A three-dimensional reconstruction method, system, device and storage medium for low-yield and small-angle neutron imaging
By combining preprocessing of neutron imaging projection data, adaptive artifact correction, and the DiffNeuVox algorithm with sample geometric information to form a three-dimensional reconstruction method, the problems of artifacts and background interference in low-yield, low-angle neutron imaging were solved, and high-quality three-dimensional reconstruction results were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF ENERGY HEFEI COMPREHENSIVE NAT SCI CENT (ANHUI ENERGY LAB)
- Filing Date
- 2026-04-24
- Publication Date
- 2026-06-05
AI Technical Summary
Under conditions of low yield and limited angle neutron imaging, existing 3D reconstruction methods are prone to artifacts, large background interference, and unclear structural details. Especially when the neutron source intensity is limited and the exposure time and sample motion accuracy are restricted, it is difficult to obtain projection data with high signal-to-noise ratio, resulting in poor reconstruction quality.
The projection data is preprocessed using uniform sampling and resampling techniques, combined with adaptive correction based on artifact evaluation metrics, and 3D reconstruction is performed using the DiffNeuVox algorithm. Background suppression is achieved by generating a mask based on sample geometry priors, and finally, image post-processing is performed to improve the reconstruction quality.
It significantly improves the structural clarity and noise resistance of the reconstruction results under low signal-to-noise ratio conditions, effectively suppresses background noise, and improves the usability and visibility of the reconstruction results. It is suitable for neutron imaging under reactor and laboratory neutron source conditions.
Smart Images
Figure CN122156491A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of neutron imaging and tomographic reconstruction technology, and in particular to a three-dimensional reconstruction method, system, device and storage medium for low-yield, low-angle neutron imaging. Background Technology
[0002] Neutron imaging technology, with its high sensitivity to light elements and strong penetrating power to high atomic number materials, has been widely used in advanced materials analysis, non-destructive testing of complex structures, and characterization of the internal structure of scientific samples. In practical applications, neutron imaging typically requires the sample to be rotated around a central axis on a rotating stage to acquire projected images from multiple angles, and then tomographic reconstruction algorithms are used to obtain the three-dimensional structural information of the sample.
[0003] However, limitations such as neutron source intensity, exposure time, sample motion precision, and experimental costs often make it difficult to obtain ideal projection data with high throughput, full angular range, and high signal-to-noise ratio in actual imaging processes. Especially under laboratory-built neutron source conditions (such as the deuterium-deuterium ICP accelerator neutron source), the total neutron yield is typically only 10. 9 Due to limitations in acquisition efficiency and data volume, the number of projection angles is often limited to around 30. Under these low-yield, limited-angle imaging conditions, the acquired projection images suffer from problems such as low signal-to-noise ratio, obvious fringe artifacts, prominent ring artifacts, and insufficient angle coverage.
[0004] To address the aforementioned issues, existing technologies often employ traditional reconstruction algorithms such as Filtered Back Projection (FBP), Simultaneous Iterative Reconstruction Technique (SIRT), and Conjugate Gradient Least Squares (CGLS). However, these algorithms are prone to background noise amplification, edge blurring, detail loss, and artifact stacking under low signal-to-noise ratio and limited-angle conditions, resulting in a significant decrease in the structural identification capability of the final 3D reconstructed object. In recent years, reconstruction methods based on Neural Radiation Field (NeRF) (such as SAX-NeRF) have been attempted for limited-angle reconstruction, but they are highly sensitive to noise in low-yield neutron imaging scenarios, have high computational complexity, and lack stability in the reconstruction results.
[0005] Furthermore, in neutron imaging, samples typically have a cylindrical structure, and their effective region is often concentrated within the sample. Existing reconstruction procedures generally do not fully utilize the sample's geometric prior information, resulting in the retention of a large amount of background noise and invalid regions in the reconstruction results, which adversely affects subsequent analysis and visualization. Therefore, this application proposes a three-dimensional reconstruction method, system, device, and storage medium for low-yield, low-angle neutron imaging. Summary of the Invention
[0006] The purpose of this invention is to address the problems in the prior art where existing three-dimensional reconstruction methods are prone to artifacts, have large background interference, and lack clear structural details under low-yield and low-angle neutron imaging conditions. This invention proposes a three-dimensional reconstruction method, system, device, and storage medium for low-yield and low-angle neutron imaging.
