Low-dose CT image denoising method based on hybrid representation learning

Abstract

This invention relates to a low-dose CT image denoising method based on hybrid representation learning. The invention employs a symmetrical encoder-decoder structure, incorporating hybrid representation blocks at each encoder and decoder stage. Convolutional mapping information is passed to a self-attention module, enabling information interaction within and between windows. Simultaneously, this is combined with image enhancement information processed by the Canny operator, maximizing denoising performance without compromising local information. By combining mean squared error and a ResNet-based multi-scale perceptual function, residual processing is performed between the final output noisy image and the low-dose CT image. This ensures denoising effectiveness while preserving more detailed texture information, thus improving the denoising effect of low-dose CT images. This invention effectively captures global and local detail information and resolves edge blurring issues, enhancing the detailed texture of CT images.

Description

Technical Field

[0001] This invention belongs to the field of medical image denoising technology, and relates to a low-dose CT image denoising method based on hybrid representation learning. Background Technology

[0002] Modern medical diagnostic methods have developed rapidly, and one of the most important tools is computed tomography (CT). It is a reliable and non-invasive medical imaging method that helps detect pathological abnormalities in the human body, including diseases of the head and neck, cardiovascular system, chest, abdomen, and pelvis. Besides diagnosis, CT is also very useful in guiding various clinical treatments, such as radiotherapy and surgery. However, the X-ray radiation from repeated CT scans can pose certain health risks, potentially leading to decreased immune function, metabolic abnormalities, genital damage, and an increased risk of leukemia, cancer, and genetic diseases. Therefore, in recent years, low-dose CT (LDCT) has been increasingly used. LDCT aims to reduce the X-ray dose during CT scans while ensuring the image quality meets diagnostic requirements. However, this can lead to problems such as increased noise, reduced contrast at edges, corners, and sharp features, and excessive image smoothing.

[0003] In recent years, extensive research has been conducted on denoising in LDCT. These methods can be broadly classified into three categories: 1. Filter-based methods: These methods address image blocks with many similarities within the same image by stacking non-local similar blocks. Examples include Non-Local Self-Similarity (NLM) image denoising algorithms, Block Matching 3D Filtering (BM3D) algorithms, and Wavelet Transform. 2. Model-based methods: Examples include the LNLTV algorithm. This is a combined local and non-local total variation regularized image denoising model that utilizes both local image structure and non-local similarity. It combines total variation (TV) and NLTV models to mitigate their shortcomings and leverage their advantages for image denoising. 3. Learning-based methods: The emergence of deep network methods has yielded more promising denoising results than filter-based, model-based, and traditional learning-based methods, and has become the mainstream approach. This method focuses on learning the latent mapping from noisy images to clean images. Due to the powerful feature learning and feature mapping capabilities of CNNs, CNN-based LDCT image denoising networks have achieved good results. However, there is still a risk that the denoised image may be over-smoothed, leading to the loss of key details and affecting the correct analysis of lesion sites. VIT (vision transformer) captures long-term dependencies in data through global self-attention, becoming another major breakthrough in the field of computer vision.

[0004] As a result, more and more transformer structures have emerged in the image processing field and achieved good results, including in medical image denoising. Recently, many studies have also attempted to fuse CNNs with transformers, achieving some success. However, combining the two may result in inaccurate information due to parameter weight sharing.

[0005] Therefore, while simple convolutional neural networks excel at local feature extraction, they suffer from weak global modeling capabilities. Window-based self-attention mechanisms, through complex spatial transformations and long-range feature dependencies, construct a global representation. However, they may overlook local feature details, potentially leading to overly smoothed CT images with insufficient detail and texture information, thus affecting diagnosis. Summary of the Invention

[0006] The purpose of this invention is to provide a low-dose CT image denoising method based on hybrid representation learning. The CT image is first subjected to edge enhancement processing, and then interaction is performed using depth convolution and local window self-attention to obtain CT image information between windows. The mean square error and ResNet-based multi-scale perceptual function are combined to finally output the noisy image and the low-dose CT image for residual processing. This method can retain more detailed texture information while ensuring the denoising effect of the denoised image, thereby improving the denoising effect of low-dose CT images.

