PET image scattering correction method and system

Abstract

The invention relates to a scattering correction method and system for a PET image. The method comprises the following steps: alternately and iteratively solving radioactivity distribution and an attenuation coefficient of a detector space based on an ADMM algorithm; mapping from PET detection data and a detector space attenuation coefficient to accurate radioactive activity distribution is realized based on the deep learning network G1; based on a deep learning network G2, mapping from a detector space attenuation coefficient to matched linear attenuation coefficient distribution is realized, and an object mask matrix is obtained; unfitted scattering estimation is obtained through single scattering simulation; keeping radioactive activity distribution and a detector space attenuation coefficient as constants, and obtaining a scattering correction factor by using unfitted scattering estimation; and obtaining a PET image subjected to scattering correction by utilizing the accurate scattering distribution after fitting of the spatial attenuation coefficient of the detector and the scattering correction factor. The beneficial effects are that compared with traditional scatter correction, the method is higher in correction precision and shorter in time consumption, further improves the image quality, and reduces the reconstruction time consumption.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a scatter correction method and system for PET images. Background Art

[0002] Positron emission tomography (PET) is a high-end nuclear medicine imaging diagnostic technology. In practical applications, it uses radioactive isotopes (such as 18F, 11C, etc.) to label metabolites and inject these labeled compounds into the human body. Subsequently, the patient is subjected to functional metabolic imaging through the PET system to reflect the metabolic activities in the living body, thereby achieving the purpose of disease diagnosis. Currently, commercial PET devices on the market are usually used in combination with other imaging modes, such as computed tomography (CT) or magnetic resonance imaging (MRI), which can not only simultaneously obtain the patient's anatomical structure information, but also improve the accuracy of lesion positioning. Ultimately, the fusion of functional and anatomical images on the same machine combines the advantages of the two modalities, allowing doctors to understand the patient's physical condition more intuitively and comprehensively, helping to detect lesions early and make accurate diagnoses, especially in the diagnosis and treatment of tumors and heart and brain diseases.

[0003] However, during the PET data acquisition process, photons may experience Compton scattering in the human body and change their direction of travel. Due to the limited energy resolution of the detector, such scattering events are sometimes mistakenly recorded as true coincidence events, thereby interfering with the position information of the original signal, forming artifacts in the generated image, and seriously affecting the image quality. Especially in the three-dimensional acquisition mode, the proportion of scattered coincidence events may be as high as 30%-60% of the total counts. Therefore, effective scatter correction has become one of the key links in the PET reconstruction process.

[0004] The currently widely used scatter correction method is Single Scatter Simulation (SSS). This method is based on the radioactivity distribution and linear attenuation coefficient distribution of the scanned object, and calculates the probability of a single gamma-ray particle being scattered once before being detected to estimate the overall scattering distribution. Since the actual measurement data contains multiple scattering components, it is difficult for the SSS method to accurately estimate the proportional relationship between the scattering component and the total coincidence. Usually, the scattering contribution is determined by analyzing the radial tail characteristics of the sinusoidal graph or using Monte Carlo simulation. In actual clinical acquisition, the accuracy of SSS is usually affected by the following factors:

[0005] First, before scatter correction, the initial distribution of radioactivity is usually approximated by a scatter-free PET reconstructed image. When the patient is heavy, the error of the scatter-free reconstructed initial image is large. Even if the PET reconstructed image is iteratively solved in the SSS process, the global convergence of the scatter parameter estimation cannot be guaranteed, and the error caused by the reconstructed image is still large. Therefore, an inaccurate initial distribution of radioactivity will have a significant impact on the accuracy of scatter correction.

[0006] Secondly, for multimodal acquisition methods, although different imaging technologies can provide the necessary attenuation information for scatter correction, there are still some challenges in the actual operation process. For example, in the PET / CT system, the coverage of PET is often larger than the working area of ​​CT or other auxiliary imaging methods; especially for those subjects with larger body size, other imaging methods may not provide enough field of view width, resulting in truncation of linear attenuation distribution, which in turn affects the quality of PET images. In addition, CT imaging is fast and completed almost instantly, while PET takes a long time to complete the scan of a position, which means that it is difficult for the patient to remain still during the whole process, and factors such as respiratory movement can also cause spatial misalignment between the two. Slight movement of the patient's body during long-term scanning can also cause time synchronization problems between PET and CT images. It is also worth noting that if there are metal objects in the body (such as pacemakers, dental implants, etc.), obvious bright spots will appear on the CT image - the so-called "metal artifacts", which will also bring difficulties to subsequent processing. In short, in clinical practice, the problem of incomplete matching of radioactivity distribution images with other modality images due to various reasons will directly affect the effect of scatter correction and may even introduce serious artifacts in the final result.

