Two-step anisotropic Gaussian splashing method and device

Abstract

The invention relates to the technical field of radiation imaging, in particular to a two-step anisotropic Gaussian spattering method and device.The method comprises the steps that Gaussian kernel representation of a target scanning object is obtained through projection data initialization, and Gaussian kernel parameters are preliminarily optimized with a reconstruction result of a traditional reconstruction method as a supervision signal; a projection image represented by a Gaussian kernel is obtained through micro-rasterization, a scanned real projection image serves as a supervision signal, Gaussian kernel parameters are further optimized, anisotropic Gaussian regularization is introduced, the representation behavior of the Gaussian kernel is standardized, if a scanning mode does not meet a standard camera model, a virtual detection plane is introduced, and the real projection image is obtained. A mapping relation between the projection data and real projection data is constructed to meet a camera model; and finally voxelizing the Gaussian representation to obtain a computed tomography reconstruction result. According to the method, the volume reconstruction effect under the condition of CL scanning mode ill-conditioned data can be effectively improved, interlayer artifacts and noise of the reconstructed image are inhibited, and the practicability is extremely high.

Description

Technical Field

[0001] The present application relates to the field of radiation imaging technology, and in particular to a two-step anisotropic Gaussian sputtering method and device. Background Art

[0002] With the development of modern industry, computed tomography (CT) has been widely used in many fields because it is not restricted by the type of workpiece material and geometric shape, and has high spatial and density resolution. However, CT is not suitable for the inspection and imaging of plate-like specimens, and the limited detection space in the scanner makes it difficult to rotate plate-like objects, making it difficult to collect complete data. The scanning method of computed laminography (CL) avoids the rays passing through the test piece in the direction of maximum thickness, which is specific in data acquisition. However, due to the limited penetration angle of the rays during CL imaging, the collected projection information is incomplete, resulting in serious inter-layer aliasing artifacts in the CL reconstructed images.

[0003] In related technologies, to address the serious inter-layer aliasing artifacts in CL reconstructed images, numerous studies have focused on adding regularization or prior maps to iterative algorithms, using deep learning methods to remove inter-layer aliasing artifacts in CL reconstructed images. In recent years, 3D Gaussian splatting (3DGS) has been a major breakthrough in 3D reconstruction and 3D scene representation, and has become a popular technology in the field of computer vision. Specifically, 3DGS models the scene as a series of 3D Gaussian kernels with learnable parameters, and uses camera imaging principles and differentiable rasterization technology to render the scene into a 2D image and perform iterative optimization, achieving realistic scene surface reconstruction. For X-ray imaging, cone-beam X-ray scanners can be compared to visible light cameras, and the constraints inherent in the Gaussian representation help improve the quality of volume reconstruction under pathological conditions. Therefore, preliminary research has been conducted on the application of 3DGS to CT imaging.

[0004] However, in actual application scenarios, the aliasing artifacts of CL images are essentially artifacts caused by insufficient projection data and thus missing information. Iterative reconstruction or deep learning methods can only improve but not completely eliminate the artifacts. CL reconstruction methods based on 3DGS are relatively scarce, and most of them focus on adjusting the Gaussian representation to adapt to X-ray imaging, rather than customizing special reconstruction strategies based on X-ray characteristics. The clarity and realism of the reconstructed images are also lacking, which needs to be urgently addressed. Summary of the Invention

[0005] The present application provides a two-step anisotropic Gaussian sputtering method and device to solve the problem in related technologies that, in actual application scenarios, the aliasing artifacts of CL images are essentially artifacts caused by insufficient projection data and thus missing information. Iterative reconstruction or deep learning methods can only improve but not completely eliminate the artifacts. CL reconstruction methods based on 3DGS are relatively scarce and most of them focus on adjusting the Gaussian representation to adapt to X-ray imaging, without customizing a special reconstruction strategy based on X-ray characteristics. The clarity and authenticity of the reconstructed image are also lacking.

[0006] A first aspect embodiment of the present application provides a two-step anisotropic Gaussian sputtering method, comprising the following steps: initializing the target scanning object through the projection data of the target scanning object to obtain an initial Gaussian representation of the target scanning object, and iteratively optimizing the Gaussian kernel parameters of the initial Gaussian representation until a first preset iterative stop condition is satisfied to obtain a preliminary Gaussian representation; when the scanning method of the target scanning object satisfies the camera model, iteratively optimizing the Gaussian kernel parameters of the preliminary Gaussian representation until a second preset iterative stop condition is satisfied to obtain a final fine Gaussian representation; otherwise, based on the scanning method, constructing a mapping relationship between a virtual detection plane and the projection data, and based on the mapping relationship, iteratively optimizing the Gaussian kernel parameters of the preliminary Gaussian representation until the second preset iterative stop condition is satisfied to obtain the final fine Gaussian representation; discretizing the three-dimensional volume corresponding to the final fine Gaussian representation into a voxel representation to generate a computer tomography reconstruction result of the target scanning object according to the voxel representation.

[0007] Optionally, in one embodiment of the present application, the target scanning object is initialized by the projection data of the target scanning object to obtain an initial Gaussian representation of the target scanning object, including: initializing the target scanning object based on the projection data to obtain a point cloud representation of the target scanning object; modeling the points corresponding to the point cloud representation as corresponding three-dimensional radiation Gaussian kernels to combine the three-dimensional radiation Gaussian kernels to obtain the initial Gaussian representation.

[0008] Optionally, in one embodiment of the present application, the iteratively optimizing the Gaussian kernel parameters of the initial Gaussian representation until a first preset iterative stopping condition is satisfied to obtain the preliminary Gaussian representation includes: discretizing the projection data and the target scanned object based on the detector resolution and preset reconstruction resolution of the target computer tomography system to generate a traditional reconstruction result in combination with the geometric relationship and target reconstruction strategy of the target computer tomography system; iteratively optimizing the Gaussian kernel parameters based on the traditional reconstruction result and the voxel representation corresponding to the initial Gaussian representation until the first preset iterative stopping condition is satisfied to obtain the preliminary Gaussian representation.

[0009] Optionally, in one embodiment of the present application, constructing a mapping relationship between a virtual detection plane and the projection data based on the scanning mode includes: determining a virtual detector according to a geometric relationship between a light source point and a real detector of the target computer tomography imaging system, wherein the combination of the virtual detector and the light source point satisfies a standard camera model; determining a coordinate mapping relationship between the virtual detector and the real detector at multiple projection angles according to the scanning mode and the imaging geometric relationship; and determining a mapping relationship between the virtual detection plane and the projection data according to the coordinate mapping relationship.

[0010] Optionally, in one embodiment of the present application, the iterative optimization of the Gaussian kernel parameters of the preliminary Gaussian representation based on the mapping relationship until the second preset iterative stop condition is met to obtain the final fine Gaussian representation includes: calculating the rendering projection value of the target scanning object at the target angle based on the mapping relationship, and constructing a loss function based on the rendering projection value and the true projection value of the target scanning object at the target angle; iteratively optimizing the Gaussian kernel parameters of the preliminary Gaussian representation according to the loss function until the second preset iterative stop condition is met to obtain the final fine Gaussian representation.

[0011] Optionally, in one embodiment of the present application, constructing a loss function based on the rendered projection value and the real projection value of the target scanned object at a target angle includes: constructing a data fidelity term between the rendered projection value and the real projection value; adding an anisotropic regularization term for constraining the Gaussian kernel's multi-directional representation behavior to the data fidelity term according to a target ratio to determine the loss function.

[0012] A second aspect of the present application provides a two-step anisotropic Gaussian sputtering device, comprising: a first optimization module, configured to initialize the target scanning object using projection data of the target scanning object to obtain an initial Gaussian representation of the target scanning object, and iteratively optimize Gaussian kernel parameters of the initial Gaussian representation until a first preset iterative stop condition is satisfied to obtain a preliminary Gaussian representation; a second optimization module, configured to iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation when the scanning method of the target scanning object satisfies a camera model until a second preset iterative stop condition is satisfied to obtain a final fine Gaussian representation; otherwise, based on the scanning method, construct a mapping relationship between a virtual detection plane and the projection data, and based on the mapping relationship, iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation until the second preset iterative stop condition is satisfied to obtain the final fine Gaussian representation; and a voxelization module, configured to discretize the three-dimensional volume corresponding to the final fine Gaussian representation into a voxel representation to generate a computer tomography reconstruction result of the target scanning object based on the voxel representation.

