Cross-CT scanner general low-dose CT denoising method

Abstract

The invention discloses a cross-CT scanner general low-dose CT denoising method based on a uniformly distributed residual denoising diffusion model. The method comprises the following steps: in a training stage, firstly, constructing a multi-source heterogeneous data set; in the forward process, normal dose CT images from different sources are degraded into potential representation of shared distribution by introducing a residual diffusion mechanism and a scanner specific attenuation item; and then, under the condition of a low-dose CT image, a reverse process is executed through a residual prediction network, prediction residual in a degraded image is learned, and network parameters are optimized by minimizing residual difference. In the reasoning stage, a to-be-denoised low-dose CT image is used as a condition, random noise is added into the low-dose CT image to construct a degraded image, and a normal-dose CT image is obtained step by step through iterative residual prediction. According to the method, the model focuses on learning universal denoising mapping through the uniformly distributed degraded images, so that the generalization problem caused by noise mode difference between scanners is solved, and universal and efficient denoising across the scanners is realized.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model. Background Technology

[0002] Computed tomography (CT) has become an indispensable imaging tool in modern medicine, playing a crucial role in clinical scenarios such as tumor screening, cardiovascular disease diagnosis, and emergency trauma assessment. Given the potential high radiation dose risk associated with CT, there has been ongoing effort to develop low-dose CT (LDCT) imaging techniques to reduce radiation exposure while maintaining image quality. Despite significant progress in LDCT technology, its widespread adoption in multicenter clinical diagnosis and treatment faces a key challenge: significantly different scanner-specific noise patterns exist between CT scanners from different manufacturers and models. These differences introduce systemic bias, affecting diagnostic consistency and complicating cross-institutional image interpretation. Therefore, developing a robust cross-scanner LDCT imaging framework is crucial for ensuring diagnostic accuracy and promoting the wider integration of LDCT technology into clinical workflows.

[0003] In recent years, LDCT reconstruction algorithms based on deep learning (DL) have developed rapidly, gradually becoming a core direction of next-generation CT imaging technology. Based on their dependence on scanner data, existing methods can be divided into two main categories: scanner-specific depth reconstruction and scanner-independent depth reconstruction. Scanner-specific depth reconstruction methods are designed and trained for specific CT scanners by fusing explicit physical forward models of a particular CT system and its inherent noise characteristics. Examples of scanner-specific depth reconstruction methods allow for deep optimization of the scanner-specific reconstruction workflow by constructing different domain network architectures (including projection domain, image domain, and dual-domain networks). Specifically, projection domain networks directly process the raw projection data to suppress electronic and quantum noise captured by the detector. Image domain networks work on the reconstructed CT image, improving image quality by reducing noise and noise-induced streak artifacts while preserving key anatomical structures. Dual-domain networks construct an end-to-end framework, achieving superior denoising performance by synergistically utilizing complementary information from the projection and image domains. Scanner-specific depth reconstruction methods exhibit excellent reconstruction quality on the corresponding scanners by learning unique noise features related to the hardware and reconstruction algorithm. However, due to differences in noise characteristics among different scanners, the generalization ability of such models is limited, and performance degradation may occur on unknown devices. Furthermore, developing separate models for each scanner significantly increases time, manpower, and computational costs.

[0004] Therefore, in order to address the shortcomings of existing technologies, it is essential to provide a universal low-dose CT denoising method based on a uniformly distributed residual denoising diffusion model across CT scanners. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model. This universal low-dose CT denoising method based on a uniformly distributed residual denoising diffusion model can achieve universal and efficient denoising across different devices.

[0006] The above-mentioned objectives of the present invention are achieved through the following technical measures: This paper presents a universal low-dose CT denoising method based on a uniformly distributed residual denoising diffusion model across CT scanners. The training phase consists of the following steps: S1. Normal-dose spiral projection data collected from different CT scanners are used to reconstruct normal-dose CT images. Then, noise was added to the normal-dose spiral projection data to reconstruct paired low-dose CT images. Through normal dose CT images and paired low-dose CT images Heterogeneous datasets consisting of multiple sources; S2, the paired low-dose CT images from S1 Normal dose CT images after subtracting S1 Obtain the residual image Then, a CT scanner-specific attenuation term is introduced. ,exist Within the range, randomly sample a time step. t Defined as the current step, normal dose CT images The corresponding residual image and attenuation term The image is input into the residual denoising diffusion model for a forward process, resulting in the degraded image of the current step. ; S3, Pair the low-dose CT images from S1 And the degraded image of S2 The residual prediction network, which is input into the residual denoising diffusion model, performs direct reconstruction from the current step to the time step. t= The prediction result is obtained by reversing step 1; S4. Construct the loss function and calculate the residual image obtained in S3 and the residual image obtained in S2. The difference between them is used to optimize the parameters of the residual prediction network using gradient descent. Then, it is determined whether the training has reached the end condition. If it has, the residual diffusion model of the residual prediction network containing the current parameters is used as the post-training residual denoising diffusion model. If it has not been reached, return to S2.

