A low-dose CT image denoising method, device, equipment and medium

Abstract

The low-dose CT image denoising method, device, equipment and medium provided by the embodiment of the application have the following beneficial effects: the application adopts a symmetrical encoder-decoder neural network model architecture, and combines a jump connection and a multi-scale attention mechanism function block to maintain the integrity of spatial features and enhance the feature representation capability. The function block enhances the denoising capability by combining a multi-scale convolution network module and a convolution block attention module with parallel spatial attention. The two modules work cooperatively to extract hierarchical features on multiple receptive fields and dynamically recalibrate the importance of channels and spaces, effectively improving the denoising precision of the low-dose CT image and enhancing the denoising effect.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method, apparatus, device, and medium for denoising low-dose CT images. Background Technology

[0002] Computed tomography (CT) is a crucial non-invasive medical imaging technique in clinical diagnosis. However, the ionizing radiation emitted by routine CT scans may pose a potential risk to patients' health, especially in cases requiring multiple or long-term follow-up examinations. To reduce radiation hazards, low-dose CT (LDCT) imaging technology has emerged. LDCT significantly reduces radiation dose by decreasing the amount of projection data acquired or lowering the tube current. However, this comes at the cost of a significant increase in noise in the reconstructed images, leading to decreased visual quality, blurred anatomical structures, and loss of detail, thus affecting diagnostic accuracy.

[0003] To overcome these limitations, researchers have introduced deep learning-based image denoising methods, particularly convolutional neural networks (CNNs). Early CNN-based LDCT denoising methods mostly employed an encoder-decoder structure, which introduced learning strategies into the encoder-decoder structure for denoising CT images. Although the algorithms have been improved, existing models still struggle to simultaneously capture both the details and overall features of an image, easily leading to the blurring or loss of fine structures, and have limited ability to preserve complex anatomical structures, making it difficult to completely retain anatomical information.

[0004] As can be seen from the above description, low-dose CT image denoising still faces difficulties in terms of insufficient accuracy and poor effect. How to improve the denoising accuracy and effect of low-dose CT images is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To overcome the shortcomings of existing low-dose CT image denoising methods in terms of poor accuracy, this invention proposes a denoising method, apparatus, device, and medium for low-dose CT images.

[0006] To achieve the above objectives, according to a first aspect of the present invention, an embodiment of the present invention provides a method for denoising low-dose CT images, the method comprising the following steps:

[0007] Low-dose CT images are input into an encoder, which consists of multiple stacked convolutional layers, and the multiple convolutional layers sequentially calculate the output feature map.

[0008] Multi-scale convolution is performed on the input feature map of each convolutional layer. The multi-scale convolution is processed in parallel using branches with different convolutional kernels to obtain multi-scale feature maps.

[0009] The multi-scale convolution includes four parallel branches:

[0010] The first scale branch processes the input feature map using a 1×1 convolution with ReLU activation.

[0011] The second scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and a 5×5 convolution with ReLU activation.

[0012] The third scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and two consecutive 3×3 convolutions with ReLU activation.

[0013] In the fourth scale branch, the input feature map is pooled.

[0014] The processing results of all branches are concatenated and multi-scale feature maps are obtained through 1×1 convolution.

[0015] Multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, and then spliced ​​and normalized to generate channel attention maps;

[0016] The channel attention map performs global average pooling and global max pooling along the channel dimension, and generates a spatial attention map by concatenation and normalization through parallel convolution branches.

[0017] The output feature map is input to the decoder, which is composed of multiple deconvolutional layers stacked sequentially, and the multiple deconvolutional layers are symmetrically arranged with the multiple convolutional layers; each deconvolutional layer performs residual calculation based on the calculation result of the upper deconvolutional layer and the spatial attention map input to the corresponding convolutional layer, and finally outputs a denoised image;

[0018] The neural network model is trained using loss functions from denoised and reference images. The trained neural network model is then deployed to denoise low-dose CT images.

[0019] Optionally, the multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, respectively, and then concatenated and normalized to generate channel attention maps, including:

[0020] The multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, respectively, to generate two compact channel feature descriptors;

[0021] The channel feature descriptors are respectively processed by two layers of multilayer perceptron with shared weights, and then fused by element-wise addition and normalized by the sigmoid function to generate channel attention vectors;

[0022] The channel attention vector is applied to the multi-scale feature map through channel multiplication to obtain the channel attention map.

[0023] Optionally, the channel attention map performs global average pooling and global max pooling along the channel dimension, and generates a spatial attention map through parallel convolution branches, concatenating and normalizing, including:

[0024] In the two parallel branches, global average pooling and global max pooling are performed along the channel dimension of the channel attention map to generate corresponding channel feature descriptors, which are then concatenated along the channel dimension to form the final descriptor.

[0025] Convolution calculations were performed on the final descriptors of each parallel branch using 3×3 and 7×7 convolution kernels, respectively.

[0026] The computation results of the two parallel branches are fused by element-wise addition, normalized, and processed by the sigmoid function to obtain the spatial attention vector;

[0027] The spatial attention vector is applied to the channel attention map through channel multiplication to obtain the spatial attention map.

[0028] Optionally, the residual calculation includes: establishing a skip connection between corresponding layers of the encoder and decoder using a residual compensation mechanism, embedding the skip connection and multi-scale attention function block between corresponding layers of the encoder and decoder to learn the residual function, and adding it to the encoder output feature map to achieve residual compensation.

[0029] According to a second aspect of the present invention, embodiments of the present invention also provide a noise reduction apparatus for low-dose CT images, comprising:

[0030] An encoder is used to receive low-dose CT images, and the encoder is composed of multiple stacked convolutional layers, which sequentially calculate the output feature map.

[0031] The skip connection and multi-scale attention functional blocks include a multi-scale convolutional network module and a convolutional block attention module with parallel spatial attention, wherein...

[0032] The multi-scale convolutional network module is used to perform multi-scale convolution on the input feature map of each convolutional layer. The multi-scale convolution uses branches with different convolutional kernels for parallel processing to obtain multi-scale feature maps.

[0033] The multi-scale convolution includes four parallel branches:

[0034] The first scale branch processes the input feature map using a 1×1 convolution with ReLU activation.

[0035] The second scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and a 5×5 convolution with ReLU activation.

[0036] The third scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and two consecutive 3×3 convolutions with ReLU activation.

