A CT Image Denoising Method, System, Electronic Device and Storage Medium

Abstract

The present invention discloses a CT image denoising method, system, electronic device and storage medium. The method includes: obtaining a normal-dose high-resolution CT image without labels, and reducing the image resolution by a preset multiple to obtain a normal-dose low-resolution CT image without labels; constructing a cascaded diffusion model by cascading an unconditional diffusion model and a conditional diffusion model; training the unconditional diffusion model with the normal-dose low-resolution CT image without labels, and using the trained model to generate a low-resolution normal-dose CT image from a random noise image; training the conditional diffusion model with the normal-dose high-resolution CT image without labels, and using the trained model to generate a high-resolution normal-dose CT image with the image generated by the unconditional diffusion model as a condition; combining the trained cascaded diffusion model with the maximum a posteriori framework to construct a denoising model, inputting a low-dose CT image, and iteratively solving multiple maximum a posteriori problems in combination with the prior provided by the cascaded diffusion model to achieve denoising of the low-dose CT image.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly relates to a CT image denoising method, system, electronic device and storage medium. Background Art

[0002] Pelvic segmentation has always been an essential step in the clinical diagnosis and surgical planning of pelvic bone diseases, and the processing of CT images is a fundamental work for this step. Low-dose computed tomography (CT) has received increasing attention due to its significant reduction in radiation dose. Although low-dose CT can reduce the risk of patients being exposed to X-rays, it usually results in poor-quality reconstructed images with obvious noise, thus hindering accurate diagnosis. Currently, various algorithms have been developed to improve the image quality of low-dose CT, including sinogram filtering, iterative reconstruction, and image denoising. Among them, low-dose CT image denoising is the most widely applied because it does not involve the reconstruction process and can be applicable to different CT imaging systems, facilitating doctors' subsequent surgical planning and postoperative effect evaluation.

[0003] The success of supervised deep learning algorithms depends to a large extent on a large amount of paired training data. However, due to adhering to the principle of "as low as reasonably achievable", it is challenging to obtain such data in a clinical environment. Therefore, it is crucial to develop methods that can utilize the high performance of deep neural networks without requiring a large amount of labeled data.

[0004] Some existing methods attempt to use different strategies to solve this problem. One class of methods relies on training data obtained through special acquisition schemes. Another class of methods further relaxes the conditions for training data, but still requires a large number of low-dose CT images for training.

[0005] By exploring existing data and basic common sense, the following problems exist in current low-dose CT image denoising:

[0006] 1. There is currently no dataset for the task of metal artifact removal, making it difficult to train conventional deep learning models.

[0007] 2. Supervised deep learning methods usually require paired low-dose and normal-dose CT images for training, and paired data is difficult to obtain in a clinical environment.

[0008] 3. In a patient's CT sequence images, when the depth exceeds 300, the starting and ending positions of artifacts are not determined, and there is a certain similarity between adjacent anatomical structures and noise, which may cause the generation model to incorrectly generate redundant structures.

[0009] 4. There are problems with traditional generation networks, such as poor generation effects in the case of insufficient labeled data and no gain for segmentation models. Summary of the Invention

[0010] In view of the deficiencies in the prior art, the present invention provides a CT image denoising method, system, electronic device, and storage medium. The denoising method only uses normal-dose CT images in a public dataset to train a cascaded diffusion model. After training, the cascaded diffusion model can gradually generate realistic low-resolution normal-dose CT images and high-resolution normal-dose CT images from random noise. By utilizing the prior of the diffusion model and combining it with the maximum a posteriori framework, zero-shot denoising of low-dose CT images is achieved.

[0011] The first object of the present invention is to provide a CT image denoising method, including:

[0012] Obtain unlabeled normal-dose high-resolution CT images from a public dataset, and reduce the resolution of the unlabeled normal-dose high-resolution CT images by a preset multiple to obtain unlabeled normal-dose low-resolution CT images;

[0013] Construct a cascaded diffusion model, which is composed of an unconditional diffusion model and a conditional diffusion model in cascade;

[0014] The unconditional diffusion model is trained with unlabeled normal-dose low-resolution CT images. After training, the unconditional diffusion model is used to generate low-resolution normal-dose CT images from random noise images;

[0015] The conditional diffusion model is trained with unlabeled normal-dose high-resolution CT images. After training, the conditional diffusion model is used to generate high-resolution normal-dose CT images with the low-resolution normal-dose CT images generated by the unconditional diffusion model as the condition;

[0016] Combine the trained cascaded diffusion model with the maximum a posteriori framework to construct a denoising model. Input the low-dose CT image into the denoising model. The denoising model combines the prior provided by the cascaded diffusion model and iteratively solves multiple maximum a posteriori problems to achieve denoising of the input low-dose CT image.

