X-ray machine projection parameter calibration method

Abstract

The application discloses an X-ray machine projection parameter calibration method and relates to the technical field of medical image and diagnosis equipment. A double-channel feature extraction module shields a high-noise area response in a spatial dimension dynamically, meanwhile, a channel gain mechanism is utilized to retain key anatomical structure features, and quantum noise interference on edge positioning is effectively inhibited. A generative adversarial framework converts parameter regression into an optimization task driven by reconstruction quality, a discriminator is based on a structural similarity constraint of a pre-trained deep feature, and a generator is forced to output an offset meeting a physical projection rule, thus breaking a bottleneck that an optimization algorithm falls into a local extreme value under low signal-to-noise ratio data. Triple constraint terms in a loss function balance parameter precision and image authenticity through a weight self-adaptive mechanism, geometric parameters are converged quickly in an early training stage, texture consistency is strengthened in a middle and late training stage, and artifact distortion caused by insufficient dose is reduced.

Description

Technical Field

[0001] This invention relates to the field of medical imaging and diagnostic equipment technology, and in particular to a method for calibrating projection parameters of an X-ray machine. Background Technology

[0002] In the fields of medical and pet diagnostics, low-dose X-ray imaging technology has become a rigid requirement to reduce the risk of radiation exposure to live subjects. Current equipment generally achieves dose control by reducing tube current or shortening exposure time, but the resulting deterioration of the signal-to-noise ratio of the projected image is becoming increasingly prominent. Especially when photographing easily moving live subjects, such as pets with rising and falling chests, the detection must be completed within an extremely short exposure window, which further limits the number of available photons, resulting in significant quantum noise and contrast loss in the projected image.

[0003] For parameter calibration in low signal-to-noise ratio environments, most current solutions rely on projection data preprocessing and feature enhancement techniques. For example, deep learning-based gain correction models are used to denoise the original projection map, or multi-scale feature fusion networks are used to extract robust marker point coordinates. Some solutions transform the geometric parameter estimation task into a sequence optimization problem and correct the projection matrix through iterative reconstruction feedback.

[0004] However, in practical applications, the noise suppression process of existing solutions tends to blur the edges of key structures, which amplifies the feature point positioning deviation; the sparsity of low-dose projection data causes the optimization algorithm to get stuck in local extrema; in scenarios lacking high-contrast references, such as soft tissue imaging of small animals, the system may not be able to converge to an effective solution, and the calibration failure rate increases significantly; in addition, the real-time requirements further restrict the deployment space of complex algorithms. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a method for calibrating X-ray machine projection parameters to address the problem that low-dose X-ray imaging leads to a deterioration in the signal-to-noise ratio of the projection image, and that existing calibration methods are susceptible to noise interference and therefore fail.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a method for calibrating projection parameters of an X-ray machine, comprising:

[0009] Step S1: Obtain X-ray projection images of the target object under low-dose conditions;

[0010] Step S2: Input the projected image into a pre-trained noise estimator to generate a corresponding noise distribution map;

[0011] Step S3: The projected image and noise distribution map are processed simultaneously by the dual-channel feature extraction module to output a fused feature map;

[0012] Step S4: Input the fused feature map into the geometric parameter regression network to generate projection parameter offset; wherein, the geometric parameter regression network is a neural network optimized through adversarial training;

[0013] Step S5: Correct the geometric parameters of the X-ray machine according to the offset.

[0014] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, the dual-channel feature extraction module includes:

[0015] The first branch of the convolutional neural network extracts deep features from the projected image;

[0016] The noise attention unit in the second branch generates a spatial mask based on the noise distribution map;

[0017] The feature fusion unit multiplies the deep features element-wise with the spatial mask to suppress the response in noisy regions;

[0018] The first branch of the convolutional neural network includes:

[0019] The four-level residual module contains a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function in each level.

[0020] The cross-level skip connection concatenates shallow edge features with deep semantic features.

