Three-dimensional sparse view medical image reconstruction method and system based on two-dimensional diffusion model, terminal and storage medium

Abstract

The invention discloses a three-dimensional sparse view medical image reconstruction method and system based on a two-dimensional diffusion model, a terminal and a storage medium, and the method comprises the steps: obtaining an original medical image under a sparse view, inputting the original medical image into the two-dimensional diffusion model, and obtaining an initial prediction image corresponding to the sparse view; curvature regularization calculation is carried out according to the initial prediction image, and a target curvature of the initial prediction image in a Z-axis direction phase is obtained; performing image iteration reconstruction processing on the target curvature by adopting an alternating direction multiplier method to obtain an initial medical reconstruction image; and performing re-iteration processing on the initial medical reconstruction image to obtain a target medical reconstruction image. According to the method, the diffusion model and the curvature prior technology are combined, the original medical image under the sparse view can be reconstructed, and the structural consistency of the image crossing multiple slices is kept, so that the reconstruction effect of the medical image is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of image reconstruction technology, and in particular to a three-dimensional sparse view medical image reconstruction method, system, terminal and computer-readable storage medium based on a two-dimensional diffusion model. Background Art

[0002] In the medical image reconstruction process of existing technologies, especially in computed tomography (CT) and magnetic resonance imaging (MRI), raw data (sinusoidal data in CT and k-space data in MRI, where k-space data refers to the raw data collected during magnetic resonance imaging, which contains all the spatial frequency information required to reconstruct the image) is a key factor in medical reconstruction. However, these raw data are usually incomplete, which leads to many challenges in the medical reconstruction process. As a generative model, the diffusion model has been successfully applied to the 2D reconstruction task of medical images, and can gradually transform noise into high-quality images that conform to the data distribution.

[0003] However, existing technologies have difficulty maintaining the structural consistency of images across multiple slices when extending the diffusion model to 3D reconstruction, resulting in poor medical image reconstruction effects and failure to meet user needs.

[0004] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0005] The main purpose of the present invention is to provide a three-dimensional sparse view medical image reconstruction method, system, terminal and computer-readable storage medium based on a two-dimensional diffusion model, aiming to solve the problem in the prior art that when the diffusion model is extended to 3D reconstruction, it is difficult to maintain the structural consistency of the image across multiple slices, resulting in poor medical image reconstruction effect and failure to meet user needs.

[0006] To achieve the above object, the present invention provides a three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model, and the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model comprises the following steps: Acquire an original medical image under a sparse view, and input the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view; Performing curvature regularization calculation according to the initial predicted image to obtain a target curvature of the initial predicted image in the phase direction of the Z axis; An alternating direction multiplier method is used to perform an iterative image reconstruction process on the target curvature to obtain an initial medical reconstructed image; The initial medical reconstructed image is iteratively processed to obtain a target medical reconstructed image.

[0007] Optionally, the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model, wherein the two-dimensional diffusion model includes a denoising model and a denoising model; The obtaining of an original medical image under a sparse view and inputting the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view specifically includes: Determining a sparse view to be reconstructed, and acquiring an original medical image corresponding to the sparse view; Inputting the original medical image into the denoising model in the two-dimensional diffusion model, performing forward denoising processing on the original medical image through the denoising model to obtain denoised Gaussian distribution data; Using a denoising score matching method to optimize and train the denoising model in the two-dimensional diffusion model to obtain an optimized and trained denoising model; The denoised Gaussian distribution data is input into an optimized and trained denoising model, and the denoised Gaussian distribution data is subjected to reverse denoising processing by optimizing the trained denoising model to obtain an initial predicted image corresponding to the sparse view.

[0008] Optionally, in the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model, the calculation expression of the forward denoising process is: ; Among them, the To add noise to Gaussian distribution data, is the time-dependent variance function, is the differential of the variance function, is the increment of Brownian motion, is the differential of time.

[0009] Optionally, in the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model, the calculation expression of the reverse denoising process is: ; Among them, the is the initial prediction image, To optimize the trained denoising model, is the image status, For time, This is the reverse standard Brownian motion.

[0010] Optionally, the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model, wherein the target curvature includes total square curvature, total regularized variation, total absolute curvature, mean square error, Geman regularization, logarithmic determinant regularization, Laplace distribution regularization and exponential energy regularization.

[0011] Optionally, the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model, wherein the alternating direction multiplier method is used to perform image iterative reconstruction processing on the target curvature to obtain an initial medical reconstructed image, specifically includes: Introducing an auxiliary variable, and calculating an augmented Lagrangian function corresponding to the auxiliary variable and the target curvature; The augmented Lagrangian function is iteratively updated and the image is iteratively reconstructed using an alternating direction multiplier method to obtain an initial medical reconstructed image.

[0012] Optionally, the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model, wherein the iterative processing of the initial medical reconstructed image to obtain the target medical reconstructed image specifically includes: Iteratively inputting the initial medical reconstruction image into the two-dimensional diffusion model, and performing the curvature regularization calculation and the image iterative reconstruction process; When the number of iterations reaches a preset iteration threshold, the iteration is stopped and the target medical reconstructed image is output.

[0013] In addition, to achieve the above-mentioned purpose, the present invention also provides a three-dimensional sparse view medical image reconstruction system based on a two-dimensional diffusion model, wherein the three-dimensional sparse view medical image reconstruction system based on a two-dimensional diffusion model comprises: An initial predicted image generation module, used for acquiring an original medical image under a sparse view, and inputting the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view; A curvature regularization calculation module, used to perform curvature regularization calculation according to the initial predicted image to obtain a target curvature of the initial predicted image in the phase in the Z-axis direction; An image iterative reconstruction processing module, used for performing image iterative reconstruction processing on the target curvature by using an alternating direction multiplier method to obtain an initial medical reconstructed image; The iterative reconstruction processing module is used to perform iterative processing on the initial medical reconstructed image to obtain a target medical reconstructed image.