[0017] As a further improvement of the present invention, the training of the unconditional diffusion model includes:

[0018] Initialize the model parameters. Use unlabeled normal-dose low-resolution CT images as the input of the forward process of the model. Add Gaussian noise to the input image step by step at each time step to obtain a series of noise-added images until the input image is completely transformed into a random noise image;

[0019] The model reverse process starts from a random noise image, predicts the noise to be removed in the current image at each time step to gradually remove the noise from the noise image and recover the input image, and updates the model parameters;

[0020] The variational bound that minimizes the negative log-likelihood is used as the loss function. Through the optimization of the loss function, the model parameters are continuously adjusted. When the model performance no longer improves or reaches the preset number of iterations, the training stops to obtain the trained unconditional diffusion model.

[0021] As a further improvement of the present invention, the training of the conditional diffusion model includes:

[0022] Initialize the model parameters. Use unlabeled high-resolution CT images of normal doses as the input of the model forward process. Add Gaussian noise to the input image step by step at each time step to obtain a series of noise-added images until the input image is completely transformed into a random noise image;

[0023] The model reverse process starts from a random noise image, conditions on the low-resolution normal-dose CT image generated by the unconditional diffusion model, predicts the noise to be removed in the current image at each time step to gradually remove the noise from the noise image and recover the input image, and updates the model parameters;

[0024] The variational bound that minimizes the negative log-likelihood is used as the loss function. Through the optimization of the loss function, the model parameters are continuously adjusted. When the model performance no longer improves or reaches the preset number of iterations, the training stops to obtain the trained conditional diffusion model.

[0025] As a further improvement of the present invention, the reverse process of the trained conditional diffusion model is combined with the maximum a posteriori framework to construct a denoising model. The low-dose CT image is used as the likelihood and input into the reverse process of the trained conditional diffusion model. Combining the prior of the trained conditional diffusion model, iterative processing is performed to achieve denoising of the input low-dose CT image.

[0026] As a further improvement of the present invention, during the iterative denoising process, the denoised image obtained in each iteration is expressed as:

[0027] ;

[0028] where is the time step; is the noise image at time step ; is the hyperparameter at time step ; is the mean conditioned on ; is the time step The denoised image obtained by iteration at time step At time step is the noise level; is the condition of the conditional diffusion model, that is, the low-resolution normal-dose CT image generated by the unconditional diffusion model; At as the condition, the denoised image obtained by iteration of the conditional diffusion model at time step is obtained.

[0029] As a further improvement of the present invention, during the iterative denoising process, an adaptive strategy is used to adjust the hyperparameters to adapt to images with different noise levels;

[0030] The adaptive strategy is to use the standard deviation of all the noise in the image as the noise level to dynamically determine the hyperparameters;

[0031] The formula for adjusting the hyperparameters with the adaptive strategy is:

[0032] ;

[0033] where is the hyperparameter adjusted by the adaptive strategy, , are both manually selected parameters, is the th standard deviation of all pixels in the image obtained by the

[0034] As a further improvement of the present invention, the inverse process of the trained conditional diffusion model is accelerated by interval sampling to reduce the number of iterations.

[0035] The second object of the present invention is to provide a CT image denoising system for implementing the above denoising method, including:

[0036] An image acquisition module, which acquires unlabeled normal-dose high-resolution CT images from a public dataset and reduces the resolution of the unlabeled normal-dose high-resolution CT images by a preset multiple to obtain unlabeled normal-dose low-resolution CT images;

[0037] A cascaded diffusion model construction module for constructing a cascaded diffusion model, which is composed of an unconditional diffusion model and a conditional diffusion model in cascade;

