Binary distribution tomographic image reconstruction method suitable for small-angle projection

By establishing a model from multi-projection attenuation rate to binary matrix in X-ray imaging technology, combined with the joint algebraic reconstruction algorithm with Sigmoid function constraints, the problem of poor image reconstruction quality under small angle projection is solved, and efficient images are achieved Reconstruction, improving computing speed and quality.

CN118918210APending Publication Date: 2024-11-08BEIHANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411173152.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

With few angle projections, it is difficult for X-ray imaging technology to effectively reconstruct high-quality images. The existing algorithms have poor reconstruction results under noise and artifact interference, and have low computational efficiency and high algorithm complexity, making it difficult to fully implement clinical applications. Get value.

Method used

Using a binary distribution tomographic image reconstruction method suitable for less angle projection, an X-ray imaging system using a ray source as a fan beam is established by establishing a model from multiple projection attenuation rate to binary matrix, combined with a joint algebraic reconstruction algorithm with Sigmoid function constraints , measure and reconstruct the attenuation coefficient distribution in the region to improve the image reconstruction quality.

Benefits of technology

Effectively utilize projection attenuation rate information, the algorithm convergence speed and calculation speed are improved, the reconstruction quality of X-ray imaging images is improved with a small angle, and the reconstruction error is reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918210A_ABST
    Figure CN118918210A_ABST
Patent Text Reader

Abstract

The invention provides a binary distribution cross-sectional image reconstruction method suitable for small-angle projection, which mainly comprises the following steps of: firstly, irradiating a two-dimensional reconstruction region by using a fan-shaped beam light source, carrying out grid division on the reconstruction region, calculating a sensitivity matrix, calculating a projection attenuation rate through a projection value obtained by measurement, and calculating a binary distribution cross-sectional image; constructing an attenuation coefficient matrix according to the material composition of the reconstruction region; then, representing attenuation coefficient distribution of a reconstruction area by using a binary matrix, establishing a linear imaging model from a multi-optical-path projection attenuation rate to attenuation coefficient distribution, and adding Sigmoid function constraint in an iterative algorithm in combination with binary matrix characteristics; and finally, according to the solved binary matrix and a set attenuation coefficient matrix, performing mapping to obtain attenuation coefficient distribution of the reconstruction region, thereby obtaining substance distribution. According to the method, the calculation cost can be reduced, the reconstruction error of small-angle projection is reduced, the image reconstruction quality is remarkably improved, and the method has important application prospects in the field of tomography.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a binary distribution tomographic image reconstruction method suitable for projections with few angles, and belongs to the field of X-ray imaging. Projection information is fully utilized to establish a material distribution model from multi-projection attenuation rate to binary matrix and then to a two-dimensional field to be measured. In the iterative solution process, the joint algebraic reconstruction algorithm is integrated with the Sigmoid function constraint, which can effectively reduce the reconstruction error and improve the image reconstruction quality in the case of few angles. Background Art

[0002] The advent of X-ray Computer Tomography (XCT) technology is another epoch-making breakthrough in the engineering field after the discovery of radiation. XCT performs multi-angle radiation scanning on the object to be measured from the outside, thereby accurately and intuitively reconstructing the distribution of the internal tissue of the object's cross section, greatly enhancing the ability of humans to observe the internal structure of the object. This technology is widely used in medical diagnosis, industrial testing, safety inspection, aerospace, national defense and military industries due to its non-invasiveness, fast measurement speed and high measurement accuracy.

[0003] The two important components of the CT detection system are the hardware system and the software algorithm. Among them, the CT image reconstruction algorithm is the core of the software algorithm and has a great influence on the reconstruction results. Therefore, when the performance of the hardware system is fixed, the main way to improve the CT reconstruction effect is to optimize and improve the reconstruction algorithm. Currently, the commonly used reconstruction algorithms include analytical algorithms (AR) based on Radon inverse transform and filtered back projection (FBP) algorithms and iterative algorithms (IR) based on solving linear equations.