[0014] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model stored in the memory and executable on the processor, wherein the three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model implements the steps of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model as described above when the three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model is executed by the processor.

[0015] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model, and when the three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model is executed by a processor, the steps of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model as described above are implemented.

[0016] In the present invention, an original medical image under a sparse view is obtained, and the original medical image is input into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view; curvature regularization calculation is performed based on the initial predicted image to obtain the target curvature of the initial predicted image in the phase in the Z-axis direction; the target curvature is iteratively reconstructed using an alternating direction multiplier method to obtain an initial medical reconstructed image; the initial medical reconstructed image is iteratively processed to obtain a target medical reconstructed image. The present invention realizes an effective combination of a two-dimensional diffusion model and curvature regularization technology by inputting the original medical image under a sparse view into a two-dimensional diffusion model and processing it using curvature regularization, thereby reconstructing the original medical image under the sparse view. There is no need to perform high-dimensional calculations of a 3D diffusion model, which greatly reduces the amount of calculation of three-dimensional data and effectively improves the image quality and reconstruction accuracy of the target medical reconstructed image. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of a preferred embodiment of the method for reconstructing a three-dimensional sparse view medical image based on a two-dimensional diffusion model of the present invention; Figure 2 It is a schematic diagram of a three-dimensional sparse view medical image reconstruction framework using a two-dimensional diffusion model with a curved prior in a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model of the present invention; Figure 3 It is a schematic diagram of the denoising and denoising process of the diffusion model of a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model of the present invention; Figure 4It is a schematic diagram of comparison of axial, sagittal and coronal slice image results of three-dimensional CT image reconstruction using 8 views on an LDCT data set according to a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model of the present invention; Figure 5 It is a schematic diagram of comparing the axial, sagittal and coronal slice image results of three-dimensional CT image reconstruction using four views on an LDCT data set according to a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model of the present invention; Figure 6 It is a schematic diagram of the comparison of axial, sagittal and coronal slice image results of three-dimensional CT image reconstruction using two views on an LDCT data set according to a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model of the present invention; Figure 7 It is a schematic diagram of the CS-MRI reconstruction result of the BraTS18 FLAIR mode of a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model of the present invention; Figure 8 It is a schematic diagram of a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model of the present invention, which uses eight views to perform quantitative comparison of three-dimensional CT image reconstruction results on an LDCT data set; Fig. 9 It is a schematic diagram of a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model of the present invention, which uses four views to perform quantitative comparison of three-dimensional CT image reconstruction results on an LDCT data set; Fig.10 It is a schematic diagram of a preferred embodiment of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model of the present invention, which uses two views to perform quantitative comparison of three-dimensional CT image reconstruction results on an LDCT data set; Fig.11 It is a structural diagram of a preferred embodiment of a three-dimensional sparse view medical image reconstruction system based on a two-dimensional diffusion model of the present invention; Fig.12 It is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] In the medical image reconstruction process of existing technologies, especially in computed tomography (CT) and magnetic resonance imaging (MRI), raw data (sinusoidal images in CT and k-space data in MRI) are key factors in medical reconstruction images. However, these raw data are usually incomplete, which leads to many challenges in the medical reconstruction process. In traditional methods, such as model-based iterative reconstruction (MBIR), data consistency is combined with prior knowledge, and the reconstruction process is optimized through regularization techniques (such as Tikhonov regularization, total variation, and low-rank methods). However, these methods often fail to effectively capture the spatial correlation in 3D images, especially in the case of sparse or limited viewing angles, and the reconstruction effect is poor.

[0020] As a generative model, the diffusion model has been successfully applied to 2D reconstruction tasks of medical images and can gradually transform noise into high-quality images that conform to the data distribution. However, extending the diffusion model to 3D reconstruction faces great challenges due to the high computational overhead and the difficulty in maintaining the structural consistency of the image across multiple slices. Although curvature regularization performs well in terms of edge and contrast preservation, combining it with the diffusion model for 3D medical image reconstruction remains an unsolved problem.

[0021] The disadvantages of existing technologies are as follows: 1. High computational complexity: Diffusion models require huge computational resources when extended to 3D, making real-time or clinical applications unfeasible. 2. Poor spatial consistency: Existing diffusion models fail to effectively utilize the spatial correlation along the z-axis in 3D reconstruction, resulting in inconsistent reconstruction between slices. 3. Inadequate performance under sparse viewing angle conditions: Current 3D reconstruction methods have difficulty recovering details and maintaining structural integrity when data is scarce or viewing angles are limited.

[0022] In view of the above problems, the present invention proposes an efficient three-dimensional medical image reconstruction method based on a pre-trained two-dimensional diffusion model and a priori curvature, aiming to solve the problems of high computational cost and poor reconstructed image effect in the prior art in three-dimensional image reconstruction under sparse views. The present invention combines the iterative reconstruction of the curvature regularized model with the denoising step of the diffusion model, and uses the alternating direction method of multipliers (ADMM) and the backtracking step rule to solve the reconstruction problem, wherein the curvature regularization captures the spatial correlation along the z-axis and enhances the consistency of the three-dimensional reconstruction. The GPU (Graphics Processing Unit) acceleration algorithm ensures efficient processing across slices while maintaining structural consistency. Experimental results show that under sparse views and limited angle CT conditions, the present invention can significantly improve detail retention and structural integrity. Compared with the baseline method, the PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) indicators are increased by 15% respectively. In MRI reconstruction, the present invention has achieved superior quantitative indicators on a variety of data sets, demonstrating its strong generalization ability. By effectively combining the diffusion model and the curvature prior, the framework of the present invention successfully addresses the key challenges in three-dimensional medical image reconstruction and provides high-fidelity reconstruction results, with potential clinical application prospects such as low-dose CT and fast MRI acquisition.

