Integrated view synthesis based on 3DGS and sparse view CT reconstruction method

Abstract

The application discloses an integrated view synthesis and sparse view CT reconstruction method based on 3DGS, which comprises the following steps: establishing internal and external parameter matrices through X-ray scanning parameters, combining sparse view projection data, and constructing an initial voxel space and initializing a 3DGS radiation field based on an ACUI strategy; modeling anisotropic radiation intensity by using a radiation intensity response function; generating virtual projection by using volume consistency rendering. The optimized radiation field is converted into a voxel grid, and density contribution is applied for voxel space reconstruction to obtain a reconstructed voxel space and generate a new 3DGS radiation field. The new 3DGS radiation field is used to re-render virtual view projection, and iterative optimization is carried out by using a loss function until a preset iteration number is reached to output a new view synthesis image and a CT reconstruction body. An efficient process of one-time training and double-output is realized. The efficiency and precision of X-ray imaging are improved, the radiation dose of patients is reduced, and a higher-quality imaging solution is provided for medical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of CT image technology, and in particular to an integrated viewpoint synthesis and sparse viewpoint CT reconstruction method based on 3DGS. Background Technology

[0002] X-ray novel viewpoint synthesis (NVS) aims to generate X-ray projections of objects from previously uncaptured viewpoints using only existing projections scanned from different angles. X-rays are widely used in medical imaging due to their strong penetrating power and ability to capture the internal structure of objects being imaged. However, X-rays are harmful to the human body due to their powerful ionizing radiation, especially with increased X-ray doses. Improving NVS technology helps reduce X-ray exposure and provides physicians and downstream tasks such as CT reconstruction with a comprehensive view of the imaging site. Therefore, X-ray NVS is very important and valuable.

[0003] Existing methods are primarily based on Neural Radiation Fields (NeRF) and 3D Gaussian Splatting (3DGS). NeRF typically uses a multilayer perceptron (MLP) to learn the mapping from point locations to radio density, then creates projections by volume rendering along the rays. This ray-tracing scheme is very time-consuming because it requires sampling many 3D points and then calculating them for each ray, slowing down the training and inference processes. Even the most efficient NeRF-based methods recently still require more than an hour of training time, increasing patient and physician wait times and leading to inefficient diagnosis. 3DGS represents the scene as an explicit 3D Gaussian set, quickly synthesizing new perspectives by projecting the Gaussian onto the image plane and performing splatting. However, this two-dimensional projection approximately ignores the density integral along the ray direction, resulting in blurred edges, surface light leakage, and requiring SfM initialization, making it prone to failure in low-contrast X-ray images. When using the projection generated by 3DGS for sparse-view CT reconstruction, the reconstructed volume still exhibits striated artifacts and structural distortion, failing to meet clinical diagnostic needs.

[0004] In clinical practice, to reduce the X-ray dose received by patients and shorten scan time, only a very small number of images (e.g., less than 25) are often acquired, far fewer than the hundreds required for conventional CT. This "sparse sampling" leads to severe stripe artifacts, edge blurring, and density distortion in traditional analytical algorithms (FDK) or iterative algorithms (SART). How to fully exploit structural priors and achieve high-fidelity three-dimensional density volume reconstruction within limited projections has become a core bottleneck in low-dose CT and intraoperative real-time imaging.

[0005] Therefore, a method based on 3DGS for integrated view synthesis and sparse view CT reconstruction is proposed to improve the efficiency and accuracy of X-ray imaging, reduce patient radiation dose, and provide higher quality imaging solutions for medical diagnosis and industrial inspection. Summary of the Invention

[0006] This invention overcomes the shortcomings of existing technologies and provides an integrated viewpoint synthesis and sparse viewpoint CT reconstruction method based on 3DGS. Its important purpose is to improve the efficiency and accuracy of X-ray imaging, reduce patient radiation dose, and provide higher quality imaging solutions for medical diagnosis and industrial inspection.

[0007] To achieve the above objectives, the first aspect of this invention provides a method for integrated viewpoint synthesis and sparse viewpoint CT reconstruction based on 3DGS, comprising:

[0008] An intrinsic and extrinsic parameter matrix is ​​established using X-ray scanning equipment parameters. Raw CT projection data under sparse viewpoints are obtained. An initial voxel space is constructed using an angle-attitude cuboid uniform initialization strategy, and initial 3DGS radiation field modeling is performed using the initial voxel space.

[0009] The initial 3DGS radiation field is imported into the 3D Gaussian rendering framework for volumetric consistency rendering to generate a virtual perspective composite image. The initial 3DGS radiation field is then combined with the original CT projection data to optimize its spatial structure and generate a spatially optimized 3DGS radiation field.

[0010] Based on the spatially optimized 3DGS radiation field, the anisotropic radiation intensity is modeled by a learnable feature vector using the radiation intensity response function. The feature weights are iteratively updated through backpropagation to output the physically optimized 3DGS radiation field.

[0011] The center position, covariance matrix and anisotropic radiation intensity value of each 3D Gaussian are extracted by the physical optimization 3DGS radiation field. The Gaussian distribution is mapped to the voxel grid and a density contribution is applied to reconstruct the voxel space, generating the reconstructed voxel space.

[0012] The reconstructed voxel space is converted back into a new 3DGS radiation field, and the virtual view projection is re-rendered based on the new 3DGS radiation field. The loss function is used for iterative optimization until a preset number of iterations are reached to output a new view synthesized image and CT reconstruction.

[0013] In this scheme, the step of establishing an intrinsic and extrinsic parameter matrix through X-ray scanning equipment parameters, obtaining raw CT projection data under sparse viewpoints, constructing an initial voxel space using an angle-attitude cuboid uniform initialization strategy, and performing initial 3DGS radiation field modeling using the initial voxel space specifically includes:

[0014] Obtain X-ray scanning equipment parameters, including the distance from the X-ray source to the object, the distance from the X-ray source to the detector, the detector size, and the rotation angle sequence corresponding to different viewing angles;

[0015] The X-ray source rotation trajectory is encoded by trigonometric function relationships, and the scale of the cone beam projection is defined by distance parameters. This generates an external parameter matrix describing the camera position and attitude and an internal parameter matrix describing the imaging characteristics, in order to establish a mapping relationship between the three-dimensional coordinate system and the two-dimensional detection plane.

[0016] Obtain the original CT projection data under sparse viewpoint, and use the generated intrinsic and extrinsic parameter matrix to perform geometric correction on the original CT projection data; construct a cuboid with size S1×S2×S3 to completely surround the target object, and use a mesh of size M1×M2×M3 to divide the constructed cuboid to generate the initial voxel space.