[0007] Employing a symmetrical encoder-decoder structure, a Mixed Block is added to each encoder and decoder stage. This Mixed Block passes convolutional mapping information to the self-attention module, enabling information interaction within and between windows. Simultaneously, it is concatenated with image enhancement information processed by the Canny operator, maximizing noise reduction without compromising local information. Furthermore, skip connections fuse each output decoder block with its corresponding input encoder block across channels, preserving more texture details through the fusion of low-level and high-level information. This also prevents the vanishing gradient problem, improving training speed.

[0008] The specific steps of this invention are as follows:

[0009] Step 1: Construct a low-dose CT dataset:

[0010] A certain number of patients were selected, and low-dose CT images and their corresponding normal-dose CT images from different sites were chosen. Poisson noise was inserted until the noise level was close to 25% of the full dose, forming a CT image dataset (x, y). Here, x is the low-dose CT image, and y is the normal-dose CT image corresponding to x.

[0011] Step 2: Construct the denoising network model:

[0012] A symmetric encoder-decoder network model is constructed based on Mixed Block learning. This network model includes a pre-processing block, an input projection block, an input encoding block (MED), an intermediate layer mixed block, an output decoding block (MEU), and an output projection block.

[0013] In the image preprocessing stage, the input image is filtered by the Canny operator to generate an image with enhanced edge features. This image is used as auxiliary information to flow into subsequent network blocks to highlight the image edge features and increase the receptive field of the model.

[0014] Considering that directly inputting low-dose CT images into a neural network model can lead to over-smoothing and loss of some detailed features, an image preprocessing layer based on the Canny operator was designed to enhance image edge features. Specifically, the image filtered by the Canny operator is directly concatenated to the input mapping block, encoding block, and decoding block as auxiliary features to achieve image denoising.

[0015] The input mapping block includes a multilayer perceptron (MLP), which maps the input to a fixed dimension through two layers of neurons, facilitating subsequent representation operations.

[0016] The input encoding block includes an input hybrid representation block, an edge enhancement feature layer, a separable convolution, and a downsampling layer. First, the output of the input mapping block is used as a query set, key, and value to be input into the hybrid representation block for representation learning. Then, the edge enhancement feature map processed by the Canny operator is concatenated with the representation result and passed through a separable convolution. Finally, it is input into the downsampling layer to realize the encoding of the input module. The downsampling convolution kernel is 3×3, the stride is 2, and the padding is 1.

[0017] The intermediate layer hybrid representation block includes Window Self-Attention (WSA) and Depthwise Convolution (Dwconv). Window Self-Attention reduces dependence on the outside of the window, capturing correlations within the feature space window while achieving efficient computation. Depthwise Convolution, on the other hand, performs individual convolution operations on each channel to extract channel information without changing the number of convolution channels.

[0018] The intermediate layer hybrid representation block is designed. Through bidirectional interaction between the channel layer and the spatial layer, it not only solves the problem of limited receptive field in the window self-attention mechanism, but also eliminates the shortcomings caused by weight sharing in depth convolution. It realizes texture information interaction and information gain within and between windows, effectively enhancing the global modeling capability of CT images. In the actual configuration, the attention module window size is 7×7, and the depth convolution kernel size is 3×3.

[0019] The output decoding block (MEU) consists of an upsampling layer, an edge enhancement feature layer, a separable convolution, and an output mixed representation block. First, it undergoes an upsampling operation with a 4×4 kernel and a stride of 2. This is because during deconvolution, a checkerboard effect may occur if the kernel size is not divisible by the stride. Then, the edge feature image processed by the Canny operator is concatenated with it, followed by a separable convolution, and finally, a mixed representation block is added to learn more texture features.