[0023] The three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model includes the following steps: Step S10, obtaining an original medical image under a sparse view, and inputting the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view, wherein the two-dimensional diffusion model includes a denoising model and a denoising model.

[0024] Most current 3D medical image reconstruction methods rely on either high-dimensional 3D diffusion models or traditional model-based iterative reconstruction (MBIR) methods. Diffusion models perform well in 2D medical image processing, but are often not efficiently applied in 3D image reconstruction due to high computational complexity. The present invention combines a 2D diffusion model with a curvature regularization prior for efficient 3D medical image reconstruction, especially for sparse-view and limited-angle CT and MRI reconstruction tasks.

[0025] The present invention is based on a pre-trained two-dimensional diffusion model and significantly reduces the computational complexity of 3D data processing by processing slices along the z-axis separately, while maintaining the structural consistency of the image.

[0026] Specifically, a sparse view to be reconstructed is determined, and an original medical image corresponding to the sparse view is obtained; the original medical image is input into the denoising model in the two-dimensional diffusion model, and the original medical image is forward denoised by the denoising model to obtain denoised Gaussian distribution data. The calculation expression of the forward denoising process is: ; wherein said To add noise to Gaussian distribution data, is the time-dependent variance function, is the differential of the variance function, is the increment of Brownian motion, is the differential of time.

[0027] The diffusion model learns the distribution of image data by gradually adding Gaussian noise in the forward process and denoising it through the inverse process. The residual noise of the data is estimated by training a neural network model, and the image is restored through the inverse process. Figure 2 As shown, the diffusion model is passed through a standard Gaussian noise image Generate the final desired result image.

[0028] like Figure 3 As shown in Figure 1, the basic idea of ​​the diffusion model is to gradually add noise to the data through a forward process, and then recover the data from the noise through a reverse process. In the forward process, Gaussian noise is gradually added to the data, which can be represented by a forward stochastic differential equation: ,in, is the original image data, The Gaussian noise added (i.e., noise-added Gaussian distribution data), wherein the added Gaussian noise (i.e., noise-added Gaussian distribution data) can also be expressed as: ,in, is the time-dependent variance function, is the differential of the variance function, is the increment of Brownian motion, describing the addition of noise. The forward process transforms the data distribution into a Gaussian distribution, which change.

[0029] The denoising model in the two-dimensional diffusion model is optimized and trained by using a denoising score matching method to obtain an optimized and trained denoising model; the denoised Gaussian distribution data is input into the optimized and trained denoising model, and the denoised Gaussian distribution data is reversely denoised by using the optimized and trained denoising model to obtain an initial predicted image corresponding to the sparse view. Figure 3 As shown, the reverse denoising process can be expressed by the reverse stochastic differential equation: , here It represents the noise-added Gaussian distribution data obtained by adding Gaussian noise data to the original image data. represents the denoising process of the denoised Gaussian distribution data, You can also use To express, equal , wherein the calculation expression of the reverse denoising process (i.e., reverse sampling process) is: ; wherein said is the initial prediction image, To optimize the trained denoising model, for The image status at the moment, is the time, which can effectively recover samples from noise. Specifically, Figure 2 The Unet network in the figure has a U-shaped structure, so it is called Unet. This is the reverse standard Brownian motion.

[0030] like Figure 3 As shown, before this, the reverse process recovers the data by reversing the dynamics of the forward process. The formula is: , in, is the drift term in the reverse process, which guides the model to generate samples towards the data distribution. In order to estimate the gradient in the reverse process, denoising score matching (DSM) is used, and its optimization goal is: ,in, Represents all adjustable parameters of the model, and the goal is to find the optimal parameter value by minimizing the loss function. For time and image status During the training process, the loss function is the expected value of all possible times. and image status To seek expectations, is a two-norm square, which means calculating the difference between the two. It is a weighting term for different time steps, the purpose of which is to allow the model to learn to recover data under different noise levels. is the neural network trained to predict the denoised score.

[0031] The present invention reduces the high computational overhead in three-dimensional image processing by means of slice-level processing. The method decomposes the processing of three-dimensional data into multiple independent two-dimensional slice processing, which greatly reduces the computational burden.

[0032] The present invention reduces the consumption of computing resources and improves computing efficiency. In particular, it can quickly achieve high-quality three-dimensional reconstruction under GPU acceleration. Compared with the traditional three-dimensional diffusion model, it can achieve more efficient image reconstruction and meet practical application needs such as low-dose CT scanning and fast MRI imaging.

[0033] Step S20, performing curvature regularization calculation according to the initial predicted image to obtain the target curvature of the initial predicted image in the phase in the Z-axis direction, wherein the target curvature includes total square curvature, total regularized variation, total absolute curvature, mean square error, Geman regularization, logarithmic determinant regularization, Laplace distribution regularization and exponential energy regularization.

[0034] Traditional medical image reconstruction methods (such as MBIR, Model-Based Iterative Reconstruction) usually rely on techniques such as Tikhonov regularization and total variation (TV) regularization to constrain the smoothness or edge features of the image, but these methods have limitations in capturing three-dimensional spatial information and enhancing image structural consistency. Especially under sparse or limited angle conditions, traditional methods often cannot fully utilize the spatial correlation between slices.

[0035] In order to enhance spatial consistency, the present invention introduces a curvature-based regularization prior. This regularization method can effectively capture and utilize spatial correlation along the z-axis, especially under sparse viewing angle or limited angle scanning, which can significantly improve the spatial consistency of three-dimensional reconstruction, reduce artifacts between slices, and enhance structure retention.