[0037] In the fourth scale branch, the input feature map is pooled.

[0038] The processing results of all branches are concatenated and multi-scale feature maps are obtained through 1×1 convolution.

[0039] The convolutional block attention module for parallel spatial attention includes a channel attention module and a parallel spatial attention module. The channel attention module performs global average pooling and global max pooling on multi-scale feature maps along the spatial dimension, generating a channel attention map through concatenation and normalization. The parallel spatial attention module performs global average pooling and global max pooling on the channel attention maps along the channel dimension, generating a spatial attention map through parallel convolutional branches, followed by concatenation and normalization.

[0040] A decoder is used to receive the output feature map, and the decoder is composed of multiple deconvolutional layers stacked in sequence, the multiple deconvolutional layers being symmetrically arranged with the multiple convolutional layers; each deconvolutional layer performs residual calculation based on the calculation result of the upper deconvolutional layer and the spatial attention map input to the corresponding convolutional layer, and finally outputs a denoised image;

[0041] The training module is used to train a neural network model using a loss function on denoised images and reference images. The trained neural network model is then deployed to denoise low-dose CT images.

[0042] Optionally, the channel attention module performs global average pooling and global max pooling on the multi-scale feature maps in the spatial dimension, respectively, and generates a channel attention map by concatenation and normalization, including:

[0043] The multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, respectively, to generate two compact channel feature descriptors;

[0044] The channel feature descriptors are respectively processed by two layers of multilayer perceptron with shared weights, and then fused by element-wise addition and normalized by the sigmoid function to generate channel attention vectors;

[0045] The channel attention vector is applied to the multi-scale feature map through channel multiplication to obtain the channel attention map.

[0046] Optionally, the parallel spatial attention module performs global average pooling and global max pooling on the channel attention map along the channel dimension, and generates a spatial attention map through parallel convolution branches, concatenation, and normalization, including:

[0047] In the two parallel branches, global average pooling and global max pooling are performed along the channel dimension of the channel attention map to generate corresponding channel feature descriptors, which are then concatenated along the channel dimension to form the final descriptor.

[0048] Convolution calculations were performed on the final descriptors of each parallel branch using 3×3 and 7×7 convolution kernels, respectively.

[0049] The computation results of the two parallel branches are fused by element-wise addition, normalized, and processed by the sigmoid function to obtain the spatial attention vector;

[0050] The spatial attention vector is applied to the channel attention map through channel multiplication to obtain the spatial attention map.

[0051] Optionally, the residual calculation includes: establishing a skip connection between corresponding layers of the encoder and decoder using a residual compensation mechanism, embedding the skip connection and multi-scale attention function block between corresponding layers of the encoder and decoder to learn the residual function, and adding it to the encoder output feature map to achieve residual compensation.

[0052] According to a third aspect of the present invention, embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the low-dose CT image denoising method as described in any of the above embodiments.

[0053] According to a fourth aspect of the present invention, embodiments of the present invention also provide a storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the steps of the denoising method for low-dose CT images as described in any of the above embodiments.

[0054] As described above, the denoising method, apparatus, device, and medium for low-dose CT images provided by the embodiments of the present invention have the following beneficial effects: A low-dose CT image is input into an encoder, which is composed of multiple stacked convolutional layers, and the multiple convolutional layers sequentially calculate output feature maps; multi-scale convolution is performed on the input feature maps of each convolutional layer, and the multi-scale convolution uses branches with different convolutional kernels for parallel processing to obtain multi-scale feature maps; global average pooling and global max pooling are performed on the multi-scale feature maps in the spatial dimension, and channel attention maps are generated by splicing and normalization; the channel attention maps are then processed along... Global average pooling and global max pooling are performed along the channel dimension, and a spatial attention map is generated through parallel convolutional branches, concatenated, and normalized. The output feature map is input to the decoder, which consists of multiple deconvolutional layers stacked sequentially, symmetrically arranged with the convolutional layers. Each deconvolutional layer performs residual calculation based on the calculation results of the previous deconvolutional layer and the spatial attention map input to the corresponding convolutional layer, finally outputting a denoised image. The neural network model is trained using the loss function of the denoised image and the reference image, and the trained neural network model is used to deploy and denoise low-dose CT images. This invention employs a novel encoder-decoder neural network architecture, combined with skip connections and multi-scale attention mechanism functional blocks, to maintain the integrity of spatial features and enhance feature representation capabilities. The SCMA functional block combines a multi-scale convolutional network module (MSCN Block) with a convolutional block attention module (CBAM-PSA Block) that integrates channel attention and parallel spatial attention. The CBAM-PSA module can simultaneously extract features in both channel and spatial dimensions, effectively preserving the structural information of the image and significantly improving the denoising accuracy and structural fidelity of low-dose CT images. It addresses the receptive field adaptation defects in low-dose CT image denoising and the problem of not utilizing the differences in importance between feature channels and spatial locations in the denoising task. Through the collaborative work of the above modules, the model can fully extract hierarchical features across multiple receptive fields and dynamically recalibrate the importance of channels and space, overcoming the limitations of traditional methods. It can simultaneously extract hierarchical features across multiple receptive fields and achieve spatial feature recalibration from coarse to fine granular, taking into account both local details and global structure, while capturing local edge structures and global contextual information, effectively improving the denoising accuracy and enhancing the denoising effect of low-dose CT images. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a method for denoising low-dose CT images provided in an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the architecture of a neural network model provided in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of the structure of an SCMA functional block provided in an embodiment of the present invention;

[0058] Figure 4 This is a flowchart illustrating a multi-scale convolution method provided in an embodiment of the present invention;

[0059] Figure 5 This is a schematic diagram of a computational framework for multi-scale convolution provided in an embodiment of the present invention;

[0060] Figure 6 This is a schematic flowchart of a channel attention calculation method provided in an embodiment of the present invention;

[0061] Figure 7 This is a schematic diagram of a computational framework for channel attention and parallel spatial attention provided in an embodiment of the present invention;

[0062] Figure 8 This is a schematic diagram of a parallel spatial attention computation process provided in an embodiment of the present invention;

[0063] Figure 9 This is a schematic diagram of a computational architecture for residual calculation provided in an embodiment of the present invention;

[0064] Figure 10 This is a comparison image of low-dose CT image denoising results provided by an embodiment of the present invention;

[0065] Figure 11 This is a schematic diagram of the structure of a noise reduction device for low-dose CT images provided in an embodiment of the present invention;

[0066] Figure 12 This is a schematic diagram of the hardware structure of an electronic device for performing a method for denoising low-dose CT images, as provided in an embodiment of the present invention. Detailed Implementation

[0067] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0068] Please see Figures 1 to 12It should be noted that the illustrations provided in this embodiment are only schematic representations of the basic concept of the present invention. Therefore, the illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0069] See Figure 1 This is a flowchart illustrating a method for denoising low-dose CT images provided in an embodiment of the present invention. Figure 1 As shown in the figure, the embodiment of the present invention illustrates the flowchart of a method for denoising low-dose CT images.

