Self-consistency constrained self-adaptive low-dose CT image reconstruction method during testing

Abstract

The invention discloses a self-consistency constrained self-adaptive low-dose CT image reconstruction method during testing, and belongs to the field of medical images. The method comprises the steps that in the training process, an affine modulation module and a latent variable coding module are frozen, and diffusion trunk parameters are optimized; all parameters of a backbone network are generated through freezing diffusion, and in the testing process, parameters of an affine modulation module and latent variables of a slice latent variable coding module are optimized; when the optimization step number reaches a preset unfreezing step number, unfreezing and diffusing to generate an output layer of the backbone network; first noise is obtained through the reference prediction path, second noise is obtained through the original low-dose CT image consistency prediction path, and a low-dose CT pixel prediction value is obtained through an affine modulation module; and constructing an adaptive joint optimization target according to the first noise, the second noise, the latent variable, the low-dose CT pixel predicted value and the low-dose CT pixel true value. According to the invention, the structure fidelity and noise suppression performance of the low-dose CT image can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and more particularly to a test-time adaptive low-dose CT image reconstruction method with self-consistency constraints. Background Technology

[0002] Computed tomography (CT) scans, as an important tool in medical imaging diagnosis, can reconstruct the internal tissue structures of the human body through multi-angle X-ray projection. However, while traditional CT imaging improves image quality, it also increases the radiation dose. Excessive doses may pose radiation risks to patients. Therefore, low-dose CT imaging is commonly used in clinical practice to reduce radiation hazards.

[0003] However, low-dose scanning significantly reduces the signal-to-noise ratio of projected data, resulting in strong random noise and blurred details in the reconstructed images, thus affecting the accuracy of doctors' diagnostic judgment and lesion detection. Traditional low-dose CT reconstruction methods typically rely on high-dose images as supervisory signals, learning to remove noise from low-dose images. However, in practical applications, acquiring high-dose images not only increases the radiation risk to patients but is also constrained by ethical restrictions, equipment limitations, and data privacy protection, making it difficult to obtain sufficient paired samples, thus limiting model training. While existing deep learning denoising methods can achieve good results under specific doses and fixed noise distributions, their parameters are highly dependent on the training data distribution. When scanning conditions, equipment models, or noise characteristics change, the model's generalization ability significantly decreases, and the reconstruction results are prone to artifacts, structural distortion, and blurred details, failing to meet the requirements of clinical diagnosis for image clarity and stability.

[0004] Therefore, a novel test-time adaptive low-dose CT image reconstruction method based on self-consistency constraints is proposed. This method can automatically optimize model parameters without high-dose supervision, and achieve self-calibration and adaptation to different scanning conditions and noise distributions. It has important application value and promotion significance. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of insufficient model generalization ability, reliance on high-dose paired data, and inability to adaptively optimize according to noise characteristics in existing low-dose CT image reconstruction methods. This invention provides a test-time adaptive low-dose CT image reconstruction method with self-consistency constraints, which can self-calibrate for different scanning conditions and noise distributions, thereby effectively improving the structural detail reconstruction capability and noise suppression performance of low-dose CT images.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A test-time adaptive low-dose CT image reconstruction method with self-consistency constraints performs the following steps based on a network structure including a diffusion generation backbone network, a slice latent variable encoding module, and an affine modulation module based on an implicit representation network:

[0008] Step 1: Construct a dataset of CT images paired with low-dose and high-dose doses, and divide it into training and testing sets;

[0009] Step 2: Using low-dose CT images from the training set as input and high-dose CT images as reference, keep the affine modulation module and latent variable coding module frozen, and only optimize the parameters of the diffusion generation backbone network;

[0010] Step 3: Using low-dose CT images from the test set as input and high-dose CT images as reference, freeze all parameters of the backbone network generated by diffusion, and optimize the parameters of the affine modulation module and the latent variables of the slice latent variable encoding module.

[0011] The input to the slice latent variable encoding module is a low-dose CT image. For each input low-dose CT image, a corresponding slice latent variable is constructed, and a latent variable vector is output.

[0012] The affine modulation module takes the latent variable vector and the pixel position encoding of the input low-dose CT image as input to simultaneously generate low-dose CT pixel prediction values ​​and corresponding pixel-by-pixel affine modulation parameters.

