Low-dose CT image denoising method based on lightweight context Transform network

Abstract

The invention provides a lightweight low-dose CT (Computed Tomography) (LDCT) image denoising method based on a Transform network, and belongs to the technical field of medical image processing. The method provided by the invention is called CT-Denomer, global context information in a CT image and local context details of complex space change can be captured, and a high-quality image is generated. A core of a framework of the CT-Denoimmer is a Transform module, and the CT-Denoimmer comprises two key components, namely a multi-head convolution transpose attention (MDTA) and a hybrid context feedforward network (MCFN). The MDTA module captures global context information in an image with linear computational complexity, while the MCFN module pays attention to multi-scale local static and dynamic context information through a series of enhanced context Transformer blocks (eCoT). Besides, an intelligent operation attention layer (OWALs) is introduced to cope with a complex and changing noise mode in the LDCT image, so that the network can optimize the denoising effect in diversified scenes. According to the method, the fine structure retention and noise suppression of the denoising result of the LDCT image reach advanced performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and more specifically, relates to a low-dose CT image denoising method based on a lightweight contextual Transformer network. Background Art

[0002] Computed tomography (CT) is a key tool in clinical diagnosis and is widely used to detect diseases such as brain hemorrhage, visceral tumors, and bone injuries. However, the use of CT scans has raised significant health concerns as increased exposure to X-rays has been associated with increased risk of genetic diseases and cancer. Therefore, reducing patient exposure to X-rays during CT scans has become an issue that the medical community has worked hard to address. A simple and cost-effective way to reduce radiation exposure is to reduce the X-ray tube current or exposure time (mAs) during data acquisition. Although this method effectively reduces the radiation dose, it also results in the generation of low-dose CT (LDCT) images. These LDCT images are often affected by artifacts caused by noise, which can significantly affect the accuracy of medical diagnosis.

[0003] In order to solve the noise problem in LDCT images, researchers have developed many image quality enhancement methods, which can be roughly divided into three categories: chord domain preprocessing, model-based iterative reconstruction (MBIR), and image domain postprocessing. Although these methods are effective in suppressing noise, they rely heavily on manually designed priors. However, due to the non-uniformity and unpredictability of noise and content distribution in CT images, it is very difficult to design a suitable prior. Therefore, traditional methods often face difficulties in effectively balancing noise suppression and fine structure preservation.

[0004] In recent years, convolutional neural networks (CNNs) have made significant breakthroughs in noise suppression of LDCT images, largely due to their ability to learn generalizable priors from large datasets. For example, Chen et al. introduced RED-CNN, a residual encoder-decoder architecture that integrates autoencoders, deconvolution, and shortcut connections to achieve competitive performance in LDCT image denoising. Yang et al. proposed a network that uses a GAN network with Wasserstein distance and perceptual loss to further improve the effectiveness of denoising. Shan et al. introduced a modular adaptive processing neural network that enables radiologists to customize the denoising network for specific tasks. Zhang et al. proposed a multi-scale feature aggregation and fusion network with self-supervised multi-layer perceptual loss for texture-preserving LDCT denoising.

[0005] Although complex CNN architectures have made significant progress in denoising, most current CNN-based LDCT denoising networks still basically use convolution as their main building block. Convolution operators mainly perform local feature aggregation, which leads to two major deficiencies: limited receptive field and lack of flexibility to adapt to different input content. These limitations hinder their ability to capture the complex structure of CT images: 1. Global correlation: CT images are reconstructed from sinusoidal data by backprojection, which produces strong global correlation in structural features. Therefore, the network module needs to have a strong ability to model global context information, which is often poorly handled by traditional CNNs. 2. Local static context information: The tissue structure and lesion changes in local areas of CT images are usually stable or have minimal changes, resulting in rich static context information in the neighborhood. Therefore, when modeling the global context, the network module must also effectively represent this local static information. 3. Local dynamic context information: The inherent spatially varying statistical characteristics of CT image data lead to significant differences in content and noise patterns in different regions. Therefore, it is crucial for the network module to capture dynamic context information to effectively manage these differences. 4. Multi-scale structural information: The irregular presence of tissues and lesions, which can appear at different scales in CT images, further complicates the task. The network module should have the ability to represent multi-scale features to accurately restore the subtle textures and structures of these tissues and lesions. In order to capture the global correlation in the image, Transformer-based networks have been used for LDCT image denoising in recent years. The Transformer module relies on the self-attention mechanism to excel in capturing the interaction between distant pixels, making it a powerful alternative to basic network components. For example, Li et al. proposed a self-attention convolutional neural network (SACNN) that uses a self-supervised perceptual loss network for LDCT denoising. Luthra et al. pioneered the use of the Transformer model for LDCT denoising and proposed a method called Eformer, which implements a window-based attention mechanism and combines a trainable Sobel-Feldman operator to enhance edges. Zhang et al. decomposed the noisy LDCT image into high-frequency and low-frequency components and processed them through a three-layer encoder and a three-layer decoder containing a Transformer. Inspired by the Swin Transformer, Zhu et al. proposed STEDNet, which first uses a fully convolutional subnetwork to extract high-frequency information and then applies the Transformer structure for denoising. In addition, Zhang et al. proposed a novel LDCT image denoising method that combines the Transformer with a convolutional neural network (Transformer-CNN). This method uses the U-Net architecture, uses CNN as the encoder, and introduces a dual-path Transformer block as the decoder.Wang et al. proposed a convolution-free Token2Token Dilated Visual Transformer (CTformer) for LDCT denoising, which expands and shifts feature maps to capture long-range interactions.

