Low-dose 3D spiral CT reconstruction method and system based on context cold diffusion model

Abstract

The invention discloses a low-dose 3D spiral CT reconstruction method and system based on a context cold diffusion model, and the method comprises the following steps: converting a linear inverse problem of CT imaging into an optimization problem, and adding a regularization term; based on a near-end gradient descent framework, solving the optimization problem by adopting an iterative shrinkage threshold algorithm, replacing a near-end operator in the near-end gradient descent method with a pre-trained two-dimensional cold diffusion model, and replacing a regularization item with the pre-trained cold diffusion model; when the cold diffusion model is pre-trained, three-dimensional context information formed by a current slice and two adjacent slices is used as the input of the cold diffusion model; taking a low-dose two-dimensional CT image as input, and performing low-dose 3D spiral CT image reconstruction by adopting the trained network; by combining the physical imaging model and the deep learning prior, the reconstruction quality and efficiency of the low-dose CT are remarkably improved, the noise and artifacts in the low-dose CT image are remarkably reduced, and the quality of the low-dose CT image is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical imaging, and particularly relates to a low-dose 3D helical CT reconstruction method and system based on a context cold diffusion model. Background Art

[0002] Computed Tomography (CT), as a core imaging technology for clinical diagnosis, the balance between its radiation dose control and image quality has always been a key issue in the field of medical imaging. Low-Dose CT (LDCT) technology reduces the patient's radiation exposure by reducing the X-ray tube current or voltage. However, problems such as increased quantum noise, projection data sparsification, and incomplete sampling caused thereby lead to streak artifacts, noise interference, and detail blurring in the reconstructed image, seriously affecting the accuracy of clinical diagnosis.

[0003] The current mainstream Filtered Back Projection (FBP) algorithm is still widely used in clinical CT systems due to its computational efficiency advantage. However, the FBP algorithm is based on the Radon inverse transform theory and is extremely sensitive to noise and undersampled data under low-dose conditions. The Signal-to-Noise Ratio (SNR) of its reconstructed image deteriorates significantly. Especially in the helical CT scanning mode, the stepped artifacts caused by the lack of continuity in the z-axis direction are particularly prominent. Therefore, iterative reconstruction algorithms (such as the ASIR and AIDR series) achieve a balance between noise suppression and detail preservation by introducing a statistical noise model and regularization constraints and performing multiple rounds of optimization in the projection domain or image domain. However, such methods have two inherent defects: firstly, the empirical setting of the regularization parameter and the number of iterations in the three-dimensional iterative process results in insufficient stability of the reconstruction result; secondly, the computational complexity increases cubically with the scale of three-dimensional data, making it difficult to meet the clinical real-time requirement.

[0004] With the development of deep learning, model-based CT reconstruction methods have gradually emerged, especially in two-dimensional geometric structures. These methods embed traditional iterative optimization processes into deep networks, such as ISTA-Net, 3pADMM, and NGIM-IRL. Although these techniques perform well under simpler geometric structures, their application in actual helical geometries is still limited. Deep learning-based reconstruction methods have made significant progress in recent years by establishing a non-linear mapping relationship between low-dose and standard-dose CT images through the construction of end-to-end neural networks (such as U-Net, GAN, VAE, etc.). Although such methods perform excellently in two-dimensional image denoising tasks, they face the following technical bottlenecks in three-dimensional helical CT reconstruction: (1) relying on strictly paired low / standard-dose training datasets, while there are ethical and operational feasibility limitations in dual-dose scans of the same patient in clinical practice; (2) the network model is sensitive to scan protocol parameters (such as tube voltage, pitch, slice thickness), and the cross-device generalization ability is insufficient; (3) existing architectures mostly adopt two-dimensional slice-level processing, ignoring the spatial continuity characteristics of three-dimensional volume data, resulting in artifacts remaining in the z-axis direction and anatomical structure breaks.

[0005] Diffusion Models model the Markov chain of forward noise addition and reverse denoising, showing the ability to reconstruct details beyond traditional methods in image generation tasks. Cold Diffusion Models significantly improve the inference efficiency by reducing the number of sampling steps from thousands to hundreds through an improved noise scheduling strategy. However, there are three major limitations in the application of existing cold diffusion technologies in the field of CT reconstruction: (1) the reverse denoising process is still based on two-dimensional convolution operations and cannot effectively capture three-dimensional anisotropic features; (2) model training requires relying on complete three-dimensional annotation data, and the data acquisition cost is high; (3) the collaborative optimization of the iterative sampling mechanism and the iterative reconstruction framework is insufficient, resulting in repeated consumption of computing resources.