[0017] Subsequently, a set of three-dimensional coordinate points is obtained by uniform sampling at intervals of d∈R in the initial voxel space. The density value of each sampling point is extracted. If the density value is greater than the preset density threshold, the 3D Gaussian function is initialized with the corresponding position coordinate as the center of the target voxel unit, and finally the initial 3DGS radiation field is generated.

[0018] Each Gaussian function is defined by a covariance matrix to determine its spatial distribution range. The initial direction is estimated by the voxel gradient, the initial opacity is obtained by linear mapping of the voxel values, and the radiation color attribute is assigned a grayscale value based on a preset material library.

[0019] In this solution, the step of importing the initial 3DGS radiation field into a 3D Gaussian rendering framework for volumetric consistency rendering to generate a virtual perspective composite image specifically includes:

[0020] The initial 3DGS radiation field is imported into the 3D Gaussian rendering framework to generate an imaging plane, in which a ray is constructed for each pixel, originating from the ray source and passing through the center of the pixel;

[0021] Perform volumetric consistency rendering on each ray, detect all 3D Gaussian primitives that intersect with the target ray and sort them from front to back to ensure that the integration order conforms to the real light propagation direction;

[0022] Subsequently, each three-dimensional Gaussian is projected onto the ray direction to obtain a one-dimensional Gaussian function. Its mean and scale along the ray direction are calculated. Based on the maximum value and scale of the Gaussian along the ray direction, the transmittance of each Gaussian is calculated using analytical integral.

[0023] The calculated transmittance is converted into opacity. Based on the forward alpha blending model, the pixel colors are synthesized using all Gaussian colors and alpha values ​​to generate a virtual perspective composite image.

[0024] In this solution, the calculation method for transmittance and opacity is specifically as follows:

[0025] The specific formula for calculating the transmittance of each Gaussian using analytical integrals is as follows:

[0026]

[0027] Among them, G j (γ j d) is the maximum value of the Gaussian distribution along the direction d of the light ray; γ j It is the mean of a 1D Gaussian distribution, representing the position of the center of the Gaussian distribution on the light ray; β j κ represents the variance of the 1D Gaussian distribution and the width of the Gaussian distribution. j It is the weight of the j-th Gaussian volume element;

[0028] The specific formula for converting the calculated transmittance of light through the Gaussian region into opacity is as follows:

[0029]

[0030] in, For transmittance, α i For transparency.

[0031] In this scheme, the spatial structure optimization of the initial 3DGS radiation field based on the original CT projection data to generate a spatially optimized 3DGS radiation field specifically includes:

[0032] Obtain geometrically corrected original CT projection data, generate a real projection image based on the geometrically corrected original CT projection data, and calculate the L1 norm difference of pixel-by-pixel gray values ​​between the real projection image and the virtual viewpoint synthesized image as an intensity fidelity term.

[0033] A multi-scale SSIM algorithm is introduced to measure the matching degree between the edge and the texture within the local window, which is used as a structural similarity term. A dual-constraint loss function is constructed based on the intensity fidelity term and the structural similarity term.

[0034] The initial 3DGS radiation field is optimized based on the dual-constraint loss function. The gradient is calculated using backpropagation, and the 3D Gaussian parameters are dynamically optimized using the stochastic gradient descent algorithm. Finally, the spatially optimized 3DGS radiation field is output.

[0035] During the optimization process, the opacity of each Gaussian point is compared with the preset opacity threshold. When the opacity is less than the preset opacity threshold, the corresponding Gaussian point is identified as a low-contribution Gaussian point and pruning is performed.

[0036] In this scheme, the spatially optimized 3DGS radiation field is modeled using a radiation intensity response function through learnable feature vectors to model anisotropic radiation intensity. The feature weights are iteratively updated via backpropagation to output a physically optimized 3DGS radiation field. Specifically, this includes:

[0037] The spatially optimized 3DGS radiation field is obtained, the spatial location and covariance parameters of each Gaussian point cloud are extracted, the anatomical tissue density region to which the Gaussian point belongs is located based on the parameters, and an independent learnable feature vector is assigned to each Gaussian point cloud according to the preset density-weight mapping rule.

[0038] A radiation intensity response function is introduced to perform a nonlinear transformation on the learnable feature vector. The Sigmoid activation function is used to compress the feature values ​​to the standard radiation intensity range. A constant weight vector is applied simultaneously to constrain the physical contribution ratio of the feature components, thereby generating viewpoint-independent radiation intensity values.

[0039] During the optimization process, a physical consistency loss term is constructed based on the difference gradient between the virtual synthetic projection and the real CT projection, combined with the prior knowledge of X-ray anisotropic attenuation, and the feature vector weights are updated through backpropagation.

[0040] By iteratively optimizing and gradually suppressing non-physically related feature components, the radiation intensity field converges to the optimal solution that conforms to the attenuation physical model, and outputs a physically optimized 3DGS radiation field with enhanced anisotropic properties.

[0041] The constant weight vector is initialized based on the physical prior of tissue density and is fixed throughout, ensuring that the radiation intensity is determined only by the spatial location density; the X-ray anisotropic attenuation prior is that the radiation intensity at the same spatial point is constant under any viewing angle.

[0042] In this scheme, the process of converting the physically optimized 3DGS radiation field into an explicit voxel representation, mapping the Gaussian distribution onto a high-resolution voxel mesh, and applying CT physical constraints to update the voxel space to generate a reconstructed voxel space specifically includes:

[0043] Based on the physical optimization of the 3DGS radiation field, the center position, covariance matrix and anisotropic radiation intensity value of each 3D Gaussian are extracted. The density contribution of each Gaussian distribution to the voxel unit in its local space is applied. The continuous radiation field is mapped to the initial three-dimensional voxel density field by the approximate integration of the Gaussian volume function to generate the first voxel representation.

[0044] A differentiable X-ray projection operator is constructed. The theoretical projection value is calculated by integrating each projection ray in the voxel space and comparing it pixel by pixel with the real X-ray projection under the actual sparse angle. The projection data fidelity term is calculated as the reconstruction loss.

[0045] Subsequently, based on the construction of a differentiable X-ray projection operator, and combined with the difference between the projection image from the real sparse viewpoint and the simulated projection result, a joint optimization objective function such as projection consistency loss is constructed. The voxel density is iteratively updated through differentiable optimization methods such as gradient descent to obtain a dense three-dimensional reconstruction result, and finally a reconstructed voxel space is generated.

