Edge enhancement based low dose CT image denoising system and method

Abstract

The application discloses an edge enhancement-based low-dose CT image denoising system and method, and belongs to the field of medical image processing. The system comprises a feature extraction module, an encoder, a decoder and an image reconstruction module. In the encoder and the decoder, each encoding layer and each decoding layer comprises a multi-domain feature module. The multi-domain feature module comprises a local feature modeling unit, a state space modeling unit and a multi-domain feature fusion unit. The local feature modeling unit is used for performing convolution operation on the input feature to obtain a local feature. The state space modeling unit is used for performing feature extraction on the input feature through a state space model to obtain a global feature, and after extracting an edge feature, the two are fused into an edge enhancement global feature. The multi-domain feature fusion unit is used for fusing the local feature and the edge enhancement global feature. The application can improve the noise suppression effect of the low-dose CT image, and significantly improve the structure consistency and edge retention capability of the denoised image.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing, and more specifically, relates to a low-dose CT image denoising system and method based on edge enhancement. Background Technology

[0002] Computed tomography (CT) technology is widely used in clinical disease screening, diagnosis, and treatment evaluation due to its high imaging speed and spatial resolution. However, CT imaging requires the use of X-rays, and the radiation dose poses a potential hazard to human health. Long-term or frequent high-dose CT scans may increase the risk of developing diseases such as cancer. Therefore, reducing CT scan doses while ensuring diagnostic quality has become an important research direction in the field of medical imaging.

[0003] Reducing the scan dose directly leads to a significant increase in quantum noise in CT images, manifesting as increased noise levels, decreased contrast, and blurred details, severely impacting the accuracy of clinical diagnosis. To address these issues, researchers have proposed various low-dose CT image denoising methods, including traditional filtering-based methods, model-based iterative reconstruction methods, and deep learning-based methods.

[0004] In recent years, deep learning models such as Convolutional Neural Networks (CNNs) and Transformers have achieved good results in LDCT image denoising tasks. Among them, CNNs are good at modeling local spatial features and have the advantage of high computational efficiency, but their receptive field is limited, making it difficult to effectively capture long-range dependencies. Although Transformer models can perform global modeling, their computational complexity is high, requiring a lot of GPU memory and computing power, which is not conducive to the efficient processing of high-resolution medical images.

[0005] Visual State Space Models (VSSMs), as a novel sequence modeling structure, have shown good performance in low-level vision tasks in recent years, achieving long-range dependency modeling while maintaining linear computational complexity. However, existing VSSMs still suffer from shortcomings in medical image denoising tasks, such as insufficient modeling of edges and high-frequency details and limited adaptability to complex noise distributions, and their denoising performance needs further improvement. Summary of the Invention

[0006] In view of the shortcomings of the existing technology and the need for improvement, the present invention provides a low-dose CT image denoising system and method based on edge enhancement. Its purpose is to improve the noise suppression effect on low-dose CT images and significantly improve the structural consistency and edge preservation ability of the denoised image.

[0007] To achieve the above objectives, according to one aspect of the present invention, a low-dose CT image denoising system based on edge enhancement is provided, comprising: a feature extraction module, an encoder, a decoder, and an image reconstruction module; The feature extraction module is used to extract features from the input low-dose CT images to obtain shallow spatial features; The encoder consists of multiple sequentially connected coding layers, which are used to downsample shallow spatial features step by step and extract semantic features to obtain coded features; the decoder consists of multiple sequentially connected decoding layers, which are used to upsample coded features step by step and recover features to obtain decoded features; the feature information output by each coding layer is introduced into the input of the corresponding decoding layer through skip connections; The image reconstruction module is used to reconstruct the decoded features and then fuse them with the input low-dose CT image through residual connections to obtain a denoised CT image. Each coding layer and each decoding layer contains a multi-domain feature module, which includes a local feature modeling unit, a state space modeling unit, and a multi-domain feature fusion unit. The local feature modeling unit is used to perform convolution operations on the input features to obtain local features; The state-space modeling unit is used to extract features from the input features through the state-space model to obtain global features, and after extracting edge features, the two are fused into edge-enhanced global features. The multi-domain feature fusion unit is used to fuse local features with edge-enhanced global features.