[0006] In summary, the existing technologies have not effectively solved the following core problems: (1) the contradiction between computational efficiency and reconstruction quality in three-dimensional helical CT reconstruction; (2) the lack of model generalization ability across devices and dose levels; (3) the collaborative optimization of maintaining the continuity of three-dimensional anatomical structures and suppressing artifacts. Therefore, there is an urgent need to develop an efficient and robust three-dimensional low-dose CT reconstruction method to achieve diagnostic-level image quality while ensuring clinical real-time performance. Summary of the Invention

[0007] To solve the problems existing in the prior art, the present invention provides a three-dimensional helical CT iterative reconstruction method based on a pre-trained cold diffusion model. A generalized CT imaging model is constructed using variational regularization techniques, and the proximal gradient descent method is used to solve the three-dimensional CT reconstruction problem. The proximal operator in the three-dimensional iterative algorithm is replaced with a pre-trained two-dimensional cold diffusion model, that is, a pre-trained context cold diffusion model, thereby effectively reducing the computational complexity and improving the reconstruction quality.

[0008] To achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a low-dose 3D helical CT reconstruction method based on a context cold diffusion model, including the following steps: Convert the linear inverse problem of CT imaging into an optimization problem and add a regularization term; Based on the proximal gradient descent framework, use the iterative shrinkage threshold algorithm to solve the optimization problem. Replace the proximal operator in the proximal gradient descent method with a pre-trained two-dimensional cold diffusion model, and replace the regularization term with a pre-trained cold diffusion model; when the cold diffusion model is pre-trained, the three-dimensional context information composed of the current slice and its two adjacent slices is used as the input of the cold diffusion model; Use the low-dose two-dimensional CT image as the input and adopt the trained network to reconstruct the low-dose 3D helical CT image.

[0009] Further, converting the linear inverse problem of CT imaging into an optimization problem and adding a regularization term specifically means:

[0010] Among them, x is the CT image to be reconstructed, y is the measurement data, is the regularization term based on prior knowledge, λ is the trade-off parameter, y = A x + b, A is the forward model of the Radon transform, and b is the noise term.

[0011] Further, the iterative shrinkage threshold algorithm is used to solve the optimization problem, and the iterative steps are:

[0012]

[0013] Among them, prox is the proximal operator.

[0014] Further, the cold diffusion model is trained using a U-Net architecture with three-channel input. In the forward diffusion process, noise is gradually added to the clean image to generate a noisy image ; During the inverse reconstruction process, the 2D cold diffusion model is optimized by minimizing the L2 distance between the restored output and the target image, gradually denoising using a pre-trained restoration network, and introducing adjacent slices ( , ) as context constraints.

[0015] Furthermore, during the inverse reconstruction process, it is optimized by minimizing the following loss function: .

[0016] Among them, represents the global minimization objective of the loss function during the entire training process, represents the image restoration network, which is responsible for restoring a clean image from a degraded image, is the context input image at the diffusion time step t, which is composed of the current slice and its adjacent slices before and after , , stacked in the channel dimension, is the clean image.

[0017] Furthermore, replacing the proximal operator in the proximal gradient descent method with a pre-trained 2D cold diffusion model includes: Replacing the traditional proximal operator with a pre-trained context cold diffusion model

[0018] During the iterative process, the pre-trained context cold diffusion model uses the context information of the image for denoising.

[0019] Furthermore, the proximal gradient descent algorithm and the pre-trained cold diffusion model network are jointly trained iteratively from end to end, and the data consistency module and the proximal operator module are updated alternately; The data consistency module updates the intermediate variable according to the projection data to ensure that the reconstructed image is consistent with the measured values; In the proximal operator module, by stacking the current slice with its upper and lower neighboring slices, a context input is formed, and the pre-trained context cold diffusion model is used to denoise the reconstructed image.