[0007] During the inference phase, the low-dose CT image to be denoised is input as a condition into the trained residual denoising diffusion model, and random noise of a specific intensity is added to the low-dose CT image to construct the initial degraded image; subsequently, the process is executed from time step... t = T proceeds sequentially to time step t= The process of step 1 is reversed to gradually reconstruct the image through iterative residual prediction; finally, the time step is subtracted from the low-dose CT image. t= The prediction result of 1 yields the reconstructed image, which is the predicted normal dose CT image.

[0008] Preferably, the above S1 is performed by the following steps: S1.1 Normal dose spiral projection data obtained from clinical CT scanners manufactured by Anke and Neusoft are used to reconstruct normal dose CT images. ; S1.2. Quantum noise and electronic noise are inserted into the normal dose spiral projection data of S1.1 to reconstruct the paired low-dose CT image. ; S1.3, from normal dose CT images and paired low-dose CT images This constitutes a multi-source heterogeneous dataset and pairs low-dose CT images. and normal dose CT images Normalization is performed.

[0009] In S2, the normal dose CT image This is the diffusion starting point of the forward process of the residual denoising diffusion model.

[0010] In S2, the forward process of the residual denoising diffusion model involves introducing residual diffusion and noise diffusion, and introducing a CT scanner-specific attenuation term. Low-dose CT images from different CT scanners Mapped to a unified distribution space.

[0011] Preferably, the forward diffusion process of the above residual denoising diffusion model at each time step is represented by equation (1): ...Equation (1); in, and These are all independent coefficients, controlling the intensity of residual diffusion and noise diffusion respectively. The attenuation coefficient used to control the intensity of specific information in a CT scanner; t For time step and t The range is ; For time steps t Degraded images, To be based on time steps t-1 Degraded image residual image and low-dose CT images Under the condition that the current time step t Degraded image The conditional probability distribution it follows.

[0012] Preferably, the above-mentioned degraded image Equation (2) represents: ...Equation (2); in, Let be the cumulative sum of the residual images added from step 1 to step t, and , The standard deviation of the total noise added from step 1 to step t, and , Let be the cumulative sum of the decay terms subtracted from step 1 to step t, and , For random standard Gaussian noise and , The covariance matrix is ​​the identity matrix.

[0013] When the time step is T Time-degraded images Equation (3) represents: ... Equation (3).

[0014] During the inference phase, each time step is sampled according to equation (4): ...Equation (4); in, k The sampling interval is... For the predicted results, Reconstruct the image for the sampling interval.

[0015] In the inference phase, the predicted normal dose CT image is based on time steps. t= Prediction results at 1 o'clock The calculation yields the result, expressed by equation (5): ...Equation (5); in, To predict normal dose CT images.

[0016] Preferably, the above loss function is expressed by equation (6): ...Equation (6); in, This is the expected value.

