A generative adversarial network processing system for radiology ct image reconstruction

Abstract

The present application relates to the technical field of image processing, and more particularly to a generative adversarial network processing system for radiology CT image reconstruction, which comprises a preprocessing module, a double-path generator module, an attention discriminator module, a consistency constraint module and an adversarial training and optimization module. The present application realizes high-quality and high-fidelity reconstruction of CT images from low-quality projection data, and improves the clinical reliability of the reconstruction results.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a generative adversarial network processing system for CT image reconstruction in radiology. Background Technology

[0002] Medical imaging technology is a core support for modern medical diagnosis and treatment, and the performance of its image reconstruction algorithms directly determines the quality, resolution, and computational efficiency of the final image. Deep learning-based image reconstruction methods have become a cutting-edge technology for improving CT image quality. This method aims to utilize the feature learning capabilities of neural networks to reconstruct CT images from low-dose or undersampled projection data, ensuring the image quality required for diagnosis.

[0003] Existing generative adversarial networks are widely used in such image reconstruction tasks. They optimize the reconstruction results through adversarial training between the generator and the discriminator. However, the design of existing generator network structures is insufficient in recovering detailed features, which can easily lead to blurred textures or artifacts in the reconstructed images. The discriminator's discrimination ability is also limited, making it difficult to accurately distinguish the subtle differences between real high-quality images and generated images in complex pathological structures, resulting in instability in the adversarial training process. Summary of the Invention

[0004] The purpose of this invention is to provide a generative adversarial network processing system for CT image reconstruction in radiology, in order to solve the technical problems of insufficient recovery of detailed features by existing generators and limited discrimination ability of discriminators, which leads to a decline in image quality.

[0005] This invention provides a generative adversarial network processing system for CT image reconstruction in radiology, including a preprocessing module, and further comprising: a dual-path generator module, an attention discriminator module, a consistency constraint module, and an adversarial training and optimization module; wherein,

[0006] The dual-path generator module is used to reconstruct high-quality CT images from low-quality projection data to be reconstructed; the dual-path generator module is an encoder-decoder architecture, and integrates local detail paths and global structure paths in the middle of the encoder; the encoder includes a convolutional downsampling block, the local detail path includes a residual dense block, the global structure path includes a non-local attention module, and the decoder includes a transposed convolutional upsampling block.

[0007] The attention discriminator module is used to discriminate the input image. The attention discriminator module has a pyramid structure and includes parallel discriminator subnetworks.

[0008] The consistency constraint module is used to force the reconstructed image and the original projection data to meet the physical model constraints of CT imaging during the training process. The consistency constraint module includes a differentiable forward projection operator, which is used to simulate the reconstruction result in the image domain back to the projection domain to generate simulated projection data. The consistency constraint module calculates the L2 norm loss of the simulated projection data and the low-quality projection data to be reconstructed in the sinusoidal domain, and calculates the structural similarity index loss between the reconstructed image and the corresponding real high-quality CT image in the image domain. The two losses are weighted and summed to generate a consistency loss value.

[0009] The adversarial training and optimization module is used to receive adversarial loss and consistency loss, and to iteratively update the network parameters of the dual-path generator module and the attention discriminator module using an alternating optimization strategy.

[0010] In some embodiments, the data preprocessing module's processing includes at least: performing logarithmic transformation and normalization on the original projection data to map the data value range to between 0 and 1; performing data augmentation operations on the normalized projection data in the sinusoidal domain, the data augmentation operations including adding Gaussian noise within a preset noise level range to simulate low-dose conditions, and randomly discarding a portion of the projection view within a preset angle range to simulate undersampling conditions.

[0011] In some embodiments, the residual dense block includes multiple convolutional layers, the output of each convolutional layer is concatenated with the input of all subsequent convolutional layers in the channel dimension, and the output of the first convolutional layer in the residual dense block is directly passed to the 1x1 convolutional fusion layer at the end of the residual dense block through a skip connection.

[0012] In some embodiments, the nonlocal attention module is used to calculate the association weight between any two positional feature vectors. The calculation process includes at least: performing a 1x1 convolution to reduce the dimensionality of the input feature map to obtain a query feature map, a key feature map, and a value feature map; performing matrix multiplication between the query feature map and the key feature map, and normalizing the result using the Softmax function to obtain an attention weight map; multiplying the attention weight map with the value feature map, performing a 1x1 convolution to increase the dimensionality of the result, and then adding it to the original input feature map.