[0046] In this scheme, the reconstructed voxel space is converted back into a new 3DGS radiation field, and the virtual view projection is re-rendered based on the new 3DGS radiation field. Iterative optimization is performed using a loss function until a preset number of iterations are reached to output a new view synthesized image and CT reconstructed volume. Specifically, this includes:

[0047] The reconstructed voxel space is obtained, and significant structural regions are screened based on a preset density threshold. Within the selected region, the voxel grid is traversed and high gradient change points are extracted as candidate centers. The principal axis direction of the covariance matrix is ​​calculated based on the three-dimensional gradient direction field of the candidate point neighborhood, and principal component analysis is used to determine the structural extension direction.

[0048] By using the principal axis of the covariance matrix to linearly map voxel values ​​to the radiation intensity range and associating them with the corresponding anisotropy properties, a new 3DGS radiation field is finally generated.

[0049] Based on the new 3DGS radiation field, volumetric consistency rendering is performed. A new synthetic virtual projection image is generated by analytical integration along the target viewpoint ray. Combined with the loss function, radiation field optimization and voxel constraint reconstruction are alternately iterated until the preset number of iterations is reached to output a new viewpoint synthetic image and CT reconstruction volume.

[0050] During the alternating iteration process, the forward optimization phase updates the Gaussian position and covariance parameters through stochastic gradient descent based on the projection error loss, while the reverse constraint phase strengthens edge continuity and suppresses outlier noise points by reapplying total variational regularization and projection fidelity constraints.

[0051] A second aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a 3DGS-based integrated view synthesis and sparse view CT reconstruction method program, wherein when the 3DGS-based integrated view synthesis and sparse view CT reconstruction method program is executed by a processor, it implements the steps of the 3DGS-based integrated view synthesis and sparse view CT reconstruction method as described in any of the preceding claims.

[0052] The purpose of this invention is to address the problems existing in the prior art, including inefficient and error-prone camera calibration, projection distortion during rendering, and viewpoint coupling issues during modeling. By proposing an integrated new viewpoint synthesis and sparse viewpoint CT reconstruction method based on 3DGS, this invention employs rapid initialization without SfM, volume-consistent rendering, and anisotropic density modeling. Through these improvements, this invention aims to enhance the efficiency and accuracy of X-ray imaging, reduce patient radiation dose, and provide higher-quality imaging solutions for medical diagnosis and industrial inspection. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0054] Figure 1 A flowchart of an integrated view synthesis and sparse view CT reconstruction method based on 3DGS is provided as an embodiment of the present invention.

[0055] Figure 2 The flowchart illustrates an optimized method for integrated view synthesis and sparse view CT reconstruction according to an embodiment of the present invention.

[0056] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0057] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0059] Figure 1 A flowchart of an integrated view synthesis and sparse view CT reconstruction method based on 3DGS is provided as an embodiment of the present invention.

[0060] like Figure 1 As shown, this invention provides a flowchart of an integrated viewpoint synthesis and sparse viewpoint CT reconstruction method based on 3DGS, including:

[0061] S102, establish the intrinsic and extrinsic parameter matrix through X-ray scanning equipment parameters, obtain the original CT projection data under sparse view, construct the initial voxel space using the angle-attitude cuboid uniform initialization strategy, and perform initial 3DGS radiation field modeling through the initial voxel space.

[0062] S104, the initial 3DGS radiation field is imported into the 3D Gaussian rendering framework for volumetric consistency rendering to generate a virtual perspective composite image, and the initial 3DGS radiation field is spatially optimized by combining the original CT projection data to generate a spatially optimized 3DGS radiation field.

[0063] S106, based on spatially optimized 3DGS radiation field, uses the radiation intensity response function to model anisotropic radiation intensity through learnable feature vectors, iteratively updates feature weights through backpropagation, and outputs physically optimized 3DGS radiation field.

[0064] S108, by extracting the center position, covariance matrix and anisotropic radiation intensity value of each 3D Gaussian through the physical optimization 3DGS radiation field, the Gaussian distribution is mapped to the voxel grid, and density contribution is applied to reconstruct the voxel space to generate the reconstructed voxel space.

[0065] S110, the reconstructed voxel space is converted into a new 3DGS radiation field again, and the virtual view projection is re-rendered based on the new 3DGS radiation field. The loss function is used for iterative optimization until the preset number of iterations is reached to output the new view synthesized image and CT reconstruction.

[0066] Furthermore, in a preferred embodiment of the present invention, the step of establishing an intrinsic and extrinsic parameter matrix through X-ray scanning equipment parameters, obtaining raw CT projection data under sparse viewing angles, constructing an initial voxel space using an angle-attitude cuboid uniform initialization strategy, and performing initial 3DGS radiation field modeling using the initial voxel space specifically includes:

[0067] Obtain X-ray scanning equipment parameters, including the distance from the X-ray source to the object, the distance from the X-ray source to the detector, the detector size, and the rotation angle sequence corresponding to different viewing angles;

[0068] The X-ray source rotation trajectory is encoded by trigonometric function relationships, and the scale of the cone beam projection is defined by distance parameters. This generates an external parameter matrix describing the camera position and attitude and an internal parameter matrix describing the imaging characteristics, in order to establish a mapping relationship between the three-dimensional coordinate system and the two-dimensional detection plane.

[0069] Obtain the original CT projection data under sparse viewpoint, and use the generated intrinsic and extrinsic parameter matrix to perform geometric correction on the original CT projection data; construct a cuboid with size S1×S2×S3 to completely surround the target object, and use a mesh of size M1×M2×M3 to divide the constructed cuboid to generate the initial voxel space.

[0070] Subsequently, a set of three-dimensional coordinate points is obtained by uniform sampling at intervals of d∈R in the initial voxel space. The density value of each sampling point is extracted. If the density value is greater than the preset density threshold, the 3D Gaussian function is initialized with the corresponding position coordinate as the center of the target voxel unit, and finally the initial 3DGS radiation field is generated.

[0071] Each Gaussian function is defined by a covariance matrix to determine its spatial distribution range. The initial direction is estimated by the voxel gradient, the initial opacity is obtained by linear mapping of the voxel values, and the radiation color attribute is assigned a grayscale value based on a preset material library.