[0013] Optionally, in one embodiment of the present application, the first optimization module includes: an initialization unit for initializing the target scanning object based on the projection data to obtain a point cloud representation of the target scanning object; a modeling unit for modeling the points corresponding to the point cloud representation into corresponding three-dimensional radiation Gaussian kernels to combine the three-dimensional radiation Gaussian kernels to obtain the initial Gaussian representation.

[0014] Optionally, in one embodiment of the present application, the first optimization module includes: a generation unit, configured to discretize the projection data and the target scanned object based on the detector resolution and preset reconstruction resolution of the target computer tomography system, so as to generate a traditional reconstruction result in combination with the geometric relationship of the target computer tomography system and the target reconstruction strategy; and a first optimization unit, configured to iteratively optimize the Gaussian kernel parameters based on the traditional reconstruction result and the voxel representation corresponding to the initial Gaussian representation until the first preset iterative stopping condition is met, so as to obtain the preliminary Gaussian representation.

[0015] Optionally, in one embodiment of the present application, the second optimization module includes: a first determination unit, used to determine a virtual detector based on the light source point of the target computer tomography system, wherein the combination of the virtual detector and the light source point satisfies a standard camera model; based on the scanning mode and the imaging geometric relationship, determining the coordinate mapping relationship between the virtual detector and the real detector under multiple projection angles; and a second determination unit, used to determine the mapping relationship between the constructed virtual detection plane and the projection data based on the coordinate mapping relationship.

[0016] Optionally, in one embodiment of the present application, the second optimization module includes: a calculation unit, used to calculate the rendering projection value of the target scanning object at the target angle based on the mapping relationship, and construct a loss function according to the rendering projection value and the real projection value of the target scanning object at the target angle; a second optimization unit, used to iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation according to the loss function until the second preset iteration stop condition is met to obtain the final fine Gaussian representation.

[0017] Optionally, in one embodiment of the present application, the computing unit includes: a construction subunit for constructing a data fidelity term between the rendered projection value and the true projection value; a determination subunit for adding an anisotropic regularization term for constraining the Gaussian kernel's multi-directional representation behavior to the data fidelity term according to a target ratio to determine the loss function.

[0018] The third aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the two-step anisotropic Gaussian sputtering method as described in the above embodiment.

[0019] A fourth aspect of the present application provides a computer-readable storage medium, which stores a computer program. When the program is executed by a processor, it implements the above two-step anisotropic Gaussian sputtering method.

[0020] The fifth aspect of the present application provides a computer program product, including a computer program, which, when executed, is used to implement the above two-step anisotropic Gaussian sputtering method.

[0021] In the embodiment of the present application, the target scanned object scanned by the target computer tomography imaging system can be modeled as an initial Gaussian representation, and by using the results of the traditional reconstruction algorithm and the scanned real projection data as supervision signals, the Gaussian kernel parameters are iteratively optimized in the image domain and the projection domain, thereby obtaining the final fine Gaussian representation of the target scanned object and the corresponding computer tomography reconstruction result. Thus, a two-step reconstruction method is designed to progressively iteratively optimize the Gaussian kernel parameters of the Gaussian kernel representation modeled according to the target scanned object. Since the traditional reconstruction result without smoothing filtering can retain high-frequency information as much as possible, the preliminary Gaussian representation obtained by iteratively optimizing the initial Gaussian representation in the image domain in this application can effectively present the detailed information of the scanned object, and then further iteratively optimize the preliminary Gaussian representation in the projection domain using the projection data information obtained by scanning to obtain the final fine Gaussian representation, which can effectively suppress the reconstructed image artifacts and noise, thereby providing a three-dimensional image with higher clarity and contrast, and thus achieving accurate high-definition and high-contrast three-dimensional image reconstruction; and the anisotropic Gaussian regularization technology designed in this application takes into account the characteristics of CL imaging and Gaussian kernel representation, standardizes the representation behavior of the Gaussian kernel in multiple directions, and thus restores the inter-layer resolution; and the mapping relationship between the virtual detection plane and the real projection data constructed in this application can ensure that the rendering formula of the three-dimensional Gaussian splash is suitable for various CL scanning methods, greatly improving the practicality of this application. This solves the problem in related technologies that, in actual application scenarios, the aliasing artifacts of CL images are essentially artifacts caused by insufficient projection data and thus missing information. Iterative reconstruction or deep learning methods can only improve but not completely eliminate the artifacts. CL reconstruction methods based on 3DGS are relatively scarce, and most of them focus on adjusting the Gaussian representation to adapt to X-ray imaging, rather than customizing special reconstruction strategies according to X-ray characteristics. The clarity and realism of the reconstructed images are also lacking.

[0022] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0024] Figure 1 A flow chart of a two-step anisotropic Gaussian sputtering method provided according to an embodiment of the present application;

[0025] Figure 2 A schematic diagram of a two-step anisotropic Gaussian sputtering method according to an embodiment of the present application;

[0026] Figure 3A schematic diagram of the construction of CL imaging geometry and a virtual detection plane under general circumstances according to one embodiment of the present application;

[0027] Figure 4 This is a schematic diagram of imaging geometry of a square cross-sectional field of view rotating CL imaging system according to one embodiment of the present application;

[0028] Figure 5 A schematic diagram of the geometric and coordinate mapping relationship of a square cross-sectional field of view rotating CL imaging system according to an embodiment of the present application;

[0029] Figure 6 Schematic diagram of the structure of a two-step anisotropic Gaussian sputtering device provided according to an embodiment of the present application;

[0030] Figure 7 A schematic diagram of the structure of an electronic device provided according to an embodiment of the present application.

[0031] Reference numerals:

[0032] 10-Two-step anisotropic Gaussian sputtering device: 100-first optimization module, 200-second optimization module and 300-voxelization module; 701-memory, 702-processor and 703-communication interface. DETAILED DESCRIPTION

[0033] The following describes in detail embodiments of the present application, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0034] The following describes the two-step anisotropic Gaussian sputtering method and device of the embodiment of the present application with reference to the accompanying drawings. In view of the related technologies mentioned in the above background technology, in actual application scenarios, the removal of inter-layer aliasing artifacts in CL reconstructed images by deep learning methods will result in the traditional CL image reconstruction method being able to improve but not completely eliminate this artifact. The CL reconstruction methods based on 3DGS are relatively scarce and most of them focus on adjusting the Gaussian representation to adapt to X-ray imaging, without customizing a special reconstruction strategy based on X-ray characteristics. The clarity and authenticity of the reconstructed image are also lacking. The present application provides a two-step anisotropic Gaussian sputtering method, in which the target scanned object scanned by the target computer tomography imaging system can be modeled as an initial Gaussian representation, and by using the results of the traditional reconstruction algorithm and the scanned real projection data as supervision signals, the Gaussian kernel parameters are iteratively optimized in the image domain and the projection domain to obtain the final fine Gaussian representation of the target scanned object and the corresponding computer tomography reconstruction result. Thus, a two-step reconstruction method is designed to progressively iteratively optimize the Gaussian parameters of the Gaussian representation. Since the traditional reconstruction results without smoothing filtering can retain high-frequency information as much as possible, the preliminary Gaussian representation obtained by iteratively optimizing the initial Gaussian representation in the image domain in this application can effectively present the detailed information of the scanned object, and then further iteratively optimize the preliminary Gaussian representation in the projection domain using the projection data information obtained by scanning to obtain the final fine Gaussian representation, which can effectively suppress the artifacts and noise of the reconstructed image, thereby providing a three-dimensional image with higher clarity and contrast, and thus achieving accurate high-definition and high-contrast three-dimensional image reconstruction; and the anisotropic Gaussian regularization technology designed in this application takes into account the characteristics of CL imaging and Gaussian kernel representation, standardizes the representation behavior of the Gaussian kernel in multiple directions, and thus restores the inter-layer resolution; furthermore, the mapping relationship between the virtual detection plane constructed in this application and the real projection data can ensure that the rendering formula of the three-dimensional Gaussian splash is suitable for various CL scanning methods, which greatly improves the practicality of this application. This solves the problem in related technologies that, in actual application scenarios, the aliasing artifacts of CL images are essentially artifacts caused by insufficient projection data and thus missing information. Iterative reconstruction or deep learning methods can only improve but not completely eliminate the artifacts. CL reconstruction methods based on 3DGS are relatively scarce, and most of them focus on adjusting the Gaussian representation to adapt to X-ray imaging, rather than customizing special reconstruction strategies according to X-ray characteristics. The clarity and realism of the reconstructed images are also lacking.

[0035] Specifically, Figure 1 A flow chart of a two-step anisotropic Gaussian sputtering method provided in an embodiment of the present application.