[0013] In some embodiments, the adversarial loss is a Wasserstein distance loss with gradient penalty; the total loss function of the adversarial training and optimization module is a weighted sum of the adversarial loss, the consistency loss, and the image domain pixel-level norm loss.

[0014] In some embodiments, the alternating optimization strategy is as follows: first, the generator parameters are fixed and the discriminator parameters are updated once, and then the discriminator parameters are fixed and the generator parameters are updated once. The network parameters are iteratively optimized using an adaptive moment estimation algorithm, the initial learning rate is set to 0.0001, and a cosine annealing strategy is used for decay.

[0015] In some embodiments, the input of the dual-path generator module is connected to the output of the preprocessing module.

[0016] In some embodiments, the input of the attention discriminator module is connected to the output of the dual-path generator module.

[0017] In some embodiments, the input of the consistency constraint module is connected to the output of the preprocessing module.

[0018] In some embodiments, the input of the adversarial training and optimization module is connected to the output of the attention discriminator module and the output of the consistency constraint module, respectively.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] 1. By constructing a system that includes a dual-path generator, an attention discriminator, and a consistency constraint module, the problem of poor detail recovery in existing GAN reconstruction methods is addressed;

[0021] 2. The dual-path generator achieves simultaneous modeling of image texture and anatomical structure through parallel local detail paths and global structure paths, ensuring clear reconstruction of key details such as minute lesions;

[0022] 3. The attention discriminator utilizes a pyramid structure to focus on subtle differences in pathological areas at different resolutions, enhancing the sensitivity to identify defects in the generated images;

[0023] 4. The consistency constraint module introduces a differentiable forward projection operator, which incorporates the imaging physical model into the training target, forcing the reconstructed image to be consistent with the real data in both the projection domain and the image domain. This effectively suppresses artifact generation and significantly improves the clinical credibility of the reconstruction results.

[0024] 5. Through the adversarial training and optimization module, a stable balance between adversarial learning and physical priors is achieved, ultimately producing high-quality CT images that meet the requirements while significantly reducing radiation dose or shortening scan time. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is an architecture diagram of the generative adversarial network processing system of the present invention;

[0027] Figure 2 This is a schematic diagram illustrating the core principle of the dual-path generator of this invention. Detailed Implementation

[0028] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0029] Partial interpretation:

[0030] 1. L1 loss: This is the most commonly used loss function in regression tasks. Its core function is to measure the error between the model's predicted value and the true label value.

[0031] 2. Consistency loss: This is a general term for a type of loss that aims to constrain the model output / features to meet consistency criteria. Its core function is to assist the main loss, allowing the model to learn more robust and essential rules, rather than overfitting the surface features of the data.

[0032] 3. Cosine annealing strategy: This is a core learning rate scheduling method in deep learning / machine learning. Its core is to make the learning rate change periodically according to a cosine function. It has the advantages of slow decay to ensure convergence and periodic restart to improve generalization. It solves the problems of conventional learning rate decay being prone to getting into local optima and slow convergence speed.

[0033] 4. GAN Reconstruction Method: This method is based on Generative Adversarial Networks (GANs) to achieve input-target mapping reconstruction. Through adversarial training between the generator and the discriminator, the generator learns the essential distribution of the data, achieving accurate reconstruction from incomplete / noisy / low-quality input to complete / clean / high-quality target. Unlike traditional interpolation, filtering, and encoding / decoding reconstruction, it has the advantages of high reconstruction fidelity, strong detail restoration, and adaptability to complex distributions.

