Projection data noise removal method for multi-energy CT (Computed Tomography)

Abstract

The invention discloses a multi-energy CT-oriented projection data noise removal method, which relates to the technical field of image processing, is used for solving the problems that the noise content of an imaging result is increased and the signal-to-noise ratio is reduced due to the fact that the radiation dose of the current multi-energy CT is reduced, and comprises the following steps: constructing a data set for low-dose multi-energy CT projection data denoising; constructing a projection data denoising network model; constructing a mixed loss function for supervising the training of a denoising network model; training data are imported, an optimizer is selected, related hyper-parameters are set, and then model training is carried out; carrying out convergence judgment on the training process of the projection data denoising network model; updating a network model parameter weight; and processing the low-dose multi-energy CT projection data by using the optimal network model. According to the multi-energy CT projection data denoising model constructed by the invention, more detail features can be reserved while low-dose multi-energy CT projection data noise is removed, and the quality of a reconstructed image is effectively improved.

Description

Technical Field

[0001] This patent relates to the field of multi-energy computed tomography (CT). Aiming at the problem that the noise content of imaging results increases and the signal-to-noise ratio decreases due to the reduction of radiation dose in multi-energy CT, a noise removal method based on deep learning technology is proposed. Background Art

[0002] Multi-energy computed tomography can distinguish the attenuation of X-ray photons of different energies in the human body. It has the characteristics of high energy spectrum resolution and radiation dose efficiency. It has become one of the main development directions in the field of CT technology. However, X-rays are a proven carcinogen. As early as 2002, the International Commission on Radiological Protection pointed out that the absorbed dose of CT to tissues usually approaches or exceeds the known level, thereby increasing the probability of cancer. In response to this problem, the International Commission on Radiological Protection recommends that the principle of radiation protection optimization should be followed, that is, to minimize the radiation dose without affecting the patient's diagnosis and prognosis. As the radiation dose decreases, non-ideal effects such as quantum noise and photon starvation have a more serious impact on imaging quality, resulting in reduced reconstructed image quality and decreased signal-to-noise ratio, which seriously affects the observation of lesion tissue during diagnosis.

[0003] At present, low-dose CT image denoising algorithms are mainly divided into projection data-based processing methods, iterative reconstruction methods, and post-processing methods based on reconstructed images. Among them, although the iterative reconstruction method directly achieves high-quality reconstruction from projection data to CT images, this type of method strongly relies on prior information and consumes a lot of computing resources. The post-processing method based on reconstructed images usually uses image processing methods such as dictionary learning and non-local filtering to process the reconstructed low-dose CT images, but it still has shortcomings in retaining detailed structures and features. Compared with these two methods, the processing method based on projection data can directly use the noise information of projection data to study the noise distribution model, which has higher computational efficiency. In recent years, data-driven deep learning technology has developed rapidly and has been widely used in the field of medical image processing. However, convolutional neural networks often use the deepest visual features through the superposition of convolution blocks, and fail to explore the intrinsic connection between feature information at different levels. In addition, the low-dose multi-energy CT projection data in each energy interval are highly correlated, and the projection data itself has self-similarity and contains special internal structures. Most of the existing low-dose multi-energy CT denoising algorithms ignore these characteristics, resulting in residual noise in the processed reconstructed images. Therefore, making full use of the multiple characteristics of projection data can help further suppress the noise in low-dose multi-energy CT projection data. Summary of the invention

[0004] The patent of this invention aims at the detection deviation of the multi-energy CT scanning system caused by noise. It removes noise based on the projection domain scanning data of the detection system, thereby reducing the impact of the detection deviation on the imaging results, improving the detection accuracy of the CT system, and improving the quality of the reconstructed image, thereby providing an accurate reference for subsequent medical diagnosis.