[0007] Finally, when using the tail fitting method in the SSS algorithm, only the data outside the object is selected as the reference sample. When the amount of available data is small, especially for large target objects, this method may become unstable and easily underestimate or overestimate the scattering distribution. Summary of the invention

[0008] Technical issues to be solved

[0009] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a scatter correction method and system for PET images, which solves the technical problem of how to further improve image quality and reduce reconstruction time.

[0010] Technical Solution

[0011] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0012] In a first aspect, the present invention provides a scatter correction method for a PET image, comprising:

[0013] Modeling the PET acquisition process;

[0014] Based on the ADMM algorithm, the radioactivity distribution and the attenuation coefficient of the detector space are solved by alternating iteration;

[0015] Based on the deep learning network G1, the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution is achieved;

[0016] Based on the deep learning network G2, the mapping of the detector spatial attenuation coefficient to the matching linear attenuation coefficient distribution is realized, and the object mask matrix is ​​obtained;

[0017] Unfitted scattering estimates are obtained through single scattering simulations based on accurate radioactivity distribution, matching linear attenuation coefficient distribution and object mask matrix;

[0018] Keeping the radioactivity distribution and the detector spatial attenuation coefficient constant, the scatter correction factor is obtained by maximizing the log-likelihood function using the unfitted scatter estimation;

[0019] The accurate scatter distribution after fitting the detector spatial attenuation coefficient and scatter correction factor is used to obtain the scatter-corrected PET image.

[0020] Alternatively, the PET acquisition process can be modeled as follows:

[0021]

[0022] In the formula y=[y 1t ,y 2t ,…,y it ,…,y NT ]' represents the detected data, represents the average value of the detection data, N represents the size of the detection data sinusoidal graph, T represents the size of the TOF discrete space, i represents the variable index of the detection data sinusoidal graph response line, and t represents the variable index of the TOF discrete space; the single quote superscript represents the matrix transposition operation; x = [x1, x2, ..., x j ,…,x M ]' represents the unknown radioactivity distribution image, M represents the size of the radioactivity distribution image space, j represents the variable index of the radioactivity distribution image space, and represents the point source at the corresponding spatial position; a = [a1, a2, ..., a j …,a N] ' represents the attenuation coefficient of the detector space; A = [A ijt] is the system matrix; S = [s 1t ,s 2t ,…,s NT ]' represents the estimated value of scattering, α = [α 1t ,α 2t ,…,α NT ]' represents the correction factor for the scatter estimation at each position.

[0023] Optionally, the radioactivity distribution and the attenuation coefficient of the detector space are solved by alternate iteration based on the ADMM algorithm, including:

[0024] According to the formula The radioactivity distribution x and the detector spatial attenuation coefficient a are solved by alternating iterations.

[0025] Optionally, based on the deep learning network G1, the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution is achieved, including:

[0026] The PET detection data y and the attenuation coefficient a in the detector space are used as input to obtain the scatter-corrected radioactivity distribution x sc As output, the activity distribution x is compared with the accurately scatter-corrected cor By comparison, we optimize the training network parameters θ1 to minimize the loss function L, and finally get x sc , the formula is as follows:

[0027]

[0028] Optionally, based on the deep learning network G2, a mapping of the detector spatial attenuation coefficient to a matching linear attenuation coefficient distribution is implemented, and an object mask matrix is ​​obtained, including:

[0029] Taking the detector spatial attenuation coefficient a as input, the linear attenuation coefficient distribution μ sc And the object mask matrix Mask as output, through the matching CT to obtain the attenuation coefficient distribution μ CT And the accurate object outline Mask drawn by the doctor label By comparison, we optimize the training network parameters θ2 to minimize the loss function L, and finally get μ sc , Mask, the formula is as follows:

[0030]

[0031] Optionally, an unfitted scattering estimate is obtained by single scattering simulation based on the accurate radioactivity distribution, the matching linear attenuation coefficient distribution and the object mask matrix, including:

[0032] The unfitted scattering distribution is obtained according to the following formula:

[0033]

[0034] Among them, R BS Represents the distance from the scattering point to the detector B, R AS represents the distance from the scattering point to the detector A; s is the distance from the annihilation point to the scattering point; I A It means that the positron annihilates at any point on S1 and emits a pair of γ photons. One photon moves along path S1 without scattering and is detected by detector A. The photon energy is 511 keV and the detection efficiency is ε AS , the linear attenuation coefficient image is μ, another photon moves along path S2 after being scattered at point S and is detected by detector B. The photon energy is less than 511 keV, and the detection efficiency is ε′ BS , the linear attenuation coefficient image is μ′; I B It means that the positron and the electron annihilate at any point on S2 to emit a pair of γ photons. One photon moves along the path S2 without scattering and is detected by detector B. The photon energy is 511keV and the detection efficiency is ε BS , the linear attenuation coefficient image is μ, another photon moves along path S1 after being scattered at point S and is detected by detector A. The photon energy is less than 511 keV, and the detection efficiency is ε′ AB , the linear attenuation coefficient image is μ′; Vs represents the total scattering volume, σ AS represents the geometric cross section of detector A along the γ-ray, σ BS represents the geometric cross section of detector B along ray γ; σ c represents the Compton scattering cross section, represents the differential cross section of Compton scattering, which can be obtained by the Klein-Nishina formula, and Ω represents the scattering solid angle.