[0072] It should be noted that existing methods rely on SfM or COLMAP for camera calibration, which is unsuitable for X-ray imaging for two reasons. First, X-ray images are grayscale and have low contrast. Second, different layers of an object may be re-viewed at the same location in the projection. These two issues reduce the accuracy of feature detection and matching in SfM. Furthermore, running the SfM algorithm typically takes a long time, which prolongs the training process. To address these issues, this invention customizes an angle-attitude cuboid uniform initialization (ACUI) strategy for conical beam X-ray scanning scenarios. First, the physical parameters of the X-ray scanning equipment are obtained, including the distance from the X-ray source to the object (SOD), the distance from the X-ray source to the detector (SDD), the detector pixel size, and the rotation angle sequence corresponding to each viewpoint. The circular motion trajectory of the X-ray source around the object is encoded using trigonometric functions. The geometrical proportions of the conical beam projection are defined by combining the SOD and SDD parameters, generating an external parameter matrix describing the camera pose and an internal parameter matrix describing the imaging characteristics.

[0073] The next step for ACUI is to initialize the center position of the 3D Gaussian function. Although the precise shape of the scanned object is not given initially, the scan space can be approximated. Therefore, a cuboid of size S1×S2×S3 (mm) is constructed to completely enclose the object. The center of this cuboid is also the center of the object and the origin of the world coordinate system. A grid of size M1×M2×M3 (voxels) is used to divide this cuboid, and points are uniformly sampled within the grid at intervals d∈R. Using the aforementioned intrinsic and extrinsic parameter matrices, geometric correction is performed on the original CT projection data based on a sparse viewpoint: affine transformations are used to eliminate projection coordinate offsets and scale distortions caused by equipment assembly errors or mechanical vibrations, outputting position-calibrated projection data. Subsequently, an initial low-resolution voxel space is generated based on the corrected projection data. Finally, within the initial voxel space, all voxel units are traversed, and a set of three-dimensional coordinate points is obtained through uniform sampling at intervals d∈R. For each sampled point, its density value is extracted; if the density value is greater than a preset threshold, a 3D Gaussian function is initialized centered on that point.

[0074] During initialization: the covariance matrix is ​​estimated from the gradient vectors of the voxels in the neighborhood of the sampling point to determine the principal direction, controlling the spatial extension shape of the Gaussian distribution; the initial opacity maps the density value to the [0,1] interval through a linear function, characterizing the material's ability to attenuate X-rays; the radiation color attribute is assigned a grayscale value according to a preset CT value-material mapping table (e.g., high grayscale for bones, low grayscale for air). This ultimately forms an initial 3DGS radiation field that explicitly characterizes the object's geometric structure and radiation physical properties.

[0075] Furthermore, in a preferred embodiment of the present invention, the step of importing the initial 3DGS radiation field into a 3D Gaussian rendering framework for volumetric consistency rendering to generate a virtual perspective composite image specifically includes:

[0076] The initial 3DGS radiation field is imported into the 3D Gaussian rendering framework to generate an imaging plane, in which a ray is constructed for each pixel, originating from the ray source and passing through the center of the pixel;

[0077] Perform volumetric consistency rendering on each ray, detect all 3D Gaussian primitives that intersect with the target ray and sort them from front to back to ensure that the integration order conforms to the real light propagation direction;

[0078] Subsequently, each three-dimensional Gaussian is projected onto the ray direction to obtain a one-dimensional Gaussian function. Its mean and scale along the ray direction are calculated. Based on the maximum value and scale of the Gaussian along the ray direction, the transmittance of each Gaussian is calculated using analytical integral.

[0079] The calculated transmittance is converted into opacity. Based on the forward alpha blending model, the pixel colors are synthesized using all Gaussian colors and alpha values ​​to generate a virtual perspective composite image.

[0080] It should be noted that this invention, in the volumetric rendering process, no longer relies on the heuristic approximation method in traditional 3DGS, but instead introduces the analytical solution of the volumetric rendering equation into the 3D Gaussian rendering framework. By directly integrating the density field onto the light rays, the contribution of each 3D Gaussian primitive to the pixel color is accurately calculated, thereby achieving physically consistent rendering. This method abandons several approximation assumptions in traditional splatting rendering, making the rendering results more accurate, and is particularly suitable for dense, opaque scenes or tasks such as medical imaging. First, the initial 3DGS radiation field is loaded into the rendering framework to construct the imaging plane (i.e., a two-dimensional pixel matrix) of the target viewpoint. For each pixel on this plane, a ray is extended from the X-ray source location, through the pixel center, and into the scene, simulating the propagation path of X-rays in physical space. Volumetric consistency rendering is performed on each ray: all 3D Gaussian primitives intersecting with the ray are detected, and the intersecting Gaussian primitives are strictly ordered from near to far according to the direction of ray propagation. This order ensures that the subsequent integration order conforms to the physical process of X-rays actually penetrating objects, that is, starting from the ray source, penetrating tissues of different depths in sequence. Then, physically precise integral calculations are performed: each 3D Gaussian unit is projected along the ray direction, transforming it into a 1D Gaussian function (i.e., the density distribution along the ray direction). The statistical properties of this 1D Gaussian function (mean location and distribution scale) are analytically calculated, and based on the maximum density value and distribution range of the Gaussian function, the transmittance of light passing through this Gaussian region is solved using closed-form analytical integration. The transmittance is then converted into an opacity value, characterizing the degree to which the ray is absorbed in this Gaussian region. Finally, based on the forward alpha blending model: the color contribution (grayscale value) of all intersecting Gaussian units is accumulated along the ray direction along the product of the corresponding opacity, while considering the occlusion effect of the Gaussian units in front on the light rays behind, to synthesize the final color value of the pixel. Although an analytical solution for volumetric rendering is used in this process, the entire workflow is implemented using a GPU-friendly rasterization method, balancing rendering speed and quality.

[0081] Furthermore, in a preferred embodiment of the present invention, the calculation method for transmittance and opacity is specifically as follows:

[0082] The specific formula for calculating the transmittance of each Gaussian using analytical integrals is as follows:

[0083]

[0084] Among them, G j (γ j d) is the maximum value of the Gaussian distribution along the direction d of the light ray; γj It is the mean of a 1D Gaussian distribution, representing the position of the center of the Gaussian distribution on the light ray; β j κ represents the variance of the 1D Gaussian distribution and the width of the Gaussian distribution. j It is the weight of the j-th Gaussian volume element;

[0085] The specific formula for converting the calculated transmittance of light through the Gaussian region into opacity is as follows:

[0086]

[0087] in, For transmittance, α i For transparency.

