Photon counting detector data correction method based on channel attention mechanism

Abstract

The invention discloses a photon counting detector data correction method based on a channel attention mechanism, relates to the field of CT detection data correction, and is used for solving the problem that an existing data correction method is low in precision due to the complexity of crosstalk distribution of a photon counting detector. The method comprises the following steps: constructing a CT scanning simulation system; preparing a data set; constructing a distortion data correction network model; importing training data, and carrying out network model training to obtain an optimal network model weight; and performing data correction by using the optimal network model. According to the method, the advantages of the channel attention mechanism in the aspect of improving the network feature extraction capability are fully utilized, energy channel correlation information in the photon counting detector data and crosstalk models distributed in the energy boxes can be extracted, crosstalk and noise in the data are effectively suppressed, and higher correction precision is achieved.

Description

Technical Field

[0001] The present invention relates to the field of CT detection data correction, and in particular to a photon counting detector data correction method based on a channel attention mechanism. Background Art

[0002] As an important development direction in the field of CT in the future, spectral CT technology based on photon counting detector (PCD) has significant advantages in improving spatial resolution, material property imaging, and low-dose imaging. However, before PCD is widely used in commercial CT systems, there are still some problems that need to be solved. PCDs made of materials such as cadmium telluride (CdTe), cadmium zinc telluride (CdZnTe, CZT), and silicon (Si) will be affected by physical effects during the detection process, resulting in distortion of detection data. Physical effects such as charge sharing, Compton scattering, and K-fluorescence will cause crosstalk to the detector array, causing some high-energy photons to be mistakenly counted as more low-energy photons in adjacent pixels, seriously affecting the subsequent reconstruction quality. Therefore, it is of great significance to perform data correction on PCD.

[0003] Many PCD data correction methods have been proposed. One of them is to compensate for the distorted signal in the readout circuit, but this method is prone to other negative effects and is relatively expensive. Later, it was proposed to correct PCD distorted data at the algorithm level, but this method relies on an accurate PCD distortion model, and its development is seriously hindered.

[0004] In recent years, deep learning, as an efficient data processing technology, has been widely used in data correction, image processing and other fields, and has achieved remarkable results. Deep learning is an end-to-end method that does not require an accurate detector model. It only needs paired data to automatically learn the hidden relationship between distorted data and clean data, thereby realizing data correction. At present, some correction methods based on fully connected networks or convolutional neural networks have been proposed. However, the fully connected network has a simple structure and is too sensitive to spatial position, which is not conducive to the extraction of crosstalk noise features; and the crosstalk of detector data usually occurs between adjacent pixels, resulting in poor performance of the fully connected network with all pixels as the receptive field when processing PCD data. Convolutional neural networks with convolution operations as the core can better extract crosstalk features distributed in adjacent pixels. However, crosstalk occurs not only between adjacent pixels, but also between different energy boxes. Convolutional neural networks often ignore the complex distribution law of crosstalk and the data correlation between energy channels, and some detailed features will be lost during the correction process. The channel attention mechanism adopted by the present invention can well extract channel correlation and adjust channel weights, which has certain advantages in utilizing the correlation of data in each energy box of PCD and extracting crosstalk noise features. Therefore, the present invention applies channel attention to the PCD data correction network to improve the correction accuracy. Summary of the invention

[0005] In order to overcome the defects in the above-mentioned prior art, the present invention provides a PCD data correction method based on an efficient channel attention mechanism. This method makes full use of the advantages of the channel attention mechanism in improving the network feature extraction capability, and can effectively extract the energy channel correlation information in the PCD data and the crosstalk model distributed in each energy box, thereby more effectively suppressing the crosstalk and noise in the data and improving the final correction accuracy.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] A method for correcting photon counting detector data based on a channel attention mechanism comprises the following steps:

[0008] Step 1. Construction of a PCD-based spectral CT scanning simulation system: Establish a spectral CT scanning simulation system consisting of an X-ray source, a human digital phantom and a PCD model.

[0009] Step 2: Dataset preparation: Run the spectral CT scanning simulation system established in step 1 to obtain PCD detection data, and randomly divide the detection data into training set, validation set and test set according to a certain ratio.