[0036] like Figure 1 As shown, the two-step anisotropic Gaussian sputtering method includes the following steps:

[0037] In step S101, the target scanning object is initialized by the projection data of the target scanning object to obtain an initial Gaussian representation of the target scanning object, and the Gaussian kernel parameters of the initial Gaussian representation are iteratively optimized until a first preset iteration stop condition is met to obtain a preliminary Gaussian representation.

[0038] It is understandable that 3D Gaussian splatting (3DGS) is a major breakthrough in 3D reconstruction and 3D scene representation. It can model the scene as a series of 3D Gaussian kernels with learnable parameters, render the scene into a 2D image using camera imaging principles and rendering equations, and perform iterative optimization to achieve real scene surface reconstruction. For X-ray imaging, a fan beam X-ray scanner can be compared to a visible light camera, and the constraints inherent in the Gaussian representation help to improve the effect of volume reconstruction under pathological conditions. Based on this, the embodiment of the present application can implement CL (computed laminography) reconstruction based on 3DGS.

[0039] In some embodiments, the present application may, but is not limited to, first initialize the target scanned object into a series of initial Gaussian representations based on the projection data of the target scanned object, and use the preliminary reconstruction results obtained using the target reconstruction strategy as a supervisory signal to guide the iterative optimization of the Gaussian kernel parameters of the initial Gaussian representation, thereby fitting a preliminary Gaussian representation.

[0040] Among them, the target scanning object here can be understood as any object scanned by CL; the projection data here can be understood as the projection data obtained after CL scans the target scanning object; the target reconstruction strategy here can be understood as some traditional reconstruction algorithms, such as the basic ART (Algebraic Reconstruction Technique) algorithm, SIRT (Simultaneous Iterative Reconstruction Technique) algorithm, ML-EM (Maximum Likelihood-Expectation Maximization) algorithm, etc.; the first preset iteration stop condition here can be understood as the condition for stopping the iterative optimization of the Gaussian kernel parameters of the pre-established initial Gaussian representation, for example, reaching the number of iterations, meeting the parameter requirements, etc., which can be specifically set or adjusted by professional and technical personnel in this field. The embodiments of this application are only for illustrative purposes and are not specifically limited.

[0041] The embodiment of the present application can initialize the target scan object into a series of initial Gaussian representations, and use Gaussian functions to characterize each point in the reconstruction area, thereby utilizing the smooth continuous constraints, flexible representation, and high resolution of the Gaussian representation to improve the effect of target scan object reconstruction under pathological conditions, and introduce image domain optimization to optimize the three-dimensional representation of the target scan object to effectively present the detailed information of the target scan object.

[0042] Optionally, in one embodiment of the present application, the target scanning object is initialized using the projection data of the target scanning object to obtain an initial Gaussian representation of the target scanning object, including: initializing the target scanning object based on the projection data to obtain a point cloud representation of the target scanning object; modeling the points corresponding to the point cloud representation as corresponding three-dimensional radiation Gaussian kernels to combine the three-dimensional radiation Gaussian kernels to obtain the initial Gaussian representation.

[0043] It can be understood based on the relevant descriptions of other embodiments that the embodiments of the present application can initialize the target scanning object through the projection data of the target scanning object, thereby obtaining an initial Gaussian representation of the target scanning object.

[0044] During the actual implementation process, the present application can, but is not limited to, initialize the target scanning object into a point cloud representation based on the projection data of the target scanning object, then model each point in the point cloud representation as a corresponding three-dimensional radiation Gaussian kernel, and then superimpose these three-dimensional radiation Gaussian kernels into an initial Gaussian representation of the target scanning object.

[0045] For example, for a CL system, the process of obtaining projection data can be expressed as, but not limited to:

[0046]

[0047] Among them, g(l) is the projection along the ray path l, f(x) is the attenuation distribution of the scanned object (target scanning object) in the form of a function expression, and the projection operator represents the pixel attenuation integral along the ray path l according to the scanner parameters. The core of CL reconstruction is to solve the inverse problem of (1) and obtain an estimate of f (the matrix representation of the target scanned object).

[0048] The embodiment of the present application can initialize the target scan object into a point cloud representation based on the information provided by the projection image, and then model each point into a three-dimensional radiation Gaussian kernel G i (i is the Gaussian kernel index), a series of Gaussian kernels can be superimposed and combined to form the initial Gaussian representation of the target scanned object, and its expression can be but is not limited to:

[0049]

[0050] in, and They represent the center intensity, position, and shape scale of the Gaussian kernel, respectively, and are the parameters to be learned and optimized. In this application, they can be uniformly expressed as φ i , is the set of all Gaussian kernels, called the initial Gaussian representation of the object. Then the intensity value at any spatial position x within the target scan object f can be expressed as, but not limited to:

[0051]

[0052] In the explanation of the subsequent steps, g and φ are used to represent all projection data g(l) and Gaussian kernel parameters φ i A collection of .

[0053] It should be noted that the embodiments of the present application will perform progressive iterative optimization on the parameters of the Gaussian kernel during the optimization phase in the image domain and the projection domain, so the different initialization methods here have little effect on the final result. Therefore, the embodiments of the present application can also randomly generate an initial point cloud in three-dimensional space according to uniform distribution, Gaussian distribution, etc. as the starting point for optimization. In summary, the embodiments of the present application can initialize the target scanned object into a point cloud representation based on the information provided by the projection data or random information or a combination of projection data and random information. The specific application can be determined by professional and technical personnel in this field based on actual needs and actual application scenarios. The embodiments of the present application are only for illustrative purposes and are not specifically limited.

[0054] The embodiment of the present application can model the target scan object in the target computer tomography as an initial Gaussian representation, and characterize it with Gaussian parameters such as position, shape, scale, and center intensity, effectively utilizing the constraints of the Gaussian representation, thereby improving the effect of reconstructing the target scan object under pathological conditions.

[0055] Optionally, in one embodiment of the present application, the Gaussian kernel parameters of the initial Gaussian representation are optimized using the traditional reconstruction result obtained by the target reconstruction strategy as a supervisory signal until a first preset iterative stop condition is met to obtain a preliminary Gaussian representation, including: discretizing the projection data and the target scanned object in the image domain based on the detector resolution and the preset reconstruction resolution of the target computer tomography system to generate a traditional reconstruction result in combination with the geometric relationship of the target computer tomography system and the target reconstruction strategy; iteratively optimizing the Gaussian kernel parameters based on the traditional reconstruction result and the voxel representation corresponding to the initial Gaussian representation until the first preset iterative stop condition is met to obtain the preliminary Gaussian representation.

[0056] In some embodiments, after obtaining the initial Gaussian representation of the target scanned object, the present application further iteratively optimizes the Gaussian kernel parameters of the initial Gaussian representation in the image domain to obtain a preliminary Gaussian representation of the target scanned object.

[0057] When iteratively optimizing the Gaussian kernel parameters of the initial Gaussian representation, the present application may, but is not limited to, discretizing the projection data and the target scanned object based on the detector resolution rate and preset reconstruction resolution of the target computer tomography system, and then generating a traditional reconstruction result using a target reconstruction strategy based on the geometric relationship of the target computer tomography system. Finally, the Gaussian kernel parameters of the initial Gaussian representation are iteratively optimized by combining the traditional reconstruction result and the voxel representation corresponding to the initial Gaussian representation to obtain a preliminary Gaussian representation of the target scanned object.

[0058] The target computed tomography system herein can be understood as a CL imaging system that scans the target scan object. The preset reconstruction resolution herein can be understood as a resolution condition pre-established based on the requirements for CL image reconstruction. This can be a single resolution value or a resolution range. This embodiment of the present application is for illustrative purposes only and does not impose any specific limitation.

[0059] It should be noted that in the embodiments of this application, there are no specific restrictions on the target scanning object; any object that meets the CL scanning conditions can be used. There are also no specific restrictions on the target CT system. That is, the two-step anisotropic Gaussian sputtering method in this application is applicable to any scannable object and CT system. In actual application, both the target scanning object and the target CT system can be selected or adjusted by professionals in this field based on the actual application scenario and actual needs. The embodiments of this application are for illustrative purposes only and do not impose specific restrictions.

[0060] Taking the commonly used square cross-sectional field of view rotation CL (RC-CL) imaging system as an example, the embodiment of the present application can discretize the projection data g and the target scan object f according to the detector resolution of the system and a certain reconstruction resolution, that is, divide the continuous object space into small units, which will form a discrete voxel grid; then, according to the geometric relationship of the system (the spatial position relationship between the ray source, detector, rotation axis and imaged object in the system), a traditional reconstruction algorithm is used to obtain a traditional reconstruction result.