[0021] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, the noise attention unit includes:

[0022] The noise distribution map is processed sequentially by convolutional layers and the Sigmoid activation function to generate a noise weight matrix in the [0,1] interval;

[0023] The noise weight matrix is ​​weighted and fused with the deep feature map.

[0024] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, step S2, the weighted fusion process of the noise weight matrix and the deep feature map includes:

[0025] After obtaining the noise weight matrix, a complementary mask is constructed to suppress the response of the high-noise region, expressed as:

[0026] M i,j =1-W n (i,j),

[0027] Among them, W n(i,j) represents the noise weight of pixel (i,j) in the noise weight matrix, taking values ​​from 0 to 1, dimensionless, where i is the row index and j is the column index, M i,j It is a complementary spatial mask, dimensionless;

[0028] Perform pointwise decay on deep features:

[0029]

[0030] Among them, F d,c,i,j Let be the normalized intensity of the deep feature map at pixel (i,j) in channel c, where c is the channel index. The intermediate feature after spatial noise suppression, and F d,c,i,j Same dimension;

[0031] An adaptive adjustment of the noise suppression ratio of each semantic channel is introduced by introducing channel gain, which is expressed as:

[0032]

[0033] Among them, g c σ is the channel gain coefficient, taking values ​​from 0 to 1, dimensionless; σ(·) is the Sigmoid function; δ(·) is the ReLU function; w1 and w2 are the weight matrices of the first and second fully connected layers, respectively; and can be trained real tensors. For the channel The result of the global average pooling operation, in scalar form;

[0034] The fusion features are obtained using the residual method:

[0035]

[0036] Among them, F f,c,i,j This is the output of the fused feature map.

[0037] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, the geometric parameter regression network is a generative adversarial network framework, comprising:

[0038] The generator network takes the fused feature map as input and outputs the projection parameter offset.

[0039] The discriminator network takes the projected image reconstructed with corrected parameters as input and outputs the probability of authenticity.

[0040] The generator network includes:

[0041] Fully connected layers flatten the fused feature maps into feature vectors;

[0042] The three-layer perceptron regresses the projection parameter offset, and the output dimension is consistent with the number of geometric parameters;

[0043] When the discriminator network constructs the supervisory signal:

[0044] Image features are extracted using a pre-trained VGG16 network;

[0045] Calculate the structural similarity index between real samples and generated samples in the feature space.

[0046] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, the training loss function of the generator network includes:

[0047] The mean square error term of the parameter offset;

[0048] Adversarial loss term output by the discriminator network.

[0049] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, wherein: the training loss function of the generator network is decomposed into three terms, including:

[0050] Parameter regression terms:

[0051]

[0052] Among them, L reg The mean squared error loss is given by K, where K is the number of geometric parameters and k is the parameter index, with values ​​1 ≤ k ≤ K. The offset of the k-th parameter predicted by the generator. This is the corresponding truth offset;

[0053] Adversarial consistency terms:

[0054] L adv = -log D(I gen ),

[0055] Among them, L adv Let I be the adversarial loss of the generator, D(·) be the probability of authenticity of the discriminator's output, and I be the probability of authenticity of the discriminator's output. gen For application Post-reconstructed low-dose projection images, This is the vector of geometric parameter offsets output by the generator;

[0056] Feature structure terms:

[0057]

[0058] Among them, L ssim For structural similarity loss, SSIM(·,·) is the structural similarity index, where SSIM(·,·)∈[0,1]. A larger output indicates greater structural similarity. For pre-training the VGG16 feature extractor, Ireal High-dose projection images in the same posture;

[0059] The comprehensive loss is defined as:

[0060] L G =αL reg +βL adv +γL ssim ,

[0061] Among them, L G The generator's total loss is represented by α, β, and γ, which are the weighting coefficients of the three terms.

[0062] Introducing a weight adaptive mechanism:

[0063] (α,β,γ)=Softmax(u),

[0064] Where Softmax(·) is the standard softmax function. The weight vector is a trainable weight vector whose elements are updated during training. The sum of the three weights is equal to 1, and each weight takes values ​​in the range (0,1). The weight coefficients are obtained by normalizing the trainable vector using the Softmax function.