[0036] Understandably, medical imaging systems such as CT (Computed Tomography) and MRI (Magnetic Resonance Imaging) are modeled using linear forward equations: ,here Represents measurement data (such as sinusoids in CT or k-space data in MRI, such as Figure 2 b, is the CT projection measurement data under sparse 8 viewing angles), is the image to be reconstructed, is an m-dimensional real vector, is the system transformation matrix (such as the Radon transform in CT or the Fourier transform in MRI, Figure 2 The a in it represents the Radon transform in CT). represents the measurement noise.

[0037] In order to solve the problem of 3D sparse view reconstruction, a method based on a two-dimensional diffusion model and model-based iterative reconstruction (MBIR) is proposed. In the reconstruction process, a curvature regularization term is introduced to enhance spatial consistency. The optimization problem formula of curvature regularization is: ;in, is the measurement data, is the system transformation matrix, is the reconstructed image, is the curvature along the z-axis, is the curvature regularization term.

[0038] The curvature regularization term smoothes artifacts between slices by capturing the spatial continuity of the image while preserving details and avoiding over-smoothing. The specific form of this term is: ,in, is the regularization weight, is a function that depends on the curvature, is the image gradient, is the spatial domain. The part that calculates the curvature is as follows Figure 2 The present invention realizes image reconstruction by calculating the curvature of adjacent images in the z-axis direction, wherein the curvature equation Including total square curvature (TSC), total regularized variation (TRV), total absolute curvature (TAC), mean square error (MS), Geman regularization, logarithmic determinant regularization (log-det), Laplace distribution regularization (Laplace) and exponential energy regularization (EE), as shown in Table 1 below.

[0039] Table 1: List of curvature functions

[0040] By introducing the curve prior, the present invention can effectively reduce the spatial artifacts between slices, maintain the edge details and contrast of the image, and achieve better results than traditional methods, especially in preserving the edge structure of medical images.

[0041] Step S30: using an alternating direction multiplier method to perform image iterative reconstruction processing on the target curvature to obtain an initial medical reconstructed image.

[0042] Specifically, an auxiliary variable is introduced, and an augmented Lagrangian function corresponding to the auxiliary variable and the target curvature is calculated; an alternating direction multiplier method is used to iteratively update the augmented Lagrangian function and iteratively reconstruct the image to obtain an initial medical reconstructed image.

[0043] The present invention takes TAC (total absolute curvature) as an example to introduce auxiliary variables and The unconstrained optimization problem can be transformed into a constrained optimization problem: ,in, is the curvature parameter, is the curvature, is the absolute value of the curvature, Auxiliary variables The absolute value of the variable indicates the size of the variable. is a constraint, that is, a condition that must be satisfied during the optimization process. The meaning of this formula is to minimize the objective function, which consists of a data fitting term and a curvature regularization term and satisfies some constraints, specifically: yes The gradient of and equal.

[0044] Augmented Lagrangian function The definition is as follows: ;in, and represents the Lagrange multiplier, and represents the penalty parameter, Indicates transpose.

[0045] ADMM (Alternating Direction Method of Multipliers) is used to iteratively solve the sub-minimization problem. ADMM converts the original nonlinear constraint problem into multiple easy-to-solve sub-problems by introducing auxiliary variables. The ADMM optimization process is updated through the following iterative steps: 1. Sub-questions: ;in, represents the updated , is the gradient operator, which means the partial derivative operation on the coordinates in the space. is the transpose of the gradient, For the The gradient in the iteration Sub-question, is the Lagrange multiplier. In the alternating direction multiplier method (ADMM), represents the penalty for the constraint.

[0046] 2. Sub-questions: , yes The purpose of updating the sub-problem is to make the gradient of the image meet the optimization constraints; 3. Sub-questions: ,in, It is updated by the shrink operator Sub-question, yes Update, Represents the shrinkage operator to enforce the curvature of the image to maintain sparsity.

[0047] 4. Lagrange multipliers : ; 5. Lagrange multipliers : .

[0048] This optimization process is done by updating Sub-problems and Lagrange multipliers integrate data fidelity with the prior knowledge of the curve to achieve high-quality 3D reconstruction while balancing computational efficiency and accuracy. This update step is Figure 2 In d.

[0049] Step S40: performing re-iterative processing on the initial medical reconstructed image to obtain a target medical reconstructed image.

[0050] Specifically, the initial medical reconstruction image is iteratively input into the two-dimensional diffusion model, and the curvature regularization calculation and the image iterative reconstruction processing are performed; when the number of iterations reaches a preset iteration threshold, the iteration is stopped and the target medical reconstructed image is output.

[0051] The image reconstruction process in the present invention is implemented step by step. In each iteration, the score is first predicted by the Unet network, and then the curvature in the z-axis direction is calculated using the measured value of the sparse view, and the regularized reconstruction is performed by the ADMM alternating direction multiplier method. The new variables obtained are passed to the next MBIR iteration process. The computational efficiency is improved by the shared variable strategy. Only one curvature and ADMM calculation need to be performed in each iteration. By combining denoising, data consistency and spatial regularization, the quality and efficiency of image reconstruction are significantly improved. Especially under sparse view conditions, it can effectively reduce artifacts and enhance the spatial consistency of images, and has strong practical application value.

[0052] Experimental verification: like Figure 4 As shown in the figure, the comparison of axial, sagittal and coronal slice image results of 3D CT image reconstruction using 8 views on the LDCT data set. From left to right, they are: back-projection reconstruction, domain-specific perception network, step-by-step wavelet optimization diffusion model, manifold-constrained diffusion model, total variation diffusion model, experimental results of the present invention and true value. By comparing with other methods, it can be found that the present invention has better reconstruction effect.