[0020] The input blending representation block, intermediate layer blending representation block, and output blending representation block have the same structure. A skip connection is used between the input encoding block and the output decoding block. In this symmetrical structure, each output decoding block is fused with its corresponding input encoding block on the same channel. By fusing low-level and high-level information, more texture details are preserved. Furthermore, this also prevents the problem of vanishing network gradients.

[0021] The Output Projection block consists of an MLP layer that maps the output to 1×H×W, restoring the original image size.

[0022] Step 3: Data Augmentation

[0023] Medical images inherently suffer from problems such as complex sample data and sparse annotations. To further increase the number of training samples, this invention uses image enhancement strategies to perform various enhancement operations on the collected dataset with different probabilities, such as horizontal-vertical flipping and random cropping, to construct a larger number of images.

[0024] Step 4: Model Optimization

[0025] Two loss functions are used to optimize the model in order to further improve its performance.

[0026] First, the mean squared error (MSE) is used to converge the error to a minimum. The L1 loss function is expressed as follows:

[0027]

[0028] L1 is used to evaluate the pixel-level similarity between the denoised image and the real image, and whether the model can accurately restore the original image;

[0029] Where R(x) i ) represents a low-dose noise image x i The pure noise image mapped after residual learning; y i This represents the CT image corresponding to the normal dose.

[0030] Secondly, a ResNet-based multi-scale sensing function is used to implement the residual function between low-dose images and noisy images. The L2 loss function expression is as follows:

[0031] The L2 loss function is mainly used to evaluate the structural similarity between the image predicted by the model and the real image;

[0032] Where α represents the use of the classic feature extraction network ResNet50 as the feature extractor. This refers to the weights of ResNet50 on the ImageNet dataset that are frozen after the pooling layer is removed, R(x). i ) represents a low-dose noise image x i The pure noise image mapped after residual learning. i Represents a low-dose noise image x i The corresponding normal dose CT image.

[0033] The loss function is L = λ1L1 + λ2L2, where λ1 and λ2 are adjustable hyperparameters.

[0034] By continuously adjusting the learning rate and adjustable hyperparameters, the model is optimized to obtain the optimal parameters and output the optimal model.

[0035] Step 5: Select any low-dose CT image and put it into the denoising model. Output the final result to obtain the denoised low-dose CT image.

[0036] Compared with existing technologies, the innovation and advantages of the present invention using the above technical solution are as follows:

[0037] Applying depthwise convolution and window self-attention mechanism to representation learning of CT images can largely realize the interaction and information gain of texture information in CT images. By stitching it with the edge information image enhanced by the Canny operator, a hybrid feature map is formed, which enables the generated denoised image to better preserve the original overall structure and local texture details.

[0038] The original low-dose denoised image is first processed by the Canny operator to generate an edge feature image. After each downsampling or upsampling, the image is fused with the edge feature image through a MED and MEU module to enhance the details of the CT image, prevent excessive smoothing, and avoid losing crucial detail features. Simultaneously, skip connections between the corresponding MED and MEU blocks enable the network to fuse the feature maps of the encoder at the corresponding channel during each upsampling stage. By fusing low-level and high-level features, the network can retain more texture detail information contained in the high-level feature maps, thereby improving feature representation capabilities. Furthermore, skip connections can alleviate the gradient vanishing problem, accelerate network training, and thus improve network performance.

[0039] We employ commonly used mean squared error and a ResNet-based multi-scale perceptual loss function for model optimization. Using mean squared error accelerates convergence, while the multi-scale perceptual loss function makes the denoising learning function similar to the identity mapping, and residual mapping is easier to optimize. Through residual learning, we implicitly remove potentially clean images from the hidden layers, better identifying noise information and achieving better denoising results. Compared to directly outputting clean images, this avoids overfitting and significantly improves denoising performance.

[0040] The model minimizes the number of parameters, achieving better denoising results with the same parameter complexity.