[0070] Step S101: Input the low-dose CT image into the encoder, which is composed of multiple stacked convolutional layers, and the multiple convolutional layers sequentially calculate the output feature map.

[0071] First, data preparation is performed using low-dose CT (LDCT) images and corresponding conventional-dose CT (NDCT) images as paired training data. The data used to train the neural network model can be obtained clinically, for example, from LDCT and NDCT images of the same patient. Alternatively, since it is difficult to obtain both low-dose and conventional-dose CT images from the same patient simultaneously in clinical practice, they can be simulated using image processing techniques. The source of the original data is not limited in this embodiment of the invention. Subsequently, the LDCT images are preprocessed, including resizing and normalization. The resulting image pairs serve as input and output samples in supervised learning for training the model.

[0072] See Figure 2 This is a schematic diagram of the architecture of a neural network model provided in an embodiment of the present invention. Figure 2 The system architecture for implementing a low-dose CT image denoising method according to an embodiment of the present invention is shown, such as... Figure 2 As shown, the system is a symmetric encoder-decoder architecture, and its specific components are as follows:

[0073] The encoder, consisting of five fully connected convolutional units (Conv1 to Conv5 in the figure) and ReLU units stacked together, is used to extract hierarchical features and achieve spatial downsampling. This structure effectively balances noise suppression and feature preservation. Since this invention additionally introduces structures such as cascaded SCMA blocks (Skip-Connection and Multi-Scale Attention), adding more convolutional layers would significantly increase computational complexity. Considering the balance between performance and efficiency, this invention chooses to simplify the encoder-decoder network structure. The definition of a stacked encoder (SE) is as follows:

[0074] ,

[0075] in, For activation function, The number of convolutional layers. As weight, For bias, Represents the convolution operator. This is an image patch extracted from the input LDCT image. The stacked encoder outputs a one-dimensional spatial feature vector. This design allows the network to learn the spatial hierarchical structure of features, thereby achieving noise suppression at multiple abstraction levels. When each convolutional layer in the encoder processes an image patch, it can extract high-level features such as edges and textures while minimizing noise and artifacts. The feature extraction process can preserve important structural details while denoising. In practice, each convolutional layer uses 64 filters, a kernel size of 7×7, and an output feature map depth of 64. The stride is set to 1, and the padding mode is set to "Same" to ensure that the input size remains unchanged after the convolution operation. ReLU is used to introduce non-linearity, making it adaptable to complex data. Thus, the low-dose CT image is input to the first convolutional layer, and then processed by five convolutional layers in the encoder to finally obtain the output feature map.

[0076] The decoder consists of five stacked deconvolutional layers, each including deconvolutional units (Deconv1 to Deconv5 in the figure) and ReLU units. The deconvolutional layers are used to recover structural details from the extracted features, i.e., reconstruction. In this embodiment, both the encoder and decoder are symmetrical, with convolutional and deconvolutional layers using convolutional kernels of the same size. The encoder-decoder data flow employs a "reverse layer mapping" design, meaning the processing flow of the convolutional layers is matched to the corresponding layers of the decoder in reverse order. This is essentially determined by the symmetrical layout of the encoder and decoder; that is, the first convolutional layer of the encoder (Conv1 + ReLU) corresponds to the fifth deconvolutional layer of the decoder (Deconv5 + ReLU), and vice versa. In the decoder, the stacked decoder (SD) is defined as follows:

[0077] ,

[0078] in, The weights of the deconvolution layer, The output feature map is the input from the encoder. The bias of the deconvolution layer, The number of deconvolution layers. This represents the deconvolution operator, which is the inverse operation of the convolution operator. Thus, the output feature map input from the encoder is processed through five deconvolution layers to finally obtain the denoised CT image.

[0079] In addition, further integration Figure 3This is a schematic diagram of the structure of an SCMA functional block provided in an embodiment of the present invention. The system also involves an SCMA functional block, which further includes a multi-scale convolutional network module (MSCN Block, whose full English name is Multi-Scale Convolutional Network Block) and a convolutional block attention module with parallel spatial attention (CBAM-PSA Block in the figure, whose full English name is Convolutional Block AttentionModule with Parallel Spatial Attention Block). The convolutional block attention module with parallel spatial attention further includes a channel attention module (CAM, whose full English name is Channel Attention Module) and a parallel spatial attention module (PSAM, whose full English name is Parallel Spatial Attention Module). From the data path perspective, the input feature map from the encoder is processed sequentially by the multi-scale convolutional network module, the channel attention module, and the parallel spatial attention module. The calculation results are finally applied to the decoder. That is, the SCMA functional block realizes the jump connection mechanism from a specific convolutional layer in the encoder to a specific deconvolutional layer in the decoder, thereby realizing the overall denoising method flow of the embodiment of the present invention. Finally, the neural network model is trained end-to-end using the loss function Loss, and the trained neural network model is deployed to denoise low-dose CT images. The specific process described above will be described in detail in subsequent embodiments.

[0080] Step S102: Perform multi-scale convolution on the input feature map of each convolutional layer. The multi-scale convolution uses branches with different convolution kernels for parallel processing to obtain multi-scale feature maps.

[0081] Similarly combined Figure 2 and Figure 3 The low-dose CT image is input into the first convolutional layer of the encoder. Therefore, the input feature map of the first convolutional layer can be understood as the low-dose CT image. Subsequently, the input feature maps of the second to fifth convolutional layers can be understood as the calculation results of the previous convolutional layer. For the input feature map of each convolutional layer, multi-scale convolution (i.e., corresponding to...) is further performed. Figure 3 (Calculation process of the MSCN module in the middle).

