A method for magnetic resonance image reconstruction based on three-dimensional Gaussian body rendering

Abstract

This invention discloses a magnetic resonance imaging (MRI) image reconstruction method based on three-dimensional Gaussian volume rendering. The method includes: first, acquiring point clouds of slice data; second, constructing a three-dimensional Gaussian primitive mesh and fitting the point cloud to obtain basic voxel intensities; then, introducing a neural residual field to learn and predict intensity residuals to correct the basic intensities; finally, fusing the basic intensities and residuals to generate high-fidelity final voxel intensities, and reconstructing the slice data into a three-dimensional volume based on this, mapping it to the target image space. By using this invention, high-resolution MRI volume reconstruction tasks can be completed quickly, meeting the time-sensitive requirements of clinical workflows. This invention can be widely applied in the field of image processing.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is an important non-invasive imaging technique in modern medicine. It offers advantages such as high soft tissue contrast and the absence of ionizing radiation, and is widely used in clinical diagnosis and treatment planning for areas such as the brain, abdomen, and spine. Ideally, MRI should acquire isotropic, high-resolution three-dimensional volumetric data to support multiplanar reconstruction in any direction and precise volumetric measurements.

[0003] However, due to the constraints of physical laws, it is difficult for MRI imaging to achieve a balance between spatial resolution, signal-to-noise ratio, and scan speed. Directly acquiring isotropic high-resolution volumes requires a long scan time, which seriously affects patient comfort and the efficiency of clinical workflows.

[0004] Therefore, a multi-stacked thick-slice acquisition strategy is commonly used in clinical practice: multiple sets of two-dimensional thick-slices are acquired from different directions (such as axial, coronal, and sagittal). This strategy can significantly shorten scanning time and maintain high in-plane resolution, but it introduces serious anisotropy problems—the slice thickness is usually 3 to 5 times the in-plane resolution. The resulting data suffers from partial volume effects, interslice gaps, and interslice misalignment caused by patient movement, which seriously affect subsequent quantitative analysis and clinical diagnosis. Reconstructing isotropic high-resolution three-dimensional volume (SVR) from anisotropic multi-stacked thick-slice data is currently a key area in medical imaging. Most existing reconstruction methods are implicit neural representation (INR) methods, such as NeSVoR. These methods parameterize the volume as a continuous function through coordinate-based neural networks, providing high-resolution reconstruction capabilities. However, INR relies on coordinate-based multilayer perceptrons (MLPs), and the computation of each voxel requires a large number of network forward passes, resulting in long training and inference times and low computational efficiency. Summary of the Invention

[0005] In view of this, in order to solve the technical problem that most existing magnetic resonance imaging methods require complete network forward propagation, resulting in low computational efficiency, this invention proposes a magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering. This method realizes the reconstruction and spatial transformation of three-dimensional volume data through the following steps: First, point clouds of slice data are acquired; second, a three-dimensional Gaussian primitive mesh is constructed, and the point cloud is fitted to obtain the basic voxel intensity; then, a neural residual field is introduced to learn and predict the intensity residual to correct the basic intensity; finally, the basic intensity and the residual are fused to generate a high-fidelity final voxel intensity, and the slice data is reconstructed into a three-dimensional volume based on this and mapped to the target image space.

[0006] In some embodiments, during the processing of the three-dimensional Gaussian element mesh, the volume space is divided into a uniform mesh, and each element is assigned to the corresponding mesh cell according to the Gaussian center position; for any query point, its mesh cell is first determined, and then the relevant Gaussian elements are retrieved only in the neighborhood cells within the preset search radius to achieve efficient local query.

[0007] Based on the above scheme, this invention provides a magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering. It employs explicit Gaussian unit representation to replace the point-by-point query mechanism of implicit neural networks, and achieves efficient local Gaussian retrieval through a block-based spatial partitioning strategy. This avoids the high computational complexity of traditional iterative optimization methods and the computational bottleneck of requiring a complete network forward propagation for each query point in implicit neural representation methods. This invention enables high-resolution MRI volume reconstruction to be completed in a shorter time, better meeting the time-sensitive requirements of clinical workflows. Attached Figure Description