[0013] Step 4: When the number of optimization steps reaches the preset number of unfreezing steps, the unfreezing diffusion generates the output layer of the backbone network, allowing it to participate in the optimization; when the number of optimization steps has not reached the preset number of unfreezing steps, proceed to Step 5.

[0014] Step 5: Input the low-dose CT images from the test set into the baseline prediction path to obtain the first noise; input the low-dose CT images from the test set after blind spot occlusion processing into the consistency prediction path to obtain the second noise; wherein, the baseline prediction path includes an optimized diffusion generation backbone network and an affine modulation module; the consistency prediction path includes an optimized diffusion generation backbone network and an affine modulation module.

[0015] Low-dose CT images from the test set are input into an optimized affine modulation module to obtain low-dose CT pixel prediction values.

[0016] Step 6: Based on the first noise, the second noise, the latent variable, the predicted value of low-dose CT pixels, and the actual value of low-dose CT pixels, construct an adaptive joint optimization objective, and perform adaptive iterative optimization on the adaptive joint optimization objective through multiple batches of test samples;

[0017] When the maximum number of iterations is reached, the currently optimized diffusion generation backbone network, slice latent variable encoding module and affine modulation module are used for the target low-dose CT image to obtain the corresponding reconstruction results;

[0018] If the maximum number of iterations has not been reached, return to step four.

[0019] In one or more embodiments, the diffusion generation backbone network is a GDM network, which achieves the mapping from noisy images to high-quality images through multi-step geometric diffusion backpropagation (i.e., achieving a mapping from low-dose CT to high-dose CT with the sharpness deviation within an acceptable range). The mapping process can be expressed by the following formula:

[0020] ;

[0021] in, Indicates diffusion time step The intermediate noise state, For the diffusion network in network parameters The following is a prediction of the noise component. It is a function of noise intensity. For time step interval, This is a constant term related to the Fisher–Rao information geometry metric.

[0022] In one or more embodiments, the slice latent variable encoding module constructs a paired latent variable for each CT slice and outputs a latent variable vector, thus obtaining a set of learnable latent vectors; wherein the output latent variable vector follows a normal distribution with a mean of 0 and a covariance matrix of σ²I. Initialize the variance for the latent variables, where I is the identity matrix.

[0023] In one or more embodiments, the pixel position encoding of the input affine modulation module is as follows: random Fourier transform is performed on the image coordinates of the input low-dose CT image to obtain random Fourier feature position encoding.

[0024] In one or more embodiments, the affine modulation module based on the implicit representation network takes slice latent variables and pixel position codes as input to simultaneously generate low-dose CT pixel prediction values ​​and corresponding pixel-by-pixel affine modulation parameters and perform modulation on the diffusion backbone features; the implicit representation network adopts a multilayer perceptron structure with a sinusoidal activation function, and the input consists of a latent variable vector. With random Fourier feature location encoding Together, they form a whole, in which positional encoding and affine generation processes are defined as follows:

[0025] ;

[0026] ;

[0027] in, Indicates normalized to Pixel coordinates in the coordinate system To follow a Gaussian distribution The frequency matrix, Indicates the number of Fourier bases; This represents the predicted value of pixels in low-dose CT scans. These are pixel-level affine parameters. d is the potential spatial dimension. This is the function corresponding to the multilayer perceptron structure.

[0028] In one or more embodiments, the pixel-wise affine modulation parameters generated by the affine modulation module are used to modulate the diffuse backbone features produced by the diffuse generation backbone network; and its implicit representation network adopts a multilayer perceptron structure with a sinusoidal activation function.

[0029] In one or more embodiments, before the affine parameters are injected into the output layer convolution of the diffusion backbone network, pixel-wise affine modulation is performed on the feature map through a forward hooking mechanism, and its calculation form is as follows:

[0030] ;

[0031] in The feature mapping for the input of the convolutional layer, This is the modulated feature map.

[0032] In one or more embodiments, in step one, all CT images are scaled to a resolution of 256×256 and linearly normalized to the [-1,1] interval; the normalized paired CT images are divided into training set and test set according to a set ratio.