[0041] This invention can remove noise and artifacts in low-dose CT images while preserving the overall structure and local texture details of the original image and solving the problem of edge blurring. It also prevents the image from being overly smoothed and losing some local information, thus facilitating clinical diagnosis. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall network structure of the present invention;

[0043] Figure 2 This is a schematic diagram of the input coding block (MED) structure of the present invention;

[0044] Figure 3 This is a schematic diagram of the output decoding block (MEU) structure of the present invention;

[0045] Figure 4 This is a schematic diagram of the intermediate layer mixed characterization block structure of the present invention;

[0046] Figure 5 This is a schematic diagram of low-dose CT images in the embodiment;

[0047] Figure 6 yes Figure 5 A schematic diagram of a low-to-medium dose CT image after denoising. Detailed Implementation

[0048] The present invention will be explained in detail below with reference to the accompanying drawings.

[0049] A low-dose CT image denoising method based on hybrid representation learning, the specific steps of which are as follows:

[0050] Step 1: Preprocess the CT image dataset:

[0051] The CT image dataset was divided into a training set, a validation set, and a test set. Each pair of paired images in the training and validation sets was randomly cropped to an input size of 128×128 pixels for training, in order to obtain local information of the images and expand the sample size.

[0052] Step 2: Construct a denoising network model, such as Figure 1 As shown, specifically:

[0053] A denoising model is constructed using an encoder-decoder symmetric structure. The aim is to input a noisy low-dose CT image into the model and denoise it to obtain a normal-dose image. The overall network model consists of six core modules: Pre-processing, Input Projection, Input Encoding (MED), Mixed Block, Output Decoding (MEU), and Output Projection.

[0054] Considering that directly inputting low-dose CT images into a neural network model may result in over-smoothing and loss of some detailed features, a Canny operator-based image preprocessing layer was first designed to enhance image edge features. In the image preprocessing stage, the input image is filtered by the Canny operator to generate an image with enhanced edge features, denoted as Hand-crafted features. This image serves as auxiliary information fed into subsequent network blocks to highlight image edge features and increase the model's receptive field.

[0055] First, input low-dose CT images X a This is placed into the input projection block, which consists of a multilayer perceptron (MLP): the input is mapped to a fixed dimension C×H×W, denoted as X, through two layers of neurons. c This facilitates subsequent representation calculations.

[0056] In this embodiment, the encoder stage includes three input code blocks (MEDs), such as... Figure 2As shown, the MED block is used to process the input low-dose CT image X. c First, the input is processed through a mixed block. The output of the input mapping block is used as a query set, key, and value to be input into the mixed block for representation learning. Then, it is concatenated with the hand-crafted features of the edge feature enhancement image. At this point, the size of the two-dimensional feature map becomes X. c ∈R 2C×H×W Then, a separable convolution is applied, in which a depthwise convolution is performed first, followed by a pointwise convolution. In the pointwise convolution, the number of output channels is set to fix the output channels, at which point the size of the 2D feature map reverts to X. c ∈R C×H×W Finally, it undergoes downsampling to extract deeper details. The downsampling convolution kernel is 3×3, with a stride of 2 and padding of 1. At this point, the size of the 2D feature map becomes... After passing through two more MED blocks, the above operation is repeated, at which point the size of the two-dimensional feature map becomes

[0057] A mixed representation block is used in the intermediate layer, consisting of window self-attention (WSA) and depthwise convolution (Dwconv). Window self-attention primarily reduces dependence on the outside of the window, capturing correlations within the feature space window while achieving efficient computation. Depthwise convolution performs individual convolution operations on each channel without changing the number of convolution channels to extract channel information. Through bidirectional interaction between the channel layer and the spatial layer, this not only solves the problem of limited receptive field in the window self-attention mechanism but also eliminates the shortcomings of weight sharing in depthwise convolution. It achieves texture information interaction and information gain within and between windows, effectively enhancing the global modeling capability of CT images. In the actual configuration, the attention module convolution kernel is 7×7, and the depthwise convolution kernel is 3×3. The image after passing through the intermediate layer is denoted as X. m .