[0034] Example 1

[0035] This embodiment includes a generative adversarial network processing system for CT image reconstruction in radiology, comprising a preprocessing module, and further comprising: a dual-path generator module, an attention discriminator module, a consistency constraint module, and an adversarial training and optimization module; wherein, the dual-path generator module is used to reconstruct high-quality CT images from low-quality projection data to be reconstructed; the dual-path generator module is an encoder-decoder architecture, and integrates local detail paths and global structure paths in the middle of the encoder; the encoder includes convolutional downsampling blocks, the local detail path includes residual dense blocks, the global structure path includes non-local attention modules, and the decoder includes transposed convolutional upsampling blocks; the attention discriminator module is used to discriminate the input image, and the attention discriminator module is a pyramid structure, including parallel discrimination... The system consists of a sub-network and a consistency constraint module. During training, the consistency constraint module forces the reconstructed image and the original projection data to satisfy the physical model constraints of CT imaging. This module includes a differentiable forward projection operator to simulate the reconstruction result from the image domain back to the projection domain, generating simulated projection data. The consistency constraint module calculates the L2 loss between the simulated projection data and the low-quality projection data to be reconstructed in the sinusoidal domain, and calculates the structural similarity index loss between the reconstructed image and the corresponding real high-quality CT image in the image domain. The two losses are weighted and summed to generate the consistency loss value. An adversarial training and optimization module receives the adversarial loss and the consistency loss, and uses an alternating optimization strategy to iteratively update the network parameters of the dual-path generator module and the attention discriminator module.

[0036] To better understand this invention, the details are as follows:

[0037] first,

[0038] The preprocessing module receives the raw projection data, which is typically in the form of a sine wave. It performs a logarithmic transformation on the raw projection data, that is, it calculates the following for each projection value p: Where p0 is the unaffected reference intensity, typically the maximum projected value. Next, the transformed data is linearly normalized, mapping it to the interval between 0 and 1. After normalization, the preprocessing module applies two data augmentation strategies to the sine wave domain:

[0039] Within a preset noise level range, such as a Gaussian distribution with a standard deviation of 0.01 to 0.05, Gaussian noise is added to the normalized projection data to simulate the statistical noise caused by insufficient photon counts in low-dose CT scans.

[0040] Within a preset angle discard range, such as randomly discarding 10% to 30% of continuous or non-continuous projection views, and setting some projection views to zero, the undersampling effect under sparse angle or fast scanning conditions is simulated.

[0041] After the above processing, the data is split into two outputs:

[0042] One path is used to send the low-quality projection data to be reconstructed into the dual-path generator module;

[0043] The other path sends the raw, low-quality projection data to the consistency constraint module for subsequent physical model consistency verification.

[0044] then,

[0045] The dual-path generator module employs an encoder-decoder architecture, introducing a dual-path parallel feature extraction mechanism in the encoder to focus on modeling local details and global structural information. The encoder consists of five cascaded convolutional downsampling blocks, each containing two convolutional layers, one batch normalization layer, and one modified linear unit activation function. All convolutional layers use 3×3 kernels with an initial channel count of 64, which doubles with each downsampling level. Specifically, the output channel counts for the first to fifth downsampling blocks are 64, 128, 256, 512, and 1024, respectively. The convolutional stride is uniformly set to 2, halving the spatial size of the feature map after each downsampling level, ultimately reducing it from the original 512×512 input to 16×16. At the third downsampling block of the encoder, where the output size is 128×128 and the number of channels is 256, the feature map is copied and distributed in parallel to both the local detail path and the global structural path.

[0046] The local detail path includes a residual dense block containing four sequentially connected convolutional layers, each employing a 3×3 convolutional kernel and 32 output channels. Furthermore, the residual dense block utilizes a dense connection strategy:

[0047] The output of layer 1 also serves as the input of layers 2, 3, and 4.

[0048] The output of layer 2 also serves as the input of layers 3 and 4;

[0049] The output of layer 3 is used as the input of layer 4.

[0050] All intermediate outputs are concatenated along the channel dimension to form a cumulative feature tensor.

[0051] In some alternative implementations, the output of the first convolutional layer is also passed directly to a 1×1 convolutional fusion layer at the end of the residual dense block via a skip connection, forming a local residual learning structure.

[0052] The global structure path is implemented by a non-local attention module, which performs a 1×1 convolution on the input 256-channel feature map to reduce its dimensionality to 128 channels, generating query, key, and value feature maps respectively.

[0053] Furthermore, the query feature map and the key feature map are multiplied by a matrix to obtain a similarity score matrix between locations. This matrix is ​​then normalized along the key dimension using the Softmax function to generate an attention weight map. This weight map is multiplied by the value feature map, and the result is then enlarged back to 256 channels via a 1×1 convolution. Finally, it is added to the original input feature map to output a feature map with global context enhancement.

