A cone beam CT reconstruction method based on multi-scale hash coding
By using a self-supervised multi-scale hash coding method to map CT projection data into a high-dimensional space and using a multilayer perceptron to predict the attenuation coefficient, the problems of slow reconstruction speed and insufficient accuracy of CBCT are solved, achieving fast and accurate CT data reconstruction and reducing the radiation risk to patients.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2023-12-18
- Publication Date
- 2026-07-31
AI Technical Summary
Existing CBCT reconstruction methods perform poorly when there are sparse angles or insufficient angle coverage. Furthermore, iterative reconstruction methods consume a lot of computation time and memory, while learning-based methods are difficult and time-consuming to collect training data.
A self-supervised multi-scale hashing encoding method is adopted to map CT projection data to a high-dimensional space. The attenuation coefficient is predicted by a multilayer perceptron and trained in combination with a self-supervised network framework to reduce feature vector updates and accelerate the network training process.
It improves the accuracy and speed of CBCT reconstruction, reduces X-ray radiation dose, and lowers the risk of disease for patients.
Smart Images
Figure CN117710343B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of CT imaging technology, specifically relating to a cone-beam CT reconstruction method based on multi-scale hash coding. Background Technology
[0002] Cone-beam computed tomography (CBCT) offers advantages over traditional CT techniques, including faster imaging speed and clearer images. CT imaging utilizes the principle that X-rays can penetrate many substances. Human tissue is irradiated with X-rays from multiple angles, and the attenuated X-ray dose is measured. CT reconstruction algorithms calculate the attenuation coefficients of different tissues based on the emitted X-ray dose, a discretized model of the object, and the measured attenuated X-ray dose, thus generating an image. To reduce the risk of cancer due to excessive X-ray exposure, the radiation dose can be lowered by reducing the number of projections in the CBCT scan.
[0003] CBCT reconstruction is the process of obtaining the attenuation coefficients of various locations in human tissue using partial projection data. Currently, CBCT reconstruction methods can be divided into three categories. First, analytical reconstruction methods estimate the attenuation coefficients through the Radon transform and its inverse transform. While they can achieve good reconstruction results under ideal conditions, they perform poorly when sparse angles or insufficient angle coverage are encountered. Second, iterative reconstruction methods introduce a regularized optimization framework, treating the reconstruction process as a minimization process. They can solve the problems of analytical reconstruction but consume significant computational time and memory. Third, learning-based methods, most of which suffer from difficulties in collecting sufficiently labeled training data and long training times.
[0004] In summary, this invention proposes a self-supervised cone-beam CT reconstruction method based on multi-scale hash coding. This method only requires X-ray projection data for training, which can accelerate the training process of the network while ensuring the accuracy of the reconstructed CT data. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention proposes a cone-beam CT reconstruction method based on multi-scale hash coding, which includes:
[0006] S1: Scan the physical object and acquire CT projection data;
[0007] S2: Generate X-rays from the X-ray source to the direction of the projection pixels based on the CT projection data, and uniformly sample N points at the intersection of the X-rays and the object;
[0008] S3: Multi-scale hash coding is used to map the location information of the sampling points to a high-dimensional space to obtain the feature vector of the sampling points;
[0009] S4: Input the feature vector of the sampling point into the multilayer perceptron for processing to obtain the attenuation coefficient of the sampling point;
[0010] S5: Calculate the total loss based on the attenuation coefficient of the sampling points, and adjust the multilayer perceptron parameters based on the total loss to obtain the trained multilayer perceptron.
[0011] S6: The trained multilayer perceptron is used to process the physical object to be reconstructed by CT to obtain the CT reconstruction result of the physical object.
[0012] Preferably, the generated ray is represented as follows:
[0013]
[0014] Where, r k,i,j Indicates the distance from the X-ray source to the projected pixel I. k,i,j Rays in the direction, O k Indicates the generation of projection I k The location of the X-ray source, t represents the distance from the sampling point to the X-ray source. Represents ray r k,i,j The direction vector is usually a unit vector.
[0015] Preferably, the process of mapping the location information of sampling points to a high-dimensional space includes:
[0016] Map all grid points of the voxel at each scale of the sampling point to a hash table to obtain the feature vector of each grid point at each scale.
