CT image reconstruction method based on three-dimensional diffusion model and related device

Abstract

The invention belongs to the field of CT image processing, and particularly relates to a CT image reconstruction method based on a three-dimensional diffusion model and a related device.The method comprises the steps that a medical CT scanner is adopted to obtain three-dimensional CT voxel data, and the three-dimensional CT voxel data is preprocessed; extracting a medical object surface point cloud from the preprocessed three-dimensional CT voxel data by adopting a voxel-level reconstruction algorithm, and generating three-dimensional sparse point cloud data; processing the three-dimensional sparse point cloud data by adopting a 3D Gaussian splashing technology, and converting the three-dimensional sparse point cloud data into three-dimensional Gaussian distribution data by using a pre-trained three-dimensional Gaussian kernel; gradually adding Gaussian noise to the three-dimensional Gaussian distribution data by adopting a three-dimensional diffusion model, and gradually removing the Gaussian noise to generate reconstructed three-dimensional Gaussian distribution data; and projecting the reconstructed three-dimensional Gaussian distribution data to a two-dimensional plane by adopting a rendering technology to generate a two-dimensional CT reconstructed image. According to the method, the advantages of the 3D Gaussian splashing technology and the three-dimensional diffusion model are combined, efficient, real-time and denoising medical image reconstruction can be achieved under the low-dose CT imaging condition, and meanwhile the problem of white floater artifacts possibly occurring in the reconstruction process is effectively solved.

Description

Technical Field

[0001] The invention belongs to the technical field of medical sparse view CT reconstruction, and in particular relates to a CT image reconstruction method based on a three-dimensional diffusion model and a related device. Background Art

[0002] In a 3D CT scan, an object or human body is irradiated with X-rays from different angles. The detector receives the X-rays that pass through the object and converts them into digital signals. These digital signals are recorded to form the original 3D voxel data. 3D voxel data is the smallest data unit in 3D space, similar to pixels in 2D images. Through specific reconstruction algorithms, these 3D voxel data can be converted into 2D CT images.

[0003] At present, three-dimensional CT reconstruction technology is an important tool for analyzing internal structures and diagnosis. Traditional three-dimensional CT reconstruction methods such as algebraic reconstruction and filtered back projection are susceptible to noise under low-dose CT, resulting in artifacts and reconstruction errors. In addition, these methods have difficulty in generating high-quality images under sparse view conditions. In recent years, deep learning methods such as convolutional neural networks have been introduced into CT reconstruction, such as Chinese patent CN116824146A. Although these solutions have improved image quality and noise suppression, their dependence on a large amount of annotated data, complex training process, and lack of generalization ability limit their application. Although generative adversarial networks have advantages in generating high-definition images, their training is unstable and prone to unrealistic artifacts. Neural implicit representation methods such as NeRF are effective in simulating complex geometry, but due to high computational costs and long training time, they are difficult to meet real-time imaging requirements.

[0004] At present, some technicians use 3D Gaussian splashing technology to solve the problem of 3D image reconstruction. The core idea of ​​3D Gaussian splashing is to represent the 3D scene as a set of 3D Gaussian functions. Each 3D Gaussian function has its specific parameters, including 3D position, covariance matrix, color and transparency. These Gaussian functions are distributed in 3D space and together constitute a continuous and realistic 3D scene. However, 3D Gaussian splashing technology cannot directly reconstruct 3D CT voxel data, and 3D Gaussian splashing technology is difficult to remove noise and white floating artifacts in CT images. Summary of the invention

[0005] In view of this, the present invention proposes a CT image reconstruction method and related devices based on a three-dimensional diffusion model, which reconstructs three-dimensional CT voxel data using a voxel-level reconstruction algorithm, and then uses a 3D Gaussian splashing technology to convert three-dimensional sparse point cloud data into high-quality three-dimensional Gaussian distribution data; the three-dimensional diffusion model is used to process the three-dimensional Gaussian distribution data, which can gradually remove noise and white floating object artifacts in CT images, significantly suppress noise and reduce artifacts under sparse view conditions, restore the original image details, and effectively maintain the structural information of the medical image.