[0054] The output feature maps of the two paths mentioned above are concatenated along the channel dimension to form a fused feature map with 512 channels, which is then input into the decoder. The decoder consists of 5 cascaded transposed convolutional upsampling blocks. Each block contains one transposed convolutional layer (kernel size 4×4, stride 2), one batch normalization layer, and one modified linear unit activation function.

[0055] The number of channels in the decoder is halved from 1024 to 64 in each step. Each upsampling block receives feature maps from the corresponding downsampling block of the encoder through skip connections. For example, the first block of the decoder (input 16×16×1024) receives the output of the fifth block of the encoder; the second block (32×32×512) receives the output of the fourth block of the encoder, and so on.

[0056] These skip connections fuse shallow detail information from the encoder with deep semantic information from the decoder through channel splicing, achieving multi-scale feature collaborative reconstruction. The final output layer of the decoder uses a 1×1 convolutional layer with a hyperbolic tangent activation function to map the feature map into a single-channel reconstructed CT image with a value range of [-1, 1], and the size is consistent with the target high-quality CT image, typically 512×512 pixels.

[0057] Immediately afterwards,

[0058] The attention discriminator module employs a pyramidal architecture, comprising three parallel discriminator sub-networks that process input images at original resolutions of 512×512, half-resolution 256×256, and quarter-resolution 128×128, respectively. Each sub-network has the same structure, consisting of four convolutional downsampling blocks. Each convolutional downsampling block includes one convolutional layer, one instance normalization layer, and one leak-corrected linear unit activation function with a negative slope of 0.2.

[0059] After the second and third convolutional layers of each subnetwork, a channel-space collaborative attention module is integrated. The workflow of this module is as follows:

[0060] The input feature map is subjected to global average pooling and global max pooling respectively, resulting in two 1×1×C vectors (C is the number of channels). The two vectors are concatenated and then input into a shared multilayer perceptron.

[0061] In some optional implementations, the multilayer perceptron consists of two fully connected layers. The first layer compresses the number of channels to C / 16, and the second layer restores it to C. Modified linear units are used for activation in between, and the final output is channel attention weights. Simultaneously, the attention discriminator module performs max pooling and average pooling on the input feature map along the channel dimension, respectively, to obtain two H×W×1 two-dimensional feature maps. These are then concatenated and used to generate spatial attention weights through a 7×7 convolutional layer.

[0062] Ultimately, the channel attention weights are broadcast along the channel dimension, and the spatial attention weights are broadcast along the spatial dimension. The two are multiplied together and then multiplied element-wise with the original input feature map to achieve adaptive feature calibration.

[0063] Each of the three subnetworks outputs a scalar score, which is then fused by a fully connected layer. The sigmoid activation function outputs a probability value between 0 and 1, representing the likelihood that the input image is a real, high-quality CT image. This probability value is fed back to the adversarial training and optimization module as part of the adversarial loss.

[0064] Then,

[0065] The consistency constraint module integrates a differentiable forward projection operator based on a distance-driven model. This model accurately simulates the attenuation process of X-rays passing through the reconstructed image by calculating the intersection length of the ray and the pixel grid, thereby improving the image domain reconstruction result I. rec Projecting back to the sinusoidal domain generates simulated projection data P. sim Subsequently, two losses were calculated: one was the L2 norm loss for the sinusoidal graph domain. , where P low The first is the low-quality projection data to be reconstructed output by the preprocessing module; the second is the structural similarity index loss in the image domain. This loss is calculated by multiplying the brightness, contrast, and structure components, and is used to measure the difference between the reconstructed image and the corresponding real high-quality CT image. gt Similarity in perceived quality. The consistency constraint module assigns these two losses to preset weights (e.g., =0.8, =0.2) Weighted summation to generate the consistency loss value: And input it into the adversarial training and optimization module.

[0066] at last,

[0067] The adversarial training and optimization module receives adversarial loss from the attention discriminator module. and physical consistency loss from consistency constraint modules The adversarial loss employs the Wasserstein distance loss with gradient penalty, defined as:

[0068] Where D() is the discriminator output (without passing through the Sigmoid); I fake To generate an image; I real It is a real image; These are random interpolation samples between the real image and the generated image; This is the gradient penalty coefficient, usually set to 10; Let be the mathematical expectation.

[0069] The generator's total loss function is defined as:

[0070]

[0071] in, It is the pixel-level norm loss in the image domain.

