A method and device for 3D CT image reconstruction based on block-based flow iteration

By using a block-based streaming iterative method combined with the CUDA streaming mechanism, the problems of memory usage and efficiency in traditional iterative reconstruction algorithms are solved, achieving efficient and low-memory-usage 3D CT image reconstruction, and supporting high-resolution industrial CT inspection.

CN123118967APending Publication Date: 2026-11-17ZHONGBEI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611356356.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional iterative reconstruction algorithms in industrial CT inspection suffer from high memory consumption and low iteration efficiency, failing to meet the requirements of high resolution and high efficiency. Furthermore, existing block-based iterative algorithms do not fully utilize the parallel performance of the CUDA architecture, resulting in low coupling between memory scheduling and parallel computing, which cannot meet the needs of rapid batch inspection in industrial applications.

Method used

A block-based streaming iterative method is adopted. By dividing the projected image and the 3D voxel space into blocks, a bidirectional index mapping table is constructed to achieve streaming parallel scheduling, reduce peak memory usage, improve iteration efficiency, and perform parallel iterative reconstruction. The CUDA streaming mechanism is used for concurrent execution of computation and data transmission.

Benefits of technology

It significantly reduces memory usage, improves image reconstruction efficiency, meets the needs of high-resolution industrial CT inspection, ensures imaging accuracy and inspection speed, and supports high-resolution reconstruction tasks of consumer-grade GPUs and embedded GPUs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118967A_ABST
    Figure CN123118967A_ABST
Patent Text Reader

Abstract

This invention discloses a method and device for three-dimensional CT image reconstruction based on block-based streaming iteration. The method includes: acquiring projection images of the target to be reconstructed at multiple projection angles; performing block segmentation on the projection images to obtain multiple projection sub-blocks; segmenting the three-dimensional voxel space with the same size as the target to be reconstructed to obtain multiple voxel sub-blocks; constructing a bidirectional index mapping table based on the multiple projection sub-blocks and the multiple voxel sub-blocks, the bidirectional index mapping table including the voxel space coordinate range corresponding to each projection sub-block and the projection sub-block corresponding to each voxel sub-block; and performing streaming parallel scheduling based on the bidirectional index mapping table to iteratively reconstruct the multiple projection sub-blocks to obtain the reconstructed image of the target to be reconstructed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and computing device for three-dimensional CT image reconstruction based on block-based streaming iteration. Background Technology

[0002] Industrial CT 3D reconstruction technology is a core method for detecting internal defects in precision castings, electronic chips, and aerospace components, enabling non-contact acquisition of the workpiece's internal 3D structural information. Among them, algebraic iterative reconstruction algorithms, represented by Simultaneous Algebraic Reconstruction (SART), are widely used in high-precision industrial inspection scenarios due to their advantages of strong sparse projection adaptability, high defect reconstruction accuracy, and few artifacts.

[0003] As industrial inspection demands increasingly higher voxel resolution and inspection efficiency, the limitations of traditional iterative reconstruction algorithms are becoming more and more apparent. Firstly, they consume a lot of video memory. Traditional iterative algorithms use a global voxel loading mode, and the amount of 3D voxel data increases cubically with resolution. Reconstructing large-size phantoms can easily exceed the GPU's video memory capacity, resulting in memory overflow and making it impossible to run on low-end or mid-range graphics cards. Moreover, the iteration efficiency is low. Traditional CPU serial iteration requires hundreds of iterations to converge, and the reconstruction of a single workpiece can take several hours. Even with global parallel acceleration, the large overhead of global data reading and writing makes it difficult to meet the needs of rapid batch inspection in industry. Summary of the Invention

[0004] To address the existing technical problems, this invention provides a three-dimensional CT image reconstruction method and computing device based on block-based streaming iteration, which can improve image reconstruction efficiency.

[0005] Firstly, a block-based streaming iterative method for three-dimensional CT image reconstruction is provided, comprising: acquiring projection images of the target to be reconstructed at multiple projection angles; performing block segmentation on the projection images to obtain multiple projection sub-blocks; segmenting the three-dimensional voxel space with the same size as the target to be reconstructed to obtain multiple voxel sub-blocks; constructing a bidirectional index mapping table based on the multiple projection sub-blocks and the multiple voxel sub-blocks, the bidirectional index mapping table including the voxel space coordinate range corresponding to each projection sub-block and the projection sub-block corresponding to each voxel sub-block; and performing streaming parallel scheduling based on the bidirectional index mapping table to iteratively reconstruct the multiple projection sub-blocks to obtain the reconstructed image of the target to be reconstructed.

[0006] In a second aspect, a computing device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the block-based streaming iteration-based three-dimensional CT image reconstruction method provided in the first aspect of this application.