[0006] In a first aspect of the present invention, the present invention provides a CT image reconstruction method based on a three-dimensional diffusion model, the method comprising:

[0007] Acquiring three-dimensional CT voxel data using a medical CT scanner, and preprocessing the three-dimensional CT voxel data;

[0008] The voxel-level reconstruction algorithm is used to extract the surface point cloud of the medical object from the preprocessed 3D CT voxel data to generate 3D sparse point cloud data;

[0009] The 3D Gaussian splashing technology is used to process the 3D sparse point cloud data, and the pre-trained 3D Gaussian kernel is used to convert it into 3D Gaussian distribution data;

[0010] Using a three-dimensional diffusion model, gradually adding Gaussian noise to the three-dimensional Gaussian distribution data, and gradually removing the Gaussian noise, to generate reconstructed three-dimensional Gaussian distribution data;

[0011] The rendering technology is used to project the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane to generate a two-dimensional CT reconstructed image.

[0012] In a second aspect of the present invention, the present invention further proposes a CT image reconstruction device based on a three-dimensional diffusion model, the device comprising:

[0013] A data acquisition module, which uses a medical CT scanner to acquire three-dimensional CT voxel data and pre-processes the three-dimensional CT voxel data;

[0014] The voxel reconstruction module uses a voxel-level reconstruction algorithm to extract the surface point cloud of the medical object from the preprocessed 3D CT voxel data and generate 3D sparse point cloud data;

[0015] The point cloud reconstruction module uses 3D Gaussian splashing technology to process 3D sparse point cloud data and converts it into 3D Gaussian distribution data using pre-trained 3D Gaussian kernel;

[0016] A diffusion reconstruction module, which uses a three-dimensional diffusion model to gradually add Gaussian noise to the three-dimensional Gaussian distribution data, and gradually removes the Gaussian noise to generate reconstructed three-dimensional Gaussian distribution data;

[0017] The rendering and reconstruction module uses rendering technology to project the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane to generate a two-dimensional CT reconstructed image.

[0018] In the third aspect of the present invention, the present invention further proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect of the present invention when executing the computer program.

[0019] In a fourth aspect of the present invention, the present invention further proposes a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in the first aspect of the present invention.

[0020] Compared with the prior art, the advantages of the present invention are:

[0021] The present invention overcomes many limitations of existing CT reconstruction technology. First, the frequency regularization technology is used to constrain the position encoding of the Gaussian kernel, which effectively prevents the overfitting phenomenon that may occur in the model during the training process, thereby improving the generalization ability of the reconstruction results. Secondly, the present invention uses a three-dimensional diffusion model for denoising, which can gradually remove the noise and white floating artifacts in the CT image, restore the original image details, and effectively maintain the structural information of the medical image, avoiding the over-smoothing phenomenon that may occur in traditional denoising methods, retaining details and edge information, and improving the diagnostic value of the image. Finally, combined with the 3D Gaussian Splatting technology, the present invention realizes efficient three-dimensional data mapping and image reconstruction, while ensuring image quality, significantly improving computational efficiency. Finally, the present invention has real-time processing capabilities, is particularly suitable for rapid reconstruction under low-dose scanning conditions, and meets the dual requirements of real-time and high quality in medical imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of a CT image reconstruction method based on a three-dimensional diffusion model according to an embodiment of the present invention;

[0023] Figure 2 This is a flow chart of a CT image reconstruction method based on a three-dimensional diffusion model according to a preferred embodiment of the present invention;

[0024] Figure 3 A schematic diagram of a CT image reconstruction structure based on a three-dimensional diffusion model according to an embodiment of the present invention;

[0025] Figure 4 The figure is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] In some embodiments, this embodiment provides a CT image reconstruction method based on a three-dimensional diffusion model, such as Figure 1 As shown, the method includes the following contents:

[0028] S101. Acquire three-dimensional CT voxel data using a medical CT scanner, and pre-process the three-dimensional CT voxel data. In some embodiments, acquire three-dimensional CT voxel data using a medical CT scanner, perform smoothing interpolation and denoising processing to improve data quality.