[0038] The cascaded diffusion model training module is used to train the cascaded diffusion model, including an unconditional diffusion model training sub-module and a conditional diffusion model training sub-module; the unconditional diffusion model training sub-module trains the unconditional diffusion model with unlabeled normal-dose low-resolution CT images, and the trained unconditional diffusion model is used to generate low-resolution normal-dose CT images from random noise images; the conditional diffusion model training sub-module trains the conditional diffusion model with unlabeled normal-dose high-resolution CT images, and the trained conditional diffusion model is used to generate high-resolution normal-dose CT images with the low-resolution normal-dose CT images generated by the unconditional diffusion model as the condition;

[0039] The denoising model construction module is used to combine the trained cascaded diffusion model with the maximum a posteriori framework to construct a denoising model;

[0040] The image denoising module is used to input the low-dose CT image into the denoising model, and the denoising model combines the prior provided by the cascaded diffusion model to iteratively solve multiple maximum a posteriori problems to achieve denoising of the input low-dose CT image.

[0041] The third object of the present invention is to provide an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit executes the above-mentioned denoising method.

[0042] The fourth object of the present invention is to provide a storage medium, which stores a computer program executable by an electronic device, and when the program runs on the electronic device, the electronic device executes the above-mentioned denoising method.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] When training the diffusion model, paired normal-dose and low-dose CT images are not required, nor any low-dose CT images as the training set. Only the normal-dose CT images are used to train the cascaded diffusion model, and the trained cascaded diffusion model can gradually generate realistic low-resolution normal-dose CT images and high-resolution normal-dose CT images from random noise.

[0045] By combining the cascaded diffusion model with the maximum a posteriori framework, using the prior of the diffusion model, iteratively solving multiple maximum a posteriori problems, zero-shot denoising of low-dose CT images is achieved, thereby improving the quality of low-dose CT images.

[0046] The denoising method provided by the present application uses the cascaded diffusion model and the maximum a posteriori framework technology to train and generate artifact-free low-dose pelvic CT images without a control data set to improve the segmentation effect of intraoperative and postoperative 3D pelvic CT images. Description of the Drawings

[0047] Figure 1 It is a flowchart of the denoising algorithm provided by the present invention;

[0048] Figure 2 It is a schematic diagram of the reverse process of the unconditional diffusion model;

[0049] Figure 3 It is a schematic diagram of the reverse process of the conditional diffusion model;

[0050] Figure 4 It is a schematic diagram of the denoising process of the denoising model;

[0051] Figure 5 It is a schematic diagram of adjusting hyperparameters with an adaptive strategy;

[0052] Figure 6 It is a flowchart of the denoising system provided by the present invention. Detailed Description of the Invention

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] The present invention will be further described in detail below with reference to the accompanying drawings:

[0055] Please refer to Figure 1 , this embodiment provides a CT image denoising method, including:

[0056] Obtain unlabeled normal-dose high-resolution CT images from a public dataset, and reduce the resolution of the unlabeled normal-dose high-resolution CT images by a preset multiple to obtain unlabeled normal-dose low-resolution CT images;

[0057] Construct a cascaded diffusion model, where the cascaded diffusion model is composed of an unconditional diffusion model and a conditional diffusion model cascaded;

[0058] The unconditional diffusion model is trained with unlabeled normal-dose low-resolution CT images, and the trained unconditional diffusion model is used to generate low-resolution normal-dose CT images from random noise images;

[0059] The conditional diffusion model is trained with unlabeled normal-dose high-resolution CT images, and the trained conditional diffusion model is used to generate high-resolution normal-dose CT images conditioned on the low-resolution normal-dose CT images generated by the unconditional diffusion model.

[0060] The trained cascaded diffusion model is combined with the maximum a posteriori framework to construct a denoising model. The low-dose CT image is input into the denoising model, and the denoising model combines the prior provided by the cascaded diffusion model to iteratively solve multiple maximum a posteriori problems to achieve denoising of the input low-dose CT image.

[0061] Given a set of normal-dose high-resolution CT images , , reduce the resolution by times to obtain normal-dose low-resolution images , obtaining an image pair , where , and two models are defined based on this.

[0062] The first model is an unconditional diffusion model with forward process parameters and backward process parameters . This model is used to generate low-resolution normal-dose CT images from random noise.

[0063] The second diffusion model is a conditional diffusion model that achieves super-resolution of normal-dose CT images conditioned on the normal-dose low-resolution CT images generated by the unconditional diffusion model. Its forward process parameters are , and the backward process parameters are .