[0035] Optionally, keeping the activity distribution and the detector spatial attenuation coefficient constant, using the unfitted scattering estimate, maximizing the log-likelihood function, and obtaining the scattering correction factor include:

[0036] Keeping the radioactivity distribution and the detector spatial attenuation coefficient constant, according to the formula The scatter correction factor α is obtained, where

[0037] In a second aspect, the present invention provides a scatter correction system for PET images, comprising:

[0038] Modeling module, which models the PET acquisition process;

[0039] Iterative module, which solves the radioactivity distribution and the attenuation coefficient of the detector space by alternating iteration based on the ADMM algorithm;

[0040] The first mapping module, based on the deep learning network G1, realizes the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution;

[0041] The second mapping module, based on the deep learning network G2, realizes the mapping of the detector spatial attenuation coefficient to the matching linear attenuation coefficient distribution and obtains the object mask matrix;

[0042] Single scattering simulation module, which obtains unfitted scattering estimates through single scattering simulation based on accurate radioactivity distribution, matching linear attenuation coefficient distribution and object mask matrix;

[0043] The scatter correction factor determination module keeps the radioactivity distribution and the detector spatial attenuation coefficient as constants, uses the unfitted scatter estimation, maximizes the log-likelihood function, and obtains the scatter correction factor;

[0044] The correction module uses the accurate scattering distribution after fitting the detector spatial attenuation coefficient and the scattering correction factor to obtain a scatter-corrected PET image.

[0045] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed, the method for scatter correction of a PET image as described in any one of the first aspects above is implemented.

[0046] In a fourth aspect, the present invention provides a storage device, comprising a storage medium and a processor, wherein the storage medium stores a computer program, and when the program is executed by the processor, the scatter correction method for a PET image described in any one of the first aspects above is implemented.

[0047] Beneficial Effects

[0048] The beneficial effects of the present invention are as follows: a scatter correction method for PET images of the present invention obtains the attenuation coefficient of the detector space from PET data, and then uses a deep learning network to obtain the matching linear attenuation coefficient distribution and the initial distribution of radioactive activity, and then performs scatter correction calculations, and performs maximum likelihood estimation iteration on the scatter correction factor to obtain the optimal solution, which not only avoids the scattering assessment error caused by inaccurate initial radioactive activity distribution and linear attenuation coefficient distribution, but also improves the stability of scatter correction for high-noise data and greatly reduces the scattering estimation time. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A schematic flow chart of a scatter correction method for PET images provided in an embodiment of the present invention;

[0050] Figure 2A PET reconstructed image obtained by calculating scattering using a traditional algorithm provided in an embodiment of the present invention;

[0051] Figure 3 The embodiment of the present invention provides a PET reconstructed image obtained by calculating the scattering distribution using the present invention. DETAILED DESCRIPTION

[0052] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0053] The present invention first obtains the attenuation coefficient of the detector space from the PET data, and then uses a deep learning network to obtain the matching linear attenuation coefficient distribution and the initial distribution of radioactivity, respectively. The two are then used for scattering correction calculations. Finally, the optimal scattering correction factor is iteratively solved through the PET detection data to obtain an accurate scattering distribution through fitting.

[0054] First, since the SSS calculation process is highly dependent on the accuracy of the initial activity distribution, especially for patients with heavier weight, more scattering iterations are usually required, which often takes a long time and still cannot obtain a good activity distribution. This patent uses a deep learning network to obtain a matching linear attenuation coefficient distribution and an accurate initial activity distribution for scattering calculations, which not only makes the SSS calculation more accurate and improves image quality, but also greatly shortens the calculation time of scattering correction.

[0055] Secondly, extracting attenuation information directly from PET acquisition data without relying on other modal imaging can effectively correct attenuation artifacts caused by mismatching data from different modalities, thereby broadening the application range of PET imaging. By applying PET acquisition data to obtain the attenuation coefficient of the detector space, it is ensured that the PET distribution and attenuation distribution are strictly matched, effectively eliminating motion artifacts and improving image quality.

[0056] Finally, in the process of fitting the scattering correction factor, this method models the scattering process and uses all the data for iterative solution. Compared with the traditional tail fitting method, it has higher redundancy for the noise of the collected data.