[0061] At this point, the embodiment of the present application can iteratively optimize the Gaussian kernel parameters of the initial Gaussian representation. In each iteration, the calculation of each discrete voxel grid point x in the reconstruction area is performed. n The intensity value, that is, the sum of the intensity values of each single Gaussian kernel at that point, can be expressed by the following formula but is not limited to:

[0062]

[0063] Among them, f(x n ) represents each discrete voxel grid point x in the reconstruction area nThe intensity value, G i Represents a three-dimensional radiation Gaussian kernel, i is the Gaussian kernel index, φ i represents the parameters of the Gaussian kernel.

[0064] Then calculate With f(x n ) loss function, where [] n Indicates extracting the nth element of the vector. The loss function can be, but is not limited to, L1 loss, L2 loss, etc., and the Gaussian kernel parameter φ is updated by the gradient descent method. i It should be noted that the loss function and update strategy here can be selected or adjusted by professionals in this field according to actual application scenarios and actual needs. The embodiments of this application are only for illustrative purposes and are not specifically limited.

[0065] Furthermore, during the iterations, professionals in this field can also utilize algorithms to intelligently and dynamically adjust the number of Gaussian kernels according to specific training strategies, such as by invoking an adaptive density control module, to enhance representational capabilities and optimize the Gaussian representation of the scene within the scanned area, maximizing the extraction and preservation of detail that can be obtained by traditional reconstruction methods. Ultimately, after a certain number of iterations, when a certain stopping condition is reached (e.g., a set number of iterations), the optimization process can be terminated, resulting in a preliminary Gaussian representation.

[0066] Step S102: When the scanning method of the target scanning object satisfies the camera model, iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation until the second preset iterative stop condition is met to obtain the final fine Gaussian representation; otherwise, based on the scanning method, construct a mapping relationship between the virtual detection plane and the projection data, and based on the mapping relationship, iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation until the second preset iterative stop condition is met to obtain the final fine Gaussian representation.

[0067] In other embodiments, after optimizing the initial Gaussian representation of the target scanned object in the image domain to obtain a preliminary Gaussian representation, the present application can further iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation in the projection domain using the actual projection data of the target scanned object as a supervisory signal to obtain the final Gaussian representation optimization result, i.e., a refined Gaussian representation, thereby achieving accurate reconstruction of the target scanned object.

[0068] Among them, when further iteratively optimizing the Gaussian kernel parameters of the preliminary Gaussian representation in the projection domain, it is necessary to obtain the rendered projection value of the target scanned object at a specific angle. The rendered projection value can be understood as splatting each three-dimensional Gaussian kernel into the screen space to become a two-dimensional Gaussian kernel by treating the target computer tomography system as a pinhole camera model, and then obtaining the rendered two-dimensional projection value through Alpha blending, that is, the result of rendering to a two-dimensional plane at a specific angle.

[0069] When the central ray of the target computer tomography system is perpendicular to the detector plane, each 3D Gaussian kernel can be directly mapped to a 2D Gaussian kernel in the screen space, and then the rendered 2D projection value can be obtained through Alpha blending.

[0070] However, when the central ray of the target CT system is not perpendicular to the detector plane, rendering becomes difficult. Based on this, the embodiment of the present application introduces a virtual detection plane perpendicular to the central ray. Through this virtual detection plane, the target CT system can be compared to a pinhole camera model, so that the fan beam scanner in the target CT system corresponds to the pinhole camera model, and then the rendering is smoothly achieved by using the differentiable rasterization of 3D Gaussian splatter. In general, the schematic diagram of the construction of CL imaging geometry and virtual detection plane is as follows: Figure 2 shown.

[0071] Specifically, the embodiment of the present application can introduce a virtual detection plane based on the scanning method of the target scanning object, that is, the scanning method of the target computer tomography system, and obtain the rendering projection value through the differentiable rasterization of 3DGS to construct a mapping relationship between the virtual detection plane and the projection data. Based on the mapping relationship, the Gaussian kernel parameters of the preliminary Gaussian representation are further iteratively optimized until the second preset iteration stop condition is met to obtain the final fine Gaussian representation.

[0072] It should be noted that the initial Gaussian representation, preliminary Gaussian representation, and refined Gaussian representation in the embodiments of the present application are all states, which are essentially Gaussian kernel representations, and iterative optimization is a process, that is, the description of the preliminary Gaussian representation is only to distinguish before and after iterative optimization, and there is no other difference from the Gaussian kernel representation.

[0073] It should also be noted that the second preset iterative stopping condition here can be understood as a pre-established condition for stopping the iterative optimization of the Gaussian kernel parameters of the preliminary Gaussian representation, for example, reaching the number of iterations. However, in the embodiment of the present application, the second preset iterative stopping condition may be the same as or different from the first preset iterative optimization condition, and may be specifically set or adjusted by professionals skilled in the art. The embodiment of the present application is for illustrative purposes only and does not impose any specific limitation.

[0074] The embodiment of the present application can perform projection domain optimization on the basis of performing image domain optimization on the initial Gaussian representation of the target scan object to present the detailed information of the target scan object, thereby effectively suppressing the inter-layer artifacts and noise of the CL reconstructed image and providing a three-dimensional image of the target scan object with higher clarity and better contrast.

[0075] Optionally, in one embodiment of the present application, based on the scanning method of the target scanning object, a mapping relationship between the virtual detection plane and the projection data is constructed, including: determining a virtual detector according to the light source point of the target computer tomography system, wherein the combination of the virtual detector and the light source point satisfies the standard camera model; determining the coordinate mapping relationship between the virtual detector and the real detector at multiple projection angles according to the scanning method and the imaging geometric relationship; and determining the mapping relationship between the virtual detection plane and the projection data based on the coordinate mapping relationship.

[0076] In certain embodiments, when constructing a mapping relationship between a virtual detection plane and projection data according to the scanning method of a target scanned object, the present application first needs to determine a virtual detector based on the light source point of the target computer tomography imaging system, and ensure that the combination of the virtual detector and the light source point can meet the standard camera model.

[0077] For example, while ensuring that the plane where the virtual detector is located is the same plane as the plane where the real detector is located, the present application can draw a perpendicular line to the plane where the real detector of the target computer tomography system is located based on the light source point of the target computer tomography system. The intersection point is the center of the virtual detector, and the perpendicular line is the central ray. The size of the virtual detector can be, but is not limited to, set to the minimum rectangular surface that can enclose the motion trajectory of the real detector.

[0078] Because the rendered projection values are obtained by differentiable rasterization, and the projection is on the virtual detector, its scale is larger than the real projection value collected by the real detector. Therefore, the embodiment of the present application can determine the coordinate mapping relationship between the virtual detector and the real detector at multiple projection angles based on the scanning method and imaging geometry of the target scanning object. In other words, the coordinate position of the real detector in the virtual detector coordinate system at each projection angle is determined. The obtained coordinate mapping relationship between the virtual detector and the real detector is the mapping relationship between the virtual detection plane and the projection data, and the area on the virtual detector corresponding to the real projection value is found.

[0079] The embodiment of the present application fully exploits the imaging geometric characteristics of CL and introduces a virtual detection plane perpendicular to the central ray. By mapping the real projection coordinates of the target scanned object in the projection data to a specific area of the virtual plane, rendering calculations are only performed on the area corresponding to the actual projection. This can achieve high-precision imaging while smoothly implementing rendering without introducing additional computing costs.

[0080] Moreover, the embodiment of the present application constructs a mapping relationship between the virtual detection plane construction and the real projection data, so that various CL scanning geometries can be easily analogized to the imaging architecture of the pinhole camera model. Therefore, in the optimization stage of the projection data, the forward projection of the object and the back propagation of the error can be efficiently performed based on the differentiable rasterization operator of the three-dimensional Gaussian splash, thereby improving the generalization ability of the reconstruction method for various CL scanning methods and effectively improving the practical ability of the present application.

[0081] Optionally, in one embodiment of the present application, based on the mapping relationship, the Gaussian kernel parameters of the preliminary Gaussian representation are iteratively optimized until a second preset iterative stop condition is met to obtain a final fine Gaussian representation, including: based on the mapping relationship, calculating the rendering projection value of the target scanning object at the target angle, and constructing a loss function based on the rendering projection value and the true projection value of the target scanning object at the target angle; iteratively optimizing the Gaussian kernel parameters of the preliminary Gaussian representation according to the loss function until the second preset iterative stop condition is met to obtain the final fine Gaussian representation.