[0033] In one or more embodiments, in step two, the diffusion generation backbone network learns the inverse diffusion mapping from low-dose to high-dose CT images by predicting noise components at each diffusion time step, and the training objective function is defined as:

[0034] ;

[0035] in, Represents high-dose CT images. To be in diffusion time step Noisy samples, The network's prediction results for the noise components. It follows a normal distribution with a mean of 0 and a covariance matrix of identity matrix I.

[0036] In one or more embodiments, in step four, a gradual unfreezing mechanism is set in the adaptive initial stage during testing, and the control strategy for freezing and unfreezing can be expressed as:

[0037] ;

[0038] in, This represents the convolution parameters of the output layer of the diffusion backbone. To optimize the current number of steps, This is the preset number of thawing steps.

[0039] In one or more embodiments, in step five, the blind spot occlusion processing adopts the Noise2Void mechanism and replaces the center pixel with randomly sampled circular neighborhood pixels.

[0040] In one or more embodiments, the adaptive joint optimization objective includes implicit pixel regression loss, latent variable regularization term, and self-consistency constraint loss, in the form of:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] in This represents the low-dose CT pixel prediction values ​​generated by the implicit network. To input the true value of the corresponding pixel in the low-dose image, This represents the pixel regression loss between the low-dose pixel predictions generated by the implicit network and the actual input; For latent variables, regularization terms; and These represent the noise components predicted by the model in the blind spot region for low-dose CT images with and without masking, respectively. The loss is due to the self-consistency constraint. These are the weighting coefficients for each loss term.

[0046] The technical solution provided by this invention brings at least the following beneficial effects:

[0047] This invention utilizes a diffusion-generated backbone network based on GDM as the basic denoising framework, and combines affine modulation parameters generated by an implicit representation network with an adaptive optimization mechanism for latent variables to achieve individualized reconstruction of low-dose CT images. Compared with traditional denoising methods that rely solely on fixed network parameters, this invention effectively reduces the model's sensitivity to changes in scanning conditions and noise differences by adaptively adjusting the affine modulation module and latent variables during the testing phase, significantly improving the robustness and detail fidelity of the reconstruction. At the same time, it introduces self-consistency constraints and a blind spot mechanism, enabling unsupervised optimization without the need for high-dose labels, reducing the dependence on paired training data. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0049] Figure 1 This is a diagram illustrating the overall structure of the self-consistency constraint-based adaptive low-dose CT image reconstruction method according to an embodiment of the present invention.

[0050] Figure 2 This is a flowchart of a test-time adaptive low-dose CT image reconstruction method with self-consistency constraints, according to an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0052] This invention provides a test-time adaptive low-dose CT image reconstruction method with self-consistency constraints. The method includes: during training, freezing the affine modulation module and the latent variable encoding module, and optimizing the parameters of the diffusion generation backbone network; freezing all parameters of the diffusion generation backbone network, and during testing, optimizing the parameters of the affine modulation module and the latent variables of the slice latent variable encoding module; when the optimization steps reach a preset unfreezing step count, unfreezing the output layer of the diffusion generation backbone network; obtaining a first noise through a baseline prediction path, obtaining a second noise through a consistency prediction path of the original low-dose CT image, and obtaining low-dose CT pixel prediction values ​​through the affine modulation module; and constructing an adaptive joint optimization objective based on the first noise, the second noise, the latent variables, the low-dose CT pixel prediction values, and the actual low-dose CT pixel values. The affine modulation module outputs affine parameters a and b, as well as the predicted pixel values ​​for low-dose CT images. The diffusion-generated backbone network outputs the denoising results. The affine modulation module calculates the predicted pixel values ​​for low-dose CT images simply to match the latent variable encoding of the slices. This allows the affine modulation module to output corresponding affine parameters based on the latent variable encoding of each slice, performing an affine transformation on the denoising results generated by the diffusion backbone network, thereby optimizing the quality of the denoising results. This invention effectively improves the structural fidelity and noise suppression performance of low-dose CT images.

[0053] like Figure 1 , Figure 2 As shown, this embodiment of the invention provides a test-time adaptive low-dose CT image reconstruction method with self-consistency constraints, including a diffusion-generated backbone network (…). Figure 1 The method includes the following steps: a diffusion network backbone, a slice latent variable coding module, and an affine modulation module based on an implicit representation network.