[0007] This application divides a single projected image into blocks to obtain projected sub-blocks, and divides the three-dimensional voxel space into blocks to obtain multiple voxel sub-blocks. A bidirectional mapping relationship is established between the projected sub-blocks and the voxel sub-blocks. The corresponding voxel sub-blocks can be located from the projected sub-blocks, and vice versa. In this way, during the iterative reconstruction process, the global voxel loading can be converted into a single-block time-sharing loading operation, without loading the entire projected image. This reduces the peak memory usage to the level of a sub-block. Based on the bidirectional index mapping table, streaming parallel scheduling can also be performed, thereby realizing parallel iterative reconstruction and improving image reconstruction efficiency. Attached Figure Description

[0008] Figure 1 This is an application environment diagram of a block-based streaming iteration-based three-dimensional CT image reconstruction method in one embodiment; Figure 2 This is a flowchart of a three-dimensional CT image reconstruction method based on block-based streaming iteration in one embodiment; Figure 3 This is a schematic diagram of the projection sub-block and voxel sub-block in one embodiment; Figure 4 This is a schematic diagram of adjacent voxel weighted fusion in one embodiment; Figure 5 This is a schematic diagram of a three-dimensional CT image reconstruction device based on block-based streaming iteration in one embodiment; Figure 6 This is a schematic diagram of a computing device in one embodiment. Detailed Implementation

[0009] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0010] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0011] In the following description, the expression “some embodiments” refers to a subset of all possible embodiments. However, it should be understood that “some embodiments” can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0012] See Figure 1 This diagram illustrates the application environment of a block-based streaming iterative method for 3D CT image reconstruction in one embodiment. The block-based streaming iterative method is applied in a computing device 10, which acquires projection images of the target object at various projection angles. Each projection angle corresponds to a projection image. In CT imaging, projection images are acquired by emitting X-ray beams from an X-ray source that penetrate the target object. The attenuated X-ray intensity is received by a flat-panel detector on the other side and converted into a digital signal. During this process, the object rotates around its rotation axis within the acquisition range at equal or non-uniform angles (or continuously). The system simultaneously acquires two-dimensional attenuation distribution maps at each projection angle. These two-dimensional (Digital Radiography, DR) projection images constitute the raw input data required for subsequent 3D reconstruction. The physical acquisition process strictly follows the Lambert-Beer attenuation law, recording the combined absorption information of the object density and atomic number along the X-ray path.

[0013] As industrial inspection demands increasingly higher voxel resolution and detection efficiency, the limitations of traditional iterative reconstruction algorithms are becoming more apparent: First, they place a heavy burden on GPU memory. Traditional iterative algorithms use a global voxel loading mode, and the amount of 3D voxel data increases cubically with resolution. Reconstructing large-size phantoms easily exceeds the GPU's memory capacity, leading to memory overflow and making it unusable on low- to mid-range (Compute Unified Device Architecture, CUDA) graphics cards. Second, their iterative efficiency is low. Traditional CPU serial iteration requires hundreds of iterations to converge, and single-workpiece reconstruction can take several hours. Even with CUDA global parallel acceleration, the large overhead of global data read and write makes it difficult to meet the needs of rapid industrial batch inspection. Third, there is an imbalance in iteration accuracy. Under the fixed regularization constraint strategy, the workpiece edges are blurred and small defects are lost at low iteration counts, while noise is amplified and overfitting is serious at high iteration counts, making it impossible to balance detail preservation and noise suppression throughout the entire iteration cycle. Fourth, the block reconstruction scheme has shortcomings. Existing block iterative algorithms mostly focus only on task splitting and do not combine the streaming scheduling and memory optimization characteristics of the CUDA architecture. Block stitching artifacts are obvious, and the coupling between memory scheduling and parallel computing is low, which cannot give full play to the parallel performance of the CUDA architecture.

[0014] Therefore, there is an urgent need for a block-based streaming iterative 3D reconstruction method that is deeply adapted to the CUDA architecture, has low memory usage, high iteration efficiency, and excellent imaging accuracy, in order to solve the problem of the difficulty in balancing lightweight and high accuracy in existing technologies.

[0015] Please see Figure 2 This is a flowchart of a block-based streaming iteration-based 3D CT image reconstruction method provided in an embodiment of this application. The block-based streaming iteration-based 3D CT image reconstruction method is applied in a computing device and includes the following steps: S11. Acquire projection images of the target to be reconstructed at multiple projection angles.

[0016] In this embodiment, the target to be reconstructed refers to the target that needs to be reconstructed. Projected images of the target to be reconstructed can be acquired at multiple projection angles. The projection angle is the spatial positioning parameter of the CT scanning system relative to the rotation sampling of the target to be reconstructed, and the projected image is the two-dimensional digital image of X-ray attenuation output by the detector at that projection angle. The projected image can be the projection data after preprocessing the acquired original projection image. The preprocessing includes, but is not limited to, image cropping, noise reduction, etc. For example, projected images of the target to be reconstructed at multiple projection angles are acquired, and all projected images are standardized to a resolution of 512×512 pixels.

[0017] S12. Perform block segmentation on the projected image to obtain multiple projected sub-blocks.

[0018] In this embodiment, for a single projected image, a uniform grid partitioning strategy is used to globally divide the image, resulting in multiple projected sub-blocks of the same size. Each projected sub-block corresponds to a unique projected sub-block number. For example, as... Figure 3 As shown, Figure 3 This is a schematic diagram of the projection sub-blocks and voxel sub-blocks in one embodiment. A 4×4 equal grid partitioning strategy is used to globally divide a single projection image, resulting in 16 projection sub-blocks with a size of 128×128 pixels.

[0019] S13. Divide the three-dimensional voxel space with the same size as the target to be reconstructed into blocks to obtain multiple voxel sub-blocks.

[0020] In this embodiment, the three-dimensional voxel space refers to a regular three-dimensional mesh data domain formed by discretizing and dividing the physical three-dimensional continuous region occupied by the target to be reconstructed in computer memory. The three-dimensional voxel space is composed of a series of cubic units (i.e., voxels) of equal size stacked in the X, Y, and Z dimensions. Each voxel unit is assigned a gray value that characterizes the average attenuation coefficient of the material at that local location to X-rays. The distribution set of all its gray values ​​is the final three-dimensional volume data to be solved by the iterative reconstruction algorithm. Each voxel sub-block can uniquely correspond to a voxel sub-block number.