[0065] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, the discriminator network constructs the supervision signal through the following steps:

[0066] High-dose projection images are used as real samples, and low-dose projection images reconstructed after parameter correction are used as generated samples.

[0067] Calculate the difference between the two in terms of structural similarity index.

[0068] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, the noise estimator is pre-trained in the following manner:

[0069] Input pairs of low-dose and conventional-dose projection images;

[0070] The training objective is to minimize the residual maps of the two objects.

[0071] The training of the noise estimator includes:

[0072] U-Net architecture encoders extract multi-scale noise patterns;

[0073] Jump connections fuse low-frequency noise distribution with high-frequency texture information.

[0074] As a preferred embodiment of the X-ray machine projection parameter calibration method of the present invention, the calibration in step S5 is completed within a single exposure frame period of the X-ray machine, and the entire process is executed automatically without manual intervention.

[0075] The beneficial effects of this invention are as follows: This invention significantly improves calibration robustness under low-dose conditions through a multi-level adaptive mechanism; the dual-channel feature extraction module dynamically shields high-noise region responses in the spatial dimension, while utilizing a channel gain mechanism to retain key anatomical structural features, effectively suppressing quantum noise interference on edge localization and avoiding feature point extraction failures caused by image blurring in traditional methods; the generative adversarial framework transforms parameter regression into a reconstruction quality-driven optimization task, with the discriminator constrained by the structural similarity of pre-trained deep features, forcing the generator to output an offset that conforms to the laws of physical projection, thus overcoming the bottleneck of optimization algorithms getting trapped in local extrema under low signal-to-noise ratio data; furthermore, the triple constraint term in the loss function balances parameter accuracy and image realism through a weight adaptive mechanism, rapidly converging geometric parameters in the early training stage and enhancing texture consistency in the mid-to-late stages, reducing artifact distortion caused by insufficient dose; the entire process is based on a lightweight network design, completing the closed-loop operation from noise estimation to parameter correction within a single exposure cycle, eliminating dependence on phantom scanning and manual intervention, and meeting the timeliness requirements of live imaging.

[0076] This invention is particularly suitable for low-dose scenarios such as pet chest cavity or children's limbs with a signal-to-noise ratio of less than 15dB, maintaining calibration reliability while ensuring radiation safety. Attached Figure Description

[0077] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0078] Figure 1 This is a flowchart illustrating the X-ray machine projection parameter calibration method in Example 1. Detailed Implementation

[0079] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0080] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0081] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0082] Example 1, referring to Figure 1 This embodiment provides a method for calibrating the projection parameters of an X-ray machine, including the following steps:

[0083] Step S1: Obtain X-ray projection images of the target object under low-dose conditions;

[0084] Step S2: Input the projected image into the pre-trained noise estimator to generate the corresponding noise distribution map;

[0085] Step S3: The projection image and noise distribution map are processed simultaneously through the dual-channel feature extraction module to output the fused feature map;

[0086] The dual-channel feature extraction module includes:

[0087] The first branch of the convolutional neural network extracts deep features from the projected image;

[0088] The noise attention unit in the second branch generates a spatial mask based on the noise distribution map;

[0089] The feature fusion unit multiplies deep features element-wise with the spatial mask to suppress the response in noisy regions;

[0090] The first branch of the convolutional neural network includes:

[0091] The four-level residual module contains a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function in each level.

[0092] The cross-level skip connection concatenates shallow edge features with deep semantic features;

[0093] The noise attention unit includes:

[0094] The noise distribution map is processed sequentially by convolutional layers and the Sigmoid activation function to generate a noise weight matrix in the [0,1] interval;

[0095] The noise weight matrix is ​​weighted and fused with the deep feature map;

[0096] In step S2, the weighted fusion process of the noise weight matrix and the deep feature map includes the following steps:

[0097] After obtaining the noise weight matrix, a complementary mask is constructed to suppress the response of the high-noise region, expressed as:

[0098] Mi,j =1-W n (i,j),

[0099] Among them, W n (i,j) represents the noise weight of pixel (i,j) in the noise weight matrix, taking values ​​from 0 to 1, dimensionless, where i is the row index and j is the column index, M i,j It is a complementary spatial mask, dimensionless;

[0100] Perform pointwise decay on deep features:

[0101]

[0102] Among them, F d,c,i,j Let be the normalized intensity of the deep feature map at pixel (i,j) in channel c, where c is the channel index. The intermediate feature after spatial noise suppression, and F d,c,i,j Same dimension;

[0103] An adaptive adjustment of the noise suppression ratio of each semantic channel is introduced by introducing channel gain, which is expressed as:

[0104]

[0105] Among them, g c σ is the channel gain coefficient, taking values ​​from 0 to 1, dimensionless; σ(·) is the Sigmoid function; δ(·) is the ReLU function; w1 and w2 are the weight matrices of the first and second fully connected layers, respectively; and can be trained real tensors. For the channel The result of the global average pooling operation, in scalar form;

[0106] The fusion features are obtained using the residual method:

[0107]

[0108] Among them, F f,c,i,j Output the fused feature map;

[0109] Specifically, this fusion strategy suppresses the activation of high-noise pixels in the spatial dimension through complementary masks, and then redistributes weights to different semantic layers with channel gain to achieve dual adaptive adjustment of space and channel. Compared with the approach that relies solely on single-scale attention, this method preserves structural edge information and reduces the interference of texture artifacts on subsequent regression networks.

[0110] Step S4: Input the fused feature map into the geometric parameter regression network to generate projection parameter offsets; wherein, the geometric parameter regression network is a neural network optimized through adversarial training;

[0111] The geometric parameter regression network is a generative adversarial network framework, including:

[0112] The generator network takes the fused feature map as input and outputs the projection parameter offset.

[0113] The discriminator network takes the projected image reconstructed with corrected parameters as input and outputs the probability of authenticity.

[0114] Generator networks include:

[0115] Fully connected layers flatten the fused feature maps into feature vectors;

[0116] The three-layer perceptron regresses the projection parameter offset, and the output dimension is consistent with the number of geometric parameters;

[0117] When the discriminator network constructs the supervisory signal:

[0118] Image features are extracted using a pre-trained VGG16 network;

[0119] Calculate the structural similarity index between real samples and generated samples in the feature space;

[0120] The training loss function for a generator network includes:

[0121] The mean square error term of the parameter offset;

[0122] The adversarial loss term output by the discriminator network;

[0123] In the training loss function of the generator network, to simultaneously constrain geometric parameter regression bias, suppress artifacts, and maintain the consistency of the discriminator, the generator loss is split into three terms, including:

[0124] Parameter regression terms:

[0125]

[0126] Among them, L reg The mean squared error loss is given by K, where K is the number of geometric parameters and k is the parameter index, with values ​​1 ≤ k ≤ K. The offset of the k-th parameter predicted by the generator. This is the corresponding truth offset;

[0127] Adversarial consistency terms:

[0128] L adv =-logD(I gen ),

[0129] Among them, L adv Let I be the adversarial loss of the generator, D(·) be the probability of authenticity of the discriminator's output, and I be the probability of authenticity of the discriminator's output. gen For application Post-reconstructed low-dose projection images, This is the vector of geometric parameter offsets output by the generator;

[0130] Feature structure terms:

[0131]

[0132] Among them, L ssim For structural similarity loss, SSIM(·,·) is the structural similarity index, where SSIM(·,·)∈[0,1]. A larger output indicates greater structural similarity. For pre-training the VGG16 feature extractor, I real High-dose projection images in the same posture;

[0133] The comprehensive loss is defined as:

[0134] L G =αL reg +βL adv +γL ssim ,

[0135] Among them, L G The generator's total loss is represented by α, β, and γ, which are the weighting coefficients of the three terms.

[0136] Introducing a weight adaptive mechanism:

[0137] (α,β,γ)=Softmax(u),

[0138] Where Softmax(·) is the standard softmax function. It is a trainable weight vector whose elements are updated during training. The sum of the three is equal to 1, and each takes a value in the range (0,1) to maintain the balance of the three losses at different stages.