[0057] The present invention only utilizes PET data to more accurately estimate the scattering component for PET reconstruction, which not only improves the applicability and accuracy for patients with large body weight, but also no longer relies on other modality images, while improving the robustness of scatter correction to data noise, so that more accurate PET images can be obtained in clinical acquisition, eliminating scatter correction artifacts.

[0058] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0059] First, refer to Figure 1 This embodiment provides a scatter correction method for a PET image, comprising:

[0060] S1, Modeling the PET acquisition process.

[0061] S2, based on the ADMM algorithm, alternately iterates to solve the radioactivity distribution and the attenuation coefficient in the detector space.

[0062] S3, based on the deep learning network G1, realizes the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution.

[0063] S4, based on the deep learning network G2, realizes the mapping of the detector spatial attenuation coefficient to the matching linear attenuation coefficient distribution, and obtains the object mask matrix.

[0064] S5, unfitted scattering estimates are obtained through single scattering simulation based on the accurate radioactivity distribution, matching linear attenuation coefficient distribution and object mask matrix.

[0065] S6, keeping the radioactivity distribution and the detector spatial attenuation coefficient constant, using the unfitted scattering estimate, maximizes the log-likelihood function and obtains the scattering correction factor.

[0066] S7, using the accurate scattering distribution fitted by the detector spatial attenuation coefficient and the scattering correction factor to obtain a scattering-corrected PET image.

[0067] Alternatively, the PET acquisition process can be modeled as follows:

[0068]

[0069] In the formula y=[y 1t ,y 2t ,…,y it ,…,y NT ]' represents the detected data, represents the average value of the detection data, N represents the size of the detection data sinusoidal graph, T represents the size of the TOF discrete space, i represents the variable index of the detection data sinusoidal response line LOR (line of response), and t represents the variable index of the TOF discrete space. The single quote superscript represents the matrix transposition operation. x=[x1,x2,…,x j ,…,x M ]' represents the unknown radioactivity distribution image, M represents the size of the radioactivity distribution image space, j represents the variable index of the radioactivity distribution image space, and represents the point source at the corresponding spatial position. a=[a1,a2,…,a j …,a N] ' represents the attenuation coefficient of the detector space. A=[A ijt ] is the system matrix, which expresses the probability that a point source j in the PET system is detected by a response line LOR i and has a flight time TOF of t in mathematical form, reflecting the physical characteristics of the system. 1t ,s 2t ,…,s NT ]' represents the estimated value of scattering, α = [α 1t ,α 2t ,…,α NT ]' represents the correction factor of the scattering estimation at each position, which is used to ensure that the scattering component estimated by the SSS method is consistent with the scattering component of the actual collected data.

[0070] PET detection data follows Poisson distribution, and the unknowns are the radioactivity distribution x, the attenuation coefficient a in the detector space, and the scatter correction factor α. The log-likelihood function of the detection data is expressed as:

[0071]

[0072] Substituting formula (1) into formula (2) and ignoring terms unrelated to the unknown, the log-likelihood function can be written as:

[0073]

[0074] Using PET detection data to estimate the attenuation coefficient a in the detector space, ignoring the scattered component, formula (3) becomes:

[0075]

[0076] Optionally, the radioactivity distribution and the attenuation coefficient of the detector space are solved by alternate iteration based on the ADMM algorithm, including:

[0077] According to the formula The radioactivity distribution x and the detector spatial attenuation coefficient a are solved by alternating iterations.

[0078] The Alternating Direction Method of Multipliers (ADMM) algorithm is used to iteratively solve the radioactivity distribution x and the attenuation coefficient a in the detector space. First, the attenuation coefficient a in the detector space is guaranteed to be a constant, and the log-likelihood function is maximized, which is the traditional MLEM algorithm:

[0079]

[0080] In the formula, n represents the current iteration number.

[0081] The initial value of x in formula (5) can be set to a constant or an image obtained by other reconstruction methods or an uncorrected PET image. In this application, it is set to a constant distribution within the imaging field of view, and the constant is selected as 1000.

[0082] Then ensure that the radioactivity distribution x is a constant, maximize the log-likelihood function for the attenuation coefficient a of the unknown detector space, and take the derivative of the attenuation coefficient a of the detector space, that is, formula (6):

[0083]

[0084] Keep all variables non-negative and monotonic:

[0085]

[0086] Substituting formula (6) into formula (7), we get:

[0087]

[0088] In the formula, n represents the current iteration number.

[0089] The initial value of a in formula (8) is usually set to a constant. Since the evaluation algorithm in the formula is only related to the radioactivity distribution, the detector spatial attenuation coefficient obtained by this method strictly matches the radioactivity distribution.