[0021] On a plane perpendicular to the main X-ray projection direction, the three-dimensional voxel space is divided into a grid of the same dimension and scale, resulting in voxel sub-blocks that correspond one-to-one with the projection sub-blocks. Each voxel sub-block contains multiple voxels.

[0022] S14. Construct a bidirectional index mapping table based on multiple projection sub-blocks and multiple voxel sub-blocks.

[0023] In this embodiment, the bidirectional index mapping table includes the voxel space coordinate range corresponding to each projection sub-block and the projection sub-block corresponding to each voxel sub-block. By constructing the bidirectional index mapping table, the forward direction records the voxel space coordinate range and projection contribution coefficient index corresponding to each projection sub-block, while the reverse direction records the number of the projection sub-block and the grayscale correction weight index corresponding to each voxel sub-block. Independent initialization of voxel sub-blocks is achieved through block mapping, replacing the traditional one-time global voxel loading mode. The projection contribution coefficient index is used for orthographic projection operations, i.e., the process of mapping from voxel space to projection space. It records the storage location or offset of the non-zero element in the system matrix corresponding to each voxel within the current projection sub-block. The grayscale correction weight index is used for backprojection correction operations, i.e., the mapping process of reverse allocation from projection error to three-dimensional voxel space. It records the storage location of the normalized weighting factor required by the current voxel sub-block when receiving projection residual feedback. The partitioning rules for the three-dimensional voxel space can be the same as those for the projection sub-blocks. On the XY plane perpendicular to the X-ray main projection direction, the same partitioning rules as those for the projected image can be used for partitioning.

[0024] For example, such as Figure 3 As shown, the 4×4 block rule of the matching projection image is used to divide the 512×512×512 three-dimensional voxel space into a 4×4 grid on the XY plane perpendicular to the X-ray main projection direction, resulting in 16 voxel sub-blocks with a size of 128×128×512 voxels.

[0025] S15. Based on the bidirectional index mapping table, perform streaming parallel scheduling to iteratively reconstruct multiple projection sub-blocks to obtain the reconstructed image of the target to be reconstructed.

[0026] By splitting the 3D voxel space and the projected image into matching sub-blocks, only a few projection sub-blocks need to be loaded in a single iteration to complete the calculation, thereby reducing peak memory usage to the level of sub-blocks and supporting high-resolution reconstruction tasks on consumer-grade GPUs. Simultaneously, the constructed bidirectional index mapping table allows the location of the target voxel space coordinate range of a projection sub-block, and conversely, the corresponding projection sub-block number can be found. This bidirectional index mapping table ensures accurate location of the corresponding data address during iterative orthographic projection operations and backprojection corrections in block-parallel update scenarios, eliminating the need to load all projection data at once. This improves the addressing efficiency of streaming scheduling and the continuous read / write performance of memory, thereby enhancing the efficiency of image reconstruction.

[0027] In the above embodiments, by dividing a single projected image into blocks to obtain projected sub-blocks, and dividing the three-dimensional voxel space into blocks to obtain multiple voxel sub-blocks, a bidirectional mapping relationship is established between the projected sub-blocks and the voxel sub-blocks. The corresponding voxel sub-blocks can be located from the projected sub-blocks, and the corresponding projected sub-blocks can be located from the voxel sub-blocks. In this way, during the iterative reconstruction process, the global voxel loading can be converted into a single-block time-sharing loading operation, without loading the entire projected image, thereby reducing the peak memory usage to the level of a sub-block. Based on the bidirectional index mapping table, streaming parallel scheduling can also be performed, thereby realizing parallel iterative reconstruction and improving image reconstruction efficiency.

[0028] In some embodiments, constructing a bidirectional index mapping table based on multiple projection sub-blocks and multiple voxel sub-blocks includes: For projection sub-blocks The mapping formula is used to determine the projected sub-blocks. The target voxel space coordinate range, where the mapping formula is as follows: .

[0029] in For projection sub-blocks The horizontal index, For projection sub-blocks The vertical index, The x-coordinate of the target voxel within the target voxel space coordinate range. The vertical coordinate of the target voxel within the target voxel space coordinate range. The target voxel's Z-coordinate is located within the target voxel space coordinate range. The total dimension of voxels on the X-axis. The total dimensions of the voxels on the Y-axis. The total dimension of voxels on the Z-axis, where This represents the total number of voxel blocks on the X-axis. This represents the total number of voxel blocks on the Y-axis; The target voxel sub-blocks and projection sub-blocks corresponding to the target voxel's x-coordinate, y-coordinate, and Z-coordinate. Correspondingly.

[0030] In this embodiment, the X-axis represents the length direction of the three-dimensional voxel space, the Y-axis represents the width direction of the three-dimensional voxel space, and the Z-axis represents the Z-axis direction of the three-dimensional voxel space. , , These are integer coordinates in 3D voxel space. The target voxel space coordinate range represents the projected sub-block in 3D voxel space. The corresponding voxel space range. The x-coordinate of the target voxel represents the corresponding x-coordinate in the target voxel space coordinate range, and the y-coordinate of the target voxel represents the corresponding y-coordinate in the target voxel space coordinate range. The Z-coordinate of the target voxel represents the corresponding Z-coordinate in the target voxel space coordinate range.

