Low-dose CT chordal graph recovery method and system based on multi-Gaussian electronic noise modeling

Abstract

The invention discloses a low-dose CT chordal graph recovery method and system based on multi-Gaussian electronic noise modeling, and the method comprises the following steps: building a low-dose CT chordal graph recovery model based on multi-Gaussian electronic noise modeling according to a CT projection noise generation model, decomposing the chordal graph recovery model into sub-problems about a plurality of to-be-estimated parameters by adopting an alternating projection method; updating and solving multiple Gaussian parameters according to an EM algorithm, iteratively solving and updating a photon number, iteratively solving chordal graph data according to a near-end operator, and calculating by using a Resnet network model by the near-end operator in chordal graph data solving; the limitation that the electronic noise in the assumed image data follows single Gaussian distribution when the electronic noise distribution is processed in the traditional technology is broken through, and more complex electronic noise distribution can be accurately estimated. Through an alternate projection algorithm, the method not only shows excellent ability in the aspect of chordal graph recovery, but also realizes significant performance improvement in the aspects of structural similarity and visual consistency.

Description

Technical Field

[0001] The invention belongs to the technical field of low-dose CT projection image restoration, and in particular relates to a low-dose CT chordogram restoration method introducing multi-Gaussian electronic noise. Background Art

[0002] Computed tomography (CT) has become one of the most common detection and diagnosis methods in modern medical imaging. However, the widespread use of CT has caused people to worry that X-rays may cause cancer or genetic damage, and low-dose CT imaging algorithms that reduce X-ray doses during scanning have received widespread attention. The main measure to reduce radiation dose is to reduce the X-ray tube current or tube voltage, thereby reducing the X-ray hazard to patients. However, reducing the tube current and tube voltage will lead to enhanced photon noise and electronic noise, making the noise background of the chordogram data (chordogram) more significant, seriously affecting image quality and diagnostic accuracy.

[0003] Chord data refers to CT scan data before logarithmic transformation. It is the basis of all CT imaging algorithms, and its quality sensitively affects the image effect of FBP (Filtered Back Projection, FBP) imaging algorithm or other imaging algorithms. Therefore, a good chord data recovery method can effectively enable other CT imaging algorithms to restore high-quality CT images while reducing the radiation dose of CT scanning.

[0004] The current chord diagram data recovery method uses a Poisson distribution and a single Gaussian distribution to characterize and model photon noise and electronic noise, respectively, and then calculates the chord diagram data. The Chinese patent with publication number: CN16312983A designs an implicit regularized low-dose CT image reconstruction method based on a noise generation mechanism, which solves the problem of high-quality CT image restoration under ultra-low dose conditions, but focuses on the assumption that the system's inherent electronic noise obeys a single Gaussian distribution. CT imaging literature research shows that a single Gaussian distribution model cannot accurately characterize electronic noise. Theoretical research has shown that multi-Gaussian mixture distributions can approximate any probability distribution. Summary of the invention

[0005] In order to better restore the chordogram data, the present invention uses a low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling, which can more accurately characterize the true distribution characteristics of the CT system electronic noise and quickly and effectively remove the noise of the chordogram data.

[0006] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a low-dose CT chordogram recovery method based on multi-Gaussian electronic noise modeling, firstly establishing an optimization model of introducing electronic noise as a multi-Gaussian electron distribution, and then designing an alternating projection iterative algorithm to solve it to obtain chordogram data that is closer to the actual one.

[0007] The specific steps include:

[0008] A low-dose CT chordogram data recovery model that introduces Gaussian electronic noise into the low-dose CT chordogram data input;

[0009] An alternating projection iterative algorithm is used to solve the low-dose CT chordogram data recovery model with Gaussian electronic noise. The solution of the low-dose CT chordogram data recovery model with Gaussian electronic noise is decomposed into the problem of solving the photon number T and the problem of solving the chordogram data Y. The proximal gradient descent method is used to solve Y to obtain the restored chordogram data Y.

[0010] The CT imaging algorithm is used to convert the chord data Y into a readable CT image.