[0054] Step 1: Construct a paired dataset of low-dose and high-dose CT images, dividing the dataset into training and testing sets. The high-dose CT images are acquired using standard clinical scanning protocols; the low-dose CT images are simulated by applying a noise insertion algorithm to the high-dose projection data to mimic the image quality degradation effect after reducing the radiation dose.

[0055] As an optional implementation, in this step, all images are scaled to a resolution of 256×256 and linearly normalized to the [-1,1] interval, and the image dataset is divided into a training set and a test set in a 4:1 ratio. Furthermore, the image dataset in this embodiment is derived from a publicly available paired high-dose and low-dose CT image dataset from a clinic.

[0056] Step 2: Using low-dose CT images from the training set as input and high-dose CT images as reference, keep the affine modulation module and latent variable coding module frozen, and optimize only the diffusion backbone parameters.

[0057] In one or more embodiments, the diffusion generation backbone network used is GDM (Geodesic Diffusion Model). This network takes low-dose CT images as input and achieves the mapping from noisy images to high-quality images through multi-step geometric diffusion backpropagation. The mapping process can be expressed by the following formula:

[0058] (1);

[0059] in, Indicates diffusion time step The intermediate noise state, For the diffusion network in network parameters The following is a prediction of the noise component. It is a function of noise intensity. For time step interval, For constant terms related to the Fisher–Rao information geometry metric, and The calculation formula is as follows:

[0060] (2);

[0061] (3);

[0062] in, The maximum noise level set for the model. The minimum noise level set for the model.

[0063] As an optional implementation, maximum noise intensity The value is 80, the minimum noise intensity. The value is 0.002.

[0064] In one or more embodiments, the slice latent variable encoding module constructs a paired latent variable for each slice to characterize the overall noise characteristics and structural distribution of the slice during the training and testing phases. The latent variable encoding module is a set of learnable latent vectors, denoted as […].

[0065] (4);

[0066] in, Indicates the relationship with the first The latent variable vector corresponding to each CT slice Initialize the variance for the latent variables. As a potential spatial dimension, The number of CT slices in the dataset. Learnable latent vectors. It follows a normal distribution with a mean of 0 and a covariance matrix of σ²I. Initialize the variance for the latent variables, where I is the identity matrix. Represents the real number field.

[0067] As an optional implementation method, spatial dimension The variance is 128, initialized. It is 0.01.

[0068] In one or more embodiments, the affine modulation module of the implicit representation network takes slice latent variables and pixel position codes as input to simultaneously generate low-dose CT pixel prediction values ​​and corresponding pixel-by-pixel affine modulation parameters and modulates them on the diffusion backbone features.

[0069] As an optional implementation, the implicit representation network employs a multilayer perceptron structure, SIREN, with a sinusoidal activation function, and the input consists of a latent variable vector. With random Fourier feature location encoding Together, they form a whole, in which positional encoding and affine generation processes are defined as follows:

[0070] (5);

[0071] (6);

[0072] in, Indicates normalized to Pixel coordinates in the coordinate system To follow a Gaussian distribution The frequency matrix, Indicates the number of Fourier bases; This represents the predicted value of pixels in low-dose CT scans. These are pixel-level affine parameters; Functions corresponding to the multilayer perceptron structure

[0073] Before the affine parameters are injected into the output layer convolution of the diffusion backbone network, pixel-wise affine modulation is performed on the feature map through a forward hook mechanism. The calculation form is as follows:

[0074] (7);

[0075] in, The feature mapping for the input of the convolutional layer, This is the modulated feature map.

[0076] As an optional implementation, the above frequency matrix The elements follow a Gaussian distribution with variance The number of Fourier bases is 15. It is 64.

[0077] In this step, the diffusion generation backbone network learns the inverse diffusion mapping from low-dose to high-dose CT images by predicting noise components at each diffusion time step. The training objective function is defined as:

[0078] (8);

[0079] in, Represents high-dose CT images. To be in diffusion time step Noisy samples, The network's prediction results for the noise components. It follows a normal distribution with a mean of 0 and a covariance matrix of identity matrix I.

[0080] Step 3: Using low-dose CT images from the test set as input and high-dose CT images as reference, freeze all parameters of the backbone network generated by diffusion, and optimize the parameters of the affine modulation module and the latent variables of the slice latent variable encoding module.