[0004] The FBP algorithm is the representative of analytical reconstruction algorithms. The FBP algorithm is based on the Randon transform. It removes noise and improves image quality by filtering the projection data, and then performs a back-projection operation. The back-projection results are superimposed to obtain a reconstructed CT image. It has the advantages of simple algorithm and fast calculation speed, and has been widely used in commercial CT. In 2008, Shtok J et al. published "Adaptive Filtered-back-projection for Computed Tomography" at the 25th IEEE Conference of Electrical and Electronics Engineers in Israel. They proposed an adaptive, spatially varying linear filter based on a two-dimensional convolution kernel, which replaced the window-shaped ramp kernel filter in the standard FBP algorithm. In addition, the scheme also includes a post-processing step, that is, using a two-dimensional adaptive filter to perform convolution filtering on the image. By optimizing the corresponding objective function, the two filters of pre-processing and post-processing are trained to perform specific reconstruction tasks. The resulting adaptive scheme can compensate for discrete errors, specifically for a given task, and effectively improves the quality of reconstruction. In 2014, Pelt, Batenburg et al. published an article titled "Improving Filtered Backprojection Reconstruction by Data-Dependent Filtering" in IEEE Transactions On Image Processing, Vol. 23, No. 11, pp. 4750-4762. They used a custom data-dependent filter to improve the filtered backprojection method, thereby minimizing the reconstruction error. The data-dependent filter is calculated by minimizing the difference between the reconstructed image and the measured projection data, and the amount of calculation is reduced by assuming that the filter is a piecewise constant function. The final reconstruction accuracy of the proposed method is similar to that of the algebraic method, but the computational effort is smaller, the generation speed is faster, and it can be extended to utilize certain forms of prior knowledge to improve the reconstruction accuracy in specific cases.In 2016, Wang ZZ et al. published a paper titled "Studies on Filtered Back-projection Imaging Reconstruction Based on a Modified Wavelet Threshold Function" at the Optoelectronic Imaging And Multimedia Technology IV conference. The paper applied the wavelet threshold denoising method to the filtered back-projection algorithm for image reconstruction. Aiming at the pseudo-Gibbs oscillation of the traditional soft and hard threshold functions, an improved wavelet threshold function with two thresholds and two variables was proposed. Different denoised reconstructed images were obtained by combining different threshold functions with the FBP algorithm. The experimental results show that the filtered back-projection reconstruction algorithm based on the improved wavelet threshold function has better denoising and reconstruction effect. In 2022, Jiang YM et al. published "A Generalized Construction Model for CT Projection-Wise Filters on the SDBP Technique" in Mathematics, Vol. 10, No. 4, supplementing the relevant knowledge of the method of constructing CT projection filters. Starting from the mathematical principles of CT reconstruction, the second-order differential back projection (SDBP) technology was proposed for the first time, and a new filter expression calculation model was proposed. The model reveals the correlation between the convolution kernel and the filter expression used for data recovery, and stipulates the construction principle of the filter, thereby realizing accurate and efficient Radon inverse transform. The proposed method has certain reference value for filter optimization research in the field of CT reconstruction.In 2023, Xia YX et al. published "Generalized-Equiangular Geometry CT: Concept and Shift-invariant FBP Algorithms" in Medical Physics, Vol. 50, No. 8, pp. 5150-5165. For generalized equiangular geometry CT (GEGCT), the normalized radial-offset distance (NROD) was used to characterize the relative position between the X-ray source and the detector focus, and a set of shift-invariant FBP algorithms with multiple weighting strategies were proposed. Three feasible weighting strategies were proposed, including the classic strategy in Besson's literature, a new strategy based on curvature fitting, and a strategy based on empirical formulas. Theoretical analysis and numerical studies have shown that the weights in the shift-invariant FBP algorithm can ensure high-precision reconstruction of GEGCT, and are robust and flexible. As a traditional CT reconstruction method, the analytical method has the advantage of high computational efficiency, but it requires high-quality projection data. The completeness and signal-to-noise ratio of the projection data directly affect the reconstruction results. In the case of incomplete data and high noise, analytical reconstruction is difficult to achieve good imaging effects, and the reconstructed image is often severely interfered by noise and artifacts.

[0005] Iterative reconstruction algorithms can effectively overcome the problems inherent in analytical methods. Iterative reconstruction methods construct models based on the physical model of the imaging system and the characteristics of the detection data, and then use iterative algorithms to iteratively solve the objective function. The results are better than traditional analytical methods. Compared with analytical reconstruction algorithms, iterative reconstruction algorithms can be used to solve the reconstruction problem of incomplete projection data. In recent years, thanks to the rapid progress of computer technology, iterative reconstruction algorithms have become a hot topic of research. Iterative reconstruction algorithms construct a system of linear equations and then use algebraic or statistical methods to iteratively solve them; however, statistical iterative algorithms usually require a large amount of data to estimate parameters, and are more sensitive to initial values ​​and outliers. Their convergence may be limited by data quality, model assumptions, and the algorithm itself; while the convergence of algebraic iterative algorithms can usually be guaranteed by mathematical proofs, and the model complexity is low and the calculation speed is faster.

[0006] The algebraic reconstruction algorithm uses the ratio (multiplication correction) or difference (addition correction) of the measured projection and the estimated projection to correct the solution of the equations, thereby updating the pixel value. Representative algorithms include the Algebraic Reconstruction Technique (ART) and the Simultaneous Algebraic Reconstruction Technique (SART). In 2007, Hou X et al. published the paper "Modified Simultaneous Algebraic Reconstruction Technique and Its Application to Image Reconstruction" at the 27th International Congress on High Speed ​​Photography and Photonics. In view of some shortcomings of the traditional SART when processing data with fewer views or limited field of view, an improved SART algorithm, MSART, was proposed. In the error correction process, the algorithm not only considers the pixel value, but also the segment length of the intersecting rays. Computer simulations have been used to compare SART and MSART. The results show that MSART effectively solves the edge distortion problem and greatly improves the reconstruction quality under sparse data. In 2011, Batenburg KJ et al. published an article entitled "DART: A Practical Reconstruction Algorithm for Discrete Tomography" in IEEE Transactions on Image Processing, Vol. 20, No. 9, pp. 2542-2553. They proposed a discrete algebraic reconstruction technique (DART). When it is known that the scanned object consists of only a few different components, DART can be applied to divide the pixels into fixed pixels and free pixels, alternately perform continuous update steps and discretization update steps, and use the prior knowledge of grayscale values ​​to turn the current reconstruction to a reconstruction containing only these grayscale values. Experiments based on CT data show that DART can calculate more accurate reconstructions from a small number of projection images or within a small angle range, can effectively process projection data with noise, and has strong robustness.In 2016, Tateishi K et al. published a paper titled "Continuous Analog of Multiplicative Algebraic Reconstruction Technique for Computed Tomography" at the International Conference on Physics of Medical Imaging (Medical Imaging 2016: Physics of Medical Imaging). They proposed a continuous-time image reconstruction (CIR) method as a continuous analog of the continuous block-iterative multiplicative algebraic reconstruction technique (BI-MART). The first-order expansion of the difference equation was performed using geometric multiplication calculus, and the same iterative formula as BI-MART was obtained. Experiments have shown that this method can provide good image quality in binary tomography even with a small number of projections. In the actual reconstruction process, in order to further improve the reconstruction quality of the traditional iterative reconstruction algorithm, a regularization term (also called a penalty term) is often added to the iterative algorithm to utilize some prior information about the reconstructed image. Commonly used regularization methods include: minimizing the total variation (TV) of the image, dictionary learning, etc.