[0029] First, a medical CT scanner is used to collect a series of two-dimensional slice image data of the chest. By setting appropriate scanning parameters, the chest is scanned from multiple angles to obtain about 20 to 30 chest CT images at different angles. Each image represents a cross-section or slice of the chest, and these two-dimensional slice data will be combined into a three-dimensional chest CT voxel data set through a stacking operation. Secondly, filtering techniques (such as Gaussian filtering, median filtering, etc.) are applied to remove noise from the image. Noise removal can improve the clarity and details of the image and reduce errors in subsequent processing.

[0030] S201, extracting a surface point cloud of a medical object from the preprocessed three-dimensional CT voxel data using a voxel-level reconstruction algorithm to generate three-dimensional sparse point cloud data;

[0031] In some embodiments, the present embodiment uses the Marching Cubes algorithm to extract the surface point cloud of the medical object from the preprocessed CT data, generate sparse point cloud data, and provide three-dimensional geometric information.

[0032] After step S101, the chest 3D CT voxel data after smooth interpolation and denoising is obtained. The voxel data set contains various structural information inside the chest, including the lungs, chest cavity, heart, and blood vessels.

[0033] First, the Marching Cubes algorithm is used to perform surface detection in 3D voxel data. For each voxel point, the algorithm analyzes other voxel points in its neighborhood to find surface boundaries that meet the set threshold T. For soft tissue in chest CT scan data, the threshold range is set to 200. When the value of a voxel point reaches or exceeds T, it is identified as a voxel point on the surface.

[0034] When there is a sign change between adjacent CT voxel values, it indicates that there is an isosurface between the two CT voxels. The intersection coordinates are calculated using a linear interpolation method to determine the exact intersection of the isosurface and the CT voxel boundary.

[0035] According to the status of each CT voxel, the connection method of the vertices is determined by referring to the lookup table of the Marching Cubes algorithm; through triangle connection, a continuous mesh is formed to reconstruct the surface structure of the chest. In chest CT data, the mesh of the lung and chest surface is mainly composed of larger triangles, while in more delicate structural vascular branches and small cavities, the mesh is composed of more and denser triangles. Through this triangular mesh connection method, the three-dimensional structure of the surface of the chest organs is fully presented.

[0036] S301, using 3D Gaussian splashing technology to process the three-dimensional sparse point cloud data, and using the pre-trained three-dimensional Gaussian kernel to convert it into three-dimensional Gaussian distribution data;

[0037] In some embodiments, a 3DGS method is used to train a Gaussian kernel from three-dimensional point cloud data to represent a real three-dimensional scene, and frequency regularization is applied to the Gaussian kernel position encoding.

[0038] In step S201, sparse point cloud data of the medical object has been extracted from the chest CT data, including the three-dimensional surface information of the chest. Next, the 3DGS method is used to convert this sparse point cloud data into a Gaussian kernel representation, thereby constructing a continuous distribution model of the three-dimensional scene. The form of the Gaussian kernel is:

[0039]

[0040] in, represents the Gaussian kernel at position P, μ = (x, y, z) is the position of the point, Σ is the covariance matrix, and |Σ| is the determinant of the covariance matrix.

[0041] This Gaussian kernel representation method can construct a three-dimensional scene with continuous characteristics based on the sparse representation of chest CT point cloud, effectively describing the geometric morphology and color characteristics of the chest anatomical structure.

[0042] In some embodiments, this embodiment applies frequency regularization to the position encoding of the Gaussian kernel to improve the model's ability to express high-frequency detail information, while maintaining overall smoothness and enhancing the capture of details, so that the model can better reproduce the fine structures in chest CT data. Frequency regularization adjusts the position encoding by introducing high-frequency components, and the specific formula is:

[0043]

[0044] Among them, f is the regularization frequency factor, which controls the frequency; λ is the regularization weight, which controls the degree of smoothness; ‖P‖ represents the modulus of the position vector, which is used to adjust the regularization effect of different positions. Here P = (x, y, z), and correspondingly, ‖(x, y, z)‖ represents the modulus of the position vector of the point (x, y, z), which measures the distance of the point in space.

[0045] After completing the frequency regularization of the position encoding, the present embodiment can use the gradient descent optimization algorithm to train the Gaussian kernel parameters so that it can best represent the three-dimensional structure of the chest. Specifically, by continuously adjusting the center position and covariance matrix of each Gaussian kernel, the error between the Gaussian kernel representation and the actual chest CT structure is minimized.