[0020] In a second aspect, the present invention provides a low-dose 3D helical CT reconstruction system based on a context cold diffusion model, including a problem construction module, a solution module, and an image generation module; The problem construction module is used to transform the linear inverse problem of CT imaging into an optimization problem and add a regularization term; The solution module is based on the proximal gradient descent framework and uses the iterative shrinkage threshold algorithm to solve the optimization problem. It replaces the proximal operator in the proximal gradient descent method with a pre-trained two-dimensional cold diffusion model and replaces the regularization term with a pre-trained cold diffusion model. When the cold diffusion model is pre-trained, the three-dimensional context information composed of the current slice and its two adjacent slices is used as the input of the cold diffusion model. The image generation module is used to take the low-dose two-dimensional CT image as the input and reconstruct the low-dose 3D helical CT image using the trained network.

[0021] In a third aspect, the present invention can also provide a computer device, including a processor and a memory. Among them, the memory is used to store computer-executable programs, and the processor reads and executes the computer-executable programs from the memory. When the processor executes the computer-executable programs, it can implement the low-dose 3D helical CT reconstruction method based on the context cold diffusion model of the present invention.

[0022] At the same time, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the low-dose 3D helical CT reconstruction method based on the context cold diffusion model of the present invention can be implemented.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects: 1) Improve the quality of low-dose CT images: By combining the physical imaging model and the deep learning prior, the reconstruction quality and efficiency of low-dose CT are significantly improved. Experiments show that this method can meet the clinical diagnosis requirements within 5 iterations, with the PSNR increased by about 10 dB, and it adapts to different dose levels and scanner configurations. The optimized cold diffusion-ISTA algorithm can significantly reduce the noise and artifacts in low-dose CT images. By effectively denoising and restoring details, it significantly improves the quality of low-dose CT images, ensuring that while reducing the radiation dose, the image clarity is maintained, especially the details in key areas such as tumors and blood vessels, which helps to improve the accuracy of early disease diagnosis.

[0024] 2) Improve the computational efficiency and generalization ability of the reconstruction algorithm: The optimized iterative reconstruction algorithm significantly shortens the image reconstruction time, meets the needs of clinical real-time diagnosis, and improves the efficiency and practicality of medical image processing. The proposed algorithm has strong generalization ability and can run stably under different clinical scenarios and devices. It is widely applicable to various three-dimensional helical CT image data, enhancing its application and promotion potential in actual medical treatment.

[0025] 3) Practical clinical application value: Especially in cancer screening and early diagnosis of cardiovascular and cerebrovascular diseases, the optimized images will provide more accurate imaging support for doctors, helping them make more accurate diagnoses. By deploying and testing in the actual clinical environment, the practicality and stability of the algorithm are verified, technical support is provided for different hospitals and devices, the wide application of low-dose CT in clinical practice is promoted, and the overall level of medical services is improved.

[0026] In summary, it is expected that the present invention can design an efficient, accurate and well-generalized three-dimensional helical CT low-dose image reconstruction algorithm based on the combination of the cold diffusion model and the ISTA algorithm, significantly improving the application effect of low-dose CT in clinical practice, reducing the risk of patient radiation exposure, and promoting the development of medical imaging technology. Brief Description of the Drawings

[0027] Figure 1 It is a schematic diagram of the context cold diffusion model and the iterative reconstruction framework: (a) Forward diffusion and backward reconstruction process, showing the adjacent slice stacking operation; (b) Iterative framework structure, including a data consistency module (gradient update) and a proximal operator module (cold diffusion model).

[0028] Figure 2 It is a comparison chart of the reconstruction results for different numbers of iterations (1, 2, 5 times): The noise of the low-dose FBP reconstruction image is significant, while the noise of the present invention is significantly reduced after 1 iteration, and the details are clear after 5 iterations, approaching the normal dose level. Detailed Embodiments

[0029] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the 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.

[0030] In the description of the present invention, it should be understood that the terms "including" and "comprising" indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0031] Example 1, referring to Figure 1 , the present invention provides a low-dose 3D helical CT reconstruction method based on a pre-trained context cold diffusion model, including the following steps: 1) Modeling of the inverse problem of CT imaging: The linear inverse problem of CT imaging can be represented by the following equation: y =A x +b where \(A\) is the forward model of the Radon transform, x is the CT image to be reconstructed, y is the measurement data, and \(b\) is the noise term. Further, the present invention transforms the linear inverse problem of CT imaging into an optimization problem and introduces prior knowledge by adding a regularization term:

[0032] where is the regularization term based on prior knowledge, and \(\lambda\) is the trade-off parameter.