[0007] Total Variation (TV) minimization is a denoising method based on image gradient information. Its core idea is to achieve denoising by minimizing the gradient change of the image. It is particularly suitable for retaining edge information in the image while suppressing noise. In 2008, Sidky EY et al. published an article entitled "Image Reconstruction in Circular Cone-beam Computed Tomography by Constrained, Total-variation Minimization" in Physics In Medicine And Biology, Vol. 53, No. 17, pp. 4777-4807. They proposed an iterative algorithm for cone-beam scanning volume image reconstruction. The algorithm minimizes the total variation (TV) of the image under the constraints that the estimated projection data is within the specified tolerance range of the available data and the value of the volume image is non-negative. The algorithm uses projection on convex sets (POCS) to implement constraints and uses adaptive step-size steepest descent to achieve the goal of minimizing the total variation. It is called adaptive steepest descent projection on convex sets (ASD-POCS). The algorithm is simple to implement and has few algorithm parameters, which makes it easy to analyze its image reconstruction performance. The feasibility of the algorithm is verified by experiments and can be applied to scanning geometries involving other X-ray source trajectories. In 2014, Liu Y et al. published an article entitled "Total Variation-Stokes Strategy for Sparse-View X-ray CT Image Reconstruction" in IEEE Transactions On Computational Imaging, Vol. 33, No. 3, pp. 749-763. In view of the shortcomings of traditional CT image reconstruction methods based on total variation (TV) when processing sparse view data, the reconstructed images often have blocky or patchy artifacts. They proposed an image reconstruction method based on total variation-Stokes-convex set projection, called TVS-POCS (Total Variation-Stokes Projection onto Convex Sets) algorithm. The TVS model recovers the information that may be missing in the case of sparse view data by introducing isoradial directions. Therefore, the required consistency along the normal and tangent directions is retained in the resulting image. Digital projection, physical projection and clinical data experiments were used to verify the effectiveness of the algorithm. Compared with the corresponding total variation-based algorithm, the algorithm has significant improvements in image quality, resolution and contrast.In 2016, Zhuge XD et al. published an article titled "TVR-DART: A More Robust Algorithm for Discrete Tomography From Limited Projection Data With Automated Gray Value Estimation" in IEEE Transactions On Image Processing, Vol. 25, No. 1, pp. 455-468. They proposed a total variation regularized discrete algebraic reconstruction technique based on automatic gray value estimation. By simultaneously utilizing two prior knowledge of the scanned object, the discrete reconstruction problem is solved within a compressed sensing-inspired optimization framework, directing the current reconstruction to a solution with a specified number of discrete gray values. As the iterative reconstruction improves, the gray value and threshold are estimated. From the experimental results, under noisy conditions, TVR-DART can provide more accurate reconstruction results from a small number of projection images or within a small angle range. In 2024, Diwakar M et al. published an article titled "Edge-guided Filtering Based CT Image Denoising Using Fractional Order Total Variation" in Volume 92 of Biomedical Signal Processing And Control. In order to solve the noise problem in CT reconstructed images, a new weighting function was introduced to denoise CT images using fractional order total variation, which solved the blocking effect problem and the non-convex optimization problem, thereby obtaining a better solution. The Split Bregman algorithm and Augmented Lagrangian algorithm were used in the study to implement the proposed weighted fractional total variation denoising. Experimental results show that the proposed method is superior to existing methods in denoising effect, image detail preservation and computational efficiency. Although TV makes a strong assumption on gradient sparsity through non-smoothed absolute values ​​that are more suitable for piecewise constant images, TV often causes artifacts such as blurred details and patchy textures in the final reconstruction.

[0008] Dictionary learning is another commonly used regularization method. Its principle is mainly based on sparse representation theory. It aims to find a dictionary matrix and a coefficient matrix so that the product of the dictionary matrix and the coefficient matrix can restore the original data as much as possible, and the coefficient matrix is ​​as sparse as possible. Compared with the discrete gradient transform used in the TV method, dictionary learning is an effective sparse representation method that can highlight the most important part of the signal and ignore unimportant or redundant information, which helps to improve the efficiency of data processing. In 2012, Xu Q et al. published an article entitled "Low-Dose X-ray CT Reconstruction via Dictionary Learning" in IEEE Transactions On Medical Imaging, Vol. 31, No. 9, pp. 1682-1697. They introduced dictionary learning constraints into the iterative reconstruction framework, constructed two reconstruction objective functions based on global dictionary learning and adaptive dictionary learning, and greatly improved the reconstruction effect. Through experiments, the proposed method was compared with filtered back projection and total variation-based reconstruction methods. The experimental results show that the proposed method can produce images with lower noise and more detailed structural features in selected cases. In 2019, Bao P et al. published an article entitled "Convolutional Sparse Coding for Compressed Sensing CT Reconstruction" in IEEE Transactions On Medical Imaging, Vol. 38, No. 11, pp. 2607-2619. They first used the convolutional sparse coding framework in CT reconstruction and proposed a convolutional sparse coding constrained reconstruction method. This method directly processes the entire image and does not need to divide the image into overlapping small blocks like dictionary learning-based methods. It can retain more details and avoid artifacts caused by small block aggregation, thus achieving noise and artifact suppression in sparse angle CT reconstruction.In 2021, Duan JY et al. published an article titled "Image Quality Guided Iterative Reconstruction for Low-dose CT Based on CT Image Statistics" in Physics In Medicine And Biology, Vol. 66, No. 18. By introducing CT image quality assessment (IQA) into the iterative reconstruction framework, the problem that the selection of regularization parameters is based on empirical rules or assumptions that require some prior knowledge is solved. A quality-guided iterative reconstruction algorithm is established, which can achieve adaptive hyperparameter adjustment and improve imaging quality. Experiments have proved the effectiveness of this method. Compared with other strategies, this method has the advantages of not requiring prior information. Iterative reconstruction algorithms based on total variation and dictionary learning can be used in combination under certain conditions to improve the performance of image restoration and denoising, but due to the low computational efficiency, high algorithm complexity, long processing time and poor stability caused by data incompatibility of such algorithms, it is difficult to give full play to their value in clinical application scenarios.