[0031] For example, = = =512; There are 16 projection sub-blocks in total, which is 4×4 projection sub-blocks. , Therefore, there are also 16 voxel blocks, that is, 4×4 voxel blocks. , , () represents the integer coordinates of a voxel in three-dimensional voxel space. , , The coordinate range definition ensures that the 16 voxel sub-blocks completely divide the entire 512×512×512 voxel space and cover it without overlap (except for the overlapping buffer area). Each voxel belongs to one and only one voxel sub-block, providing a spatial basis for subsequent block-based independent iterative reconstruction.

[0032] In some embodiments, based on a bidirectional index mapping table, streaming parallel scheduling is performed to iteratively reconstruct multiple projected sub-blocks to obtain a reconstructed image of the target to be reconstructed, including: Configure the data transfer stream and the computation stream, where the data transfer stream is responsible for copying data between the host's memory and the graphics processor's video memory, and the computation stream is used to perform iterative reconstruction; Based on the order of each projection sub-block in the bidirectional index mapping table, the current projection data and the corresponding current voxel sub-block data corresponding to the current projection sub-block are obtained from the bidirectional index mapping table through the data transmission stream and loaded into the video memory; Through computational flow, in the current iteration, based on the current projection data, the voxels of each current voxel sub-block in the current voxel sub-block data are updated in parallel iteratively. After the current voxel sub-block completes the update of the current iteration, the resources occupied by the current voxel sub-block are released in the video memory.

[0033] In this embodiment, the current projection sub-block represents the projection sub-block currently being processed. The current projection data represents the projection image corresponding to the current projection sub-block. The current voxel sub-block data represents the identifier data of the voxel sub-block corresponding to the current projection sub-block, determined according to the bidirectional index mapping table. The voxel sub-block corresponding to the current projection sub-block is the current voxel sub-block. The current voxel sub-block data represents the data required to reconstruct the voxel values ​​of each current voxel in the current voxel sub-block, such as the current voxel value of each current voxel, the identifier of each current voxel, the identifier of each current voxel sub-block, etc. When processing a current projection sub-block, it is first loaded into the video memory via a data transmission stream, and then reconstructed via a computation stream, that is, the voxels of each current voxel sub-block in the current voxel sub-block data are iteratively updated.

[0034] In a current iteration, for a given current projection sub-block, the voxels of each corresponding current voxel sub-block can be updated in parallel. Once the voxels of a current voxel sub-block have been updated in the current iteration, the video memory resources occupied by that current voxel sub-block can be released in the current video memory.

[0035] Optionally, the method further includes: While the computation stream is performing iterative reconstruction of the current projection sub-block, the data transmission stream synchronization copies the projection data of the next projection sub-block and the corresponding voxel sub-block data from the host memory to the graphics processor's video memory in sequence.

[0036] The CUDA streaming mechanism is used to construct a streaming memory management logic of "single-block loading - parallel iteration - immediate release - cyclic scheduling": multiple CUDA streams are configured to achieve concurrent execution of computation and data transmission. According to the order of the projected sub-blocks in the bidirectional index mapping table, only one set of projected sub-blocks and their corresponding voxel sub-block data are loaded into the GPU memory at a time. After the CUDA parallel kernel function is called to complete the iterative correction operation of the current voxel sub-block, the memory resources corresponding to the current voxel sub-block are immediately released, and then the next set of projected sub-blocks is loaded, and the cycle is repeated. In a single iteration, the iterative update of all voxel sub-blocks corresponding to a set of projected sub-blocks is completed in sequence.

[0037] A streaming scheduling logic is constructed based on the CUDA streaming mechanism, configuring two CUDA streams: a data transfer stream and a computation stream. The computation stream is responsible for executing iterative kernel function operations; the data transfer stream is responsible for copying data between the host and GPU memory. Through streaming concurrency technology, when the computation stream is executing the iterative reconstruction operation of the Nth group of projected sub-blocks, the data transfer stream simultaneously copies the projection data of the N+1th group of projected sub-blocks and the corresponding voxel sub-block data from the host memory to the GPU memory. Both are executed in parallel in different CUDA streams without blocking each other, thus completely hiding the data transfer latency within the iterative computation time, allowing the GPU computing unit to continuously execute iterative operations without waiting for data loading. Since the Nth group of projected sub-blocks can release the resources occupied in GPU memory after completing the update of a current voxel sub-block, the N+1th group of projected sub-blocks can be gradually loaded. After the voxel updates of all current voxel sub-blocks corresponding to the Nth group of projected sub-blocks are completed, the GPU memory space occupied by the Nth group of projected sub-blocks can be completely released. During the release process, the data of the N+1th group of projected sub-blocks is simultaneously loaded through the data transfer stream, forming concurrency with the computation stream. The above process is repeated until all projected sub-blocks have completed the current iteration, at which point the updated voxel data for each voxel is sent back to the host memory. This block scheduling process is repeated round by round until all iterations are completed.

