmax function is used to normalize the dot product between the features after the linear transformation of the feature map. The output of the softmax function is a matrix with all values ​​ranging from 0 to 1. Multiplying this matrix with v yields the filtered local features. The formula is:

[0120]

[0121] It should be noted that this operation automatically captures key information from the feature map. Although windowed multi-head self-attention (W-MSA) reduces computational overhead by dividing the windows, its receptive field can be considered reduced due to the lack of information exchange between different windows. This limitation hinders the acquisition of global feature information. To facilitate information exchange between different windows, cyclic shifting is used to address this issue. Smaller windows are shifted into other windows, forming four regular windows. For example, windows I, G, C, and A form a regular window. Multi-head self-attention (MSA) can be computed within these regular windows. However, since some regular windows are not adjacent in the original feature map after cyclic shifting, MSA cannot be computed between them. Therefore, it is necessary to use masked multi-head self-attention (masked MSA) on the newly formed windows and then shift them back to their original positions after the calculation is complete. This approach achieves continuity between local feature extraction and contextual information, thus overcoming the limitation of convolutional neural networks (CNNs) that can only extract local features. Furthermore, by stacking multiple windowed self-attention layers, the network can gradually expand the receptive field, ultimately achieving global information modeling. This design not only improves computational efficiency, but also enhances the model's ability to capture the global structure and local details of the image.

[0122] Further, such as Figure 6 The processing of the adaptive feature fusion module shown includes:

[0123] The input feature maps L1 and L2 of the two resolution data streams are added pixel by pixel to generate the output feature map L in ;

[0124] To aggregate the feature maps in each channel, the maximum pooling and average pooling are applied to the output feature maps to obtain channel statistics s1 and s2;

[0125] The channel statistics s1 and s2 are added together and then passed through a 1×1 convolution and activated using the LeakyReLU activation function. Then two parallel channel upsampling convolution layers and the softmax function are applied to obtain the attention activation value. and

[0126] Attention activation value and Multiply the original input feature map channel by channel to obtain two channel attention feature maps, and then add the two channel attention feature maps to obtain the final output feature map. The formula is:

[0127]

[0128] It should be noted that by introducing learnable weight parameters, different importance weights can be adaptively assigned to each feature channel. By analyzing the global contextual information of the feature maps in different channels, the contribution of each channel is dynamically adjusted: feature channels containing important structural information are given higher weights, while channels dominated by noise are suppressed. Through this feature fusion strategy, the network can effectively suppress noise while maximizing the preservation of structural details that are critical for diagnosis, significantly improving the clinical usability of denoised images.

[0129] In the embodiment of the present application, the above step S400 inputs the third feature map into the second convolution module for image reconstruction, and the fourth feature map obtained includes:

[0130] The second convolution module includes a DPC module, a transposed convolution operation, and a LeakyReLU activation function;

[0131] The processed feature RHA4 is then input into the DPC module for further feature extraction to generate the feature map Z;

[0132] The feature map Z output by the DPC module is reconstructed through a transposed convolution operation and a LeakyReLU activation function to generate the fourth feature map.

[0133] In the embodiment of the present application, the above step S500 adds the low-dose computed tomography slice to the fourth feature map to obtain the final denoised image, which is expressed as follows:

[0134]

[0135] Among them, ConvT represents the transposed convolution (ConvTranspose) operation, Represents the output denoised image.

[0136] Furthermore, the mean square error (MSE) loss is used to optimize the proposed network, which is expressed as follows:

[0137]

[0138] Among them, Y i represents the NDCT image, Represents the LDCT image after denoising.

[0139] As can be seen from the above embodiments, this embodiment provides a residual hybrid attention denoising method, system, device, and medium for LDCT. It utilizes a binary parallel convolution module to extract shallow-level information from LDCT images. Considering the low contrast, high noise, and low resolution characteristics of LDCT images, a residual hybrid attention group is proposed. The core of the residual hybrid attention module is a multi-scale feature extraction module, which generates complementary features at multiple scales while preserving the original high-resolution features. In the MSFE module, a designed global-local feature extraction block and a novel adaptive feature fusion module work together to effectively extract image features. The global-local feature extraction block uses an attention mechanism based on continuous and variable window techniques to achieve continuity between local feature extraction and contextual information, thereby overcoming the limitation of convolutional neural networks that only extract local features. Furthermore, the adaptive feature fusion module uses a self-attention mechanism to adaptively select and fuse key information from information streams at different resolutions. Experiments demonstrate that the proposed RHA-Net outperforms existing methods in content preservation and denoising. This not only improves the quality of LDCT images but also enhances clinical practice and protects the health of patients and medical staff.