[0007] In a first aspect, this application provides a three-dimensional reconstruction method for low-yield, low-angle neutron imaging, comprising the following steps:
[0008] Acquire neutron imaging projection data of the sample under test at multiple rotation angles to form a projection sequence;
[0009] The projection sequence is preprocessed, including normalization, uniform angle sampling, spatial clipping, and resolution uniform resampling, to obtain standardized projection data.
[0010] Based on the standardized projection data, an artifact evaluation index is calculated, and the standardized projection data is adaptively suppressed according to the artifact evaluation index to obtain corrected projection data.
[0011] A voxel-level differentiable optimization algorithm is used to reconstruct the three-dimensional projected data after correction. During the optimization process, data consistency constraints and regularization constraints are used in combination, and the optimization parameters are dynamically adjusted to control the termination of the iteration to obtain the initial reconstructed volume data.
[0012] A three-dimensional region of interest mask is generated based on the geometric prior information of the sample to be tested. The mask is then used to suppress the background of the initial reconstructed volume data to obtain the target reconstructed volume data.
[0013] The target reconstructed body data is post-processed to output the three-dimensional reconstruction result.
[0014] Optionally, the low-yield, low-angle neutron imaging conditions include: a total neutron yield of 10 from the neutron source. 9 The scale is significant, and the number of projection angles collected is 30.
[0015] Optionally, the uniform angle sampling includes: when the number of original projected angles is greater than the number of preset target angles, selecting a subset from the projection sequence at equal angular intervals so that the number of angles in the subset meets the condition for low-angle reconstruction.
[0016] Optionally, the artifact evaluation metrics include stripe artifact score and ring artifact score;
[0017] The stripe artifact score is obtained by calculating the statistical difference between the mean values of the column direction and the mean values of the row direction of multiple sampling slices.
[0018] The ring artifact score is obtained by statistically analyzing the fluctuation of the circumferential mean curve from the pixel point to the center radius of the slice.
[0019] Based on the stripe artifact score and the ring artifact score, the destripe intensity and the ring artifact intensity are determined respectively, and adaptive artifact suppression is performed.
[0020] Optionally, the voxel-level differentiable optimization algorithm is the DiffNeuVox algorithm, which uses three-dimensional volume data as optimization variables to construct a total loss function; the total loss function includes:
[0021] A data consistency term is used to constrain the consistency between the forward projection of the reconstructed volume data and the corrected projection data.
[0022] Total variation regularization term is used to suppress noise in volume data and maintain structural edge continuity;
[0023] The voxel-level differentiable optimization algorithm also includes:
[0024] Batch edge estimation is used to simultaneously constrain edge gradients of multiple slices in volume data.
[0025] Learning rate scheduling adaptively adjusts the optimization step size based on changes in the loss function;
[0026] Early stopping control terminates the optimization process when the loss function decreases below a threshold after multiple iterations.
[0027] Optionally, the geometric prior information includes the center position, radius, and axial range of the cylindrical sample; the generation of the three-dimensional region of interest mask includes:
[0028] The center and radius of a cylindrical sample are detected using an averaged projection image.
[0029] Construct a three-dimensional cylindrical mask based on the axial range of the sample;
[0030] Set the voxel values outside the mask to preset background values for background suppression.
[0031] Optionally, the image post-processing includes detail enhancement, microscale sharpening, and contrast stretching; the output 3D reconstruction results include exporting to NPZ format, TIFF sequence, or reconstruction analysis map.