[0090] Optionally, based on the deep learning network G1, the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution is achieved, including:

[0091] The PET detection data y and the attenuation coefficient a in the detector space are used as input to obtain the scatter-corrected radioactivity distribution x sc As output, the activity distribution x is compared with the accurately scatter-corrected corBy comparison, we optimize the training network parameters θ1 to minimize the loss function L, and finally get x sc , the formula is as follows:

[0092]

[0093] Optionally, based on the deep learning network G2, a mapping of the detector spatial attenuation coefficient to a matching linear attenuation coefficient distribution is implemented, and an object mask matrix is ​​obtained, including:

[0094] Taking the detector spatial attenuation coefficient a as input, the linear attenuation coefficient distribution μ sc And the object mask matrix Mask as output, through the matching CT to obtain the attenuation coefficient distribution μ CT And the accurate object outline Mask drawn by the doctor label By comparison, we optimize the training network parameters θ2 to minimize the loss function L, and finally get μ sc , Mask, the formula is as follows:

[0095]

[0096] In order to reduce the calculation time of scatter correction and increase the accuracy of scatter correction, this application uses two deep learning networks to obtain the linear attenuation coefficient distribution μ sc and the activity distribution x for scatter correction sc First, network G1 is used to realize the PET detection data y, the detector spatial attenuation coefficient a to x sc Secondly, we get a through formula (8), and use network G2 to realize the mapping from a to μ sc mapping, and obtain the accurate object mask matrix Mask.

[0097] First, the PET detection data y and the attenuation coefficient a in the detector space are used as input to obtain the scatter-corrected radioactivity distribution x sc As output, the activity distribution x is compared with the accurately scatter-corrected cor By comparison, we optimize the training network parameters θ1 to minimize the loss function L, and finally get x sc ,Right now:

[0098]

[0099] Then take the detector spatial attenuation coefficient a as input, and the linear attenuation coefficient distribution μ sc And the object mask matrix Mask as output, through the matching CT to obtain the attenuation coefficient distribution μ CT And the accurate object outline Mask drawn by the doctor label By comparison, we optimize the training network parameters θ2 to minimize the loss function L, and finally get μsc , Mask, that is:

[0100]

[0101] Mask label It represents the patient contour drawn by a doctor or professional, which is used for network training to obtain contour information.

[0102] The training data set can come from simulation or actual acquisition. The training data set needs to be preprocessed and screened to ensure that the PET and CT images do not have truncation or motion artifacts and can be fully matched. For large patient data, accurate scatter correction should also be ensured, and inaccurate scatter correction should not occur due to too few scatter calculation iterations.

[0103] The deep learning network G can select a CNN network, a Unet network, a GAN network or other networks.

[0104] Optionally, an unfitted scattering estimate is obtained by single scattering simulation based on the accurate radioactivity distribution, the matching linear attenuation coefficient distribution and the object mask matrix, including:

[0105] The unfitted scattering distribution is obtained according to the following formula:

[0106]

[0107] Among them, R BS Represents the distance from the scattering point to the detector B, R AS represents the distance from the scattering point to the detector A; s is the distance from the annihilation point to the scattering point; I A It means that the positron annihilates at any point on S1 and emits a pair of γ photons. One photon moves along path S1 without scattering and is detected by detector A. The photon energy is 511 keV and the detection efficiency is ε AS , the linear attenuation coefficient image is μ, another photon moves along path S2 after being scattered at point S and is detected by detector B. The photon energy is less than 511 keV, and the detection efficiency is ε′ BS , the linear attenuation coefficient image is μ′; I B It means that the positron and the electron annihilate at any point on S2 to emit a pair of γ photons. One photon moves along the path S2 without scattering and is detected by detector B. The photon energy is 511keV and the detection efficiency is ε BS , the linear attenuation coefficient image is μ, another photon moves along path S1 after being scattered at point S and is detected by detector A. The photon energy is less than 511 keV, and the detection efficiency is ε′ AB , the linear attenuation coefficient image is μ ′ ; Vs represents the total scattering volume, σ ASrepresents the geometric cross section of detector A along the γ-ray, σ BS represents the geometric cross section of detector B along ray γ; σ c represents the Compton scattering cross section, represents the differential cross section of Compton scattering, which can be obtained by the Klein-Nishina formula, and Ω represents the scattering solid angle.

[0108] In the process of PET image reconstruction, the single scatter simulation (SSS) method is usually used to estimate the scattering distribution. The general idea of ​​SSS is: first determine all possible scattering points, and then calculate the scattering value of each scattering point on the broken line formed by any two crystals. After traversing all combinations of scattering points and crystal pairs, the scattering value on each crystal pair is calculated. Therefore, the number of scattering points directly affects the calculation time and accuracy of scattering correction. The number of scattering points is determined by the size of the object. The larger the object, the more scattering points there are, and the longer the time required for scattering correction, and vice versa. When performing scattering correction, the image of other modalities is usually used to approximate the maximum circumscribed cube range as the object range for scattering correction. However, this range often contains more scattering points than the number of scattering points of the actual object, which not only increases the calculation time of scattering, but also reduces the accuracy of scattering correction. Obtaining the accurate contour range of the object through deep learning can significantly reduce the number of scattering points required for calculation and reduce the calculation time. Further using the network to obtain accurate radioactivity distribution can greatly reduce the calculation time of scattering iteration for patients with heavier weight, while the scattering estimation is more accurate and the image quality is greatly improved.