[0140] Example 2, reference Figures 7 to 10 Based on the previous embodiment, this embodiment provides an application example of the residual hybrid attention denoising method, system, device and medium for LDCT to verify and illustrate the technical effects adopted in this method.

[0141] To fairly compare existing low-dose CT (LDCT) denoising algorithms, this example uses the 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge dataset (Mayo dataset), the LDCT-and-Projection-data dataset, and the Piglet dataset. The Mayo dataset is from the 2016 National Institutes of Health-American Association of Physicists in Medicine-Mayo Clinic Low Dose CT Challenge and is licensed from the Mayo Clinic. This dataset contains standard-dose CT (NDCT) images and their corresponding low-dose CT (LDCT) images, all with a slice thickness of 3 mm. For training and validation, 2167 image pairs were selected from 9 patients. In addition, data from one patient (L506) was reserved for testing, providing a total of 211 image pairs. To further validate the practicality of the model proposed in this example, the LDCT-and-Projection-data dataset and the Piglet dataset were used as additional test sets. In the experiment, 793 pairs of piglet images were selected, with a radiation dose of 300 mAs as the standard dose and 150 mAs as the low dose. A test set of 477 pairs of head CT images was selected from the LDCT-and-Projection-data dataset. The LDCT-and-Projection-data dataset was also provided by the Mayo team, but its imaging protocol differs from that of the Mayo dataset. Due to differences in imaging protocols, denoising LDCT images with varying noise characteristics is a challenging task. The piglet dataset was acquired using a GE scanner (Discovery CT750 HD) and contains images from a range of doses. The dataset is provided in DICOM format.

[0142] Evaluation metrics include peak signal-to-noise ratio (PSNR), structural similarity (SSIM), feature similarity (FSIM), and root mean square error (RMSE). PSNR measures the quality of the denoised image; SSIM assesses image similarity based on brightness, contrast, and structure, with a value range of [0, 1]. FSIM evaluates images based on phase congruence and gradient magnitude, with higher values ​​indicating greater feature similarity and better quality. RMSE is extremely sensitive to the error between each pair of pixels in the image; lower values ​​indicate greater similarity between the two images. Generally speaking, higher PSNR, SSIM, and FSIM values, as well as lower RMSE values, indicate better performance.

[0143] This example experiment uses an end-to-end denoising architecture without any pre-training. The model is trained using the Adam optimizer, with an initial learning rate of 2×10 -4 During the training process, the cosine annealing strategy is applied to slowly reduce the learning rate from the initial value to 1×10 -6 . In addition, a progressive training method is adopted. In the early stage of training, the network is trained with smaller image patches, and gradually transitions to larger image patches as training progresses. The image sizes are [128, 160, 192, 256, 320, 364], and the corresponding number of iterations is [48000, 24000, 24000, 24000, 24000, 24000]. To prevent overfitting and enhance the robustness of the model, the data is flipped horizontally and vertically to achieve data augmentation, and finally tested on (512×512) images.

[0144] This example compares eight mainstream algorithms from both quantitative and qualitative aspects. These algorithms are: RED-CNN, FourLLIE, Noise2Sim, U-Net-MA, CTformer, Hformer, ASCON, and MWDCNN. First, internal testing is performed on the Mayo dataset, and L506 is selected as the test set. The test set mainly contains 211 pairs of 3mm slice CT images. Secondly, in order to test the performance of the RHA-Net method on different datasets, 477 pairs of head CT images are selected from the LDCT-and-Projection-data dataset, and 793 pairs of images are selected from the Piglet dataset for external testing. It is worth noting that these datasets used for external testing did not appear in the training process.