[0011] Furthermore, the low-dose CT chordogram data recovery model with the introduction of Gaussian electronic noise is:

[0012]

[0013] in represents the number of photons received by the detector, where N is the total number of X-ray incident paths in the scan, i represents the i-th incident X-ray path, i = 1, 2, ..., N. K represents the number of multi-Gaussians, k = 1, 2, ..., K, σ k , μ k Respectively represent the variance and mean of the kth Gaussian distribution, π k represents the weight of the kth Gaussian distribution, I 0i is the intensity of the i-th incident X-ray. i is the number of photons received by the detector when incident along the projection path i, Y is the chord diagram data, and Y i is the chord diagram data on the detector when incident along the projection path i, g1(Y) is the regularization function with respect to Y, and λ1 is the penalty coefficient of the regularization function.

[0014] It should be noted that, for the purpose of algorithm design, it is temporarily assumed that K multi-Gaussian parameters are given.

[0015] Further, the training of the low-dose CT chordogram data recovery model with the introduction of Gaussian electronic noise comprises the following steps:

[0016] Acquire several conventional-dose CT images, convert them into chord data by inverse Radon transform, and superimpose a mixed distribution consisting of K Gaussian distributions in the projection domain to simulate a more realistic electronic noise distribution, thereby forming low-dose projection domain data;

[0017] The alternating projection optimization algorithm is used to decompose the chord diagram data recovery model into several sub-problems about parameters to be estimated. The EM algorithm and the proximal gradient descent algorithm are used to iteratively update the parameters. After the algorithm converges, the chord diagram data Y and the chord diagram data recovery model are obtained.

[0018] The normal-dose real CT images are converted into chord data through inverse Radon transform, and then mixed multi-Gaussian distribution noise simulation is introduced into the chord data to simulate the data characteristics under real low-dose CT imaging conditions. The step of adding noise to the chord data not only considers the influence of photon noise, but also covers the diversity of electronic noise.

[0019] The alternating projection optimization algorithm is further used to decompose the chord diagram data recovery model into several sub-problems about the parameters to be estimated. The EM algorithm and the proximal gradient descent algorithm are used to iteratively update the parameters, including:

[0020] The model (1) is solved using the alternating projection iterative algorithm.

[0021] S3.1. Fixed multi-Gaussian parameter σ k , μ k , the weight coefficient of the kth Gaussian distribution and the chord diagram data Y, establish a sub-optimization model for solving the photon number T, and use T i The variable separation property converts the original problem into a single variable optimization problem T i Optimization problem:

[0022]

[0023] Apply Newton's method to update the objective function. The first and second order derivatives f′(T i )、f″(T i )

[0024]

[0025]

[0026]

[0027]

[0028] Here R is the rounding function.

[0029] S3.2, use the proximal gradient method to update the chord diagram data Y, fix the chord diagram data Y, use the proximal gradient descent method to solve Y, the update iterative formula of Y is as follows:

[0030]

[0031] Among them, the proximal operator The residual network (ResNet) can be used for learning, and the Resnet model is used to calculate the proximal operator determined by the regularization term g1(·).

[0032] Step 3.3, calculate the multi-Gaussian parameters Π, Λ, Σ, and use the obtained photon number T and the original data S, with the help of the residual data, that is, the difference between the low-dose chord data S and the number of photons T received by the detector, to estimate the multi-Gaussian distribution parameters using the EM algorithm. The EM algorithm iteration process is set to 10 rounds, and the formula for iteratively updating the multi-Gaussian parameters Π, Λ, Σ is as follows:

[0033]

[0034]

[0035]

[0036]

[0037] After the iteration is completed, the estimated multi-Gaussian parameters will be regarded as fixed values ​​and will be further used in subsequent algorithm processing.

[0038] The iteration is stopped after the T and Y iteration process reaches a fixed number of rounds, otherwise the obtained multi-Gaussian parameters Π, Λ, Σ, T, and Y are brought into S3.1 to continue execution.

[0039] Furthermore, the residual network is composed of 4 residual blocks, each of which is composed of two layers of convolution, batch normalization and nonlinear activation function ReLU. The filter size of each convolution layer is 3×3, and the number of channels varies according to the dimension of the input data. After each convolution, the output result is normalized, nonlinearly transformed by ReLU activation function, and the input data is directly passed to the output layer using residual connection, thereby alleviating the gradient vanishing problem in the deep network; at the same time, in order to ensure that the Y update process contains richer feature information, the historical update value Y plus Concatenate it to the current Y, and use the concatenated result as input to the residual network for further processing.

[0040] Furthermore, the CT imaging algorithm is used to convert the chordogram data Y into a readable CT image, including: fixing the proximal operator learned in S3.2, substituting the scanned low-dose chordogram data S, the incident light intensity I0, and the multi-Gaussian parameters Π, Λ, Σ into the algorithm of step 3, converting the restored chordogram data Y into the final CT image through the FBP algorithm, calculating the structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) compared with the conventional dose CT image, and evaluating the chordogram data recovery results.