[0009] In order to solve the above-mentioned problems of long calculation time and high algorithm complexity, deep learning (DL) technology can be combined with CT image reconstruction. Deep learning methods have better performance and faster processing speed. The DL model can demonstrate its ability to learn uncertain noise distribution by capturing high-level features of the image, so it can effectively adapt to any noise type. According to the way DL is used in iterative reconstruction algorithms, it can be divided into model-based methods and iterative expansion-based methods. The core of the prior constraint model-based method is to use deep neural networks to replace the prior constraints in iterative reconstruction. In 2021, Ye SQ et al. published an article entitled "Unified Supervised-Unsupervised (SUPER) Learning for X-Ray CT Image Reconstruction" in IEEE Transactions On Medical Imaging, Vol. 40, No. 11, pp. 2986-3001, proposing a unified supervised-unsupervised (SUPER) learning reconstruction framework. Traditional model-based image reconstruction (MBIR) methods have limitations when dealing with incomplete measurement data. The SUPER learning framework combines the advantages of supervised learning and unsupervised learning, combining deep network priors with prior-based regularization terms in a unified MBIR framework through fixed-point iterative analysis. Numerical and visual results show that the proposed unified SUPER method outperforms independent supervised learning methods and iterative MBIR methods, while also demonstrating the fast convergence of the proposed algorithm. The iterative expansion class method is another deep reconstruction method, and its core idea is to use deep learning to achieve iterative solutions. In 2022, Su T et al. published an article entitled "Generalized Deep Iterative Reconstruction for Sparse-view CT Imaging" in Physics in Medicine and Biology, Volume 67, Issue 2, exploring a generalized expansion scheme for a model-driven deep learning CT image reconstruction method to further improve its performance in sparse-view CT imaging. In this scheme, iteration parameters, regularization terms, data fidelity terms and even mathematical operations are assumed to be learned and optimized through network training. Three deep iterative networks DIR-I, DIR-II and DIR-III are proposed, which learn the parameters and operations in the iteration process to different degrees. Numerical and experimental results of sparse view CT show that the network has the greatest generalization ability and excellent reconstruction performance.Although CT reconstruction algorithms based on deep learning have shown superior performance in some aspects, in order to obtain better reconstruction results, such algorithms often require a large amount of training data, have high computational costs, and are highly data-dependent. At the same time, the algorithm has limited stability and generalization capabilities, and artifacts and noise may still exist when processing sparse views, limited angle measurements, and noise interference.

[0010] Based on the above background, the present invention proposes a binary distribution tomographic image reconstruction method suitable for small-angle projection. According to the material composition in the two-dimensional reconstruction area, the corresponding X-ray linear attenuation coefficient is selected, and the projection value at different projection angles is obtained by measurement using an X-ray imaging system with a fan-shaped beam as the ray source; then, according to the sensor optical path layout, a binary matrix is ​​introduced to represent the distribution position of the attenuation coefficient in the reconstruction area, and the constraint of the attenuation distance is integrated to establish a solution model for the projection attenuation rate on each ray to the binary matrix. Combined with the characteristics of the binary matrix itself, an iterative algorithm with a Sigmoid function is used to solve the binary matrix; through binary matrix mapping, the attenuation coefficient distribution of the reconstruction area, that is, the material distribution, is finally obtained. The advantage of this method is that it effectively utilizes the information of the projection attenuation rate, combines the characteristics of the reconstruction matrix, adds the Sigmoid function constraint in the iterative algorithm, accelerates the convergence of the algorithm, improves the calculation speed, and improves the quality of X-ray imaging image reconstruction in the case of small angles. Summary of the invention

[0011] 1. Technical issues to be resolved

[0012] The purpose of the present invention is to propose a binary distribution tomographic image reconstruction method suitable for small-angle projection. The method integrates prior information, makes full use of projection information, and adds Sigmoid function constraints in the iterative algorithm in combination with the reconstruction matrix characteristics. The reconstruction accuracy is high, and the purpose is to improve the reconstruction quality of X-ray imaging in small-angle conditions.

[0013] (II) Technical solution

[0014] The present invention is a binary distribution tomographic image reconstruction method suitable for small-angle projection, which mainly comprises the following steps:

[0015] Step 1: Obtain the integrated attenuation rate projections on multiple X-ray paths through an X-ray reconstruction system; the reconstruction system includes an X-ray source, a stage for the object to be measured, and a detector array, wherein s X-ray sources are evenly arranged on a circle centered on the object to be measured, each ray source corresponds to a detector array, and each detector array consists of d detectors arranged in a row; the X-ray source is located on one side of the object to be measured, and the corresponding detector array is located on the other side. The rays of each X-ray source are fan-shaped beams, covering only the detector array on the opposite side, and the sensor arranges a total of Q = s × d rays; the detector is used to obtain the X-ray projection values ​​when there is no object to be measured and when there is an object to be measured, which are used as the incident and outgoing energies of the X-rays, respectively. According to the Lambert-Beer law, the attenuation rate p is expressed as

[0016]

[0017] Where I0 is the incident intensity of the X-ray, I is the intensity of the X-ray after attenuation, L is the distance the X-ray passes through the object, and u is the linear attenuation coefficient of the object;

[0018] When an X-ray passes through a non-uniform field to be measured, the attenuation rate projection on the X-ray path is the superposition of the attenuation rates of different substances. If the field to be measured consists of only two substances, and the linear attenuation coefficients of each substance are calculated from the database, the attenuation rate p is expressed as

[0019]

[0020] Among them, u m is the attenuation coefficient of the mth substance per unit path length, Δx m is the path length occupied by the mth material distribution. Formula (2) can be written in matrix form, which is the single light path material distribution histogram model.