[0082] Specifically, see Figure 4 This is a flowchart illustrating a multi-scale convolution method provided in an embodiment of the present invention, and is combined with... Figure 5This is a schematic diagram of a multi-scale convolution computation framework provided by an embodiment of the present invention. The diagram shows that the MSCN module uses parallel (1×1, 3×3, 5×5) convolution kernels and mean pooling to extract multi-scale features. These features are then concatenated, processed by the ReLU activation function, and refined by (1×1) convolution before being passed to the CBAM-PSA module. The detailed process is as follows:

[0083] Step S1021: In the first scale branch, the input feature map is processed by a (1×1) convolution with ReLU activation function.

[0084] Combination Figure 5 In the 1×1 branch, channel-level operations are performed using (1×1) convolutions with ReLU activation, while preserving spatial dimension information.

[0085] Step S1022: In the second scale branch, the input feature map is processed sequentially by (1×1) convolution with ReLU activation function and (5×5) convolution with ReLU activation function.

[0086] Similarly combined Figure 5 In the 5×5 branch, dimensionality reduction is first achieved through (1×1) convolution, followed by (5×5) convolution and ReLU activation to capture richer details.

[0087] The two parallel branch structures above follow the following convolution formula:

[0088] ,

[0089] in, The kernel size is the convolution kernel size. Introducing nonlinear characteristics, For the input feature map, The corresponding element at a specific location in the feature map. is the weight of the convolution kernel or filter, is the learnable parameter of the convolutional layer, and b is the bias. yes Weights in a convolutional filter used to extract features at a specific scale.

[0090] Step S1023: In the third scale branch, the input feature map is processed sequentially by a (1×1) convolution with ReLU activation function and two consecutive 3×3 convolutions with ReLU activation function.

[0091] exist Figure 5 In this approach, the 3×3 stacked branch employs a deep structure consisting of a (1×1) convolutional layer followed by two consecutive (3×3) convolutional layers. This effectively reduces computational costs while achieving continuous feature extraction, as shown in the following formula:

[0092] ,

[0093] in, The representative pooling window is the average pooling operation. The area covered by the kernel with each slide, the window contains 1 pixel, used to calculate the average value.

[0094] Step S1024: In the fourth scale branch, perform pooling processing on the input feature map.

[0095] same Figure 5 In this embodiment, the pooling branch processes the input feature map using pooling and (1×1) convolution with ReLU activation function. The specific pooling process will not be described in detail in this embodiment.

[0096] Step S1025: Concatenate all branch processing results and obtain a multi-scale feature map through (1×1) convolution.

[0097] After multi-scale processing, the outputs of each branch are spliced ​​along the channel dimension to form a multi-scale tensor that integrates fine-grained to coarse-grained features. Finally, the spliced ​​feature map is compressed and integrated through (1×1) convolution to generate the output of the multi-scale convolutional network (MSCN) to ensure its compatibility with downstream layers.

[0098] In summary, MSCN processes the input feature map through four parallel branches: a 1×1 branch, a 5×5 branch, a 3×3 stacked branch, and a pooling branch. Each branch aims to extract useful spatial information at different receptive field scales. Furthermore, by concatenating the outputs of multi-scale convolutions and then performing feature compression and dimensionality alignment using the ReLU activation function and (1×1) convolutions, the model achieves a balance between computational efficiency and rich spatial encoding.

[0099] Step S103: Perform global average pooling and global max pooling on the multi-scale feature maps in the spatial dimension, and generate channel attention maps by splicing and normalization.

[0100] See Figure 6 This is a flowchart illustrating a channel attention calculation method provided in an embodiment of the present invention, and is combined with... Figure 7 This is a schematic diagram of a computational framework for channel attention and parallel spatial attention provided in an embodiment of the present invention. The computational mechanism for channel attention is as follows:

[0101] Step S1031: Perform global average pooling and global max pooling on the multi-scale feature map in the spatial dimension to generate two compact channel feature descriptors.

[0102] The Channel Attention Module (CAM) enhances feature discrimination by strengthening key features in the channel dimension, thereby optimizing network performance. For the input multi-scale feature map (… Figure 7 China and Israel (Representation) CAM first performs global average pooling and global max pooling in the spatial dimension, respectively, to generate two compact channel feature descriptors. Its mathematical expression is as follows:

[0103] ,

[0104] ,

[0105] ,

[0106] ,

[0107] in, Represents the full-space-channel feature map. These represent the height, width, and number of channels of the feature map, respectively. express The descriptor feature vector is obtained by aggregating and compressing feature space information along the channel dimension. This is an index for the channel dimension.

[0108] Step S1032: The channel feature descriptors are respectively processed by two layers of multilayer perceptron with shared weights, and then fused by element-wise addition and normalized by the sigmoid function to generate channel attention vectors.

[0109] Global average pooling output With global max pooling output Each layer independently passes through a two-layer multilayer perceptron (MLP) with shared weights. The weight matrices of this MLP are as follows: and To improve parameter efficiency, the channel dimension was reduced to... (c is the number of channels in the input feature map,) For the reduction ratio, (This represents the ReLU activation function). The output expression of the MLP is as follows:

[0110] ,

[0111] ,

[0112] ,

[0113] ,

[0114] Subsequently, the two channel descriptors are fused by element addition to obtain , Then through the Sigmoid function (equation) (Representation) Normalized generation of channel attention vectors .

[0115] Step S1033: Apply the channel attention vector to the multi-scale feature map through channel multiplication to obtain the channel attention map.

[0116] The channel attention vectors generated in the above embodiments Spatial information is applied to the input multi-scale feature map using channel multiplication (i.e., This yields the channel attention map, calculated using the following formula:

[0117] ,

[0118] This process enables the Channel Attention Module (CAM) to enhance information-rich channels while suppressing secondary feature channels.

[0119] Step S104: Perform global average pooling and global max pooling along the channel dimension of the channel attention map, and generate a spatial attention map by concatenation and normalization through parallel convolution branches.

[0120] See Figure 8 This is a schematic diagram of a parallel spatial attention computation process provided in an embodiment of the present invention. Simultaneously, combined with... Figure 7 The parallel spatial attention calculation mechanism is as follows:

[0121] Step S1041: In the two parallel branches, global average pooling and global max pooling are performed along the channel dimension of the channel attention map to generate corresponding channel feature descriptors, which are then concatenated along the channel dimension to form the final descriptor.