[0010] Step 3: Construct a PCD distorted data correction network model: The model consists of an input module, an efficient channel attention encoder, an efficient channel attention decoder, and an output module.

[0011] Step 4: Construct a composite loss function: select L1 loss for regression loss and multi-scale structural similarity index loss for perceptual loss.

[0012] Step 5: Preprocess the data set.

[0013] Step 6: Import training data, select the optimizer and set relevant hyperparameters to train the network model.

[0014] Step 7: Conduct convergence judgment on the training process of the data correction network model: if the loss function does not converge, adjust the learning rate and modify the hyperparameter settings, and go to step 6; if the loss function converges, stop training, save the best model weights, and go to step 8.

[0015] Step 8: Use the best network model weight to correct the PCD data: Import the PCD distortion data into the best network model saved in step 7 for calculation to obtain the corrected PCD data, which effectively suppresses crosstalk and noise.

[0016] Furthermore, the specific implementation method of step 2 is: using the PCD model to calculate the PCD detection data containing crosstalk and noise and the corresponding ideal data, and obtain a paired data set consisting of input data and label data. The data set is divided into a training set, a validation set, and a test set in a ratio of 4:1:1.

[0017] Furthermore, the specific implementation method of step 3 includes the following steps:

[0018] Step 3.1. The input module of the multi-channel data correction network performs shallow feature extraction on the input distorted data: The first part of the input module consists of two 3×3 convolutional layers. The input data is input into these two convolutional layers to generate shallow features for each energy bin data. The second part is the SE channel attention module. In the Squeeze part of the SE module, the shallow features of the energy bin data are passed through a global average pooling layer to compress the global spatial information of each feature channel into a channel descriptor. In the Excitation part, the SE block uses two fully connected layers and a ReLU activation layer to explicitly model the correlation between channels and generate appropriate weights. The number of channels is reduced in the first fully connected layer with a reduction rate of 16, and the features are re-expanded to the original size in the second fully connected layer. The weights are then normalized using the Sigmoid function. Finally, the output of the input module is the shallow feature map of the PCD data that is refined according to the generated weight map.

[0019] Step 3.2, the efficient channel attention encoder part further extracts features from the shallow feature map to obtain more abstract deep features: the shallow feature map output by the input module in step 3.1 is used as the input of the encoder part. The encoder part consists of 4 efficient channel attention (ECA) encoder modules. The encoder module consists of a maximum pooling downsampling layer, two 3×3 convolution blocks and an ECA block. The convolution block consists of a 3×3 standard convolution, a batch normalization layer and a ReLU activation function. The input features are encoded after passing through the maximum pooling downsampling layer and two 3×3 convolution blocks, and then input into the ECA block for channel refinement. The ECA block uses a global average pooling layer to obtain the global spatial features of the input features, and then uses a one-dimensional convolution layer to efficiently learn local cross-channel correlations and generate channel weights. The weights are normalized using the sigmoid activation function, and the normalized weights are used to refine the input features of the ECA block again. The obtained channel refinement features are used as the output of each ECA encoder module. The shallow feature map passes through the 4 ECA encoder modules in the encoder part to generate the deep feature map.

[0020] Step 3.3, the efficient channel attention decoder part decodes the deep feature map: the deep feature map output by the encoder part in step 2.2 is used as the input of the ECA decoder part. The decoder part consists of 4 ECA decoder modules. Each decoder module consists of a 3×3 deconvolution block, a 4×4 deconvolution block, and an ECA block. The 3×3 deconvolution block consists of a standard 3×3 deconvolution layer, a batch normalization layer, and a ReLU activation function, while the 4×4 deconvolution block consists of a 4×4 deconvolution layer with a stride of 2, a batch normalization layer, and a ReLU activation function. The input features are decoded and upsampled through these two deconvolution blocks to obtain features with a lower level of abstraction. The features after the deconvolution block are input into the ECA block to obtain features after channel refinement. The deep feature map is restored to a shallow feature map corresponding to the output of the input module through the ECA decoder part.

[0021] Step 3.4: The output module restores the shallow feature map to PCD data: The output of the ECA decoder in step 3.3 is used as the input of the output module. The output module consists of two 3×3 convolutional layers, an ECA module, and a 1×1 convolutional layer. The shallow feature map is further restored through two 3×3 convolutional layers, further refined on the channel through the ECA module, and then adjusted on the number of channels through the 1×1 convolutional layer, so as to obtain output data of the same size as the input data.