[0041] Based on the concept of the method, the present invention also provides a low-dose CT chordogram recovery system based on multi-Gaussian electronic noise modeling, a data acquisition module, a chordogram data recovery module and an image conversion module;

[0042] The data acquisition module is used for acquiring low-dose CT chordogram data and inputting the low-dose CT chordogram data into a low-dose CT chordogram data recovery model into which Gaussian electronic noise is introduced;

[0043] The chord data recovery module uses an alternating projection iterative algorithm to solve the low-dose CT chord data recovery model with Gaussian electronic noise. The solution of the low-dose CT chord data recovery model with Gaussian electronic noise is decomposed into the problem of solving the photon number T and the problem of solving the chord data Y. The proximal gradient descent method is used to solve Y to obtain the restored chord data Y.

[0044] The image conversion module uses a CT imaging algorithm to convert the chord data Y into a readable CT image.

[0045] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, the low-dose CT chordogram recovery method based on multi-Gaussian electronic noise modeling can be implemented.

[0046] A computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, the low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling can be implemented.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention introduces a multi-Gaussian distribution that more realistically expresses electronic noise into the CT projection noise generation mechanism, which can more accurately capture the real distribution characteristics of the electronic noise of the CT system. The constructed noise model is more in line with the actual physical laws of CT images, breaking through the limitation of the chord diagram data recovery technology that electronic noise is represented by a single Gaussian distribution.

[0049] The present invention proposes a low-dose CT chordogram restoration method based on multi-Gaussian electronic noise, which cleverly realizes the update of multi-Gaussian distribution parameters by introducing the expectation maximization algorithm, thus overcoming the dilemma that the multi-Gaussian parameters are optimized as unknown variables when establishing an optimization model, making it difficult for the alternating projection algorithm to solve.

[0050] The present invention proposes a low-dose CT image restoration method that can be applied to the study of simulation data when it is not possible to directly perform low-dose scanning on patients; it lays a foundation for directly performing low-dose scanning on patients in the future and further carrying out low-dose CT imaging and algorithm research. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is represented as a framework diagram of the low-dose CT chordogram recovery method based on multi-Gaussian electronic noise modeling.

[0052] Figure 2 Chord diagram showing conventional dose levels (200mAs, 120kVp).

[0053] Figure 3 It is represented as the chord diagram restored by the algorithm of the present invention.

[0054] Figure 4 Shown as a chord diagram for low dose level (20mAs, 120kVp).

[0055] Figure 5 The images are shown as scanned at (200 mAs, 120 kVp) and reconstructed using FBP after chordogram restoration according to the present invention.

[0056] Figure 6 The images reconstructed by FBP at conventional dose level (200 mAs, 120 kVp) are shown.

[0057] Figure 7 Shown are FBP reconstructed images at low dose (20 mAs) level. DETAILED DESCRIPTION

[0058] The specific implementation modes of the present invention are described in detail below with reference to the accompanying drawings.

[0059] Embodiment 1, the method for recovering low-dose CT chordogram data by introducing multi-Gaussian electronic noise provided by the present invention comprises the following steps:

[0060] (1) Design of a low-dose CT chordogram recovery algorithm that introduces multi-Gaussian electronic noise

[0061] The generative model of CT projection data S can be expressed as:

[0062] S=T+ε (1)

[0063] in, represents the number of photons received by the detector, represents the electronic background noise, where N is the total number of X-ray incident paths in the scan.

[0064] In order to more accurately describe the electronic noise of the CT system, the electronic noise of the i-th incident path can be expressed as the following mixed Gaussian distribution:

[0065]

[0066] Among them, ε i is the electronic noise of the ith incident path, i = 1, ..., N; k = 1, ..., K, K represents the number of Gaussian distributions, π k represents the weight coefficient of the kth Gaussian distribution; μ k ,σ k denote the mean and variance of the kth Gaussian distribution, denoted by Π={π1,…,π k}, Λ={μ1,…,μ k},∑={σ1,…,σ k}, its kth Gaussian component is:

[0067]

[0068] The number of photons T received by the CT detector is the result of X-ray photon fluctuations. Related research results show that the number of photons obtained by CT scanning approximately follows the Poisson distribution law:

[0069]

[0070] G i Represents the average number of photons received under ideal conditions.

