Gaussian splash sparse-view x-ray 3d reconstruction method guided by anatomical prior

Abstract

The application discloses a Gauss splash sparse view X-ray three-dimensional reconstruction method based on an anatomical prior guide, and sequentially constructs a three-dimensional Gauss attenuation field and an X-ray differentiable projection model; a projection edge is estimated and fused to generate a projection saliency map through adaptive double-window estimation, and a three-channel anatomical prior body containing intensity, structural importance and tissue weight is obtained through back projection; Gauss full life cycle control is realized relying on the prior, multi-index importance scoring is designed to complete spatial barrel constraint proliferation, and a structure guardian prior protection pruning mechanism is proposed; hybrid ray sampling is carried out based on the saliency map, a time sequence weighted reconstruction loss is constructed, and optimization is completed in combination with TV regularization. Through integrated prior closed-loop modeling from the projection domain to the body space, the Gauss distribution is controlled in multiple dimensions, Gauss resources are preferentially allocated to clinical anatomical areas such as bones and soft tissues, invalid Gauss proliferation in the artifact area is inhibited, and the detail precision and image signal-to-noise ratio of CT reconstruction under sparse view are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of CT three-dimensional reconstruction technology, specifically involving a three-dimensional reconstruction method based on anatomical prior guidance using Gaussian splashing sparse perspective X-rays. Background Technology

[0002] Sparse-view CT reconstruction is a challenging task with significant application value in the field of medical image reconstruction. This task aims to recover diagnostically significant three-dimensional volumetric data using a limited number of X-ray projections, thereby reducing patient radiation dose, shortening scan time, and improving the applicability of CT imaging in low-dose, rapid acquisition scenarios. However, due to the insufficient number of projection views, sparse-view CT reconstruction is inherently a highly pathological inverse problem. Limited angle sampling leads to severe fringe artifacts and geometric ambiguity, while low-contrast structures such as soft tissue boundaries are often only clearly visible under specific diagnostic windows, further increasing the difficulty of accurate reconstruction. Existing sparse-view CT reconstruction methods can be broadly classified into the following four categories:

[0003] Traditional analytical and iterative reconstruction methods: Traditional CT reconstruction methods mainly rely on imaging physics models and mathematical optimization strategies. Typical analytical methods, such as the FDK algorithm, quickly recover 3D volume data through filtered backprojection, demonstrating high efficiency in conventional dense-view CT reconstruction. Iterative reconstruction methods, such as SART, continuously minimize projection errors to ensure consistency between the reconstructed volume and the acquired projection. To alleviate artifacts under sparse views, methods such as ASD-POCS further introduce total variational regularization, utilizing the sparsity of image gradients to constrain the reconstruction results, thereby suppressing stripe artifacts caused by undersampling to some extent. In recent years, some methods have also attempted to combine depth priors or learned regularization terms with traditional iterative frameworks to improve reconstruction quality under low-dose or sparse-view conditions.

[0004] Although traditional analytical and iterative methods have clear physical meaning and good interpretability, they still have significant shortcomings under severely sparse perspective conditions. Analytical methods are prone to strong aliasing artifacts due to insufficient projection perspective; while iterative methods can improve projection consistency, they are computationally expensive and have slow convergence speeds. Furthermore, methods based on total variational regularization often over-smooth the real anatomical structure while removing noise and artifacts, resulting in weakened soft tissue texture, weak boundaries, and fine bone structures, and even step-like artifacts, making it difficult to meet the structural fidelity requirements of clinical diagnosis.

[0005] Methods based on implicit neural representations: With the development of neural radiation fields and implicit neural representations, CT reconstruction methods based on coordinate mapping have gradually gained attention. These methods typically use multilayer perceptrons or coding networks to map three-dimensional spatial coordinates to attenuation coefficients at corresponding locations, and are supervised by differentiable rendering or ray integration and input projection. Methods such as NeAT, MedNeRF, and NAF introduce NeRF-style continuous field representations into medical CT reconstruction, recovering continuous three-dimensional attenuation fields by optimizing neural network parameters. Methods such as SAX-NeRF further combine structure-aware constraints to improve reconstruction results under sparse perspectives. In addition, multi-resolution hash coding, geometric priors, and large-scale pre-trained models have also been used to accelerate the convergence of implicit representations and enhance reconstruction capabilities.

[0006] While implicit neural representation-based methods can represent 3D volumetric data as continuous functions and achieve good results in some sparse viewpoint tasks, they still have several limitations. First, these methods typically rely on extensive ray sampling and dense network queries, resulting in long optimization times and difficulty meeting the demands for efficient reconstruction. Second, the implicit network representation lacks explicit geometric structure, making it difficult to directly control the generation, deletion, and refinement of local anatomical structures. Third, when strong learning or generation priors are introduced, the model may produce "illusionary" structures that do not perfectly match the patient's true projection, posing a potential risk in medical imaging scenarios. For clinically critical areas such as soft tissue boundaries and fine bone structures, implicit methods may also exhibit oversmoothing or boundary discontinuities.

[0007] The 3D Gaussian splashing method is an explicit 3D representation method that uses a set of 3D Gaussian kernels to represent scene or volume data. Each Gaussian kernel contains parameters such as position, scale, rotation, and density, and can be efficiently optimized through differentiable rendering. This type of method initially achieved significant results in novel perspective synthesis tasks and was subsequently extended to the fields of X-ray imaging and CT reconstruction. Methods such as [list of methods] transform the Gaussian splashing form in traditional graphics rendering into a radiation integral model suitable for X-ray absorption imaging. By analytically calculating the attenuation contribution of rays passing through the Gaussian kernel, CT reconstruction based on Gaussian representation is achieved. Methods such as X-Field further combine Gaussian representation with other reconstruction processes, demonstrating certain advantages in terms of efficiency and explicit representation.

[0008] However, most existing CT reconstruction methods based on 3D Gaussian splashing follow the gradient-driven proliferation and pruning strategy used in natural image rendering. This strategy determines the replication, splitting, and deletion of Gaussian kernels based on local rendering errors or positional gradients. In sparse-view CT, this strategy is unreliable: undersampling-induced stripe artifacts may have strong gradients, causing the model to incorrectly allocate a large number of Gaussian kernels to artifact regions; while the gradients of real soft tissue boundaries, low-contrast lesions, or delicate anatomical structures may be weaker, making them more susceptible to underestimation or deletion during pruning. Therefore, relying solely on gradient-based Gaussian lifecycle management can lead to Gaussian resource misallocation, resulting in problems such as floating artifacts, structural breaks, excessive smoothing of soft tissues, and insufficient detail in bone structures.

[0009] Methods guided by anatomical or structural priors: To improve the structural reliability of sparse-view CT reconstruction, existing research has attempted to introduce geometric constraints, structural priors, learned priors, or volume data initialization results to assist reconstruction. For example, some methods provide initial volume data through coarse reconstruction results, while others utilize edge, gradient, segmentation, or multi-view feature backprojection to enhance structural perception capabilities. Still others attempt to highlight specific anatomical regions using tissue window width and level or clinical priors. The basic idea behind these methods is to reduce the uncertainty caused by sparse projection through additional structural information, enabling the reconstruction process to not only minimize projection errors but also maintain a more reasonable anatomical morphology.

[0010] Furthermore, existing prior-guided methods typically only function at a specific stage of reconstruction, such as initialization, regularization, or post-processing enhancement, and are not yet fully involved in the complete lifecycle management of the 3D Gaussian representation. Especially in Gaussian splash CT reconstruction, effective solutions remain lacking for how to automatically extract reliable tissue-related structural priors from the projection domain, how to transform 2D projection saliency information into 3D volume space priors, and how to utilize these priors to simultaneously guide Gaussian initialization, proliferation, pruning, and optimized sampling.

[0011] Therefore, it is necessary to propose an anatomically prior-guided Gaussian splashing method for sparse-view CT reconstruction, so that the distribution and evolution of the Gaussian kernel are more consistent with the real anatomical structure, thereby reducing artifacts while preserving soft tissue boundaries and bone structure details. Summary of the Invention

[0012] Purpose of the invention: To retain the advantages of the three-dimensional Gaussian splashing method, such as explicit representation, efficient differentiable rendering, and ease of local refinement, while solving its problems of being easily misled by stripe artifacts, insufficient representation of weak soft tissue boundaries, and lack of anatomical constraints in Gaussian proliferation and pruning in sparse-view CT reconstruction, this invention proposes a sparse-view X-ray three-dimensional reconstruction method based on anatomical prior guidance. The initialization, proliferation, pruning, and optimization process of Gaussian kernels is modeled as a structure-aware resource allocation problem, rather than a refinement process that simply relies on local rendering gradients.