[0058] like Figure 4As shown, specifically: Figure (a) represents the channel interaction stage, using a Global Average Convergence (GAP) layer, followed by two convolutional layers with a kernel size of 1×1 and a stride of 1. Batch normalization and simple gating block processing are performed within this stage, and finally, a sigmoid activation function is used to input the V of the self-attention module. The simple gating block performs channel bisection on the features, dividing them into two C / 2 channels, which are then multiplied. This replaces some linear functions with a non-linear approach, thus reducing computational cost. Figure (b) represents the spatial interaction stage, also using two convolutional layers with a kernel size of 1×1 and a stride of 1. Batch normalization and simple gating block processing are also performed within this stage, and finally, a sigmoid activation function is used to generate a spatial self-attention map, providing a larger spatial range for the depthwise convolutional part. This new module is named MixedBlock and can be represented as:

[0059] X′ i+1 =MIX(LN(X) l ),WS1,D345nv)+X l ;

[0060] X l+1 =FFN(LN(X′l +1 ))+X′ l+1 ;

[0061] The `MIX()` function represents the function for fusing features from WSA and Dwconv. `LN` represents layer normalization, while `FFN()` is an MLP consisting of two linear layers with a simple gating block in between. l X′ represents the input feature tensor of the l-th layer. l+1 This represents the result after processing through the interaction of window self-attention and depthwise convolution, and is denoted as an intermediate state; X l+1 This indicates that the output after l layers of hybrid representation processing can also be used as the input for the next hybrid representation block.

[0062] In the decoder stage, such as Figure 3 As shown, it specifically consists of three output decoding blocks (MEUs). Each MEU block is a block that processes the image X after the intermediate Mixed Block. ( First, an upsampling operation is performed with a 4×4 kernel and a stride of 2. This upsampling is achieved using transposed convolution. The feature map size then... After one upsampling and amplification At this point, after concatenating with the downsampled edge feature-enhanced image (Hand-crafted features), the feature map size becomes... Then, a separable convolution is applied to fix the output channels, at which point the size of the 2D feature map reverts to its original size. Finally, an output hybrid representation block is added to learn more texture features. The same steps are then performed on the two MEU blocks to revert the size of their 2D feature maps back to R. C×H×W The resulting image is denoted as X. u .

[0063] Next, image X u The output is placed in the Output Projection block, which consists of a single MLP layer. This maps the output to 1×H×W, restoring the original image size. The output is a purely noisy image, denoted as X. o .

[0064] The input mixed representation block, intermediate layer mixed representation block, and output mixed representation block have the same structure. A skip connection is used between the input encoding block and the output decoding block. In the symmetrical structure, each output decoding block is fused with its corresponding input encoding block on the channel. By fusing low-level and high-level information, more texture details are preserved. Secondly, it can also prevent the problem of vanishing gradients in the network and accelerate network training.

[0065] Finally, a residual processing is performed between the original image and the denoised CT image X. i .

[0066] Step 3: Data Augmentation

[0067] A certain number of patients were selected, and low-dose CT images from different parts of their bodies were selected. Data augmentation operations were performed on the collected dataset with different probabilities, such as horizontal-vertical flipping and random cropping, to construct a larger number of images to form the training set and the test set.

[0068] Step 4: Model Optimization

[0069] To improve the model's performance, two loss functions are constructed to optimize it and further enhance its capabilities.

[0070] First, the mean squared error (MSE) is used to converge the error to a minimum. The L1 loss function is expressed as follows:

[0071]

[0072] L1 is used to evaluate the pixel-level similarity between the denoised image and the real image, and whether the model can accurately restore the original image;

[0073] Where R(x) i ) represents a low-dose noise image x iThe pure noise image mapped after residual learning; y i This represents the CT image corresponding to the normal dose.