[0071] Let Y represent the noise-free "chord diagram data". According to the Lambert-Beer law, Y satisfies:

[0072]

[0073] Combining (4) and (5), we can obtain:

[0074]

[0075] Combining (1) and (2), we can obtain the following conditional joint distribution:

[0076]

[0077] In summary, the posterior conditional distribution of S and T with respect to Y is:

[0078]

[0079] According to the Bayesian conditional probability formula, the complete posterior distribution of the estimated parameters T and Y can be obtained:

[0080]

[0081] (9) Since S is the measured data, p(S) is a constant. So the maximum a posteriori estimation model of the estimated parameters T and Y is obtained:

[0082]

[0083] Taking the negative logarithm of the objective function of (10), it can be transformed into the following equivalent optimization problem:

[0084]

[0085] Due to the convexity of the logarithmic function, according to Jensen's inequality, solving the minimum value of the objective function is transformed into solving the minimum value of its upper bound:

[0086]

[0087] Furthermore, the chord graph implicit regularization term g1(Y) is added to the above model to construct the final CT chord graph data reconstruction model:

[0088]

[0089] (2) Update the parameters in the iterative CT chord recovery model

[0090] The present invention uses the alternating projection method to decompose the parameter problem in the chord diagram recovery model into sub-problems about the parameters to be estimated T and Y, and then uses the EM algorithm to update the multi-Gaussian parameters Π, Λ, Σ that are temporarily assumed to be fixed. The specific calculation process is as follows:

[0091] Update T

[0092] Solve the following subproblems about T:

[0093]

[0094] Using T i The variable separation property of the above N-dimensional optimization problem can essentially be converted into N single-variable optimization problems:

[0095]

[0096] For the above N single variable optimization problems, the integer variable T i Relaxation is a continuous variable, and equation (17) can be generalized to the following problem:

[0097]

[0098] Solve the first-order derivative and second-order derivative of equation (18) respectively:

[0099]

[0100] Among them, Γ(T i +1) is the generalized factorial, which is used to replace T in continuous space i ! . Since f″(T i )>0, in continuous space, integer programming is a strictly convex function, which means that any local optimal solution is also a global optimal solution, and Newton's method can be applied to update T:

[0101]

[0102] Where superscript n is the iteration index, η1 is the step size, and in the last iteration, rounding is used to approximate T i The real integer value of , thus obtaining an approximate integer solution to the global optimal solution.

[0103] T i n =R(T i n ) (twenty one)

[0104] Here R is the rounding function.

[0105] Update Y

[0106] When the values ​​of other parameters are determined, the update of parameter Y is equivalent to solving the following sub-problems:

[0107]

[0108] The first term in the above formula is at a given point Y( k ) is approximated by the second-order Taylor expansion, then the second-order approximation of the subproblem is:

[0109]

[0110] Among them, Y (n-1) is the updated result of the chord diagram data after the n-1th iteration, η2 is the step length, solve (22), Y (n) The update can be expressed using the proximal operator as:

[0111]

[0112] in is the proximal operator determined by the regularization term g1(·). The calculation formula is as follows:

[0113]

[0114] Combining equation (23) with equation (24), Y (n) The complete update iteration formula is as follows:

[0115]

[0116] For (25), the proximal operator Residual network (ResNet) can be used for learning.

[0117] The update of Y applies the "identical shortcut link" in the deep residual network, and uses ResNet to expand the proximal operator, so that the residual network automatically learns the prior information of the image from the training data and gradually optimizes the objective function. The residual network used consists of 4 residual blocks, each of which consists of two layers of convolution, batch normalization, and nonlinear activation function ReLU. The filter size of each layer of convolution is 3×3, and the number of channels varies according to the dimension of the input data. After each convolution, the output result is normalized, nonlinearly transformed by the ReLU activation function, and the residual connection is used to pass the input data directly to the output layer, thereby alleviating the gradient disappearance problem in the deep network.

[0118] Calculate Π, Λ, Σ:

[0119] Assume hidden variables It indicates the probability that the implicit category of each sample is the kth Gaussian distribution, and its posterior expectation is as follows:

[0120]

[0121] Update iterative Π, Λ, ∑ parameters:

[0122]

[0123]

[0124]

[0125] in, is the sum of the probabilities belonging to the kth Gaussian distribution. Repeat steps E and M until the maximum number of iterations is reached or T and Y converge. The final parameters Π, Λ, Σ, i.e., the estimated values ​​of the multi-Gaussian model in this round, are introduced as constants in the rest of the parameter update process.