[0013] Technical solution: The present invention provides a method for three-dimensional reconstruction using Gaussian splashing sparse-view X-rays based on anatomical prior guidance, comprising the following steps:

[0014] Step 1: Construct a three-dimensional Gaussian decay field and X-ray differentiable projection rendering model;

[0015] Step 2: Construct the anatomical prior. First, based on the input sparse view X-ray projection, adaptively estimate the soft tissue window and hard tissue window. Then, fuse the diagnostic window response and the projection edge response to generate a projection-level saliency map. Based on the CT scan geometric parameters, back-project the projection-level saliency map to the volume space to finally generate a three-channel anatomical prior containing intensity channel, structural importance channel, and tissue weight channel.

[0016] Step 3: Perform structure-aware Gaussian lifecycle management. Utilize the three-dimensional volumetric three-channel anatomical prior obtained in Step 2 to generate an anatomical confidence map. This map guides the initialization, proliferation, and pruning of three-dimensional Gaussian primitives. Specifically, first, determine the Gaussian primitives to be proliferated and pruned through multi-index importance scoring. Then, constrain the number of Gaussian proliferations in different spatial regions through spatial bucket budget allocation. A structure-protected prior protection pruning mechanism sets protection markers for Gaussian primitives located in high-confidence anatomical structures. For unprotected Gaussian primitives with low contribution, low visibility, or high redundancy, perform patient soft pruning. Output the structure-aware updated set of three-dimensional Gaussian primitives and their lifecycle indices as the basis for subsequent importance-aware optimization. This ensures that Gaussian resources are preferentially distributed in real anatomical structures such as soft tissue boundaries and bone structures, reducing erroneous Gaussian proliferation and floating artifact Gaussian in spatial or stripe artifact regions, and preventing premature pruning of real weak structures, thereby improving the stability of subsequent optimization and the structural fidelity of the three-dimensional CT volumetric data.

[0017] Step 4: Importance perception optimization. First, perform hybrid ray sampling based on the projection-level saliency map and construct a saliency-weighted reconstruction loss. Combine this with the total variation regularization term to complete joint optimization. While ensuring global projection consistency, enhance the reconstruction effect of soft tissue boundaries and key clinical areas of bone structure.

[0018] The optimized 3D Gaussian ensemble output can be used to generate 2D X-ray projections from any viewpoint through ray tracing, or to obtain complete 3D CT volume data through voxel interpolation and mesh query.

[0019] This invention sequentially constructs a three-dimensional Gaussian attenuation field and an X-ray differentiable projection model; adaptive dual-window estimation and fusion of projection edges generate a projection saliency map, and back-projection yields a three-channel anatomical prior containing intensity, structural importance, and tissue weight; relying on the prior, Gaussian full life-cycle control is achieved, a multi-index importance score is designed to complete spatial bucket constraint proliferation, and a structural guardian prior protection pruning mechanism is proposed to avoid accidental deletion of key soft tissue structures; finally, hybrid ray sampling is carried out based on the saliency map, a time-weighted reconstruction loss is constructed, and optimization is completed by combining TV regularization.

[0020] Furthermore, the detailed method of step 1 is as follows:

[0021] Step 1.1, Representation of the three-dimensional Gaussian decay field;

[0022] Given a sparse viewpoint X-ray projection and its corresponding CT scan geometry, a set of three-dimensional Gaussian elements is used. To represent the CT attenuation field to be reconstructed;

[0023] No. Gao Siyuan Including the central location Covariance matrix and Gaussian decay density ,and ;

[0024] To ensure the positive semidefiniteness of the covariance matrix and improve optimization stability, the covariance matrix is ​​decomposed into:

[0025] ;

[0026] in, Let be a rotation matrix. It is a three-dimensional principal axis scale diagonal matrix; and They are and The transpose of the matrix;

[0027] Here, a small number of learnable Gaussian kernels can be used to express anatomical structures with different orientations and scales, such as bone structures, blood vessels, and organ boundaries;

[0028] Step 1.2: X-ray projection rendering model;

[0029] Unlike the synthesis of new perspectives from natural images, CT reconstruction requires line integration of the attenuation coefficients along the path of the ray. Therefore, this invention employs a differentiable projection model suitable for X-ray absorption imaging, for a ray passing through volume space... The predicted X-ray intensity obtained from its rendering is represented in Beer-Lambert form, as shown below:

[0030] ;

[0031] in, Indicates the intensity of incident X-rays. Indicates ray Intersecting sets of Gaussian elements, Indicates the first A Gaussian nuclear ray The analytical line integral;

[0032] This step allows you to render a 3D Gaussian representation as a 2D projection I(u,v) and then use the real projection to supervise Gaussian parameter optimization.

[0033] In clinical CT, different tissues are typically observed under different window widths and levels. Soft tissue boundaries are often clearer under soft tissue windows, while high-attenuation areas such as bone structures are more prominent under bone windows. Fixed window widths and levels rely on well-calibrated CT intensity and may be unreliable in sparse viewpoints or inadequately calibrated projection data.

[0034] Therefore, step 2 proposes an adaptive dual-window estimation method based on maximizing information entropy to automatically obtain the soft tissue window and the hard tissue window. The detailed method is as follows:

[0035] Step 2.1: Adaptive dual-window weight estimation, converting the original single-view projection... Attenuation values ​​are mapped to a pseudo-HU intensity space, a one-dimensional intensity histogram is calculated and smoothed; low-attenuation soft tissue intervals are defined based on global intensity quantiles. High-attenuation hard tissue region ], extract soft tissue peak values ​​respectively Hard tissue peak (Corresponding to soft tissue dominant response and hard tissue dominant response respectively). It is the 0th percentile. It is the 50th percentile. 100th percentile;

[0036] Then, a foreground region mask is generated to remove atmospheric background interference.

[0037] ;

[0038] Represents the projected pixel coordinates. Indicates the background threshold. This represents the projection attenuation value at pixel coordinate u of the projected image, and the background threshold is dynamically determined based on the global intensity range of the projection. The formula is:

[0039] ;

[0040] and These are the global minimum and maximum attenuation values ​​for projection, respectively;

[0041] Then, the standard window for projection is executed. Transformation:

[0042] ;

[0043] In the formula, WW is the window width and WL is the window level; the response after window transformation Discretized A grayscale range, and only in the foreground mask Internal Statistics The probability of each interval ;

[0044] Next, based on the information content of the window parameters quantized by Shannon entropy:

[0045] ;

[0046] It is a minimal numerical stability constant;

[0047] For both soft tissue and hard tissue , respectively in the peak Constrained search space Find the window parameters that maximize entropy within the soft tissue constraint search space. Hard-structured constrained search space Solving for the optimal window parameters using the inner maximum entropy function:

[0048] ;

[0049] Substituting the optimal window parameters into the window transform operator yields the soft tissue window response. and hard tissue window response ;

[0050] Since the two emphasize low-contrast soft tissue structures and high-contrast bone structures respectively, pixel-by-pixel max pooling fusion is used to obtain a tissue-perceived response with high recall. :

[0051] ;

[0052] Step 2.2: Projection-level saliency map generation; The obtained dual-window fusion response can highlight tissue-related areas, but it cannot fully express the local boundaries in the projection domain. To further introduce structural information, the dual-window fusion response is fused with the projection edge response to improve the original projection. Solving for the magnitude of the Sobel gradient , and dual-window fusion response Weighted fusion and min-max normalization are used to generate projection-level saliency maps. :

[0053] ;

[0054] In the formula, These are the weighting coefficients for edge response and tissue window response. Represents the min-max normalization operation; saliency map Larger weights are assigned to projected pixels with strong structural changes or organizational correlations, providing two-dimensional projection priors for subsequent prior construction and importance-aware sampling;

[0055] Step 2.3: Construction of the three-channel anatomical prior body. First, a small number of SART iterations are used to generate a coarse strength reference body through coarse reconstruction. (The coarse reconstruction is not the final result, but rather a rough attenuation scaffold for constructing anatomical priors); projecting saliency maps from all perspectives. The three-dimensional structural importance field is obtained by geometric backprojection from CT scans. Then, a three-dimensional three-channel anatomical prior body is constructed. Its value at the three-dimensional voxel coordinate x is expressed as:

[0056] ;

[0057] This represents the three-channel prior vector of the three-dimensional three-channel anatomical prior body at voxel coordinate x; The value for the intensity channel is used to characterize the coarse attenuation intensity at this voxel location; The value for the structural importance channel is used to characterize the degree of structural support obtained by backprojection of the multi-view projection saliency map at the voxel location; The tissue weight channel values ​​are used to characterize the tissue relevance of the voxel location under the responses of soft and hard tissue windows. The intensity channel values, structural importance channel values, and tissue weight channel values ​​together constitute a three-dimensional volumetric three-channel anatomical prior for subsequent Gaussian initialization, proliferation, pruning, and optimization.