[0006] While Transformer-based methods have shown promising results for denoising LDCT images, they still face some challenges due to their reliance on large-scale attention maps. These maps are crucial for evaluating the relationship between query-key pairs but are also computationally demanding, especially in low-level image denoising tasks. To alleviate this burden, current strategies typically restrict Transformer operations to small windows. However, this approach undermines the core advantage of Transformer: capturing important long-range dependencies, which is critical for understanding image structure. Another major challenge is capturing complex local contextual information, which is critical for detailed image recovery. Standard Transformers often have difficulty effectively modeling local nuances. To address these limitations, some recent methods combine convolution operations with Transformer mechanisms to form hybrid architectures to enhance modeling of global and local context. Despite these advances, the reliance on convolution operations remains a limitation. Convolution operations are inherently constrained by local translation invariance, limiting their ability to effectively capture local dynamic context and multi-scale features. As a result, the preservation of textures and details may be affected in LDCT image denoising. Summary of the invention

[0007] 1. Problem that the invention aims to solve The present invention aims to overcome the shortcomings of the prior art and introduces a low-dose CT image denoising method based on a lightweight contextual Transformer network, called CT-Denoimer. The core of the CT-Denoimer denoising network framework is a Transformer module consisting of two key building blocks: a multi-head self-attention (SA) mechanism to capture global context information, and a mixed context feedforward network (MCFN) to extract local multi-scale context information, including static and dynamic information, to effectively capture global correlations in the image as well as complex, spatially varying local context features, thereby improving the denoising performance of LDCT images. To cope with the high computational burden associated with generating query-specific attention maps in the standard Transformer, we adopt a multi-head convolutional transposed attention (MDTA) module to replace the traditional multi-head self-attention (SA). This MDTA module generates attention maps by computing the cross-covariance between feature channels. Importantly, before performing the cross-covariance calculation, MDTA integrates local context through deep convolutions to ensure that the global contextual relationship between pixels is implicitly modeled during the attention process. Secondly, the feed-forward network (FFN) in the standard Transformer usually relies on two fully connected layers separated by a nonlinear function, which is not enough to capture complex local context features. To address this limitation, the invention draws on existing achievements and introduces a contextual Transformer (CoT) module, which is designed to effectively handle static and dynamic local contexts. It is worth noting that the invention enhances CoT by aggregating static contexts from a wider receptive field, providing additional attention to establishing the relationship between queries and keys in a relatively small neighborhood, thereby enriching the representation of dynamic context. In addition, this module uses deep convolution instead of group convolution to reduce the computational burden of the network and form an enhanced CoT (eCoT). In order to deal with the inherent multi-scale structural complexity of CT images, this invention adopts a lightweight eCoT module to simulate multi-scale receptive fields. Finally, in order to cope with complex and changing noise patterns in LDCT images, such as stripe artifacts and granular noise, smart operational attention layers (OWALs) are introduced. This addition enables CT-Denoimer to dynamically adapt to different noise conditions, thereby optimizing noise removal in a variety of scenarios.

[0008] 2. Technical solution In order to achieve the above purpose, the present invention constructs a low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network, which mainly includes the following steps: Step 1: Collect conventional dose CT (NDCT) projection data from different patients, reconstruct them using the filtered back projection (FBP) method, obtain the corresponding NDCT images, and build the NDCT image library. Subsequently, add noise to the same set of NDCT projection data, reconstruct them using the FBP method, simulate the corresponding quarter-dose CT images, and build the LDCT image library. The NDCT and LDCT images form a paired training sample set for the training of the network model. Each training sample consists of paired LDCT and NDCT images, represented as (x, y), where x is the LDCT image and y is the corresponding NDCT image.