[0122] Parallel Spatial Attention Module (PSAM) enhances feature representations by focusing on key spatial locations that improve network performance (see...). Figure 7 Given an input channel attention map, in tensor... This module first performs global max pooling and global average pooling along the channel dimension to generate two independent channel feature descriptors. Each element of the descriptor is composed of The calculations yielded the results for these two descriptors. and The final descriptor is formed by splicing along the channel dimension. The mathematical representation of this process is as follows:

[0123] ,

[0124] ,

[0125] ,

[0126] ,

[0127] ,

[0128] in, This represents the full-space-channel feature map, where H, W, and C represent the height, width, and number of channels of the feature map, respectively. A spatial feature map is used to encode the spatial location information of features; and for Spatial subset; This indicates element-wise addition.

[0129] Step S1042: Perform convolution calculations on the corresponding final descriptors of each parallel branch using (3×3) and (7×7) convolution kernels respectively.

[0130] Subsequently, PSAM processes these descriptors through two parallel convolutional branches to capture multi-scale spatial dependencies. The mathematical expression for this process is as follows:

[0131] ,

[0132] ,

[0133] in, and Filters and Learnable convolutional kernels, Represents the convolution operator. and Filters and The bias term is used to capture local fine-grained structural features and coarse-grained structural features.

[0134] Step S1043: The computation results of the two parallel branches are fused by element-wise addition, normalized, and processed by the sigmoid function to obtain the spatial attention vector.

[0135] The output of the above steps is obtained through the following concatenation and normalization:

[0136] ,

[0137] ,

[0138] ,

[0139] ,

[0140] in, For learnable scaling factor, This is the offset. This is the batch average. For batch variance, For a small positive number, ensure numerical stability during batch normalization calculations, and its value range is... Then through the Sigmoid function ( Map the normalized values ​​to Interval. Generate spatial attention vectors. .

[0141] Step S1044: Apply the spatial attention vector to the channel attention map through channel multiplication to obtain the spatial attention map.

[0142] Finally, the channel attention map of the input. The spatial attention vector output by the above embodiments Perform element-wise multiplication to obtain the spatial attention map. The mathematical expression for this process is as follows:

[0143] ,

[0144] ,

[0145] Thus, through the processing in steps S103 and S104 above, a parallel attention fusion strategy is implemented, and the present invention dynamically reweights the features:

[0146] In the process of processing channel attention features (corresponding to CAM, step S103), it is implemented by combining global average pooling and max pooling with shared multilayer perceptrons (MLPs);

[0147] In the process of processing parallel spatial attention features (corresponding to PSAM, step S104), it is achieved through parallel (3×3) and (7×7) convolution branches, thereby realizing the refinement of spatial features from coarse to fine.

[0148] Step S105: The output feature map is input into the decoder, which is composed of multiple deconvolutional layers stacked sequentially, and the multiple deconvolutional layers are symmetrically arranged with the multiple convolutional layers; each deconvolutional layer performs residual calculation based on the calculation result of the upper deconvolutional layer and the spatial attention map input to the corresponding convolutional layer, and finally outputs a denoised image.

[0149] The structure of the decoder has been described in detail in the above embodiments and will not be repeated here. See also Figure 9 This is a schematic diagram of a computational architecture for residual computation provided in an embodiment of the present invention. A residual compensation mechanism is used to establish skip connections between corresponding layers of the encoder and decoder, enabling the network to learn the residual mapping between input and output, rather than a direct mapping. By avoiding a forced one-to-one correspondence between convolutional and deconvolutional layers, skip connections achieve direct flow of encoder information to the decoder, ensuring the preservation of key spatial information during feature transformation. It should be noted that this skip connection method can be set between all convolutional layers and their corresponding convolutional layers, or between some convolutional and deconvolutional layers. Furthermore, the calculation process of skip connections can include one or any number of steps S102 to S104. The number of such calculations is not inherently limited and can be adjusted according to available computing resources (such as GPU memory and processing power) and the performance requirements of specific applications. All the above specific implementation processes should fall within the protection scope of this invention. The residual compensation mechanism can improve model performance in three aspects: improving model training efficiency by mitigating gradient decay in deep networks, preventing gradient vanishing by constructing cross-layer identity mapping paths, and improving reconstruction quality by balancing feature recovery capability and noise suppression effect. From a mathematical perspective, in the residual compensation mechanism, the input... With output Through residual function By establishing a correlation, the residuals can be approximated as:

[0150] ,

[0151] The network learns the residual function by stacking convolutional layers, and the original mapping is reconstructed through the identity mapping path. Figure 9 ), specifically:

[0152] ,

[0153] Encoder Feature Map: This is the module's input. The output feature map comes from the encoder. The dashed box in the Skip-Connection Module contains the core logic of the skip connection, including an N×SCMA Block (which can be understood as an N-layer SCMA functional block), whose function is to learn and obtain the residual function. To the features of subsequent inputs Perform transformation processing. Input It will "propagate directly" to the module backend, and the features processed by the SCMA Block (denoted as...) ), will be different from the original input Added together, they become ( ), to achieve "residual compensation". After being calculated by all the deconvolution layers in the decoder, the Decoder Feature Map (decoder feature map) corresponds to the final output ( ), that is, the finally generated denoised image (Reconstructed Output).

[0154] This design of the embodiment of the present invention enables the model to effectively retain the key underlying features in the input image and at the same time achieve output reconstruction with higher accuracy. The convolutional layer and the deconvolution layer adopt a symmetric structure and are configured with the same-size convolutional kernels. The overall architecture being symmetric is conducive to the efficient recovery of spatial features, ensuring that the feature map extracted by the encoder can be accurately reconstructed by the decoder; in addition, the model introduces the ReLU function to introduce non-linear characteristics, thereby enhancing the feature expression ability. The skip connections between corresponding layers are enhanced by the SCMA block to achieve deep residual learning and maintain spatial details during the denoising process, and the skip connections allow information to bypass the intermediate layers and be directly added to the decoding output, thereby significantly improving the structural similarity of the reconstructed image and optimizing the overall parameter learning efficiency and representation generalization ability of the model.

[0155] Step S106: Train the neural network model through the loss function of the denoised image and the reference image, and deploy the trained neural network model. The trained neural network model is used for deployment and denoising of low-dose CT images.