[0032] Secondly, this application provides a three-dimensional reconstruction system for low-yield, low-angle neutron imaging, used to implement the three-dimensional reconstruction method for low-yield, low-angle neutron imaging as described in the first aspect, comprising:
[0033] The data reading module is used to acquire neutron imaging projection sequences;
[0034] The preprocessing module is used to perform normalization, uniform sampling, cropping, and resampling;
[0035] The artifact evaluation and suppression module is used to calculate artifact scores and adaptively perform stripe removal and ring removal artifacts;
[0036] The voxel optimization and reconstruction module is used for 3D reconstruction using a voxel-level differentiable optimization algorithm. This module includes a data consistency constraint unit, a regularization constraint unit, a learning rate scheduling unit, and an early stopping control unit.
[0037] The ROI constraint and post-processing module is used to generate ROI masks based on sample geometry priors and perform background suppression, as well as to perform detail enhancement and contrast optimization.
[0038] The results export module is used to output the reconstructed volume data and visualization results.
[0039] Thirdly, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.
[0040] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.
[0041] Compared with the prior art, this application includes at least one of the following beneficial technical effects:
[0042] By employing uniform sampling, cropping, and resampling, it is possible to achieve a total neutron yield of 10 9 It can effectively reconstruct data under low signal-to-noise ratio conditions with only 30 projection angles, reducing data acquisition requirements.
[0043] By introducing fringe artifact scoring and ring artifact scoring, the intensity of defringe and dering artifacts is adaptively adjusted according to the scores to achieve accurate correction and avoid overcorrection or undercorrection.
[0044] The method employs the DiffNeuVox algorithm in conjunction with data consistency constraints and total variation regularization, and introduces batch edge estimation, learning rate scheduling and early stopping control to balance fidelity and noise resistance. It outperforms existing methods such as SAX-NeRF, SIRT, CGLS and FDK in terms of SSIM and PSNR metrics.
[0045] The system automatically detects the geometric parameters of cylindrical samples and generates a 3D ROI mask to suppress background interference and focus the reconstruction results on the target area.
[0046] It performs detail enhancement, micro-scale sharpening, and contrast stretching to improve visibility; it supports multiple export formats such as NPZ, TIFF sequences, and analysis graphs.
[0047] In reactor neutron sources and laboratory deuterium-deuterium ICP accelerator neutron sources (10 9 Good reconstruction results were achieved under both conditions (scale) and magnitude, proving its versatility and practicality.
[0048] In summary, this invention significantly improves the structural clarity, noise resistance, and usability of 3D reconstruction results from low-yield, low-angle neutron imaging by organically combining preprocessing, adaptive artifact suppression, DiffNeuVox voxel optimization reconstruction, ROI constraints, and post-processing. Attached Figure Description
[0049] Figure 1 A block diagram of the DiffNeuVox neutron imaging three-dimensional reconstruction system provided in an embodiment of the present invention;
[0050] Figure 2 A flowchart of the DiffNeuVox 3D reconstruction method provided in this embodiment of the invention;
[0051] Figure 3 This is a flowchart of the DiffNeuVox optimized reconstruction sub-process provided in an embodiment of the present invention;
[0052] Figure 4 A comparison chart of reconstruction results of DiffNeuVox, SAX-NeRF, SIRT, CGLS and FDK algorithms provided in the embodiments of the present invention;
[0053] Figure 5 A comparison chart of SSIM and PSNR quantization for different algorithms provided in embodiments of the present invention;
[0054] Figure 6 A comparison of the physical image of the metal clock experimental phantom under reactor-type neutron source conditions and the 3D reconstruction results from DiffNeuVox;
[0055] Figure 7 This is the neutron source for the deuterium-deuterium ICP accelerator in our laboratory (10 9 A comparison of the actual image of the metal shell with polyethylene block under the condition of (scale) and the 3D reconstruction result of DiffNeuVox. Detailed Implementation
[0056] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.
[0057] Example 1: This example provides a 3D reconstruction method for low-yield, low-angle neutron imaging, specifically for a total neutron yield of 10... 9 Low signal-to-noise ratio, few-angle imaging scenarios with a scale of 30 and a projection angle of approximately 30. For example... Figure 2 As shown, the method includes the following steps S101 to S108.