[0005] In order to achieve the above object, the present invention adopts the following technical scheme:

[0006] The present invention provides a method for removing noise from projection data of low-dose multi-energy CT, comprising the following steps:

[0007] Step 1: Construct a dataset for denoising low-dose multi-energy CT projection data;

[0008] Step 2: Construct a projection data denoising network model: the network model structure includes m shift and local enhancement Transformer window modules and n residual convolution modules;

[0009] Step 3: Construct a hybrid loss function for supervised denoising network model training: The hybrid loss function consists of Huber loss, feature-aware loss, and non-local total variation loss;

[0010] Step 4: Import training data, select the optimizer and set relevant hyperparameters, and then perform model training;

[0011] Step 5: Convergence judgment of the training process of the projection data denoising network model: if the loss function does not converge, adjust the learning rate and modify the hyperparameter settings, and go to step 4; if the loss function converges, stop training, save the best model weights, and go to step 6;

[0012] Step 6: Update the network model parameter weights and save the model when the loss function converges;

[0013] Step 7: Use the optimal network model to process the low-dose multi-energy CT projection data, import the noisy projection data into the optimal network model for calculation, and then use the filtered back projection algorithm to obtain the multi-energy CT image after noise removal.

[0014] Furthermore, the method for implementing the low-dose multi-energy CT projection data denoising data set in step 1 is: using phantom generation software to obtain a human body model containing X-ray energy attenuation information of various parts of the human body, and using a multi-energy CT detector simulation model to scan the human body model under conventional dose and low dose conditions, respectively, to obtain multi-energy CT projection data under conventional dose and low dose.

[0015] Furthermore, the specific implementation method of step 2 includes the following steps:

[0016] Step 2.1, construct a shifted windows Transformer Unit (STU). The STU part contains a shifted-window-based multi-head self-attention (SW-MSA) unit, a LayerNorm layer, and a multi-layer perceptron (MLP) layer. The input feature vector is sent to the SW-MSA unit through the LayerNorm layer, and the feature vector processed by the SW-MSA unit is residually connected to the input. The SW-MSA unit divides the input features into different window partitions, and shifts the window partitions during the calculation process, and then moves the window position in the next layer to generate new window partitions. The self-attention calculation in the new window spans the window boundary in the previous layer to achieve connection with the window of the previous layer. The LayerNorm layer is used before each SW-MSA unit and MLP layer for normalization. The SW-MSA unit adopts a multi-head self-attention mechanism structure, which can extract multiple groups of features in parallel in several feature spaces. The self-attention calculation process of each feature space is as follows:

[0017]

[0018] Among them, Q τ , K τ and V τ are the query vector, key vector and value vector of the feature space τ respectively, and B τ is the relative position deviation;

[0019] Step 2.2, build a locally enhanced window Transformer unit (LTU): This part introduces the convolution operator based on the STU structure, and replaces the original MLP unit with a locally enhanced feed-forward network (LeFF) unit built on a deep convolutional layer: the unit first uses a fully connected layer to increase the input data to a higher dimension, then performs deep convolution on the high-dimensional feature data, and finally uses a fully connected layer to restore it to the initial dimension and add it to the input data;

[0020] Step 2.3, construct the shift and local enhancement Transformer window module: This module first extracts the input data features through a 1×1 convolution layer, then splits the data features through a split operation, and inputs the feature information into STU and LTU respectively: This operation converts the feature information X of the i-th (1<i≤m) level SLBi Split equally into two feature map groups X i1 and X i2 , and processed by STU and LTU respectively to obtain output features and The process is represented as follows:

[0021] X i1 , X i2 =Split(Conv(X i ))

[0022]

[0023] The output features of the two parts and After being merged through the connection layer and feature fused through the convolution layer, it is added with the input data of the SLB at this level and output to the next module; the final output Z of the i-th level SLB is i It is expressed as follows:

[0024]

[0025] Step 2.4: Build a residual convolution module: This module consists of a 3×3 convolution layer, a batch normalization layer, and a Leaky ReLU activation layer.

[0026] Step 2.5: Build a denoising network model based on the shifted and locally enhanced Transformer window module and residual convolution module constructed in steps 2.3 and 2.4.

[0027] Furthermore, the specific implementation method of step 3 includes the following steps:

[0028] Step 3.1, construct the Huber loss function, which can be expressed as:

[0029]

[0030] Among them, P GT and P Predict They represent the label and denoising network model output data respectively, and ε is the regularization constant;

[0031] Step 3.2, construct feature-aware loss function: Feature-aware loss uses a network model with pre-trained weights as a feature extractor, inputs the model output data and labels to be trained into the feature extraction model respectively, and extracts the output features of some hidden layers, describes the difference between the two and minimizes it to reduce the loss of detail features: First, construct a feature extraction model, and then use P GT and P PredictThey are input into the feature extraction model respectively and the feature maps of the two sets of data in the feature extraction model are obtained respectively. Finally, the content loss and style loss are calculated to describe the difference between the feature maps of the two sets of data. The feature perception loss is obtained after weighted average. The calculation process is as follows:

[0032]

[0033] Among them, L is the total number of layers of feature maps extracted in the feature extraction model, h l 、w l and c l The output data and label are the lth layer feature maps in the feature extraction model respectively and The length, width and number of channels of , G(·) is the operator for calculating the Gram matrix of the extracted feature map data, and λ is the weight parameter after the normalization of the two parts of the loss;

[0034] Step 3.3: Construct a non-local total variation loss function, which can be expressed as follows:

[0035] L NLTV =Z NLTV (P Predict )+μZ NLTV (P Predict -P GT )

[0036] Among them, Z NLTV (·) is the operator for calculating the non-local total variation of a given image I, μ is the weight parameter, when calculating the non-local total variation of a given image I, assuming that the size of the divided image block is p×p, for the target image block in the image The middle pixel is located at image position x. By calculating the similarity weight, the local range r×r around x and the pixels in U are screened. x Similar image patches U x with U y The similarity weight is expressed as follows:

[0037]

[0038] In the formula, g σ is a Gaussian convolution kernel with a standard deviation of σ, h0 is the global filter parameter, and the NLTV of image I is expressed as:

[0039]

[0040] In the formula, Ω represents the overall space of a given image, Ω x Represents the image domain space around pixel x;

[0041] Step 3.4: Construct a mixed loss function based on the losses of each part in steps 3.1 to 3.3: The specific implementation method is to weighted average the losses of each part, which is specifically expressed as follows:

[0042] L Total =L Huber +λ1L Feature +λ2L NLTV

[0043] Among them, λ1 and λ2 are weight parameters respectively.

[0044] Furthermore, the specific implementation method of step 4 is: importing training data, selecting an optimizer and setting the momentum and weight decay size, selecting a learning rate adjustment strategy and setting the initial learning rate size, and using a hybrid loss function to guide model training.

[0045] The beneficial effects of the present invention are as follows:

[0046] The present invention proposes a projection data noise removal method based on a two-stream Transformer to improve the quality of reconstructed images. This method utilizes the non-local self-similarity of projection data to enable the network model to remove projection data noise while retaining the original internal structure, thereby effectively improving the signal-to-noise ratio and detail structure clarity of the reconstructed image. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0048] Figure 1 A flowchart of a method for removing noise from projection data of multi-energy CT is provided for an embodiment of the present invention;

[0049] Figure 2 A network model structure diagram of a projection data noise removal method for multi-energy CT is provided for an embodiment of the present invention;

[0050] Figure 3 A structural diagram of the shifting and local enhancement Transformer window module in a network model of a projection data noise removal method for multi-energy CT is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0052] The present invention is further described below, but the present invention is not limited to the scope described here. Simple changes made by ordinary technicians in the field based on the concept of the present invention should be within the scope of protection claimed by the present invention. The following is a detailed description in conjunction with the above-mentioned drawings:

[0053] Step 1: Construct a dataset for denoising low-dose multi-energy CT projection data.

[0054] In this embodiment, the required human body model is generated using XCAT phantom software, an edge-incident detector simulation model is used to simulate the X-ray detection process, and a GE_Maxiray_125 tube is used as the ray source, with a driving voltage of 120 kVP. The energy thresholds of the detector are set to 80 and 100 keV, respectively, and the three energy intervals are 20-80 keV, 80-100 keV, and 100-120 keV, respectively. The slice thickness is 1.5 mm, the scanning range is 360°, the position of the X-ray source is 59.5 cm away from the rotation center, and the low-dose scanning condition is simulated by reducing the photon flux, and the initial incident photon number is set to 1×10 5 , the number of incident photons under conventional dose scanning conditions is 1×10 7 The number of human models generated is 10, and the number of slices obtained for each human model is 120. The sampling condition is set to evenly distribute 367 detector boxes between 0 and 359° and sample 360 ​​projection views. The size of the sinogram after each slice scan is 367×360. A total of 1200 data sets are obtained and divided into training sets and test sets at a ratio of 9:1. The training set needs to be further divided into training sets and validation sets at a ratio of 9:1.