[0064] By cascading two diffusion models, first generate normal-dose low-resolution CT images from random noise, and then use the generated normal-dose low-resolution CT images as the condition in the conditional diffusion model to generate normal-dose high-resolution CT images.

[0065] When training the unconditional diffusion model:

[0066] Use as the input data of the model to learn the distribution characteristics and generation process of the data. Among them, is an unlabeled normal-dose low-resolution CT image, usually clean data without noise pollution, is the data distribution.

[0067] The forward process of the unconditional diffusion model: gradually add Gaussian noise to the input data to obtain a series of gradually noised data , until is completely transformed into random noise data .

[0068] During the forward process, at each time step (from 1 to ), the data for the current time step is generated based on the data from the previous time step . The forward process of the unconditional diffusion model is defined as a Markov chain (Equation 1), and the equation is:

[0069] ;

[0070] where is the noise-added data at time step , is the noise-added data at time step , is the predefined noise variance, is the identity matrix, is the Gaussian distribution.

[0071] The reverse process of the unconditional diffusion model: Starting from the random noise data , at each time step, the noise to be removed from the current image is predicted to gradually remove the noise from the noisy image, and finally the original input data is recovered .

[0072] The reverse process of the unconditional diffusion model is defined as a Markov chain (Equation 2) with learnable parameters (reverse process parameters). The reverse process is implemented through a series of conditional probability distributions , and the equation is:

[0073] ;

[0074] where is the mean, usually predicted by U-nets; is the noise level at time step , is the variance of the noise level at time step , is defined as , , respectively represent the forward process parameters at time steps , ; The training loss is obtained by minimizing the variational bound of the negative log-likelihood .

[0075] By optimizing the loss function, the parameters of the unconditional diffusion model are continuously adjusted. When the model performance no longer improves or reaches the preset number of iterations, the training is stopped to obtain the trained unconditional diffusion model for generating normal-dose low-resolution CT images from random noise images.

[0076] When training the conditional diffusion model:

[0077] Take as the input data of the model to learn the distribution characteristics and generation process of the data. Among them, is the unlabeled normal-dose high-resolution CT image (resolution 512*512), usually clean data without noise pollution, is the data distribution of.

[0078] The forward process of the conditional diffusion model: Gradually add Gaussian noise to the input data to obtain a series of gradually noised data , until is completely transformed into random noise data , which is the same as the forward process of the unconditional diffusion model.

[0079] The reverse process of the conditional diffusion model: Starting from the random noise data , taking the low-resolution normal-dose CT image generated by the unconditional diffusion model as the condition, predict the noise to be removed in the current image at each time step to gradually remove the noise of the noisy image, and finally recover the original input data .

[0080] The reverse process of the conditional diffusion model is defined as (Equation 3):

[0081] ;

[0082] Among them, is the condition of the conditional diffusion model, that is, the low-resolution normal-dose CT image generated by the unconditional diffusion model; the training loss is obtained by minimizing the variational bound of the negative log-likelihood .

[0083] By optimizing the loss function, the parameters of the conditional diffusion model are continuously adjusted. When the model performance no longer improves or reaches the preset number of iterations, the training is stopped to obtain the trained conditional diffusion model that generates normal-dose high-resolution CT images with the low-resolution normal-dose CT image generated by the unconditional diffusion model as the condition.

[0084] Two diffusion models are cascaded to generate normal-dose CT images. The first model is an unconditional diffusion model that generates normal-dose low-resolution CT images from random noise. The generation process is as Figure 2 shown; the second model is a conditional diffusion model that generates high-resolution normal-dose CT images of size 512×512 conditioned on the normal-dose low-resolution CT images generated by the unconditional diffusion model. The generation process is as Figure 3 shown.

[0085] When training the diffusion model, paired normal-dose and low-dose CT images are not required, nor are any low-dose CT images needed as the training set. Only normal-dose CT images are used to train the cascaded diffusion model. The trained cascaded diffusion model can gradually generate realistic low-resolution and high-resolution normal-dose CT images from random noise. The entire model is trained unsupervised using only unlabeled normal-dose CT images, effectively solving the problem in the prior art that paired low-dose CT images and normal-dose CT images are required for training, and it is difficult to obtain paired images.