[0109] Scattering estimation uses the radioactivity distribution and the linear attenuation coefficient distribution to calculate the scattering distribution based on the analytical method. It can be seen from formula (11) that SSS needs to integrate the radioactivity distribution and the linear attenuation coefficient distribution, so it is insensitive to the details of the linear attenuation coefficient distribution, the resolution requirement is reduced, and the robustness is higher. The scattering distribution estimated by SSS can be expressed as

[0110]

[0111] Optionally, keeping the activity distribution and the detector spatial attenuation coefficient constant, using the unfitted scattering estimate, maximizing the log-likelihood function, and obtaining the scattering correction factor include:

[0112] Keeping the radioactivity distribution and the detector spatial attenuation coefficient constant, according to the formula The scatter correction factor α is obtained, where

[0113] Considering the influence of multiple scattering, in order to determine the proportional relationship between the unfitted scattering component and the actual scattering component, it is necessary to fit it, that is, to solve the scattering correction factor α.

[0114] Formula (1) is written in matrix form as shown in formula (12):

[0115] Y=A×x+S×α (12)

[0116] A is the system matrix, S is the diagonal matrix It can be seen from formula (12) that x and α are symmetrical in the formula. Therefore, the iterative algorithm of the unknown x can be used to solve the unknown α. Therefore, formula (12) can be simplified to the following formula (13), that is, the radioactivity distribution x is kept, the attenuation coefficient a in the detector space is constant, the log-likelihood function is maximized for the unknown scatter correction factor α, and the new scatter correction factor α is directly calculated using PET data. The corresponding algorithm is:

[0117]

[0118] in

[0119] The initial value of α in formula (13) can be set to a constant or a scattering correction factor distribution obtained by other fitting methods. In this application, it is set to a constant of 1.

[0120] In the implementation process of the present invention, firstly, in order to obtain the attenuation coefficient a of the detector space, the objective function is maximized for the radioactivity distribution x, that is, the traditional MLEM iterative reconstruction algorithm, and then the radioactivity distribution x is kept constant, and the objective function is maximized for the attenuation coefficient a of the unknown detector space to obtain the attenuation coefficient a of the detector space; secondly, two deep learning networks are applied to respectively obtain the activity distribution x for scatter correction sc , linear attenuation coefficient distribution μ sc and the object edge information Mask; again, the radioactivity distribution x sc , linear attenuation coefficient distribution μ sc The object edge information Mask is directly applied to the subsequent SSS calculation to evaluate the unfitted scattering distribution, which not only increases the accuracy of the scattering distribution evaluation but also reduces the scattering time. Finally, the radioactivity distribution x and the attenuation coefficient a in the detector space are kept constant, and the objective function is maximized for the scattering correction factor α. The unfitted scattering distribution is used for iterative calculation, and the scattering correction factor is continuously corrected to make it close to the actual situation, so as to obtain the scattering correction factor that meets the requirements of the maximized objective function. Finally, the matched attenuation correction and accurate scattering estimation are applied to the reconstruction to obtain high-quality, artifact-free PET images.

[0121] Figure 2 For PET reconstructed images obtained by calculating scatter using traditional algorithms, the patient’s body is obviously white due to the large body weight and small amount of data, and the scatter assessment is inaccurate; Figure 3 In order to calculate the scattering distribution using the patented method, the PET images obtained under the same reconstruction method and parameters have significantly improved the abdominal whitening phenomenon, more accurate scattering distribution, and better image quality.

[0122] The present invention proposes a method for scatter correction of PET images, which overcomes the scatter estimation error caused by inaccurate initial activity distribution, and also avoids the serious time-consuming problem caused by multiple scattering iterative calculations; secondly, the method no longer relies on other modality images, and can obtain a more accurate scatter distribution even in the absence of other modality images or in the presence of obvious artifacts in these images; finally, an iterative algorithm is used to calculate the scatter correction factor, which solves the problem of insufficient stability when the amount of data is small and the patient is heavy. In addition, this method significantly reduces the patient's radiation dose while significantly reducing the influence of the patient's body shape on the scatter estimation, and the estimation accuracy of the scatter distribution is higher, so it has stronger applicability. Compared with traditional scatter correction algorithms, the method proposed by the present invention has higher correction accuracy and takes less time, which helps to further improve image quality and reduce reconstruction time.