[0145] ①Mayo dataset results:

[0146] Quantitative Evaluation: To quantitatively evaluate the denoising performance of the proposed RHA-Net method and eight comparison algorithms, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), feature similarity index (FSIM), and root mean square error (RMSE) were used as evaluation metrics. A summary of the comparison of different algorithms on the Mayo dataset is shown in Table 1. Experimental results are presented as a±b, where a represents the mean and b represents the standard deviation. Results highlighted in bold indicate the best performance among all methods. The results show that the method provided in this embodiment significantly outperforms the compared denoising algorithms. For example, RHA-Net achieves an average PSNR that is approximately 4.67% higher than RED-CNN and approximately 3.00% higher than Noise2Sim. In addition, RHA-Net achieves an average SSIM of 0.9339, which is approximately 0.2% to 1.2% higher than the other methods, reflecting its better ability to preserve image structural similarity. RHA-Net achieves an average FSIM of 0.7684, which is approximately 0.5% to 3.2% higher than the other algorithms, demonstrating its superior feature preservation capabilities. Furthermore, the RMSE of RHA-Net is 5.3150, which is approximately 0.9% to 15.6% lower than that of other methods, meaning that the reconstruction error is extremely small. Overall, RHA-Net performs well in terms of image quality, structure preservation, feature preservation, and error minimization.

[0147] Table 1: Comparison with different algorithms on the MAYO dataset (bold represents the best performance)

[0148]

[0149]

[0150] Qualitative evaluation: In order to evaluate the effectiveness of different denoising algorithms, a sample is randomly selected from the test set for analysis. Figure 7Images processed by various denoising algorithms are shown. In low-dose computed tomography (LDCT) images, noise affects the boundaries of organs with similar density and tissue characteristics, making it difficult for physicians to accurately assess organ morphology and structural changes. Different denoising algorithms reduce noise to varying degrees, making organ boundaries more distinct. However, each algorithm differs in its ability to process detailed information. For example, the Hformer method processes images by dividing them into non-overlapping blocks. While it achieves significant denoising results, it fails to effectively maintain continuity between blocks, resulting in noticeable grid-like artifacts at image boundaries (indicated by the purple arrows). Overall, the RED-CNN, CTformer, and U-Net-MA methods also achieve some noise reduction, but they can still produce patchy artifacts that obscure image detail. Detailed imaging of blood vessels is crucial in abdominal CT scans, as it clearly demonstrates the integrity of vascular structures and blood flow, which is essential for assessing organ health and diagnosing vascular diseases. However, methods such as Noise2Sim and MWDCNN result in noticeable smoothing of vascular structures in the spleen (indicated by the red arrows). For the area indicated by the yellow box, among all the denoising algorithms, the result obtained by the method provided in this embodiment is closest to the result of normal dose computed tomography (NDCT).

[0151] In addition, if Figure 7 ROI 3 Figure 8 As shown in Figure 3, the MWDCNN and RHA-Net methods perform well in maintaining vascular continuity, while other denoising algorithms lead to varying degrees of vascular discontinuity, which may be misjudged as pathological changes such as vascular occlusion or stenosis. Figure 8 This may affect diagnosis and treatment decisions. In addition to vascular analysis, such as Figure 8 As shown by the green arrows in the middle, normal-dose CT scans clearly show regular, uniform low-density lesions, while the noise from low-dose CT scans blurs the boundaries of these lesions. Among different denoising algorithms, the method in this embodiment can better reconstruct the overall structure of the lesion.

[0152] ②LDCT-and-projection-data dataset results:

[0153] Quantitative Evaluation: Results on low-dose computed tomography (LDCT) and projection data datasets demonstrate that RHA-Net outperforms other algorithms across all evaluation metrics. Specifically, RHA-Net achieves an average peak signal-to-noise ratio (PSNR) of 44.1536, approximately 10.3% higher than RED-CNN, highlighting its superior image quality restoration capabilities. Its average structural similarity index (SSIM) is 0.9826, approximately 0.6% higher than RED-CNN, demonstrating its superior ability to preserve image structural similarity. RHA-Net's average feature similarity index (FSIM) is 0.8250, approximately 4.4% higher than FourLLIE, demonstrating its superior ability to preserve image features. Furthermore, RHA-Net achieves an average root mean square error (RMSE) of 1.6563, approximately 25.3% lower than FourLLIE, demonstrating minimal error in image reconstruction. Overall, RHA-Net demonstrates significant advantages over other methods in terms of image quality, structural similarity, feature preservation, and error minimization.