[0058] I. Projection Data Acquisition and Organization (S101)
[0059] First, neutron imaging projection data of the sample under test at multiple rotation angles are acquired. In this embodiment, the sample under test is fixed on a rotating stage and rotated at equally spaced angles within a range of 0° to 360°, acquiring a total of 30 projection images, each stored in TIFF format. The TIFF projection sequence is sorted according to the filename to ensure consistency of the angle order.
[0060] II. Preprocessing: Normalization, uniform sampling, intelligent cropping and resampling (S102, S103)
[0061] After reading the TIFF projection sequence, normalization is performed to unify the projection data range to a preset interval (e.g., [0,1]) to eliminate energy differences between different projection images.
[0062] Since the number of original projection angles obtained may be greater than the number of target angles (the number of target angles is 30 in this embodiment), uniform sampling is performed: 30 projection images are selected from the complete projection sequence at equal angular intervals to form the projection subset required for small-angle reconstruction.
[0063] Furthermore, intelligent cropping is performed on each projected image: the effective sample area in the projected image is automatically detected, invalid dark areas around the edges are removed, and the smallest rectangular area containing the main sample is retained. Then, resolution resampling is performed to unify all projected images and subsequent reconstructed volume data to a preset voxel resolution. In this embodiment, the resampled input volume data is unified to a voxel resolution of 512×512×512.
[0064] III. Adaptive artifact assessment and suppression (S104, S105)
[0065] To address the common streak and ring artifacts in neutron imaging, this embodiment designs an adaptive evaluation and correction process.
[0066] Select multiple sampling slices (e.g., 10 slices at equal intervals), and calculate the fringe artifact score for each slice: calculate the ratio of the standard deviation of the column direction mean to the standard deviation of the row direction mean; the larger the ratio, the more severe the fringe artifact. Average the scores of all sampling slices to obtain the final fringe artifact score.
[0067] The ring artifact score is calculated as follows: based on the radius distribution from the pixel to the center of the sliced image, the mean circumferential curves at different radii are statistically analyzed, and the ring artifact score is calculated based on the degree of curve undulation (such as variance or peak-to-valley difference). The higher the score, the more obvious the ring artifact.
[0068] Based on the fringe artifact score and ring artifact score, the defringe intensity (e.g., filter kernel size or threshold) and the ring artifact intensity (e.g., circumferential filter parameters) are adaptively determined. Then, the standardized input data is sequentially processed for defringe and ring artifact removal to obtain corrected projection data. This enables targeted artifact correction for different samples and under different acquisition conditions.
[0069] IV. DiffNeuVox Voxel-Optimized Reconstruction (S106)
[0070] This embodiment uses the DiffNeuVox (Differential Neutron Voxel Optimization) algorithm to perform 3D reconstruction on the corrected projection data. The DiffNeuVox optimization reconstruction module employs a voxel-level difference optimization framework, using the input volume data as the optimization variable. For example... Figure 3 As shown, the algorithm includes the following sub-processes:
[0071] (1) Constructing data consistency items
[0072] To maintain consistency between the optimized 3D volume and the input volume data in terms of overall distribution, the mean square error between the forward projection (simulated projection) of the current reconstructed volume data and the measured corrected projection data is calculated to ensure that the reconstruction results match the actual collected data.
[0073] (2) Construct the total variation regularization term
[0074] The gradient magnitude of the 3D volume data is constrained by the L1 norm to suppress random noise while maintaining the structural continuity of the volume data.
[0075] (3) Batch edge estimation
[0076] During the optimization process, edge gradient estimation is performed simultaneously on multiple consecutive slices in the 3D volume data, and it is used as an auxiliary constraint to enhance the structural consistency of the reconstruction results between layers.
[0077] (4) Learning rate scheduling strategy
[0078] The Adam optimizer is used to iteratively update the 3D volume data. The initial learning rate is set to 0.001. The learning rate scheduling adopts a descent strategy based on plateau detection: when the loss function decreases less than a preset threshold after several consecutive iterations, the learning rate is multiplied by a decay factor (e.g., 0.5).