[0053] like Figure 5 The figure shows the comparison of axial, sagittal and coronal slice image results of 3D CT image reconstruction using 4 views on the LDCT dataset. From left to right, they are: back-projection reconstruction, domain-specific perception network, step-by-step wavelet optimization diffusion model, manifold-constrained diffusion model, total variation diffusion model, experimental results of the present invention and true value.

[0054] like Figure 6 As shown in the figure, the comparison of axial, sagittal and coronal slice image results of 3D CT image reconstruction using 8 views on the LDCT data set. From left to right, they are: back-projection reconstruction, domain-specific perception network, step-by-step wavelet optimization diffusion model, manifold-constrained diffusion model, total variation diffusion model, experimental results of the present invention and true value. By comparing with other methods, it can be found that the present invention has better reconstruction effect.

[0055] like Figure 7 The figure shows the reconstruction results of CS-MRI (CS-MRI is an important branch of computer science. CS-MRI reconstructs the original signal from a small number of signal samples by utilizing the principle of signal sparsity) in BraTS18 FLAIR (brain tumor segmentation imaging) mode. From left to right, they are: total square curvature + 2D diffusion, total regularized variation + 2D diffusion, total absolute curvature + 2D diffusion, Geman regularization + 2D diffusion, logarithmic determinant regularization + 2D diffusion, and exponential energy regularization + 2D diffusion.

[0056] As shown in Table 2 below, for quantitative evaluation of sparse view reconstruction (2, 4, and 8 views) on LDCT, the axial-plane (on which the diffusion model was applied) and the coronal and sagittal views were compared. Bold indicates the best performance, and underlined indicates the suboptimal performance. The table highlights the comparative performance of the different views (axial, coronal, sagittal), the comparison of sagittal views in terms of image quality, and highlights which configurations produce the best and suboptimal results.

[0057] Table 2: Quantitative evaluation of sparse view reconstruction (2, 4 and 8 views) on LDCT

[0058] Table 2 continued:

[0059] like Figure 8 As shown, the quantitative comparison of the 3D CT image reconstruction results using 8 views on the LDCT dataset ( Figure 8 a in the figure is a quantitative comparison of the peak signal-to-noise ratio and the three-dimensional CT image reconstruction results corresponding to the axial plane based on 8 views on the LDCT dataset; Figure 8 b in the figure is a quantitative comparison of the structural similarity and the 3D CT image reconstruction results corresponding to the axial plane based on 8 views on the LDCT dataset); Fig. 9 As shown, the quantitative comparison of the 3D CT image reconstruction results using 4 views on the LDCT dataset ( Fig. 9 a in the figure is a quantitative comparison of the peak signal-to-noise ratio and the three-dimensional CT image reconstruction results corresponding to the sagittal plane based on the use of 4 views on the LDCT dataset; Fig. 9 b in the figure is a quantitative comparison of the structural similarity and the 3D CT image reconstruction results corresponding to the sagittal plane based on the 4-view LDCT dataset); Fig.10 As shown, the quantitative comparison of the 3D CT image reconstruction results using 2 views on the LDCT dataset ( Fig.10 a in the figure is a quantitative comparison of the peak signal-to-noise ratio and the corresponding coronal plane 3D CT image reconstruction results based on 2 views on the LDCT dataset; Fig.10 (b) is a quantitative comparison of the structural similarity and the 3D CT image reconstruction results corresponding to the coronal plane based on 2 views on the LDCT dataset.

[0060] Compared with the prior art, the advantages of the present invention are as follows: 1. Traditional MBIR methods and diffusion model-based image generation methods often work independently. MBIR achieves data consistency through iterative optimization, while diffusion models are mainly used for denoising or image generation. The combination of the two requires optimization at multiple levels. Existing diffusion model-based methods focus more on denoising and ignore data consistency.

[0061] This paper combines MBIR and diffusion model in a hybrid optimization framework, and proposes a hybrid optimization framework based on "Alternating Direction Method of Multipliers (ADMM)". By alternating data consistency update and denoising steps, the two can complement each other, thereby achieving higher accuracy and consistency in the image reconstruction process. This hybrid optimization framework achieves a balance between data consistency and image denoising, can improve the reconstruction quality of three-dimensional images, and reduce artifacts and noise that may occur in traditional methods, especially under limited and sparse viewing angles.

[0062] 2. Traditional three-dimensional image reconstruction methods are computationally intensive and require a long time to process. Especially when using three-dimensional diffusion models, due to high-dimensional calculations, real-time performance is poor and it is difficult to meet the clinical demand for rapid processing.

[0063] The present invention adopts a GPU-accelerated algorithm framework, and significantly improves the processing speed by performing parallel calculations on each slice in the image reconstruction process. In particular, a calculation method adapted to GPU acceleration is adopted for the data consistency update and diffusion denoising steps in the iterative optimization process. Through GPU acceleration, not only the image reconstruction time is greatly shortened, but also a larger medical image data set can be processed, thereby meeting the needs of clinical real-time imaging and large-scale medical data processing.

[0064] 3. Most current medical image reconstruction methods, especially those based on machine learning, are usually optimized for specific imaging devices or data sets and lack multimodal generalization capabilities. It is often difficult to achieve the same effect under different scanning devices or different imaging conditions.

[0065] The method of the present invention has strong multimodal adaptability, can adapt to different CT and MRI scanning conditions, and can be applied to different data sets. Based on the combination of data consistency and a priori, it can maintain good reconstruction effects in different imaging modes. The present invention can maintain high reconstruction quality under different data sets and imaging conditions, shows good generalization ability, and can be applied to different types of medical imaging tasks.