[0021] p=X 1×2 ·U 2×1 (3)

[0022] Where p is the total attenuation rate of the optical path, U 2×1 The rows represent the linear attenuation coefficients of different materials, X 1×2 The columns of represent the path lengths occupied by different substances, which is defined as a substance distribution histogram;

[0023] For two-dimensional imaging, a multi-angle X-ray optical path covering the reconstruction area is arranged, and the attenuation rate projection along multiple X-ray paths is measured. If the total number of X-ray paths is Q, the attenuation rate projection along Q paths constitutes the attenuation vector p Q×1 According to formula (3), the material distribution histogram X of Q X-ray paths is Q×2 satisfy

[0024] pQ×1 =X Q×2 ·U 2×1 (4)

[0025] Step 2: Establish a binary matrix solution model for material distribution; discretize the reconstruction area into D×D grids, and consider the material composition in each grid to be unique, that is, the attenuation coefficient in each grid is a constant. According to the sensor optical path layout and grid division, calculate the path length of each X-ray passing through all grids, which is recorded as the sensitivity matrix W Q×DD , sensitivity matrix W Q×DD The element w in the i-th row and j-th column of i,j represents the optical path length of the i-th X-ray passing through the j-th grid, and thus the total optical path length constraint satisfied by each optical path is obtained, that is, the sum of the histogram values ​​along each X-ray path is equal to the sum of the optical path lengths of the optical path passing through all grids, expressed as

[0026] X Q×2 ·e 2×1 =W Q×DD ·e DD×1 (5)

[0027] Among them, e DD×1 and e 2×1 is a column vector whose elements are all 1, so equations (4) and (5) form a new set of equations with distance constraints:

[0028] X Q×2 ·[U 2×1 e 2×1 ]=[p Q×1 W Q×DD ·e DD×1 ] (6)

[0029] In each grid of the reconstruction area, the material attenuation coefficient is one of the two specific attenuation coefficients, which can be represented by a binary matrix Y with elements 0 and 1. DD×2 To represent the attenuation coefficient distribution, each column of the matrix represents a corresponding attenuation coefficient value, and each row represents a grid in the reconstruction area;

[0030] The distribution of each material can be regarded as an object with a uniform attenuation coefficient. The attenuation distance occupied by the material on each X-ray optical path is the projection of the X-ray passing through the uniform object. Therefore, the sensitivity matrix W can be used to calculate the attenuation distance of the material on each X-ray optical path. Q×DD Establish the relationship between histogram information and binary distribution

[0031] X Q×2 =W Q×DD ·Y DD×2 (7)

[0032] Combining equation (6) and equation (7) we can form the projection attenuation rate p on each X-ray beam: Q×1 To the binary matrix Y DD×2 The new set of equations

[0033] W Q×DD ·Y DD×2 ·[U 2×1 e 2×1 ]=[p Q×1 W Q×DD ·e DD×1 ] (8)

[0034] in, Since u1≠u2, the matrix is a full rank matrix, equation (8) can be written as

[0035]

[0036] That is, we get the binary matrix Y DD×2 The final solution model;

[0037] Step 3: Set the maximum number of iterations of the outer loop L m And the maximum number of iterations K of the inner loop m , combined with the binary matrix Y DD×2 The characteristics of are solved; first, for the solution equation of formula (9), the binary matrix Y is reconstructed using the joint algebraic reconstruction algorithm fused with the Sigmoid function constraint DD×2 Solve each column in turn, and in (k+1) iterations, get the binary matrix Y DD×2 The jth element of a column The calculation formula is

[0038]

[0039] Among them, w ij is the sensitivity matrix W Q×DD The element in row i and column j, w i is the i-th row of the sensitivity matrix, x i For the matrix The i-th element of the column, y k is the binary vector of the kth iteration; when the matrix Y DD×2 After all elements in this column are solved by equation (10), the column is processed by Sigmoid function to find the maximum value y in the column j,max With the minimum value y j,min , each element in the column is processed as follows

[0040]

[0041] Among them, a is the parameter of the Sigmoid function, which is used to adjust the degree of binarization approximation; for the binary matrix Y DD×2 This column loops through the calculations of equations (10) and (11) until the number of iterations meets the maximum number of iterations K of the inner loop. m ;

[0042] For binary matrix Perform the above operation column by column to obtain the binary matrix in the lth iteration.