[0033] 2) Construct the proximal gradient descent framework: The iterative shrinkage threshold algorithm (ISTA) is used to solve the optimization problem, and the iterative steps are as follows:

[0034]

[0035] where \(prox\) is the proximal operator. Traditional methods need to manually design the regularization term (such as TV sparsity), while the present invention replaces the regularization term with a pre-trained cold diffusion model and implicitly learns prior knowledge by using its powerful image generation ability.

[0036] 3) Construct the context cold diffusion model: The cold diffusion model is pre-trained using a U-Net architecture with three-channel input. The U-Net input consists of the current slice and its two adjacent slices to form three-dimensional context information. In the forward diffusion process, noise is gradually added to the clean image to generate the noisy image :

[0037] where is the noise scheduling parameter, which controls the amount of noise added at each time step. In the reverse reconstruction process, the model is optimized by minimizing the L2 distance between the output of the recovery network and the target image, and the pre-trained cold diffusion model is used to gradually denoise, optimizing the following loss function:

[0038] In the reverse reconstruction process, the pre-trained recovery network is used to gradually denoise. To enhance three-dimensional continuity, adjacent slices ( , ) are introduced as context constraints. The input feature is the context slice with channel stacking:

[0039] The network is optimized by minimizing the following loss function:

[0040] Among them, represents the global minimization objective of the loss function throughout the training process, denotes the image restoration network, which is responsible for restoring clean images from degraded images, is the context input image at the diffusion time step t, which is composed of the current slice and its adjacent front and rear slices , , stacked in the channel dimension, is the clean image.

[0041] 4) Transfer the pre-trained 2D diffusion model to the 3D CT reconstruction framework. Traditional diffusion models are usually trained and inferred based on 2D images. However, since CT images are inherently 3D, directly applying 2D models to 3D image reconstruction will face many challenges. To solve this problem, the present invention proposes an innovative transfer path to effectively apply the pre-trained 2D diffusion model to 3D CT reconstruction.

[0042] The pre-trained diffusion model for 2D CT images restores images by gradually denoising over multiple time steps and uses context information, that is, each image slice combines the data of adjacent front and rear slices during the restoration process. The pre-trained diffusion model for 2D CT images can learn the structural information in the images, thereby effectively removing noise and maintaining details.

[0043] To apply the 2D model to 3D CT reconstruction, the present invention forms a 3D context by combining three consecutive 2D slices and inputs it into the pre-trained context cold diffusion model. The expression is as follows:

[0044] By stacking the current slice with its upper and lower neighborhood slices to form a context input, in this way, the pre-trained context cold diffusion model can handle the spatial continuity in 3D images, especially the anatomical structures along the z-axis. This strategy enables the 2D pre-trained model to maintain the structural coherence of 3D images by introducing context information and solves the problem that the 3D anatomical continuity cannot be effectively captured when processing 2D images alone.

[0045] Among them, the pre-trained context cold diffusion model is composed of the following functional modules for denoising and reconstructing 3D images of low-dose CT (LDCT). Specifically, it includes: Feature extraction layer, the front end composed of a convolutional neural network, is used to extract multi-scale deep features from the input CT images. For three-dimensional data, the pre-trained context cold diffusion model takes each axial slice as the basic processing unit, and takes it together with adjacent slices as input channels to capture local texture and structural information while integrating cross-slice anatomical correlation features.

[0046] Context fusion mechanism, which is used to fuse the context information of adjacent slices into the feature representation of the current slice to enhance the model's perception of the continuity of three-dimensional anatomical structures. In implementation, the pre-trained context cold diffusion model concatenates the target slice with its upper and lower adjacent slices in the channel dimension to form a multi-channel input containing context. This context fusion strategy ensures that adjacent slices remain unchanged throughout the denoising process, thus constraining the continuity of the output structure in the Z-axis direction during network inference; Differential propagation module, a unique error modulation unit and iterative propagation mechanism in the pre-trained context cold diffusion model. At each time step of diffusion sampling, the model calculates the difference (error) between the current reconstruction and the input, and injects this difference into the feature representation of the model through the Error-Modulated Module (EMM). EMM belongs to a per-feature-channel linear modulation unit. It uses a shallow network to extract the error features between the current prediction result and the original input, generates a set of modulation coefficients, and is used to correct the embedding features of the next time step. Through the propagation and modulation of this differential information, the model can correct the sampling bias in successive iterations and reduce the error accumulation caused by multiple iterations. This module essentially ensures that at each iteration, the internal state of the network is aligned with the input and the reconstruction result of the previous stage.