[0066] Technical effects of the present invention: 1. Reduced computational complexity: By using a pre-trained two-dimensional diffusion model, the reconstruction process does not need to directly perform high-dimensional calculations of the 3D diffusion model. Instead, the reconstruction process is decomposed into independent processing of multiple two-dimensional slices. This design greatly reduces the amount of computation of three-dimensional data and can maintain the spatial consistency of the image while significantly reducing the consumption of computing resources. By using a two-dimensional diffusion model to process a single slice, the high computational overhead in three-dimensional data is avoided. This greatly reduces computing resources and adapts to clinical applications that have high requirements for computational efficiency, such as low-dose CT scanning and fast MRI acquisition.

[0067] 2. Enhanced spatial consistency and structural retention: By introducing curvature regularization prior, the present invention can effectively capture and retain the spatial correlation between slices, reduce artifacts between slices, and ensure the structural consistency of the three-dimensional reconstruction results. Curvature regularization helps to improve the reconstruction quality under sparse viewing conditions, especially in the reconstruction of limited viewing angle CT and low-dose CT, and can effectively retain details and edges. By introducing curvature-based regularization prior, the spatial correlation between slices along the z-axis can be captured. It effectively avoids artifacts and structural inconsistencies between slices, improves the image quality of sparse viewing angle CT and limited angle CT reconstruction, and makes the reconstructed image more in line with clinical needs.

[0068] 3. Improved reconstruction accuracy and image quality: The present invention combines MBIR (model-based iterative reconstruction) with the denoising process of the diffusion model and optimizes it through the alternating direction multiplier method (ADMM). Through this optimization process, data consistency can be maintained while the denoising step is used to improve reconstruction accuracy. Experimental results show that in CT and MRI reconstruction, the framework of the present invention can achieve higher PSNR and SSIM values ​​than traditional methods on multiple data sets. By combining MBIR with the denoising of the diffusion model and optimizing the framework through ADMM, the reconstruction accuracy and image quality are improved, and the details and structure of the original medical image can be better restored, especially under sparse and limited viewing angles, to obtain clearer images.

[0069] 4. Enhanced generalization ability: The framework of the present invention demonstrates superior generalization ability, can adapt to different CT and MRI scan data sets, and can maintain good performance under different imaging conditions. This shows that the technical solution of the present invention is not only effective on specific data sets, but also has strong cross-platform application capabilities. The present invention combines a data-driven diffusion model with a model-based regularization method, which enhances the generalization ability of the model and can provide stable reconstruction effects under different data sets and different imaging conditions.

[0070] 5. Improved feasibility of clinical applications: The algorithm of the present invention uses GPU acceleration to optimize the calculation process and improve the efficiency of image reconstruction. For real-time imaging needs in clinical environments, it can provide high-quality image reconstruction results in a limited time, providing support for clinical applications such as low-dose CT scanning and fast MRI imaging. By combining GPU acceleration optimization with an efficient denoising framework, the present invention shortens the reconstruction time and meets the needs of clinical environments for fast and efficient medical image reconstruction.

[0071] In addition, the technical solution of the present invention can be implemented through a variety of implementations. Several possible deformation methods and similar methods are listed below: 1. Use different diffusion models for reconstruction: Although the present invention uses a pre-trained two-dimensional diffusion model for image reconstruction, other types of generative models can also be considered, such as variational autoencoders (VAE), generative adversarial networks (GAN), etc. Through these alternative generative models, the reconstruction process can still be decomposed into multiple slices, and curvature regularization is used to enhance spatial consistency. The implementation method is as follows: replace the two-dimensional diffusion model with VAE or GAN, maintain the image generation ability, and explore the combination effect of other models.

[0072] 2. Implementation based on other optimization methods: The present invention uses the alternating direction method of multipliers (ADMM) as the optimization framework, but other common optimization methods can also be used, such as the conjugate gradient method (CG), stochastic gradient descent (SGD), and gradient descent. Different optimization methods may have different effects in specific application scenarios, but their core goal is to maintain data consistency while optimizing image quality. The implementation method is as follows: Replace ADMM with the conjugate gradient method, SGD or other optimization methods, and select the most appropriate optimization algorithm based on different computing resources and reconstruction accuracy requirements.

[0073] 3. Multimodal medical image reconstruction: In addition to CT and MRI, the method framework of the present invention can also be applied to other medical imaging tasks, such as positron emission tomography (PET) and single photon emission computed tomography (SPECT). In order to adapt to these imaging modes, the forward operator and data consistency update strategy can be adjusted so that the algorithm can adapt to different imaging features and data structures. The implementation method is as follows: According to different imaging modes, the forward model (for example, the forward projection operator of PET) is adjusted, and the same diffusion model and curvature regularization strategy are applied.

[0074] 4. Combination of deep learning and traditional methods: Although the present invention has achieved good results through diffusion model and curvature regularization, it is possible to consider combining deep learning methods with this framework. For example, in the process of image denoising or reconstruction, a convolutional neural network (CNN) or a traditional low-pass filter is used to assist the denoising step, thereby further improving the reconstruction accuracy. The implementation method is as follows: A convolutional neural network (CNN) module is added to the process of diffusion denoising, and adaptive denoising and feature extraction are performed through deep learning.

[0075] 5. Expansion to large-scale data sets: For large-scale medical data sets, the algorithm may need to perform distributed computing or adopt a more efficient memory management strategy. In order to adapt to these large-scale applications, the framework of the present invention can be deployed in a distributed computing environment, using multiple GPUs or cloud computing resources to accelerate the data processing process. The implementation method is as follows: deploy a distributed computing framework to process large-scale medical image data through cloud computing or multi-GPU clusters to further improve efficiency.

[0076] Possible combinations of key technical points are as follows: 1. Two-dimensional diffusion model + curvature regularization prior: Combining the two-dimensional diffusion model with the curvature regularization prior, through slice-level processing and spatial consistency enhancement, efficient and high-quality three-dimensional image reconstruction can be achieved.