[0017] This invention discloses a universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model. The training phase consists of the following steps: S1, collecting normal-dose spiral projection data from different CT scanners and reconstructing normal-dose CT images. Then, noise was added to the normal-dose spiral projection data to reconstruct paired low-dose CT images. Through normal dose CT images and paired low-dose CT images Constructing a heterogeneous dataset from multiple sources; S2, pairing low-dose CT images from S1. Normal dose CT images after subtracting S1 Obtain the residual image Then, a CT scanner-specific attenuation term is introduced. ,exist Within the range, randomly sample a time step. t Defined as the current step, normal dose CT images The corresponding residual image and attenuation term The image is input into the residual denoising diffusion model for a forward process, resulting in the degraded image of the current step. S3, Pair the low-dose CT images from S1 And the degraded image of S2 The residual prediction network, which is input into the residual denoising diffusion model, performs direct reconstruction from the current step to the time step. t= The reverse process of step 1 yields the prediction result; step S4 involves constructing the loss function and calculating the residual image obtained in step S3 and the residual image obtained in step S2. The differences between the parameters are analyzed, and the parameters of the residual prediction network are optimized using gradient descent. Then, it is determined whether the training has reached the end-of-training condition. If it does, the residual diffusion model of the residual prediction network containing the current parameters is used as the post-trained residual denoising diffusion model; otherwise, the process returns to S2. During the inference phase, the low-dose CT image to be denoised is input as a condition into the post-trained residual denoising diffusion model, and random noise of a specific intensity is added to the low-dose CT image to construct the initial degraded image. Subsequently, the process is executed from time step... t= T proceeds sequentially to time step t=The process of step 1 is reversed to gradually reconstruct the image through iterative residual prediction; finally, the time step is subtracted from the low-dose CT image. t= The prediction result of step 1 yields the reconstructed image, which is the predicted normal-dose CT image. This invention introduces degradation information from low-dose CT images during diffusion by constructing a residual denoising diffusion model based on a unified distribution. Utilizing a unified distribution diffusion mechanism, it gradually transforms the data distribution from different scanners into a unified distribution space by integrating scanner-specific attenuation terms. Furthermore, in the reverse process, this invention uses low-dose CT images as explicit conditions to reconstruct high-quality images faithful to the original scanner characteristics from a shared distribution, ultimately achieving universal and efficient denoising across devices. This provides a new technical path for multi-center clinical diagnosis and treatment, possessing broad application prospects and promotional value. Attached Figure Description

[0018] The invention will be further described with reference to the accompanying drawings, but the contents of the drawings do not constitute any limitation on the invention.

[0019] Figure 1 This is a flowchart illustrating a universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model.

[0020] Figure 2 These are CT images before and after denoising from simulation data from multiple scanners with known low dose levels, as presented in this embodiment of the invention.

[0021] Figure 3 These are CT images before and after denoising from simulation data from multiple scanners with unknown low dose levels, as presented in this embodiment of the invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described in conjunction with the following embodiments.

[0023] Example 1

[0024] A universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model, such as... Figure 1 The training phase consists of the following steps: S1. Normal-dose spiral projection data collected from different CT scanners are used to reconstruct normal-dose CT images. Then, noise was added to the normal-dose spiral projection data to reconstruct paired low-dose CT images. Through normal dose CT images and paired low-dose CT images Heterogeneous datasets consisting of multiple sources; S2, the paired low-dose CT images from S1 Normal dose CT images after subtracting S1 Obtain the residual image Then, a CT scanner-specific attenuation term is introduced. ,exist Within the range, randomly sample a time step. t Defined as the current step, normal dose CT images The corresponding residual image and attenuation term The image is input into the residual denoising diffusion model for a forward process, resulting in the degraded image of the current step. ; S3, Pair the low-dose CT images from S1 And the degraded image of S2 The residual prediction network, which is input into the residual denoising diffusion model, performs direct reconstruction from the current step to the time step. t= The prediction result is obtained by reversing step 1; S4. Construct the loss function and calculate the residual image obtained in S3 and the residual image obtained in S2. The difference between them is used to optimize the parameters of the residual prediction network using gradient descent. Then, it is determined whether the training has reached the end condition. If it has, the residual diffusion model of the residual prediction network containing the current parameters is used as the post-training residual denoising diffusion model. If it has not reached the end condition, return to S2.

[0025] The present invention provides a universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model. During the inference phase, the low-dose CT image to be denoised is input as a condition into the trained residual denoising diffusion model, and random noise of a specific intensity is added to the low-dose CT image to construct an initial degraded image; subsequently, the method executes from time step... t= T proceeds sequentially to time step t= The process of step 1 is reversed to gradually reconstruct the image through iterative residual prediction; finally, the time step is subtracted from the low-dose CT image. t= The prediction result of 1 yields the reconstructed image, which is the predicted normal dose CT image.

[0026] It should be noted that the heterogeneous dataset in S1 of this invention refers to the original normal-dose CT images. Data from different CT scanners has different distributions, hence the term heterogeneous data.

[0027] S2 is the degraded image obtained by forward processing the heterogeneous data collected in S1, thus transforming the originally heterogeneous data containing multiple noise patterns and inconsistent distributions into degraded images in a uniform distribution space. .

[0028] It should also be noted that, during the training phase, the data in each batch of this invention is randomly sampled from all available CT scanner types. The heterogeneous dataset in this embodiment contains CT data from four scanners. During training, each iteration extracts a certain amount of data from the dataset for training (also called a batch). The extracted data should simultaneously include data from all four scanners to ensure that each training batch represents the overall data distribution. The training termination condition of this invention can be reaching the maximum number of training iterations or loss function convergence. The specific termination condition, whether it's the maximum number of training iterations or loss function convergence, depends on the actual situation.