[0017] Based on the feature vectors of each grid point, linear interpolation is used to obtain the initial feature vectors of the sampling points at each scale.
[0018] By concatenating the initial feature vectors of the sampling points at all scales, the final feature vector of the sampling points is obtained.
[0019] Furthermore, the formula for mapping grid points to hash tables is:
[0020]
[0021] Where h(Z) represents the eigenvector of grid point X, and x1, x2, and x3 represent the x, y, and y coordinates of grid point X, respectively. π1, π2, and π3 represent the first, second, and third largest prime numbers, respectively, and T represents the number of entries in the hash table.
[0022] Preferably, the process of calculating the total loss based on the attenuation coefficient of the sampling points includes:
[0023] The attenuation coefficients of all sampling points on the composite ray are used to obtain the composite projection of the ray;
[0024] The total loss is calculated based on the composite projection and the true projection of the rays.
[0025] Furthermore, the formula for the attenuation coefficient of all sampling points on the synthetic ray is:
[0026]
[0027] Where I represents the composite projection of the ray, I0 represents the initial intensity of the ray, and σ i δ represents the attenuation coefficient at the i-th sampling point, N represents the number of sampling points on the ray, and δ i This represents the distance between the i-th sampling point and the (i+1)-th sampling point.
[0028] Furthermore, the formula for calculating the total loss is:
[0029]
[0030] Where Loss represents the total loss, I r (r) represents the true projection of ray r, I s (r) represents the composite projection of the true projection, r represents ray r, and R represents the set of rays.
[0031] The beneficial effects of this invention are as follows: This invention represents CT data as a mapping from spatial coordinates to attenuation coefficients and combines it with a self-supervised network framework to support the reconstruction from two-dimensional projection to three-dimensional CT data. Multi-scale hashing encoding is used to map the input of the neural network to a high-dimensional space to better fit the frequently changing data, thereby ensuring the accuracy of the reconstructed CT data. This invention uses multi-scale hashing encoding to combine hash tables and neural networks, storing feature vectors in a structured manner, allocating most of the learning task to the data structure. Since only a small number of feature vectors need to be updated during each backpropagation, the training process of the network is accelerated. Attached Figure Description
[0032] Figure 1 This is a flowchart of the cone-beam CT reconstruction method based on multi-scale hash coding in this invention;
[0033] Figure 2 This is a schematic diagram of the geometric structure of the CBCT scanner during scanning in this invention;
[0034] Figure 3 This is a schematic diagram of the multi-scale CT data grid in this invention. Detailed Implementation
[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0036] This invention proposes a cone-beam computed tomography (CBCT) reconstruction method based on multi-scale hash coding, such as... Figure 1 As shown, the method includes the following:
[0037] S1: Scan the physical object and acquire CT projection data.
[0038] To successfully acquire projection data when scanning a physical object using a CBCT scanner, the geometry of the CBCT scanner needs to be determined; for example... Figure 2 As shown, the geometric structure of this invention when using a CBCT scanner for scanning is as follows:
[0039] The X-ray source S is located at a distance DSO from the center of rotation O, and the origin of the Cartesian coordinate system is located at the center of rotation O. The X-ray source irradiation includes CT data. A cone-shaped region and a detector The detector measures the intensity of photons that strike it, photons that have decayed according to Beer-Lambert's law. The CT data is centered at position O', which is offset from the origin of the coordinate system by... The detector is located at a distance of DSD from the X-ray source, with its center at D′, and the offset between it and D is... D is a point in the xy-plane at a distance of DSD-DSO from the origin. The projected coordinate system uv is defined with the lower left corner of the detector as its center. During the acquisition of projected data, the X-ray source and detector rotate around the z-axis, making an angle α (rotation interval α degrees) with the initial position.
[0040] After determining the geometry of the CBCT scanner, the X-ray source rotates around the object and emits a cone-shaped X-ray beam, detecting the projection data of the X-rays at equal angular intervals in a 2D plane.
[0041] S2: Generate X-rays from the X-ray source to the direction of the projection pixels based on the CT projection data, and uniformly sample N points at the intersection of the X-rays and the object.