[0058] Through rough reconstruction After cropping out 0.1% to 99.9% of the extreme values, normalize to [0,1] to obtain the intensity channel. For example, Extreme values ​​outside the 0.1% to 99.9th percentile are clipped and then normalized. To suppress the effects of anomalous projection values ​​and strong artifacts; the structural importance field obtained from backprojection Normalization yields structural importance channels Structural importance channel Used to indicate whether a three-dimensional location is continuously supported by significant structures across multiple projection views; in coarse-grained bodies Adaptive soft tissue window and hard tissue window estimation are performed again, and tissue weight channels are obtained by fusion through max pooling. , used to represent the tissue correlation of this three-dimensional location.

[0059] The obtained three-channel prior can simultaneously obtain coarse attenuation intensity, cross-view structural importance, and tissue window correlation, providing anatomical guidance for subsequent Gaussian initialization, proliferation, pruning, and optimization.

[0060] Traditional 3D Gaussian splashing typically uses random or uniform initialization methods, which can easily lead to the placement of Gaussian kernels in the spatial domain, artifact regions, or regions with insufficient constraints, resulting in instability in the early optimization. The structure-aware Gaussian lifecycle management in step 3 specifically includes the following processes:

[0061] Step 3.1: Anatomical perception Gaussian initialization, first fusing the structural importance of the three-channel prior. ) and organizational weight Generate anatomical confidence maps During the initialization phase, instead of uniformly sampling Gaussian centers across the entire reconstructed volume space, high-confidence regions are selected as the set of candidate Gaussian initialization centers. By using coarse strength channels Gaussian initial decay density is obtained by interpolation sampling. The initial scale is set based on the voxel neighborhood spacing;

[0062] ; The larger the value, the more likely the location is to be a structurally significant or tissue-related region;

[0063] ;

[0064] To initialize the confidence threshold; based on Probabilistic sampling or selecting the voxel with the highest confidence as the Gaussian center This design ensures that the Gaussian kernel is concentrated in a reasonably anatomically appropriate region from the outset, reducing erroneous initialization caused by fringe artifacts or spatial response.

[0065] Step 3.2: Since the candidate centers originate from discrete voxel meshes, directly calculating the neighborhood distance in the original coordinates may be affected by differences in scale across axes. To obtain more stable isolated point identification, anisotropic sensing initialization and cleanup are performed:

[0066] Map the candidate centers to a scale-normalized isotropic coordinate system and define the candidate points. K-nearest neighbor isolation:

[0067] ;

[0068] express of A set of nearest neighbor candidate points; if It was identified as an isolated noise point and removed. The isolation distance threshold is used; this step can remove discrete high-response points caused by noise or stripe artifacts, and improve the spatial continuity of the initial Gaussian distribution.

[0069] Step 3.3: If Gaussian proliferation and pruning are determined solely by instantaneous position gradients, they are easily affected by sparse view artifacts. Therefore, further tracking of the EMA lifecycle index is required here.

[0070] Maintenance contribution for each Gaussian element Residual response Location gradient Visibility and normalization scale Historical indicators are smoothed using exponential moving averages; any indicator Updated formula:

[0071] ;

[0072] In the formula, Here, t is the EMA decay coefficient, and t is the current iteration step number.

[0073] The contribution measure is the Gaussian's contribution to the rendered projection; the residual response measures its correlation with the current projection error; the position gradient measures the local optimization needs; visibility measures the frequency of its observation by rays; and the scale reflects its spatial coverage. EMA smoothing avoids anomalous gradients dominating lifecycle decisions in a single iteration.

[0074] Step 3.4: Perform multi-index importance scoring to filter out noise interference from sparse projection and achieve adaptive refinement of multi-scale human anatomical structures.

[0075] Step 3.5: If the Gaussian with the highest score is directly selected globally for multiplication, the new Gaussian may be concentrated in a few high-gradient artifact regions. Therefore, spatial bucket budget constraint multiplication is performed here. First, the 3D reconstruction space is divided into... Rule space bucket Given a global proliferation budget The local proliferation budget for the k-th spatial bucket is:

[0076] ;

[0077] For each bucket, within a single bucket, proliferation importance is ranked... Before selecting descending order A Gaussian multiplication process is performed; based on the normalization scale. With cloning threshold Distinguishing operations:

[0078] when Perform cloning: Add a small perturbation to the center of the original Gaussian to generate a new Gaussian;

[0079] when Perform splitting: along the largest principal axis of the covariance ellipsoid Generate two sub-Gaussians and scale them down proportionally, for example... The original Gaussian density is then distributed among the sub-Gaussians to avoid sudden increases in local decay values. This strategy can simultaneously improve the problems of underexpression of fine structures and oversmoothing of large-scale Gaussians.

[0080] Simultaneously along the principal direction of the original covariance ellipsoid Place a new center ;

[0081] ;

[0082] Step 3.6: Since conventional pruning methods typically remove Gaussians based on density, contribution, or gradient magnitude, but in sparse-view CT, the true soft tissue boundary may have a weak response in the early optimization stages and is easily mistakenly removed, StructureGuardian prior protection pruning is performed here. For Gaussians... Perform Gaussian protection marking: ;

[0083] For indicator functions, This is the threshold for anatomical protection. When the Gaussian is located in a high-confidence anatomical region, it is directly preserved during the routine pruning stage;

[0084] For unprotected Gaussians, a patient, soft pruning method is used, which is as follows:

[0085] A single degradation is achieved by multiplying the Gaussian density by the attenuation factor. And accumulate the number of penalties; when the number of penalties reaches... Then, score according to pruning. Delete the top M redundant Gaussians from highest to lowest; periodically perform KNN outlier cleanup to remove floating artifact Gaussians.

[0086] To determine which Gaussian nuclei should proliferate, the dissection prior and multiple optimization indicators are incorporated into the importance score. The multi-indicator importance score includes the Gaussian proliferation importance score and the Gaussian pruning score.

[0087] For the center is located Gauss The formula for calculating the Gaussian proliferation importance score is:

[0088] ;

[0089] This represents the anatomical prior guidance intensity coefficient. It is the numerical stability constant; This is the scale adaptation factor;

[0090] The first term increases the priority of Gaussian proliferation located in high anatomical confidence regions; gradient term This indicates the need for local optimization; the geometric mean of contribution, visibility, and residual response ensures that the selected Gaussian simultaneously contributes to rendering, supports observations, and is correlated with errors.

[0091] For cloning operations Splitting operation ;

[0092] This allows for higher splitting priority to larger Gaussian structures, which are then used to refine excessively large Gaussian structures into smaller structural units.

[0093] Accordingly, to determine which Gaussian kernels should be pruned, a Gaussian pruning score is used. The formula for calculating the Gaussian pruning score is as follows:

[0094] ;

[0095] The higher the pruning score, the weaker the Gaussian is in terms of contribution, residual correlation, gradient requirement and visibility, and it is more likely to be a redundant Gaussian or a floating Gaussian induced by artifacts.

[0096] The final output of this invention is the reconstructed three-dimensional CT attenuation field, specifically represented as an optimized set of three-dimensional Gaussian elements. Through joint optimization of Gaussian parameters (including center position, covariance matrix, and attenuation density), a continuous representation of the complete three-dimensional volume data is achieved. The detailed method for step 4 is as follows:

[0097] Step 4.1: In sparse-view CT projection, a large number of rays correspond to background or low-information areas, while key structures such as soft tissue boundaries, bone structures, and blood vessels only account for a small proportion. If uniform ray sampling is used, the optimization budget will be consumed by a large number of easy-to-access areas, resulting in slow convergence of structural regions. Therefore, importance-aware hybrid ray sampling is performed here (to obtain the sampled ray set and its probability distribution):

[0098] For the The mixed sampling probability of ray r under each viewpoint:

[0099] In the formula, For saliency sampling weights, For the first Total number of pixels per viewpoint For rays The saliency value of the corresponding pixel;

[0100] Importance-aware hybrid ray sampling allows the model to focus on optimizing structure-related rays while maintaining global coverage through uniform sampling, thus avoiding ignoring low-response regions.