[0086] After the cascaded diffusion model is trained, the reverse process of the trained conditional diffusion model is combined with the maximum a posteriori framework to construct a denoising model. The low-dose CT image is used as the likelihood input into the reverse process of the trained conditional diffusion model. Combining the prior of the trained conditional diffusion model, iterative processing is performed to achieve denoising of the input low-dose CT image. The denoising process is as Figure 4 shown.

[0087] In the image denoising problem, the noisy image is modeled as follows (Equation 4):

[0088] ;

[0089] where, is the input noisy image, and is the underlying clean image we want to obtain.

[0090] To achieve denoising of low-dose CT images, a well-trained diffusion model is used to capture the prior distribution of the underlying clean image. The diffusion model uses the forward process parameters and the reverse process parameters to accurately capture the prior distribution of . The goal is to effectively denoise using the diffusion prior.

[0091] The specific method is outlined as follows:

[0092] In the diffusion model, the image at the time step can be obtained from Generate, image The calculation formula (Formula 5) is as follows:

[0093] ;

[0094] Among them, is obtained from the normal distribution.

[0095] To introduce the likelihood of the input noisy image at each time step and generate the denoised image of the current time step through from the input noisy image , the calculation formula (Formula 6) of the image at the time step is as follows: The calculation formula (Formula 6) is as follows:

[0096] ;

[0097] Substitute Formula 3 and Formula 4 into Formula 5 and set For The following equation (Formula 7) is obtained:

[0098] ;

[0099] Formula 7 indicates that at each time step , there is a new denoising problem of noise .

[0100] Since decreases as increases, when , Therefore, we can use as the initialization; then for , we can incorporate the diffusion prior into the maximum a posteriori (MAP) framework to solve , and the calculation formula (Formula 8) is as follows:

[0101] ;

[0102] Among them, is the coefficient that balances the likelihood and the prior at the time step . In the inverse process of the diffusion model, follows the Gaussian distribution with the mean and the diagonal covariance matrix .

[0103] When , Formula 9 is obtained:

[0104] ;

[0105] When , it indicates that the final generation step of diffusion is deterministic, and formula 10 is obtained:

[0106] ;

[0107] Formula 9 is a convex optimization problem with a closed-form solution. By solving 9 and combining it with formula 10, we can derive the solution of (formula 11) at each iteration:

[0108] ;

[0109] Formula 11 illustrates the denoising process using a pre-trained unconditional diffusion model.

[0110] Similarly, if the pre-trained diffusion model is conditioned on , the solution of becomes (formula 12):

[0111] ;

[0112] where is the time step; is the noise image at time step ; is the hyperparameter at time step ; is the mean conditioned on ; is the denoised image obtained by iteration at time step ; is the noise level at time step ; is the condition of the conditional diffusion model, i.e., the low-resolution normal-dose CT image generated by the unconditional diffusion model; is the denoised image obtained by iteration of the conditional diffusion model conditioned on at time step .

[0113] Formula 12 illustrates the denoising process using a pre-trained conditional diffusion model, that is, in this embodiment, the normal-dose low-resolution CT image generated by the unconditional diffusion model is used as the condition, and the expression of the denoised image obtained at each iteration in the iterative denoising process.

[0114] The reverse process of the trained conditional diffusion model is combined with the maximum a posteriori framework, and the prior of the conditional diffusion model is utilized to iteratively solve multiple maximum a posteriori problems, achieving zero-shot denoising of low-dose CT images, thereby improving the quality of low-dose CT images.

[0115] During the iterative denoising process, an adaptive strategy is used to adjust the hyperparameters (i.e., the coefficients that balance likelihood and prior), in order to balance the likelihood and prior in the MAP estimation and adapt to images with different noise levels.

[0116] Considering that different regions of low-dose CT images exhibit different degrees of noise, the adaptive strategy estimates the standard deviation of all the noise in the image as the noise level to dynamically determine the hyperparameters;

[0117] The formula for adjusting the hyperparameters with the adaptive strategy is:

[0118] ;

[0119] where is the hyperparameter adjusted by the adaptive strategy, and are both manually selected parameters, is the standard deviation of all pixels in the image obtained at the -th iteration. The process of refining the hyperparameters is as shown in Figure 5 .