[0047] Reconstruction output module, the back end composed of an upsampling layer and a convolutional layer, is used to reconstruct the representation that has fused multi-scale features and context information into a denoised CT image. Usually, a U-Net-style encoding-decoding structure is adopted to gradually restore the spatial resolution in the decoding stage. The final output module generates a reconstruction result that matches the normal-dose CT (NDCT), and applies the denoising increment to the input LDCT through a residual connection or an addition operation, thereby obtaining a high-quality CT reconstruction image.

[0048] The network architecture of the pre-trained context cold diffusion model is designed with full consideration of the connection compatibility with the two-dimensional pre-trained model. On the one hand, the main architecture of CLEAR-Net is based on a two-dimensional convolutional neural network (i.e., U-Net), so the pre-trained weights of existing two-dimensional image denoising or diffusion models can be directly used for initialization. Specifically, by simply expanding the input layer of the pre-trained model from a single channel to a three-channel one (for receiving the target slice and its context slices), it can be evolved into the basic backbone network of the pre-trained context cold diffusion model, and other convolutional layers and parameters can follow the pre-trained parameters of the two-dimensional model. This connection method enables the model to draw on the feature extraction ability obtained from training with two-dimensional large-scale data, accelerating convergence and improving the generalization performance under limited medical image data.

[0049] On the other hand, due to the introduction of three-dimensional context fusion, the pre-trained context cold diffusion model shows obvious advantages in three-dimensional reconstruction tasks compared with the pure two-dimensional model. The context fusion mechanism ensures the structural consistency across slices, reducing the tomographic artifacts and discontinuity problems caused by layer-by-layer independent denoising. Compared with directly using a three-dimensional convolutional network, the method of the pre-trained context cold diffusion model effectively utilizes the information of adjacent slices while maintaining computational efficiency, improving the signal-to-noise ratio and detail fidelity of the reconstructed image.

[0050] 5) Fusion of Diffusion Model and Traditional Optimization Algorithm A key innovation of the present invention lies in the fusion of the diffusion model and the traditional optimization algorithm ISTA; the Iterative Shrinkage Thresholding Algorithm (ISTA) is often used to introduce prior knowledge and perform regularization in CT image reconstruction. However, these algorithms usually rely on manually designed regularization terms, such as Total Variation (TV) regularization, to suppress noise and preserve the details of the image. In contrast, the diffusion model can adaptively learn the prior of the image and restore the image quality through a step-by-step denoising process. The following is the specific content of the fusion of the diffusion model and the traditional optimization algorithm: The goal of the ISTA algorithm is to solve the optimization problem through proximal gradient descent. In CT reconstruction, the traditional ISTA algorithm combines the projection data with the regularization term and iteratively solves the image to be reconstructed:

[0051] where, is the image at the k-th iteration, is the step size, is the gradient of the objective function, and is the proximal operator, usually related to the sparsity or smoothness of the image.

[0052] The pre-trained context cold diffusion model replaces the traditional regularization term: In the present invention, the pre-trained context cold diffusion model replaces the traditional proximal operator , the pre-trained context cold diffusion model can more effectively restore the details and structure of the image by learning the context information of the image and adaptively denoising. During the iteration process, the pre-trained context cold diffusion model uses the context information of the image for denoising instead of relying on manually designed regularization terms. The reconstruction process can not only better suppress noise but also retain more details and structure, especially in low-dose CT images.

[0053] The joint proximal gradient descent algorithm and the pre-trained cold diffusion model network are used for end-to-end iterative training: the iterative reconstruction framework includes an alternating update mechanism of a data consistency module and a proximal operator module. In the data consistency module, the intermediate variable is updated according to the projection data , ensuring that the reconstructed image is consistent with the measured values, and the formula is:

[0054] In the proximal operator module, by stacking the current slice with its upper and lower neighboring slices, a context input is formed, and the pre-trained cold diffusion model is used to denoise the reconstructed image. The specific expression is:

[0055] Then, through the cold diffusion model restore a high-quality image :

[0056] The method of end-to-end iterative training by combining the joint proximal gradient descent algorithm and the pre-trained cold diffusion model network incorporates an alternating update mechanism of a data consistency module and a proximal operator module. In the data consistency module, each update of the image must ensure consistency with the measured projection data; while in the proximal operator module, the pre-trained context cold diffusion model is used to gradually denoise and restore the image. Apply the trained network to reconstruct low-dose 3D spiral CT images: the input is low-dose two-dimensional CT images, without three-dimensional paired data, and the spiral CT images after three-dimensional reconstruction can be obtained, and it can adapt to different doses and scanner configurations.