[0029] Specifically, S1 is performed through the following steps: S1.1 Normal dose spiral projection data obtained from clinical CT scanners manufactured by Anke and Neusoft are used to reconstruct normal dose CT images. ; S1.2. Quantum noise and electronic noise are inserted into the normal dose spiral projection data of S1.1 to reconstruct the paired low-dose CT image. ; S1.3, from normal dose CT images and paired low-dose CT images This constitutes a multi-source heterogeneous dataset and pairs low-dose CT images. and normal dose CT images Normalization is performed.

[0030] In S2, the normal dose CT image This is the diffusion starting point of the forward process of the residual denoising diffusion model.

[0031] In S2, the forward process of the residual denoising diffusion model involves introducing residual diffusion and noise diffusion, and introducing a CT scanner-specific attenuation term. Low-dose CT images from different CT scanners Mapped to a unified distribution space.

[0032] Therefore, this invention designs a forward diffusion process, introducing a specific attenuation term for CT scanners to transform low-dose CT data from different CT scanners into a unified distribution space. The forward diffusion process of the residual denoising diffusion model at each time step is represented by equation (1): ...Equation (1); in, and All are independent coefficients, which respectively control the intensity of residual and noise propagation; The attenuation coefficient used to control the intensity of specific information in a CT scanner; tFor time step and t The range is ; For time steps t Degraded images, To be based on time steps t-1 Degraded image residual image and low-dose CT images Under the condition that the current time step t Degraded image The conditional probability distribution it follows.

[0033] By attenuating CT scanner-specific features during forward diffusion, it prompts data from different CT scanners to map to a shared underlying distribution, specifically degraded images. Equation (2) represents: ...Equation (2); in, Let be the cumulative sum of the residual images added from step 1 to step t, and , The standard deviation of the total noise added from step 1 to step t, and , Let be the cumulative sum of the decay terms subtracted from step 1 to step t, and , For random standard Gaussian noise and , The covariance matrix is ​​the identity matrix.

[0034] It should be noted that equation (2) is a closed-form solution to equation (1), and the time step can be calculated. t Degraded image status, in which For random standard Gaussian noise and , The covariance matrix is ​​the identity matrix, meaning it follows a standard Gaussian distribution with a mean of 0 and a covariance matrix of identity.

[0035] Where, when the time step is T , The increase from 0 to 0.9 gradually diminishes the CT scanner-specific information at the diffusion endpoint when the time step is... T Time-degraded images Equation (3) represents: ... Equation (3).

[0036] It should be noted that the diffusion endpoint is approximately a non-pure Gaussian distribution, allowing data from different CT scanners to share a uniform distribution space.

[0037] The reverse process in the inference phase is the inverse process of diffusion, which involves sampling in a uniform distribution space that is independent of the CT scanner.

[0038] In existing technologies, a pre-trained residual prediction network is used to estimate the residual during the reverse process. The predicted residual image is used to obtain the predicted normal dose CT image. and noisy images The DDIM sampling strategy is employed to accelerate the sampling process, which can be divided into multiple steps: in, k This is the sampling interval. In each step, the network... and Predicting input and The images are gradually refined into clean, normal-dose CT images.

[0039] Because in practice, The value is almost zero and has no impact on performance, so this item is omitted in this invention.

[0040] During the inference phase, each time step is sampled according to equation (4): ...Equation (4); in, k The sampling interval is... For the predicted results, Reconstruct the image for the sampling interval.

[0041] This invention introduces an attenuation term. and cascaded low-dose CT images As a strongly conditional guiding term, it realizes the mapping from a uniform distribution space to a CT scanner-specific data space. This iterative denoising of the uniform distribution space achieves robust cross-CT scanner generalization and effective noise suppression.

[0042] In the inference phase, the predicted normal dose CT image is based on time steps. t= Prediction results at 1 o'clock The calculation yields the result, expressed by equation (5): ...Equation (5); in, To predict normal dose CT images.