[0038] First, by implementing a pipelined operation of single-block loading—parallel iteration—instant release—loop scheduling through the CUDA streaming mechanism, the peak memory usage required for a single iteration is compressed from the global voxel level to the single sub-block level, effectively overcoming the bottleneck of limited GPU memory capacity in high-resolution CT reconstruction. This allows consumer-grade and even embedded GPUs to handle iterative reconstruction tasks with large-size volumetric data. Second, multiple CUDA streams support highly concurrent execution of data transmission and iterative reconstruction computation, enabling the prefetching operations of projection sub-blocks and voxel sub-blocks from the host to the device to overlap with the SART iterative correction of the current sub-block. This effectively hides the PCIe bus transmission latency within the computation time, significantly improving the overall iterative throughput efficiency. Furthermore, this... The scheduling strategy, combined with a block mapping table, enables precise reuse and rapid reclamation of GPU memory, avoiding system overhead and fragmentation issues caused by frequent memory allocation and release, and ensuring the stability of long-term iterative operation. Simultaneously, due to the natural decoupling of iterative computation between sub-blocks, CUDA kernel functions can utilize high utilization rates to call GPU massively parallel threads to perform independent projection and back-projection operations, fully leveraging the GPU's dense computing power while reducing global memory random access conflicts, improving data locality and GPU memory bandwidth utilization. Furthermore, this mechanism provides the reconstruction process with flexible breakpoint continuation and abnormal interruption response capabilities, allowing for safe resource release and saving of intermediate states after any sub-block iteration is completed, enhancing the robustness of engineering implementation. This application, while ensuring mathematically equivalent convergence of the SART iteration, achieves a comprehensive technical effect of minimizing GPU memory usage, maximizing computational transfer overlap, maximizing parallel efficiency, and optimizing system stability, providing reliable technical support for the engineering deployment of high-resolution, large-field-of-view industrial CT.

[0039] In some embodiments, the parallel iterative reconstruction of each current voxel sub-block in the current voxel sub-block data based on the current projection data, via computational flow, includes: For each current voxel in the current voxel sub-block, in the current iteration, the absolute value of the current neighborhood grayscale difference of each current voxel in the three coordinate axis directions is calculated in parallel through the computation flow based on the current voxel value of each current voxel. The absolute values ​​of the gray-level differences of the three current neighborhoods corresponding to each current voxel are summed to obtain the current gradient value corresponding to each current voxel. Based on the current gradient value corresponding to each current voxel, update the current voxel value corresponding to each current voxel in parallel.

[0040] In this embodiment, the current voxel refers to the voxel whose voxel value is being updated in the current voxel sub-block. The three coordinate axes are the X-axis, Y-axis, and Z-axis.

[0041] For each current voxel, based on the gray levels of its six surrounding neighborhoods, and the absolute value of the gray level difference in each coordinate axis direction, the formula for calculating the current gradient value corresponding to the current voxel is as follows: Current gradient value Used to quantify the current voxel In the The intensity of local grayscale change during the next iteration. As shown in the equation, it is based on the neighboring voxels of the current voxel in the positive X-axis direction. and neighboring voxels in the negative X-axis direction Calculate the absolute value of the gray-level difference of the current neighborhood along the X-axis. Based on the current voxel's neighboring voxels along the positive Y-axis. and neighboring voxels in the negative Y-axis direction Calculate the absolute value of the gray-level difference of the current neighborhood along the Y-axis. Based on the neighboring voxels of the current voxel in the positive Z-axis direction. and neighboring voxels in the negative Z-axis direction Calculate the absolute value of the gray-level difference of the current neighborhood along the Z-axis. . Indicates the current voxel No. The current voxel value in the next iteration. A larger current gradient value indicates that the current voxel is in a region of abrupt gray-level changes, such as workpiece edges, defect boundaries, or material interfaces; a smaller current gradient value indicates that the current voxel is in a region of uniform gray-level changes. By embedding this current gradient value as a regularization constraint term into the iterative update formula, enhanced protection of edge details or differentiated control of noise smoothing can be achieved during the iteration process. The gradient constraint weights are dynamically assigned based on the current iteration round.

[0042] Optionally, based on the current gradient value corresponding to each current voxel, updating the current voxel value in parallel includes: For the The current voxel, at the _ ... During the iteration, based on the first The current voxel is in the... The current gradient value of the nth iteration is used to update the 1st iteration using the reconstruction formula. The current voxel is in the... The current voxel value in the nth iteration will be updated to the nth voxel value. The current voxel value of the nth iteration is used as the... The current voxel The current voxel value in the next iteration, where the reconstruction formula is as follows: in Indicates the first The current voxel The current voxel value in the next iteration. Indicates the first The current voxel The current voxel value in the next iteration. To preset the iterative relaxation factor, Indicates the first The numbering of the projection rays, Indicates passing through the first The set of indices of all projected rays for a current voxel. Indicates the first The projection ray to the first The projection contribution coefficient of each current voxel Indicates the first Measured projection values ​​of the projected rays. For the first The set of voxels through which a projection ray passes. Represents a set of voxels The Middle The index number of the individual element, Indicates the first The projection ray to the first The projection contribution coefficient of individual prime blocks. Indicates the first Dynamic gradient constraint weights for each iteration. Indicates the first The current voxel is in the... The current gradient value in the next iteration. Represents a set of voxels The Middle Individual element The voxel value of the next iteration.

[0043] Can The corresponding projection contribution coefficient index is obtained from the bidirectional index mapping table. similar.