[0043] Step 4: Use the Sigmoid function to transform the binary matrix Y DD×2 The rows of the matrix Y are constrained; DD×2 Each row represents a grid, and each grid has a unique attenuation coefficient, that is, the matrix Y DD×2 Each row has only one 1 value, so the binary matrix obtained in step 3 needs to be Each row is processed by Sigmoid function

[0044]

[0045] Among them, y i is a binary matrix The i-th element of a row, y i,max With y i,min are the maximum and minimum values ​​in the row respectively. After the calculation of formula (12), we can get the binary matrix after the lth iteration: The i-th element of the row For the binary matrix obtained in step 3 The elements of each row of are operated as above in turn to obtain the binary matrix after the lth iteration The updated value of

[0046] If the current number of iterations is l <L m , then repeat steps 3 and 4; when the maximum number of iterations of the outer loop L is met m After that, we can get the binary matrix Y DD×2 The calculation results of

[0047] Step 5: Calculate the material distribution in the reconstruction area; obtain the binary matrix Y through iterative calculation DD×2 After that, the attenuation coefficient Z in all grids DD×1 It can be calculated by the following formula

[0048] Z DD×1 =Y DD×2 ·U 2×1 (13)

[0049] Finally, the material distribution in the reconstructed area can be obtained through the attenuation coefficient distribution.

[0050] (III) Beneficial effects

[0051] The beneficial effect of the present invention is that a binary distribution tomographic image reconstruction method suitable for small-angle projection is proposed, a binary matrix is ​​used to represent the material distribution of the field to be measured, the material distribution solution is converted into the solution of the binary matrix, and a solution model is established using a histogram. According to the X-ray multi-projection attenuation rate information and the binary matrix characteristics, an iterative algorithm integrating Sigmoid function constraints is used to solve the binary matrix. According to the set linear attenuation coefficient composition, the attenuation coefficient is mapped one-to-one with the grid, thereby obtaining the attenuation coefficient distribution of the field to be measured, that is, the material distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Attached Figure 1 :The specific implementation diagram of this method

[0053] Attached Figure 2 : Schematic diagram of the reconstruction system light source (101), the field to be measured (102) and the detection array (103)

[0054] Attached Figure 3 :The two-dimensional attenuation coefficient distribution of the field to be measured under the 120×120 grid division obtained by simulation is used as the simulation reconstruction object Figure 4 : Attenuation coefficient distribution obtained by simulation reconstruction under 50×50 grid division DETAILED DESCRIPTION

[0055] Reference Figure 1 This is a specific implementation diagram of this method. Figure 2 Attached is a schematic diagram of the reconstruction system. Figure 3 is the original attenuation coefficient distribution, Figure 4 To reconstruct the attenuation coefficient distribution, the specific steps combined with an example are as follows:

[0056] Step 1: Theoretically calculate the projection attenuation rate along all light rays under the current sensor optical path layout. Figure 2 The figure shows a schematic diagram of the field to be measured, an X-ray source and the corresponding detection array. Eight groups of X-ray sources and detection arrays are arranged at equal angles around the field to be measured. The X-ray source emits fan-shaped beams that irradiate the detection array on the opposite side. 150 detection pixels on the detection array are selected, and the entire reconstruction system arranges 1200 rays in total. When calculating the projection attenuation rate, the two-dimensional field to be measured is evenly divided into 120×120 grids. It is assumed that there is a uniform attenuation coefficient distribution in each grid. The attenuation coefficient distribution of the original model is shown in the attached figure. Figure 3 As shown, the attenuation coefficient of the white part is 1, and the attenuation coefficient of the black part is 0.

[0057] The number of grids inside the reconstruction area is 11304. Based on the 120×120 grid division, the sensitivity matrix W of 1200 X-ray rays passing through 11304 grids is calculated. 1200×11304 According to the original distribution, the attenuation coefficient matrix C corresponding to 11304 grids is obtained 11304×1 , thus obtaining the projection attenuation rate along all rays

[0058] p 1200×1 =W 1200×11304 ·C 11304×1 (1)

[0059] Set the attenuation coefficient value during reconstruction. The attenuation coefficient matrix under the material unit path is:

[0060]

[0061] Step 2: Construct a binary matrix solution model. Further establish the material distribution histogram X of all X-ray paths 1200×2 With the projection attenuation rate p 1200×1 The linear equations between

[0062] p 1200×1 =X 1200×2 ·U 2×1 (3)

[0063] The two-dimensional reconstruction area is evenly divided into 50×50 grids, where the number of grid points inside the sensor is 1976. The sensitivity matrix W′ of 1200 X-rays passing through 1976 grids under the 50×50 grid division is calculated. 1200×1976 , from which we can get the total length constraint of each light path that each light path satisfies, that is, the sum of the histogram values ​​along each X-ray path is equal to the sum of the lengths of the light path passing through all the grids, expressed as

[0064] X 1200×2 ·e 2×1 =W' 1200×1976 ·e 1976×1 (4)

[0065] Among them, e 1976×1 and e 2×1 is a column vector whose elements are all 1. Therefore, equations (3) and (4) form a new system of equations:

[0066] X 1200×2 ·[U 2×1 e 2×1 ]=[p 1200×1 W1' 200×1976 ·e 1976×1 ] (5)

[0067] In each grid of the reconstruction area, the material attenuation coefficient is one of the two specific attenuation coefficients, using a binary matrix Y with elements 0 and 1 1976×2 To represent the attenuation coefficient distribution, each column of the matrix represents a corresponding attenuation coefficient value, and each row represents a grid in the reconstruction area.

[0068] The distribution of each material can be regarded as an object with a uniform attenuation coefficient. The attenuation distance occupied by the material on each X-ray path is the projection of the X-ray passing through the uniform object. Therefore, the sensitivity matrix W1′ can be used 200×1976 Establish the relationship between histogram and binary distribution

[0069] X 1200×2 =W' 1200×1976 ·Y 1976×2 (6)

[0070] Combining equation (5) and equation (6) we can form the projection attenuation rate p on each X-ray beam: 1200×1 To the binary matrix Y 1976×2 The new set of equations

[0071] W' 1200×1976 ·Y 1976×2 ·[U 2×1 e 2×1 ]=[p 1200×1 W' 1200×1976 ·e 1976×1 ] (7)

[0072] in, matrix is a full rank matrix, so formula (7) can be written as

[0073]

[0074] That is, we get the binary matrix Y 1976×2 The final solution model.