[0126] The above-mentioned multi-Gaussian update module can effectively suppress the noise and artifacts in the chord data while retaining the detailed texture of the reconstructed image. Compared with the deep network based on the assumption that electronic noise obeys a single Gaussian noise model, this method can show better performance in terms of chord structure similarity and visual consistency, and more effectively suppress the noise of chord data.

[0127] Example

[0128] S1, the CT images under conventional doses are converted into chordogram data by inverse Radon transform, and the electronic noise obeying multi-Gaussian distribution is introduced into the chordogram domain to simulate the data characteristics under real low-dose CT imaging conditions.

[0129] S2, collects chest CT chordogram images formed using low-dose technology and paired chest CT chordogram images formed using normal dose; there are 700 simulated low-dose CT chordogram images and 700 corresponding normal-dose CT chordogram images, and the size of the collected single image is 736×1152, which together constitute a 700 paired CT chordogram data set.

[0130] S3, the paired CT chord image dataset is randomly divided into a training set and a test set according to a ratio of 6:1, where the training set contains 600 pairs of CT chord image images and the test set contains 100 pairs of CT chord image images;

[0131] S4. The proximal operator is trained by using the chord diagram data recovery algorithm with the introduction of multi-Gaussian electronic noise, and the proximal operator is calculated using the deep Resnet network model.

[0132] S5. Use the Adam optimizer with an adaptive learning rate decay strategy to train a low-dose CT image denoising model. The hyperparameters set include: 20 training epochs, 2 batches, and an initial learning rate of 0.0001.

[0133] S6. After the model training is completed, call the model weight file saved during the training process, input the low-dose CT chordogram image in the test set, and you can get the denoised high-quality CT chordogram. Among them, the chordogram at the conventional dose (200mAs, 120kVp) level is as follows Figure 2 As shown, the restored chord diagram is Figure 3 The chord diagram for low dose level (20mAs, 120kVp) is shown in Figure 4 As shown; the FBP reconstructed image at the conventional dose level (200mAs, 120kVp) is as follows Figure 5 As shown, the chord diagram at the low dose (20mAs, 120kVp) level is scanned, and the FBP reconstructed image after restoration through multi-Gaussian error modeling is as follows Figure 6 As shown, the FBP reconstructed images at low dose (20mAs, 120kVp) level are as follows Figure 7 shown.

[0134] On the test set, the PSNR (Peak Signal to Noise Ratio) of the FBP reconstructed image restored by multi-Gaussian error modeling reached 36.29, and the SSIM (Structure Similarity Index Measure) reached 0.948. The results show that this method can effectively denoise low-dose CT chordograms.

[0135] Embodiment 2, the present invention also provides a low-dose CT chordogram recovery system based on multi-Gaussian electronic noise modeling, a data acquisition module, a chordogram data recovery module and an image conversion module;

[0136] The data acquisition module is used for acquiring low-dose CT chordogram data and inputting the low-dose CT chordogram data into a low-dose CT chordogram data recovery model into which Gaussian electronic noise is introduced;

[0137] The chord data recovery module uses an alternating projection iterative algorithm to solve the low-dose CT chord data recovery model with Gaussian electronic noise. The solution of the low-dose CT chord data recovery model with Gaussian electronic noise is decomposed into the problem of solving the photon number T and the problem of solving the chord data Y. The proximal gradient descent method is used to solve Y to obtain the restored chord data Y.

[0138] The image conversion module uses a CT imaging algorithm to convert the chord data Y into a readable CT image.

[0139] The method also includes a model training module, which is used to obtain a number of conventional dose CT images, convert the conventional dose CT images into chord data through inverse Radon transform, and superimpose a mixed distribution consisting of K Gaussian distributions in the projection domain to simulate a more realistic electronic noise distribution, thereby forming low-dose projection domain data; an alternating projection optimization algorithm is used to decompose the chord data recovery model into several sub-problems about parameters to be estimated, and the EM algorithm and the proximal gradient descent algorithm are used to iteratively update the parameters, including:

[0140] Fixed the multi-Gaussian parameters Π, Λ, Σ and the chord diagram data Y, established a sub-optimization model for solving the photon number T, and used T i The variable separation property converts the original problem into a single variable T i The optimization problem is:

[0141]

[0142] Apply Newton's method to update the objective function and calculate f(T i )’s first and second order derivatives f′(T i )、f″(Ti ), and use Newton's method to update T:

[0143]

[0144]

[0145]

[0146]

[0147] Where R is the rounding function, the superscript n is the iteration index, η1 is the step size, and in the last iteration, rounding is used to approximate T i The real integer value of ;

[0148] The number of photons T is fixed, and the proximal gradient descent method is used to solve Y. The update iteration of Y is as follows:

[0149]

[0150] Among them, the proximal operator Using residual network learning;

[0151] Using the photon number T and the original data S, and the difference between the original chord data and the number of photons received by the detector, the EM algorithm is used to estimate the parameters of the multi-Gaussian distribution, and the multi-Gaussian parameters Π, Λ, Σ are obtained and fixed.

[0152] Repeat the above steps until T and Y converge.

[0153] In summary, the present invention provides a low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling. According to the CT projection noise generation model, a low-dose CT chordogram restoration model based on multi-Gaussian noise modeling is established, the maximum posterior distribution under known chordogram data conditions is constructed, and the chordogram restoration model is decomposed into several sub-problems about parameters to be estimated by using an alternating projection method; the multi-Gaussian parameters are updated and solved according to the EM algorithm, the number of photons is updated and solved iteratively, and the chordogram data is solved iteratively according to the proximal operator. The proximal operator in the chordogram data solution is calculated using the Resnet network model; the present invention breaks through the traditional technology that is limited to assuming that the image data is in the distribution of electronic noise when processing Electronic noise follows the limitations of a single Gaussian distribution, and can accurately estimate more complex electronic noise distributions; a Poisson distribution + multi-Gaussian mixed noise model is proposed, which cleverly incorporates multi-Gaussian parameters but does not model them as variables, overcoming the dilemma of modeling multi-Gaussian parameters as variables in the model and making it difficult to design a model solution algorithm. The multi-Gaussian mixed distribution approximates the real electronic noise, so that it can better retain the real structure of the chord data, thereby reconstructing a clearer CT image through the FBP imaging algorithm or other new imaging algorithms, not only showing excellent capabilities in chord recovery, but also achieving significant performance improvements in structural similarity and visual consistency. In addition, the present invention has better adaptability and ease of operation, and is particularly suitable for processing large-scale data sets, especially in the field of clinical computed tomography, and can effectively meet the urgent needs for fast imaging processing.

[0154] In addition, the present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, the low-dose CT chordogram recovery method based on multi-Gaussian electronic noise modeling described in the present invention can be implemented.

[0155] On the other hand, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the low-dose CT chordogram recovery method based on multi-Gaussian electronic noise modeling described in the present invention can be implemented.

[0156] The computer device may be a laptop computer, a desktop computer or a workstation.

[0157] The processor described in the present invention may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC) or a readily available field programmable gate array (FPGA).

[0158] The memory of the present invention may be an internal storage unit of a laptop computer, a desktop computer or a workstation, such as a memory or a hard disk; or an external storage unit, such as a mobile hard disk or a flash memory card.

[0159] Computer-readable storage media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media may include: read-only memory (ROM), random access memory (RAM), solid-state drive (SSD) or optical disk, etc. Among them, random access memory may include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0160] The above description shows and describes several preferred embodiments of the invention, but as mentioned above, it should be understood that the invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the invention concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall be within the scope of protection of the claims attached to the invention without departing from the spirit and scope of the invention.

Claims

1. A low-dose CT chordogram recovery method based on multi-Gaussian electronic noise modeling, characterized in that: The following steps are involved: A low-dose CT chordogram data recovery model that introduces Gaussian electronic noise into the low-dose CT chordogram data input; An alternating projection iterative algorithm is used to solve the low-dose CT chordogram data recovery model with Gaussian electronic noise. The solution of the low-dose CT chordogram data recovery model with Gaussian electronic noise is decomposed into the problem of solving the photon number T and the problem of solving the chordogram data Y. The proximal gradient descent method is used to solve Y to obtain the restored chordogram data Y. The CT imaging algorithm is used to convert the chord data Y into a readable CT image.