[0043] loss function Equation (6) represents: ...Equation (6); in, This represents the expected value. This corresponds to the training process in S2 and S3 above, where a time step is randomly selected during training. t Randomly select normal-dose CT images and corresponding low-dose CT images, and obtain residual images by the difference between the low-dose CT images and normal-dose CT images. During the S2 forward propagation process, the normal dose CT image is transformed into a degraded image through t-step degradation. The network in S3 predicts the residual image from the degraded image to the clean image. , with the true residual image Calculate the mean error. The model continuously trains to reduce this error until it eventually converges.

[0044] This universal low-dose CT denoising method based on a unified distribution residual denoising diffusion model incorporates degradation information from low-dose CT images during diffusion. It utilizes a unified distribution diffusion mechanism to integrate scanner-specific attenuation terms, gradually transforming data distributions from different scanners into a unified distribution space. Furthermore, in the reverse process, the invention uses low-dose CT images as explicit conditions to reconstruct high-quality images faithful to the original scanner characteristics from a shared distribution, ultimately achieving universal and efficient denoising across devices. This provides a new technical approach for multi-center clinical diagnosis and treatment, demonstrating broad application prospects and promotional value.

[0045] Example 2 Example 1 describes the application of a universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model. This example utilizes multi-source datasets obtained from clinical CT scanners manufactured by Anke and Neusoft, including normal-dose spiral projection data from two key anatomical regions: the head and abdomen.

[0046] To obtain LDCT data at different dose levels, quantum noise and electronic noise were inserted into the normal dose projection data to simulate head data at 1 / 2, 1 / 4, 1 / 6 and 1 / 8 dose levels, as well as abdominal data at 1 / 4, 1 / 6, 1 / 8 and 1 / 10 dose levels.

[0047] Considering the differences in data distribution among different manufacturers and different anatomical sites, the data are divided into four scanner types: Anke abdominal data is scanner 1, Neusoft abdominal data is scanner 2, Anke head data is scanner 3, and Neusoft head data is scanner 4.

[0048] In the experiments, the training dataset consisted of 88,920 image pairs, including 1 / 4 and 1 / 8 dose data from scanners 1 and 2, and 1 / 2 and 1 / 6 dose data from scanners 3 and 4. Test dataset A contained 8,506 image pairs with dose levels consistent with those in the training dataset. Test dataset B consisted of 9,306 image pairs, including 1 / 6 and 1 / 10 dose data from scanners 1 and 2, and 1 / 4 and 1 / 18 dose data from scanners 3 and 4, used to validate the model's generalization performance at unseen dose levels.

[0049] During training, the Adam optimizer is used with an initial learning rate of The batch size was 10. Training was performed for 300,000 iterations, with randomly cropped 256×256 image patches as input. To mitigate the imbalance of sample sizes from different scanners in the multi-source dataset, different weights were assigned based on the size of each subset: Scanner 1 was assigned 0.2, Scanner 2 0.5, Scanner 3 0.1, and Scanner 4 0.2. During inference, all tests were performed on 512×512 images, and a consistent generation strategy with 3 time steps was applied to the data from different scanners.

[0050] This embodiment is illustrated by... Figure 2 The denoising results of different methods on test dataset A are presented. Test dataset A contains LDCT images from four scanners with dose levels consistent with the training dataset. It can be observed that LDCT images from different scanners and anatomical structures exhibit varying noise intensities and artifact distributions, reflecting inherent differences between scanners. REDCNN effectively suppresses noise artifacts in abdominal images from scanners 1 and 2, but significant noise remains in head images from scanners 3 and 4. From the region of interest (ROI) images and residual images, it can be seen that images generated using LIT-Former result in blurred tissue structures, uneven denoising, and loss of structural details. The denoised DU-GAN and MTD-GAN images show high-contrast structural tissues, but struggle to suppress noise-induced structural artifacts, with significant noise distribution still present in head images. When applied to data from different scanners, CoreDiff exhibits a tendency for over-smoothing, thus reducing the fidelity of key anatomical structures. In contrast, our proposed method performs well in both noise suppression and structural contrast enhancement, demonstrating robust performance generalization across different scanners.

[0051] This embodiment is illustrated by... Figure 3The denoising results of different methods on test dataset B containing unknown dose levels are presented. For scanner 2, 1 / 10 of the test dose data is outside the training range, resulting in residual stripe artifacts in the results of the comparative methods. On scanner 3, the results from REDCNN and LIT-Former exhibit oversmoothing and blurring features, while the outputs of DU-GAN and MTD-GAN retain a significant amount of fine-grained noise. The CoreDiff method is prone to structural artifacts. Our proposed method achieves an excellent balance between effective noise suppression and preservation of key texture details across different scanners and dose levels.