[0042] Each pixel value in the projection data is the result of X-rays passing through the object (CT data) and being attenuated by the internal medium. X-ray source to projection I k pixel I k i, j The ray r in the direction is represented as:
[0043]
[0044] Where, r k,i,j Indicates the distance from the X-ray source to the projected pixel I. k,i,j A ray in the direction of projection, where k is the index of the projection and (i, j) is the index of the projected pixel; O k Indicates the generation of projection I k The location of the X-ray source is given by t, where t represents the distance from the sampling point to the X-ray source. Represents ray r k,i,j The direction vector, usually a unit vector, is calculated from the difference between the pixel position on the detector plane and the position of the X-ray source.
[0045] A layered sampling method is used to sample N points at the intersection of the X-ray and the object. Specifically, the X-ray is divided into N uniform intervals, and one point is sampled within each interval. N is set to be larger than the required CT data size to ensure that at least one sampling point is allocated to each voxel through which the X-ray passes.
[0046] S3: Multi-scale hash coding is used to map the location information of the sampling points to a high-dimensional space to obtain the feature vector of the sampling points.
[0047] For the CT data to be reconstructed, consider that it has L different scales, and each scale is associated with a hash table with T entries, each entry being a feature vector of dimension F.
[0048] Each scale corresponds to a different resolution N. l , l∈[1,L]; the resolution range is [N min N max The formula for calculating the resolution of each layer is:
[0049]
[0050] Where b is a smaller growth factor, its calculation formula is:
[0051]
[0052] The resolution determines which voxel the sampling point belongs to.
[0053] The sampling point is mapped to all grid points of the voxel at each scale into a hash table, obtaining the feature vector of each grid point at each scale. Based on the feature vector of each grid point, linear interpolation is used to obtain the initial feature vector of the sampling point at each scale. The hash function h used in this invention, i.e., the formula for mapping grid points to the hash table, is:
[0054]
[0055] Where h(X) represents the eigenvector of grid point X, and x1, x3, and x4 represent the x, y, and y coordinates of grid point X, respectively. The expression represents an element-wise XOR operation, where π1, π2, and π3 represent the first, second, and third largest prime numbers, and T represents the number of entries in the hash table. At coarser scales, the total number of grid points in the CT data is less than the number of hash table entries T, resulting in a 1:1 mapping. At finer scales, hash collisions may occur. Since a collision may occur at scale l, but is unlikely to occur at all scales, hash collisions are not addressed.
[0056] For example, such as Figure 3 As shown, for sampling point P∈R 3 At scale l, it is located at voxel V l In this model, the voxel has 8 grid points. Each grid point is mapped to a hash table to obtain its respective feature vector {c1, c2, ..., c8}. Linear interpolation is then used to obtain the feature vector f of the sampling point P at scale l. l (The distance from the sampling point to the grid point is used as the weight, and a weighted sum is made with the feature vector of the grid point).
[0057] By concatenating the initial feature vectors of the sampling points at all scales, the final feature vector of the sampling points is obtained.
[0058] S4: Input the feature vector of the sampling point into the multilayer perceptron for processing to obtain the attenuation coefficient of the sampling point.
[0059] For a sampling point P, its hash-encoded feature is f. The attenuation coefficient σ is predicted using a multilayer perceptron (MLP), expressed as:
[0060] σ=MLP(f)
[0061] S5: Calculate the total loss based on the attenuation coefficient of the sampling points, and adjust the multilayer perceptron parameters according to the total loss to obtain the trained multilayer perceptron.
[0062] The attenuation coefficients at all sampling points along the composite ray are used to obtain the composite projection of the ray. Specifically, according to Beer's Law, the attenuation of X-rays penetrating matter is the exponential integral of the attenuation coefficients along its propagation path, and its discrete form is expressed as:
[0063]
[0064] Where I represents the composite projection of the ray, I0 represents the initial intensity of the ray, and σ i δ represents the attenuation coefficient at the i-th sampling point, and N represents the number of sampling points on the ray; i δ represents the distance between the i-th sampling point and the (i+1)-th sampling point. i =‖P i+1 -Pi ‖.
[0065] The total loss is calculated based on the synthetic projection and the true projection (CT projection data obtained by the CBCT scanner from scanning physical objects); specifically, the model is optimized by minimizing the L2 loss between the true projection and the synthetic projection. The total model loss is:
[0066]
[0067] Where Loss represents the total loss, i r (r) represents the true projection of ray r, I s (r) represents the composite projection of the true projection, r represents ray r, and R represents the set of rays.