[0075] Step 3: Set the maximum number of iterations of the outer loop L m And the maximum number of iterations K of the inner loop m , combined with the binary matrix Y 1976×2 First, for the solution equation of formula (8), the joint algebraic reconstruction algorithm integrating Sigmoid function constraints is used to reconstruct the binary matrix Y 1976×2 Solve each column in turn, and in (k+1) iterations, get the binary matrix Y 1976×2 The jth element of a column The calculation formula is

[0076]

[0077] Among them, wij is the sensitivity matrix W′ 1200×1976 The element in row i and column j, w i is the i-th row of the sensitivity matrix, x i For the matrix The i-th element of the column, y k is the binary vector of the kth iteration. 1976×2 After all elements in the column are solved by equation (9), the column is processed by Sigmoid function to find the maximum value y in the column j,max With the minimum value y j,min , each element in the column is processed as follows

[0078]

[0079] Among them, a is the parameter of Sigmoid function, which is used to adjust the degree of binarization. 1976×2 This column loops through the calculations of equations (9) and (10) until the number of iterations meets the maximum number of iterations K of the inner loop. m .

[0080] For binary matrix By performing the above operations column by column, we can obtain the binary matrix in the lth iteration.

[0081] Step 4: Use the Sigmoid function to constrain the binary matrix Y 1976×2 The rows of the matrix Y are constrained. 1976×2 Each row represents a grid, and each grid has a unique attenuation coefficient, that is, the matrix Y 1976×2 Each row has only one 1 value, so the binary matrix obtained in step 3 needs to be Each row is processed by Sigmoid function

[0082]

[0083] Among them, y i is a binary matrix The i-th element of a row, y i,max With y i,min are the maximum and minimum values ​​of the two elements in the row respectively. After the calculation of formula (11), we can get the binary matrix after the lth iteration: The i-th element of the row For the binary matrix obtained in step 3 The elements of each row of are operated as above in turn to obtain the binary matrix after the lth iteration Update value of

[0084] If the current number of iterations is l <Lm , then repeat steps 3 and 4. When the maximum number of iterations of the outer loop L is met m After that, we can get the binary matrix Y 1976×2 The calculation result of .

[0085] Step 5: Calculate the material distribution in the reconstruction area. Iterate and calculate to get the binary matrix Y 1976×2 After that, the attenuation coefficient Z of all grids inside the circular reconstruction area 1967×1 It can be calculated by the following formula

[0086] Z 1976×1 =Y 1976×2 ·U 2×1 (12)

[0087] Finally, the material distribution in the reconstructed area can be obtained through the attenuation coefficient distribution.

Claims

1. A binary distribution tomographic image reconstruction method suitable for few-angle projections, Features: A fan beam light source is used to illuminate the two-dimensional reconstruction area, and the projection values ​​at different projection angles are measured. According to the sensor optical path layout and grid division, the sensitivity matrix of the reconstruction area is calculated, and the corresponding attenuation coefficient matrix is ​​obtained according to the material composition in the reconstruction area. A binary matrix with elements of 0 and 1 is used to represent the attenuation coefficient distribution of the reconstruction area. According to the fact that the sum of the path lengths occupied by each attenuation coefficient on each X-ray is equal to the sum of the optical path lengths of the optical path passing through all grids, a distance constraint of the sensitivity matrix is ​​introduced, and a model between the projection attenuation rate and the attenuation coefficient distribution on each light ray is established. In view of the characteristics that the binary matrix to be solved only contains two elements of 0 and 1 and each row has only one element with a value of 1, a Sigmoid function constraint is added in the iterative process, and the binary matrix is ​​solved using the established model. According to the set attenuation coefficient, the attenuation coefficient distribution result of the two-dimensional reconstruction area is obtained by binary matrix mapping, thereby realizing the reconstruction of material distribution with few angles.