[0057] In the numerical experiment, the 2016 NIH-AAPM-Mayo Clinic low-dose CT challenge dataset was used for the experiment. This dataset contains the low-dose CT and normal-dose CT projection data of 10 patients. The detector parameters were set as 736×64 channels, pitch 0.8, source-to-isocenter distance 595mm, and the size of the original projection data was [48590,64,736]. In the data preprocessing stage, the projection data was first rearranged into a format conforming to the fan-beam geometry through a helical trajectory, and then the FBP method was used to preliminarily reconstruct the low-dose data to obtain 2D slices of 512×512. Finally, the pixel values were scaled to the range of 0-1 through normalization processing, and a random flipping operation was introduced to enhance data diversity.

[0058] In the training process of the cold diffusion model, the Adam optimizer was used, and the initial learning rate was set to 1×10 -4 and a weight decay of 1×10 -5 was imposed. The noise scheduling parameter was gradually increased from 1×10 -4 to 0.02 through a linear strategy, and a total of 150,000 iterations of training were completed. The model was optimized by minimizing the L2 distance between the output of the recovery network and the target image, and its loss function was defined as:

[0059] In the data consistency module, the projection space constraint was realized through the forward model A of the Radon transform and its transpose operation A^T. The proximal operator module stacked the current slice with its upper and lower neighboring slices to form a three-dimensional input, and the image was restored through the pre-trained cold diffusion model. Referring to Figure 2 , the entire process started from the FBP initial reconstruction of the low-dose CT, and the reconstruction quality was gradually improved after 5 iterations. Table 1 is a quantitative comparison table: the comparison of the execution time of the FBP method and our method in 1 iteration.

[0060] Table 1

[0061] Through the three-dimensional helical CT iterative reconstruction method based on the pre-trained cold diffusion model, by combining the physical imaging model with the deep learning prior, the reconstruction quality and efficiency of the low-dose CT have been significantly improved. Experiments show that this method can meet the clinical diagnosis requirements within 5 iterations, the PSNR is improved by about 10dB, and it adapts to different dose levels and scanner configurations. In the future, the real-time performance of the cold diffusion model will be further optimized to promote its application in mobile CT devices.

[0062] Example 2, the present invention provides a low-dose 3D helical CT reconstruction system based on a context cold diffusion model, including a problem construction module, a solution module, and an image generation module; The problem construction module is used to transform the linear inverse problem of CT imaging into an optimization problem and add a regularization term. The solution module is based on the proximal gradient descent framework and uses the iterative shrinkage threshold algorithm to solve the optimization problem. It replaces the proximal operator in the proximal gradient descent method with a pre-trained two-dimensional cold diffusion model and replaces the regularization term with a pre-trained cold diffusion model. When the cold diffusion model is pre-trained, three-dimensional context information composed of the current slice and its two adjacent slices is used as the input of the cold diffusion model. The image generation module is used to take the low-dose two-dimensional CT image as the input and perform low-dose 3D spiral CT image reconstruction using the trained network.

[0063] In addition, the present invention can also provide a computer device, including a processor and a memory. The memory is used to store computer-executable programs. The processor reads part or all of the computer-executable programs from the memory and executes them. When the processor executes part or all of the computer-executable programs, it can implement the low-dose 3D spiral CT reconstruction method based on the context cold diffusion model of the present invention.

[0064] The present invention also provides a computer device, which includes a processor and a memory. The memory is used to store computer-executable programs. The processor reads and executes the computer-executable programs from the memory. When the processor executes the computer-executable programs, it can implement the low-dose 3D spiral CT reconstruction method based on the context cold diffusion model of the present invention.

[0065] The computer device can be a laptop computer, a desktop computer or a workstation.

[0066] The processor described in the present invention can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0067] For the memory described in the present invention, it can be an internal storage unit of a laptop computer, a desktop computer or a workstation, such as a memory or a hard disk; it can also use an external storage unit, such as a mobile hard disk or a flash card.

[0068] A computer-readable storage medium may include a computer storage medium and a communication medium. The computer storage medium includes volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), solid state drives (SSD), or optical discs, etc. Among them, the random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0069] The above content is only for explaining the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention falls within the protection scope of the claims of the present invention.