[0068] S6: The trained multilayer perceptron is used to process the physical object to be reconstructed by CT to obtain the CT reconstruction result of the physical object.
[0069] The final reconstructed CT data should be a discrete 3D matrix. Therefore, a mesh with the required resolution is constructed; a trained multilayer perceptron is used to process the object to be reconstructed from the CT image, and the attenuation coefficients of the predicted sampling points are output; the mesh coordinates are then passed to the attenuation coefficients to obtain the reconstructed CT data.
[0070] In summary, this invention represents CT data as a mapping from spatial coordinates to attenuation coefficients and incorporates a self-supervised network framework to support the reconstruction from two-dimensional projection to three-dimensional CT data. Multi-scale hashing encoding is employed to map the neural network input to a high-dimensional space to better fit the frequently changing data, thereby ensuring the accuracy of the reconstructed CT data. CT data reconstruction using this invention achieves a good balance between speed and accuracy, and the use of CBCT projection helps reduce the likelihood of cancer in patients during medical examinations.
[0071] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A cone-beam computed tomography (CBCT) reconstruction method based on multi-scale hash coding, characterized in that, include: S1: Scan the physical object and acquire CT projection data; S2: Generate X-rays from the X-ray source to the direction of the projection pixels based on the CT projection data, and uniformly sample N points at the intersection of the X-rays and the object; S3: Multi-scale hash coding is used to map the location information of the sampling points to a high-dimensional space to obtain the feature vector of the sampling points; S4: Input the feature vector of the sampling point into the multilayer perceptron for processing to obtain the attenuation coefficient of the sampling point; S5: Calculate the total loss based on the attenuation coefficient of the sampling points, and adjust the multilayer perceptron parameters based on the total loss to obtain the trained multilayer perceptron; S6: The trained multilayer perceptron is used to process the physical object to be reconstructed by CT to obtain the CT reconstruction result of the physical object.
2. The cone-beam CT reconstruction method based on multi-scale hash coding according to claim 1, characterized in that, The generated ray is represented as: ； in, Indicates the distance from the X-ray source to the projected pixel. Rays in the direction, Indicates the generation of projection The location of the X-ray source This indicates the distance from the sampling point to the radiation source. Represents rays The direction vector.
3. The cone-beam CT reconstruction method based on multi-scale hash coding according to claim 1, characterized in that, The process of mapping the location information of sampling points to a high-dimensional space includes: Map all grid points of the voxel at each scale of the sampling point to a hash table to obtain the feature vector of each grid point at each scale. Based on the feature vectors of each grid point, linear interpolation is used to obtain the initial feature vectors of the sampling points at each scale. By concatenating the initial feature vectors of the sampling points at all scales, the final feature vector of the sampling points is obtained.
4. The cone-beam CT reconstruction method based on multi-scale hash coding according to claim 3, characterized in that, The formula for mapping grid points to a hash table is: ； in, The eigenvectors of lattice point X are represented. , and These represent the x, y, and vertices of the grid point X, respectively. This represents the element-wise XOR operation. , , Representing the first, second, and third largest prime numbers, Indicates the number of entries in the hash table.
5. The cone-beam CT reconstruction method based on multi-scale hash coding according to claim 1, characterized in that, The process of calculating the total loss based on the attenuation coefficient of the sampling points includes: The attenuation coefficients of all sampling points on the composite ray are used to obtain the composite projection of the ray; The total loss is calculated based on the composite projection and the true projection of the rays.
6. The cone-beam CT reconstruction method based on multi-scale hash coding according to claim 5, characterized in that, The formula for the attenuation coefficient of all sampling points on the synthetic ray is: ； in, Represents the composite projection of rays. Indicates the initial intensity of the ray. This represents the attenuation coefficient at the i-th sampling point. Indicates the number of sampling points on the ray. This represents the distance between the i-th sampling point and the (i+1)-th sampling point.
7. The cone-beam CT reconstruction method based on multi-scale hash coding according to claim 5, characterized in that, The formula for calculating the total loss is: ； in, Indicates the total loss. This represents the true projection of ray r. A composite projection representing a true projection. Represents rays , Represents a set of rays.