CN-122023667-A - CT reconstruction method, system, equipment and medium for three-dimensional sparse view angle

CN122023667ACN 122023667 ACN122023667 ACN 122023667ACN-122023667-A

Abstract

The invention provides a CT reconstruction method, a system, equipment and a medium of a three-dimensional sparse view. The method comprises the steps of obtaining projection data and corresponding imaging geometric parameters formed under a sparse view angle scanning condition, determining a ray path set based on the projection data, generating a plurality of three-dimensional sampling point coordinates on each ray path according to the imaging geometric parameters corresponding to the ray path, performing feature mapping and interpolation processing on the three-dimensional sampling point coordinates based on hash codes to generate point feature vectors corresponding to all three-dimensional sampling points, inputting all the point feature vectors into a neural network model to obtain attenuation coefficients corresponding to all the three-dimensional sampling points, integrating and accumulating the attenuation coefficients of all the sampling points along a ray direction to obtain a projection result of the ray path, and generating a three-dimensional reconstruction image based on the projection result of all the ray paths. The invention can realize high-quality low-power-consumption three-dimensional CT reconstruction processing under the sparse view angle scanning condition.

Inventors

LOU XIN
LI HAOYAN
WAN HAOCHUAN
ZHANG XIANGYU

Assignees

上海科技大学

Dates

Publication Date: 20260512
Application Date: 20260212

Claims (10)

1. A method for CT reconstruction of a three-dimensional sparse view, the method comprising: acquiring projection data formed under a sparse view angle scanning condition and corresponding imaging geometric parameters; Determining a set of ray paths based on the projection data; the following is done for each ray path: Generating a plurality of three-dimensional sampling point coordinates on the ray path according to imaging geometric parameters corresponding to the ray path; based on hash coding, performing feature mapping and interpolation processing on the coordinates of each three-dimensional sampling point to generate point feature vectors corresponding to each three-dimensional sampling point; Inputting the feature vectors of each point into a neural network model to obtain attenuation coefficients corresponding to each three-dimensional sampling point; Integrating and accumulating attenuation coefficients of all sampling points along the ray direction to obtain a projection result of the ray path; A three-dimensional reconstructed image is generated based on the projection results of the respective ray paths.
2. The method of three-dimensional sparse view CT reconstruction of claim 1, wherein generating a plurality of three-dimensional sampling point coordinates on the ray path according to imaging geometry parameters corresponding to the ray path comprises: determining starting point coordinates and direction vectors of the ray paths according to imaging geometric parameters; And taking the starting point coordinates as a starting point, and equidistant sampling the ray paths along the direction of the direction vector according to a preset sampling step length to obtain the coordinates of each three-dimensional sampling point.
3. The CT reconstruction method for three-dimensional sparse view according to claim 1, wherein the point feature vector is obtained by performing hash coding on a multi-scale mapping layer, and performing feature mapping and interpolation processing on coordinates of three-dimensional sampling points based on the hash coding, and the step of generating the point feature vector corresponding to the three-dimensional sampling points comprises: for each mapping layer: mapping the coordinates of the three-dimensional sampling points to a space voxel unit corresponding to the mapping layer, and determining a plurality of voxel vertexes adjacent to the three-dimensional sampling points in the space voxel unit; Calculating hash indexes of each voxel vertex, and reading feature vectors from a hash table prestored in the mapping layer according to the hash indexes; Calculating interpolation weights corresponding to the vertexes of all voxels according to the relative position relation of the three-dimensional sampling points in the space voxel units corresponding to the mapping layer; according to the interpolation weight of each voxel vertex, carrying out weighted fusion on the corresponding feature vector to obtain a fusion result corresponding to the mapping layer; And performing splicing processing on fusion results corresponding to the mapping layers to generate point feature vectors corresponding to the three-dimensional sampling points.
4. A CT reconstruction method for three-dimensional sparse view according to claim 3, wherein the step of mapping the three-dimensional sampling point coordinates to a spatial voxel unit corresponding to the mapped layer and determining a plurality of voxel vertices adjacent to the three-dimensional sampling point in the spatial voxel unit comprises: performing scale transformation corresponding to the mapping layer on the three-dimensional sampling points to obtain mapped coordinates; performing rounding operation on the mapped coordinate values to determine the index of the spatial voxel unit of the three-dimensional sampling point in the mapping layer; Determining a boundary range of the spatial voxel unit containing the three-dimensional sampling point based on the spatial voxel unit index; A plurality of voxel vertices corresponding to the spatial voxel unit are generated based on the boundary range.
5. The method for CT reconstruction of a three-dimensional sparse view of claim 4, wherein for each voxel vertex, the step of calculating an interpolation weight corresponding to each voxel vertex according to a relative positional relationship of a three-dimensional sampling point in a spatial voxel unit corresponding to the mapping layer comprises: reading the decimal part of the mapped coordinate value; according to the decimal part, respectively calculating weight components of the space voxel units in different dimension directions; and carrying out combination operation on the weight components in each dimension direction to generate interpolation weights corresponding to the voxel vertexes.
6. The CT reconstruction method for three-dimensional sparse view according to claim 3, wherein the step of calculating a hash index of each voxel vertex for each voxel vertex and reading the feature vector from the hash table pre-stored in the mapping layer according to the hash index comprises: calculating a hash index corresponding to the voxel vertex according to a preset hash mapping rule; Address mapping processing is carried out on the hash index, and an effective index address for storage access is generated; And reading the feature vector associated with the voxel vertex from a hash table prestored in the corresponding mapping layer according to the effective index address.
7. The method for reconstructing the CT of the three-dimensional sparse view according to claim 1, wherein the step of inputting the point feature vector into the neural network model to obtain the attenuation coefficient corresponding to the three-dimensional sampling point comprises the steps of inputting the point feature vector into the neural network model, loading the point feature vector into a parallel multiplication and addition operation array based on a weight multiplexing mechanism, and performing parallel matrix multiplication and addition calculation on the point feature vector by performing block storage and repeated calling on network weights to obtain the attenuation coefficient corresponding to the three-dimensional sampling point.
8. A CT reconstruction system for three-dimensional sparse viewing angles, the system comprising: The data acquisition module is used for acquiring projection data and corresponding imaging geometric parameters formed under the sparse view angle scanning condition; A ray path determination module for determining a set of ray paths based on the projection data; A ray processing module, configured to perform the following processing for each ray path: Generating a plurality of three-dimensional sampling point coordinates on the ray path according to imaging geometric parameters corresponding to the ray path; based on hash coding, performing feature mapping and interpolation processing on the coordinates of each three-dimensional sampling point to generate point feature vectors corresponding to each three-dimensional sampling point; Inputting the feature vectors of each point into a neural network model to obtain attenuation coefficients corresponding to each three-dimensional sampling point; Integrating and accumulating attenuation coefficients of all sampling points along the ray direction to obtain a projection result of the ray path; and the reconstruction module is used for generating a three-dimensional reconstruction image based on the projection result of each ray path.
9. An electronic device, the electronic device comprising: one or more processors; storage means for storing one or more programs which, when executed by the one or more processors, cause the electronic device to implement the CT reconstruction method of three-dimensional sparse view as claimed in any one of claims 1 to 7.
10. A computer-readable storage medium, having stored thereon a computer program which, when executed by a processor of a computer, causes the computer to perform the CT reconstruction method of three-dimensional sparse view according to any one of claims 1 to 7.