[0101] Step 4.2: For the importance-aware hybrid ray sampling set, perform a saliency-weighted reconstruction loss. First, calculate the dynamic weight of each ray as iterative scheduling changes:

[0102] ;

[0103] These are the parameters for iterative timing scheduling;

[0104] Early training can make A value as small as or even 0 allows the model to initially achieve a globally consistent coarse reconstruction; the value is then gradually increased in the later stages of training. This shifts the focus of optimization to structurally significant and organizationally relevant regions; Normalized weighted L1 reconstruction loss:

[0105] ;

[0106] In the formula, Render the predicted intensity for the model. For true projected intensity, For the set of sampled rays from the v-th viewpoint;

[0107] loss While retaining all sampled ray supervision, higher weights are assigned to anatomically relevant regions to improve detail reconstruction;

[0108] Step 4.3: To further suppress local oscillations and noise under sparse supervision, a volume total variational regularization term is introduced in addition to the reconstruction loss to suppress noise artifacts, resulting in the complete loss function. :

[0109] ;

[0110] In the formula, V is the three-dimensional density mesh obtained by Gaussian voxelization. For the total variational regularization term, The regularization term weight coefficient;

[0111] Iteratively optimize all three-dimensional Gaussian elements; the optimized set of Gaussian elements is the final three-dimensional CT attenuation field reconstruction result.

[0112] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0113] (1) Improve the adaptability and reliability of anatomical prior extraction in sparse-view X-ray 3D reconstruction.

[0114] This invention automatically estimates soft tissue and hard tissue windows based on input sparse X-ray projection, and determines window width and level parameters by maximizing information entropy. It does not rely on fixed empirical window widths and levels, nor does it require additional manual annotation, organ segmentation results, or external anatomical models. By fusing the soft tissue window response, hard tissue window response, and Sobel edge response, this invention can simultaneously highlight low-contrast soft tissue boundaries and high-attenuation bone structure regions, extracting more salient information from sparse projection that better conforms to medical anatomy, thus providing stable structural priors for subsequent 3D reconstruction.

[0115] (2) Realize the anatomical prior closed-loop modeling from the two-dimensional projection domain to the three-dimensional volume space, and improve the three-dimensional structural constraint capability.

[0116] This invention backprojects a projection-level saliency map onto a three-dimensional volumetric space according to the geometry of a CT scan, and combines this with a coarse intensity reference obtained from a small number of SART iterations to construct a three-dimensional volumetric three-channel anatomical prior, including an intensity channel, a structural importance channel, and a tissue weight channel. This three-channel anatomical prior simultaneously characterizes the coarse attenuation distribution, the degree of consistent structural support across multiple views, and tissue correlation. This ensures that subsequent Gaussian initialization, proliferation, pruning, and optimization no longer rely solely on local projection errors or instantaneous gradients, but are continuously constrained by the volumetric anatomical structural information, which is beneficial for improving the structural consistency and anatomical reliability of three-dimensional CT volumetric data.

[0117] (3) Improves the problem that Gaussian elements are easily misled by artifact gradients in the traditional Gaussian splashing method, and improves the accuracy of Gaussian resource allocation.

[0118] This invention transforms the proliferation process of Gaussian units into a resource allocation process based on anatomical structure perception through anatomical confidence maps, multi-index importance scoring, and spatial bucket budget allocation. This allows newly added Gaussian units to be preferentially distributed in real anatomical structure regions such as soft tissue boundaries, bone structures, and organ contours, reducing invalid Gaussian units, floating Gaussian units, and background Gaussian units induced by stripe artifacts, thereby improving the spatial compactness and structural alignment of Gaussian representations.

[0119] (4) It can protect the boundaries of weak soft tissues and small anatomical structures, reducing the risk of real structures being accidentally pruned.

[0120] This invention employs a structure-protective prior protection pruning mechanism. Gaussian elements located in high-confidence anatomical regions are marked with protective tags, ensuring their preservation during routine pruning. For unprotected Gaussian elements with low contribution, low visibility, or high redundancy, patient soft pruning and periodic KNN isolated point cleanup are used. This effectively removes redundant Gaussian elements in artifact and isolated regions while preventing premature deletion of genuine weak structures, thus improving the continuity and integrity of soft tissue boundaries, cortical bone edges, and fine structures.

[0121] (5) The optimization process has increased its focus on key clinical areas while also taking into account the consistency of global projection.

[0122] This invention performs importance-aware hybrid ray sampling based on projection-level saliency maps, allowing more sampling rays to pass through high-information regions such as soft tissue boundaries and bone structures, while retaining a certain proportion of uniform sampling rays to cover the global region. This avoids focusing only on local salient structures and compromising overall projection consistency. Furthermore, this invention employs a saliency-weighted reconstruction loss that varies with the iteration process, prioritizing the learning of global projection consistency in the early stages of training and gradually strengthening the reconstruction constraints of structurally salient regions in the later stages. Combined with a total variational regularization term, it suppresses local noise and density oscillations under sparse supervision, thereby improving optimization stability and reconstruction detail quality.

[0123] (6) Improve the image quality, structural fidelity and applicability of sparse view CT reconstruction results.

[0124] This invention effectively suppresses streak artifacts, suspension artifacts, and background noise under conditions of limited projection viewpoints, reducing problems such as overly smoothed soft tissues, blurred bone structure edges, and local structural fractures. Furthermore, due to the use of explicit three-dimensional Gaussian primitive representation, the optimized Gaussian set can be used to render X-ray projections from any viewpoint, and complete three-dimensional CT volume data can be obtained through voxel queries or interpolation. Verification through examples shows that this invention achieves superior performance in three-dimensional volume data metrics, two-dimensional projection metrics, soft tissue window metrics, and bone window metrics, especially improving the reconstruction effect of low-contrast soft tissues and high-frequency bone structures. It is suitable for sparse-view X-ray CT three-dimensional reconstruction scenarios such as low-dose, limited projection, and rapid acquisition. Attached Figure Description

[0125] Figure 1 This is a schematic diagram of the overall network of the present invention;

[0126] Figure 2 This is a visual comparison of the existing solutions of the present invention in the embodiments;

[0127] Figure 3 This is a schematic diagram illustrating the visualization effect of the present invention on an additional evaluation set in the embodiments;

[0128] Figure 4This is a schematic diagram of the peak detection process based on the projection intensity histogram in the embodiment;

[0129] Figure 5 This is a schematic diagram of the window width and window level search process based on maximizing information entropy in the embodiment.

[0130] Figure 6 This is a schematic diagram of the projection response results obtained through the optimal soft tissue window in the embodiment;

[0131] Figure 7 This is a schematic diagram of the projection response results obtained through the optimal hard tissue window in the embodiment;

[0132] Figure 8 This is a schematic diagram of the result of multi-window fusion of the soft tissue window response and the hard tissue window response in the embodiment. Detailed Implementation

[0133] The technical solution of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0134] This invention presents a Gaussian splash sparse-view X-ray 3D reconstruction method based on anatomical prior guidance. It extracts diagnostically relevant structural information from the projection through adaptive dual-window and edge fusion; constructs a 3D anatomical prior through back-projection; and then, through prior-guided initialization, multi-index proliferation, protective pruning, and importance-aware optimization, allows the Gaussian kernel to be allocated and evolved around the real anatomical structure throughout its entire lifecycle. Compared with traditional gradient-driven Gaussian methods, this invention reduces ineffective Gaussian growth in artifact regions, protects weak but important soft tissue structures, and improves the reconstruction quality of bone structures and low-contrast boundaries. Using sparse-view X-ray projection as input, the anatomical prior is embedded into the complete lifecycle of the 3D Gaussian unit, thereby achieving structure-aware reconstruction for sparse-view CT. The overall process is as follows: Figure 3 As shown.

[0135] This includes anatomical prior construction, structure-aware Gaussian lifecycle management, and importance-aware optimization. First, tissue-related and structure-related salient information is extracted from the input projection to construct a 3D anatomical prior. Subsequently, this prior is used to guide the initialization, proliferation, and pruning of Gaussian units, ensuring that Gaussian resources are preferentially distributed in real anatomical structural regions. Finally, through saliency-guided sampling and weighted optimization, the model further refines soft tissue boundaries and bone structures while maintaining global projection consistency.