[0022] Step 3.5: Add the output data of the output module and the input distorted data to form a global residual structure. The summation result is used as the output of the entire network.

[0023] Furthermore, the composite loss function of step 4 is composed of L1 regression loss and perceptual loss based on structural similarity index, and the two losses are added according to a certain weight to obtain the composite loss function.

[0024] Furthermore, the specific implementation method of step 6 is: importing training set data and label data without crosstalk and noise, 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 composite loss function to guide model training.

[0025] The beneficial effects of the present invention are as follows:

[0026] The photon counting detector distorted data correction method provided by the present invention adopts a crosstalk feature extraction network with an efficient channel attention module as the core to maximize the extraction of crosstalk and noise distribution in the data, thereby enhancing the key feature information required for the correction task; a convolution layer with a convolution kernel size of 3 is used to better extract the PCD data crosstalk distribution characteristics occurring between adjacent pixels; a global residual connection is introduced to better extract the data distortion model; by using a composite loss function, not only the matching of each pixel value of the corrected data with the label data is effectively achieved, but also the corresponding relationship in the structural features is maintained; and finally, the crosstalk and noise in the data are effectively suppressed. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] 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.

[0028] Figure 1 A flow chart of a photon counting detector data correction method based on an efficient channel attention mechanism provided by an embodiment of the present invention;

[0029] Figure 2 An overall structural diagram of a PCD distortion data correction network model based on an efficient channel attention mechanism provided by an embodiment of the present invention;

[0030] Figure 3 A structural diagram of a SE module provided in an embodiment of the present invention;

[0031] Figure 4 The ECA module structure diagram provided by the embodiment of the present invention; DETAILED DESCRIPTION

[0032] 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 making creative work belong to the scope of protection of the present invention.

[0033] At present, the distortion problem of the detection data of the photon counting detector (PCD) has seriously affected the reconstruction accuracy of spectral CT. The method based on improving the readout circuit is costly and has negative effects such as reduced resolution; the method of correction at the algorithm level is heavily dependent on the accuracy of the PCD physical effect simulation model, and the correction accuracy is limited; the method based on the fully connected network has limited performance, and the existing methods based on convolutional neural networks often ignore the complex distribution of crosstalk and the data correlation between energy channels, and often lose some detailed features during the correction process. There is still room for improvement in the correction accuracy.

[0034] In order to solve the above problems, an embodiment of the present invention provides a PCD data correction method based on an efficient channel attention mechanism.

[0035] Figure 1 A flow chart of a photon counting detector data correction method provided by an embodiment of the present invention is shown, and the method comprises the following steps:

[0036] Step 1. Construction of a PCD-based spectral CT scanning simulation system: Establish a spectral CT scanning simulation system consisting of an X-ray source, a human digital phantom and a PCD model.

[0037] In this embodiment, a PCD-based spectral CT scanning simulation system is built to obtain PCD data required for network training and testing. The photon counting toolkit (PcTK) is used as the PCD simulation model to generate PCD data with spectral distortion and multiple counting phenomena. The human head phantom is generated using the XCAT phantom generation software. The phantom size is 256×256×150. The first two items are the size of a single slice, and the last item is the number of slices. Due to the need for PcTK input data, the XCAT head phantom is materially decomposed to generate soft tissue and bone maps. The specific process is: first obtain two high and low energy E high and E low The phantom attenuation coefficient μ(E high , r) and μ(E low , r), where r is the voxel position of the phantom. Then the ICRU Report 44 database is queried to obtain the attenuation coefficients μ of soft tissue and bone at these two energies.soft (E high ), μ bone (E high ), μ soft (E low ), μ bone (E low ). Each voxel in the phantom is regarded as a linear combination of soft tissue and bone with a certain weight, and the attenuation coefficient of each voxel in the phantom is regarded as the attenuation coefficient of soft tissue and the attenuation coefficient of bone with the same weight w. soft (r), w bone (r), thus obtaining a system of equations:

[0038]

[0039] Solving this set of equations can obtain the weight of soft tissue and bone in each voxel in the head phantom. Forward projecting the weighted data of these two substances can obtain their line integrals, that is, the projection data of soft tissue and bone from 360 cone beam projections. After obtaining the projection data of soft tissue and bone, it is necessary to configure PcTK related parameters to prepare to generate PCD data.