[0154] Table 2: Comparison with different algorithms on low-dose computed tomography (LDCT) and projection data (bold indicates best performance)

[0155]

[0156] Qualitative assessment: e.g. Figure 9As shown, compared to low-dose computed tomography (LDCT) images, the denoised images show significant improvements in various tissue structures. In particular, the areas highlighted by the red rectangles show significant improvements in detail and clarity. Zoomed-in images are shown above each figure. In images processed using the RED-CNN method, significant artifacts appear in the spinal cord region. These artifacts may obscure the true structure of the spinal cord or underlying lesions, compromising accurate assessment of the spinal cord and surrounding tissues, as indicated by the yellow arrows. In this head CT scan, the bright spots and radial streaks in the dental area are likely metal artifacts caused by metal objects worn by the patient, such as dental restorations. In low-dose CT scans, these metal artifacts are more pronounced because they amplify the noise present in the scan, further degrading image quality. Metal artifacts remain clearly visible in images processed using various denoising algorithms, such as FourLLE, CTformer, and Hformer, as indicated by the blue arrows. These artifacts can interfere with accurate diagnosis of soft tissue contrast or bone structure. In images processed using the Noise2Sim method, the cyan arrows indicate loss of detail in the vertebral protrusion, which can lead to diagnostic errors. In addition, as shown in the orange box, the muscle texture appears relatively smooth in the images processed by RED-CNN, Noise2Sim, and U-Net-MA. In contrast, the method in this embodiment can present clearer muscle boundaries and more obvious textures.

[0157] ③Piglet dataset results:

[0158] Quantitative Evaluation: As shown in Table 3, the RHA-Net algorithm performs exceptionally well in a comparison of the piglet dataset, significantly outperforming most of the eight state-of-the-art denoising models tested. Specifically, RHA-Net achieves a peak signal-to-noise ratio (PSNR) of 33.8745 dB, surpassing models such as RED-CNN (33.1468 dB, 2.2% lower), Hformer (33.2728 dB, 1.8% lower), and CTformer (33.6914 dB, 0.54% lower). In terms of the structural similarity index (SSIM), RHA-Net scores 0.9395, nearly on par with the best-performing U-Net-MA (0.9414), with a difference of only 0.2%. In addition, it also surpasses models such as CTformer (0.9343, 0.56% lower), FourLLIE (0.9368, 0.29% lower), and Noise2Sim (0.9369, 0.28% lower). Overall, the proposed algorithm performs well in terms of image quality and structural information preservation, achieving impressive results in low-dose CT image denoising.

[0159] Table 3: Comparison with different algorithms on the Piglet dataset. Bold represents the best (result), gray font represents the second best (result)

[0160] Method PSNR(a±b)↑ SSIM(a±b)↑ RED-CNN 33.1468±3.1111 0.9384±0.0287 FourLLIE 33.7157±3.4247 0.9368±0.0300 Noise2Sim 33.3896±3.0052 0.9369±0.0274 U-Net-MA 33.1133±3.3358 0.9414±0.0276 CTformer 33.6914±3.5282 0.9343±0.0309 Hformer 33.2728±3.0211 0.9387±0.0290 ASCON 33.6598±3.3393 0.9387±0.0291 MWDCNN 33.8627±3.4956 0.9388±0.0297 RHA-Net(ours) 33.8745±3.5511 0.9395±0.0294

[0161] Qualitative assessment: Figure 10 This paper demonstrates the exemplary results of the proposed method and eight state-of-the-art methods on a test image from the dataset. A low-dose computed tomography (LDCT) image is used as the network input, along with a corresponding normal-dose computed tomography (NDCT) image for comparison. It can be seen that methods such as RED-CNN, ASCON, FourLLIE, U-Net-MA, and Noise2Sim often lose image details and result in blurred edges. Hformer still produces artifacts. In contrast, RHA-Net and MWDCNN effectively remove most noise while preserving sufficient image texture.