[0008] Figure 1 This is a flowchart of the steps of a magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to the present invention; Figure 2 This is a schematic diagram of the data flow of the method of the present invention, including (a) the training process, (b) the point cloud construction process, and (c) the inference process; Figure 3 This is a visual comparison diagram of the reconstruction results of the present invention and the comparison method on three datasets (FeTA, FaBiAN, HCP); Figure 4 This is a visual comparison diagram of the ablation experiment results. Detailed Implementation

[0009] In addition to the problems mentioned in the background section, existing SVR reconstruction methods also include iterative optimization methods, such as NiftyMIC and SVRTK. These methods use an iterative optimization framework combined with explicit regularization terms for reconstruction. Although they have been applied clinically, these methods are computationally expensive, and the computation time increases dramatically with the increase of output resolution, resulting in poor scalability. For example, when processing high-resolution adult brain data, the NiftyMIC method can take tens of hours to reconstruct a single sample, making it difficult to meet the timeliness requirements of clinical practice.

[0010] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0012] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0013] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0014] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0015] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0016] Reference Figure 1 The diagram below illustrates an optional example of the magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering proposed in this invention. This method can be applied to computer devices, and the magnetic resonance image reconstruction method proposed in this embodiment may include, but is not limited to, the following steps: Step S1: Collect point cloud data based on slice data; Step S2: Process the point cloud data based on the three-dimensional Gaussian meta-mesh to calculate the basic strength; Step S3: Process the point cloud data based on the neural residual field to generate residual correction; Step S4: Calculate the final voxel strength based on the basic strength and the residual correction; in This represents the residual predicted by the neural residual field. This boundary ensures that the residual correction remains small relative to the Gaussian basis strength, preventing overfitting noise.

[0017] Step S5: Based on the final voxel intensity, assemble the slice data into a three-dimensional volume and convert it to the target image space.

[0018] In some feasible embodiments, step S1 specifically includes: Multiple sets of input thick slices are registered to a unified RAS anatomical coordinate space. Foreground pixels are sampled to form a point cloud, and each sampled point contains data normalized to the standard space. The three-dimensional coordinates and their corresponding signal strength values.

[0019] This embodiment uses thick slices acquired in three orthogonal directions (axial, coronal, and sagittal) as input data. Each set of slices has high in-plane resolution and a large slice thickness (the slice thickness is usually 3 to 5 times the in-plane resolution).

[0020] First, the slices from each group are rigidly registered and transformed to a unified RAS (Right-Anterior-Superior) anatomical coordinate space. During the registration process, the rigid transformation parameters for each slice are recorded. This is used for joint optimization in subsequent training phases.

[0021] The registered foreground pixels are sampled to form a point cloud dataset for initialization. Each sampled point contains the following information: 3D spatial coordinates: normalized to standard space. Signal intensity value: the MRI intensity of the corresponding pixel.

[0022] In some feasible embodiments, step S2 specifically includes: Based on the point cloud, a uniformly distributed 3D Gaussian primitive mesh is initialized at low resolution. Each Gaussian primitive is parameterized with three attributes: spatial center position, covariance matrix, and MRI signal intensity. The covariance matrix is ​​decomposed using a rotated quaternion and logarithmic scaling parameter to ensure numerical stability. The signal intensity is represented by a single scalar value instead of the view-dependent spherical harmonic function color representation in the original 3D Gaussian sputtering. The normalized volume space is divided into a uniform mesh, and each primitive is assigned to the corresponding mesh cell according to the Gaussian center position. For each sampling point from the k-th slice, a corresponding rigid transformation is applied to the reconstructed volume space, and the signal intensity at that point is calculated by aggregating the weighted contributions of each Gaussian primitive in the neighborhood.

[0023] Gaussian mesh initialization: based on sampled point clouds, at low resolution (e.g., A uniformly distributed 3D Gaussian element mesh is initialized. During initialization, the parameters of each Gaussian element are set as follows: spatial center... : Normalized coordinates of corresponding grid points; rotation quaternion Initialize to a unit quaternion Logarithmic scaling parameter : Set initial values ​​based on grid spacing; signal strength : Obtained from point cloud data through nearest neighbor interpolation.

[0024] Magnetic Gaussian representation: Each magnetic Gaussian element The parameters are defined as follows: in: The location of the spatial center in the normalized coordinate space; The covariance matrix encodes the shape and orientation of the primitives; The normalized MRI signal intensity.