[0040] The embodiment of the present invention uses a GE_Maxiray_125 tube with a tube voltage of 120kVp to generate the X-ray energy spectrum required for the simulation. The size of the electron cloud generated by the X-ray photons is set to 24μm, and the electron noise parameter is set to 2.0keV. The detector material is CdTe, the material density is 5.85g / cm3, the detector thickness is 1.6mm, and the pixel size is 450μm. The detector array used is 832×832. The energy thresholds of PCD are set to 20, 35, 50, 65, and 80keV, respectively, so that data of 5 energy boxes with energy ranges of 20-35keV, 35-50keV, 50-65keV, 65-80keV, and 80-120keV are obtained. The number of photons per pixel at each angle is set to 5×10 6 PcTK also requires the linear attenuation coefficient (LAC) of soft tissue and bone at different energies, which can still be obtained by querying the ICRUReport 44 database.

[0041] Based on the above data and related parameters, PcTK uses a random number generator to associate the projection data of each angle with the PCD crosstalk and energy box counts, and generates noisy spectral CT distorted projection data, that is, the contaminated photon count results of each energy box. The crosstalk-free data output by PcTK pure mode is used as label data. 10 human head phantoms of different genders, ages, and body shapes are used for network training and testing.

[0042] Step 2: Dataset preparation.

[0043] In this embodiment, the PCD-based spectral CT scanning simulation system established in step 1 is run to obtain PCD detection data. The PCD data size is 832×832, which is too large and can easily cause memory explosion problems. The detected data is divided into overlapping three-dimensional data blocks. The block size is set to 64×64×5, and the step size is 48. Among the generated PCD data blocks, 120,000 pairs of data blocks from 8 patient phantoms are selected for network training, and 30,000 pairs of blocks are used as verification data sets to monitor the fitting of the correction network during training. In addition, 30,000 pairs of data blocks from two other patient head phantoms are used to test the performance of the network. These data blocks are converted into LAC data for correction. No information is lost in this conversion process.

[0044] Step 3. Construct a PCD distorted data correction network model. The model consists of an input module, an efficient channel attention encoder, an efficient channel attention decoder, and an output module. The input module performs shallow feature extraction on the input PCD distorted data and extracts the channel correlation of feature channels and crosstalk noise. The efficient channel attention encoder part further extracts features from the shallow feature map to obtain more abstract deep features, extracts the correlation of feature channels, assigns appropriate weights to each feature channel, and highlights the channels that are important for the data correction task. The efficient channel attention decoder part allocates weights to feature channels while decoding the deep feature map, and further refines the features in the channel dimension. The output module restores the shallow feature map to PCD data. Finally, the output data of the output module is added to the input distorted data to form a global residual structure. The summed result is used as the output of the entire network.

[0045] Step 3.1. The input module of the multi-channel data correction network performs shallow feature extraction on the input distorted data: The first part of the input module consists of two 3×3 convolution blocks. The input data is input into these two convolution blocks to generate shallow features for each energy bin data. The second part is the SE channel attention module. In the Squeeze part of the SE module, the shallow features of the energy bin data are passed through a global average pooling layer to compress the global spatial information of each feature channel into a channel descriptor. In the Excitation part, the SE block uses two fully connected layers and a ReLU activation layer to explicitly model the correlation between channels and generate appropriate weights. The number of channels is compressed in the first fully connected layer with a compression rate of 16, and the features are re-expanded to the original size in the second fully connected layer. The weights are then normalized using the Sigmoid function. Finally, the output of the input module is a shallow feature map of the PCD data that is refined based on the generated weight map.