[0162] As shown in the experimental results above, this embodiment first utilizes a binary parallel convolution module to extract shallow information from LDCT images. Considering the challenges faced by LDCT images, such as low contrast, high noise, and low resolution—factors that complicate feature extraction in medical images—a residual hybrid attention (RHA) network is proposed. The core of RHA is the multi-scale feature extraction (MSFE) module, which generates complementary features at multiple scales while preserving the original high-resolution features. Within the MSFE module, a designed global-local feature extraction block (GLB) and a novel adaptive feature fusion (AFF) module work together to effectively extract image features. The GLB utilizes an attention mechanism based on continuous and offset window techniques to achieve continuity between local feature extraction and contextual information, thus overcoming the limitation of convolutional neural networks that only extract local features. Furthermore, the AFF module uses a self-attention mechanism to intelligently select and fuse key information from information streams at different resolutions. Experiments demonstrate that the proposed RHA-Net outperforms existing methods in content preservation and noise reduction. This approach not only improves the quality of low-dose images but also protects the health of patients and medical staff by reducing radiation dose during CT imaging. In the future, the proposed RHA-Net network can be further improved and optimized to enhance its performance under different noise levels.

[0163] Example 3: This embodiment provides a residual hybrid attention denoising system for LDCT, including:

[0164] a first extraction unit, configured to input the low-dose computed tomography slice into a binary parallel convolution module for feature extraction to generate a first feature map;

[0165] A second extraction unit is used to perform feature extraction on the first feature map through the first convolution module to obtain a second feature map;

[0166] A third extraction unit is used to extract features of the second feature map using a residual mixed attention group, where the residual mixed attention group includes multiple residual mixed attention modules, and generates a third feature map after processing;

[0167] a fourth extraction unit, configured to input the third feature map into the second convolution module for image reconstruction to obtain a fourth feature map;

[0168] The denoising output unit is configured to add the low-dose computed tomography slice to the fourth feature map to obtain a final denoised image.

[0169] It should be noted that the technical solution of the residual mixed attention denoising system for LDCT and the technical solution of the residual mixed attention denoising method for LDCT mentioned above belong to the same concept. For the details not described in detail in the technical solution of the residual mixed attention denoising system for LDCT in this embodiment, please refer to the description of the technical solution of the residual mixed attention denoising method for LDCT mentioned above.

[0170] The above-mentioned unit modules can be embedded in or independent of the processor in the electronic device in the form of hardware, or can be stored in the memory of the electronic device in the form of software, so that the processor can call and execute the corresponding operations of the above-mentioned modules.

[0171] This embodiment also provides an electronic device, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor of the electronic device is configured to provide computing and control capabilities. The memory of the electronic device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The communication interface of the electronic device is configured to communicate with an external terminal via wired or wireless communication, where the wireless communication can be achieved via Wi-Fi, a carrier network, NFC (near field communication), or other technologies. When the computer program is executed by the processor, a residual mixed attention denoising method for LDCT is implemented. The display of the electronic device can be a liquid crystal display or an electronic ink display, and the input device of the electronic device can be a touch layer covering the display, or a key, trackball, or touchpad provided on the housing of the electronic device, or an external keyboard, touchpad, or mouse.

[0172] This embodiment further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method proposed in the above embodiment is implemented.

[0173] The storage medium proposed in this embodiment and the method proposed in the above embodiment belong to the same inventive concept. For technical details not fully described in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0174] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general-purpose hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling an electronic device (which can be a personal computer, server, or network device, etc.) to execute the method of the embodiment of the present invention.

[0175] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. The residual hybrid attention denoising method for LDCT is characterized by: include: The low-dose computed tomography slices are input into a binary parallel convolution module for feature extraction to generate a first feature map; Performing feature extraction on the first feature map through a first convolution module to obtain a second feature map; Extracting features of the second feature map using a residual mixed attention group, wherein the residual mixed attention group includes a plurality of residual mixed attention modules, and generating a third feature map after processing; Inputting the third feature map into the second convolution module for image reconstruction to obtain a fourth feature map; The low-dose computed tomography slice is added to the fourth feature map to obtain a final denoised image.

2. The residual hybrid attention denoising method for LDCT according to claim 1, wherein: Generating the first feature map includes: Low-dose computed tomography slices are input into a binary parallel convolution module and extracted in parallel using multi-scale filters with different convolution kernel sizes to obtain the output of each layer. The output result of each layer is added to the low-dose computed tomography slice to obtain a first feature map.