[0077] 2. MBIR optimization framework + diffusion model denoising: The alternating direction method of multipliers (ADMM) optimization framework is used to combine the MBIR and diffusion model denoising processes to ensure data consistency and improve the image denoising effect at the same time.

[0078] 3. GPU acceleration + hybrid optimization framework: GPU acceleration is used to achieve fast processing, and the hybrid optimization framework combined with MBIR and diffusion model can ensure image quality while efficiently reconstructing and meet clinical needs.

[0079] The overall replacement or modification scheme is as follows: 1. Multi-scale model: The core technical feature of this invention is to use 2D diffusion model and curve prior to perform 3D reconstruction, and using multi-scale feature extraction at different levels may be an alternative idea. By decomposing the image reconstruction process into multiple scale levels for processing, different effects or performance advantages may be produced in some applications.

[0080] 2. Combination of deep learning and traditional methods: Although the present invention adopts an iterative reconstruction method based on ADMM, it is possible to consider combining the deep learning model with the traditional iterative method. For example, using a convolutional neural network (CNN) to automatically extract and learn image features, and then combining it with MBIR and diffusion models. This method may have better reconstruction quality in specific scenarios.

[0081] 3. Different types of diffusion models: Currently, the present invention adopts 3D reconstruction based on a 2D diffusion model. Can a diffusion model based on 3D convolution or other types of generation models (such as variational autoencoder VAE, generative adversarial network GAN, etc.) be used to directly model and optimize in 3D space to achieve similar effects?

[0082] In addition, considering that the core technical feature of the present invention is to combine the diffusion model with the prior knowledge of the curve to perform 3D reconstruction, there may be some feature substitution or deformation methods, which are mainly reflected in the following aspects: 1. Variation of diffusion model: The current solution uses score-based diffusion models. Can it be replaced with other types of generative models, such as models based on generative adversarial networks (GANs) or variational inference? These generative models can also generate samples through the training process and be applied to medical image reconstruction tasks, although more training data or computing resources may be required under certain conditions.

[0083] 2. Replacement of curvature prior: Curvature regularization is a key feature of the present invention, but can it be replaced by other prior knowledge to enhance the quality of reconstruction? For example, an automatic feature learning method based on deep learning can be used to replace the traditional curvature prior, and the spatial correlation of medical images can be directly learned using the network structure. Alternatively, regularization methods based on image gradients or local structures can be tried, which can sometimes achieve better results in the case of sparse data.

[0084] 3. Variation of ADMM optimization method: This invention uses ADMM (alternating direction multiplier method) for optimization. Can other types of optimization algorithms, such as gradient descent method, conjugate gradient method, L-BFGS, etc., be used to replace ADMM, and compare the performance differences of these methods under different conditions? It is also possible to consider optimizing computing efficiency through algorithm parallelization or graphics processing unit (GPU) acceleration.

[0085] Possible design solutions from competitors: 1. Non-curvature regularization: Competitors may choose other types of regularization methods instead of curvature regularization. For example, low-frequency components, texture information, or gradient information of the image can be used to replace the curvature prior to achieve similar reconstruction effects and avoid directly using curvature-related regularization terms.

[0086] 2. Methods based entirely on deep learning: Competitors may also propose solutions based entirely on deep learning, using an end-to-end deep learning framework to process sparse data, avoiding the use of traditional physical modeling and optimization methods (such as ADMM), and directly recovering images from sparse data by training neural networks, which may avoid the methods of the present invention.

[0087] 3. Other types of generative models: As mentioned earlier, competitors may use generative adversarial networks (GANs) or variational autoencoders (VAEs) to replace diffusion models for medical image reconstruction. These generative models can be learned and optimized end-to-end based on a large amount of data without traditional regularization terms, which may become a way to avoid design.

[0088] 4. Expand to real-time imaging tasks: In order to achieve real-time imaging, we can consider accelerating the algorithm and using FPGA or dedicated hardware accelerators to achieve real-time image reconstruction. Such an implementation can greatly improve the efficiency of clinical imaging and meet the dual requirements of real-time and high quality. A smaller model architecture can be used to achieve faster 3D medical image reconstruction.

[0089] 5. Optimization of automatic hyperparameter search: For example, Bayesian optimization, grid search, or random search methods are used to automatically adjust the hyperparameters in the model (such as learning rate, regularization strength, and diffusion model sampling steps, etc.), thereby optimizing the performance of the model in different scenarios. By automatically adjusting the parameters, the tedious work of manual parameter adjustment can be avoided, and the accuracy and robustness of the algorithm can be improved. The hyperparameter optimization algorithm applicable to the present invention can be further explored to optimize the reconstruction quality, reduce the calculation time, and reduce the memory usage.

[0090] Furthermore, if Fig.11 As shown, based on the above-mentioned three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model, the present invention also provides a three-dimensional sparse view medical image reconstruction system based on the two-dimensional diffusion model, wherein the three-dimensional sparse view medical image reconstruction system based on the two-dimensional diffusion model includes: An initial predicted image generation module 51 is used to obtain an original medical image under a sparse view, and input the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view; A curvature regularization calculation module 52, configured to perform curvature regularization calculation according to the initial predicted image to obtain a target curvature of the initial predicted image in the phase in the Z-axis direction; An image iterative reconstruction processing module 53 is used to perform image iterative reconstruction processing on the target curvature by using an alternating direction multiplier method to obtain an initial medical reconstructed image; The iterative reconstruction processing module 54 is used to perform iterative processing on the initial medical reconstructed image to obtain a target medical reconstructed image.

[0091] Furthermore, if Fig.12 As shown, based on the above-mentioned three-dimensional sparse view medical image reconstruction method and system based on the two-dimensional diffusion model, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Fig.12 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0092] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal. Further, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a three-dimensional sparse view medical image reconstruction program 40 based on a two-dimensional diffusion model is stored on the memory 20, and the three-dimensional sparse view medical image reconstruction program 40 based on a two-dimensional diffusion model can be executed by the processor 10, thereby realizing the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model in the present application.