[0074] Secondly, a ResNet-based multi-scale sensing function is used to achieve residual functionality between low-dose images and noisy images. The L2 loss function expression is as follows:

[0075] The L2 loss function is used to evaluate the structural similarity between the image predicted by the model and the real image;

[0076] Where α represents the use of the classic feature extraction network ResNet50 as the feature extractor. This refers to the weights of ResNet50 on the ImageNet dataset that are frozen after the pooling layer is removed, R(x). i ) represents a low-dose noise image x i The pure noise image mapped after residual learning. i This represents the CT image corresponding to the normal dose.

[0077] The loss function is L = λ1L1 + λ2L2, where λ1 and λ2 are adjustable hyperparameters.

[0078] This implementation uses the Adam optimizer to train and update the network weights. The model is continuously optimized by adjusting the learning rate and adjustable hyperparameters. The denoising performance of the trained model is evaluated from multiple dimensions. The parameters of the best-performing denoising model from the training set are retained.

[0079] Step 5: Denoising low-dose CT images: Use a trained denoising network to denoise the images. Figure 5 The low-dose CT images shown are denoised to generate images like... Figure 6 The CT image shown has good noise reduction.

[0080] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A low-dose CT image denoising method based on hybrid representation learning, characterized in that: Specifically, the steps include the following: Step 1: Construct a low-dose CT dataset: A certain number of patients were selected, and low-dose CT images and their corresponding normal-dose CT images from different parts of their bodies were selected. Poisson noise was inserted until the noise level was close to 25% of the full dose, forming a CT image dataset (x, y); where x is the low-dose CT image and y is the normal-dose CT image corresponding to x. Step 2: Construct the denoising network model: A symmetric encoder-decoder network model is constructed based on hybrid representation learning. This model includes a preprocessing block, an input mapping block, an input encoding block, an intermediate layer hybrid representation block, an output decoding block, and an output mapping block. In the image preprocessing stage, the input image is filtered by the Canny operator to generate an image with enhanced edge features. This image is used as auxiliary information to flow into subsequent network blocks to highlight the image edge features and increase the receptive field of the model. The input mapping block includes a multilayer perceptron: it maps the input to a fixed dimension through two layers of neurons, which facilitates subsequent representation operations; The input encoding block includes an input hybrid representation block, an edge enhancement feature layer, a separable convolution, and a downsampling layer: First, the output of the input mapping block is used as a query set, key, and value input to the hybrid representation block for representation learning. Then, the edge enhancement feature map processed by the Canny operator is concatenated with the representation result, passed through a separable convolution, and finally input into the downsampling layer to realize the encoding of the input module. The intermediate layer hybrid representation block includes window self-attention and depthwise convolution; window self-attention can reduce the dependence on the outside of the window and capture the correlation inside the feature space window while efficiently computing; depthwise convolution performs a separate convolution operation on each channel to mine channel information without changing the number of convolution channels. The design of the intermediate layer hybrid representation block, through the bidirectional interaction of the channel layer and the spatial layer, not only solves the problem of the limited receptive field of the window self-attention mechanism, but also eliminates the shortcomings caused by the weight sharing of deep convolution. It realizes the interaction and information gain of texture information inside and between windows, effectively enhancing the global modeling capability of CT images. The output decoding block includes an upsampling layer, an edge enhancement feature layer, a separable convolution, and an output hybrid representation block. First, it undergoes an upsampling operation, then it is concatenated with the edge feature image processed by the Canny operator, then passes through a separable convolution, and finally adds a hybrid representation block to learn more texture feature knowledge. The input blending representation block, intermediate layer blending representation block, and output blending representation block have the same structure; a skip connection is used between the input encoding block and the output decoding block. In the symmetrical structure, each output decoding block is fused with its corresponding input encoding block on the channel. By fusing the low-level information with the high-level information, more texture details are preserved. Secondly, it can also prevent the problem of vanishing gradients in the network and accelerate network training; The output mapping block includes an MLP layer that maps the output to 1×H×W, restoring the original image size; Step 3: Data Augmentation To further increase the number of training samples, the collected dataset is horizontally and vertically flipped or randomly cropped to construct a larger number of images; Step 4: Model Optimization Two loss functions are used to optimize the model in order to further improve its performance; First, the mean squared error (MSE) is used to converge the error to a minimum. The L1 loss function is expressed as follows: L1 is used to evaluate the pixel-level similarity between the denoised image and the real image, and whether the model can accurately restore the original image; Where R(x) i ) represents a low-dose noise image x i The pure noise image mapped after residual learning; y i Represents a low-dose noise image x i Corresponding normal dose CT images; Secondly, a ResNet-based multi-scale sensing function is used to implement the residual function between low-dose images and noisy images. The L2 loss function expression is as follows: The L2 loss function is used to evaluate the structural similarity between the image predicted by the model and the real image; Where α represents the use of the classic feature extraction network ResNet50 as the feature extractor. This refers to the weights of ResNet50 on the ImageNet dataset that are frozen after the pooling layer is removed, R(x). i ) represents a low-dose noise image x i The pure noise image mapped after residual learning; y i Represents a low-dose noise image x i Corresponding normal dose CT images; The loss function is L = λ1l1 + λ2l2, where λ1 and λ2 are adjustable hyperparameters. The model is optimized by continuously adjusting the learning rate and adjustable hyperparameters to obtain the optimal parameters and output the optimal model. Step 5: Select any low-dose CT image and put it into the optimized denoising model. Output the final result to obtain the denoised low-dose CT image.