[0088] Furthermore, in a preferred embodiment of the present invention, the step of optimizing the spatial structure of the initial 3DGS radiation field by combining the original CT projection data to generate a spatially optimized 3DGS radiation field specifically includes:

[0089] Obtain geometrically corrected original CT projection data, generate a real projection image based on the geometrically corrected original CT projection data, and calculate the L1 norm difference of pixel-by-pixel gray values ​​between the real projection image and the virtual viewpoint synthesized image as an intensity fidelity term.

[0090] A multi-scale SSIM algorithm is introduced to measure the matching degree between the edge and the texture within the local window, which is used as a structural similarity term. A dual-constraint loss function is constructed based on the intensity fidelity term and the structural similarity term.

[0091] The initial 3DGS radiation field is optimized based on the dual-constraint loss function. The gradient is calculated using backpropagation, and the 3D Gaussian parameters are dynamically optimized using the stochastic gradient descent algorithm. Finally, the spatially optimized 3DGS radiation field is output.

[0092] During the optimization process, the opacity of each Gaussian point is compared with the preset opacity threshold. When the opacity is less than the preset opacity threshold, the corresponding Gaussian point is identified as a low-contribution Gaussian point and pruning is performed.

[0093] It should be noted that traditional rendering methods, due to insufficient optimization of geometric and physical properties, result in systematic deviations between the synthesized image and the true projection, failing to meet the dual requirements of structural fidelity and radiometric accuracy in clinical diagnosis. Therefore, to improve the consistency between the virtual projection image and the true CT projection, geometrically corrected sparse-view original CT projection data is used as the true projection image. Simultaneously, it is compared with a virtual-view synthesized image generated through volumetric rendering based on the initial 3DGS radiation field to optimize spatial distribution.

[0094] The intensity fidelity term is calculated by using the L1 norm difference of pixel-by-pixel grayscale values ​​between the real projection and the virtual synthetic image. This term quantifies the deviation between the two in the global grayscale distribution, forcing the synthetic image to approximate the real data in terms of radiation intensity (such as high attenuation areas of bone and low attenuation areas of lungs). A multi-scale SSIM (structural similarity) algorithm is introduced to analyze edge sharpness, structural contrast, and detail matching within a local sliding window to generate a structural similarity term. This term evaluates the accuracy of local anatomical feature restoration, compensating for the insensitivity of the L1 norm to structural distortion, and constructs a dual-constraint loss function by combining the intensity fidelity term and the structural similarity term. Based on this loss function, the gradient of the 3D Gaussian parameters (including position, covariance, opacity, and radiation intensity) is calculated using backpropagation. The parameters are dynamically adjusted using the stochastic gradient descent (SGD) algorithm: through iterative updates, the Gaussian position is made to better fit the anatomical structure, the covariance matrix accurately describes the density diffusion range, and the opacity is made to better match the local attenuation characteristics, ultimately outputting a spatially optimized 3DGS radiation field. Dynamic pruning is performed simultaneously during the optimization process: the opacity value of each Gaussian point is monitored, and if it is lower than the preset threshold, it is determined to be a low contribution point and removed from the radiation field, thereby enhancing the expressive power of the effective structure.

[0095] Furthermore, in a preferred embodiment of the present invention, the step of optimizing the spatial 3DGS radiation field, which involves modeling anisotropic radiation intensity using a radiation intensity response function through learnable feature vectors, iteratively updating feature weights through backpropagation, and outputting a physically optimized 3DGS radiation field, specifically includes:

[0096] The spatially optimized 3DGS radiation field is obtained, the spatial location and covariance parameters of each Gaussian point cloud are extracted, the anatomical tissue density region to which the Gaussian point belongs is located based on the parameters, and an independent learnable feature vector is assigned to each Gaussian point cloud according to the preset density-weight mapping rule.

[0097] A radiation intensity response function is introduced to perform a nonlinear transformation on the learnable feature vector. The Sigmoid activation function is used to compress the feature values ​​to the standard radiation intensity range. A constant weight vector is applied simultaneously to constrain the physical contribution ratio of the feature components, thereby generating viewpoint-independent radiation intensity values.

[0098] During the optimization process, a physical consistency loss term is constructed based on the difference gradient between the virtual synthetic projection and the real CT projection, combined with the prior knowledge of X-ray anisotropic attenuation, and the feature vector weights are updated through backpropagation.

[0099] By iteratively optimizing and gradually suppressing non-physically related feature components, the radiation intensity field converges to the optimal solution that conforms to the attenuation physical model, and outputs a physically optimized 3DGS radiation field with enhanced anisotropic properties.

[0100] The constant weight vector is initialized based on the physical prior of tissue density and is fixed throughout, ensuring that the radiation intensity is determined only by the spatial location density; the X-ray anisotropic attenuation prior is that the radiation intensity at the same spatial point is constant under any viewing angle.

[0101] It is important to note that the core objective of this step is to completely decouple the radiation properties of the spatially optimized 3D Gaussian radiation field from the viewing direction. By replacing the spherical harmonic function with density, the influence of the viewing direction is eliminated, better aligning with the anisotropic physical characteristics of X-ray penetration imaging and accelerating convergence. Furthermore, it generates a radiation intensity distribution that strictly follows the X-ray attenuation law based on physically driven modeling. First, the spatial coordinates and covariance parameters of each Gaussian point cloud are extracted to locate the anatomical tissue density region to which the Gaussian point belongs. According to a pre-defined density-weight mapping rule, an independent learnable feature vector is assigned to each Gaussian point cloud, which serves as the fundamental parameter for radiation intensity. A radiation intensity response function (RIRF) is introduced to achieve a nonlinear transformation of physical constraints: the sigmoid activation function is used to compress the feature vector values ​​to the standard radiation intensity range, while a constant weight vector is simultaneously applied—this vector is initialized and fixed throughout based on prior knowledge of X-ray attenuation in different tissues (e.g., the bone attenuation coefficient μ = 0.16 / cm), forcing the contribution ratio of the feature components to conform to physical laws, ensuring that the output radiation intensity is determined solely by the spatial location density, completely eliminating the influence of the viewing direction. During the optimization process, a physical consistency loss term is constructed based on the difference gradient between the virtual synthetic projection and the real CT projection: adhering to the core prior of anisotropic X-ray attenuation, it penalizes radiation intensity fluctuations between different viewpoints. Through backpropagation, only the weights of the learnable feature vectors are updated (constant weight vectors remain unchanged), gradually suppressing feature components unrelated to physical laws. After multiple iterations, the radiation intensity field converges to the optimal solution, and the final output physical-optimized 3DGS radiation field provides a viewpoint-bias-free density field foundation for subsequent voxel reconstruction.