[0046] In this step, the 3×3 convolution block consists of a 3×3 convolution layer, a ReLU activation layer, and a batch normalization layer. The input data passes through two 3×3 convolution blocks to obtain the shallow features of each energy box data. The SE module consists of two parts: Squeeze and Excitation. The Squeeze part contains a global average pooling layer, which compresses the global spatial information of each feature channel into a channel descriptor. This operation performs feature compression in the spatial dimension, converting each two-dimensional feature channel into a real number with a global receptive field, characterizing the global distribution of responses on the feature channel, and the output dimension matches the number of input feature channels. The calculation formula for global average pooling is:

[0047]

[0048] The Excitation part is implemented by two fully connected layers. The first fully connected layer compresses C feature channels into C / r channels to reduce the amount of calculation, and then passes through a ReLU layer to keep the dimension unchanged; the second fully connected layer restores the feature map back to C channels, where r refers to the compression ratio. In this step, r = 16. The weights are then normalized using the Sigmoid function. The calculation formula for Excitation is:

[0049] s=F ex (z, W)=σ(g(z, W))=σ(W2δ(W1z))

[0050] Step 3.2, the efficient channel attention encoder part further extracts features from the shallow feature map to obtain more abstract deep features: the shallow feature map output by the input module in step 3.1 is used as the input of the encoder part. The encoder part consists of 4 efficient channel attention (ECA) encoder modules. The encoder module consists of a maximum pooling downsampling layer, two 3×3 convolution blocks and an ECA block. The convolution block consists of a 3×3 standard convolution, a batch normalization layer and a ReLU activation function. The input features are encoded after the maximum pooling downsampling layer and two 3×3 convolution blocks, and then input into the ECA block for channel refinement. The ECA block uses a global average pooling layer to obtain the global spatial features of the input features, and then uses a one-dimensional convolution layer to efficiently learn local cross-channel correlations and generate channel weights. The weights are normalized using the sigmoid activation function, and the normalized weights are used to refine the input features of the ECA block again. The obtained channel refinement features are used as the output of each ECA encoder module. The shallow feature map passes through the 4 ECA encoder modules in the encoder part to generate the deep feature map.

[0051] In this step, the ECA block consists of a global average pooling layer, a one-dimensional convolutional layer, and a sigmoid activation layer. Similar to the SE block, the global average pooling layer is still used to compress the global spatial information of each feature channel into a channel descriptor. The fully connected layer previously used in the SE module is replaced by a one-dimensional convolutional layer with a convolution kernel size of k, where k is adaptively determined by the number of channels C:

[0052]

[0053] |t| odd Represents the odd number closest to t. The one-dimensional convolutional layer can effectively learn local cross-channel correlations and generate channel weights. The calculation formula is:

[0054] ω=σ(ClD k (y))

[0055] Then the Sigmoid activation function is still used to normalize the weights, and the normalized weights are used to recalibrate the input features of the ECA module as the output of the ECA block.

[0056] Step 3.3, the efficient channel attention decoder part decodes the deep feature map: the deep feature map output by the encoder part in step 3.2 is used as the input of the ECA decoder part. The decoder part consists of 4 ECA decoder modules. Each decoder module consists of a 3×3 deconvolution block, a 4×4 deconvolution block, and an ECA block. The 3×3 deconvolution block consists of a standard 3×3 deconvolution layer, a batch normalization layer, and a ReLU activation function, while the 4×4 deconvolution block consists of a 4×4 deconvolution layer with a stride of 2, a batch normalization layer, and a ReLU activation function. The input features are decoded and upsampled through these two deconvolution blocks to obtain features with a lower level of abstraction. The features after the deconvolution block are input into the ECA block to obtain features after channel refinement. The deep feature map is restored to a shallow feature map corresponding to the output of the input module through the ECA decoder part.

[0057] Step 3.4: The output module restores the shallow feature map to PCD data: The output of the ECA decoder in step 3.3 is used as the input of the output module. The output module consists of two 3×3 convolution blocks, an ECA block, and a 1×1 convolution layer. The shallow feature map is further restored through two 3×3 convolution blocks, further refined on the channel through the ECA module, and then adjusted on the number of channels through the 1×1 convolution layer, so as to obtain output data of the same size as the input data.

[0058] Step 3.5: Add the output data of the output module and the input distorted data to form a global residual structure. The summation result is used as the output of the entire network.

[0059] In this step, a global residual connection is performed on the network, which is equivalent to making the network extract the difference between the data containing crosstalk and noise and the clean data, so that the network explicitly extracts the PCD crosstalk noise model.

[0060] Step 4: Construct a composite loss function: select L1 loss for regression loss and multi-scale structural similarity index loss for perceptual loss.