[0081] Step 4: When the number of optimization steps reaches the preset number of unfreezing steps, the unfreezing diffusion generates the output layer of the backbone network, allowing it to participate in the optimization; when the number of optimization steps does not reach the preset number of unfreezing steps, proceed to Step 5.

[0082] In one or more embodiments, the control strategy for freezing and thawing can be expressed as:

[0083] (9);

[0084] in, This represents the convolution parameters of the output layer of the diffusion backbone. To optimize the current number of steps, This is the preset number of thawing steps.

[0085] As an optional implementation, this step presets the number of thawing steps. It is 15.

[0086] Step 5: Input the original low-dose CT image into the baseline prediction path including the optimized diffusion generation backbone network and the affine modulation module to obtain the first noise. Input the original low-dose CT image after blind spot occlusion processing into the consistency prediction path including the optimized diffusion generation backbone network and the affine modulation module to obtain the second noise. Input the original low-dose CT image into the optimized affine modulation module to obtain the low-dose CT pixel prediction value.

[0087] In one or more embodiments, the baseline prediction path is taken as input from the original low-dose CT image, and the consistency prediction path is taken as input from the masked low-dose CT image after blind spot occlusion processing; the blind spot region is constructed using the Noise2Void mechanism, and the center pixel is replaced by randomly sampled annular neighbor pixels to prevent information leakage.

[0088] Step Six: Based on the first noise, the second noise, the latent variable, the predicted value of low-dose CT pixels, and the actual value of low-dose CT pixels, construct an adaptive joint optimization objective. Perform adaptive iterative optimization on the adaptive joint optimization objective using multiple batches of test samples. When the maximum number of iterations is reached, output the corresponding reconstruction result and evaluate the performance of the reconstruction result using paired high-dose CT images from the test set. If the maximum number of iterations is not reached, return to Step Four.

[0089] In one or more embodiments, the adaptive joint optimization objective includes implicit pixel regression loss, latent variable regularization term, and self-consistency constraint loss, in the form of:

[0090] (10);

[0091] (11);

[0092] (12);

[0093] (13);

[0094] in, This represents the low-dose CT pixel predictions generated by the implicit network, i.e., the reconstructed LDCT image. To input the true value of the corresponding pixel in the low-dose image, This represents the pixel regression loss between the low-dose pixel predictions generated by the implicit network and the actual input; For latent variables, regularization terms; and These represent the noise components predicted by the model in the blind spot region for low-dose CT images with and without masking, respectively. The loss is due to the self-consistency constraint. These are the weighting coefficients for each loss term.

[0095] As an optional implementation, in this step, Set them to 0.01, 0.0001 and 1 respectively.

[0096] Reaching the maximum number of iterations The corresponding reconstruction result, i.e. the predicted clear CT image, is then output; and the performance of the reconstruction result is evaluated using paired high-dose CT images from the test set.

[0097] As an optional implementation, the batch size is 4, and the maximum number of iterations is [not specified]. It is 30.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0099] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A test-time adaptive low-dose CT image reconstruction method with self-consistency constraints, characterized in that, Based on a network structure comprising a diffusion-generated backbone network, a slice latent variable coding module, and an affine modulation module based on an implicit representation network, the following steps are performed: Step 1: Construct a dataset of CT images paired with low-dose and high-dose doses, and divide it into training and testing sets; Step 2: Using low-dose CT images from the training set as input and high-dose CT images as reference, keep the affine modulation module and latent variable coding module frozen, and only optimize the parameters of the diffusion generation backbone network; Step 3: Using low-dose CT images from the test set as input and high-dose CT images as reference, freeze all parameters of the backbone network generated by diffusion, and optimize the parameters of the affine modulation module and the latent variables of the slice latent variable encoding module. The input to the slice latent variable encoding module is a low-dose CT image. For each input low-dose CT image, a corresponding slice latent variable is constructed, and a latent variable vector is output. The affine modulation module takes the latent variable vector and the pixel position encoding of the input low-dose CT image as input to simultaneously generate low-dose CT pixel prediction values ​​and corresponding pixel-by-pixel affine modulation parameters. Step 4: When the number of optimization steps reaches the preset number of unfreezing steps, the unfreezing diffusion generates the output layer of the backbone network, allowing it to participate in the optimization; when the number of optimization steps has not reached the preset number of unfreezing steps, proceed to Step 5. Step 5: Input the low-dose CT images from the test set into the baseline prediction path to obtain the first noise; input the low-dose CT images from the test set after blind spot occlusion processing into the consistency prediction path to obtain the second noise; wherein, the baseline prediction path includes an optimized diffusion generation backbone network and an affine modulation module; the consistency prediction path includes an optimized diffusion generation backbone network and an affine modulation module. Low-dose CT images from the test set are input into an optimized affine modulation module to obtain low-dose CT pixel prediction values. Step 6: Based on the first noise, the second noise, the latent variable, the predicted value of low-dose CT pixels, and the actual value of low-dose CT pixels, construct an adaptive joint optimization objective, and perform adaptive iterative optimization on the adaptive joint optimization objective through multiple batches of test samples; When the maximum number of iterations is reached, the currently optimized diffusion generation backbone network, slice latent variable encoding module and affine modulation module are used for the target low-dose CT image to obtain the corresponding reconstruction results; If the maximum number of iterations has not been reached, return to step four.