[0055] Step 2: Follow Figure 2 The structure shown here constructs a projection data denoising network model. The specific processing method of this step is as follows:

[0056] Step 2.1, build STU, the structure is as follows Figure 3As shown. The STU part includes SW-MSA, LayerNorm layer and MLP layer. In this embodiment, the window size of SW-MSA is divided into 4×4, the window shift step is set to 1, the number of self-attention layer branches of the multi-head self-attention mechanism is 8, and the number of output channels of SW-MSA is equal to the input. The MLP layer consists of 2 fully connected layers and 1 ReLU layer. The number of output channels of the first fully connected layer is 4 times the number of input channels, and the number of output channels of the second fully connected layer is equal to the number of input channels of the first fully connected layer.

[0057] Step 2.2: Build LTU, the structure is as follows Figure 3 As shown. The LTU part includes SW-MSA, LayerNorm layer and LeFF layer. In this embodiment, the window size of SW-MSA is divided into 4×4, the window shift step is set to 1, the number of self-attention layer branches of the multi-head self-attention mechanism is 8, and the number of input channels of SW-MSA is equal to the output. The LeFF layer consists of 2 fully connected layers, 1 deep convolutional layer, 3 Leaky ReLU activation layers and 3 batch normalization layers. The number of output channels of the first fully connected layer is 4 times the number of input channels, the number of output channels of the deep convolutional layer is equal to the input, and the number of output channels of the second fully connected layer is equal to the number of input channels of the first fully connected layer.

[0058] Step 2.3: Construct a shift and local enhancement Transformer window module. In this embodiment, the input data features are first extracted through a 1×1 convolutional layer, and then the data features are split through a split cascade operation, and the feature information is input to STU and LTU respectively. This operation evenly splits the feature information X into two feature map groups X1 and X2, and the output features X are obtained by STU and LTU processing respectively. S and X L Then the output features X of the two parts are S and X L After merging through the connection layer and fusion through the 1×1 convolution layer, the output features are added to the input data. The number of output channels of the two 1×1 convolution layers is equal to the input, and the splitting strategy of the split cascade operation is average random splitting.

[0059] Step 2.4: Construct a residual convolution module. In this embodiment, the module consists of two groups of 3×3 convolution layers, a batch normalization layer, and a Leaky ReLU activation layer. The number of output channels of the two 3×3 convolution layers is equal to the number of input channels.

[0060] Step 2.5: Follow Figure 1The structure shown in FIG. 2 is a denoising network model based on the shifted and locally enhanced Transformer window modules and residual convolution modules constructed in steps 2.3 and 2.4. In this embodiment, the number of shifted and locally enhanced Transformer window modules is 4, and the number of residual convolution modules is 4. The denoising network model first uses a 3×3 convolution layer for feature extraction, with an input channel number of 3, a convolution step size of 1, and an output channel number of 16. Then, it is input into the cascaded 4 groups of shifted and locally enhanced Transformer window modules, and the number of channels of the 4 groups of shifted and locally enhanced Transformer window modules is set to 32, 64, 128, and 256, respectively. The output feature map is input into the cascaded 4 groups of residual convolution modules, and the number of channels of the 4 groups of residual convolution modules is set to 256, 128, 64, and 32, respectively. Finally, a 3×3 convolution layer is used for feature compression, with an input channel number of 32, a convolution step size of 1, and an output channel number of 3.

[0061] Step 3: Construct a hybrid loss function for supervised denoising network model training. The hybrid loss function consists of Huber loss, feature-aware loss, and non-local total variation loss. The specific processing method of this step is as follows:

[0062] Step 3.1: Construct Huber loss function. In this embodiment, the regularization constant ε is set to 1×10 -6 .

[0063] Step 3.2, construct a feature perception loss function. In this embodiment, the feature extraction model used is ResNet-50, the pre-trained weight is ImageNet, the number of layers used to extract feature maps is 4, and the extracted layers are the 7th, 16th, 26th and 32nd layers respectively. The normalized weight parameter λ of the feature content loss and feature style is set to 0.5.

[0064] Step 3.3, construct a non-local total variation loss function. In this embodiment, the weight parameter μ is set to 0.5. When calculating a given non-local total variation, the size of the divided image block is 4×4. The surrounding local range of the similarity weight of the middle pixel of the target image block in the image is calculated to be 16×16, and the global filter parameter h0 is set to 0.5.