[0025] Covariance matrix decomposition: To ensure numerical stability, the covariance matrix is ​​decomposed as follows: in, For rotation matrices, use unit quaternions. ( Parameterization; It is a diagonal scaling matrix. It is a learnable logarithmic scaling parameter.

[0026] During training, unnormalized quaternions are optimized and normalized before being converted to rotation matrices. The logarithmic scaling parameter is ensured to always have a positive scaling value through exponential mapping, while also providing numerically stable gradients.

[0027] Mesh generation: normalizing volumetric space Divide the grid into a uniform grid according to the given grid resolution. Each Gaussian element... According to its center Assign to the corresponding grid cells: Neighborhood search: For any query point First, determine the index of the grid cell in which it is located. Then, define a set of neighborhood cells within the search radius block_radius: The set of active Gaussian meta-elements used for rendering is defined as: In this embodiment, the default search radius is set to 5.

[0028] MRI volume rendering: for each sampling point from the k-th slice Applying the corresponding rigid transformation yields the coordinates in the reconstructed volume space: Signal strength is calculated by weighting the contributions of each Gaussian element within the aggregated neighborhood: rigid transformation parameters The Gaussian parameters are jointly optimized during training to achieve slice alignment and volume reconstruction.

[0029] In some feasible embodiments, step S3 specifically includes: Based on Gaussian representation, a lightweight multilayer perceptron is introduced as a neural residual field to capture high-frequency anatomical details that are difficult to express by Gaussian smooth basis functions. The neural residual field uses Fourier position encoding to map the input coordinates to a high-dimensional feature space, and the output is a residual correction amount limited to a small range by a scaling hyperbolic tangent function.

[0030] The Neural Residual Field (NRF) employs a lightweight multilayer perceptron architecture: Input layer: Fourier position encoding with 6 frequency bands maps 3D coordinates to a high-dimensional feature space; Hidden layers: 4 fully connected layers, each with 64 neurons, using the SiLU activation function; Output layer: Each neuron outputs a residual value, which is constrained by a scaled hyperbolic tangent function. Within the range.

[0031] The neural residual field employs a delayed activation mechanism. On the FeTA and FaBiAN datasets, the NRF is activated at the 2000th iteration; on the HCP dataset, the NRF is activated at the 4000th iteration. This strategy ensures that Gaussian elements first establish a robust coarse-scale representation, which is then refined by the NRF for high-frequency details.

[0032] In some feasible embodiments, related training steps are also included: Multi-resolution progressive training: A coarse-to-fine training strategy is adopted, and the resolution of the Gaussian grid is gradually increased at preset iteration milestones. During resolution conversion, the Gaussian parameters of the new grid are obtained by trilinear interpolation of the previous resolution grid to obtain the intensity and scale parameters, and the rotation quaternion parameters are obtained by normalized linear interpolation.

[0033] Resolution scheduling: Gaussian grid resolution is progressively increased at preset iteration milestones: FeTA and FaBiAN datasets: from Initially, it was gradually improved to [a higher level] at the 500th, 1000th, 2000th, and 3000th iterations. The total number of iterations was 4000; HCP dataset: using an extended scheduling scheme, the final resolution reached... The highest resolution was reached on the 6000th iteration, with a total of 8000 iterations.

[0034] Parametric interpolation: During resolution conversion, the Gaussian parameters of the new mesh are initialized as follows: Intensity and scale parameters: obtained from the previous resolution mesh using trilinear interpolation; Rotation quaternions: obtained using normalized linear interpolation.

[0035] This progressive densification strategy accelerates the convergence of the model to fine anatomical details by providing strong initialization from coarse-scale features.

[0036] Based on the above training process, the loss function is expressed as follows: The total training loss includes reconstruction loss and regularization loss: The reconstruction loss consists of smoothed L1 loss and SSIM loss: The smoothed L1 loss (Huber loss) is defined as follows: The regularization loss employs anisotropic regularization: In this embodiment, the loss weight is set as follows: , , .

[0037] During training, the model renders the sampled points in each round to perform calculations. Loss and Regularization is applied, and SSIM loss is calculated by periodically sampling slices. The loss only propagates the gradient to the 3D Gaussian field, while the SSIM loss propagates the gradient to both the 3D Gaussian field and the neural residual field.