[0156] The neural network model provided by the embodiment of the present invention (characterized by ED-SCMA, English: Encoder-Decoder with Skip-Connection and Multi-Scale Attention, Chinese: Encoder-decoder architecture combining skip connection and multi-scale attention mechanism) is trained in an end-to-end manner. Here, paired low-dose computed tomography (LDCT) images and corresponding normal-dose computed tomography (NDCT) real images ( ) are used as the network input. Operations such as intensity normalization and data augmentation (such as random flipping, rotation, intensity perturbation) are performed on the low-dose CT images to improve the generalization ability and robustness of the model. Each normalized low-dose CT image is propagated through the encoder, the embedded SCMA attention module, and the decoder in sequence to generate a reconstructed output . The training objective is set as the reconstruction loss, which penalizes the voxel-level deviation of the reference NDCT image to ensure the numerical stability of the optimization process. The loss function is defined as follows:

[0157] ,

[0158] Loss function The gradients of all network parameters (convolutional, deconvolutional, and residual skip connection parameters) *b* are estimated using the standard backpropagation algorithm with the chain rule. Parameter updates are performed using the Adam optimizer: the learning rate is set to... Exponential decay coefficient Set as The parameters are iteratively updated on small batches of data until the model converges. Overall, the ED-SCMA neural network model learns the mapping relationship for hierarchical denoising during training, while dynamically adjusting the SCMA attention weights, thereby accurately highlighting key spatial-channel features.

[0159] This invention employs an end-to-end training strategy, which minimizes the loss function between the denoised prediction image and its corresponding NDCT real image, thereby better preserving image details, edge structures, and texture information, and improving the robustness and generalization ability of the model under different noise levels.

[0160] Once training is complete, the model can be deployed in real-world clinical applications. For new LDCT images, after appropriate preprocessing, the model can output high-fidelity, denoised images that preserve anatomical structures, thereby improving the diagnostic quality of the images while reducing radiation dose.

[0161] The low-dose CT image denoising method provided in this invention evaluates the system's denoising performance using objective image quality metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Root Mean Square Error (RMSE) to ensure its reliability in terms of denoising accuracy and structural fidelity. The quantitative evaluation results are shown in the table below, which also presents a comparison between the method of this invention (ED-SCMA) and other mainstream methods such as RED-CNN.

[0162] Table: Performance Comparison of the Invention and Mainstream Methods on AAPM, LDCT, and ACR Virtual Datasets

[0163]

[0164] Where PSNR is the peak signal-to-noise ratio, SSIM is the structural similarity index, and RMSE is the root mean square error;

[0165] "↑" indicates that a larger value indicates better performance, and "↓" indicates that a smaller value indicates better performance. The ED-SCMA method proposed in this invention outperforms other current mainstream methods such as RED-CNN in both image quality and quantization metrics. Figure 10This is a comparison of low-dose CT image denoising results provided by an embodiment of the present invention, demonstrating the superior resolution performance of ED-SCMA on the AAPM, LDCT, and ACR virtual datasets: compared with the real labeled NDCT image, ED-SCMA improves the image sharpness in the high-resolution contrast region, which is particularly evident in the magnified region marked by the dashed rectangle. Compared with the results of RED-CNN, ED-SCMA successfully recovers more details (Figure 10). The quantization results in the table above further verify this conclusion: ED-SCMA achieves the highest PSNR (44.1894) and SSIM (0.9474), while having the lowest RMSE (0.0030), fully demonstrating its ability to accurately reconstruct images highly similar to the target real labeled NDCT. The comprehensive improvement in visual effect and quantization performance confirms the effectiveness of this method in low-dose CT denoising tasks and has the potential for translational application in the field of medical imaging.

[0166] In summary, this implementation method effectively improves the denoising capability of low-dose CT images by introducing a deep learning network with skip connections and multi-scale attention mechanisms, and has good application prospects for medical image enhancement and clinical diagnostic support.

[0167] As can be seen from the description of the above embodiments, the denoising method for low-dose CT images provided by the embodiments of the present invention involves inputting the low-dose CT image into an encoder, which is composed of multiple stacked convolutional layers. The multiple convolutional layers sequentially calculate output feature maps. Multi-scale convolution is performed on the input feature map of each convolutional layer. The multi-scale convolution uses branches with different convolutional kernels for parallel processing to obtain multi-scale feature maps. Global average pooling and global max pooling are performed on the multi-scale feature maps in the spatial dimension, and the resulting maps are concatenated and normalized to generate channel attention maps. The channel attention maps are then divided along the channel dimension. Global average pooling and global max pooling are not performed. Spatial attention maps are generated through parallel convolutional branches, concatenated, and normalized. The output feature map is input to a decoder, which consists of multiple deconvolutional layers stacked sequentially, symmetrically arranged with the convolutional layers. Each deconvolutional layer performs residual calculation based on the calculation results of the previous deconvolutional layer and the spatial attention map input to the corresponding convolutional layer, finally outputting a denoised image. A neural network model is trained using the loss function of the denoised image and a reference image. The trained neural network model is then deployed to denoise low-dose CT images. This invention employs a novel encoder-decoder neural network architecture, combining skip connections and multi-scale attention mechanism functional blocks to maintain the integrity of spatial features and enhance feature representation capabilities. The SCMA functional block enhances denoising capabilities by combining a multi-scale convolutional network module (MSCN Block) and a convolutional block attention module with parallel spatial attention (CBAM-PSA Block). These two modules work together to extract hierarchical features across multiple receptive fields and dynamically recalibrate the importance of channels and space. This integrated approach effectively addresses the limitations of traditional methods, achieving efficient denoising and enhanced feature representation in low-dose CT images while maintaining structural integrity.

[0168] Through the description of the above method embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms, 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, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0169] This invention provides a non-volatile computer storage medium storing computer-executable instructions that can execute the denoising method for low-dose CT images in any of the above method embodiments.

[0170] Corresponding to the denoising method embodiment for low-dose CT images provided by the present invention, the present invention also provides a denoising device for low-dose CT images.