[0123] The following is a specific implementation case of clinical data collection to further illustrate the scatter correction method of a PET image proposed by the present invention:

[0124] The patient was scanned with a PET device. The PET images were reconstructed using the ordered subsets expectation-maximization (OSEM) method with 17 subsets, 2 iterations, an image matrix of 192x192, and a Gaussian filter half-height width of 4 mm.

[0125] Deep learning network model parameters: the network structure is 3D Unet, the depth is 4, the input size is 32x 144x144, the number of channels is 2, the initial number of feature channels is 64, as the depth increases, downsampling is performed, and the number of feature channels doubles, the loss function is set to L1+SSIM, the batch size is set to 8, the learning rate is set to 1e-3, and the Adam optimizer is used to optimize the network parameters.

[0126] x is set to the default initial value of 1000, a is set to the initial value of 1, α is set to the initial value of 1, the number of iterations of the attenuation coefficient in the detector space is 5, the number of iterations of the scatter correction factor is 3, and the reconstruction results are shown in Figure 2 and Figure 3 . Figure 2In order to calculate the PET reconstructed images obtained by scattering using traditional algorithms, the patient has a large body weight and a small amount of data, and there are obvious whitening artifacts on the patient's body; Figure 3 In order to calculate the scattering distribution using the patented method, the PET images obtained under the same reconstruction method and parameters show a significant improvement in the whitening phenomenon of the abdomen of patients with heavy weight.

[0127] In a second aspect, the present embodiment provides a scatter correction system for PET images, comprising: a modeling module for modeling the PET acquisition process; an iteration module for alternately iterating and solving the radioactivity distribution and the attenuation coefficient of the detector space based on the ADMM algorithm; a first mapping module for mapping the PET detection data and the detector space attenuation coefficient to the accurate radioactivity distribution based on the deep learning network G1; a second mapping module for mapping the detector space attenuation coefficient to the matching linear attenuation coefficient distribution based on the deep learning network G2, and obtaining an object mask matrix; a single scattering simulation module for obtaining an unfitted scattering estimate through single scattering simulation based on the accurate radioactivity distribution, the matching linear attenuation coefficient distribution and the object mask matrix; a scattering correction factor determination module for keeping the radioactivity distribution and the detector space attenuation coefficient as constants, and using the unfitted scattering estimate to maximize the log-likelihood function to obtain a scattering correction factor; a correction module for obtaining a scatter-corrected PET image using the accurate scattering distribution fitted with the detector space attenuation coefficient and the scattering correction factor. According to a PET image scattering correction system provided by this embodiment, since it is used to implement the steps of a PET image scattering correction method provided by the first aspect of the present invention, the PET image scattering correction system has all the technical effects of the PET image scattering correction method, which will not be repeated here.

[0128] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed, the method for scatter correction of a PET image as described in any one of the first aspects above is implemented.

[0129] In a fourth aspect, an embodiment of the present invention provides a storage device, including a storage medium and a processor, wherein the storage medium stores a computer program, and when the program is executed by the processor, the scatter correction method for a PET image described in any one of the first aspects above is implemented.

[0130] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention should also include these modifications and variations.

[0132] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A scatter correction method for PET images, characterized in that: include: Modeling the PET acquisition process; Based on the ADMM algorithm, the radioactivity distribution and the attenuation coefficient of the detector space are solved by alternating iteration; Based on the deep learning network G1, the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution is achieved; Based on the deep learning network G2, the mapping of the detector spatial attenuation coefficient to the matching linear attenuation coefficient distribution is realized, and the object mask matrix is ​​obtained; Unfitted scattering estimates are obtained through single scattering simulations based on accurate radioactivity distribution, matching linear attenuation coefficient distribution and object mask matrix; Keeping the radioactivity distribution and the detector spatial attenuation coefficient constant, the scatter correction factor is obtained by maximizing the log-likelihood function using the unfitted scatter estimation; The accurate scatter distribution after fitting the detector spatial attenuation coefficient and scatter correction factor is used to obtain the scatter-corrected PET image.

2. A scatter correction method for PET images according to claim 1, characterized in that: The PET acquisition process can be modeled as follows: In the formula y=[y 1t ,y 2t ,…,y it ,…,y NT ]' represents the detected data, represents the average value of the detection data, N represents the size of the detection data sinusoidal graph, T represents the size of the TOF discrete space, i represents the variable index of the detection data sinusoidal graph response line, and t represents the variable index of the TOF discrete space; the single quote superscript represents the matrix transposition operation; x = [x1, x2, ..., x j ,…,x M ]' represents the unknown radioactivity distribution image, M represents the size of the radioactivity distribution image space, j represents the variable index of the radioactivity distribution image space, and represents the point source at the corresponding spatial position; a = [a1, a2, ..., a j …,a N] ' represents the attenuation coefficient of the detector space; A = [A ijt ] is the system matrix; S = [s 1t ,s 2t ,…,s NT ]' represents the estimated value of scattering, α = [α 1t ,α 2t ,…,α NT ]' represents the correction factor for the scatter estimation at each position.