[0072] In some optional implementations, the calculation method is to sum the absolute values ​​of the differences between corresponding pixels in the reconstructed image and the real high-quality CT image, and then take the average. The weighting coefficient λ phys and λ pixel Let these values ​​be 1.0 and 100 respectively. The total loss function of the discriminator is:

[0073]

[0074] During training, the adversarial training and optimization modules employ an alternating optimization strategy:

[0075] In each iteration, the generator parameters are fixed first, and the discriminator parameters are updated once.

[0076] Then fix the discriminator parameters and update the generator parameters once.

[0077] The network parameters were optimized using an adaptive moment estimation algorithm. The initial learning rates of both the generator and the discriminator were set to 0.0001 and gradually decayed to 0 over 200 training epochs using a cosine annealing strategy. The batch size was set to 4. The training data came from anonymized CT scan datasets from public or partner hospitals, including thousands of cases covering different anatomical sites.

[0078] Example 2

[0079] Based on Example 1, this example further expands the application scenarios of the system and optimizes its generalization ability for different CT equipment manufacturers, different scanning protocols and different anatomical sites. The improvement of this example lies in the dynamic adaptation mechanism of the preprocessing module and the consistency constraint module.

[0080] In this embodiment, the preprocessing module incorporates a device fingerprinting subunit. This subunit analyzes metadata of the input raw projection data, such as device model, tube voltage, tube current, detector type, reconstruction algorithm identifier, etc., or the statistical characteristics of the data itself, such as noise power spectrum, projection value distribution skewness, etc., to automatically identify the manufacturer and model category of the CT scanning device. Based on the identification results, the preprocessing module dynamically adjusts the reference intensity p0, normalization range, and data augmentation parameters in the logarithmic transformation. For example, for GE Healthcare devices, whose projection data has a wider dynamic range, a 99th percentile truncation is used during normalization; while for Siemens devices, a 95th percentile truncation is used. The standard deviation of the noise addition is also scaled according to the inherent noise level of the device to ensure that the simulated low-dose conditions match the actual device performance.

[0081] The angle discarding strategy can also be adaptively adjusted according to the maximum scanning angle supported by the device (such as 180 degrees vs 360 degrees) to avoid generating undersampling modes that are not physically feasible.

[0082] In this embodiment, the consistency constraint module is upgraded to a configurable forward projection operator library. This library pre-stores geometric parameter templates corresponding to mainstream CT equipment manufacturers (such as GE, Siemens, Philips, and Canon), including source-axis distance, axis-detector distance, detector unit size, focal spot size, and scan trajectory type. When the preprocessing module identifies the equipment type, the consistency constraint module automatically loads the matching geometric parameters and initializes the corresponding differentiable forward projection operator to ensure the accuracy of the physical model constraints and avoid reconstruction deviations caused by geometric parameter mismatch.

[0083] In this embodiment, the dual-path generator module adds an input condition embedding mechanism. Before the first convolutional layer of the encoder, the device type identifier (one-hot encoded) and scanning protocol parameters (such as dose level and slice thickness) are mapped into feature vectors through an embedding layer and broadcast to the entire input projection data tensor as an additional channel input. This condition persists throughout the generation process, enabling the generator to adaptively adjust the reconstruction strategy for different input conditions. For example, under extremely low dose conditions, the generator tends to enhance the gain of local detail paths to compensate for signal-to-noise ratio loss; while under high-resolution scanning conditions, it strengthens the weight of global structural paths to maintain anatomical coherence.

[0084] The attention discriminator module also incorporates a conditional discrimination mechanism. The first convolutional layer of each of its three discriminator sub-networks receives the same conditional embedding vector as the generator, and adjusts the bias term of the convolutional kernel through affine transformation, enabling the discriminator to dynamically adjust its discrimination criteria based on the input conditions. For example, for lung CT images, the discriminator focuses more on the clarity of blood vessel and nodule edges; for abdominal CT images, it focuses more on the contrast between organ boundaries and adipose tissue.

[0085] In this embodiment, the adversarial training and optimization module adopts a course learning strategy. The training data is divided into multiple difficulty levels according to the dose level or undersampling rate. In the early stage of training, only medium difficulty samples (such as 50% dose, 20% angle drop) are used. After the model converges, high difficulty samples (such as 10% dose, 50% angle drop) are gradually introduced. This strategy effectively alleviates the gradient explosion or mode collapse problem caused by poor input quality in the early stage of training.