[0038] The Adam optimizer is used, with different learning rates set for different parameters: position parameters. Learning rate 0.001; Strength parameter Learning rate 0.05; Scale parameter Learning rate 0.005; rotation parameters Learning rate: 0.001; Neural residual field parameters: learning rate: 0.0001.

[0039] The overall data flow of the above method is referenced Figure 2 Based on the overall process of the above method, this embodiment is validated on three representative MRI datasets: FeTA dataset: Fetal brain MRI volumetric data, with an original resolution of 0.5 mm and isotropic. Downsampling was used to simulate clinical thick-slice acquisition, resulting in a resolution of [missing information]. Motion artifacts were added. The target reconstruction resolution was 0.5 mm. FaBiAN dataset: a synthetic fetal brain dataset with a resolution of [missing information]. This dataset contains k-space noise and random inter-slice motion. It provides controlled ground truth for quantitative evaluation. The target reconstruction resolution is 1.1 mm. HCP dataset: Adult brain MRI data from the Human Connectome Project, original resolution 0.7 mm isotropic, downsampled to... The data was collected in a simulated clinical setting. The target reconstruction resolution was 0.7 mm.

[0040] For all datasets, each volume generates a stack of slices in three orthogonal directions (axial, coronal, and sagittal) to simulate clinical multi-stack acquisition protocols.

[0041] Experimental environment: All experiments were conducted on a computing system equipped with two AMD EPYC 7352 CPUs and one NVIDIA RTX A6000 GPU (48GB VRAM). The method of this invention was implemented using the PyTorch framework.

[0042] Compare with the following three representative methods: SVRTK: an iterative reconstruction toolkit using robust statistics; NiftyMIC: a complete processing flow integrating motion correction, bias field estimation, and intensity normalization; and NeSVoR: a reconstruction method based on implicit neural representations (implemented in pure PyTorch without using a custom CUDA kernel).

[0043] The following metrics were used to evaluate reconstruction quality: Peak Signal-to-Noise Ratio (PSNR, dB): a higher value indicates a smaller reconstruction error; Structural Similarity Index (SSIM): a higher value indicates better structure preservation; Normalized Cross-Correlation Coefficient (NCC): a higher value indicates a stronger correlation with the reference image; Normalized Root Mean Square Error (NRMSE): a lower value indicates higher reconstruction accuracy; Runtime (minutes): a lower value indicates higher computational efficiency.

[0044] Table 1. Quantitative comparison of the method of this invention with the baseline method on three datasets. As shown in Table 1, the method of the present invention achieves the best balance between reconstruction quality and computational efficiency.

[0045] On the FeTA dataset, this invention achieved a PSNR of 40.31 dB, which is 4.23 dB higher than the second-best method (NiftyMIC), while its runtime was only 5.63 minutes, representing an approximately 14-fold speedup compared to NiftyMIC's 79.12 minutes. This demonstrates the significant advantage of this invention when processing small-volume fetal brain data with limited anatomical background.

[0046] The efficiency advantage is even more pronounced on the high-resolution HCP adult brain dataset. This invention completes reconstruction in just 17.28 minutes, achieving approximately 78 times the speedup of NiftyMIC's 1353.48 minutes, while maintaining the highest NCC index and near-optimal PSNR index. This demonstrates the excellent scalability of this invention's method for high-resolution adult brain reconstruction.

[0047] On the FaBiAN dataset, this invention achieved the highest PSNR and NCC metrics in less than 5 minutes. Although NiftyMIC's SSIM is slightly higher, considering its nearly 14x longer runtime, this invention offers a better quality-efficiency trade-off.

[0048] Figure 3A comparison of the visualization reconstruction results on three datasets is presented. As can be seen from the figures, the images reconstructed by this invention have clearer tissue boundaries and more consistent anatomical structures. The baseline methods exhibit varying degrees of blurring (NiftyMIC, SVRTK) or noise (NeSVoR), particularly noticeable in complex anatomical regions or near tissue interfaces.

[0049] Ablation experiment: To verify the contribution of each technical component, a systematic ablation experiment was conducted, and the results are shown in Table 2.

[0050] Table 2 Ablation Experiment Results The effect of multi-resolution progressive training: Comparing the rows "No SSIM loss" and "No SSIM and no progressive training" in Table 2, it can be seen that progressive training brings PSNR improvements of 1.17dB, 2.58dB, and 1.20dB on the FeTA, FaBiAN, and HCP datasets, respectively. Directly optimizing millions of primitives at high resolution leads to gradient dynamic instability, while the coarse-to-fine strategy stabilizes the training process by gradually increasing the representation complexity.