[0008] Furthermore, the state-space modeling unit includes a global feature extraction branch, an edge enhancement branch, and an activation function branch; The global feature extraction branch is used to extract features from the input features using the state-space model to obtain global features; The edge enhancement branch is used to extract edge features of the input features through the edge enhancement operator; The activation function branch is used to activate the features obtained by weighted summation of global and edge features.

[0009] Furthermore, the edge enhancement operator is a differential convolution, a directional convolution, or a combination of differential convolution and directional convolution.

[0010] Furthermore, before performing convolution operations on the input features, the local feature modeling unit also includes: The input features are subjected to discrete wavelet transform, which decomposes them into multiple sub-feature components of different frequency bands, and then the multiple sub-feature components of different frequency bands are concatenated.

[0011] Furthermore, the multi-domain feature fusion unit includes: a splicing unit, a dual-branch unit, and a residual connection unit; The stitching unit is used to stitch together local features and edge-enhanced global features along the channel dimension; The first branch in the dual-branch unit is used to perform a 3×3 convolution operation on the concatenated features to achieve feature transformation and channel compression. The second branch is used to perform two 3×3 convolution operations on the concatenated features to achieve feature refinement and enhancement. The residual connection unit is used to add the features output by the two branches in the dual-branch unit to the features output by the splicing unit, thereby achieving multi-domain feature fusion.

[0012] Furthermore, between the extraction module and the encoder, there is also a first operational-level adaptive learning module; Furthermore, a second operational-level adaptive learning module is also included between the decoder and the image reconstruction module.

[0013] According to another aspect of the present invention, a training method for the above-described edge enhancement-based low-dose CT image denoising system is provided, comprising: A training dataset was constructed using low-dose CT images and corresponding normal-dose CT images. Using low-dose CT images as input information and normal-dose CT images as supervision information, the above-mentioned low-dose CT image denoising system based on edge enhancement is trained end-to-end to obtain a trained low-dose CT image denoising system.

[0014] Furthermore, the training loss function is: ; in, Indicates the losses incurred during reconstruction. ; Represents structural similarity loss. ; This represents the denoised CT image output by the low-dose CT image denoising system. This represents the corresponding normal dose image. express……, Indicating structural similarity, and These represent the weights of the reconstruction loss and the structural similarity loss, respectively.

[0015] According to another aspect of the present invention, a low-dose CT image denoising method based on edge enhancement is provided, comprising: The low-dose CT image to be processed is input into a pre-trained low-dose CT image denoising system to obtain a denoised CT image. The low-dose CT image denoising system that has been trained is obtained by the training method provided by the present invention.

[0016] According to another aspect of the present invention, a computer-readable storage medium is provided, including a stored computer program that, when executed by a processor, implements the edge-enhanced low-dose CT image denoising method provided by the present invention.

[0017] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects: (1) The low-dose CT image denoising system based on edge enhancement provided by the present invention contains a multi-domain feature module in the encoder layer and the decoder layer. The multi-domain feature module extracts local features through convolution operation, extracts the long-range modeling dependency of low-dose CT image through state space model to obtain global features, and performs display modeling on edge information to obtain edge enhancement global features. Furthermore, the local features and edge enhancement global features are fused to realize the joint expression of local spatial information and global context information. While effectively suppressing noise, it avoids the loss of detailed structure and significantly improves the structural consistency and edge preservation ability of the denoising results.

[0018] (2) In a further preferred embodiment of the low-dose CT image denoising system based on edge enhancement provided by the present invention, the multi-domain feature module first performs discrete wavelet transform on the features, decomposes them into multiple sub-feature components of different frequency bands, and splices the multiple sub-feature components of different frequency bands. Then, it performs convolution operation on the spliced ​​features to extract local features. Thus, while extracting local features, it realizes multi-frequency band decomposition, explicitly divides the image features into low-frequency structural information and high-frequency edge details and noise information, effectively realizes the distinction between noise components and anatomical structural information in the frequency domain, avoids the problem of excessive smoothing of structural information caused by the mixing of high-frequency details and noise in existing methods, thereby improving the integrity and discriminability of the overall feature expression and providing a more reliable feature basis for subsequent applications.