2. The low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling according to claim 1, characterized in that: The low-dose CT chordogram data recovery model with the introduction of Gaussian electronic noise is: in, represents the number of photons received by the detector, N is the total number of X-ray incident paths in the scan, i represents the i-th incident X-ray path, i=1,2,…,N, K represents the number of multi-Gaussians, k=1,2,…,K, σ k , μ k Respectively represent the variance and mean of the kth Gaussian distribution, π k represents the weight of the kth Gaussian distribution, I 0i is the intensity of the i-th incident X-ray, T i is the number of photons received by the detector when incident along the projection path i, Y is the chord diagram data, and Y i is the chord diagram data on the detector when incident along the projection path i, g1(Y) is the regularization function with respect to Y, and λ1 is the penalty coefficient of the regularization function.

3. The low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling according to claim 2, characterized in that: The training of the low-dose CT chordogram data recovery model with Gaussian electronic noise introduced comprises the following steps: Acquire several conventional-dose CT images, convert them into chord data by inverse Radon transform, and superimpose a mixed distribution consisting of K Gaussian distributions in the projection domain to simulate a more realistic electronic noise distribution, thereby forming low-dose projection domain data; The alternating projection optimization algorithm is used to decompose the chord diagram data recovery model into several sub-problems about parameters to be estimated. The EM algorithm and the proximal gradient descent algorithm are used to iteratively update the parameters. After the algorithm converges, the chord diagram data Y and the chord diagram data recovery model are obtained.

4. The low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling according to claim 1, characterized in that: The alternating projection optimization algorithm is used to decompose the chord diagram data recovery model into several sub-problems about the parameters to be estimated. The EM algorithm and the proximal gradient descent algorithm are used to iteratively update the parameters, including: Fixed the multi-Gaussian parameters Π, Λ, Σ and the chord diagram data Y, established a sub-optimization model for solving the photon number T, and used T i The variable separation property converts the original problem into a single variable T i The optimization problem is: Apply Newton's method to update the objective function and calculate f(T i )’s first and second order derivatives f′(T i )、f″(T i ), and use Newton's method to update T: T i n =R(T i n ) Where R is the rounding function, the superscript n is the iteration index, η1 is the step size, and in the last iteration, rounding is used to approximate T i The real integer value of ; The number of photons T is fixed, and the proximal gradient descent method is used to solve Y. The update iteration of Y is as follows: Among them, the proximal operator Using residual network learning; Using the photon number T and the original data S, and the difference between the original chord data and the number of photons received by the detector, the EM algorithm is used to estimate the parameters of the multi-Gaussian distribution, and the multi-Gaussian parameters Π, Λ, Σ are obtained and fixed. Repeat the above steps until T and Y converge or the iteration reaches the maximum number.

5. The low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling according to claim 1, characterized in that: The residual network is composed of four residual blocks, each of which consists of two layers of convolution, batch normalization and nonlinear activation function ReLU. After each convolution, the output result is normalized, nonlinear transformation is performed through ReLU activation function, and the input data is directly transferred to the output layer using residual connection to update the historical value Y plus Concatenate it to the current Y, and use the concatenated result as input to the residual network for further processing.

6. The low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling according to claim 1, characterized in that: The CT imaging algorithm is converted into the final CT image through the FBP algorithm.

7. The low-dose CT chordogram restoration method based on multi-Gaussian electronic noise modeling according to claim 1, characterized in that: The low-dose CT image denoising model is trained using the Adam optimizer with an adaptive learning rate decay strategy.

8. A low-dose CT chordogram recovery system based on multi-Gaussian electronic noise modeling, characterized in that: Data acquisition module, chord diagram data recovery module and image conversion module; The data acquisition module is used for acquiring low-dose CT chordogram data and inputting the low-dose CT chordogram data into a low-dose CT chordogram data recovery model into which Gaussian electronic noise is introduced; The chord data recovery module uses an alternating projection iterative algorithm to solve the low-dose CT chord data recovery model with Gaussian electronic noise. The solution of the low-dose CT chord data recovery model with Gaussian electronic noise is decomposed into the problem of solving the photon number T and the problem of solving the chord data Y. The proximal gradient descent method is used to solve Y to obtain the restored chord data Y. The image conversion module uses a CT imaging algorithm to convert the chord data Y into a readable CT image.

9. A computer device, characterized in that: The invention comprises a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, the low-dose CT chordogram recovery method based on multi-Gaussian electronic noise modeling as described in any one of claims 1 to 7 can be implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, the method for restoring low-dose CT chordograms based on multi-Gaussian electronic noise modeling as described in any one of claims 1 to 7 can be implemented.

Images