[0044] Parallel reconstruction is performed based on the SART iterative framework, with a fixed total number of lightweight iterations. During the voxel grayscale correction process in each iteration, an adaptive 3D gradient constraint mechanism is introduced. The current gradient value is generated by calculating the grayscale gradient difference of each voxel in the 3D six-neighborhood in parallel. An adaptive scheduling function for constraint weights driven by the iteration round is constructed, dividing the total iteration cycle into three stages: feature enhancement, transition decay, and noise reduction convergence. In the early stage of iteration, high gradient constraint weights are used to amplify the edge gradient difference response and enhance the feature preservation of workpiece edges and minor defects. In the middle stage of iteration, a continuous decay strategy is used to achieve a smooth transition of weights to maintain numerical stability. In the later stage of iteration, low gradient constraint weights are used to weaken the penalty for high-frequency components and suppress the accumulation of iteration noise and overfitting. The dynamic constraint terms are weighted and superimposed into the SART voxel grayscale update formula and completed synchronously with the iterative correction, realizing the adaptive evolution of constraint strength with the iteration process.

[0045] The iterative process can be divided into a first iteration and a second iteration. The dynamic gradient constraint weight corresponding to the first iteration is greater than that corresponding to the second iteration. The second iteration follows the first iteration. For example, iterations 1-8 are the feature enhancement stage, where the dynamic gradient constraint weight is set to 0.8 to amplify the gradient differences in edge regions, enhance the feature preservation of workpiece edges, micro-cracks, and pores, and avoid blurring of details; iterations 9-15 are the noise reduction and convergence stage, where the dynamic gradient constraint weight is reduced to 0.3 to smooth gradient fluctuations, suppress noise accumulated during the iteration process, and alleviate overfitting.

[0046] In some embodiments, the method further includes: Obtain the overlapping buffer region between adjacent projected sub-blocks, and determine the overlapping voxel sub-block corresponding to the overlapping buffer region according to the bidirectional index mapping table; For the first and second voxel sub-blocks that are adjacent in the overlapping voxel sub-blocks; determine the overlapping voxels of the first and second voxel sub-blocks according to the bidirectional index mapping table; From the reconstructed image, obtain the first voxel value of the overlapping voxel in the first voxel sub-block and the second voxel value of the overlapping voxel in the second voxel sub-block. Based on the first voxel value and the second voxel value, a weighted average is performed to determine the corrected voxel value for the overlapping voxels.

[0047] In this embodiment, a pre-defined overlap buffer area is reserved around the edges of each projection sub-block to eliminate artifacts during subsequent block stitching. Simultaneously, the original floating-point projection data is converted to a low-precision integer storage format, reducing data storage footprint and read / write bandwidth while preserving defect features and grayscale gradient information. For example, an 8-pixel-wide overlap buffer area is reserved around the edges of each projection sub-block, ensuring pixel overlap between adjacent sub-blocks and providing a transition zone for subsequent stitching and fusion. The original 32-bit floating-point projection data is converted to an unsigned short 16-bit unsigned integer storage format, reducing data storage footprint and read / write bandwidth by 50% while preserving 5μm-level defect features and projection grayscale gradient information. Finally, a standardized block projection dataset is generated and stored in the host memory.

[0048] Optionally, based on the first voxel value and the second voxel value, a weighted average is performed to determine the corrected voxel value of the overlapping voxel, including: Obtain the first distance between the center of the overlapping voxel and the center of the first voxel sub-block, and obtain the second distance between the center of the overlapping voxel and the center of the second voxel sub-block; Based on the first distance and the second distance, determine the first weight corresponding to the first distance and the second weight corresponding to the second distance, where the greater the distance, the smaller the corresponding weight; The correction voxel value is determined by weighted summation of the first weight and the first voxel value, the second weight and the second voxel value.

[0049] Overlapping voxels represent the voxels corresponding to the overlapping buffer regions. For the overlapping buffer regions reserved at the edges of adjacent projected sub-blocks, a Gaussian weighted fusion algorithm is used to correct the voxel values ​​of the overlapping voxels. For each overlapping voxel, a portion is located in the first voxel sub-block, and a portion is located in the second voxel sub-block. Gaussian weights are assigned according to the distance of the overlapping voxel from the center of its respective voxel sub-block. The voxel values ​​of the same overlapping voxel output from adjacent voxel sub-blocks are summed according to the weights to obtain the final corrected voxel value. By smoothing the gray-level abrupt changes at the boundaries of the overlapping voxel sub-blocks, block stitching artifacts are eliminated. After all voxel sub-blocks are fused and stitched, complete and continuous 3D volume data is generated, i.e., the reconstructed image is corrected, resulting in the corrected reconstructed image. Corrected voxel value = First weight × First voxel value + Second weight × Second voxel value.

[0050] like Figure 4 As shown, Figure 4This is a schematic diagram of adjacent voxel weighted fusion in one embodiment. A Gaussian weighted fusion algorithm is used for correction. Taking two horizontally adjacent voxel sub-blocks as an example, the right edge of the left sub-block and the left edge of the right sub-block have 8 columns of voxels that coincide in spatial position, forming an overlapping buffer region with a width of 8 voxels. For each overlapping voxel in the overlapping buffer region, its distance to the center of the left sub-block and the center of the right sub-block is calculated. Based on the Gaussian function, corresponding weights are assigned; the closer to the center of the sub-block, the higher the weight of the gray value of that sub-block. The voxel values ​​of the same overlapping voxel output from two adjacent voxel sub-blocks are summed according to the weights to obtain the corrected voxel value of the overlapping voxel. The overlapping voxels of all adjacent voxel sub-blocks are corrected in the above way, smoothing boundary gray-level abrupt changes and eliminating block stitching tomographic artifacts. Finally, a complete and continuous corrected reconstructed image is generated by stitching together, outputting CT tomographic imaging results in the axial, coronal, and sagittal directions.