[0079] (5) Early stop strategy
[0080] The maximum number of iterations is set, and the loss function is monitored. The optimization automatically terminates when the loss function stops decreasing after several consecutive iterations, improving optimization stability and convergence efficiency.
[0081] In one specific implementation, iterative updates to the 3D volume data can be performed based on a differentiable optimization framework. The optimizer uses the Adam optimizer, and the learning rate scheduling employs a descent strategy based on plateau detection; iteration automatically terminates when the loss value no longer decreases after several consecutive iterations. To reduce computation time, parallel computing can be performed using a graphics processing unit (GPU), thereby improving the speed of voxel optimization. Through the above process, initial reconstructed volume data is obtained.
[0082] V. 3D ROI Mask Generation and Background Suppression (S107)
[0083] In this embodiment, the sample to be tested is a cylindrical structure. The sample's Region of Interest (ROI) is automatically detected and a three-dimensional cylindrical mask is generated: using the average projection of all projected images, the center position and radius parameters of the cylindrical sample are automatically detected (the center coordinates and radius are obtained through image binarization and a circle fitting algorithm); a three-dimensional cylindrical ROI mask is generated based on the sample's axial range (the minimum and maximum slice indices present in the sample from top to bottom). All voxel values outside the ROI mask are suppressed to a preset background value (e.g., 0), thereby reducing the impact of stripe noise and invalid background outside the sample on the reconstruction results.
[0084] VI. Post-processing and result export (S108)
[0085] Post-processing is performed on the volume data constrained by the ROI before the reconstruction results are output:
[0086] Detail enhancement: A conservative unsharp masking technique is used to enhance weak structural details, restoring them while suppressing artifacts.
[0087] Microscale sharpening: Applying a small-scale Laplacian sharpening filter improves the visibility of fine structures.
[0088] Contrast stretching: The grayscale histogram of the reconstructed slice is truncated and stretched, and 2% to 98% of the grayscale range is linearly mapped to [0,255] or a similar range to improve the readability of the reconstructed slice.
[0089] The results can be exported in multiple formats: the complete reconstructed data can be saved as an NPZ format (NumPy compressed file); each slice can be saved as a TIFF sequence (compatible with image analysis software such as ImageJ); and the front view, side view, top view and key slice analysis diagram of the reconstruction results can be generated at the same time.
[0090] VII. Experimental Results and Comparative Verification
[0091] To verify the effectiveness of the invention, experiments were conducted under two different neutron source conditions.
[0092] Experiment 1: Reactor-type Neutron Source
[0093] The neutron imaging data used in this experiment was provided by Burkhard Schillinger of the Technical University of Munich, Germany. The data originated from a reactor-type neutron source, and the experimental phantom used was a metal clock structure. The actual image and the DiffNeuVox 3D reconstruction results are shown below. Figure 6 As shown. During the data acquisition process, the phantom was fixed on a rotating platform and rotated at equal intervals within a range of 0° to 360°, acquiring a total of 30 projection data images. The 3D result reconstructed using the method of this embodiment is shown below. Figure 6 As shown, the main outline and internal structure (gears, pointers, etc.) of the sample can be clearly restored.
[0094] Experiment 2: Laboratory-scale deuterium-deuterium ICP accelerator neutron source
[0095] To further verify the applicability of the method of this invention under the conditions of a laboratory-built low-yield neutron source, a 10-ton neutron source was used in this laboratory. 9 A high-precision accelerator (deuterium-deuterium ICP neutron source) imaged a sample with a metal shell and a polyethylene block, acquiring 30 projection images at equal intervals within a range of 0° to 360°. The resulting images and corresponding DiffNeuVox 3D reconstructions are shown below. Figure 7 As shown. Reconstruction results ( Figure 7 The results show that the ability to distinguish between the metal outer shell boundary and the internal polyethylene material is significantly better than that of traditional methods.