[0102] In this scheme, the process of converting the physically optimized 3DGS radiation field into an explicit voxel representation, mapping the Gaussian distribution onto a high-resolution voxel mesh, and applying CT physical constraints to update the voxel space to generate a reconstructed voxel space specifically includes:

[0103] Based on the physical optimization of the 3DGS radiation field, the center position, covariance matrix and anisotropic radiation intensity value of each 3D Gaussian are extracted. The density contribution of each Gaussian distribution to the voxel unit in its local space is applied. The continuous radiation field is mapped to the initial three-dimensional voxel density field by the approximate integration of the Gaussian volume function to generate the first voxel representation.

[0104] A differentiable X-ray projection operator is constructed. The theoretical projection value is calculated by integrating each projection ray in the voxel space and comparing it pixel by pixel with the real X-ray projection under the actual sparse angle. The projection data fidelity term is calculated as the reconstruction loss.

[0105] Subsequently, based on the construction of a differentiable X-ray projection operator, and combined with the difference between the projection image from the real sparse viewpoint and the simulated projection result, a joint optimization objective function such as projection consistency loss is constructed. The voxel density is iteratively updated through differentiable optimization methods such as gradient descent to obtain a dense three-dimensional reconstruction result, and finally a reconstructed voxel space is generated.

[0106] It should be noted that, firstly, based on the optimized 3DGS radiation field (preserving anisotropic characteristics), the center position, anisotropic covariance matrix (reflecting the directional differences in density diffusion), and view-dependent radiation intensity values ​​(encoded by the radiation intensity response function to represent directional attenuation characteristics) of each 3D Gaussian are extracted. The density contribution is calculated using a direction-sensitive Gaussian-voxel transformation function: according to the ellipsoidal direction defined by the covariance matrix (the principal axis corresponds to the structural extension direction), a directional density distribution is applied to neighboring voxels in the 3D mesh. The continuous radiation field is mapped to an initial 3D voxel density field through anisotropic Gaussian volume integral, generating the first voxel representation. This process explicitly preserves direction dependence (such as bone texture orientation or anisotropic attenuation of metal implants). A differentiable X-ray projection operator is constructed, introducing directional physical constraints: when generating projection rays along each real scanning angle, the attenuation coefficient is modulated according to the cosine of the angle between the ray incident direction and the Gaussian principal axis within the voxel. Theoretical projection values ​​are calculated by direction-sensitive integrals on voxels through which light passes, and compared pixel-by-pixel with actual sparse projections to construct a direction-aware projection data fidelity term. This term forces the reconstructed volume to match the true attenuation characteristics at specific angles. Furthermore, combining the differences between the actual projection and direction-sensitive simulation results, an anisotropic joint optimization objective function (including direction fidelity and anisotropic regularization terms) is constructed. Voxel density is iteratively updated using a gradient descent algorithm. Directional gradient calculation: analyzing the partial derivative of the incident angle with respect to the density gradient during backpropagation; anisotropic constraint: penalizing artifacts caused by abrupt changes in principal axis directions; finally, a directionally consistent reconstructed voxel space is generated, whose density field accurately reflects the anisotropic structure. Thus, a physically accurate 3D Gaussian radiation field (3DGS) is converted into a voxel space that meets clinical diagnostic standards, and the noise, artifacts, and edge blurring problems in sparse view reconstruction are resolved.

[0107] Figure 2 The flowchart illustrates an optimized method for integrated view synthesis and sparse view CT reconstruction according to an embodiment of the present invention.

[0108] like Figure 2 As shown, this invention provides a flowchart of an optimized method for integrated view synthesis and sparse view CT reconstruction, comprising:

[0109] S202, obtain the reconstructed voxel space, filter significant structural regions based on the preset density threshold, traverse the voxel grid within the selected region and extract high gradient change points as candidate centers, calculate the principal axis direction of the covariance matrix based on the three-dimensional gradient direction field of the candidate point neighborhood, and use principal component analysis to determine the structural extension direction.

[0110] S204 uses the principal axis direction of the covariance matrix to linearly map voxel values ​​to the radiation intensity range and associates them with the corresponding anisotropic properties to finally generate a new 3DGS radiation field.

[0111] S206: Perform volumetric consistency rendering based on the new 3DGS radiation field, generate a new synthetic virtual projection image by analyzing and integrating the light rays along the target viewpoint, and alternately iterate the radiation field optimization and voxel constraint reconstruction using the loss function until the preset number of iterations is reached to output the new viewpoint synthetic image and CT reconstruction volume.

[0112] In S208, during the alternating iteration process, the forward optimization stage updates the Gaussian position and covariance parameters through stochastic gradient descent based on the projection error loss, while the reverse constraint stage strengthens edge continuity and suppresses outlier noise points by reapplying total variational regularization and projection fidelity constraints.

[0113] It should be noted that, to achieve the conversion from reconstructed voxel space to an anisotropic 3D Gaussian radiation field and generate diagnostic-grade output through bidirectional optimization, the following steps are taken: First, the reconstructed voxel space is acquired. High-density regions are selected based on a preset density threshold. Within the selected region, the voxel mesh is traversed, and extreme points with gradient values ​​greater than the preset threshold are extracted as candidate centers—these points are typically located at anisotropic structural interfaces (such as trabecular bone junctions or the edges of metal implants). For each candidate point, the 3D gradient direction field of its neighborhood is calculated. Principal component analysis (PCA) is used to determine the physical extension principal axis of the local structure, thereby defining the direction of the anisotropic covariance matrix. Using the principal axis direction of the covariance matrix (preserving the anisotropic topology), the voxel values ​​are linearly mapped to the radiation intensity range, and the view-related radiation properties are correlated to generate a new 3DGS radiation field. Based on the new radiation field, direction-aware volumetric rendering is performed: when integrating rays along the target viewpoint, the radiation intensity is modulated according to the angle θ between the ray incident direction and the Gaussian principal axis. After generating the virtual projection image, alternating iterative optimization is initiated: In the forward optimization phase, the orientation-sensitive loss between the virtual projection and the real data is calculated, and the Gaussian position and covariance anisotropy ratio are dynamically adjusted through stochastic gradient descent to accurately restore anisotropic features; In the reverse constraint phase, the updated radiation field is converted to voxel space, and orientation constraint TV regularization is applied. Edge penalties are weakened along the principal axes of the structure (preserving anisotropic transitions), and smoothing and noise suppression are enhanced in the vertical direction. Simultaneously, the projection fidelity term is calibrated to eliminate angle-specific biases. When the preset number of iterations is reached, the final result—a synthesized image from a new perspective and a CT reconstructed body—is output.