[0136] For the input sparse X-ray projection, this invention first obtains the soft tissue window and hard tissue window responses through adaptive dual-window estimation, and then generates a projection-level saliency map by combining Sobel edge information. Subsequently, the projection-level saliency map is back-projected onto the three-dimensional volume space according to the CT scan geometry to obtain the structural importance volume; simultaneously, a coarse intensity volume is obtained using a small number of SART iterations. Finally, the coarse intensity information, structural importance information, and tissue window response information are fused into a three-channel anatomical prior. This prior not only reflects the coarse attenuation distribution of the reconstructed region, but also encodes cross-viewpoint consistent structural saliency and tissue correlation, providing a basis for subsequent Gaussian initialization, protective pruning, and budget-constrained proliferation. A unified anatomical confidence map is generated through the anatomical prior, and the importance of each Gaussian unit is scored by combining multiple indicators such as contribution, residual response, position gradient, visibility, and scale. For Gaussians requiring proliferation, the system selects candidate Gaussians using a space bucket budget allocation strategy and performs cloning or splitting operations based on scale. For Gaussians requiring pruning, the system protects real structures located in high anatomical confidence regions using a Structure Guardian mechanism, while removing low-contribution, low-visibility, or isolated artifact Gaussians. This process reduces the erroneous allocation of Gaussian resources to artifact regions and improves the three-dimensional continuity of soft tissue and bone structures.

[0137] This invention does not merely improve the X-ray projection rendering model, but rather incorporates the anatomical priors extracted from the projection domain throughout the entire lifecycle of the Gaussian representation, so that initialization, proliferation, pruning, and optimization are all constrained by the anatomical structure, thereby obtaining CT reconstruction results with higher structural fidelity under sparse view conditions.

[0138] This embodiment specifically includes the following steps:

[0139] Step 1: Construct a three-dimensional Gaussian decay field and X-ray differentiable projection rendering model;

[0140] Step 2: Construct the anatomical prior. First, based on the input sparse X-ray projection, adaptively estimate the soft tissue window and hard tissue window. Then, fuse the diagnostic window response and the projection edge response to generate a projection-level saliency map. Based on the CT scan geometry, back-project the projection-level saliency map to the volume space to finally generate a three-channel anatomical prior containing intensity channel, structural importance channel, and tissue weight channel.

[0141] Step 3: Perform structure-aware Gaussian lifecycle management. Utilize the three-dimensional volumetric three-channel anatomical prior obtained in Step 2 to generate an anatomical confidence map. This map guides the initialization, proliferation, and pruning of three-dimensional Gaussian primitives. Specifically, first, determine the Gaussian primitives to be proliferated and pruned through multi-index importance scoring. Then, constrain the number of Gaussian proliferations in different spatial regions through a spatial bucket budget allocation. Finally, set protection markers for Gaussian primitives located in high-confidence anatomical structural regions using a structure-guarding prior protection pruning mechanism. Perform patient soft pruning on unprotected Gaussian primitives with low contribution, low visibility, or high redundancy. Output the structure-aware updated set of three-dimensional Gaussian primitives and their lifecycle indices.

[0142] Step 4: Importance-aware optimization. First, hybrid ray sampling is performed based on the projection-level saliency map, and a saliency-weighted reconstruction loss is constructed. Combined with the total variational regularization term, joint optimization is completed. Under the premise of ensuring global projection consistency, the reconstruction effect of soft tissue boundaries and key clinical areas of bone structure is enhanced, and finally complete three-dimensional CT volume data is obtained.

[0143] This embodiment of the anatomical prior-guided Gaussian splash sparse-view X-ray 3D reconstruction method involves an anatomical prior volume construction module, a structure-aware Gaussian lifecycle management module, and an importance-aware optimization module. The anatomical prior volume construction module adaptively estimates soft tissue and hard tissue windows from the input sparse X-ray projection, fuses the diagnostic window response with the projection edge response to generate a projection-level saliency map, and back-projects it onto volume space under CT scan geometry to construct a three-channel anatomical prior volume containing intensity, structural importance, and tissue weight. The structure-aware Gaussian lifecycle management module utilizes the anatomical prior to guide Gaussian initialization, proliferation, and pruning. Through multi-index importance scoring, space bucket budget allocation, and the Structure Guardian protection mechanism, it prioritizes the allocation of Gaussian resources to real anatomical structure regions and suppresses erroneous proliferation in artifact regions. The importance-aware optimization module uses the projection-level saliency map for mixed ray sampling and saliency-weighted reconstruction loss, ensuring that the optimization process maintains global projection consistency while paying more attention to clinically critical areas such as soft tissue boundaries and bone structures.

[0144] The detailed method for step 1 in this embodiment is as follows:

[0145] Step 1.1, Representation of the three-dimensional Gaussian decay field;

[0146] Given a sparse viewpoint X-ray projection and its corresponding CT scan geometry, a set of three-dimensional Gaussian elements is used. To represent the CT attenuation field to be reconstructed; It is the number of Gaussian elements in the set;

[0147] No. Gao Siyuan Including the central location Covariance matrix Non-negative decay density ;

[0148] To ensure the positive semidefiniteness of the covariance matrix and improve optimization stability, the covariance matrix is ​​decomposed into:

[0149] ;

[0150] in, Let be a rotation matrix. It is a three-dimensional principal axis scale diagonal matrix;

[0151] Step 1.2, X-ray projection rendering model; that is, using a differentiable projection model suitable for X-ray absorption imaging, for a ray passing through volume space. The predicted X-ray intensity obtained from its rendering is represented in Beer-Lambert form, as shown below:

[0152] ;

[0153] in, Indicates the intensity of incident X-rays. Indicates ray Intersecting sets of Gaussian elements, Indicates the first A Gaussian nuclear ray The analytical line integral.

[0154] The detailed method for step 2 in this embodiment is as follows:

[0155] Step 2.1, Adaptive dual-window weight estimation: First, the original single-view projection is... Attenuation values ​​are mapped to a pseudo-HU intensity space, a one-dimensional intensity histogram is calculated and smoothed; low-attenuation soft tissue intervals are defined based on global intensity quantiles. High-attenuation hard tissue region ], extract soft tissue peak values ​​respectively Hard tissue peak ;

[0156] Then generate a foreground region mask. To eliminate atmospheric background interference:

[0157] ;

[0158] Represents the projected pixel coordinates. This represents the background threshold, which is dynamically determined based on the global intensity range of the projection. The formula is:

[0159] ;

[0160] and These are the global minimum and maximum attenuation values ​​for projection, respectively;

[0161] Secondly, regarding the candidate window width and window position Perform a standard window transformation, using the following formula:

[0162] ;

[0163] Then the response after window transformation Discretized A grayscale range, and only in the foreground mask Internal Statistics The probability of each interval The information content of the candidate window parameters is quantified using Shannon entropy, and the quantification formula is as follows: ;

[0164] It is a minimal numerical stability constant;

[0165] Next, in the soft tissue constraint search space Hard-structured constrained search space Solving for the optimal window parameters using the inner maximum entropy function:

[0166] ;

[0167] Substituting the optimal window parameters into the window transform operator yields the soft tissue window response. and hard tissue window response ;

[0168] Finally, pixel-by-pixel max pooling is used to fuse the dual-window response:

[0169] ;

[0170] Step 2.2, Projection-level saliency map generation, including:

[0171] For the original projection Solving for the magnitude of the Sobel gradient , and dual-window fusion response Weighted fusion and min-max normalization are used to generate projection-level saliency maps. :

[0172] ;

[0173] In the formula, These are the weighting coefficients for edge response and tissue window response. This represents the min-max normalization operation;

[0174] Step 2.3: Construction of the three-channel anatomical prior body. First, a small number of SART iterations are used to generate a coarse strength reference body through coarse reconstruction. Projecting saliency maps from all viewpoints The three-dimensional structural importance field is obtained by geometric backprojection from CT scans. Then, a three-dimensional, three-channel anatomical prior body is constructed:

[0175] ;

[0176] In the formula, x represents the three-dimensional voxel coordinates; the three channels mentioned above include the intensity channel. Structural importance channel and organizational weight channel .