[0019] (3) The edge-enhanced low-dose CT image denoising system provided by the present invention, in its further preferred embodiment, includes a multi-domain feature fusion unit comprising a stitching unit, a dual-branch unit and a residual connection unit, which realizes joint modeling of spatial domain, frequency domain, edge domain and long-range dependency information, so that spatial domain features, frequency domain features and edge enhancement features can complement and synergize with each other, thereby enhancing the model's adaptability to complex noise distribution, different tissue structures and different imaging conditions of CT images, improving the robustness and generalization performance of the model under multiple scene and multiple data distribution conditions. Furthermore, by using multi-domain and long- and short-range dependency joint modeling, the present invention significantly improves the image edge clarity and structural restoration quality while ensuring noise suppression performance, so that the denoising results show obvious advantages in both subjective visual quality and objective evaluation indicators.

[0020] (4) The edge enhancement-based low-dose CT image denoising system provided by the present invention, in its further preferred embodiment, introduces an operation-level adaptive learning mechanism at the input and output ends, which can dynamically adjust the weights of various feature processing operators according to different feature response intensities, so that the model can adaptively select a more suitable feature processing method under different noise levels, different image regions and different structural complexities, thereby further improving the denoising effect and reducing the damage to image details.

[0021] (5) The training method of the low-dose CT image denoising system based on edge enhancement provided by the present invention constructs a supervised training mechanism between low-dose CT images and corresponding standard-dose CT images to train the denoising system end-to-end, so that the obtained model can accurately learn the statistical characteristics of low-dose imaging noise and its mapping relationship with the real structure, thereby obtaining stable and reliable denoising performance.

[0022] (6) The training method of the low-dose CT image denoising system based on edge enhancement provided by the present invention, in its preferred embodiment, includes structural similarity loss in addition to reconstruction loss, which can constrain brightness, contrast and structure, and further improve the denoising effect of the system.

[0023] Overall, this invention can balance noise suppression and structure preservation when denoising low-dose CT images, resulting in high-quality reconstruction results. It has good engineering practical value and broad clinical application prospects. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a low-dose CT image denoising system based on edge enhancement provided in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of the state-space modeling unit provided in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of a training method for a low-dose CT image denoising system based on edge enhancement, provided in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of a low-dose CT image denoising method based on edge enhancement provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0029] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0030] To address the challenges of existing low-dose CT image denoising methods in balancing noise suppression and structural detail preservation, their insufficient ability to model long-range image dependencies, and their limited adaptability to complex and non-stationary noise distributions, this invention provides a low-dose CT image denoising system and method based on edge enhancement. Implemented using a deep learning framework, this system constructs a denoising network that integrates an edge enhancement state space model and a multi-domain feature modeling mechanism. This effectively suppresses quantum noise, structure-dependent noise, and non-uniform noise in low-dose CT images while preserving tissue structure and edge detail information as much as possible.

[0031] It should be noted that the low-dose CT images involved in this invention are CT images with a reduced scanning dose compared to clinical normal-dose CT images.

[0032] The following is an example.

[0033] Example 1: A low-dose CT image denoising system based on edge enhancement, such as Figure 1 As shown, it includes: a feature extraction module, a first-operational-level adaptive learning module, an encoder, a decoder, a second-operational-level adaptive learning module, and an image reconstruction module. The specific implementation methods of each module are as follows.

[0034] In this embodiment, the feature extraction module is used to extract features from the input low-dose CT image to obtain shallow spatial features. Optionally, in this embodiment, the feature extraction module achieves shallow spatial feature extraction through a two-dimensional convolution operation with a kernel size of 3×3.

[0035] In this embodiment, an operation-level adaptive learning module is introduced in the feature extraction stage. By dynamically weighting and combining various feature processing operators, the system can adaptively select the appropriate operation method according to the statistical characteristics of the input features.