[0120] By using dynamic coefficients to resume the denoising process at intermediate time steps, the denoising model can adaptively process low-dose CT images with different noise levels.

[0121] During the denoising process, the reverse process of the trained conditional diffusion model is accelerated by interval sampling to reduce the number of iterations and speed up the denoising process of low-dose CT images.

[0122] Specifically, we extract a subsequence of length from , and then perform sampling from to . The denoised image is represented as:

[0123] ;

[0124] This non-uniform sampling method is used because the likelihood term plays a more prominent role in smaller time steps, and denser sampling can better utilize the input of low-dose CT images, thereby obtaining more accurate denoising results.

[0125] The denoising method provided in this embodiment can train and generate artifact-free low-dose pelvic CT images without a control data set (paired low-dose CT images and normal-dose CT images) by using a cascaded diffusion model and a maximum a posteriori framework technology, so as to improve the segmentation effect of 3D pelvic CT images during and after surgery.

[0126] Please refer to Figure 6 , this embodiment provides a CT image denoising system for implementing the above denoising method, including:

[0127] An image acquisition module that acquires unlabeled normal-dose high-resolution CT images from a public data set, and reduces the resolution of the unlabeled normal-dose high-resolution CT images by a preset multiple to obtain unlabeled normal-dose low-resolution CT images;

[0128] A cascaded diffusion model construction module for constructing a cascaded diffusion model, which is composed of an unconditional diffusion model and a conditional diffusion model in cascade;

[0129] A cascaded diffusion model training module for training the cascaded diffusion model, including an unconditional diffusion model training sub-module and a conditional diffusion model training sub-module; the unconditional diffusion model training sub-module trains the unconditional diffusion model with unlabeled normal-dose low-resolution CT images, and the trained unconditional diffusion model is used to generate low-resolution normal-dose CT images from random noise images; the conditional diffusion model training sub-module trains the conditional diffusion model with unlabeled normal-dose high-resolution CT images, and the trained conditional diffusion model is used to generate high-resolution normal-dose CT images with the low-resolution normal-dose CT images generated by the unconditional diffusion model as the condition;

[0130] A denoising model construction module for combining the trained cascaded diffusion model with the maximum a posteriori framework to construct a denoising model;

[0131] An image denoising module for inputting low-dose CT images into the denoising model, and the denoising model combines the prior provided by the cascaded diffusion model to iteratively solve multiple maximum a posteriori problems to achieve denoising of the input low-dose CT images.

[0132] The denoising system provided in this embodiment can train and generate artifact-free low-dose pelvic CT images without a control data set (paired low-dose CT images and normal-dose CT images) by using a cascaded diffusion model and a maximum a posteriori framework technology, so as to improve the segmentation effect of 3D pelvic CT images during and after surgery.

[0133] This embodiment provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit is caused to execute the above denoising method.

[0134] This embodiment provides a storage medium that stores a computer program executable by an electronic device. When the program runs on the electronic device, the electronic device is caused to execute the above denoising method.