[0114] It's worth noting that the transformation from 3D to 2D images involves: projection transformation – rasterization – alpha blending to synthesize the final image. After the transformation, the image is first converted from the world coordinate system to the camera coordinate system, and then compressed from perspective projection to orthographic projection. Because this process is non-linear, a Jacobian matrix is ​​introduced to account for the covariance. After compression, viewport transformation and rasterization are performed. During rasterization, the image is first blurred by a low-pass filter before sampling.

[0115] Another aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a 3DGS-based integrated view synthesis and sparse view CT reconstruction method program, wherein when the 3DGS-based integrated view synthesis and sparse view CT reconstruction method program is executed by a processor, it implements the steps of the 3DGS-based integrated view synthesis and sparse view CT reconstruction method as described in any of the preceding claims.

[0116] This invention proposes an integrated new perspective synthesis and sparse perspective CT reconstruction method based on 3DGS. Compared with existing technologies, this invention has the following advantages: Rapid initialization: This invention adopts Angle-Attitude Cube Unified Initialization (ACUI), directly utilizing scanner geometric parameters to generate the camera matrix and Gaussian center, avoiding matching failures and lengthy processing times in low-contrast X-ray images using SfM. Volumetric Consistent Rendering: This invention approximates the 2D projection of 3DGS with ray direction analytical integrals, avoiding edge blurring and stripe artifacts, and better achieving the high-fidelity requirements of sparse perspective CT reconstruction. Viewpoint-independent density: This invention replaces the spherical harmonic function with density, eliminating the influence of viewpoint direction, better conforming to the anisotropic physical characteristics of X-ray penetration imaging and accelerating convergence.

[0117] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0118] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0119] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0120] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0121] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0122] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for integrated viewpoint synthesis and sparse viewpoint CT reconstruction based on 3DGS, characterized in that, include: An intrinsic and extrinsic parameter matrix is ​​established using X-ray scanning equipment parameters to obtain raw CT projection data under sparse viewpoints. An initial voxel space is constructed using an angle-attitude cuboid uniform initialization strategy, and an initial 3DGS radiation field model is performed using the initial voxel space. The initial 3DGS radiation field is imported into the 3D Gaussian rendering framework for volumetric consistency rendering to generate a virtual perspective composite image. The initial 3DGS radiation field is then combined with the original CT projection data to optimize its spatial structure and generate a spatially optimized 3DGS radiation field. Based on the spatially optimized 3DGS radiation field, the anisotropic radiation intensity is modeled by a learnable feature vector using the radiation intensity response function. The feature weights are iteratively updated through backpropagation to output the physically optimized 3DGS radiation field. The center position, covariance matrix and anisotropic radiation intensity value of each 3D Gaussian are extracted by the physical optimization 3DGS radiation field. The Gaussian distribution is mapped to the voxel grid and a density contribution is applied to reconstruct the voxel space, generating a reconstructed voxel space. The reconstructed voxel space is converted back into a new 3DGS radiation field, and the virtual view projection is re-rendered based on the new 3DGS radiation field. The loss function is used for iterative optimization until a preset number of iterations are reached to output a new view synthesized image and CT reconstruction. The process involves establishing an intrinsic and extrinsic parameter matrix using X-ray scanning equipment parameters, acquiring raw CT projection data under sparse viewing angles, constructing an initial voxel space using an angle-attitude cuboid uniform initialization strategy, and performing initial 3DGS radiation field modeling using this initial voxel space. Specifically, this includes: Obtain X-ray scanning equipment parameters, including the distance from the X-ray source to the object, the distance from the X-ray source to the detector, the detector size, and the rotation angle sequence corresponding to different viewing angles; The X-ray source rotation trajectory is encoded by trigonometric function relationships, and the scale of the cone beam projection is defined by distance parameters. This generates an external parameter matrix describing the camera position and attitude and an internal parameter matrix describing the imaging characteristics, in order to establish a mapping relationship between the three-dimensional coordinate system and the two-dimensional detection plane. Obtain the original CT projection data under sparse viewpoint, and use the generated intrinsic and extrinsic parameter matrix to perform geometric correction on the original CT projection data; construct a cuboid with size S1×S2×S3 to completely surround the target object, and use a mesh of size M1×M2×M3 to divide the constructed cuboid to generate the initial voxel space. Subsequently, a set of three-dimensional coordinate points is obtained by uniform sampling at intervals of d∈R in the initial voxel space. The density value of each sampling point is extracted. If the density value is greater than the preset density threshold, the 3D Gaussian function is initialized with the corresponding position coordinate as the center of the target voxel unit, and finally the initial 3DGS radiation field is generated. Each Gaussian function is defined by a covariance matrix to define its spatial distribution range. The initial direction is estimated by the voxel gradient, the initial opacity is obtained by linear mapping of the voxel values, and the radiation color attribute is assigned a gray value according to a preset material library. The process of importing the initial 3DGS radiation field into a 3D Gaussian rendering framework for volumetric consistency rendering to generate a virtual perspective composite image specifically includes: The initial 3DGS radiation field is imported into the 3D Gaussian rendering framework to generate an imaging plane, in which a ray is constructed for each pixel, originating from the ray source and passing through the center of the pixel; Perform volumetric consistency rendering on each ray, detect all 3D Gaussian primitives that intersect with the target ray and sort them from front to back to ensure that the integration order conforms to the real light propagation direction; Subsequently, each three-dimensional Gaussian is projected onto the ray direction to obtain a one-dimensional Gaussian function. Its mean and scale along the ray direction are calculated. Based on the maximum value and scale of the Gaussian along the ray direction, the transmittance of each Gaussian is calculated using analytical integral. The calculated transmittance is converted into opacity. Based on the forward alpha blending model, the pixel colors are synthesized using all Gaussian colors and alpha values ​​to generate a virtual perspective composite image.