[0036] In practical applications, a predefined set of candidate operators with multiple scales and receptive fields can be used. Optionally, the feature processing operators in this candidate operator set may include depthwise separable convolutions with different kernel sizes, dilated convolutions with different dilation rates, average pooling operators, etc. , Indicates the first A feature processing operator, , Indicates the number of feature processing operators; For the input feature map First, the channel-level description vector is obtained through global average pooling: ; in, and Representing feature maps respectively Width and height, Representation of feature map The number of channels, express The pixel value at that location.

[0037] Then, the operation weight tensor is generated via the fully connected layer network as follows: ; in, This indicates the Softmax operation. This indicates the number of operation layers in the stack.

[0038] Finally, the outputs of each operator are weighted and summed using the calculated operation weights. Accordingly, the output of each operation layer is defined as follows: ; in, For correspondence operators Adaptive weights, This indicates the use of feature processing operators. For the input feature map The result of the processing This represents a nonlinear activation function. This mechanism enables the network to dynamically select the optimal feature processing method based on different noise characteristics.

[0039] It should be noted that the types, quantities, and combinations of the above-mentioned operators can be adjusted according to actual application requirements. This is only an optional implementation method and should not be construed as the sole limitation of the present invention.

[0040] In this embodiment, the encoder includes multiple sequentially connected encoding layers, which are used to downsample the shallow spatial features processed by the first operational-level adaptive learning module and extract semantic features to obtain encoded features; the decoder includes multiple sequentially connected decoding layers, which are used to upsample the encoded features and recover features to obtain decoded features; the feature information output by each encoding layer is introduced into the input of the corresponding decoding layer through skip connections.

[0041] like Figure 1 As shown, in this embodiment, each encoding layer and each decoding layer includes a multi-domain feature module (i.e., Figure 1 The Multi-domain Block in the model includes local feature modeling units (i.e., Figure 1 WRB), state-space modeling unit (i.e. Figure 1 The system includes an EE-VSSM and a multi-domain feature fusion unit. The local feature modeling unit performs convolution operations on the input features to obtain local features; the state space modeling unit extracts features from the input features through the state space model to obtain global features, and after extracting edge features, it fuses the two into edge-enhanced global features; the multi-domain feature fusion unit fuses the local features with the edge-enhanced global features.

[0042] To further improve image structure consistency and edge preservation capabilities, in this embodiment, the local feature modeling unit first performs multi-scale frequency domain decomposition on the input features. Specifically, it uses Discrete Wavelet Transform (DWT) to decompose the features into multiple sub-feature components of different frequency bands. The decomposition process can be represented as follows: ; in, This represents low-frequency subband features, used to characterize the overall structure and low-frequency information of an image; Indicates the first A high-frequency sub-band feature is used to characterize the edges, texture details, and noise components of an image. , Indicates the number of frequency bands.

[0043] Based on discrete wavelet decomposition, the local feature modeling unit concatenates the sub-feature components of each frequency band and performs a convolution operation to obtain local features.

[0044] In this embodiment, the local feature modeling unit achieves explicit modeling of features at different scales and frequencies through multi-layer wavelet decomposition, thereby enhancing the model's ability to distinguish between noise and structural information, and introducing residual enhancement structures in the frequency domain to strengthen texture and edge detail information.

[0045] It is easy to understand that when performing discrete wavelet transform, the wavelet basis type, the number of wavelet decomposition levels, and the subband processing method can all be adjusted according to the specific application scenario.

[0046] In this embodiment, as Figure 2 As shown, the state-space modeling unit includes a global feature extraction branch, an edge enhancement branch, and an activation function branch; The global feature extraction branch is used to extract features from the input features using the state-space model to obtain global features; The edge enhancement branch is used to extract edge features of the input features through the edge enhancement operator; The activation function branch is used to activate the features obtained by weighted summation of global and edge features.