[0171] See Figure 11 This is a schematic diagram of a noise reduction device for low-dose CT images provided in an embodiment of the present invention, and can also be referred to simultaneously. Figure 2 As shown in the figure, the device includes:

[0172] Encoder 1 is used to receive low-dose CT images, and encoder 1 is composed of multiple convolutional layers (not shown) stacked together, the multiple convolutional layers sequentially calculating the output feature map;

[0173] Skip connections and multi-scale attention functional block 2, including multi-scale convolutional network module 21 and convolutional block attention module 22 with parallel spatial attention, wherein:

[0174] The multi-scale convolutional network module 21 is used to perform multi-scale convolution on the input feature map of each convolutional layer. The multi-scale convolution uses branches with different convolutional kernels for parallel processing to obtain multi-scale feature maps.

[0175] The multi-scale convolution includes four parallel branches:

[0176] The first scale branch processes the input feature map using a 1×1 convolution with ReLU activation.

[0177] The second scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and a 5×5 convolution with ReLU activation.

[0178] The third scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and two consecutive 3×3 convolutions with ReLU activation.

[0179] In the fourth scale branch, the input feature map is pooled.

[0180] The processing results of all branches are concatenated and multi-scale feature maps are obtained through 1×1 convolution.

[0181] The convolutional block attention module 22 of the parallel spatial attention includes a channel attention module 221 and a parallel spatial attention module 222; wherein, the channel attention module 221 is used to perform global average pooling and global max pooling on the multi-scale feature maps in the spatial dimension, and generate a channel attention map by concatenation and normalization; the parallel spatial attention module 222 is used to perform global average pooling and global max pooling on the channel attention maps along the channel dimension, and generate a spatial attention map by concatenation and normalization through parallel convolutional branches;

[0182] Decoder 3 is used to receive the output feature map, and the decoder 3 is composed of multiple deconvolution layers (not shown) stacked in sequence, the multiple deconvolution layers are symmetrically arranged with the multiple convolution layers; each deconvolution layer performs residual calculation based on the calculation result of the upper deconvolution layer and the spatial attention map input of the corresponding convolution layer, and finally outputs a denoised image;

[0183] Training module 4 is used to train a neural network model using loss functions of denoised images and reference images. The trained neural network model is then deployed to denoise low-dose CT images.

[0184] Optionally, the channel attention module 221 performs global average pooling and global max pooling on the multi-scale feature maps in the spatial dimension, respectively, and generates a channel attention map by concatenation and normalization, including:

[0185] The multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, respectively, to generate two compact channel feature descriptors;

[0186] The channel feature descriptors are respectively processed by two layers of multilayer perceptron with shared weights, and then fused by element-wise addition and normalized by the sigmoid function to generate channel attention vectors;

[0187] The channel attention vector is applied to the multi-scale feature map through channel multiplication to obtain the channel attention map.

[0188] Optionally, the parallel spatial attention module 222 performs global average pooling and global max pooling on the channel attention map along the channel dimension, and generates a spatial attention map through parallel convolution branches, concatenating and normalizing the data, including:

[0189] In the two parallel branches, global average pooling and global max pooling are performed along the channel dimension of the channel attention map to generate corresponding channel feature descriptors, which are then concatenated along the channel dimension to form the final descriptor.

[0190] Convolution calculations were performed on the final descriptors of each parallel branch using 3×3 and 7×7 convolution kernels, respectively.

[0191] The computation results of the two parallel branches are fused by element-wise addition, normalized, and processed by the sigmoid function to obtain the spatial attention vector;

[0192] The spatial attention vector is applied to the channel attention map through channel multiplication to obtain the spatial attention map.

[0193] Optionally, the residual calculation includes: establishing skip connections between corresponding layers of the encoder and decoder using a residual compensation mechanism; embedding the skip connections and multi-scale attention function blocks between corresponding layers of the encoder and decoder to learn the residual function; and adding the residual function to the encoder output feature map to achieve residual compensation. The skip connections include an integrated SCMA function block, which combines a multi-scale convolutional network module (MSCN Block) and a convolutional block attention module with parallel spatial attention (CBAM-PSA Block).

[0194] Figure 12 This is a schematic diagram of the hardware structure of the electronic device for performing a low-dose CT image denoising method according to an embodiment of the present invention, as shown below. Figure 12 As shown, the device includes:

[0195] One or more processors 510 and memory 520, Figure 12 Take the 510 processor as an example.

[0196] The apparatus for performing a denoising method on low-dose CT images may further include an input device 530 and an output device 540.

[0197] The processor 510, memory 520, input device 530, and output device 540 can be connected via a bus or other means. Figure 12 Taking the example of a connection between China and Israel via a bus.

[0198] The memory 520, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the low-dose CT image denoising method in this embodiment of the invention (e.g., attached...). Figure 11 The encoder 1, skip connection and multi-scale attention function block 2, decoder 3, and training module 4 are shown. The processor 510 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 520, thereby implementing the denoising method for low-dose CT images described in the above method embodiment.

[0199] The memory 520 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the low-dose CT image denoising processing device. Furthermore, the memory 520 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 520 may optionally include memory remotely located relative to the processor 510, and these remote memories may be connected to the low-dose CT image denoising processing device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0200] Input device 530 can receive input digital or character information, as well as key signal inputs related to user settings and function control of the low-dose CT image denoising processing device. Output device 540 may include a display device such as a display screen.

[0201] The one or more modules are stored in the memory 520 and, when executed by the one or more processors 510, perform the low-dose CT image denoising method in any of the above method embodiments.

[0202] The above-described product can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0203] The electronic devices of this invention exist in various forms, including but not limited to:

[0204] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include: smartphones (e.g., iPhones), multimedia phones, feature phones, and low-end phones, etc.

[0205] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, such as the iPad.

[0206] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes: audio and video players (such as iPods), handheld game consoles, e-books, as well as smart toys and portable car navigation devices.

[0207] (4) Server: A device that provides computing services. The components of a server include a processor, hard disk, memory, system bus, etc. Servers are similar to general computer architectures, but because they need to provide highly reliable services, they have higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0208] (5) Other electronic devices with data interaction functions.