[0096] Comparative experiment: The DiffNeuVox method of this invention is compared with SAX-NeRF, SIRT, CGLS and FDK algorithms on the same dataset. Figure 4The reconstruction results of DiffNeuVox are compared with those of SAX-NeRF, SIRT, CGLS, and FDK algorithms. As can be seen from the front, side, and top views, DiffNeuVox outperforms the compared algorithms in terms of structural boundary sharpness, preservation of internal details, and artifact suppression. Figure 5 The results show a comparison of SSIM and PSNR quantization results for different algorithms. The results indicate that DiffNeuVox outperforms DiffNeuVox in both structural similarity and peak signal-to-noise ratio, validating the effectiveness of the proposed method under low-yield, low-angle reconstruction conditions.
[0097] In summary, this invention organically combines low-yield, low-angle neutron imaging data preprocessing, adaptive artifact scoring and correction, DiffNeuVox voxel-optimized reconstruction, ROI prior constraints, and post-processing derivation to form a complete 3D reconstruction workflow suitable for low-yield, low-angle neutron imaging scenarios. This technical solution can improve upon existing methods that suffer from significant artifacts, large background interference, and unclear structural boundaries under low signal-to-noise ratio conditions, and has good engineering application value.
[0098] Example 2: A 3D Reconstruction System for Low-Yield, Few-Angle Neutron Imaging
[0099] This embodiment provides a three-dimensional reconstruction system for low-yield, low-angle neutron imaging, such as... Figure 1 As shown, it includes:
[0100] Data reading module: Used to read neutron projection sequences in TIFF format and organize them in angular order.
[0101] Preprocessing module: Used to perform normalization, uniform sampling, intelligent cropping, and resolution resampling on the projected sequence.
[0102] The artifact evaluation and suppression module is used to calculate the stripe artifact score and ring artifact score on multiple slices, and adaptively determine the destripe artifact intensity and dering artifact intensity based on the score results, and perform artifact correction.
[0103] The DiffNeuVox optimization and reconstruction module is used to complete voxel-level differentiable optimization reconstruction. It contains a data consistency constraint unit, a total variation regularization unit, a batch edge estimation unit, a learning rate scheduling unit, and an early stopping control unit.
[0104] ROI Constraint and Post-processing Module: Used to automatically detect prior geometric information of the sample (center, radius, and axial range of the cylindrical sample) and generate a 3D ROI mask, perform background suppression, and enhance details, microscale sharpening, and contrast optimization.
[0105] Results Export Module: Used to output reconstructed body data (NPZ, TIFF sequences) and visualization analysis charts.
[0106] The modules are connected via data flow interfaces and can be deployed on GPU-accelerated computing devices to achieve end-to-end automated reconstruction.
[0107] Example 3: Electronic Equipment
[0108] This embodiment provides an electronic device, including a processor (e.g., an Intel Xeon CPU or an NVIDIA GPU), a memory (e.g., DDR4 RAM or an SSD), and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the three-dimensional reconstruction method as described in Embodiment 1.
[0109] Example 4: Computer-readable storage medium
[0110] This embodiment provides a computer-readable storage medium, such as an HDD, SSD, flash memory, or DVD-ROM, on which a computer program is stored. When the program is executed by a processor, it implements the three-dimensional reconstruction method as described in Embodiment 1.
[0111] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.
Claims
1. A three-dimensional reconstruction method for low-yield, low-angle neutron imaging, characterized in that, Includes the following steps: Acquire neutron imaging projection data of the sample under test at multiple rotation angles to form a projection sequence; The projection sequence is preprocessed, including normalization, uniform angle sampling, spatial clipping, and resolution uniform resampling, to obtain standardized projection data. Based on the standardized projection data, an artifact evaluation index is calculated, and the standardized projection data is adaptively suppressed according to the artifact evaluation index to obtain corrected projection data. A voxel-level differentiable optimization algorithm is used to reconstruct the three-dimensional projected data after correction. During the optimization process, data consistency constraints and regularization constraints are used in combination, and the optimization parameters are dynamically adjusted to control the termination of the iteration to obtain the initial reconstructed volume data. A three-dimensional region of interest mask is generated based on the geometric prior information of the sample to be tested. The mask is then used to suppress the background of the initial reconstructed volume data to obtain the target reconstructed volume data. The target reconstructed body data is post-processed to output the three-dimensional reconstruction result.