[0047] In this embodiment, the state-space modeling unit introduces an edge enhancement branch based on the existing global feature extraction and activation function branches in the traditional visual state-space model. This enhances edge perception capabilities while maintaining the ability to sequentially model features and capture long-range dependencies in images using a two-dimensional state-space model. The improved model is an edge-enhanced state-space model. The weights of the weighted sum of global and edge features are learnable parameters. This embodiment uses a state-space model to model long-range dependencies. Compared to denoising methods based on self-attention mechanisms, it eliminates the need for explicit calculation of the global attention matrix when processing high-resolution, low-dose CT images, significantly reducing computational complexity and memory overhead. While maintaining long-range dependency modeling capabilities, it improves overall computational efficiency and model stability, making it more suitable for practical applications involving large-size medical images.

[0048] In practical applications, edge enhancement operators can be implemented using differential convolution, oriented convolution, or a combination thereof. As an optional implementation, in this embodiment, the edge enhancement operator is specifically multi-angle differential convolution (MDConv).

[0049] In order to better achieve multi-domain feature fusion and make full use of the complementary and synergistic effects between spatial domain features, frequency domain features and edge enhancement features, the multi-domain feature fusion unit in this embodiment specifically includes: a splicing unit, a dual-branch unit and a residual connection unit. The stitching unit is used to stitch together local features and edge-enhanced global features along the channel dimension; The first branch in the dual-branch unit is used to perform a 3×3 convolution operation on the concatenated features to achieve feature transformation and channel compression. The second branch is used to perform two 3×3 convolution operations on the concatenated features to achieve feature refinement and enhancement. The residual connection unit is used to add the features output by the two branches in the dual-branch unit to the features output by the splicing unit, thereby achieving multi-domain feature fusion.

[0050] In this embodiment, the spatial domain features, frequency domain features and edge enhancement features are fused together by the multi-domain feature fusion unit, which can enhance the denoising system's adaptability to complex noise distribution, different tissue structures and CT images under different imaging conditions, and improve the robustness and generalization performance of the model under multiple scene and multiple data distribution conditions.

[0051] The skip connections between each coding layer and its corresponding decoding layer can be represented as follows: ; in, Indicates the decoder's first The output feature map of the layer, Indicates the decoder's first Feature maps output by the layer Indicates the encoder's first... Feature maps output by the layer Indicates the first Layer upsampling operation, Indicates the first Convolutional fusion operation in layer decoder, This indicates feature splicing or weighted fusion.

[0052] In this embodiment, the second operation-level adaptive learning module processes the features output by the decoder based on the operation-level adaptive learning mechanism. The implementation of the second operation-level adaptive learning module can refer to the implementation of the first operation-level adaptive learning module.

[0053] The image reconstruction module reconstructs the decoded features processed by the second-level adaptive decoding module, and then fuses them with the input low-dose CT image through residual connections to obtain a denoised CT image. Optionally, in this embodiment, the image reconstruction module performs the reconstruction operation through a 3×3 two-dimensional convolution operation.

[0054] The low-dose CT image denoising system based on edge enhancement provided in this embodiment, in order to This represents the input low-dose CT image. and Representing the height and width of the image respectively, in If the output is a denoised image, then: ; in, This embodiment describes the low-dose CT image denoising system based on edge enhancement provided. This represents the set of learnable parameters in the denoising system. By introducing a residual learning mechanism, the network focuses primarily on estimating the noise components, which helps improve training stability and mitigate the oversmoothing problem.

[0055] Overall, this embodiment is based on a deep learning framework. By constructing a denoising network that integrates an edge enhancement state space model and a multi-domain feature modeling mechanism, it effectively suppresses quantum noise, structure-related noise, and non-uniform noise in low-dose CT images, while preserving tissue structure and edge details as much as possible.

[0056] In practical applications, the convolution operations in the feature extraction module and the decoding and reconstruction module adopt a two-dimensional convolution structure, and the kernel size can be selected as 3×3, 5×5 or 7×7. Preferably, a 3×3 convolution kernel is used in the shallow feature extraction stage to enhance the ability to express local details, and a 5×5 or 7×7 convolution kernel can be used in combination in the encoding and decoding stage to expand the receptive field.