[0009] Step 2: Construct CT-Denoimer low-dose CT denoising network; Step 3: Train the CT-Denoimer low-dose CT denoising network; Step 4: Use the trained CT-Denoimer low-dose CT denoising network to denoise low-dose CT images.

[0010] The training set consists of a dataset containing some noisy LDCT images x X and a dataset of some clean NDCT images y Y The goal is to learn the mapping ,The CT-Denoimer network consists of four main parts: input embedding, latent feature encoding and decoding, feature refinement and output projection. Specifically, for a given size H×W LDCT images, a 3×3 convolution is first applied to extract shallow features. These features are then passed through a series of consecutive N The proposed method is enhanced by 0 smart operational attention layers (OWALs) that perform adaptive feature embedding to accommodate the diverse and complex noise patterns commonly found in LDCT images. The resulting embedded features are denoted as , and then through a four-layer symmetrical encoder-decoder for large-scale compression and reconstruction, it is converted into high-level features , where skip connections are used between the encoder and decoder stages to improve gradient backpropagation and image detail recovery. Then, another set of consecutive N 0 smart operation attention layers (OWALs) are used to further refine the high-level latent features so that they can robustly adapt to changing noise patterns. Finally, these refined features are projected into a residual image R through convolutional layers. The denoised image is obtained by adding this residual image to the initial LDCT image. ,Right now ;by L 1 Loss and Structural Similarity Index Metric Loss L SSIMA composite loss function is constructed to train the network model.

[0011] Step 3: Train the CT-Denoimer low-dose CT denoising network, and use paired LDCT and NDCT images as training sets to train and evaluate the constructed model. The images are cropped into small blocks of 64×64 pixels as input to improve the training effect. For the network structure configuration, the number of layers of the Transformer block is configured to {4, 6, 6, 8}, the number of attention heads per layer is set to {1, 2, 4, 8}, the number of OWALs blocks is set to 4, and the number of channels is set to 48. The initial learning rate is set to 1 / 10 during training. 4 , the learning rate is halved after every 6000 iterations. The batch size is set to 4 and the network is trained until no improvement in the loss is observed after 200 iterations.

[0012] Step 4 is the use of CT-Denoimer, a low-dose CT image denoising method based on a lightweight contextual Transformer network. First, a real low-dose CT image is obtained, and then it is input into the trained denoising network to obtain a denoised CT image.

[0013] 3. Beneficial effects Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects: (1) This paper proposes a low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network. This method can effectively capture the global correlation and complex and spatially varying local context details in CT images, thereby generating high-quality images and improving the preservation of subtle structures and details. (2) The present invention embeds a lightweight Transformer module in the network, which consists of two main components: a multi-convolutional head transposed attention (MDTA) module and a mixed context feedforward network (MCFN) module. The MDTA module efficiently captures the global correlation in the CT image structure with linear computational complexity. At the same time, the MCFN module manages multi-scale local static and dynamic context information through a series of enhanced contextual Transformer (eCoT) modules. This comprehensive approach provides a detailed representation of local and global structural features, effectively addressing the inherent spatial variation data distribution characteristics of CT images; (3) The present invention can demonstrate the most advanced performance in LDCT denoising applications. The LDCT denoised image obtained using the present invention can effectively preserve fine structures and suppress noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1It is the overall network structure diagram of CT-Denoimer of the present invention; Figure 2 Detailed structural diagram of the multi-head convolution transposed attention block of the present invention; Figure 3 Detailed structural diagram of the hybrid context feed-forward network block of the present invention; Figure 4 This is a detailed structural diagram of the intelligent operation attention layer of the present invention. DETAILED DESCRIPTION

[0015] The present invention is a low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network. The specific implementation process of the present invention will be described in detail below with reference to the accompanying drawings.

[0016] like Figure 1 As shown in the figure, the CT-Denoimer low-dose CT denoising network designed by the present invention first uses 3×3 convolution to extract shallow features, and then passes the features through a set of continuous N 0 intelligent operation attention layers (OWALs) for enhancement, followed by feature encoding and reconstruction through four layers of symmetrical encoder-decoder, each layer of which contains a set of proposed Transformer modules and uses pixel-unshuffle and pixel-shuffle for downsampling and upsampling operations; finally, another set of consecutive N 0 smart operation attention layers are used to further refine high-level latent features, and skip connections are added for improving gradient back-propagation and image detail recovery.