[0065] Step 3.4: construct a mixed loss function based on the partial losses described in steps 3.1 to 3.3. In this embodiment, the weight parameters λ1 and λ2 of the partial loss functions are 0.3 and 0.2 respectively.

[0066] Step 4: Import training data, select an optimizer and set relevant hyperparameters, and then perform model training. The network model training in this embodiment uses the AdamW optimizer to optimize the model parameters. The parameters of the AdamW optimizer are set to (0.9, 0.999), and the weight decay is 0.02. The training is performed for 1000 epochs, during which the cosine decay strategy is used to increase the learning rate from the initial value of 1×10 -5 Reduced to 1×10 -6 The network model was trained on an NVIDIA GeForce RTX 2080 GPU powered by an Intel i5-9400f CPU, using Python version 3.9.16 and TensorFlow-GPU version 2.6.0.

[0067] Step 5: Convergence judgment is performed on the training process of the projection data denoising network model. In this embodiment, if the loss function value is greater than 1×10 -2 , then adjust the learning rate and related hyperparameters and return to step 4. If the loss function value is less than or equal to 1×10 -2 , then stop training and go to step 6.

[0068] Step 6: Update the network model parameter weights and save the model when the loss function converges.

[0069] Step 7: Use the optimal network model to process the low-dose multi-energy CT projection data, import the noisy projection data into the optimal network model for calculation, and then use the filtered back-projection algorithm to obtain the multi-energy CT image after noise removal. In this embodiment, the multi-energy CT projection data used is 367×360 in size, in the bin binary file format, and in the 32-bit floating point number format.

[0070] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manner. Any other implementation manners derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.

Claims

1. A method for removing noise from projection data of multi-energy CT, characterized by: The following steps are involved: Step 1: construct a data set for multi-energy CT projection data denoising; Step 2: Construct a projection data denoising network model. The network model structure includes m shift and local enhancement Transformer window modules and n residual convolution modules; Step 3: Construct a hybrid loss function for supervised denoising network model training: The hybrid loss function consists of Huber loss, feature-aware loss, and non-local total variation loss. Step 4: Import training data, select the optimizer and set relevant hyperparameters, and then perform model training; Step 5: Convergence judgment of the training process of the projection data denoising network model: if the loss function does not converge, adjust the learning rate and modify the hyperparameter settings, and go to step 4; if the loss function converges, stop training, save the best model weights, and go to step 6; Step 6: Update the network model parameter weights and save the model when the loss function converges; Step 7: Use the optimal network model to process the multi-energy CT projection data, import the projection data containing noise into the optimal network model for calculation, and then use the filtered back projection algorithm to obtain the multi-energy CT image after noise removal.

2. The method for removing noise from projection data of multi-energy CT according to claim 1, characterized in that: The method for implementing the multi-energy CT projection data denoising data set in step 1 is: using phantom generation software to obtain a human body model containing X-ray energy attenuation information of various parts of the human body, and using a multi-energy CT detector simulation model to scan the human body model under conventional dose and low dose conditions respectively, to obtain multi-energy CT projection data under conventional dose and low dose.

3. The method for removing noise from projection data of multi-energy CT according to claim 1, characterized in that: The projection data denoising network model constructed in step 2 includes m shift and local enhancement Transformer window modules and n residual convolution modules.