2. The three-dimensional reconstruction method for low-yield, low-angle neutron imaging according to claim 1, characterized in that, The low-yield, low-angle neutron imaging conditions include: a total neutron yield of 10 from the neutron source. 9 The scale is significant, and the number of projection angles collected is 30.
3. The three-dimensional reconstruction method for low-yield, low-angle neutron imaging according to claim 1, characterized in that, The uniform angle sampling includes: when the number of original projected angles is greater than the number of preset target angles, selecting a subset from the projection sequence at equal angular intervals so that the number of angles in the subset meets the condition of low-angle reconstruction.
4. The three-dimensional reconstruction method for low-yield, low-angle neutron imaging according to claim 1, characterized in that, The artifact evaluation metrics include stripe artifact score and ring artifact score; The stripe artifact score is obtained by calculating the statistical difference between the mean values of the column direction and the mean values of the row direction of multiple sampling slices. The ring artifact score is obtained by statistically analyzing the fluctuation of the circumferential mean curve from the pixel point to the center radius of the slice. Based on the stripe artifact score and the ring artifact score, the destripe intensity and the ring artifact intensity are determined respectively, and adaptive artifact suppression is performed.
5. A three-dimensional reconstruction method for low-yield, low-angle neutron imaging according to claim 1, characterized in that, The voxel-level differentiable optimization algorithm is the DiffNeuVox algorithm, which uses three-dimensional volume data as optimization variables to construct a total loss function; the total loss function includes: A data consistency term is used to constrain the consistency between the forward projection of the reconstructed volume data and the corrected projection data. Total variation regularization term is used to suppress noise in volume data and maintain structural edge continuity; The voxel-level differentiable optimization algorithm also includes: Batch edge estimation is used to simultaneously constrain edge gradients of multiple slices in volume data. Learning rate scheduling adaptively adjusts the optimization step size based on changes in the loss function; Early stopping control terminates the optimization process when the loss function decreases below a threshold after multiple iterations.
6. A three-dimensional reconstruction method for low-yield, low-angle neutron imaging according to claim 1, characterized in that, The geometric prior information includes the center position, radius, and axial range of the cylindrical sample; the generation of the three-dimensional region of interest mask includes: The center and radius of a cylindrical sample are detected using an averaged projection image. Construct a three-dimensional cylindrical mask based on the axial range of the sample; Set the voxel values outside the mask to preset background values for background suppression.
7. A three-dimensional reconstruction method for low-yield, low-angle neutron imaging according to claim 1, characterized in that, The image post-processing includes detail enhancement, microscale sharpening, and contrast stretching; the output 3D reconstruction results include exporting to NPZ format, TIFF sequence, or reconstruction analysis map.
8. A three-dimensional reconstruction system for low-yield, low-angle neutron imaging, used to implement the three-dimensional reconstruction method for low-yield, low-angle neutron imaging as described in any one of claims 1-7, characterized in that, include: The data reading module is used to acquire neutron imaging projection sequences; The preprocessing module is used to perform normalization, uniform sampling, cropping, and resampling; The artifact evaluation and suppression module is used to calculate artifact scores and adaptively perform stripe removal and ring removal artifacts; The voxel optimization and reconstruction module is used for 3D reconstruction using a voxel-level differentiable optimization algorithm. This module includes a data consistency constraint unit, a regularization constraint unit, a learning rate scheduling unit, and an early stopping control unit. The ROI constraint and post-processing module is used to generate ROI masks based on sample geometry priors and perform background suppression, as well as to perform detail enhancement and contrast optimization. The results export module is used to output the reconstructed volume data and visualization results.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method of any one of claims 1 to 7.