[0017] The novel Transformer module designed in this invention includes a multi-head convolutional transposed attention block (MDTA) and a mixed context feedforward network block (MCFN) to help the network effectively represent the complex spatial variation local context information inherent in CT images with acceptable computational complexity. Given the input feature X from the (l-1) layer (l-1) , the calculation of the lth layer Transformer is as follows:

[0018] Where X ′ l and X l Represent the outputs of MDTA and MCFN respectively; LN represents layer normalization.

[0019] Figure 2The module structure of MDTA is shown. The MDTA module first integrates local context through deep convolution and then performs cross-channel attention calculation instead of traditional self-attention in the spatial dimension. This method can efficiently capture global associations with linearly scalable computational complexity. It also ensures that the local contextual relationship between pixels is implicitly modeled when generating covariance-based attention maps. Specifically, given a feature map input , the MDTA module first normalizes it, and then generates the query Q, key K and value V through a combination of 1×1 convolution and 3×3 depth convolution, and the output dimension is , the process of MDTA is defined as:

[0020] in, φ First reshape the matrix into C×H×W dimension, and then perform a 1×1 convolution; α is a learnable scaling parameter. Multi-head attention is used to learn independent attention maps in parallel.

[0021] Figure 3 The module structure of MCFN is shown, which is to capture multi-scale local contexts through different receptive field sizes, which is used to effectively model complex, spatially varying local context structures in images. MCFN combines static and dynamic representations of different scales, enabling it to handle complex local structural patterns unique to images. Specifically, the module contains a series of small and lightweight 3×3 enhanced context transformer (eCoT) modules to simulate multi-scale receptive fields. The output features of the first and second eCoT modules in the figure essentially imitate the operations of 3×3 and 5×5 receptive fields, respectively. This process can be expressed as:

[0022] Among them, F in Indicates the content input to MCFN, Con 1×1 Represents 1×1 convolution, eCoT 3×3 Represents a 3×3 eCoT module, Cat represents a connection operation, ξ Represents the ReLU activation function.

[0023] The eCoT block in this module can effectively capture two types of local spatial context: static context is achieved through convolution, and dynamic context utilizes contextualized self-attention. These contexts are adaptively integrated to form a comprehensive representation of the input feature map. Unlike the original 3×3 CoT module, this module restricts convolutions to a 3×3 grid for static feature embedding. By enriching the representation of static details with contextual information from a wider neighborhood, it is possible to guide the generation of more detailed dynamic context information. To this end, this module uses 3×3 and 5×5 convolutions to aggregate keys with different receptive fields around the query, and utilizes the aggregated static context to guide the generation of dynamic context. In addition, in this eCoT design, depthwise convolutions are used to replace group convolutions to reduce computational overhead. Specifically, given an input vector , firstly, F c It is directly used as query Q, key K, and a 1×1 convolution is applied to obtain the value V. The key K is further processed by separate 3×3 and 5×5 depthwise convolutional layers to generate contextual representations within 3×3 and 5×5 grids, respectively, denoted as K 1 and K 2 . Some contextualized keys K 1 and K 2 is concatenated with the query Q to produce a query with 3 C Then, a 1×1 convolution with an activation function is used to extract rich contextual information and reduce the channel dimension to C Another 1×1 convolution without activation function is used to calculate the context matrix of dynamic attention. The learned dynamic matrix is ​​then combined with the value V to generate the dynamic attention output K 3 Finally, the two static contexts K 1 , K 2 and dynamic context K 3 are fused to produce the final output of the eCoT module. This module can be represented as

[0024] Here, Conv 1×1 ,DwConv 3×3 and Dwconv 5×5 They represent convolution operations with corresponding kernel sizes. ξ is the ReLU activation function. A is a local attention matrix with dimension , where C h is the number of dimensions, square brackets [] denote concatenation operations, and ⊗ denotes a local matrix multiplication operation.