[0082] Based on the relevant descriptions of other embodiments, it can be understood that the present application can introduce a virtual detection plane to analogize the target computer tomography system to a pinhole camera model, and use differentiable rasterization to achieve two-dimensional projection, thereby obtaining a mapping relationship between the virtual detection plane and the real projection data.

[0083] In some embodiments, when iteratively optimizing the Gaussian kernel parameters of the preliminary Gaussian representation based on the mapping relationship, the present application primarily optimizes based on the projection angle of the target scanned object. Specifically, the present application can utilize differentiable rasterization to obtain the rendered projection values corresponding to the true projection value region of the target scanned object at the target angle, and construct a loss function based on the rendered projection values and the true projection values of the target scanned object at the target angle. Ultimately, the loss function is used to further iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation.

[0084] The target angles herein may be understood as a plurality of angles used by the target computer tomography system when scanning the target scanning object to generate projection data.

[0085] For example, by analogizing the target CT system to a pinhole camera model, the present application can consider the X-ray source of the cone-beam scanner in the target CT system to be the pinhole of the pinhole camera. Subsequently, the present application can adjust the pinhole camera's extrinsic parameters, which can be understood as adjusting the position and posture of the cone-beam scanner when scanning the object, including but not limited to the rotation matrix R and the translation matrix T, so that the center of the pinhole camera is aligned with the X-ray source and directly opposite the virtual detector plane. This process can be achieved through affine transformation, and the formula can be, but is not limited to, expressed as follows:

[0086]

[0087] Among them, x o 、y o 、z o and x s 、y s 、z s They are the coordinate values in the world coordinate system and the camera coordinate system respectively.

[0088] Next, the present embodiment can adjust the camera's internal parameters (focal length F and imaging screen size W9 (width) and H (height), corresponding to the distance from the X-ray source to the detector, the detector size, and the detector's pixel matrix, etc.) to ensure that the camera's main axis is consistent with the scanner's central ray length, and the imaging screen size is consistent with the virtual detector size. After completing these steps, the present embodiment can directly use Alpha blending to obtain the rendered projection value of the target scanned object.

[0089] Furthermore, considering that X-rays have strong penetrability, when the ray r emitted from the light source passes through M three-dimensional Gaussian kernels and intersects the x-ray on the detector plane, 2D When , the Alpha blending formula can be expressed as the superposition of two-dimensional Gaussian kernels on the corresponding plane, which can be, but is not limited to, expressed as follows:

[0090]

[0091] Among them, x 2D , are x, μ respectively i , ρ i ,∑ i Projection value on the detector plane. x 2D and It can be obtained by formula (6) and perspective projection transformation. The transformation formula can be expressed as but not limited to:

[0092]

[0093] Among them, x f and y fis the corresponding perspective projection value in the virtual detection plane coordinate system, and w represents the depth scale.

[0094] Then we can use Σ i After affine approximation of perspective projection transformation, the formula can be, but is not limited to, expressed as follows:

[0095] Σ′=JRΣ i R T J T (9)

[0096]

[0097] Among them, [:m,:n] represents a new matrix consisting of only the elements of the first m rows and n columns of the matrix, J is the corresponding Jacobian matrix, and the projection result is in μ s The partial derivative at is determined, and the formula can be, but is not limited to, expressed as follows:

[0098]

[0099] in, and ρ i Satisfies the following relationship:

[0100]

[0101] Among them, ρ i Represents the central intensity of the three-dimensional Gaussian kernel.

[0102] After obtaining the rendering projection value of the target scanned object, the embodiment of the present application can calculate the loss function of the rendering projection value and the real projection value, and iteratively update the Gaussian kernel parameter φ by the gradient descent method. i During the optimization process, the optimizer may be, but is not limited to, the Adam optimizer.

[0103] Also, during the iteration process, just like the optimization process of the initial Gaussian representation, professional and technical personnel in this field can also call some algorithms to intelligently and dynamically adjust the number of Gaussian kernels according to certain training strategies, such as calling an adaptive density control module, etc., to improve its representation capability and ultimately obtain a fine Gaussian representation of the scene.

[0104] Finally, after a certain number of iterations, a certain iteration stop condition is reached. For example, when the number of iterations has reached the set number of iterations, the optimization process can be stopped and the final fine Gaussian representation can be obtained.

[0105] The embodiment of the present application can optimize the Gaussian representation of the scene within the scanned area of the target scanned object through the projection domain during the iteration based on the smoothing characteristics of the Gaussian kernel, thereby extracting and retaining the detail information that can be obtained by the traditional reconstruction method to the greatest extent, and can effectively retain the overall structure and detail information of the target scanned object, while effectively suppressing inter-layer artifacts and noise, and effectively improving the contrast of the reconstructed image.

[0106] Optionally, in one embodiment of the present application, a loss function is constructed based on the rendered projection value and the real projection value of the target scanned object at the target angle, including: constructing a data fidelity term between the rendered projection value and the real projection value; adding an anisotropic regularization term for constraining the Gaussian kernel's multi-directional representation behavior to the data fidelity term according to the target ratio to determine the loss function.

[0107] As a possible implementation method, the present application can first construct a data fidelity term to measure the difference between the rendered projection value and the true projection value, and then construct a loss function based on the data fidelity term. Alternatively, the present application can also determine the loss function based on the data fidelity term and incorporating an anisotropic regularization term according to the target scale.

[0108] The data fidelity term can be understood as a mathematical expression used to measure the difference between the rendered projection value and the true projection value. Anisotropic regularization can be understood as the strategy used in this embodiment of the application to constrain the Gaussian kernel's representation behavior in all directions, and the regularization term is a mathematical expression of regularization. The target ratio can be understood as the weight ratio of the data fidelity term and the regularization term in the loss function.

[0109] Although projection domain optimization has imposed significant constraints on computer tomography reconstruction, inter-layer aliasing artifacts may still exist, manifesting as blurred inter-layer structures of the reconstructed object. Therefore, when determining the loss function, in addition to using the data fidelity term as a necessary functional expression of the loss function and the commonly used Lp norm and feature space indicators (SSIM, etc.), the embodiments of this application also comprehensively consider the characteristics of CL imaging and Gaussian kernel representation, and design anisotropic Gaussian regularization.

[0110] Therefore, the embodiment of the present application can add an anisotropic Gaussian regularization term on the basis of the data fidelity term to constrain the representation behavior of the Gaussian kernel in all directions, prompting the model to learn more high-frequency information and help restore resolution. However, it should be noted that the addition of an anisotropic Gaussian regularization term to the loss function in the embodiment of the present application can prompt the model to learn more high-frequency information and help restore resolution, but it is not a necessary term in the loss function. The data fidelity term is a necessary term in the loss function in the embodiment of the present application.

[0111] For example, the anisotropic Gaussian regularization loss in the embodiment of the present application is It includes but is not limited to two parts: one is the scale constraint loss function of the Gaussian kernel in the axial direction Here, the axis can be understood as the z direction among the three directions x, y, and z of the Gaussian kernel. This direction is usually also the direction in which the information collected by the CL system is the least complete, which causes the reconstruction result of the scanned object to easily degenerate into blurred aliasing artifacts. z The penalty imposed by the mean of can drive the Gaussian kernel to become narrower in the z direction, so as to improve the possibility of recovering the axial high-frequency features of the scanned object.

[0112] The other part is the cross-sectional total variation (TV) loss function designed by this application to prevent the influence of this constraint on the reconstruction effect within the layer. The cross section here can be understood as a cross section (xy plane) consisting of the x and y directions perpendicular to the z direction of the scanned object, excluding the z-direction component lost by traditional TV, while using its ability to reduce noise and preserve edges in the xy direction to reconcile Possible adverse effects.

[0113] It should be noted that in order to reduce the reconstruction time cost and video memory usage, each iteration only selects a part of the voxel block v(x, y, z) in the image domain to calculate

[0114]

[0115] Among them, λ scale ,λ TV They represent the weights of the two loss functions, which can be but are not limited to 0.05 and 0.1 respectively.

[0116] It should be noted that the anisotropic regularization in this application is mainly used to constrain the representation behavior of the Gaussian kernel in various directions. Therefore, in actual application, the z direction is used as the axial direction, or the x direction and y direction are used as the axial direction, or the xz and yz are used as cross sections, or the directions are directly constrained without being constrained through cross sections, or any other constraints that can limit the representation behavior of the Gaussian kernel to improve the reconstruction effect can be determined or adjusted by professional and technical personnel in this field according to actual conditions. The embodiments of this application are only illustrative and not specifically limited.