[0086] Through the above improvements, the system in this embodiment demonstrates good generalization ability on test sets across devices and anatomical sites. On a validation set including devices from 5 different manufacturers, 10 scanning protocols, and 6 anatomical sites, the peak signal-to-noise ratio of the reconstructed images is improved by an average of 2.1 dB, the structural similarity index is improved by 0.05, and no serious anatomical distortion is observed. This embodiment is suitable for large medical groups or multi-center clinical research scenarios.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0088] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A generative adversarial network processing system for radiology CT image reconstruction, comprising a pre-processing module, characterized in that, Also includes: The system includes a dual-path generator module, an attention discriminator module, a consistency constraint module, and an adversarial training and optimization module; among which... The dual-path generator module is used to reconstruct high-quality CT images from low-quality projection data to be reconstructed; the dual-path generator module is an encoder-decoder architecture, and integrates local detail paths and global structure paths in the middle of the encoder; the encoder includes a convolutional downsampling block, the local detail path includes a residual dense block, the global structure path includes a non-local attention module, and the decoder includes a transposed convolutional upsampling block. The attention discriminator module is used to discriminate the input image. The attention discriminator module has a pyramid structure and includes parallel discriminator subnetworks. The consistency constraint module is used to force the reconstructed image and the original projection data to meet the physical model constraints of CT imaging during the training process. The consistency constraint module includes a differentiable forward projection operator, which is used to simulate the reconstruction result in the image domain back to the projection domain to generate simulated projection data. The consistency constraint module calculates the norm loss of the simulated projection data and the low-quality projection data to be reconstructed in the sinusoidal domain, and calculates the structural similarity index loss of the reconstructed image and the corresponding real high-quality CT image in the image domain. The two losses are weighted and summed to generate a consistency loss value. The adversarial training and optimization module is used to receive adversarial loss and consistency loss, and to iteratively update the network parameters of the dual-path generator module and the attention discriminator module using an alternating optimization strategy.

2. The system according to claim 1, characterized in that, The data preprocessing module's processing steps include at least: The original projection data is subjected to logarithmic transformation and normalization to map the data value range to between 0 and 1; The normalized projection data is subjected to a sinusoidal domain data augmentation operation, which includes adding Gaussian noise within a preset noise level range to simulate low-dose conditions, and randomly discarding a portion of the projection view within a preset angle range to simulate undersampling conditions.

3. The system according to claim 1, characterized in that, The residual dense block includes multiple convolutional layers. The output of each convolutional layer is concatenated with the input of all subsequent convolutional layers in the channel dimension. The output of the first convolutional layer in the residual dense block is directly passed to the 1x1 convolutional fusion layer at the end of the residual dense block through a skip connection.

4. The system according to claim 1, characterized in that, The nonlocal attention module is used to calculate the correlation weight between any two location feature vectors, and the calculation process includes at least the following: The input feature map is reduced in dimensionality by performing a 1x1 convolution to obtain the query feature map, key feature map, and value feature map. The query feature map and the key feature map are multiplied by a matrix and then normalized using the Softmax function to obtain the attention weight map. The attention weight map is multiplied by the value feature map, and the result is increased in dimensionality by a 1x1 convolution before being added to the original input feature map.

5. The system according to claim 1, characterized in that, The adversarial loss is a Wasserstein distance loss with gradient penalty; the total loss function of the adversarial training and optimization module is a weighted sum of the adversarial loss, consistency loss, and image domain pixel-level norm loss.

6. The system of claim 1, wherein, The alternating optimization strategy is as follows: First, fix the generator parameters and update the discriminator parameters once, then fix the discriminator parameters and update the generator parameters once. The network parameters were iteratively optimized using an adaptive moment estimation algorithm, with the initial learning rate set to 0.0001 and decayed using a cosine annealing strategy.

7. The system of claim 1, wherein, The input of the dual-path generator module is connected to the output of the preprocessing module.

8. The system according to claim 1, characterized in that, The input of the attention discriminator module is connected to the output of the dual-path generator module.

9. The system according to claim 1, characterized in that, The input of the consistency constraint module is connected to the output of the preprocessing module.

10. The system according to claim 1, characterized in that, The input of the adversarial training and optimization module is connected to the output of the attention discriminator module and the output of the consistency constraint module, respectively.

Images