[0025] Figure 4The figure shows the structure of the smart operation attention layer module (OWALs), which uses eight parallel operations to ensure comprehensive coverage of potential noise variations, including: standard depthwise convolutions with kernel sizes of 1×1, 3×3, 5×5, and 7×7, dilated depthwise convolutions with kernel sizes of 3×3, 5×5, and 7×7 (all using dilation rate 2), and average pooling with a 3×3 receptive field. The attention mechanism is used as a dynamic selector for these operation layers to adaptively determine the relevance of each operation based on the input features. For a given input feature , the OWALs module first performs channel-level global average pooling to obtain the feature vector z of a specific channel. This vector z then guides the generation of attention weights The results of these operations are weighted by the attention scores and concatenated in the channel dimension to form F m , the final output of the OWALs module is calculated as:

[0026] CT-Denoimer adopts L 1 Loss and L SSIM The model is trained using a composite loss function composed of L 1 loss is used to improve the fidelity of the generated network data. Compared with the mean square error loss, it does not over-punish large deviations or small errors between images, thereby better preserving details and textures. Its mathematical representation is:

[0027] in For denoised images, is the corresponding conventional dose image, Expected value The SSIM metric is designed to capture the perceptual difference between the denoised image and the reference image, making it consistent with human visual perception. L SSIM It is introduced to improve the recovery of detail texture and structural details in CT images, and its mathematical expression is: ; The mixed loss is expressed as:

[0028] Here, λ is a manually set parameter.

[0029] In the training phase, paired LDCT and NDCT images are used as training sets to train and evaluate the constructed model. The images are cropped into small patches of 64×64 pixels as input to improve the training effect, and the initial learning rate is set to 1 / 10 4 , the learning rate is halved after every 6000 iterations. The batch size is set to 4 and the network is trained until no improvement in the loss is observed after 200 iterations.

[0030] When using the trained CT-Denoimer low-dose CT denoising network: first obtain a real low-dose CT image, and then input it into the trained CT-Denoimer network to obtain a denoised CT image.

[0031] In summary, the present invention proposes a novel and efficient Transformer module, which aims to capture global associations and complex spatially varying local context features in CT images. Specifically, a Multi-Dconv Head Transposed Attention (MDTA) module is introduced to model global context information, and a Mixed Contextual Feed-Forward Network (MCFN) module is introduced to capture multi-scale local complex context features. In MCFN, local static and dynamic context information is extracted by using an Enhanced Contextual Transformer (eCoT) module. In addition, this invention design proposes a multi-scale structure to simulate different receptive fields through a series of lightweight eCoT modules. Based on this Transformer module, a low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network is constructed for LDCT image denoising. CT-Denoimer adopts an encoder-decoder architecture, and each encoder layer contains a series of proposed Transformer modules. The jump connections from shallow to deep layers facilitate better gradient back-propagation and enhance the preservation of structural details at both coarse and fine levels, thus helping to recover subtle details in CT images. Operation-Wise Attention Layers (OWALs) are introduced at the front and back ends of the network to further cope with the complex and diverse noise patterns inherent in CT images.

Claims

1. The present invention discloses a low-dose CT image denoising method based on a lightweight contextual Transformer network (abbreviated as CT-Denoimer), which is characterized by the following: step: Step 1: Prepare training samples: The data set is NDCT projection data from different patients, and the filtered back projection (FBP) method is used to reconstruct it to obtain the corresponding conventional dose CT (NDCT) images and build an NDCT image library; then, add noise to the same set of NDCT projection data, reconstruct it by the FBP method, simulate the corresponding quarter-dose CT images, and build a low-dose CT (LDCT) image library; NDCT and LDCT images form a paired training sample set for network model training. Each training sample consists of paired LDCT and NDCT images, represented as (x, y), where x is the LDCT image and y is the corresponding NDCT image; Step 2: Construct CT-Denoimer low-dose CT image denoising network; Step 3: Train the CT-Denoimer low-dose CT image denoising network; Step 4: Use the trained base CT-Denoimer low-dose CT image denoising network to achieve low-dose CT image denoising.

2. According to claim 1, the low-dose CT image denoising method CT-Denoimer based on lightweight contextual Transformer network is characterized in that: In step 2: Step 2.1, the constructed CT-Denoimer low-dose CT image denoising network consists of four main parts: input embedding, latent feature encoding and decoding, feature refinement and output projection, which is used to learn the dataset x containing some noisy LDCT images. X and a dataset of some clean NDCT images y Y Mapping of datasets composed of ; Step 2.2, the constructed lightweight low-dose CT image denoising network CT-Denoimer based on the Transformer network first uses a 3×3 convolution to extract shallow features; then, these features are passed through a set of continuous N 0 intelligent operational attention layers (OWALs) are used to enhance the feature embedding. These OWALs perform adaptive feature embedding, and the embedded features are recorded as , and then pass through a four-layer symmetrical encoder-decoder for feature encoding and reconstruction, converting it into high-level features , where skip connections are used between the encoder and decoder stages to improve gradient backpropagation and image detail recovery; then, another set of consecutive N 0 smart operational attention layers (OWALs) further refine the high-level latent features to make them robust to changing noise patterns; finally, these refined features are projected into a residual image R through convolutional layers; the denoised image is obtained by adding this residual image to the initial LDCT image ,Right now .