[0093] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model.

[0094] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface.

[0095] In one embodiment, when the processor 10 executes the three-dimensional sparse view medical image reconstruction program 40 based on the two-dimensional diffusion model in the memory 20, the steps of the three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model as described above are implemented.

[0096] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model, and when the three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model is executed by a processor, the steps of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model as described above are implemented.

[0097] In summary, the present invention provides a method, system, terminal and computer-readable storage medium for reconstructing a three-dimensional sparse view medical image based on a two-dimensional diffusion model, the method comprising: obtaining an original medical image under a sparse view, and inputting the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view; performing curvature regularization calculation according to the initial predicted image to obtain a target curvature of the initial predicted image in the phase in the Z-axis direction; performing image iterative reconstruction processing on the target curvature using an alternating direction multiplier method to obtain an initial medical reconstructed image; and performing re-iterative processing on the initial medical reconstructed image to obtain a target medical reconstructed image. The present invention realizes an effective combination of a two-dimensional diffusion model and a curvature regularization technology by inputting the original medical image under a sparse view into a two-dimensional diffusion model and processing it using curvature regularization, thereby reconstructing the original medical image under the sparse view. There is no need to perform high-dimensional calculations of a 3D diffusion model, which greatly reduces the amount of calculation of three-dimensional data and effectively improves the image quality and reconstruction accuracy of the target medical reconstructed image.

[0098] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0099] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium that can be read by a computer, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0100] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model, characterized in that: The three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model includes: Acquire an original medical image under a sparse view, and input the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view; Performing curvature regularization calculation according to the initial predicted image to obtain a target curvature of the initial predicted image in the phase in the Z-axis direction; An alternating direction multiplier method is used to perform an iterative image reconstruction process on the target curvature to obtain an initial medical reconstructed image; The initial medical reconstructed image is iteratively processed to obtain a target medical reconstructed image.

2. The three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model according to claim 1, characterized in that: The two-dimensional diffusion model includes a noise enhancement model and a noise removal model; The obtaining of an original medical image under a sparse view and inputting the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view specifically includes: Determining a sparse view to be reconstructed, and acquiring an original medical image corresponding to the sparse view; Inputting the original medical image into the denoising model in the two-dimensional diffusion model, performing forward denoising processing on the original medical image through the denoising model to obtain denoised Gaussian distribution data; Using a denoising score matching method to optimize and train the denoising model in the two-dimensional diffusion model to obtain an optimized and trained denoising model; The denoised Gaussian distribution data is input into an optimized and trained denoising model, and the denoised Gaussian distribution data is subjected to reverse denoising processing by optimizing the trained denoising model to obtain an initial predicted image corresponding to the sparse view.

3. The three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model according to claim 2, characterized in that: The calculation expression of the forward noise enhancement process is: ； Among them, the To add noise to Gaussian distribution data, is the time-dependent variance function, is the differential of the variance function, is the increment of Brownian motion, is the differential of time.

4. The method for reconstructing a three-dimensional sparse view medical image based on a two-dimensional diffusion model according to claim 3, characterized in that: The calculation expression of the reverse denoising process is: ； Among them, the is the initial prediction image, To optimize the trained denoising model, is the image status, For time, This is the reverse standard Brownian motion.

5. The three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model according to claim 1, characterized in that: The target curvatures include total squared curvature, total regularized variation, total absolute curvature, mean square error, Geman regularization, logarithmic determinant regularization, Laplace distribution regularization, and exponential energy regularization.

6. The three-dimensional sparse view medical image reconstruction method based on the two-dimensional diffusion model according to claim 1, characterized in that: The iterative image reconstruction process of the target curvature by using the alternating direction multiplication method to obtain an initial medical reconstructed image specifically includes: Introducing an auxiliary variable, and calculating an augmented Lagrangian function corresponding to the auxiliary variable and the target curvature; The augmented Lagrangian function is iteratively updated and the image is iteratively reconstructed using an alternating direction multiplier method to obtain an initial medical reconstructed image.

7. The method for reconstructing a three-dimensional sparse view medical image based on a two-dimensional diffusion model according to claim 1, characterized in that: The iterative processing of the initial medical reconstructed image to obtain a target medical reconstructed image specifically includes: Iteratively inputting the initial medical reconstruction image into the two-dimensional diffusion model, and performing the curvature regularization calculation and the image iterative reconstruction process; When the number of iterations reaches a preset iteration threshold, the iteration is stopped and the target medical reconstructed image is output.

8. A three-dimensional sparse view medical image reconstruction system based on a two-dimensional diffusion model, characterized in that: The three-dimensional sparse view medical image reconstruction system based on the two-dimensional diffusion model includes: An initial predicted image generation module, used for acquiring an original medical image under a sparse view, and inputting the original medical image into a two-dimensional diffusion model to obtain an initial predicted image corresponding to the sparse view; A curvature regularization calculation module, used to perform curvature regularization calculation according to the initial predicted image to obtain a target curvature of the initial predicted image in the phase in the Z-axis direction; An image iterative reconstruction processing module is used to perform image iterative reconstruction processing on the target curvature by using an alternating direction multiplier method to obtain an initial medical reconstructed image; The iterative reconstruction processing module is used to perform iterative processing on the initial medical reconstructed image to obtain a target medical reconstructed image.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model stored in the memory and executable on the processor. When the three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model is executed by the processor, the steps of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model. When the three-dimensional sparse view medical image reconstruction program based on a two-dimensional diffusion model is executed by a processor, the steps of the three-dimensional sparse view medical image reconstruction method based on a two-dimensional diffusion model are implemented.

Images