[0139] Specifically, the loss function structure suppresses geometric parameter drift through mean squared error, the adversarial term maintains the authenticity of the reconstructed image in the discriminator feature distribution, and the structural term uses SSIM in the pre-trained feature space to measure texture consistency. The three complement each other. The softmax weight vector tends to favor the regression term in the early stage to accelerate parameter convergence. In the middle and late stages, the proportion of adversarial and structural terms is gradually increased, so that the generator can suppress noise texture while maintaining fine edges.

[0140] The discriminator network constructs the supervision signal through the following steps:

[0141] High-dose projection images are used as real samples, and low-dose projection images reconstructed after parameter correction are used as generated samples.

[0142] Calculate the difference between the two indices on structural similarity.

[0143] Step S5: Correct the geometric parameters of the X-ray machine based on the offset;

[0144] The noise estimator is pre-trained in the following way:

[0145] Input pairs of low-dose and conventional-dose projection images;

[0146] The training objective is to minimize the residual maps of the two objects.

[0147] Training the noise estimator includes:

[0148] U-Net architecture encoders extract multi-scale noise patterns;

[0149] Jump connections fuse low-frequency noise distribution with high-frequency texture information;

[0150] The correction in step S5 is completed within a single exposure frame period of the X-ray machine, and the entire process is executed automatically without manual intervention.

[0151] In summary, this embodiment significantly improves calibration robustness under low-dose conditions through a multi-level adaptive mechanism; the dual-channel feature extraction module dynamically shields the response of high-noise regions in the spatial dimension, while retaining key anatomical structural features using a channel gain mechanism, effectively suppressing the interference of quantum noise on edge localization and avoiding feature point extraction failures caused by image blurring in traditional methods; the generative adversarial framework transforms parameter regression into an optimization task driven by reconstruction quality, and the discriminator, based on the structural similarity constraints of pre-trained deep features, forces the generator to output an offset that conforms to the laws of physical projection, thus breaking the bottleneck of optimization algorithms getting trapped in local extrema under low signal-to-noise ratio data; furthermore, the triple constraint term in the loss function balances parameter accuracy and image realism through a weight adaptive mechanism, rapidly converging geometric parameters in the early stages of training and strengthening texture consistency in the mid-to-late stages, reducing artifact distortion caused by insufficient dose; the entire process is based on a lightweight network design, completing the closed-loop operation from noise estimation to parameter correction within a single exposure cycle, eliminating dependence on phantom scanning and manual intervention, and meeting the timeliness requirements of live imaging.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method of calibrating projection parameters of an X-ray machine, characterized by, The method comprises the following steps: Step S1, acquiring an X-ray projection image of a target object under a low-dose condition; Step S2, inputting the projection image into a pre-trained noise estimator to generate a corresponding noise distribution map; Step S3, synchronously processing the projection image and the noise distribution map through a double-channel feature extraction module to output a fused feature map; Step S4, inputting the fused feature map into a geometric parameter regression network to generate a projection parameter offset; wherein the geometric parameter regression network is a neural network optimized through adversarial training; Step S5, correcting the geometric parameters of an X-ray machine according to the offset; The double-channel feature extraction module comprises: A convolutional neural network of a first branch extracts deep features of the projection image; A noise attention unit of a second branch generates a spatial mask according to the noise distribution map; A feature fusion unit element-wise multiplies the deep features and the spatial mask to suppress the response of the noise region; The convolutional neural network of the first branch comprises: A four-level residual module, each level containing a 3x3 convolutional layer, a batch normalization layer and a ReLU activation function; A cross-level skip connection splices shallow edge features and deep semantic features; The geometric parameter regression network is a generative adversarial network framework, comprising: A generator network takes the fused feature map as input and outputs a projection parameter offset; A discriminator network takes a projection image reconstructed by the corrected parameters as input and outputs a probability of authenticity; The generator network comprises: A fully connected layer flattens the fused feature map into a feature vector; A three-layer perceptron regresses the projection parameter offset, and the output dimension is consistent with the number of geometric parameters; When constructing a supervision signal, the discriminator network: Uses a pre-trained VGG16 network to extract image features; Calculates the structural similarity index of real samples and generated samples in the feature space.