[0051] In some embodiments, on the XY plane formed by the X and Y axes, each thread traverses sequentially along the Z-axis for one voxel point. For example, for a voxel sub-block with a size of 128×128×512, a two-dimensional thread mesh is used to map the XY plane of the voxel cross-section, with each thread corresponding to one voxel point and traversing sequentially along the Z-axis. The thread block size is configured as 16×16, with each block containing 256 threads, corresponding to a 16×16 voxel region; the mesh size is configured as 8×8, covering a 128×128 voxel cross-section. This configuration conforms to the CUDA warp scheduling characteristics (NVIDIA GPUs schedule execution in units of 32 threads per warp, and 16×16=256 threads are exactly 8 complete warps, with no fragmentation waste), enabling efficient parallel scheduling of threads and avoiding branch divergence. The thread design here, in the SART iterative correction formula mentioned above, independently performs three steps of computation for each voxel: projection error backprojection, voxel correction, and gradient constraint superposition, with no data dependency between the three. The CUDA thread grid maps each voxel to an independent thread, and 128×128 voxels are processed in parallel by 8×8=64 thread blocks. Within each thread block, 256 threads cooperate through shared memory to complete the six-neighbor gradient calculation of a 16×16 voxel region (corresponding to the parallel calculation of the current gradient value mentioned above). This transforms the described voxel-by-voxel serial algorithm into large-scale parallel execution on the GPU, achieving logical unification between the algorithm description and the thread design.

[0052] In another aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the block-based streaming iteration-based three-dimensional CT image reconstruction method of any embodiment of this application.

[0053] Among them, in the computer program product, the optional implementation form of the program module architecture of the computer program that implements each step of the target recognition method can be a three-dimensional CT image reconstruction device based on block-based streaming iteration.

[0054] Please see Figure 5 One embodiment of this application provides a three-dimensional CT image reconstruction device based on block-based streaming iteration, comprising: an acquisition module 51, used to acquire projection images of the target to be reconstructed at multiple projection angles; a partitioning module 52, used to perform block partitioning on the projection images to obtain multiple projection sub-blocks; the partitioning module 52 is further used to partition the three-dimensional voxel space with the same size as the target to be reconstructed to obtain multiple voxel sub-blocks; a determination module 53, used to construct a bidirectional index mapping table based on the multiple projection sub-blocks and the multiple voxel sub-blocks, the bidirectional index mapping table including the voxel space coordinate range corresponding to each projection sub-block and the projection sub-block corresponding to each voxel sub-block; and a reconstruction module 54, used to perform streaming parallel scheduling based on the bidirectional index mapping table to iteratively reconstruct the multiple projection sub-blocks to obtain a reconstructed image of the target to be reconstructed.

[0055] Please see Figure 6 In another aspect of this application, a computing device 10 is also provided, including a memory 3011 and a processor 3012. The memory 3011 stores a computer program, and when the computer program is executed by the processor, the processor 3012 performs the steps of the block-based streaming iterative three-dimensional CT image reconstruction method provided in any of the above embodiments of this application. The computing device 10 can be (e.g., a desktop computer, laptop computer, tablet computer, handheld computer, smart speaker, server, etc.), a mobile phone (e.g., a smartphone, cordless phone, etc.), a wearable device (e.g., a pair of smart glasses or a smartwatch), or a similar device.

[0056] The processor 3012 is the control center, connecting various parts of the computer device via various interfaces and lines. It executes software programs and / or modules stored in the memory 3011, and calls data stored in the memory 3011 to perform various functions and process data. Optionally, the processor 3012 may include one or more processing cores; preferably, the processor 3012 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user page, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 3012.

[0057] The memory 3011 can be used to store software programs and modules. The processor 3012 executes various functional applications and data processing by running the software programs and modules stored in the memory 3011. The memory 3011 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 3011 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 3011 may also include a memory controller to provide the processor 3012 with access to the memory 3011.

[0058] In another aspect, this application also provides a storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the block-based streaming iteration-based three-dimensional CT image reconstruction method provided in any of the above embodiments of this application.

[0059] Those skilled in the art will understand that all or part of the processes in the methods provided in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0060] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A three-dimensional CT image reconstruction method based on block-based streaming iteration, characterized in that, include: Acquire projection images of the target to be reconstructed at multiple projection angles; The projected image is divided into blocks to obtain multiple projected sub-blocks; The three-dimensional voxel space with the same size as the target to be reconstructed is divided into blocks to obtain multiple voxel sub-blocks; Based on the multiple projection sub-blocks and the multiple voxel sub-blocks, a bidirectional index mapping table is constructed. The bidirectional index mapping table includes the voxel space coordinate range corresponding to each projection sub-block and the projection sub-block corresponding to each voxel sub-block. Based on the bidirectional index mapping table, streaming parallel scheduling is performed to iteratively reconstruct multiple projection sub-blocks to obtain the reconstructed image of the target to be reconstructed.

2. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 1, characterized in that, The construction of the bidirectional index mapping table based on multiple projection sub-blocks and multiple voxel sub-blocks includes: For projection sub-blocks The projection sub-block is determined using a mapping formula. The target voxel space coordinate range, wherein the mapping formula is as follows: , in For the projection sub-block The horizontal index, For the projection sub-block The vertical index, The x-coordinate of the target voxel within the target voxel space coordinate range. The vertical coordinate of the target voxel is defined as the target voxel's coordinate within the target voxel's spatial coordinate range. The target voxel's Z-coordinate is defined within the target voxel's spatial coordinate range. The total dimension of voxels on the X-axis. The total dimensions of the voxels on the Y-axis. The total dimension of voxels on the Z-axis, where This represents the total number of voxel blocks on the X-axis. This represents the total number of voxel blocks on the Y-axis; The target voxel sub-blocks corresponding to the target voxel's x-coordinate, y-coordinate, and Z-coordinate, and the projection sub-blocks. Correspondingly.

3. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 1, characterized in that, The process of performing streaming parallel scheduling based on the bidirectional index mapping table to iteratively reconstruct multiple projection sub-blocks and obtain the reconstructed image of the target to be reconstructed includes: Configure data transfer streams and computation streams, wherein the data transfer stream is responsible for copying data between the host's memory and the graphics processor's video memory, and the computation stream is used to perform iterative reconstruction; Based on the order of each projection sub-block in the bidirectional index mapping table, the current projection data and the corresponding current voxel sub-block data corresponding to the current projection sub-block are obtained from the bidirectional index mapping table through the data transmission stream and loaded into the video memory. Through the computational flow, in the current iteration, based on the current projection data, the voxels of each current voxel sub-block in the current voxel sub-block data are updated in parallel iteratively. After the current voxel sub-block completes the update of the current iteration, the resources occupied by the current voxel sub-block are released in the video memory.

4. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 3, characterized in that, The method further includes: While the computation stream is performing iterative reconstruction of the current projection sub-block, the data transmission stream synchronously copies the projection data of the next projection sub-block and the corresponding voxel sub-block data of the current projection sub-block in the sequence from the host memory to the graphics processor's video memory.

5. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 3, characterized in that, The step of reconstructing each current voxel sub-block in the current voxel sub-block data in parallel iteratively based on the current projection data through the computational flow in the current iteration includes: For each current voxel in the current voxel sub-block, in the current iteration, the absolute value of the current neighborhood grayscale difference of each current voxel in the three coordinate axis directions is calculated in parallel through the computation flow based on the current voxel value of each current voxel. The absolute values ​​of the gray-level differences of the three current neighborhoods corresponding to each current voxel are summed to obtain the current gradient value corresponding to each current voxel. Based on the current gradient value corresponding to each current voxel, update the current voxel value corresponding to each current voxel in parallel.

6. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 5, characterized in that, The parallel updating of the current voxel value based on the current gradient value of each current voxel includes: For the The current voxel, at the _ ... During the iteration, based on the first The current voxel is in the... The current gradient value of the nth iteration is used to update the 1st iteration using the reconstruction formula. The current voxel is in the... The current voxel value in the nth iteration will be updated to the nth voxel value. The current voxel value of the nth iteration is used as the... The current voxel The current voxel value of the next iteration, wherein the reconstruction formula is as follows: in Indicates the first The current voxel The current voxel value in the next iteration. Indicates the first The current voxel The current voxel value in the next iteration. To preset the iterative relaxation factor, Indicates the first The numbering of the projection rays, Indicates passing through the first The set of indices of all projected rays for a current voxel. Indicates the first The projection ray to the first The projection contribution coefficient of each current voxel Indicates the first Measured projection values ​​of the projected rays. For the first The set of voxels through which a projection ray passes. Represents a set of voxels The Middle The index number of the individual element, Indicates the first The projection ray to the first The projection contribution coefficient of individual prime blocks. Indicates the first Dynamic gradient constraint weights for each iteration. Indicates the first The current voxel is in the... The current gradient value in the next iteration. Represents a set of voxels The Middle Individual element The voxel value of the next iteration.

7. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 6, characterized in that, The iterative process includes a first iterative process and a second iterative process, wherein the dynamic gradient constraint weights corresponding to the first iterative process are greater than the dynamic gradient constraint weights corresponding to the second iterative process, and the second iterative process is located after the first iterative process.

8. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 1, characterized in that, The method further includes: Obtain the overlapping buffer region between adjacent projected sub-blocks, and determine the overlapping voxel sub-block corresponding to the overlapping buffer region according to the bidirectional index mapping table; For the first and second voxel sub-blocks that are adjacent in the overlapping voxel sub-blocks; the overlapping voxels of the first and second voxel sub-blocks are determined according to the bidirectional index mapping table; From the reconstructed image, obtain the first voxel value of the overlapping voxel in the first voxel sub-block, and obtain the second voxel value of the overlapping voxel in the second voxel sub-block; Based on the first voxel value and the second voxel value, a weighted average is performed to determine the corrected voxel value of the overlapping voxel.

9. The three-dimensional CT image reconstruction method based on block-based streaming iteration as described in claim 8, characterized in that, The step of performing weighted averaging based on the first voxel value and the second voxel value to determine the corrected voxel value of the overlapping voxel includes: Obtain a first distance between the center of the overlapping voxel and the center of the first voxel sub-block, and obtain a second distance between the center of the overlapping voxel and the center of the second voxel sub-block; Based on the first distance and the second distance, a first weight corresponding to the first distance and a second weight corresponding to the second distance are determined, wherein the greater the distance, the smaller the corresponding weight; The corrected voxel value is determined by weighting and summing the first weight and the first voxel value, and the second weight and the second voxel value.

10. A computing device, characterized in that, The system includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the block-based streaming iteration-based three-dimensional CT image reconstruction method as described in any one of claims 1 to 9.