3. The low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network according to claim 2, characterized in that: Step 3.1, the encoder-decoder in the constructed CT-Denoimer low-dose CT image denoising network is embedded with a lightweight Transformer module, which consists of two main components: a multi-head convolutional transposed attention (MDTA) module and a mixed context feedforward network (MCFN) module to help the network effectively represent the complex spatial variation local context information inherent in CT images under acceptable computational complexity, and use pixel-unshuffle and pixel-shuffle to perform downsampling and upsampling operations; Step 3.2, the intelligent operation attention layer OWALs module uses eight parallel operations to ensure comprehensive coverage of potential noise changes, including: standard depth convolution with kernel sizes of 1×1, 3×3, 5×5, and 7×7, dilated depth convolution with kernel sizes of 3×3, 5×5, and 7×7 (all using dilation rate 2), and average pooling with a 3×3 receptive field; the attention mechanism is used as a dynamic selector for these operation layers to adaptively determine the relevance of each operation based on the input features; for a given input feature , the OWALs module first performs channel-level global average pooling to obtain the feature vector z of a specific channel; this vector z then guides the generation of attention weights Assigned to each operation in the set Ω, the results of these operations are weighted by the attention score and combined by concatenation in the channel dimension to form F m .

4. The low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network according to claim 3, characterized in that: Step 4.1, in the embedded lightweight Transformer module, the MDTA module first integrates the local context through deep convolution, and then performs cross-channel attention calculation. For a given feature map input ,The MDTA module first normalizes it, and then generates the query Q, key K and value V through a combination of 1×1 convolution and 3×3 depthwise convolution. This method can efficiently capture global associations with linearly scalable computational complexity; Step 4.2, in the embedded lightweight Transformer module, MCFN contains a series of small and lightweight 3×3 enhanced context transformer (eCoT) modules to simulate multi-scale receptive fields. The eCoT block can effectively capture two types of local spatial context: static context is implemented through convolution, and dynamic context uses the self-attention of the context to obtain these contexts and adaptively integrate them to form a comprehensive representation of the input feature map; In this eCoT design, for a given input vector First, F c It is directly used as query Q, key K, and a 1×1 convolution is applied to obtain the value V; the key K is further processed by independent 3×3 and 5×5 deep convolutional layers to generate contextual representations within 3×3 and 5×5 grids, respectively, denoted as K 1 and K 2 ; some contextualized keys K 1 and K 2 is concatenated with the query Q to produce a query with 3 C Then, a 1×1 convolution with an activation function is used to extract rich contextual information and reduce the channel dimension to C ; Another 1×1 convolution without activation function is used to calculate the context matrix of dynamic attention; then, the learned dynamic matrix is ​​combined with the value V to generate the dynamic attention output K 3 ; Finally, the two static contexts K 1 , K 2 and dynamic context K 3 The fusion is performed to produce the final output of the eCoT module; Step 4.3, when training the CT-Denoimer network, use L 1 loss and structural similarity loss to train the model, where L 1 loss is used to improve the fidelity of the generated network data, L SSIM The loss is introduced to improve the recovery of detailed texture and structural details in CT images.

5. The low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network according to claim 1, characterized in that: In step 3: Step 5.1: When training the CT-Denoimer low-dose CT image denoising network, use paired LDCT and NDCT images as training sets to train and evaluate the constructed model; crop the image into small blocks of 64×64 pixel size as input to improve the training effect. For the network structure configuration, the number of layers of the Transformer block is configured to {4, 6, 6,8}, the number of attention heads per layer is set to {1, 2, 4, 8}, the number of OWALs blocks is set to 4, and the number of channels is set to 48; during training, the initial learning rate is set to 1 / 10 4 , the learning rate is halved after every 6000 iterations; the batch size is set to 4, and the network is trained until no improvement in the loss is found after 200 iterations.

6. The low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network according to claim 1, characterized in that: In step 4: Step 6.1: Use of the low-dose CT image denoising method CT-Denoimer based on a lightweight contextual Transformer network: First, obtain real low-dose CT images. Next, input these real images into the trained CT-Denoimer denoising network to obtain the corresponding denoised CT images.

Images