[0209] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0210] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0211] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0212] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for denoising low-dose CT images, characterized in that, include: Low-dose CT images are input into an encoder, which consists of multiple stacked convolutional layers, and the multiple convolutional layers sequentially calculate the output feature map. Multi-scale convolution is performed on the input feature map of each convolutional layer. The multi-scale convolution is processed in parallel using branches with different convolutional kernels to obtain multi-scale feature maps. The multi-scale convolution includes four parallel branches: The first scale branch processes the input feature map using a 1×1 convolution with ReLU activation. The second scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and a 5×5 convolution with ReLU activation. The third scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and two consecutive 3×3 convolutions with ReLU activation. In the fourth scale branch, the input feature map is pooled. The processing results of all branches are concatenated and multi-scale feature maps are obtained through 1×1 convolution. Multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, and then spliced ​​and normalized to generate channel attention maps; The channel attention map performs global average pooling and global max pooling along the channel dimension, and generates a spatial attention map by concatenation and normalization through parallel convolution branches. The output feature map is input to the decoder, which is composed of multiple deconvolutional layers stacked sequentially, and the multiple deconvolutional layers are symmetrically arranged with the multiple convolutional layers; each deconvolutional layer performs residual calculation based on the calculation result of the upper deconvolutional layer and the spatial attention map input to the corresponding convolutional layer, and finally outputs a denoised image; The neural network model is trained using loss functions from denoised and reference images. The trained neural network model is then deployed to denoise low-dose CT images.

2. The method for denoising low-dose CT images according to claim 1, characterized in that, The multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, respectively, and are then concatenated and normalized to generate channel attention maps, including: The multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, respectively, to generate two compact channel feature descriptors; The channel feature descriptors are respectively processed by two layers of multilayer perceptron with shared weights, and then fused by element-wise addition and normalized by the sigmoid function to generate channel attention vectors; The channel attention vector is applied to the multi-scale feature map through channel multiplication to obtain the channel attention map.

3. The method for denoising low-dose CT images according to claim 1, characterized in that, The channel attention map is subjected to global average pooling and global max pooling along the channel dimension, and a spatial attention map is generated by concatenation and normalization through parallel convolution branches, including: In the two parallel branches, global average pooling and global max pooling are performed along the channel dimension of the channel attention map to generate corresponding channel feature descriptors, which are then concatenated along the channel dimension to form the final descriptor. Convolution calculations were performed on the final descriptors of each parallel branch using 3×3 and 7×7 convolution kernels, respectively. The computation results of the two parallel branches are fused by element-wise addition, normalized, and processed by the sigmoid function to obtain the spatial attention vector; The spatial attention vector is applied to the channel attention map through channel multiplication to obtain the spatial attention map.

4. The method for denoising low-dose CT images according to claim 1, characterized in that, The residual calculation includes: establishing a skip connection between corresponding layers of the encoder and decoder using a residual compensation mechanism, embedding the skip connection and multi-scale attention function block between corresponding layers of the encoder and decoder to learn the residual function, and adding it to the encoder output feature map to achieve residual compensation.

5. A noise reduction device for low-dose CT images, characterized in that, include: An encoder is used to receive low-dose CT images, and the encoder is composed of multiple convolutional layers stacked together, the multiple convolutional layers sequentially calculating to obtain an output feature map; The skip connection and multi-scale attention functional blocks include a multi-scale convolutional network module and a convolutional block attention module with parallel spatial attention, wherein... The multi-scale convolutional network module is used to perform multi-scale convolution on the input feature map of each convolutional layer. The multi-scale convolution uses branches with different convolutional kernels for parallel processing to obtain multi-scale feature maps. The multi-scale convolution includes four parallel branches: The first scale branch processes the input feature map using a 1×1 convolution with ReLU activation. The second scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and a 5×5 convolution with ReLU activation. The third scale branch processes the input feature map sequentially through a 1×1 convolution with ReLU activation and two consecutive 3×3 convolutions with ReLU activation. In the fourth scale branch, the input feature map is pooled. The processing results of all branches are concatenated and multi-scale feature maps are obtained through 1×1 convolution. The convolutional block attention module for parallel spatial attention includes a channel attention module and a parallel spatial attention module. The channel attention module performs global average pooling and global max pooling on multi-scale feature maps along the spatial dimension, generating a channel attention map through concatenation and normalization. The parallel spatial attention module performs global average pooling and global max pooling on the channel attention maps along the channel dimension, generating a spatial attention map through parallel convolutional branches, followed by concatenation and normalization. A decoder is used to receive the output feature map, and the decoder is composed of multiple deconvolutional layers stacked in sequence, the multiple deconvolutional layers being symmetrically arranged with the multiple convolutional layers; each deconvolutional layer performs residual calculation based on the calculation result of the upper deconvolutional layer and the spatial attention map input to the corresponding convolutional layer, and finally outputs a denoised image; The training module is used to train a neural network model using a loss function on denoised images and reference images. The trained neural network model is then deployed to denoise low-dose CT images.

6. The noise reduction device for low-dose CT images according to claim 5, characterized in that, The channel attention module performs global average pooling and global max pooling on the multi-scale feature maps in the spatial dimension, and generates channel attention maps by concatenation and normalization, including: The multi-scale feature maps are subjected to global average pooling and global max pooling in the spatial dimension, respectively, to generate two compact channel feature descriptors; The channel feature descriptors are respectively processed by two layers of multilayer perceptron with shared weights, and then fused by element-wise addition and normalized by the sigmoid function to generate channel attention vectors; The channel attention vector is applied to the multi-scale feature map through channel multiplication to obtain the channel attention map.

7. The noise reduction device for low-dose CT images according to claim 5, characterized in that, The parallel spatial attention module performs global average pooling and global max pooling on the channel attention map along the channel dimension, and generates a spatial attention map through parallel convolution branches, concatenation, and normalization, including: In the two parallel branches, global average pooling and global max pooling are performed along the channel dimension of the channel attention map to generate corresponding channel feature descriptors, which are then concatenated along the channel dimension to form the final descriptor. Convolution calculations were performed on the final descriptors of each parallel branch using 3×3 and 7×7 convolution kernels, respectively. The computation results of the two parallel branches are fused by element-wise addition, normalized, and processed by the sigmoid function to obtain the spatial attention vector; The spatial attention vector is applied to the channel attention map through channel multiplication to obtain the spatial attention map.

8. The noise reduction device for low-dose CT images according to claim 5, characterized in that, The residual calculation includes: establishing a skip connection between corresponding layers of the encoder and decoder using a residual compensation mechanism, embedding the skip connection and multi-scale attention function block between corresponding layers of the encoder and decoder to learn the residual function, and adding it to the encoder output feature map to achieve residual compensation.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for denoising low-dose CT images as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or instruction set is loaded and executed by a processor to implement the steps of the method for denoising low-dose CT images as described in any one of claims 1 to 4.

Images