4. The method for removing noise from projection data of multi-energy CT according to claim 3, characterized in that: The specific implementation method of step 2 includes the following steps: Step 2.1, construct the shifted windows Transformer Unit (STU): The STU part includes a shifted-window-based multi-head self-attention (SW-MSA) unit, a LayerNorm layer and a multi-layer perceptron (MLP) layer: the input feature vector is sent to the SW-MSA unit through the LayerNorm layer, and the feature vector processed by the SW-MSA unit is residually connected with the input; the SW-MSA unit divides the input features into different window partitions, and shifts the window partitions during the calculation process, and then moves the window position in the next layer to generate new window partitions. The self-attention calculation in the new window spans the window boundary in the previous layer to achieve connection with the window of the previous layer; the LayerNorm layer is used before each SW-MSA unit and MLP layer for normalization; the SW-MSA unit adopts a multi-head self-attention mechanism structure, which can extract multiple groups of features in parallel in several feature spaces. The self-attention calculation process of each feature space is as follows: Among them, Q τ , K τ and V τ are the query vector, key vector and value vector of the feature space τ respectively, and B τ is the relative position deviation; Step 2.2, build a locally enhanced window Transformer unit (LTU): This part introduces the convolution operator based on the STU structure, and replaces the original MLP unit with a locally enhanced feed-forward network (LeFF) unit built on a deep convolutional layer: the unit first uses a fully connected layer to increase the input data to a higher dimension, then performs deep convolution on the high-dimensional feature data, and finally uses a fully connected layer to restore it to the initial dimension and add it to the input data; Step 2.3, construct the shift and local enhancement Transformer window module: This module first extracts the input data features through a 1×1 convolution layer, then splits the data features through a split operation, and inputs the feature information into STU and LTU respectively: This operation converts the feature information X of the i-th (1<i≤m) level SLB i Split equally into two feature map groups X i1 and X i2 , and processed by STU and LTU respectively to obtain output features and The process is represented as follows: X i1 ,X i2 =Split(Conv(X i )); The output features of the two parts and After being merged through the connection layer and feature fused through the convolution layer, it is added with the input data of the SLB at this level and output to the next module. The final output Z of the i-th level SLB is i It is expressed as follows: Step 2.4: Build a residual convolution module: This module consists of a 3×3 convolution layer, a batch normalization layer, and a Leaky ReLU activation layer. Step 2.5: Build a denoising network model based on the shifted and locally enhanced Transformer window module and residual convolution module constructed in steps 2.3 and 2.

4.

5. The method for removing noise from projection data of multi-energy CT according to claim 1, characterized in that: The hybrid loss function constructed in step 3 is composed of Huber loss, feature-aware loss and non-local total variation loss.

6. The method for removing noise from projection data of multi-energy CT according to claim 5, characterized in that: The specific implementation method of step 3 includes the following steps: Step 3.1, construct the Huber loss function, which can be expressed as: Among them, P GT and P Predict They represent the label and denoising network model output data respectively, and ε is the regularization constant; Step 3.2, construct feature-aware loss function: Feature-aware loss uses a network model with pre-trained weights as a feature extractor, inputs the model output data and labels to be trained into the feature extraction model respectively, and extracts the output features of some hidden layers, describes the difference between the two and minimizes it to reduce the loss of detail features: First, construct a feature extraction model, and then use P GT and P Predict They are input into the feature extraction model respectively and the feature maps of the two sets of data in the feature extraction model are obtained respectively. Finally, the content loss and style loss are calculated to describe the difference between the feature maps of the two sets of data. The feature perception loss is obtained after weighted average. The calculation process is as follows: Among them, L is the total number of layers of feature maps extracted in the feature extraction model, h l 、w l and c l The output data and label are the lth layer feature maps in the feature extraction model respectively and The length, width and number of channels of , G(·) is the operator for calculating the Gram matrix of the extracted feature map data, and λ is the weight parameter after the normalization of the two parts of the loss; Step 3.3: Construct a non-local total variation loss function, which can be expressed as follows: L NLTV =Z NLTV (P Predict )+μZ NLTv (P Predict -P GT ), Among them, Z NLTV (·) is the operator for calculating the non-local total variation of a given image I, μ is the weight parameter, when calculating the non-local total variation of a given image I, assuming that the size of the divided image block is p×p, for the target image block in the image The middle pixel is located at the image position x, and the similarity weight is calculated to filter the local range r×r around x that is similar to U x Similar image patches U x with U y The similarity weight is expressed as follows: In the formula, g σ is a Gaussian convolution kernel with a standard deviation of σ, and h0 is the global filter parameter. The NLTV of image I is expressed as: In the formula, Ω represents the overall space of a given image, Ω x Represents the image domain space around pixel x; Step 3.4: Construct a mixed loss function based on the losses of each part in steps 3.1 to 3.3: The specific implementation method is to weighted average the losses of each part, which is specifically expressed as follows: Among them, λ1 and λ2 are weight parameters respectively.

7. The method for removing noise from projection data of multi-energy CT according to any one of claims 1 to 6, characterized in that: The specific implementation method of step 4 is: importing training data, selecting an optimizer and setting the momentum and weight decay size, selecting a learning rate adjustment strategy and setting the initial learning rate size, and using a mixed loss function to guide model training.

Images