3. The scatter correction method for PET images according to claim 2, characterized in that: Based on the ADMM algorithm, the radioactivity distribution and the attenuation coefficient of the detector space are solved by alternating iteration, including: According to the formula The radioactivity distribution x and the detector spatial attenuation coefficient a are solved by alternating iterations.

4. The scatter correction method for PET images according to claim 3, characterized in that: Based on the deep learning network G1, the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution is realized, including: The PET detection data y and the attenuation coefficient a in the detector space are used as input to obtain the scatter-corrected radioactivity distribution x sc As output, the activity distribution x is compared with the accurately scatter-corrected cor By comparison, we optimize the training network parameters θ1 to minimize the loss function L, and finally get x sc , the formula is as follows:

5. The scatter correction method for PET images according to claim 4, characterized in that: Based on the deep learning network G2, the mapping of the detector spatial attenuation coefficient to the matching linear attenuation coefficient distribution is realized, and the object mask matrix is ​​obtained, including: Taking the detector spatial attenuation coefficient a as input, the linear attenuation coefficient distribution μ sc And the object mask matrix Mask as output, through the matching CT to obtain the attenuation coefficient distribution μ CT And the accurate object outline Mask drawn by the doctor label By comparison, we optimize the training network parameters θ2 to minimize the loss function L, and finally get μ sc , Mask, the formula is as follows:

6. The scatter correction method for PET images according to claim 5, characterized in that: Based on the accurate radioactivity distribution, matching linear attenuation coefficient distribution and object mask matrix, the unfitted scattering estimate is obtained through single scattering simulation, including: The unfitted scattering distribution is obtained according to the following formula: Among them, R BS Represents the distance from the scattering point to the detector B, R AS represents the distance from the scattering point to the detector A; s is the distance from the annihilation point to the scattering point; I A It means that the positron annihilates at any point on S1 and emits a pair of γ photons. One photon moves along path S1 without scattering and is detected by detector A. The photon energy is 511 keV and the detection efficiency is ε AS , the linear attenuation coefficient image is μ, another photon moves along path S2 after being scattered at point S and is detected by detector B. The photon energy is less than 511 keV, and the detection efficiency is ε ′ BS , the linear attenuation coefficient image is μ ′ ;I B It means that the positron and the electron annihilate at any point on S2 to emit a pair of γ photons. One photon moves along the path S2 without scattering and is detected by detector B. The photon energy is 511keV and the detection efficiency is ε BS , the linear attenuation coefficient image is μ, another photon moves along path S1 after being scattered at point S and is detected by detector A. The photon energy is less than 511 keV, and the detection efficiency is ε ′ AB , the linear attenuation coefficient image is μ ′ ; Vs represents the total scattering volume, σ AS represents the geometric cross section of detector A along the γ-ray, σ BS represents the geometric cross section of detector B along ray γ; σ c represents the Compton scattering cross section, represents the differential cross section of Compton scattering, which can be obtained by the Klein-Nishina formula, and Ω represents the scattering solid angle.

7. The scatter correction method for PET images according to claim 6, characterized in that: Keeping the radioactivity distribution and the detector spatial attenuation coefficient constant, using the unfitted scattering estimate, maximizing the log-likelihood function, the scattering correction factor is obtained, including: Keeping the radioactivity distribution and the detector spatial attenuation coefficient constant, according to the formula The scatter correction factor α is obtained, where 8. A scatter correction system for PET images, characterized in that: include: Modeling module, which models the PET acquisition process; Iterative module, which solves the radioactivity distribution and the attenuation coefficient of the detector space by alternating iteration based on the ADMM algorithm; The first mapping module, based on the deep learning network G1, realizes the mapping of PET detection data and detector spatial attenuation coefficient to accurate radioactivity distribution; The second mapping module, based on the deep learning network G2, realizes the mapping of the detector spatial attenuation coefficient to the matching linear attenuation coefficient distribution and obtains the object mask matrix; Single scattering simulation module, which obtains unfitted scattering estimates through single scattering simulation based on accurate radioactivity distribution, matching linear attenuation coefficient distribution and object mask matrix; The scatter correction factor determination module keeps the radioactivity distribution and the detector spatial attenuation coefficient as constants, uses the unfitted scatter estimation, maximizes the log-likelihood function, and obtains the scatter correction factor; The correction module uses the accurate scattering distribution after fitting the detector spatial attenuation coefficient and the scattering correction factor to obtain a scatter-corrected PET image.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the scatter correction method for a PET image as claimed in any one of claims 1 to 7 is implemented.

10. A storage device comprising a storage medium and a processor, wherein the storage medium stores a computer program, characterized in that: When the processor executes the computer program, the scatter correction method for a PET image according to any one of claims 1 to 7 is implemented.

Images