2. The low-dose CT image denoising method based on hybrid representation learning as described in claim 1, characterized in that: In step two, during the image preprocessing stage, the image after Canny operator filtering is directly concatenated to the input mapping block, encoding block, and decoding block as auxiliary features to achieve image denoising.

3. The low-dose CT image denoising method based on hybrid representation learning as described in claim 1, characterized in that: The input encoding block first processes the input image through an input mixing representation block, then concatenates it with the image filtered by the Canny operator to enhance detailed texture information. Next, it passes through a separable convolution, which decomposes the standard convolution into depthwise convolution and pointwise convolution to process spatial and channel information respectively. Using separable convolution reduces computation while adjusting the number of channels. Finally, it performs downsampling to reduce the size of the feature map, thereby extracting more detailed features. The downsampling convolution kernel is 3×3, with a stride of 2 and padding of 1. The output decoding block first upsamples the image with a 4×4 convolution kernel and a stride of 2. Then it is stitched with an edge enhancement feature image, followed by a separable convolution, a fixed output size, and finally an output hybrid representation block to further enhance the feature representation capability of the CT image.

4. The low-dose CT image denoising method based on hybrid representation learning as described in claim 1, characterized in that: The bidirectional interaction between the channel layer and the space layer is specifically as follows: In the channel interaction stage, a global average convergence layer is used, followed by two convolutional layers with a kernel size of 1×1 and a stride of 1. Batch normalization and simple gating block processing are performed within these layers. Finally, a sigmoid activation function is used to pass the data into the V of the self-attention module. The simple gating block divides the features into two C / 2 channels, which are then multiplied. This non-linear approach replaces some linear functions, reducing computational cost. In the spatial interaction stage, two convolutional layers with a kernel size of 1×1 and a stride of 1 are also used. Batch normalization and simple gating block processing are performed within these layers, and a sigmoid activation function is used to generate a spatial self-attention map, providing a larger spatial range for the depthwise convolutional part. This new module is named the hybrid representation block and is represented as follows: X′ l+1 =MIX(LN(X l ),WSA,Dwconv)+X l ; X l+1 =FFN(LN(X′ l+1 ))+X′ l+1 ; The MIX() function represents the function that fuses features from local window self-attention WSA and depthwise convolution Dwconv; LN represents layer normalization, while FFN() is an MLP consisting of two linear layers with a simple gating block in between; X l X′ represents the input feature tensor of the first layer. l+1 This represents the result after processing through the interaction of window self-attention and depthwise convolution, and is denoted as an intermediate state; X l+1 This indicates that the output after one layer of hybrid representation processing can also be used as the input for the next hybrid representation block.

Images