3. The residual hybrid attention denoising method for LDCT according to claim 2, wherein: The residual hybrid attention module includes two multi-scale feature extraction modules, a convolution module with a LeakyReLU activation function, and a residual connection, wherein the multi-scale feature extraction module consists of a first convolution activation module, a global-local feature extraction block, an adaptive feature fusion module, and a first pooling module.

4. The residual hybrid attention denoising method for LDCT according to claim 3, wherein: The processing of the multi-scale feature extraction module includes: The first layer data flow includes a first convolution activation module, a global-local feature extraction block, and an adaptive feature fusion module. The first convolution activation module extracts shallow information of the second feature map. The global-local feature extraction block uses a multi-head attention strategy to extract global features and local features. The adaptive feature fusion module fuses the current layer features with the next layer features and outputs an adjusted multi-scale feature map. The second layer of data flow includes a first pooling module, a global-local feature extraction block, and an adaptive feature fusion module. The first pooling module reduces the size of the feature map through a pooling operation; The third layer of data flow includes the first pooling module and the global-local feature extraction block, which further reduces the size of the feature map to capture more global feature information; The output feature maps of the three data streams are fused through residual connections to generate the final output feature map of the multi-scale feature extraction module.

5. The residual hybrid attention denoising method for LDCT according to claim 4, wherein: The processing of the global-local feature extraction block includes: The input feature x is adjusted by convolution and activation function to obtain the feature map z. l-1 ; The feature map z l-1 Through layer normalization, window multi-head self-attention module and residual connection processing, the feature map is obtained The window multi-head self-attention module extracts local features of the CT image by calculating self-attention within the window; the feature map The feature map z is obtained by layer normalization, multi-layer perceptron and residual connection processing l ; The feature map z l After layer normalization, variable window multi-head self-attention module and residual connection processing, the feature map z is obtained by layer normalization, multi-layer perceptron and residual connection processing. l+1 , wherein the variable window multi-head self-attention module extracts the global features of the CT image through a variable window operation; By l-1 , feature map z l+1 Apply residual connection to the input feature x to obtain the final output feature.

6. The residual hybrid attention denoising method for LDCT according to claim 5, wherein: The processing of the window multi-head self-attention module includes: The input feature map is divided into four regular and non-overlapping windows, and self-attention is calculated independently in each window. The formula defines how the attention weight is calculated for each divided window; The softmax function is used to normalize the dot product between the features after the linear transformation of the feature map to obtain the filtered local features.

7. The residual hybrid attention denoising method for LDCT according to claim 6, wherein: The processing of the adaptive feature fusion module includes: Add the input feature maps of the two resolution data streams pixel by pixel to generate the output feature map; To aggregate the feature maps in each channel, the output feature maps are processed by applying maximum pooling and average pooling at the same time to obtain channel statistics s1 and s2; The channel statistics s1 and s2 are added together and then passed through a 1×1 convolution and activated using the LeakyReLU activation function. Then two parallel channel upsampling convolution layers and the softmax function are applied to obtain the attention activation value. and The attention activation value and Multiply the original input feature map channel by channel to obtain two channel attention feature maps, and then add the two channel attention feature maps to obtain the final output feature map.

8. A residual hybrid attention denoising system for LDCT, applying the residual hybrid attention denoising method for LDCT according to any one of claims 1 to 7, characterized in that: include: a first extraction unit, configured to input the low-dose computed tomography slice into a binary parallel convolution module for feature extraction to generate a first feature map; A second extraction unit, configured to perform feature extraction on the first feature map through a first convolution module to obtain a second feature map; a third extraction unit, configured to extract features of the second feature map using a residual mixed attention group, wherein the residual mixed attention group includes a plurality of residual mixed attention modules, and generates a third feature map after processing; a fourth extraction unit, configured to input the third feature map into the second convolution module for image reconstruction to obtain a fourth feature map; A denoising output unit is configured to add the low-dose computed tomography slice to the fourth feature map to obtain a final denoised image.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the residual mixed attention denoising method for LDCT according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer executable instructions are executed by a processor, the steps of the residual hybrid attention denoising method for LDCT according to any one of claims 1 to 7 are implemented.

Images