2. The method of claim 1, wherein the step of calibrating the projection parameters of the X-ray machine is performed by using a calibration phantom. The noise attention unit comprises: The noise distribution map is sequentially processed by a convolutional layer and a Sigmoid activation function to generate a noise weight matrix in the interval [0, 1]; The noise weight matrix and the deep feature map are weighted and fused.

3. The method of claim 2, wherein the step of determining the projection parameter is performed by using a calibration phantom. In step S3, the process of weighting and fusing the noise weight matrix and the deep feature map comprises: After obtaining the noise weight matrix, a complementary mask is constructed to suppress the response of the high-noise region, represented as: ， wherein, is the noise weight of pixel of the noise weight matrix, taking values 0-1, dimensionless, is the row index, is the column index, is the complementary spatial mask, dimensionless; Point-by-point attenuation is performed on the deep features: ， wherein, is the normalized intensity of the deep feature map at the channel , pixel , is the channel index, is the intermediate feature after spatial noise suppression, and is of the same dimension. Channel gain is introduced to adaptively adjust the noise suppression ratio of each semantic channel, and the gain is represented as: ， where, is the channel gain coefficient, taking values 0-1, dimensionless, is the Sigmoid function, is the ReLU function, and are the first and second fully connected layer weight matrices, trainable real tensors, is the global average pooling result applied to the channel , scalar; The fused features are obtained in a residual manner: ， wherein, is the output of the fused feature map.

4. The method of claim 3, wherein the step of determining the projection parameter is performed by using a calibration phantom. The training loss function of the generator network includes: The mean square error term of the parameter offset; The adversarial loss term output by the discriminator network.

5. The method of claim 4, wherein the step of determining the projection parameter is performed by using a calibration phantom having a plurality of known shapes and sizes. 5 In the training loss function of the generator network, the generator loss is divided into three terms, including: The parameter regression term: ， wherein, is a mean squared error loss, is a number of geometric parameters, is a parameter index, taking values , is a generator-predicted offset of the th parameter, is a corresponding ground truth offset; The adversarial consistency term: ， wherein, is an adversarial loss for the generator, is a probability of authenticity output by the discriminator, is an application of a post-reconstruction low-dose projection image, is a geometry parameter offset vector output by the generator; The feature structure term: ， wherein, is a structural similarity loss, is a structural similarity index, the larger the output, the more consistent the structure, is a pre-trained VGG16 feature extractor, is a high-dose projection image of the same pose; The comprehensive loss is defined as: ， wherein, is the generator total loss, , , are weight coefficients for the three terms; A weight adaptive mechanism is introduced: ， wherein, is a standard softmax function, is a trainable weight vector whose elements are updated with training, the sum of which equals 1, and each of which takes a value range (0, 1); the weight coefficient is obtained by normalizing the trainable vector through the softmax function.

6. The method of claim 5, wherein the step of determining the projection parameter is performed by using a calibration phantom. The discriminator network constructs a supervision signal through the following steps: High-dose projection images are taken as real samples, and low-dose projection images reconstructed after parameter correction are taken as generated samples; Calculate the difference in structural similarity index between the two.

7. The method of claim 1, wherein the step of determining the projection parameters is performed by a computer program. The noise estimator is pre-trained in the following way: Input a pair of low-dose and regular-dose projection images; Minimize the residual error between the two as the training target; The training of the noise estimator comprises: The U-Net architecture encoder extracts multi-scale noise patterns; The skip connection fuses low-frequency noise distribution and high-frequency texture information.

8. The method of claim 1, wherein the step of calibrating the projection parameters of the X-ray machine is performed by a computer program. The correction of step S5 is completed within a single exposure frame period of the X-ray machine, and the entire process is automatically executed without human intervention.

Images