[0052] Finally, 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 the scope of protection of the present invention. 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 essence and scope of the technical solutions of the present invention.

Claims

1. A universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model, characterized in that, The training phase consists of the following steps: S1. Normal-dose spiral projection data collected from different CT scanners are used to reconstruct normal-dose CT images. Then, noise was added to the normal-dose spiral projection data to reconstruct paired low-dose CT images. Through normal dose CT images and paired low-dose CT images Heterogeneous datasets consisting of multiple sources; S2, the paired low-dose CT images from S1 Normal dose CT images after subtracting S1 Obtain the residual image ; Then, a CT scanner-specific attenuation term is introduced. ,exist Within the range, randomly sample a time step. t Defined as the current step, normal dose CT images The corresponding residual image and attenuation term The image is input into the residual denoising diffusion model for a forward process, resulting in the degraded image of the current step. ; S3, Pair the low-dose CT images from S1 And the degraded image of S2 The residual prediction network, which is input into the residual denoising diffusion model, performs direct reconstruction from the current step to the time step. t= The prediction result is obtained by reversing step 1; S4. Construct the loss function and calculate the residual image obtained in S3 and the residual image obtained in S2. The difference between them is used to optimize the parameters of the residual prediction network using gradient descent. Then it is determined whether the training has reached the end condition. If it has, the residual denoising diffusion model of the residual prediction network containing the current parameters is used as the post-trained residual denoising diffusion model. If it has not reached the end condition, return to S2. During the inference phase, the low-dose CT image to be denoised is input as a condition into the trained residual denoising diffusion model, and random noise is added to the low-dose CT image to construct the initial degraded image. Then execute from time step t= T proceeds sequentially to time step t= The process of step 1 is reversed to gradually reconstruct the image through iterative residual prediction; finally, the time step is subtracted from the low-dose CT image. t= The prediction result of 1 yields the reconstructed image, which is the predicted normal dose CT image.

2. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 1, characterized in that, S1 is specifically performed by the following steps: S1.1 Normal dose spiral projection data obtained from clinical CT scanners manufactured by Anke and Neusoft are used to reconstruct normal dose CT images. ; S1.

2. Quantum noise and electronic noise are inserted into the normal dose spiral projection data of S1.1 to reconstruct the paired low-dose CT image. ; S1.3, from normal dose CT images and paired low-dose CT images This constitutes a multi-source heterogeneous dataset and pairs low-dose CT images. and normal dose CT images Normalization is performed.

3. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 2, characterized in that: In S2, the normal dose CT image This is the diffusion starting point of the forward process of the residual denoising diffusion model.

4. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 3, characterized in that: In S2, the forward process of the residual denoising diffusion model involves introducing residual diffusion and noise diffusion, and introducing a CT scanner-specific attenuation term. Low-dose CT images from different CT scanners Mapped to a unified distribution space.

5. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 4, characterized in that: The forward diffusion process of the residual denoising diffusion model at each time step is represented by equation (1): ...Equation (1); in, and These are all independent coefficients, controlling the intensity of residual diffusion and noise diffusion respectively. The attenuation coefficient used to control the intensity of specific information in a CT scanner; t For time step and t The range is ; For time step t Degraded images, To be based on time steps t-1 Degraded image residual image and low-dose CT images Under the condition that the current time step t Degraded image The conditional probability distribution it follows.

6. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 5, characterized in that: The degraded image Equation (2) represents: ...Equation (2); in, Let be the cumulative sum of the residual images added from step 1 to step t, and , The standard deviation of the total noise added from step 1 to step t, and , Let be the cumulative sum of the decay terms subtracted from step 1 to step t, and , For random standard Gaussian noise and , The covariance matrix is ​​the identity matrix.

7. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 6, characterized in that: When the time step is T Time-degraded images Equation (3) represents: ... Equation (3).

8. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 7, characterized in that: During the inference phase, sampling is performed at each time step according to equation (4): ...Equation (4); in, k The sampling interval is... For the predicted results, Reconstruct the image for the sampling interval.

9. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 8, characterized in that: In the inference phase, the predicted normal dose CT image is based on time steps. t= Prediction results at 1 o'clock The calculation yields the result, expressed by equation (5): ...Equation (5); in, To predict normal dose CT images.

10. The universal low-dose CT denoising method across CT scanners based on a uniformly distributed residual denoising diffusion model according to claim 9, characterized in that: The loss function is expressed by equation (6): ...Equation (6); in, This is the expected value.

Images