[0117] The actual projection value of the target scanning volume at a certain angle is g real , rendering projection value g render For example, the loss function can be, but is not limited to, expressed as a weighted combination of L1 loss, SSIM loss, and anisotropic Gaussian regularization loss:

[0118]

[0119] in, is the data fidelity term, λ D-SSIM is the weight of SSIM loss, which can be but not limited to 0.2; λ1 is the weight of L1 loss, which cannot be 0; λ AGR is the weight of the anisotropic Gaussian regularization loss, which can be 0; and λ1 and λ AGR The size of can be adjusted by professionals in this field according to a certain ratio based on actual conditions. The embodiments of this application are only for illustrative purposes and are not specifically limited.

[0120] In the embodiment of the present application, during the optimization process of the Gaussian kernel parameters, the anisotropic Gaussian regularization can be used to constrain the axial (z-direction) behavior of the Gaussian kernel, thereby prompting the model to learn more high-frequency information in the z-direction and helping to restore inter-layer resolution.

[0121] It should be noted that the anisotropic Gaussian regularization loss is only one of the loss functions included in the loss function of the embodiment of the present application, which is used to constrain the Gaussian kernel behavior and suppress inter-layer artifacts. The loss function may also include any other loss function that helps to constrain the Gaussian kernel behavior and thus suppress inter-layer artifacts. The embodiment of the present application is only for illustrative purposes and is not specifically limited.

[0122] In step S103 , the three-dimensional volume corresponding to the final refined Gaussian representation is discretized into a voxel representation, so as to generate a computer tomography reconstruction result of the target scanned object according to the voxel representation.

[0123] In certain embodiments, after obtaining the final fine Gaussian representation of the target scanned object, the present application can discretize the three-dimensional volume corresponding to the final fine Gaussian representation into a voxel representation again, and then convert it into a visual image through some professional software to obtain the computer tomography reconstruction results of the target scanned object as a whole or different sections.

[0124] The present application is explained in detail below using a specific embodiment.

[0125] Figure 3 This is a schematic diagram of the framework of a two-step anisotropic Gaussian sputtering method according to one embodiment of the present application. Figure 3 As shown in Figure 3, this method is a two-step method consisting of two main stages: image domain optimization stage and projection domain optimization stage.

[0126] Among them, in the image domain optimization stage, the scanned object is first initialized into a series of three-dimensional Gaussian kernel representations, and the traditional reconstruction algorithm is used to obtain the traditional reconstruction results. It is used as a supervisory signal to guide the optimization of the Gaussian kernel parameters and fit the preliminary Gaussian representation. In the projection domain optimization stage, for the specific CL scanning method, a virtual detection plane is introduced to construct a mapping relationship corresponding to the real projection data so that it satisfies the pinhole camera model and ensures that the rendered projection value is aligned with the real projection value. On this basis, according to the scanning characteristics of the CL system, an anisotropic regularization loss function is designed to constrain the axial optimization of the Gaussian kernel in the reconstruction stage, further suppressing inter-layer artifacts and restoring axial resolution. By further iteratively optimizing the parameters of the Gaussian kernel in the projection domain, the final fine Gaussian representation optimization result is obtained to achieve accurate object reconstruction. The specific process can be, but is not limited to, expressed as follows:

[0127] (1) Setting the imaging system and geometric parameters

[0128] First of all, it should be noted that the two-step anisotropic Gaussian sputtering method proposed in this application is applicable to any scanning CL imaging system. Without loss of generality, the embodiment of this application uses a square cross-sectional field of view rotating CL imaging system (RC-CL), the imaging geometry diagram of which is shown in FIG. Figure 4 Its main feature is that the detector plane is perpendicular to the rotation axis. During data acquisition, the world coordinate system O-xyz and the detector coordinate system D-uv do not rotate relative to each other. As a result, the imaging field of view has a rectangular cross-section, making it ideal for imaging planar objects such as printed circuit boards and integrated circuits, while requiring less installation space.

[0129] Next, set the source-object distance |DSO|, source-detector distance |DSD|, tilt angle α, and detector pixel size U × V. During data acquisition, the light source-detector pair rotates once, uniformly capturing projection data. During reconstruction, the reconstruction center is defined at the rotation center.

[0130] (2) Initialize the scanned object to the initial Gaussian representation

[0131] As an example, for the real projection data obtained in the data acquisition phase, the present application can use the FDK algorithm adapted to CL geometry (in actual application, professional and technical personnel in this field can determine any other reconstruction algorithm according to actual conditions and actual needs. The embodiment of the present application is only for illustrative purposes and is not specifically limited) to obtain a preliminary reconstruction result.

[0132] Then, the embodiments of the present application can randomly sample several points within the reconstruction area (determined by the reconstruction resolution, and the proportion can be selected from 0.1% to 0.5%), and use them as the initial position μ0 of the Gaussian kernel; the central intensity value ρ0 can be set as the product of the FDK reconstruction intensity value at the corresponding position and a constant k (0 < k < 1) to prevent the overall intensity value from being too large due to the overlap of Gaussian kernels; the covariance matrix ∑0 is used to describe the distribution characteristics of the Gaussian kernel in each dimension, which determines the shape and scale of the Gaussian kernel. Therefore, the embodiments of the present application can, but are not limited to, define it as a diagonal matrix, and the scales in the three dimensions can be defined as the distances in each dimension between the initial position of the Gaussian kernel and its nearest neighbor:

[0133]

[0134] where NN(x) represents the nearest neighbor of point x.

[0135] It should be noted that the embodiments of the present application can also use other initialization methods, such as setting the central intensity as a random number within a fixed range, etc. Specifically, those skilled in the art can determine any other reconstruction algorithms according to the actual situation and actual needs. The embodiments of the present application are only for illustrative purposes and are not specifically limited. Thus, the initial Gaussian representation can be obtained.

[0136] (3) Image domain optimization (image domain fitting)

[0137] First, the embodiments of the present application can set the number of iterations, and the training method within each iteration period is the same as that described in the previous embodiments. Here, the traditional reconstruction method can select the FDK algorithm adapted to the CL geometry as the traditional reconstruction algorithm, and use to represent the corresponding reconstruction result after discretization. The optimizer can select the Adam optimizer, and the loss function can select the L1 loss. By minimizing the loss function, the optimal Gaussian kernel parameters and the preliminary Gaussian representation of the corresponding scene The formula can, but is not limited to, be expressed as follows:

[0138]

[0139] The iteration stop condition can be achieved by setting the number of iterations. Also, the iteration process is the same as the optimization process of the initial Gaussian representation. Those skilled in the art can call some algorithms to intelligently and dynamically adjust the number of Gaussian kernels according to a certain training strategy, such as calling an adaptive density control module, etc., to improve its representation ability.

[0140] (4) Coordinate mapping between the virtual detection plane and the real projection data

[0141] The coordinate mapping here is the coordinate mapping under the RC-CL system. Figure 5 A schematic diagram of the geometric and coordinate mapping relationship of a square cross-sectional field of view rotational CL imaging system (RC-CL) for this application is shown in FIG. Figure 5 As shown. In order to make the RC-CL system meet the pinhole camera model conditions, the embodiment of the present application introduces a virtual detection plane, and then the differentiable rasterization of the camera model can be used to obtain the rendering result of the entire plane. The scales H and W of the plane can be expressed as:

[0142] H=2|DSD|sinα+V,W=2|DSD|sinα+U(23)

[0143] Where H and W represent the height and width of the virtual detector.

[0144] In order to avoid introducing additional computational costs, the embodiment of the present application constructs a mapping relationship between the virtual detector coordinate system and the real detector coordinate system, aiming to map the real detector area to the corresponding area in the virtual detector coordinate system through coordinate mapping to perform rendering calculations. For example, under the projection angle θ, the mapping relationship can be expressed as the center coordinates (x d ,y d ) in the virtual detector coordinate system:

[0145]

[0146] Therefore, the area where rendering needs to be performed is the area with (x d ,y d ) as the center and the size is [U×V].

[0147] (5) Projection domain optimization

[0148] Then, the preliminary Gaussian representation is optimized in the projection domain based on the mapping relationship between the virtual detector coordinate system and the real detector coordinate system. For example, in each iteration cycle, the present application can randomly select a projection angle θ, and the real projection value at this angle is represented by g real The corresponding rotation matrix R and translation matrix T in formula (6) can be expressed as, but not limited to:

[0149]

[0150] For the camera internal parameters, the imaging screen size is determined by formula (14); the focal length F calculation formula can be expressed as, but not limited to:

[0151] F=|DSD|cosα(27)

[0152] where α represents the tilt angle in imaging geometry.

[0153] After adjusting the internal and external parameters of the camera, the differentiable rasterization technology can be used to obtain the rendering projection value g at the angle θ render .