[0177] Step 3 of this embodiment, the structure-aware Gaussian lifecycle management, specifically includes the following processes:

[0178] Step 3.1: Anatomical perception Gaussian initialization, firstly fusing the structural importance of the three-channel anatomical prior. With organizational weight Generate anatomical confidence maps Then, high-confidence regions are selected as the candidate center set for Gaussian initialization. By using coarse strength channels Gaussian initial decay density is obtained by interpolation sampling. The initial scale is set based on the voxel neighborhood spacing;

[0179] ;

[0180] ;

[0181] To initialize the confidence threshold; based on Probabilistic sampling or selecting the voxel with the highest confidence as the Gaussian center ;

[0182] Step 3.2, Anisotropic sensing initialization and cleanup:

[0183] Map the candidate centers to a scale-normalized isotropic coordinate system and define the candidate points. K-nearest neighbor isolation:

[0184] ;

[0185] express of A set of nearest neighbor candidate points; if It was identified as an isolated noise point and removed. The isolation distance threshold;

[0186] Step 3.3, EMA Lifecycle Indicator Tracking:

[0187] Maintenance contribution for each Gaussian element Residual response Location gradient Visibility and normalization scale Historical indicators are smoothed using exponential moving averages; any indicator Updated formula:

[0188] ;

[0189] In the formula, Here, t is the EMA decay coefficient, and t is the current iteration step number.

[0190] Step 3.4: Perform multi-index importance scoring to filter out noise interference from sparse projection and achieve adaptive refinement of multi-scale human anatomical structures.

[0191] Step 3.5, Space Bucket Budget Constraint Proliferation: First, divide the 3D reconstructed space into... Rule space bucket Given a global proliferation budget The local proliferation budget for the k-th spatial bucket is:

[0192] ;

[0193] For each bucket, within a single bucket, proliferation importance is ranked... Before selecting descending order A Gaussian multiplication process is performed; based on the normalization scale. With cloning threshold Distinguishing operations:

[0194] when When cloning is performed, a small perturbation is added to the center of the original Gaussian to generate a new Gaussian.

[0195] when Perform the split, then proceed along the largest principal axis of the covariance ellipsoid. Generate two sub-Gaussians, and then scale these two sub-Gaussians proportionally.

[0196] Simultaneously along the principal direction of the original covariance ellipsoid Place a new center ;

[0197] ;

[0198] Step 3.6: Perform StructureGuardian prior protection pruning. For Gaussian prior protection... Perform Gaussian protection marking: ;

[0199] For indicator functions, This is the threshold for anatomical protection. When the Gaussian is located in a high-confidence anatomical region, it is directly preserved during the routine pruning stage;

[0200] For unprotected Gaussians, a patient, soft pruning method is used, which is as follows:

[0201] A single degradation is achieved by multiplying the Gaussian density by the attenuation factor. And accumulate the number of penalties; when the number of penalties reaches... Then, scores were given based on pruning. Delete the top M redundant Gaussians from highest to lowest; periodically perform KNN outlier cleanup to remove floating artifact Gaussians.

[0202] The detailed method for step 4 of this embodiment is as follows:

[0203] Step 4.1, Importance-Aware Hybrid Ray Sampling:

[0204] For the The mixed sampling probability of ray r under each viewpoint:

[0205] ;

[0206] In the formula, For saliency sampling weights, For the first Total number of pixels per viewpoint For rays The saliency value of the corresponding pixel;

[0207] Step 4.2: Perform a significant weighted reconstruction loss. First, calculate the change of the dynamic weight of the single ray as iterative scheduling occurs:

[0208] ;

[0209] For iterative timing scheduling parameters; normalized weighted L1 reconstruction loss:

[0210] ;

[0211] In the formula, Render the predicted intensity for the model. For true projected intensity, For the set of sampled rays from the v-th viewpoint;

[0212] Step 4.3: The overall optimization objective is to introduce a volume total variation regularization term to suppress noise artifacts in addition to the reconstruction loss, thereby obtaining the complete loss function. :

[0213] ;

[0214] In the formula, V is the three-dimensional density mesh obtained by Gaussian voxelization. For the total variational regularization term, This is the weight coefficient for the regularization term.

[0215] Iterative optimization of all three-dimensional Gaussian elements yields the final three-dimensional CT attenuation field reconstruction result.

[0216] To verify the effectiveness of the method of this invention, this embodiment first performs a quantitative comparison with existing methods on the publicly available Human Organs (HO) sparse-view CT reconstruction dataset, including traditional reconstruction methods FDK, SART, and ASD-POCS, implicit neural representation-based methods NAF and SAX-NeRF, and methods based on three-dimensional Gaussian representation. And X-Field. In addition, this embodiment also selected typical methods for visualization result comparison to verify the advantages of the present invention in stripe artifact suppression, soft tissue boundary preservation, and bone structure detail restoration.

[0217] Figure 2 As shown, existing FDK and SART methods are prone to producing obvious needle-like stripe artifacts and structural breaks under sparse viewing conditions; although implicit representation methods such as SAX-NeRF can reduce artifacts to some extent, they still exhibit oversmoothing in complex soft tissue regions and bone structure boundaries. Gaussian methods such as X-Field have advantages in explicit representation and rendering efficiency, but due to the lack of anatomical prior constraints, they are prone to producing floating density fragments in the spatial domain or artifact regions, the so-called "Gaussian fog" phenomenon.

[0218] In contrast, the present invention can generate a cleaner background area and maintain clearer boundaries and more continuous structural morphology in complex anatomical areas such as the pelvis, abdominal soft tissue, and vascular structures. This indicates that anatomical prior-guided Gaussian lifecycle management can effectively reduce erroneous Gaussian assignments and improve the structural reliability of the reconstruction results.

[0219] Table 1. Quantitative comparison results of this invention with other methods on the Human Organs dataset.

[0220]

[0221] The aforementioned evaluation indicators include three-dimensional volumetric data indicators PSNR, SSIM, and LPIPS; two-dimensional projection indicators PSNR, SSIM, and LPIPS; and soft tissue window and bone window indicators, which are more in line with clinical observation needs. Through Figure 1 As shown in Table 1, the best results were achieved in overall 3D reconstruction quality, with a 3D PSNR of 33.367, a 3D SSIM of 0.901, and a 3D LPIPS reduced to 0.124, all of which are superior to FDK, SART, ASD-POCS, NAF, SAX-NeRF, and others. Compared with methods such as X-Field, this invention demonstrates significant advantages, particularly in clinically critical soft tissue and bone windows: the PSNR for soft tissue windows reaches 25.924, and for bone windows, it reaches 30.240, significantly higher than existing methods. This indicates that the invention not only improves global numerical accuracy but also better preserves diagnostically significant details such as low-contrast soft tissue and high-frequency bone structures. Since medical CT data often exhibit patient-specific, anatomical-site-specific, and scan-distribution-specific variations, validating the method's performance on a single dataset is insufficient to demonstrate its generalization ability. Therefore, this paper further constructs an additional cross-dataset evaluation set to verify the reconstruction robustness of this invention across different patients, anatomical regions, and low-contrast soft tissue structures. This evaluation set includes chest CT volume data from LIDC-IDRI and abdominal CT volume data from Pancreas-CT, employing scan geometry synthesis sparse view projection consistent with the HO dataset.

[0222] Table 2. Quantitative results of the present invention on the additional evaluation set.

[0223]

[0224] pass Figure 3 As can be seen from Table 2, in cross-dataset scenarios, existing technologies are still significantly affected by undersampling artifacts. Implicit representation methods suffer from oversmoothing of details, while some Gaussian methods are prone to producing floating artifacts or density discontinuities near structural boundaries.

[0225] The technical solution of this invention maintains a leading or near-optimal performance in most core indicators. For example, the 3D PSNR reaches 31.090, the 3D SSIM reaches 0.873, and the 3D LPIPS is reduced to 0.123. Under soft tissue window conditions, the PSNR reaches 23.609 and the SSIM reaches 0.850, indicating that the technical solution of this invention has good generalization ability for low-contrast tissue areas. In bone window indicators, the PSNR of the technical solution of this invention reaches 23.022, which is also superior. Gaussian contrast methods such as X-Field show that anatomical priors also have a positive effect on the restoration of high-frequency bone structures.

[0226] Figures 4 to 8 Together, they demonstrate the specific process of adaptive dual-window weight estimation in this invention, which is used to automatically extract diagnostic window responses related to soft and hard tissues from the input projection, providing a foundation for subsequent projection-level saliency map generation and three-dimensional anatomical prior construction.