[0061] In this step, the calculation formula of the L1 loss function is:

[0062]

[0063] The multi-scale structural similarity index loss comes from the image quality evaluation index structural similarity index (SSIM). The SSIM expression is as follows:

[0064]

[0065] Among them, x and y represent the output image and label image respectively, μ x and μ y is the mean of the two pixel values, σ x , σ y are the standard deviations of the pixel values ​​of the two images, σ xy is the covariance of the two pixel values, C1 and C2 are constant terms, which can avoid the denominator being 0.

[0066] Multi-scale SSIM obtains image information at multiple scales through multiple downsampling operations, calculates their structural similarity and multiplies them. M represents the number of scales, and the expression is as follows:

[0067]

[0068] The loss function based on MS-SSIM is:

[0069] L MS-SSIM = 1-MS-SSIM(x, y)

[0070] The final composite loss function used is:

[0071] L hybrid =αL1+(1-α)L MS-SSIM

[0072] Among them, α is a parameter that balances the weights of the two, and is adjusted according to the scale of the two and the training results. The goal of the network is to make the composite loss function L hybrid Minimize. In this step, α is 0.15.

[0073] Step 5: Preprocess the data set.

[0074] In this step, the PCD data is normalized to the maximum and minimum values ​​to stabilize the data value range.

[0075] Step 6: Import training data, select the optimizer and set relevant hyperparameters to train the network model.

[0076] In this step, the Adam optimizer was used for network training, with exponential decay parameters α1 = 0.9, α.2 = 0.999. The initial learning rate was set to 1×10 -4 , and gradually decreases as the training process progresses. The minimum batch used in each iteration is 32. A composite loss function is used to guide model training.

[0077] Step 7: Conduct convergence judgment on the training process of the data correction network model: if the loss function does not converge, adjust the learning rate and modify the hyperparameter settings, and go to step 6; if the loss function converges, stop training, save the best model weights, and go to step 8.

[0078] In this step, the model convergence is judged. If the loss function value is greater than 10-3 after training for 500 epochs, the learning rate and related hyperparameters are adjusted, and the process returns to step 6. If the loss function value is less than or equal to 10-3, the training is stopped and the optimal model weights are saved, and the process goes to step 8.

[0079] Step 8: Use the best network model weight to correct the PCD data: import the PCD distortion data into the best network model saved in step 7 for calculation, and obtain the corrected PCD data with suppressed crosstalk and noise.

[0080] In this step, the distorted PCD data is adjusted to the LAC form and divided into three-dimensional data blocks of size 64×64×5, which are then imported into the optimal network model for calculation and the corresponding clean PCD data is output. Finally, it is restored to the original size of 832×832×5 PCD data, which can provide high-quality original detection data for spectral CT reconstruction.

[0081] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for correcting distorted photon counting detector (PCD) data based on a channel attention mechanism, characterized in that: The following steps are involved: Step 1: Establishment of a PCD-based spectral CT scanning simulation system: Establish a spectral CT scanning simulation system consisting of an X-ray source, a human digital phantom, and a PCD model; Step 2: Dataset preparation: Run the spectral CT scanning simulation system established in step 1 to obtain PCD detection data, and randomly divide the detection data into training set, validation set and test set according to a certain ratio; Step 3: Construct a PCD distorted data correction network model: The model consists of an input module, an efficient channel attention encoder, an efficient channel attention decoder, and an output module; Step 4: Construct a composite loss function: select L1 loss for regression loss and multi-scale structural similarity index loss for perceptual loss; Step 5: Preprocess the data set; Step 6: Import training data, select optimizer and set relevant hyperparameters to train the network model; Step 7: Convergence judgment of the training process of the data correction network model: if the loss function does not converge, adjust the learning rate and modify the hyperparameter settings, and go to step 6; if the loss function converges, stop training, save the best model weights, and go to step 8; Step 8: Use the best network model weight to correct the PCD data: Import the PCD distortion data into the best network model saved in step 7 for calculation to obtain the corrected PCD data, which effectively suppresses crosstalk and noise.

2. The distorted PCD data correction method based on the channel attention mechanism according to claim 1 is characterized in that: The specific implementation method of step 2 is: use the PCD model to calculate the PCD detection data containing crosstalk and noise and the corresponding ideal data, and obtain a paired data set consisting of input data and label data. Divide the data set into a training set, a validation set, and a test set in a ratio of 4:1:

1.