[0154] Then, when calculating the loss function between the rendered projection value and the real projection value, the embodiment of the present application can, but is not limited to, express the loss function as L1 loss, SSIM loss, and heterogeneous Gaussian regularization loss. The anisotropic Gaussian regularization loss here is As explained above, it includes but is not limited to two parts: one is the Gaussian kernel axial (z direction) scale constraint loss By directly calculating the z-scale σ of the Gaussian kernel z The mean of is penalized, driving the Gaussian kernel to become narrower in the z direction to increase the possibility of recovering the high-frequency features of the axial direction of the scanned object. In order to prevent this constraint from affecting the reconstruction effect within the layer, another loss function is the cross-sectional (xy plane) total variation (TV) loss Eliminate the z-direction component of traditional TV loss, and use its ability to reduce noise and preserve edges in the xy direction to reconcile It should be noted that in order to reduce the reconstruction time cost and video memory usage, each iteration only selects a part of the voxel block v(x,y,z) in the image domain to calculate

[0155]

[0156] Among them, λ Scale ,λ TV They can be set to, but are not limited to, 0.05 and 0.1 respectively.

[0157] Exemplarily, the loss function is expressed as a weighted combination of L1 loss, SSIM loss, and anisotropic Gaussian regularization loss:

[0158]

[0159] Among them, λ1, λ D-SSIM ,λ AGR It can be set to, but is not limited to, 1.0, 0.2, and 1.0 respectively.

[0160] The Adam optimizer can be used as the optimizer. By minimizing the loss function, the optimal Gaussian kernel parameters can be obtained. And the corresponding scene fine Gaussian representation

[0161]

[0162] The iteration stop condition can be achieved by setting the number of iterations. Furthermore, the iteration process is similar to the optimization process of the initial Gaussian representation. Professionals in this field can use certain algorithms to intelligently and dynamically adjust the number of Gaussian kernels according to certain training strategies, such as using an adaptive density control module, to obtain a final, refined Gaussian representation of the scene.

[0163] Finally, the three-dimensional volume corresponding to the final fine Gaussian representation after reconstruction is discretized into voxel representation through formula (5), and then converted through professional software to obtain the visual three-dimensional reconstruction results of the entire object or different sections.

[0164] According to the two-step anisotropic Gaussian sputtering method proposed in the embodiment of the present application, the target scanned object scanned by the target computer tomography imaging system can be modeled as an initial Gaussian representation, and by using the results of the traditional reconstruction algorithm and the scanned real projection data as supervision signals, the Gaussian kernel parameters are iteratively optimized in the image domain and the projection domain to obtain the final fine Gaussian representation of the target scanned object and the corresponding computer tomography reconstruction result. Thus, a two-step reconstruction method is designed to progressively iteratively optimize the Gaussian kernel parameters of the Gaussian kernel representation modeled according to the target scanned object. Since the traditional reconstruction result without smoothing filtering can retain high-frequency information as much as possible, the preliminary Gaussian representation obtained by iteratively optimizing the initial Gaussian representation in the image domain can effectively present the detailed information of the scanned object. The final refined Gaussian representation obtained by further iteratively optimizing the preliminary Gaussian representation in the projection domain using the projection data information obtained by the scan can effectively suppress artifacts and noise in the reconstructed image, thereby providing a three-dimensional image with higher clarity and contrast, and thus achieving accurate high-definition and high-contrast three-dimensional image reconstruction; and the anisotropic Gaussian regularization technology designed in this application takes into account the characteristics of CL imaging and Gaussian kernel representation, standardizes the representation behavior of the Gaussian kernel in multiple directions, and thus restores the inter-layer resolution; and the mapping relationship between the virtual detection plane constructed in this application and the real projection data can ensure that the rendering formula of the three-dimensional Gaussian splash is adaptable to various CL scanning methods and applicable to various scanning scenarios, greatly improving the practicality of this application. This solves the problem in related technologies that, in actual application scenarios, the aliasing artifacts of CL images are essentially artifacts caused by insufficient projection data and thus missing information. Iterative reconstruction or deep learning methods can only improve but not completely eliminate this artifact, and the reconstructed image also lacks clarity and realism.

[0165] Next, a two-step anisotropic Gaussian sputtering device according to an embodiment of the present application will be described with reference to the accompanying drawings.

[0166] Figure 6Schematic diagram of the structure of a two-step anisotropic Gaussian sputtering device according to an embodiment of the present application.

[0167] like Figure 6 As shown, the two-step anisotropic Gaussian sputtering device 10 includes: a first optimization module 100 , a second optimization module 200 and a voxelization module 300 .

[0168] Among them, the first optimization module 100 is used to initialize the target scanning object through the projection data of the target scanning object to obtain an initial Gaussian representation of the target scanning object, and iteratively optimize the Gaussian kernel parameters of the initial Gaussian representation until a first preset iteration stop condition is met to obtain a preliminary Gaussian representation.

[0169] The second optimization module 200 is used to iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation when the scanning method of the target scanning object satisfies the camera model until a second preset iterative stop condition is met to obtain a final fine Gaussian representation; otherwise, based on the scanning method, a mapping relationship between the virtual detection plane and the projection data is constructed, and based on the mapping relationship, the Gaussian kernel parameters of the preliminary Gaussian representation are iteratively optimized until the second preset iterative stop condition is met to obtain a final fine Gaussian representation.

[0170] The voxelization module 300 is used to discretize the three-dimensional volume corresponding to the final refined Gaussian representation into a voxel representation, so as to generate a computer tomography reconstruction result of the target scanned object according to the voxel representation.

[0171] Optionally, in one embodiment of the present application, the first optimization module 100 includes: an initialization unit and a modeling unit.

[0172] The initialization unit is used to initialize the target scan object based on the projection data to obtain a point cloud representation of the target scan object.

[0173] The modeling unit is used to model the points corresponding to the point cloud representation as corresponding three-dimensional radiation Gaussian kernels, so as to combine the three-dimensional radiation Gaussian kernels to obtain an initial Gaussian representation.

[0174] Optionally, in one embodiment of the present application, the first optimization module 100 includes: a generation unit and a first optimization unit.

[0175] The generating unit is configured to discretize the projection data and the target scanned object in the image domain based on the detector resolution and the preset reconstruction resolution of the target computer tomography system, so as to generate an initial reconstruction result in combination with the geometric relationship of the target computer tomography system.

[0176] The first optimization unit is configured to iteratively optimize Gaussian kernel parameters based on the initial reconstruction result and the voxel representation corresponding to the initial Gaussian representation until a first preset iterative stopping condition is satisfied, so as to obtain a preliminary Gaussian representation.

[0177] Optionally, in one embodiment of the present application, the second optimization module 200 includes: a first determination unit and a second determination unit.

[0178] The first determining unit is configured to determine a virtual detector according to a light source point of a target computer tomography imaging system, so as to determine a coordinate mapping relationship between the virtual detector and a real detector at multiple projection angles based on a scanning mode and an imaging geometric relationship.

[0179] The second determining unit is configured to determine a mapping relationship between the virtual detection plane and the projection data according to the coordinate mapping relationship.

[0180] Optionally, in one embodiment of the present application, the second optimization module 200 includes: a calculation unit and a second optimization unit.

[0181] Among them, the calculation unit is used to calculate the rendering projection value of the target scanned object at the target angle based on the mapping relationship, and construct a loss function according to the rendering projection value and the real projection value of the target scanned object at the target angle.

[0182] The second optimization unit is used to iteratively optimize the Gaussian kernel parameters of the preliminary Gaussian representation according to the loss function until a first preset iteration stopping condition is met to obtain a final refined Gaussian representation.

[0183] Optionally, in one embodiment of the present application, the calculation unit includes: a construction subunit and a determination subunit.

[0184] The construction subunit is used to construct a data fidelity item between the rendered projection value and the real projection value.

[0185] A subunit is determined, which is used to determine the loss function based on the data fidelity term, or to determine the loss function by combining the data fidelity term and the anisotropic regularization term for constraining the representation behavior of the Gaussian kernel in multiple directions according to the target ratio.

[0186] It should be noted that the above explanation of the embodiment of the two-step anisotropic Gaussian sputtering method is also applicable to the two-step anisotropic Gaussian sputtering device of this embodiment, and will not be repeated here.