[0046] During the training process, the real point cloud data of chest CT is used as the target, and the loss function L is defined to represent the difference between the reconstruction result and the real scene. The loss function is defined as the error combination of the Gaussian kernel and the point cloud data in geometry and color:

[0047]

[0048] Among them, P i,model and P i,target Respectively represent the Gaussian kernel position and target point cloud position generated by the model; C i,model and C i,target They represent the Gaussian kernel color and target point cloud color generated by the model respectively; α is the weight of the color error, which controls the relative importance of the geometric error and the color error.

[0049] S401, using a three-dimensional diffusion model to gradually add Gaussian noise to the three-dimensional Gaussian distribution data, and gradually remove the Gaussian noise to generate reconstructed three-dimensional Gaussian distribution data;

[0050] In some embodiments, Gaussian noise is added to features such as Gaussian kernel position, covariance, and color, and denoised step by step through a three-dimensional diffusion model to restore the original sparse point cloud.

[0051] In some embodiments, in the forward noise addition process, Gaussian noise is independently added to the position encoding vector, covariance matrix, and color eigenvalue of the three-dimensional Gaussian distribution data to generate three-dimensional noisy point cloud data;

[0052] In the reverse denoising process, Gaussian noise is removed independently on the position encoding vector, covariance matrix and color eigenvalue of the three-dimensional noisy point cloud data to generate reconstructed three-dimensional Gaussian distribution data.

[0053] In step S301, the 3D sparse point cloud data of the chest CT data set has been represented as 3D Gaussian distribution data by the 3DGS method, and a 3D model with features such as position, covariance, and color has been generated. Gaussian noise is added to each feature of the 3D Gaussian kernel one by one to simulate the noise and verify the denoising ability of the model.

[0054] For each position point of the Gaussian kernel, add independent and identically distributed Gaussian noise at the position. The noise equation is as follows:

[0055]

[0056] Among them, η P is the position noise term with zero mean and variance Gaussian distribution.

[0057] For each Gaussian kernel's covariance matrix, Gaussian noise is added to the matrix to simulate uncertainty:

[0058]

[0059] Among them, η Σ is the covariance noise matrix with zero mean and variance Gaussian distribution of , I is the identity matrix.

[0060] For each color feature of the Gaussian kernel, add Gaussian noise to the color value:

[0061]

[0062] Among them, η C is the color noise term with zero mean and variance Gaussian distribution, acting on the red, green, and blue channels respectively.

[0063] In some embodiments, the three-dimensional diffusion model is used to gradually remove the Gaussian kernel features after adding noise. The three-dimensional diffusion model gradually removes the Gaussian noise through the steps of forward noise addition and reverse denoising to restore the original three-dimensional Gaussian distribution data.

[0064] In the training phase of the diffusion model, the forward process generates training samples with different noise levels by gradually adding noise to the three-dimensional Gaussian distribution data at each time step t. The noise addition formula of this process is:

[0065]

[0066] Among them, α t Controls the intensity of the noise at each time step, σ 2 is the noise variance, represents the noisy Gaussian kernel at time step t.

[0067] In the denoising stage, the trained three-dimensional diffusion model is used to gradually denoise the noisy Gaussian kernel features. At each time step t, the model predicts the denoised image based on the current noisy image and gradually removes the noise in the reverse diffusion process. The denoising formula of the reverse process is:

[0068]

[0069] Among them, f θ is the trained diffusion model, is the noisy Gaussian kernel feature of the tth step, It is the feature generated after denoising, which gradually approaches the original sparse point cloud.

[0070] In some embodiments, the noise of each time step is gradually removed through the iterative steps of the reverse diffusion process. The denoised Gaussian kernel features obtained in each step will be used as the input of the next step to continue denoising. After multiple iterations, the Gaussian kernel features close to the original sparse point cloud are finally restored. The point cloud data finally obtained has high-quality 3D geometric information and color information, which is suitable for subsequent 3D reconstruction and medical analysis.

[0071] S501 , projecting the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane using rendering technology to generate a two-dimensional CT reconstructed image.

[0072] In some embodiments, this embodiment uses rendering technology to map the three-dimensional Gaussian kernel to a two-dimensional plane to complete the conversion of three-dimensional data into a two-dimensional image.