[0051] The effect of SSIM loss: Compared with the full model and the configuration without SSIM loss, the introduction of SSIM loss brought significant improvements of 8.07dB on FeTA, 1.25dB on FaBiAN, and 0.54dB on HCP. SSIM loss enhanced the preservation of anatomical boundaries and tissue contrast, maintaining clinically relevant morphological features that are difficult to capture with pixel-level loss.

[0052] The role of neural residual fields (NRF): Removing the neural residual field (NRF) resulted in a decrease in PSNR of 1.20 dB (FeTA), 1.33 dB (FaBiAN), and 1.14 dB (HCP), respectively. Visualization results showed that the variant without NRF exhibited more noise near tissue boundaries, which is attributed to the inherent smoothness of the Gaussian function, making it difficult to model sharp intensity transitions. NRF addresses this limitation by providing a complementary representation suitable for high-frequency modes.

[0053] The effect of anisotropic regularization: Disabling anisotropic regularization resulted in a PSNR decrease of 0.82 dB (FeTA), 2.24 dB (FaBiAN), and 0.28 dB (HCP), respectively. The lack of this regularizer leads to striped and grainy noise artifacts, attributed to needle-like Gaussians (extreme aspect ratios) that appear during optimization. Regularization penalizes errors exceeding a threshold. The aspect ratio prevents the formation of such degenerate geometric shapes.

[0054] Reference Figure 4This demonstrates the impact of each component on the reconstruction quality.

[0055] Downstream segmentation task evaluation: To validate the clinical applicability of the reconstructed volumes, an automated brain segmentation task evaluation was performed on the FeTA dataset. The nnU-Net framework was used to train 3D segmentation models on ground-valued isotropic high-resolution volumes, which were then directly applied to the volumes reconstructed by each method (without fine-tuning). Seven anatomical structures were segmented: external cerebrospinal fluid, gray matter, white matter, ventricles, cerebellum, deep gray matter, and brainstem.

[0056] Table 3 Evaluation results of downstream segmentation task (FeTA dataset) As shown in Table 3, this invention achieved the highest average Dice coefficient (0.914), demonstrating optimal performance on most structures, including external cerebrospinal fluid, gray matter, white matter, ventricles, and brainstem. For structures with large spatial extents (such as white matter and gray matter), this invention performs particularly well, thanks to the efficient modeling ability of Gaussian units for smooth volume distributions. For geometrically complex small structures such as the cerebellum and deep gray matter, this invention performs comparably to NeSVoR, indicating good consistency in overall segmentation accuracy.

[0057] In summary, the beneficial effects of this invention also include: Addressing the physical essence of MRI signals reflecting inherent tissue characteristics rather than surface reflectivity, this invention designs a dedicated magnetic Gaussian primitive representation. Its volumetric rendering method directly performs Gaussian mixture evaluation in three-dimensional space, conforming to the physical signal formation process of MRI. The inherent smoothness of the Gaussian function naturally simulates the point spread function characteristics of MRI acquisition, and the weighted combination of overlapping Gaussians provides a physically reasonable representation of some volumetric effects, thereby effectively improving the physical consistency and reliability of the reconstruction results. This invention simplifies the viewpoint-dependent spherical harmonic function representation in the original three-dimensional Gaussian sputtering to a single intensity value, significantly reducing the number of parameters and memory usage per Gaussian primitive. This allows for the deployment of more primitives to capture fine anatomical structures under the same hardware conditions, or the achievement of high-quality reconstruction in resource-constrained computing environments, reducing dependence on high-end computing equipment and facilitating the widespread application of the method. Through the synergistic design of Gaussian primitives and neural residual fields, this invention enables Gaussian primitives to focus on efficiently capturing smooth volumetric structures, while neural residual fields supplement high-frequency details such as tissue boundaries that are difficult for Gaussian functions to express. The delayed activation strategy and output boundary constraints ensure training stability, achieving a good balance between overall structural accuracy and local boundary clarity in the reconstruction results, which is beneficial for subsequent quantitative analysis and clinical diagnosis. This invention is applicable to reconstruction tasks at different resolutions. This invention employs a multi-resolution progressive training strategy, gradually increasing representation complexity from coarse to fine and using parameter interpolation initialization. This avoids gradient instability caused by direct optimization at high resolution, enabling the model to converge stably and providing reliable technical support for isotropic volume reconstruction with different resolution requirements.