2. The integrated view synthesis and sparse view CT reconstruction method based on 3DGS according to claim 1, characterized in that, The specific methods for calculating transmittance and opacity are as follows: The specific formula for calculating the transmittance of each Gaussian using analytical integrals is as follows: ， in, It is the maximum value of the Gaussian distribution along the direction d of the light ray; It is the mean of a 1D Gaussian distribution, representing the position of the center of the Gaussian distribution on the light ray; It is the variance of the 1D Gaussian distribution, representing the width of the Gaussian distribution; It is the weight of the j-th Gaussian volume element; The specific formula for converting the calculated transmittance of light through the Gaussian region into opacity is as follows: ， in, Transmittance, For transparency.

3. The integrated viewpoint synthesis and sparse viewpoint CT reconstruction method based on 3DGS according to claim 1, characterized in that, The process of optimizing the spatial structure of the initial 3DGS radiation field by combining the original CT projection data to generate a spatially optimized 3DGS radiation field specifically includes: Obtain geometrically corrected original CT projection data, generate a real projection image based on the geometrically corrected original CT projection data, and calculate the L1 norm difference of pixel-by-pixel gray values ​​between the real projection image and the virtual viewpoint synthesized image as an intensity fidelity term. A multi-scale SSIM algorithm is introduced to measure the matching degree between the edge and the texture within the local window, which is used as a structural similarity term. A dual-constraint loss function is constructed based on the intensity fidelity term and the structural similarity term. The initial 3DGS radiation field is optimized based on the dual-constraint loss function. The gradient is calculated using backpropagation, and the 3D Gaussian parameters are dynamically optimized using the stochastic gradient descent algorithm. Finally, the spatially optimized 3DGS radiation field is output. During the optimization process, the opacity of each Gaussian point is compared with the preset opacity threshold. When the opacity is less than the preset opacity threshold, the corresponding Gaussian point is identified as a low-contribution Gaussian point and pruning is performed.

4. The integrated view synthesis and sparse view CT reconstruction method based on 3DGS according to claim 1, characterized in that, The spatially optimized 3DGS radiation field utilizes a radiation intensity response function to model anisotropic radiation intensity through learnable feature vectors, iteratively updates feature weights through backpropagation, and outputs a physically optimized 3DGS radiation field, specifically including: The spatially optimized 3DGS radiation field is obtained, the spatial location and covariance parameters of each Gaussian point cloud are extracted, the anatomical tissue density region to which the Gaussian point belongs is located based on the parameters, and an independent learnable feature vector is assigned to each Gaussian point cloud according to the preset density-weight mapping rule. A radiation intensity response function is introduced to perform a nonlinear transformation on the learnable feature vector. The Sigmoid activation function is used to compress the feature values ​​to the standard radiation intensity range. A constant weight vector is applied simultaneously to constrain the physical contribution ratio of the feature components, thereby generating viewpoint-independent radiation intensity values. During the optimization process, a physical consistency loss term is constructed based on the difference gradient between the virtual synthetic projection and the real CT projection, combined with the prior knowledge of X-ray anisotropic attenuation, and the feature vector weights are updated through backpropagation. By iteratively optimizing and gradually suppressing non-physically related feature components, the radiation intensity field converges to the optimal solution that conforms to the attenuation physical model, and outputs a physically optimized 3DGS radiation field with enhanced anisotropic properties. The constant weight vector is initialized based on the physical prior of tissue density and is fixed throughout, ensuring that the radiation intensity is determined only by the spatial location density; the X-ray anisotropic attenuation prior is that the radiation intensity at the same spatial point is constant under any viewing angle.

5. The integrated view synthesis and sparse view CT reconstruction method based on 3DGS according to claim 1, characterized in that, The process involves extracting the center position, covariance matrix, and anisotropic radiation intensity values ​​of each 3D Gaussian sphere from the physically optimized 3DGS radiation field, mapping the Gaussian distribution onto a voxel grid, and applying a density contribution to reconstruct the voxel space, thereby generating a reconstructed voxel space. Specifically, this includes: Based on the physical optimization of the 3DGS radiation field, the center position, covariance matrix and anisotropic radiation intensity value of each 3D Gaussian are extracted. The density contribution of each Gaussian distribution to the voxel unit in its local space is applied. The continuous radiation field is mapped to the initial three-dimensional voxel density field by the approximate integration of the Gaussian volume function to generate the first voxel representation. A differentiable X-ray projection operator is constructed, and the theoretical projection value is calculated by integrating each projection ray in the voxel space. The theoretical projection value is then compared pixel by pixel with the real X-ray projection under the actual sparse angle, and the projection data fidelity term is calculated as the reconstruction loss. Subsequently, based on the construction of a differentiable X-ray projection operator, and combined with the difference between the projection image of the real sparse viewpoint and the simulated projection result, a joint optimization objective function such as projection consistency loss is constructed. The voxel density is iteratively updated through differentiable optimization methods such as gradient descent to obtain a dense three-dimensional reconstruction result, and finally the reconstructed voxel space is generated.

6. The integrated view synthesis and sparse view CT reconstruction method based on 3DGS according to claim 1, characterized in that, The process of converting the reconstructed voxel space back into a new 3DGS radiation field, re-rendering the virtual view projection based on the new 3DGS radiation field, and iteratively optimizing using a loss function until a preset number of iterations are reached to output a new view synthesized image and CT reconstructed volume, specifically includes: The reconstructed voxel space is obtained, and significant structural regions are screened based on a preset density threshold. Within the selected region, the voxel grid is traversed and high gradient change points are extracted as candidate centers. The principal axis direction of the covariance matrix is ​​calculated based on the three-dimensional gradient direction field of the candidate point neighborhood, and principal component analysis is used to determine the structural extension direction. By using the principal axis of the covariance matrix to linearly map voxel values ​​to the radiation intensity range and associating them with the corresponding anisotropy properties, a new 3DGS radiation field is finally generated. Based on the new 3DGS radiation field, volumetric consistency rendering is performed. A new synthetic virtual projection image is generated by analytical integration along the target viewpoint ray. Combined with the loss function, radiation field optimization and voxel constraint reconstruction are alternately iterated until the preset number of iterations is reached to output a new viewpoint synthetic image and CT reconstruction volume. During the alternating iteration process, the forward optimization phase updates the Gaussian position and covariance parameters through stochastic gradient descent based on the projection error loss, while the reverse constraint phase strengthens edge continuity and suppresses outlier noise points by reapplying total variational regularization and projection fidelity constraints.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a 3DGS-based integrated view synthesis and sparse view CT reconstruction method program. When the 3DGS-based integrated view synthesis and sparse view CT reconstruction method program is executed by a processor, it implements the steps of the 3DGS-based integrated view synthesis and sparse view CT reconstruction method as described in any one of claims 1 to 6.

Images