3. The distorted PCD data correction method based on the channel attention mechanism according to claim 1 is characterized in that: The training model uses the SE channel attention module and the ECA efficient channel attention module, and the network loss function is a composite loss function.

4. The distorted PCD data correction method based on the channel attention mechanism according to claim 3 is characterized in that: The step 3 comprises the following specific steps: Step 3.

1. The input module of the multi-channel data correction network performs shallow feature extraction on the input distorted data: The first part of the input module consists of two 3×3 convolutional layers. The input data is input into these two convolutional layers to generate shallow features for each energy box data. The second part is the SE channel attention module. In the Squeeze part of the SE module, the shallow features of the energy box data pass through the global average pooling layer to compress the global spatial information of each feature channel into a channel descriptor. In the Excitation part, the SE block uses two fully connected layers and a ReLU activation layer to explicitly model the correlation between channels and generate appropriate weights. The number of channels is reduced in the first fully connected layer with a reduction rate of 16, and the features are re-expanded to the original size in the second fully connected layer. The weights are then normalized using the Sigmoid function. Finally, the output of the input module is a shallow feature map of the PCD data that is refined based on the generated weight map; Step 3.2, the efficient channel attention encoder part further extracts features from the shallow feature map to obtain more abstract deep features: the shallow feature map output by the input module in step 3.1 is used as the input of the encoder part. The encoder part consists of 4 efficient channel attention (ECA) encoder modules. The encoder module consists of a maximum pooling downsampling layer, two 3×3 convolution blocks and an ECA block. The convolution block consists of a 3×3 standard convolution, a batch normalization layer and a ReLU activation function. The input features are encoded after passing through the maximum pooling downsampling layer and two 3×3 convolution blocks, and then input into the ECA block for channel refinement. The ECA block uses a global average pooling layer to obtain the global spatial features of the input features, and then uses a one-dimensional convolution layer to efficiently learn local cross-channel correlations and generate channel weights. The weights are normalized using the sigmoid activation function, and the normalized weights are used to refine the input features of the ECA block again. The obtained channel refinement features are used as the output of each ECA encoder module. The shallow feature map is passed through the four ECA encoder modules in the encoder part to generate the deep feature map; Step 3.3, the efficient channel attention decoder part decodes the deep feature map: the deep feature map output by the encoder part in step 2.2 is used as the input of the ECA decoder part. The decoder part consists of 4 ECA decoder modules. Each decoder module consists of a 3×3 deconvolution block, a 4×4 deconvolution block and an ECA block. The 3×3 deconvolution block consists of a standard 3×3 deconvolution layer, a batch normalization layer and a ReLU activation function, while the 4×4 deconvolution block consists of a 4×4 deconvolution layer with a stride of 2, a batch normalization layer and a ReLU activation function. The input features are decoded and upsampled through these two deconvolution blocks to obtain features with a lower level of abstraction. The features passed through the deconvolution block are input into the ECA block to obtain features after channel refinement. The deep feature map is restored to a shallow feature map corresponding to the output of the input module through the ECA decoder part; Step 3.4, the output module restores the shallow feature map to PCD data: the output of the ECA decoder in step 3.3 is used as the input of the output module. The output module consists of two 3×3 convolutional layers, an ECA module, and a 1×1 convolutional layer. The shallow feature map is further restored through two 3×3 convolutional layers, further refined on the channel through the ECA module, and then adjusted on the number of channels through the 1×1 convolutional layer, so as to obtain output data of the same size as the input data; Step 3.5: Add the output data of the output module and the input distorted data to form a global residual structure. The summation result is used as the output of the entire network.

5. The distorted PCD data correction method based on the channel attention mechanism according to claim 3 is characterized in that: The composite loss function of step 4 is composed of L1 regression loss and perceptual loss based on structural similarity index. The two losses are added according to a certain weight to obtain the composite loss function.

6. The distorted PCD data correction method based on the channel attention mechanism according to any one of claims 1 to 5, characterized in that: The specific implementation method of step 6 is: importing training set data and label data without crosstalk and noise, 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 composite loss function to guide model training.

Images