[0073] In step S401, the sparse point cloud of the chest CT data is denoised using a three-dimensional diffusion model to obtain high-quality sparse point cloud data. Next, the three-dimensional point cloud data is mapped to a two-dimensional plane using 3DGS technology to generate CT image slices.

[0074] When projecting a 3D Gaussian kernel onto a 2D image plane, multiple Gaussian kernels have an impact on the same 2D pixel position, so these impacts need to be weighted averaged. The intensities of multiple Gaussian kernels are mixed at the same position using a weighted average. The weighting coefficient of each Gaussian kernel is related to its own intensity density:

[0075]

[0076] Among them, w i Represents the weight, and the density intensity of the Gaussian kernel Related; represents the contribution of the Gaussian kernel at the projected position.

[0077] In some embodiments, after all three-dimensional Gaussian kernels are projected onto a two-dimensional plane, different z-axis positions are selected as slice depths, multiple slice depths are set, and planes at different depths are projected to obtain a series of CT image slices.

[0078] In the application of chest CT data, these slices represent different anatomical levels, such as slice views of the lungs, chest cavity, heart, etc. The image pixel value on each slice plane is generated by the three-dimensional Gaussian kernel projection of the corresponding depth, and finally a two-dimensional CT reconstructed image is obtained.

[0079] In some preferred embodiments, this embodiment provides a CT image reconstruction method based on a three-dimensional diffusion model, such as Figure 2 As shown, the method includes the following contents:

[0080] S111, using a medical CT scanner to acquire three-dimensional CT voxel data, and preprocessing the three-dimensional CT voxel data;

[0081] S211, extracting a surface point cloud of a medical object from the preprocessed three-dimensional CT voxel data using a voxel-level reconstruction algorithm to generate three-dimensional sparse point cloud data;

[0082] S311, using 3D Gaussian splashing technology to process three-dimensional sparse point cloud data, and using the pre-trained three-dimensional Gaussian kernel to convert it into three-dimensional Gaussian distribution data;

[0083] S411, using a three-dimensional diffusion model to gradually add Gaussian noise to the three-dimensional Gaussian distribution data, and gradually remove the Gaussian noise to generate reconstructed three-dimensional Gaussian distribution data;

[0084] S511, projecting the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane using a rendering technique to generate a two-dimensional CT reconstructed image;

[0085] S611 . Use a two-dimensional diffusion model to gradually add Gaussian noise to the two-dimensional CT reconstructed image, and gradually remove the Gaussian noise to generate an updated two-dimensional CT reconstructed image.

[0086] Among them, the specific implementation of steps S211 to S511 can refer to steps S201 to S501, which will not be repeated in this embodiment. Step S611 will be specifically described below.

[0087] In this embodiment, a two-dimensional diffusion model is used to assist in denoising a two-dimensional CT image, thereby generating a high-quality medical CT image for medical diagnosis and analysis.

[0088] In step S501 and step S511, the 3D Gaussian Splatting technology is used to map the three-dimensional Gaussian kernel to the two-dimensional plane, and two-dimensional CT image slices of different depths are generated. Due to the uncertainty in the Gaussian kernel mapping process and various factors in medical imaging, these CT images may contain noise such as Gaussian noise, artifacts, scanning errors, etc. When the image quality of the two-dimensional CT reconstructed image generated by S501 or S511 is less than the preset threshold, for example, the contrast of the two-dimensional CT reconstructed image, when the contrast exceeds 80%, we believe that the generated two-dimensional CT reconstructed image can meet the demand; when the contrast is less than 80%, we believe that the generated two-dimensional CT reconstructed image needs to continue to use the two-dimensional diffusion model for denoising and reconstruction. Therefore, the generated two-dimensional CT image slice is used as the denoising input and enters the two-dimensional diffusion model for denoising.

[0089] The input noisy CT image slices are denoised by a two-dimensional diffusion model to gradually restore the details and clarity of the image. The diffusion model is trained to identify and remove noise through a forward noise addition and reverse denoising removal process.