[0227] like Figure 4 As shown, for each input X-ray projection, the original attenuation value is first mapped to a pseudo-HU intensity space, and a one-dimensional intensity histogram is calculated. To reduce local fluctuations caused by noise, the histogram can be smoothed. Subsequently, the main peaks are detected in the low-attenuation and high-attenuation regions, respectively, where the low-attenuation peaks correspond to the soft tissue-dominated response and the high-attenuation peaks correspond to the hard tissue or bone structure-dominated response. These two peaks provide data-adaptive constraints for the subsequent soft tissue and hard tissue window search ranges, so that the window width and window level no longer depend on fixed empirical parameters.

[0228] like Figure 5 As shown, within the search space defined by soft tissue peak values ​​and hard tissue peak values ​​respectively, the system performs window transformation on different candidate window width and window level combinations, and calculates the information entropy of the window transformation results in the foreground region. A higher information entropy indicates that the window parameters can preserve richer tissue intensity distributions and structural differences. Ultimately, the system selects the window width and window level combination with the highest entropy value as the optimal parameters for the soft tissue window and hard tissue window, respectively. This strategy can automatically adjust the diagnostic window for different patients, different projection views, and different intensity distributions, avoiding the failure of fixed window parameters in sparse views or uncalibrated data.

[0229] like Figure 6 and Figure 7 As shown, the soft tissue window primarily highlights low-contrast tissue areas and weak boundary information, helping to preserve fine structures such as muscle and organ boundaries; the hard tissue window primarily highlights high-attenuation areas and high-frequency structural information, helping to emphasize bone structures, calcified areas, and other strong-contrast structures. The anatomical information focused on by the two types of windows is complementary, and using one type of window alone may result in insufficient response of the other type of structure.

[0230] like Figure 8 As shown, the invention employs pixel-wise max pooling to fuse the soft tissue window response and the hard tissue window response, obtaining a comprehensive tissue-aware response. This fusion result can simultaneously preserve salient regions under either the soft or hard tissue window, thereby forming a structural prior with higher recall. Subsequently, this fused response is further combined with the Sobel edge response to generate a projection-level saliency map, which is then back-projected and aggregated into a three-dimensional volume space to form structure- and tissue-related channels in the anatomical prior volume.

[0231] Regarding the use of datasets, this embodiment uses the publicly available Human Organs (HO) dataset for standard evaluation of the invention.

[0232] The Human Organs dataset is a commonly used public benchmark dataset in the field of sparse-view CT reconstruction, containing multiple representative 3D human organ scans. To ensure a fair comparison with existing methods, eight typical anatomical scenes were selected for experiments, including Abdomen, Aneurism, Chest, Head, Jaw, Leg, Pancreas, and Pelvis. These datasets cover different anatomical regions such as the abdomen, chest, head, bone structures, and soft tissues, providing a comprehensive validation of the reconstruction capabilities of the method presented in this invention for complex medical structures.

[0233] To further verify the cross-dataset generalization ability of the method of the present invention, an additional sparse-view CT evaluation set, CrossCT, was constructed in the embodiments. CrossCT was synthesized from publicly available CT volume data, including three chest CT volume data sets, Chest_0002, Chest_0005, and Chest_0010, from the LIDC-IDRI dataset, and two abdominal CT volume data sets, Pancreas_0002 and Pancreas_0005, from the Pancreas-CT dataset. Using CT scan geometry consistent with the Human Organs dataset, sparse-view X-ray projections were synthesized from these three-dimensional volume data to test the robustness of the present invention on patients not seen, different data sources, and different anatomical regions.

[0234] The proposed anatomically prior-guided Gaussian splash sparse-view CT reconstruction method can also be implemented using the PyTorch framework and trained and tested on a single NVIDIA RTX 3090 GPU. During training, all models were optimized for 40k iterations, with the learning rate employing an exponential decay strategy, gradually decreasing from an initial learning rate of 1.6 × 10^-4 to 1.6 × 10^-6. For the hybrid ray sampler, the ray batch size for each iteration was set to 12288, and the importance sampling ratio η was set to 0.8, meaning that most of the sampling budget was allocated to the structurally relevant regions in the projection saliency map, while a certain proportion of uniform sampling was retained to maintain global projection consistency.

[0235] In the anatomical prior construction phase, pseudo-HU intensity mapping is first performed on the input sparse projection, and the candidate window ranges related to soft and hard tissues are determined by histogram peak detection. Subsequently, the parameters of the soft and hard tissue windows are automatically determined by maximizing Shannon entropy within the constrained search space, and the dual-window response is fused with the Sobel gradient response to generate a projection-level saliency map. Next, a small number of SART iterations are used to obtain a coarse intensity volume, and the saliency maps of each viewpoint are back-projected onto the three-dimensional volume space according to the same CT geometry, finally constructing a three-channel anatomical prior volume containing an intensity channel, a structural importance channel, and a tissue weight channel.

[0236] During the Gaussian initialization phase, a unified anatomical confidence map is calculated based on the structural importance channel and tissue weight channel in the anatomical prior, and initial Gaussian centers are selected from the high-confidence regions. The initial density of each Gaussian is determined by the value at the corresponding position of the coarse strength volume, and the initial scale is set according to the voxel spacing. To avoid fringe artifacts or isolated noise points being mistakenly used as initialization centers, we further perform K-nearest neighbor isolation point cleaning in the normalized coordinate system to remove candidate Gaussian centers that are too spatially isolated.

[0237] In terms of Gaussian lifecycle management, Gaussian proliferation is performed starting from the 500th iteration and continuing until approximately the 33kth iteration, with a proliferation interval of about 160 iterations. For each Gaussian primitive, the exponential moving average of indicators such as contribution, residual response, position gradient, visibility, and scale is maintained during training, with an EMA decay coefficient set to 0.9. Subsequently, Gaussians to be proliferated are selected based on the importance scores of multiple indicators, and a space bucket budget allocation strategy is used to avoid the concentration of new Gaussians in local artifact regions. Cloning operations are performed on compact and important Gaussians, while splitting operations are performed on Gaussians with larger scales. During the pruning phase, the Structure Guardian mechanism is used to protect Gaussians located in high anatomical confidence regions, preventing weak soft tissue boundaries or real fine structures from being mistakenly deleted in early optimization; a patient soft pruning strategy is used for unprotected Gaussians with low contributions, and KNN isolation cleaning is periodically performed to remove floating artifact points.

[0238] For the optimization objective, a saliency-weighted projection reconstruction loss is adopted, with a total variational regularization term for volume data added. The first 7500 iterations of training serve as a warm-up phase, with the saliency weight factor β(t) set to 0 to prioritize the model learning global projection consistency. After the warm-up, β(t) is set to 0.5, gradually focusing the optimization process on soft tissue boundaries, bone structures, and other anatomically significant regions. The total variational regularization term weight λ_tv is set to 0.05 to suppress local noise and density oscillations under sparse viewpoint supervision. Finally, the model outputs the reconstructed 3D CT attenuation volume, and volume data can be quickly extracted using native 3D Gaussian queries.

[0239] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for three-dimensional reconstruction using Gaussian splashing sparse-view X-rays based on anatomical prior guidance, characterized in that, Includes the following steps: Step 1: Construct a three-dimensional Gaussian decay field and X-ray differentiable projection rendering model; Step 2: Construct the anatomical prior volume. Based on the input sparse view X-ray projection, adaptively estimate the soft tissue window and hard tissue window. Weight the dual-window fusion response and the projection edge response and normalize them to generate a projection domain saliency weight map. Then, based on the CT scan geometric parameters, backproject the projection domain saliency weight map to the volume space to finally generate a three-channel anatomical prior volume containing intensity channel, structural importance channel, and tissue weight channel. The tissue weight channel is used to characterize the tissue relevance of voxel locations under the responses of soft tissue windows and hard tissue windows; The detailed method for obtaining the dual-window fusion response through adaptive dual-window weight estimation is as follows: First, the original single-view projection is... The attenuation value is mapped to the pseudo-HU intensity space, and a one-dimensional intensity histogram is calculated and smoothed. Low-attenuation soft tissue regions were defined based on global intensity quantiles. High-attenuation hard tissue region ], extract soft tissue peak values ​​respectively Hard tissue peak ; It is the 0th percentile. It is the 50th percentile. The 100th percentile; then a foreground region mask is generated. To eliminate atmospheric background interference: ； Represents the projected pixel coordinates. Indicates the background threshold. This represents the projection attenuation value at pixel coordinate u of the projected image, and the background threshold is dynamically determined based on the global intensity range of the projection. The formula is: ； and These are the global minimum and maximum attenuation values ​​for projection, respectively; Secondly, regarding the candidate window width and window position Perform a standard window transformation, using the following formula: ； Then the response after window transformation Discretized A grayscale range, and only in the foreground mask Internal Statistics The probability of each interval The information content of the candidate window parameters is quantified using Shannon entropy, and the quantification formula is as follows: ; This is a preset positive numerical stability constant; Next, in the soft tissue constraint search space Hard-structured constrained search space Solving for the optimal window parameters using the inner maximum entropy function: ； Substituting the optimal window parameters into the window transform operator yields the soft tissue window response. and hard tissue window response Finally, pixel-by-pixel max pooling dual-window fusion response: ； Step 3: Perform structure-aware Gaussian lifecycle management, using a three-channel anatomical prior to generate an anatomical confidence map, which guides the initialization, proliferation, and pruning of three-dimensional Gaussian units. Step 4: Importance-aware optimization. First, hybrid ray sampling is performed based on the saliency weight map of the projection domain, and a saliency-weighted reconstruction loss is constructed. Combined with the total variation regularization term, joint optimization is completed. Under the premise of ensuring global projection consistency, the reconstruction effect of soft tissue boundaries and key clinical areas of bone structure is enhanced, and finally complete three-dimensional CT volume data is obtained.