Description

CT reconstruction method, system, equipment and medium for three-dimensional sparse view angle Technical Field The invention relates to the technical field of CT data analysis, in particular to a CT reconstruction method, a system, equipment and a medium of a three-dimensional sparse view angle. Background Computer tomography (Computed Tomography, CT) is widely used as an important medical imaging technique in clinical diagnosis, disease screening, and intra-operative navigation. With increasing requirements of clinical applications on radiation safety, sparse View CT (SVCT) is one of the important development directions of CT imaging by reducing the number of X-ray projections to reduce patient radiation exposure. However, sparse sampling makes projection data underdetermined, and the traditional analytic reconstruction algorithm is easy to generate problems of artifacts, structural distortion, detail loss and the like in the scene, so that a high-quality reconstruction result meeting clinical diagnosis requirements is difficult to obtain. In order to improve sparse view CT imaging quality, in recent years, research work begins to introduce a deep learning and implicit neural representation (Implicit Neural Representation, INR) method for modeling three-dimensional CT volume data, the method generally takes continuous space coordinates as input, obtains attenuation coefficients through feature mapping and neural network reasoning, and performs integration or consistency constraint by combining a ray geometric relationship, so that expression of a continuous volume density field is realized. However, if the INR sparse view CT reconstruction is to be towards engineering deployment, especially towards embedded, portable or low-power consumption scenes, the method still faces outstanding hardware implementation challenges, on one hand, the method generally needs to repeatedly execute processing links such as coordinate generation, feature construction, network reasoning, ray accumulation and the like on a large number of sampling points, calculation and data carrying overheads are large, on the other hand, the feature construction and network reasoning process often accompanies frequent storage access and data interaction, and high requirements are put on storage bandwidth, on-chip cache capacity and data multiplexing strategies. Disclosure of Invention The invention provides a CT reconstruction method, a system, equipment and a medium of a three-dimensional sparse view, which are used for solving the technical problem that the existing reconstruction mode is difficult to balance in terms of reconstruction precision and resource occupation. The invention provides a CT reconstruction method of a three-dimensional sparse view angle, which comprises the steps of obtaining projection data and corresponding imaging geometric parameters formed under a sparse view angle scanning condition, determining a ray path set based on the projection data, generating a plurality of three-dimensional sampling point coordinates on each ray path according to the imaging geometric parameters corresponding to the ray path, performing feature mapping and interpolation processing on each three-dimensional sampling point coordinate based on hash codes to generate point feature vectors corresponding to each three-dimensional sampling point, inputting each point feature vector into a neural network model to obtain attenuation coefficients corresponding to each three-dimensional sampling point, integrating and accumulating the attenuation coefficients of each sampling point along the ray direction to obtain a projection result of the ray path, and generating a three-dimensional reconstruction image based on the projection result of each ray path. In an embodiment of the invention, the step of generating a plurality of three-dimensional sampling point coordinates on the ray path according to the imaging geometric parameters corresponding to the ray path includes determining starting point coordinates and direction vectors of the ray path according to the imaging geometric parameters, and equidistant sampling the ray path along the direction of the direction vectors by taking the starting point coordinates as the starting point according to a preset sampling step length to obtain each three-dimensional sampling point coordinate. In one embodiment of the invention, the point feature vector is obtained by carrying out hash coding on a multi-scale mapping layer, and the step of generating the point feature vector corresponding to the three-dimensional sampling point by carrying out feature mapping and interpolation processing on the three-dimensional sampling point coordinate based on the hash coding comprises the steps of mapping the three-dimensional sampling point coordinate to a space voxel unit corresponding to the mapping layer and determining a plurality of voxel vertexes adjacent to the three-dimensional sampling point in the space vox