[0090] In the training phase of the diffusion model, the forward process generates training samples with different noise levels by gradually adding noise to the noise-free image, so that the model can learn the denoising characteristics. In the denoising phase, the trained two-dimensional diffusion model is used to gradually restore a clear image from the noisy image. Through the step-by-step denoising operation of the reverse diffusion process, the model can gradually eliminate noise and restore image details, retaining the edge details and structural information of the chest CT image.

[0091] Through multiple iterations of the reverse diffusion process, the noise in the image is gradually removed, and the denoising result of each step will be used as the input for the next step. After multiple iterations, the final denoised image I clean Close to the original noise-free medical CT image.

[0092] After denoising, the restored chest CT images no longer contain artifacts and noise, and retain clear edge and structural information, providing reliable imaging support for clinical medicine.

[0093] See also Figure 3 , Figure 3 FIG. 1 is a schematic diagram of the structure of a CT image reconstruction device based on a three-dimensional diffusion model in an embodiment of the present invention. Figure 3 As shown, a CT image reconstruction device based on a three-dimensional diffusion model comprises: a data acquisition module 121, which uses a medical CT scanner to acquire three-dimensional CT voxel data and pre-processes the three-dimensional CT voxel data;

[0094] A voxel reconstruction module 221 extracts a surface point cloud of a medical object from the preprocessed three-dimensional CT voxel data using a voxel-level reconstruction algorithm to generate three-dimensional sparse point cloud data;

[0095] The point cloud reconstruction module 321 uses a 3D Gaussian splashing technique to process the three-dimensional sparse point cloud data and converts it into three-dimensional Gaussian distribution data using a pre-trained three-dimensional Gaussian kernel;

[0096] A diffusion reconstruction module 421 uses a three-dimensional diffusion model to gradually add Gaussian noise to the three-dimensional Gaussian distribution data, and gradually remove the Gaussian noise to generate reconstructed three-dimensional Gaussian distribution data;

[0097] The rendering and reconstruction module 521 uses rendering technology to project the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane to generate a two-dimensional CT reconstructed image.

[0098] The embodiment of the present invention further provides an electronic device, such as Figure 4 The electronic device includes a processor 402, a memory 403, an input unit 404, a display unit 405 and other devices. Those skilled in the art will understand that Figure 4 The device structure device shown does not constitute a limitation on all devices, and may include more or fewer components than shown, or combine certain components. The memory 403 can be used to store the application 401 and various functional modules, and the processor 402 runs the application 401 stored in the memory 403, thereby executing various functional applications and data processing of the device. The memory can be an internal memory or an external memory, or include both internal and external memories. The internal memory may include a read-only memory (ROM), a programmable ROM (PROM), an electrically programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a flash memory, or a random access memory. The external memory may include a hard disk, a floppy disk, a ZIP disk, a U disk, a magnetic tape, etc. The memory disclosed in the present invention includes but is not limited to these types of memories. The memory disclosed in the present invention is only used as an example and not as a limitation.

[0099] An embodiment of the present invention provides a computer-readable storage medium, on which an application is stored, and when the program is executed by a processor, a CT image reconstruction method based on a three-dimensional diffusion model of any one of the above embodiments is implemented. The computer-readable storage medium includes, but is not limited to, any type of disk (including floppy disk, hard disk, optical disk, CD-ROM, and magneto-optical disk), ROM (Read-Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic card or optical card. That is, the storage device includes any medium that can store or transmit information in a readable form by a device (for example, a computer, a mobile phone), which can be a read-only memory, a disk or an optical disk, etc.

[0100] The embodiment of the present invention also provides a computer application program, which runs on a computer, and the computer application program is used to execute a CT image reconstruction method based on a three-dimensional diffusion model of any of the above embodiments. Although the embodiments of the present invention have been shown and described, it is understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

[0101] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0102] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A CT image reconstruction method based on a three-dimensional diffusion model, characterized in that: The following steps are involved: Acquiring three-dimensional CT voxel data using a medical CT scanner, and preprocessing the three-dimensional CT voxel data; The voxel-level reconstruction algorithm is used to extract the surface point cloud of the medical object from the preprocessed 3D CT voxel data to generate 3D sparse point cloud data; The 3D Gaussian splashing technology is used to process the 3D sparse point cloud data, and the pre-trained 3D Gaussian kernel is used to convert it into 3D Gaussian distribution data; Using a three-dimensional diffusion model, gradually adding Gaussian noise to the three-dimensional Gaussian distribution data, and gradually removing the Gaussian noise, to generate reconstructed three-dimensional Gaussian distribution data; The rendering technology is used to project the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane to generate a two-dimensional CT reconstructed image.