Claims

1. A low-dose 3D helical CT reconstruction method based on a context cold diffusion model, characterized in that, Including the following steps: Transform the linear inverse problem of CT imaging into an optimization problem and add a regularization term; Based on the proximal gradient descent framework, use the iterative shrinkage threshold algorithm to solve the optimization problem, replace the proximal operator in the proximal gradient descent method with a pre-trained two-dimensional cold diffusion model, and replace the regularization term with a pre-trained cold diffusion model; when pre-training the cold diffusion model, the three-dimensional context information composed of the current slice and its two adjacent slices is used as the input of the cold diffusion model; Using the low-dose two-dimensional CT image as the input, use the trained network to reconstruct the low-dose 3D helical CT image.

2. The low-dose 3D helical CT reconstruction method based on the context cold diffusion model according to claim 1, wherein Transforming the linear inverse problem of CT imaging into an optimization problem and adding a regularization term specifically: Among them, x is the CT image to be reconstructed, and y is the measurement data. is the regularization term based on prior knowledge, and λ is the trade-off parameter. y = A x + b, where A is the forward model of the Radon transform and b is the noise term.

3. The low-dose 3D helical CT reconstruction method based on the context cold diffusion model according to claim 1, wherein, Using the iterative shrinkage threshold algorithm to solve the optimization problem, and the iterative steps are: wherein, prox is the proximal operator.

4. The low-dose 3D helical CT reconstruction method based on the context cold diffusion model according to claim 1, wherein The cold diffusion model is trained using a U-Net architecture with three-channel input. During the forward diffusion process, noise is gradually added to the clean image through the time step t to generate a noisy image ; In the reverse reconstruction process, the two-dimensional cold diffusion model is optimized by minimizing the L2 distance between the restored output and the target image, and the pre-trained restoration network is used to gradually denoise. Adjacent slices are introduced( , ) as context constraints.​ 5. The low-dose 3D helical CT reconstruction method based on the context cold diffusion model according to claim 4, wherein During the inverse reconstruction process, optimize by minimizing the following loss function: Among them, represents the global minimization objective of the loss function throughout the entire training process, denotes the image restoration network, which is responsible for restoring a clean image from a degraded image, is the context input image at the diffusion time step t, which is composed of the current slice and its adjacent slices before and after , , stacked in the channel dimension, is the clean image.

6. The low-dose 3D helical CT reconstruction method based on the context cold diffusion model according to claim 1, wherein Replacing the proximal operator in the proximal gradient descent method with a pre-trained two-dimensional cold diffusion model includes: Replace the proximal operator with a pre-trained context cold diffusion model ; During the iteration process, the pre-trained context cold diffusion model uses the context information of the image for denoising.

7. The low-dose 3D helical CT reconstruction method based on the context cold diffusion model according to claim 1, characterized in that, Jointly perform end-to-end iterative training on the proximal gradient descent algorithm and the pre-trained cold diffusion model network, and alternately update the data consistency module and the proximal operator module; The data consistency module updates intermediate variables according to the projection data , ensuring the consistency between the reconstructed image and the measured values; In the proximal operator module, by stacking the current slice with its upper and lower neighboring slices, form a context input, and use the pre-trained context cold diffusion model to denoise the reconstructed image.

8. A low-dose 3D helical CT reconstruction system based on a context cold diffusion model, characterized in that, Including a problem construction module, a solution module, and an image generation module; The problem construction module is used to transform the linear inverse problem of CT imaging into an optimization problem and add a regularization term; The solution module is based on the proximal gradient descent framework, uses the iterative shrinkage threshold algorithm to solve the optimization problem, replaces the proximal operator in the proximal gradient descent method with a pre-trained two-dimensional cold diffusion model, and replaces the regularization term with a pre-trained cold diffusion model; when pre-training the cold diffusion model, the three-dimensional context information composed of the current slice and its two adjacent slices is used as the input of the cold diffusion model; The image generation module is used to use the low-dose two-dimensional CT image as the input and use the trained network to reconstruct the low-dose 3D helical CT image.

9. A computer device, characterized in that, Including a processor and a memory, the memory is used to store computer-executable programs, the processor reads the computer-executable programs from the memory and executes them, and when the processor executes the computational executable programs, it can implement the low-dose 3D helical CT reconstruction method based on the context cold diffusion model described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the low-dose 3D helical CT reconstruction method based on the context cold diffusion model described in any one of claims 1 to 7.

Images