[0187] According to the two-step anisotropic Gaussian sputtering device proposed in the embodiment of the present application, the target scanned object scanned by the target computer tomography imaging system can be modeled as an initial Gaussian representation, and by using the results of the traditional reconstruction algorithm and the scanned real projection data as supervision signals, the Gaussian parameters are iteratively optimized in the image domain and the projection domain to obtain the final fine Gaussian representation of the target scanned object and the corresponding computer tomography reconstruction result. Thus, a two-step reconstruction method is designed to progressively iteratively optimize the Gaussian kernel parameters of the Gaussian kernel representation modeled according to the target scanned object. Since the traditional reconstruction result without smoothing filtering can retain high-frequency information as much as possible, the preliminary Gaussian representation obtained by iteratively optimizing the initial Gaussian representation in the image domain in this application can effectively present the detailed information of the scanned object, and then further iteratively optimize the preliminary Gaussian representation in the projection domain using the projection data information obtained by scanning to obtain the final fine Gaussian representation, which can effectively suppress the artifacts and noise of the reconstructed image, thereby providing a three-dimensional image with higher clarity and contrast, and thus achieving accurate high-definition and high-contrast three-dimensional image reconstruction; and the anisotropic Gaussian regularization technology designed in this application takes into account the characteristics of CL imaging and Gaussian kernel representation, standardizes the representation behavior of the Gaussian kernel in multiple directions, and thus restores the inter-layer resolution; and the mapping relationship between the virtual detection plane and the real projection data constructed in this application can ensure that the rendering formula of the three-dimensional Gaussian splash is suitable for various CL scanning methods, greatly improving the practicality of this application. This solves the problem in related technologies that, in actual application scenarios, the aliasing artifacts of CL images are essentially artifacts caused by insufficient projection data and thus missing information. Iterative reconstruction or deep learning methods can only improve but not completely eliminate this artifact. CL reconstruction methods based on 3DGS are relatively scarce, and most of them focus on adjusting the Gaussian representation to adapt to X-ray imaging, rather than customizing special reconstruction strategies based on X-ray characteristics. The clarity and realism of the reconstructed images are also lacking.

[0188] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device may include:

[0189] A memory 701 , a processor 702 , and a computer program stored in the memory 701 and executable on the processor 702 .

[0190] When the processor 702 executes the program, the two-step anisotropic Gaussian sputtering method provided in the above embodiment is implemented.

[0191] Furthermore, the electronic device further includes:

[0192] The communication interface 703 is used for communication between the memory 701 and the processor 702 .

[0193] The memory 701 is used to store computer programs that can be run on the processor 702 .

[0194] The memory 701 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0195] If the memory 701, processor 702, and communication interface 703 are implemented independently, the communication interface 703, memory 701, and processor 702 can be connected to each other via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0196] Optionally, in a specific implementation, if the memory 701, the processor 702 and the communication interface 703 are integrated on a chip, the memory 701, the processor 702 and the communication interface 703 can communicate with each other through an internal interface.

[0197] The processor 702 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0198] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned two-step anisotropic Gaussian sputtering method.

[0199] An embodiment of the present application also provides a computer program product, including a computer program, which can run computer instructions. When the computer instructions are executed by a processor, the two-step anisotropic Gaussian sputtering method provided in the embodiment of the present application is implemented.

[0200] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0201] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0202] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing a custom logical function or process step, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed in a different order than shown or discussed, including performing functions in a substantially simultaneous manner or in a reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application pertain.

[0203] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or N wires (electronic devices), a portable computer disk cartridge (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically by optically scanning the paper or other medium and then editing, interpreting, or otherwise processing in a suitable manner as necessary, and then storing it in a computer memory.

[0204] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. If implemented using hardware, as in another embodiment, it can be implemented using any one or a combination of the following technologies known in the art: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0205] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0206] In addition, the functional units in the various embodiments of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into a module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0207] The storage medium mentioned above may be a read-only memory, a magnetic disk, or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application. Persons skilled in the art may make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A two-step anisotropic Gaussian sputtering method, characterized in that: The following steps are involved: Initializing the target scan object using projection data of the target scan object to obtain an initial Gaussian representation of the target scan object, and iteratively optimizing Gaussian kernel parameters of the initial Gaussian representation until a first preset iteration stop condition is satisfied to obtain a preliminary Gaussian representation; If the scanning mode of the target scanning object satisfies the camera model, iteratively optimizing Gaussian kernel parameters of the preliminary Gaussian representation until a second preset iterative stop condition is satisfied, so as to obtain a final refined Gaussian representation; otherwise, constructing a mapping relationship between a virtual detection plane and the projection data based on the scanning mode, and iteratively optimizing Gaussian kernel parameters of the preliminary Gaussian representation based on the mapping relationship until the second preset iterative stop condition is satisfied, so as to obtain the final refined Gaussian representation; The three-dimensional volume corresponding to the final refined Gaussian representation is discretized into a voxel representation, so as to generate a computer tomography reconstruction result of the target scanned object according to the voxel representation.

2. The method according to claim 1, characterized in that Initializing the target scan object using projection data of the target scan object to obtain an initial Gaussian representation of the target scan object includes: Initializing the target scan object based on the projection data to obtain a point cloud representation of the target scan object; Points corresponding to the point cloud representation are modeled as corresponding three-dimensional radiation Gaussian kernels, so as to combine the three-dimensional radiation Gaussian kernels to obtain the initial Gaussian representation.

3. The method according to claim 1, characterized in that The iterative optimization of the Gaussian kernel parameters of the initial Gaussian representation until a first preset iteration stop condition is satisfied to obtain a preliminary Gaussian representation includes: discretizing the projection data and the target scanned object based on a detector resolution and a preset reconstruction resolution of a target computed tomography system to generate a conventional reconstruction result in combination with a geometric relationship of the target computed tomography system and a target reconstruction strategy; Based on the traditional reconstruction result and the voxel representation corresponding to the initial Gaussian representation, the Gaussian kernel parameters are iteratively optimized until the first preset iterative stopping condition is met to obtain the preliminary Gaussian representation.

4. The method according to claim 1, wherein The constructing of a mapping relationship between a virtual detection plane and the projection data based on the scanning mode includes: determining a virtual detector according to a geometric relationship between a light source point and a real detector of the target computer tomography imaging system, wherein a combination of the virtual detector and the light source point satisfies a standard camera model; Determining a coordinate mapping relationship between the virtual detector and the real detector at multiple projection angles according to the scanning mode and the imaging geometric relationship; A mapping relationship between a virtual detection plane and the projection data is determined according to the coordinate mapping relationship.

5. The method according to claim 1, characterized in that The iteratively optimizing the Gaussian kernel parameters of the preliminary Gaussian representation based on the mapping relationship until the second preset iteration stop condition is satisfied to obtain the final refined Gaussian representation includes: Based on the mapping relationship, calculating the rendered projection value of the target scanned object at the target angle, and constructing a loss function according to the rendered projection value and the real projection value of the target scanned object at the target angle; The Gaussian kernel parameters of the preliminary Gaussian representation are iteratively optimized according to the loss function until the second preset iteration stopping condition is met to obtain the final refined Gaussian representation.

6. The method according to claim 5, characterized in that The constructing a loss function according to the rendered projection value and the real projection value of the target scanned object at the target angle includes: Constructing a data fidelity term between the rendered projection value and the real projection value; The loss function is determined based on the data fidelity term, or the loss function is determined by combining the data fidelity term and the anisotropic regularization term for constraining the multi-directional representation behavior of the Gaussian kernel according to the target ratio.

7. A two-step anisotropic Gaussian sputtering device, characterized in that: include: a first optimization module, configured to initialize the target scan object using projection data of the target scan object to obtain an initial Gaussian representation of the target scan object, and iteratively optimize Gaussian kernel parameters of the initial Gaussian representation until a first preset iteration stop condition is satisfied, thereby obtaining a preliminary Gaussian representation; a second optimization module, configured to, if the scanning mode of the target scanning object satisfies the camera model, iteratively optimize Gaussian kernel parameters of the preliminary Gaussian representation until a second preset iterative stopping condition is satisfied, so as to obtain a final refined Gaussian representation; otherwise, construct a mapping relationship between a virtual detection plane and the projection data based on the scanning mode, and iteratively optimize Gaussian kernel parameters of the preliminary Gaussian representation based on the mapping relationship until the second preset iterative stopping condition is satisfied, so as to obtain the final refined Gaussian representation; A voxelization module is used to discretize the three-dimensional volume corresponding to the final fine Gaussian representation into a voxel representation, so as to generate a computer tomography reconstruction result of the target scanned object according to the voxel representation.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the two-step anisotropic Gaussian sputtering method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the two-step anisotropic Gaussian sputtering method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed, it is used to implement the two-step anisotropic Gaussian sputtering method according to any one of claims 1 to 6.

Images