2. The CT image reconstruction method based on the three-dimensional diffusion model according to claim 1, characterized in that: The position encoding vector of the three-dimensional Gaussian kernel is processed by frequency regularization.

3. The CT image reconstruction method based on the three-dimensional diffusion model according to claim 2, characterized in that: The position encoding vector of the three-dimensional Gaussian kernel is expressed as: in, Represents the position encoding vector of the three-dimensional Gaussian kernel at (x, y, z), f is the regularization frequency factor, which is used to control the frequency; λ is the regularization weight, which is used to control the degree of smoothness; ‖P‖ represents the modulus of the position encoding vector, which is used to adjust the regularization effect of different positions.

4. The CT image reconstruction method based on the three-dimensional diffusion model according to claim 3, characterized in that: The regularization frequency factor is a high-frequency frequency factor, and the high-frequency frequency factor controls the high-frequency component to adjust the position encoding vector.

5. The CT image reconstruction method based on the three-dimensional diffusion model according to claim 1, characterized in that: The step of gradually adding Gaussian noise to the three-dimensional Gaussian distribution data and gradually removing the Gaussian noise using the three-dimensional diffusion model to generate reconstructed three-dimensional Gaussian distribution data comprises: In the forward noise addition process, Gaussian noise is independently added to the position encoding vector, covariance matrix and color eigenvalue of the three-dimensional Gaussian distribution data to generate three-dimensional noisy point cloud data; In the reverse denoising process, Gaussian noise is removed independently on the position encoding vector, covariance matrix and color eigenvalue of the three-dimensional noisy point cloud data to generate reconstructed three-dimensional Gaussian distribution data.

6. The CT image reconstruction method based on the three-dimensional diffusion model according to claim 5, characterized in that: The step of independently adding Gaussian noise to the position encoding vector, the covariance matrix, and the color eigenvalue of the three-dimensional Gaussian distribution data includes: For the position encoding vector of three-dimensional Gaussian distribution data, an independent first Gaussian noise is added at each position; For the covariance matrix of three-dimensional Gaussian distribution data, add independent second Gaussian noise to each matrix; For the color features of three-dimensional Gaussian distribution data, an independent third Gaussian noise is added to each color channel; The first Gaussian noise, the second Gaussian noise and the third Gaussian noise are identically distributed.

7. The CT image reconstruction method based on the three-dimensional diffusion model according to claim 1, characterized in that: After the rendering technology is used to project the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane to generate a two-dimensional CT reconstructed image, the following steps are also included: It is determined whether the image quality of the two-dimensional CT reconstructed image exceeds a preset threshold. If it does not exceed the preset threshold, a two-dimensional diffusion model is used to gradually add Gaussian noise to the two-dimensional CT reconstructed image, and the Gaussian noise is gradually removed to generate an updated two-dimensional CT reconstructed image.

8. A CT image reconstruction device based on a three-dimensional diffusion model, characterized in that: The device comprises: A data acquisition module, which uses a medical CT scanner to acquire three-dimensional CT voxel data and pre-processes the three-dimensional CT voxel data; The voxel reconstruction module uses a voxel-level reconstruction algorithm to extract the surface point cloud of the medical object from the preprocessed 3D CT voxel data and generate 3D sparse point cloud data; The point cloud reconstruction module uses 3D Gaussian splashing technology to process 3D sparse point cloud data and converts it into 3D Gaussian distribution data using pre-trained 3D Gaussian kernel; A diffusion reconstruction module, which uses a three-dimensional diffusion model to gradually add Gaussian noise to the three-dimensional Gaussian distribution data, and gradually removes the Gaussian noise to generate reconstructed three-dimensional Gaussian distribution data; The rendering and reconstruction module uses rendering technology to project the reconstructed three-dimensional Gaussian distribution data onto a two-dimensional plane to generate a two-dimensional CT reconstructed image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images