[0058] A magnetic resonance image reconstruction system based on three-dimensional Gaussian volume rendering, comprising: The sampling module is used to execute step S1; The Gaussian rendering module is used to execute step S2; The residual correction module is used to execute step S3; The superposition calculation module is used to execute step S4; The conversion module is used to perform step S5.

[0059] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0060] A magnetic resonance image reconstruction device based on 3D Gaussian volume rendering: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering as described above.

[0061] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0062] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering as described above.

[0063] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0064] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering, characterized in that, Includes the following steps: Point cloud data was collected based on sliced ​​data collection; The point cloud data is processed based on a three-dimensional Gaussian meta-mesh to calculate the basic strength; The point cloud data is processed based on the neural residual field to generate residual correction; The final voxel strength is calculated based on the base strength and the residual correction. Based on the final voxel intensity, the slice data is assembled into a three-dimensional volume and transformed into the target image space.

2. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 1, characterized in that, The step of collecting point cloud data based on slice data specifically includes: Based on the slice data, a three-dimensional voxel mesh coordinate system is generated according to the target resolution. The coordinates of the three-dimensional voxel mesh are normalized to a preset dimensional space.

3. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 1, characterized in that, The step of processing the point cloud data based on a 3D Gaussian meta-mesh to calculate the basic strength specifically includes: Based on the point cloud data, a Gaussian mesh is set to obtain a three-dimensional Gaussian meta-mesh; A magnetic Gaussian representation is generated based on the three-dimensional Gaussian element mesh; The magnetic Gaussian representation is used to assign each magnetic Gaussian element to its corresponding grid. The fundamental strength is calculated by aggregating the weighted contributions of each Gaussian element within the neighborhood.

4. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 3, characterized in that, The parameters of the magnetic Gaussian element are defined as follows: in, Indicates the first The parameters of a magnetic Gaussian element, Indicates the location of the center of space in coordinate space. Represents the covariance matrix. Indicates MRI signal intensity. Represents the rotation matrix. This represents the diagonal scaling matrix.

5. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 3, characterized in that, The step of assigning each magnetic Gaussian element to the corresponding grid in conjunction with the magnetic Gaussian representation specifically includes: Divide the volumetric space into a uniform grid according to the set grid resolution; Each magnetic Gaussian element is assigned to a corresponding grid cell based on its spatial center position.

6. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 5, characterized in that, Also includes: For any query point, determine the index of the grid cell it belongs to, and define the set of neighboring cells in combination with the search radius.

7. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 3, characterized in that, The formula for calculating the foundation strength is as follows: in, Indicates the foundation strength. Indicates the transformed sampling points. Indicates the first One slice, Indicates the sampling point. express The neighborhood set.

8. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 1, characterized in that, The neural residual field comprises an input layer, a hidden layer, and an output layer, wherein: The input layer employs Fourier position encoding to map three-dimensional coordinates to a high-dimensional feature space; The hidden layer uses the SiLU activation function; The output layer is used to output residual values ​​and to limit the residual values ​​to a preset range using a scaled hyperbolic tangent function.

9. The magnetic resonance image reconstruction method based on three-dimensional Gaussian volume rendering according to claim 1, characterized in that, It also includes pre-training the three-dimensional Gaussian mesh and the neural residual field, with the loss function shown below: in, Indicates total training loss. Indicates the losses incurred during reconstruction. Indicates the regularization loss. Indicates smoothed L1 loss, Indicates SSIM loss, This indicates the corresponding weight.

10. A magnetic resonance image reconstruction system based on three-dimensional Gaussian volume rendering, characterized in that, include: The sampling module collects point cloud data based on sliced ​​data; The Gaussian rendering module processes the point cloud data based on a three-dimensional Gaussian primitive mesh to calculate the basic intensity; The residual correction module processes the point cloud data based on the neural residual field to generate residual corrections; The superposition calculation module is used to calculate the final voxel strength based on the base strength and the residual correction. The conversion module assembles the slice data into a three-dimensional volume based on the final voxel intensity and converts it to the target image space.

Images