[0057] In other alternative implementations, the convolution operation can also take the form of depth-separable convolution, group convolution, or dilated convolution, wherein the dilation rate of dilated convolution can be set to an integer value between 2 and 4, and the choice of the above convolution method is not limited.

[0058] In the multi-domain feature module, the number of decomposition layers of the discrete wavelet transform can be set according to the input feature resolution and noise intensity. In one embodiment, the number of wavelet decomposition layers is set to 1 to 3, wherein: When the number of decomposition layers is 1, the focus is on suppressing high-frequency random noise; When the number of decomposition layers is 2 to 3, structure-related noise and texture information can be further separated.

[0059] The wavelet basis functions used can be any one or a combination of Haar wavelet, Daubechies wavelet, or Symlet wavelet, and the specific selection method is not limited.

[0060] In one specific embodiment, the number of channels in the feature maps of each layer in the network is set as follows: The number of feature channels in the input layer is 1 or 2, used to represent single-channel or dual-channel CT images; The number of channels in the feature extraction module is set to 32–64; The number of channels in the intermediate feature extraction layer increases progressively to 64–128; In the multi-domain feature fusion module, the number of channels in each parallel branch can be set to 32 to 64, and after fusion, they are uniformly mapped to a 64- or 128-dimensional feature space through channel compression.

[0061] The above channel configuration can be adjusted according to computing resources and application scenarios. For example, in edge computing or real-time processing scenarios, the number of channels can be appropriately reduced to reduce the number of model parameters. This adjustment method is not limited.

[0062] In the edge enhancement state-space model, the state dimension can be set to 16–64, and the state transition parameters and input mapping parameters are adaptively learned through end-to-end training. The differential convolution operator used for edge enhancement has a kernel size of 3×3 or 5×5, and the number of directions can be set to 4–8.

[0063] It should be noted that the above parameter ranges do not constitute a limitation on the present invention.

[0064] Example 2: Regarding the training method for the low-dose CT image denoising system based on edge enhancement provided in Embodiment 1 above, such as... Figure 3 As shown, it includes: A training dataset was constructed using low-dose CT images and corresponding normal-dose CT images. Using low-dose CT images as input information and normal-dose CT images as supervision information, the above-mentioned low-dose CT image denoising system based on edge enhancement is trained end-to-end to obtain a trained low-dose CT image denoising system.

[0065] This embodiment constructs a supervised training mechanism between low-dose CT images and corresponding standard-dose CT images to train the denoising system end-to-end. This enables the obtained model to accurately learn the statistical characteristics of low-dose imaging noise and its mapping relationship with the real structure, thereby obtaining stable and reliable denoising performance.

[0066] In a preferred embodiment, the training loss function is: ; in, Indicates the losses incurred during reconstruction. ; Represents structural similarity loss. ; This represents the denoised CT image output by the low-dose CT image denoising system. This represents the corresponding normal dose image. express……, Indicating structural similarity, and These represent the weights of the reconstruction loss and the structural similarity loss, respectively.

[0067] The loss function designed in this embodiment further includes structural similarity loss on the basis of reconstruction loss, which can constrain brightness, contrast and structure, thereby further improving the denoising effect of the system.

[0068] Example 3: A low-dose CT image denoising method based on edge enhancement, such as Figure 4As shown, it includes: The low-dose CT image to be processed is input into a pre-trained low-dose CT image denoising system to obtain a denoised CT image. The low-dose CT image denoising system that has been trained is obtained by the training method described in Example 2.

[0069] It is easy to understand that the low-dose CT images to be processed can be obtained by preprocessing the original low-dose CT images, such as normalization.

[0070] Example 4: A computer-readable storage medium includes a stored computer program that, when executed by a processor, implements the low-dose CT image denoising method based on edge enhancement provided in Embodiment 3 above.