2. The method for three-dimensional reconstruction based on anatomical prior-guided Gaussian splashing sparse-view X-ray according to claim 1, characterized in that: The detailed method for step 1 is as follows: Step 1.1, Representation of the three-dimensional Gaussian decay field; Given a sparse viewpoint X-ray projection and its corresponding CT scan geometry, a set of three-dimensional Gaussian elements is used. To represent the CT attenuation field to be reconstructed; It is the number of Gaussian elements in the set; No. Gao Siyuan Including the central location Covariance matrix and Gaussian decay density , ; Decompose the covariance matrix as follows: ； in, Let be a rotation matrix. It is a three-dimensional principal axis scale diagonal matrix; and They are and The transpose of the matrix; Step 1.2, X-ray projection rendering model; that is, using a differentiable projection model suitable for X-ray absorption imaging, for a ray passing through volume space. The predicted X-ray intensity obtained from its rendering is represented in Beer-Lambert form, as shown below: ； in, Indicates the intensity of incident X-rays. Indicates ray Intersecting sets of Gaussian elements, Indicates the first A Gaussian nuclear ray The analytical line integral.

3. The method for three-dimensional reconstruction based on anatomical prior-guided Gaussian splash sparse-view X-ray according to claim 1, characterized in that: Step 2 also includes the following steps: The generation of the saliency weight map in the projection domain includes: For the original projection Solving for the magnitude of the Sobel gradient , and dual-window fusion response Weighted fusion and min-max normalization are used to generate a saliency weight map of the projection domain. : ； In the formula, These are the weighting coefficients for edge response and tissue window response. This represents the min-max normalization operation; A three-channel anatomical prior body was constructed, and a coarse strength reference body was generated by a small number of SART iterations for coarse reconstruction. ; Project saliency weights of all view domains The three-dimensional structural importance field is obtained by geometric backprojection from CT scans. Then, a three-dimensional three-channel anatomical prior body is constructed. Its value at the three-dimensional voxel coordinate x is expressed as: ； In the formula, This represents the three-channel prior vector of the three-channel anatomical prior at voxel coordinate x; Values ​​are assigned to the intensity channel; Values ​​are assigned to the structural importance channel; To determine the values ​​for the organization's weight channels; The values ​​of the intensity channel, the structural importance channel, and the tissue weight channel together constitute the three-dimensional volumetric three-channel anatomical prior for subsequent Gaussian initialization, proliferation, pruning, and optimization.

4. The method for three-dimensional reconstruction based on anatomical prior-guided Gaussian splash sparse-view X-ray according to claim 1, characterized in that, Step 3, the structure-aware Gaussian lifecycle management, specifically includes the following processes: Step 3.1: Anatomical perception Gaussian initialization, firstly fusing the structural importance of the three-channel anatomical prior. ) and organizational weight Generate anatomical confidence maps Then, high-confidence regions are selected as the candidate center set for Gaussian initialization. By using coarse strength channels Gaussian initial decay density is obtained by interpolation sampling. The initial scale is set based on the voxel neighborhood spacing; ； ； To initialize the confidence threshold; based on Probabilistic sampling or selecting the voxel with the highest confidence as the Gaussian center ; Step 3.2, Anisotropy-sensing initialization and cleanup: Map the candidate centers to a scale-normalized isotropic coordinate system and define the candidate points. K-nearest neighbor isolation: ； express of A set of nearest neighbor candidate points; if It was identified as an isolated noise point and removed. The isolation distance threshold; Step 3.3, EMA Lifecycle Indicator Tracking: Maintenance contribution for each Gaussian element Residual response Location gradient Visibility and normalization scale Historical indicators are smoothed using exponential moving averages; any indicator Updated formula: ； In the formula, Here, t is the EMA decay coefficient, and t is the current iteration step number. Step 3.4: Perform multi-index importance scoring to filter out noise interference from sparse projection and achieve adaptive refinement of multi-scale human anatomical structures. Step 3.5, Space Bucket Budget Constraint Proliferation: First, divide the 3D reconstructed space into... Rule space bucket Given a global proliferation budget The local proliferation budget for the k-th spatial bucket is: ； For each bucket, within a single bucket, proliferation importance is ranked... Before selecting descending order A Gaussian multiplication process is performed; based on the normalization scale. With cloning threshold Distinguishing operations: when When cloning is performed, the original Gaussian center is perturbed to generate a new Gaussian. when Perform the split, then proceed along the largest principal axis of the covariance ellipsoid. Generate two sub-Gaussians, and then scale these two sub-Gaussians proportionally. Simultaneously along the principal direction of the original covariance ellipsoid Place a new center ; ； Step 3.6: Perform structural protection prior pruning for Gaussian structures. Perform Gaussian protection marking: ; For indicator functions, This is the threshold for anatomical protection. At this point, the Gaussian anatomical region is located, and it is directly preserved during the routine pruning stage; For unprotected Gaussians, a patient, soft pruning method is used, which is as follows: A single degradation is achieved by multiplying the Gaussian density by the attenuation factor. And accumulate the number of penalties; when the number of penalties reaches... Then, scores were given based on pruning. Delete the top M redundant Gaussians from highest to lowest; periodically perform KNN outlier cleanup to remove floating artifact Gaussians.

5. The anatomically prior-guided Gaussian splash sparse-view X-ray three-dimensional reconstruction method according to claim 4, wherein the multi-index importance score includes Gaussian proliferation importance score and Gaussian pruning score; For the center is located Gao Siyuan The formula for calculating its Gaussian proliferation importance score is as follows: ； This represents the anatomical prior guidance intensity coefficient. ϵ is a preset positive numerical stability constant, and ϵ>0; This is the scale adaptation factor; For cloning operations Splitting operation ; The formula for calculating the Gaussian pruning score is as follows: ； A higher pruning score indicates higher Gaussian redundancy and makes the pruning process easier to remove.

6. The method for three-dimensional reconstruction based on anatomical prior-guided Gaussian splash sparse-view X-ray according to claim 1, characterized in that, The detailed method for step 4 is as follows: Step 4.1, Importance-Aware Hybrid Ray Sampling: For the Mixed sampling probability of ray r from each viewpoint The expression is as follows: ； In the formula, For saliency sampling weights, For the first Total number of pixels per viewpoint For rays The saliency value of the corresponding pixel; Step 4.2: Perform a significant weighted reconstruction loss. First, calculate the change of the dynamic weight of the single ray as iterative scheduling occurs: ； These are the parameters for iterative timing scheduling; Normalized weighted L1 reconstruction loss: ； In the formula, Render the predicted intensity for the model. For true projected intensity, For the set of sampled rays from the v-th viewpoint; ϵ is a preset positive numerical stability constant, and ϵ>0; Step 4.3: The overall optimization objective is to introduce a volume total variation regularization term to suppress noise artifacts in addition to the reconstruction loss, thereby obtaining the complete loss function. : ； In the formula, V is the three-dimensional density mesh obtained by Gaussian voxelization. For the total variational regularization term, The regularization term weight coefficient; Iterative optimization of all three-dimensional Gaussian elements yields the final three-dimensional CT attenuation field reconstruction result.

Images