2. The binary distribution tomographic image reconstruction method suitable for small-angle projection according to claim 1, Features Make full use of the projection information and use the binary matrix to establish a linear imaging model from the multi-light path projection attenuation rate to the attenuation coefficient distribution of the field to be measured. Combined with the matrix characteristics, the Sigmoid function constraint is added to the iterative algorithm to solve the binary matrix to achieve a high-precision reconstruction of the material distribution of the field to be measured with few angles. The specific steps include the following: Step 1: Obtain the integrated attenuation rate projections on multiple X-ray paths through an X-ray reconstruction system; the reconstruction system includes an X-ray source, a stage for the object to be measured, and a detector array, wherein s X-ray sources are evenly arranged on a circle centered on the object to be measured, each ray source corresponds to a detector array, and each detector array consists of d detectors arranged in a row; the X-ray source is located on one side of the object to be measured, and the corresponding detector array is located on the other side. The rays of each X-ray source are fan-shaped beams, covering only the detector array on the opposite side, and the sensor arranges a total of Q = s × d rays; the detector is used to obtain the X-ray projection values ​​when there is no object to be measured and when there is an object to be measured, which are used as the incident and outgoing energies of the X-rays, respectively. According to the Lambert-Beer law, the attenuation rate p is expressed as Among them, I 0 is the incident intensity of X-rays, I is the intensity of X-rays after attenuation, L is the distance that X-rays pass through the object, and u is the linear attenuation coefficient of the object; When an X-ray passes through a non-uniform field to be measured, the attenuation rate projection on the X-ray path is the superposition of the attenuation rates of different substances. If the field to be measured consists of only two substances, and the linear attenuation coefficients of each substance are calculated from the database, the attenuation rate p is expressed as Among them, u m is the attenuation coefficient of the mth substance per unit path length, Δx m is the path length occupied by the mth material distribution. Formula (2) can be written in matrix form, which is the single light path material distribution histogram model. p=X 1×2 ·U 2×1 (3) Where p is the total attenuation rate of the optical path, U 2×1 The rows represent the linear attenuation coefficients of different materials, X 1×2 The columns of represent the path lengths occupied by different substances, which is defined as a substance distribution histogram; For two-dimensional imaging, a multi-angle X-ray optical path covering the reconstruction area is arranged, and the attenuation rate projection along multiple X-ray paths is measured. If the total number of X-ray paths is Q, the attenuation rate projection along Q paths constitutes the attenuation vector p Q×1 According to formula (3), the material distribution histogram X of Q X-ray paths is Q×2 satisfy p Q×1 =X Q×2 ·U 2×1 (4) Step 2: Establish a binary matrix solution model for material distribution; discretize the reconstruction area into D×D grids, and consider the material composition in each grid to be unique, that is, the attenuation coefficient in each grid is a constant. According to the sensor optical path layout and grid division, calculate the path length of each X-ray passing through all grids, which is recorded as the sensitivity matrix W Q×DD , sensitivity matrix W Q×DD The element w in the i-th row and j-th column of i,j represents the optical path length of the i-th X-ray passing through the j-th grid, and thus the total optical path length constraint satisfied by each optical path is obtained, that is, the sum of the histogram values ​​along each X-ray path is equal to the sum of the optical path lengths of the optical path passing through all grids, expressed as X Q×2 ·have been 2×1 =W Q×DD ·have been DD×1 (5) Among them, e DD×1 and e 2×1 is a column vector whose elements are all 1, so equations (4) and (5) form a new set of equations with distance constraints: X Q×2 ·[U 2×1 have been 2×1 ]=[p Q×1 W Q×DD ·have been DD×1 ] (6) In each grid of the reconstruction area, the material attenuation coefficient is one of the two specific attenuation coefficients, which can be represented by a binary matrix Y with elements 0 and 1. DD×2 To represent the attenuation coefficient distribution, each column of the matrix represents a corresponding attenuation coefficient value, and each row represents a grid in the reconstruction area; The distribution of each material can be regarded as an object with a uniform attenuation coefficient. The attenuation distance occupied by the material on each X-ray optical path is the projection of the X-ray passing through the uniform object. Therefore, the sensitivity matrix W can be used to calculate the attenuation distance of the material on each X-ray optical path. Q×DD Establish the relationship between histogram information and binary distribution X Q×2 =W Q×DD ·Y DD×2 (7) Combining equation (6) and equation (7) we can form the projection attenuation rate p on each X-ray beam: Q×1 To the binary matrix Y DD×2 The new set of equations W Q×DD ·Y DD×2 ·[U 2×1 have been 2×1 ]=[p Q×1 W Q×DD ·have been DD×1 ] (8) in, Because u 1 ≠u 2 , so the matrix is a full rank matrix, equation (8) can be written as That is, we get the binary matrix Y DD×2 The final solution model; Step 3: Set the maximum number of iterations of the outer loop L m And the maximum number of iterations of the inner loop K m , combined with the binary matrix Y DD×2 The characteristics of are solved; first, for the solution equation of formula (9), the binary matrix Y is reconstructed using the joint algebraic reconstruction algorithm fused with the Sigmoid function constraint. DD×2 Solve each column in turn, and in (k+1) iterations, get the binary matrix Y DD×2 The jth element of a column The calculation formula is Among them, w ij is the sensitivity matrix W Q×DD The element in row i and column j, w i is the i-th row of the sensitivity matrix, x i For the matrix The i-th element of the column, y k is the binary vector of the kth iteration; when the matrix Y DD×2 After all the elements in this column are solved by formula (10), the column is processed by Sigmoid function to find the maximum value y in the column j,max With the minimum value y j,min , each element in the column is processed as follows Among them, a is the parameter of the Sigmoid function, which is used to adjust the degree of binarization approximation; for the binary matrix Y DD×2 This column loops through the calculations of equations (10) and (11) until the number of iterations meets the maximum number of iterations K of the inner loop. m ; For binary matrix Perform the above operation column by column to obtain the binary matrix in the lth iteration. Step 4: Use the Sigmoid function to transform the binary matrix Y DD×2 The rows of the matrix Y are constrained; DD×2 Each row represents a grid, and each grid has a unique attenuation coefficient, that is, the matrix Y DD×2 Each row has only one 1 value, so the binary matrix obtained in step 3 needs to be Each row is processed by Sigmoid function Among them, y i is a binary matrix The i-th element of a row, y i,max With y i,min are the maximum and minimum values ​​in the row respectively. After the calculation of formula (12), we can get the binary matrix after the lth iteration: The i-th element of the row For the binary matrix obtained in step 3 The elements of each row of are operated as above in turn to obtain the binary matrix after the lth iteration The updated value of If the current number of iterations is l <L m , then repeat steps 3 and 4; when the maximum number of iterations of the outer loop L is met m After that, we can get the binary matrix Y DD×2 The calculation results of Step 5: Calculate the material distribution in the reconstruction area; obtain the binary matrix Y through iterative calculation DD×2 After that, the attenuation coefficient Z in all grids DD×1 It can be calculated by the following formula Z DD×1 =Y DD×2 ·U 2×1 (13) Finally, the material distribution in the reconstructed area can be obtained through the attenuation coefficient distribution.