[0135] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A CT image denoising method, characterized in that, Including: Obtain unlabeled normal-dose high-resolution CT images from a public dataset, and reduce the resolution of the unlabeled normal-dose high-resolution CT images by a preset multiple to obtain unlabeled normal-dose low-resolution CT images; Construct a cascaded diffusion model, which is composed of an unconditional diffusion model and a conditional diffusion model in cascade; The unconditional diffusion model is trained with unlabeled normal-dose low-resolution CT images, and the trained unconditional diffusion model is used to generate low-resolution normal-dose CT images from random noise images; The training of the unconditional diffusion model includes: initializing model parameters, using unlabeled normal-dose low-resolution CT images as the input of the forward process of the model, adding Gaussian noise to the input image step by step at each time step to obtain a series of noisy images until the input image is completely transformed into a random noise image; the reverse process of the model starts from the random noise image, predicts the noise to be removed in the current image at each time step to gradually remove the noise of the noisy image and restore the input image, and updates the model parameters; using the variational bound of minimizing the negative log-likelihood as the loss function, continuously adjusting the model parameters through the optimization of the loss function, and stopping training when the model performance no longer improves or reaches the preset number of iterations to obtain the trained unconditional diffusion model; The conditional diffusion model is trained with unlabeled normal-dose high-resolution CT images, and the trained conditional diffusion model is used to generate high-resolution normal-dose CT images conditional on the low-resolution normal-dose CT images generated by the unconditional diffusion model; The training of the conditional diffusion model includes: initializing model parameters, using unlabeled normal-dose high-resolution CT images as the input of the forward process of the model, adding Gaussian noise to the input image step by step at each time step to obtain a series of noisy images until the input image is completely transformed into a random noise image; the reverse process of the model starts from the random noise image, conditional on the low-resolution normal-dose CT images generated by the unconditional diffusion model, predicts the noise to be removed in the current image at each time step to gradually remove the noise of the noisy image and restore the input image, and updates the model parameters; using the variational bound of minimizing the negative log-likelihood as the loss function, continuously adjusting the model parameters through the optimization of the loss function, and stopping training when the model performance no longer improves or reaches the preset number of iterations to obtain the trained conditional diffusion model; Combine the trained cascaded diffusion model with the maximum a posteriori framework to construct a denoising model, input the low-dose CT image into the denoising model, and the denoising model combines the prior provided by the cascaded diffusion model to iteratively solve multiple maximum a posteriori problems to achieve denoising of the input low-dose CT image; Combine the reverse process of the trained conditional diffusion model with the maximum a posteriori framework to construct a denoising model, input the low-dose CT image as the likelihood into the reverse process of the trained conditional diffusion model, and combine the prior of the trained conditional diffusion model to continuously iterate to achieve denoising of the input low-dose CT image; During the iterative denoising process, an adaptive strategy is used to adjust hyperparameters to adapt to images with different noise levels; the adaptive strategy is to use the standard deviation of all the noise in the image as the noise level to dynamically determine the hyperparameters; the formula for adjusting the hyperparameters with the adaptive strategy is: ； Among them, is the hyperparameter adjusted by the adaptive strategy, , are all manually selected parameters, is the standard deviation of all pixels in the image obtained in the -th iteration. Accelerate the inverse process of the trained conditional diffusion model through non-uniform interval sampling to reduce the number of iterations.

2. The CT image denoising method according to claim 1, characterized in that During the iterative denoising process, the denoised image obtained in each iteration is represented as: ； Among them, is the time step; is the time step at the noisy image; is the time step at the hyperparameter; is with as the conditional mean; is the time step when the denoised image obtained by iteration; is at the time step when the noise level; is the condition of the conditional diffusion model, that is, the low-resolution normal-dose CT image generated by the unconditional diffusion model; is with as the condition of the conditional diffusion model at the time step when the denoised image obtained by iteration.

3. A CT image denoising system for implementing the denoising method according to any one of claims 1 to 2, characterized in that, Including: An image acquisition module that acquires unlabeled normal-dose high-resolution CT images from a public dataset and reduces the resolution of the unlabeled normal-dose high-resolution CT images by a preset multiple to obtain unlabeled normal-dose low-resolution CT images; A cascaded diffusion model construction module for constructing a cascaded diffusion model, which is composed of an unconditional diffusion model and a conditional diffusion model in cascade; A cascaded diffusion model training module for training the cascaded diffusion model, including an unconditional diffusion model training sub-module and a conditional diffusion model training sub-module; the unconditional diffusion model training sub-module trains the unconditional diffusion model with the unlabeled normal-dose low-resolution CT images, and the trained unconditional diffusion model is used to generate low-resolution normal-dose CT images from random noise images; the conditional diffusion model training sub-module trains the conditional diffusion model with the unlabeled normal-dose high-resolution CT images, and the trained conditional diffusion model is used to generate high-resolution normal-dose CT images conditional on the low-resolution normal-dose CT images generated by the unconditional diffusion model; A denoising model construction module for combining the trained cascaded diffusion model with the maximum a posteriori framework to construct a denoising model; An image denoising module for inputting the low-dose CT image into the denoising model, and the denoising model combines the prior provided by the cascaded diffusion model to iteratively solve multiple maximum a posteriori problems to achieve denoising of the input low-dose CT image.

4. An electronic device, characterized in that, Including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit executes the denoising method according to any one of claims 1 to 2.

5. A storage medium, characterized in that, It stores a computer program executable by an electronic device, and when the program runs on the electronic device, the electronic device executes the denoising method according to any one of claims 1 to 2.

Images