2. The method according to claim 1, characterized in that, The diffusion generation backbone network is a GDM network. Through multi-step geometric diffusion backpropagation, it achieves a mapping from low-dose CT images to images whose sharpness deviates from that of high-dose CT images by no more than a preset value. The mapping process is expressed by the following formula: ； in, Indicates diffusion time step The intermediate noise state, For the diffusion network in network parameters The following is a prediction of the noise component. It is a function of noise intensity. For time step interval, This is a constant term related to the Fisher–Rao information geometry metric.

3. The method according to claim 1, characterized in that, The latent variable vector output by the slice latent variable encoding module is a learnable object that follows a normal distribution with a mean of 0 and a covariance matrix of σ²I. Initialize the variance for the latent variables, where I is the identity matrix.

4. The method according to claim 1, characterized in that, The pixel position encoding of the input affine modulation module is as follows: The image coordinates of the input low-dose CT image are subjected to a random Fourier transform to obtain the random Fourier feature location encoding: ;in, Indicates normalized to Pixel coordinates in the coordinate system To follow a Gaussian distribution The frequency matrix, This represents the number of Fourier bases.

5. The method according to claim 1 or 4, characterized in that, The pixel-wise affine modulation parameters generated by the affine modulation module are used to modulate the diffuse backbone features produced by the diffuse generation backbone network; and its implicit representation network adopts a multilayer perceptron structure with a sinusoidal activation function.

6. The method according to claim 5, characterized in that, Before the affine parameters are injected into the output layer convolution of the diffuse backbone network, pixel-wise affine modulation is performed on the feature map through a forward hook mechanism.

7. The method according to claim 1, characterized in that, In step two, the diffusion generation backbone network learns the inverse diffusion mapping from low-dose to high-dose CT images by predicting noise components at each diffusion time step.

8. The method according to claim 1, characterized in that, In step four, a gradual unfreezing mechanism is set up in the adaptive initial phase of testing. The control strategy for freezing and unfreezing can be expressed as follows: ； in, This represents the convolution parameters of the output layer of the diffusion backbone. To optimize the current number of steps, This is the preset number of thawing steps.

9. The method according to claim 1, characterized in that, In step five, the blind spot occlusion processing adopts the Noise2Void mechanism, and replaces the center pixel with randomly sampled circular neighboring pixels.

10. The method according to claim 1, characterized in that, The adaptive joint optimization objective includes implicit pixel regression loss, latent variable regularization term, and self-consistency constraint loss, and its form is as follows: ； ； ； ； in, This represents the low-dose CT pixel prediction values ​​generated by the implicit network. To input the true value of the corresponding pixel in the low-dose image, This represents the pixel regression loss between the low-dose pixel predictions generated by the implicit network and the actual input; For latent variables, regularization term, Initialize the variance for the latent variables. A vector of latent variables; and These represent the noise components predicted by the model in the blind spot region for low-dose CT images with and without masking, respectively. The loss is due to the self-consistency constraint. These are the weighting coefficients for each loss term.

Images