[0071] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-dose CT image denoising system based on edge enhancement, characterized in that, include: Feature extraction module, encoder, decoder, and image reconstruction module; The feature extraction module is used to extract features from the input low-dose CT image to obtain shallow spatial features; The encoder includes multiple sequentially connected encoding layers for downsampling the shallow spatial features and extracting semantic features to obtain encoded features; the decoder includes multiple sequentially connected decoding layers for upsampling the encoded features and recovering features to obtain decoded features. Feature information from the output of each coding layer is introduced into the input of the corresponding decoding layer through skip connections; The image reconstruction module is used to reconstruct the decoded features and then fuse them with the input low-dose CT image through residual connection to obtain a denoised CT image. Each encoding layer and each decoding layer contains a multi-domain feature module, which includes a local feature modeling unit, a state space modeling unit, and a multi-domain feature fusion unit. The local feature modeling unit is used to perform convolution operations on the input features to obtain local features; The state space modeling unit is used to extract features from the input features through the state space model to obtain global features, and after extracting edge features, the two are fused into edge-enhanced global features. The multi-domain feature fusion unit is used to fuse local features with edge-enhanced global features.

2. The low-dose CT image denoising system based on edge enhancement as described in claim 1, characterized in that, The state space modeling unit includes a global feature extraction branch, an edge enhancement branch, and an activation function branch; The global feature extraction branch is used to extract features from the input features using a state-space model to obtain global features; The edge enhancement branch is used to extract edge features of the input features through the edge enhancement operator; The activation function branch is used to activate the features obtained by weighted summation of global and edge features.

3. The low-dose CT image denoising system based on edge enhancement as described in claim 2, characterized in that, The edge enhancement operator is a differential convolution, a directional convolution, or a combination of differential convolution and directional convolution.

4. The low-dose CT image denoising system based on edge enhancement as described in any one of claims 1 to 3, characterized in that, Before performing convolution operations on the input features, the local feature modeling unit further includes: The input features are subjected to discrete wavelet transform, which decomposes them into multiple sub-feature components of different frequency bands, and then the multiple sub-feature components of different frequency bands are concatenated.

5. The low-dose CT image denoising system based on edge enhancement as described in claim 4, characterized in that, The multi-domain feature fusion unit includes: a splicing unit, a dual-branch unit, and a residual connection unit; The stitching unit is used to stitch local features and edge-enhanced global features in the channel dimension; The first branch of the dual-branch unit is used to perform a 3×3 convolution operation on the spliced ​​features to achieve feature transformation and channel compression, and the second branch is used to perform two 3×3 convolution operations on the spliced ​​features to achieve feature refinement and enhancement. The residual connection unit is used to add the features output by the two branches in the dual-branch unit to the features output by the splicing unit, thereby achieving multi-domain feature fusion.

6. The low-dose CT image denoising system based on edge enhancement as described in any one of claims 1 to 3, characterized in that, Between the extraction module and the encoder, there is also a first operational-level adaptive learning module; Furthermore, a second operational-level adaptive learning module is also included between the decoder and the image reconstruction module.

7. The training method for the low-dose CT image denoising system based on edge enhancement as described in any one of claims 1 to 6, characterized in that, include: A training dataset was constructed using low-dose CT images and corresponding normal-dose CT images. Using low-dose CT images as input information and normal-dose CT images as supervision information, the low-dose CT image denoising system based on edge enhancement as described in any one of claims 1 to 6 is trained end-to-end to obtain a trained low-dose CT image denoising system.

8. The training method as described in claim 7, characterized in that, The training loss function is: in, Indicates the losses incurred during reconstruction. ; Represents structural similarity loss. ; This represents the denoised CT image output by the low-dose CT image denoising system. This represents the corresponding normal dose image. express……, Indicating structural similarity, and These represent the weights of the reconstruction loss and the structural similarity loss, respectively.

9. A low-dose CT image denoising method based on edge enhancement, characterized in that, include: The low-dose CT image to be processed is input into a pre-trained low-dose CT image denoising system to obtain a denoised CT image. The trained low-dose CT image denoising system is obtained by the training method described in claim 7 or 8.

10. A computer-readable storage medium, characterized in that, It includes a stored computer program, which, when executed by a processor, implements the low-dose CT image denoising method based on edge enhancement as described in claim 9.

Images