CBCT scattering adaptive correction method based on deep learning

By using deep learning technology to adaptive correction of beam barriers in CBCT system, the problems of low accuracy of CBCT scattering correction and time-consuming parameter optimization are solved, and efficient and stable improvement of CBCT image quality is achieved.

CN119006639APending Publication Date: 2024-11-22ZHEJIANG NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411150897.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing CBCT scattering correction technology has problems such as low accuracy, time-consuming and labor-intensive optimization of parameters, and inability to adapt, especially when imaging in large field of vision areas.

Method used

Adaptive correction method of CBCT scattering based on deep learning is adopted. By scanning the front and back of the beam blocking plate, the projection data is obtained, the grating area and void area data are divided, and the deep learning local filtering model is input to optimize the adaptive parameters, and the final scattering estimate data is calculated and The correction is carried out, and the projection data is adaptively completed, and the corrected CBCT reconstruction image is finally obtained through noise reduction processing.

Benefits of technology

The CBCT scattering correction accuracy and image quality are improved, and the long-term process of iterative optimization of parameter selection is avoided, ensuring the stability and accuracy of the correction effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006639A_ABST
    Figure CN119006639A_ABST
Patent Text Reader

Abstract

The invention relates to the field of CBCT imaging, and aims to provide a CBCT scattering adaptive correction method based on deep learning. Comprising the following steps: obtaining projection data of an independent beam barrier plate and combined with scanning of a measured object, performing segmentation to obtain measurement scattering data and measurement total projection data, inputting the measurement scattering data and the measurement total projection data into a local filtering scattering estimation model based on deep learning, outputting adaptive optimization parameters, and performing calculation to obtain final scattering estimation data; meanwhile, global total projection data obtained through self-adaptive completion of the measurement total projection data in the scattering estimation process is reserved; and subtracting the final scattering estimation data from the global total projection data to obtain scattering correction projection data, and reconstructing after noise reduction to obtain a corrected CBCT image. According to the method, the scattering estimation result based on the specific angle and the internal structure of the object can be efficiently and accurately obtained, the problems that parameter selection is time-consuming and labor-consuming and cannot be adaptively determined, and the empirical optimization parameter correction effect is unstable are effectively solved, and the scattering correction precision is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a CBCT image correction method based on a beam blocking plate and deep learning, and belongs to the field of X-ray CBCT imaging. Background Art

[0002] Cone beam computed tomography (CBCT) systems have the advantages of easy integration, single-circle three-dimensional volumetric imaging, and high cost-effectiveness. They are widely used in image-guided radiotherapy, oral imaging, extremity imaging, intraoperative diagnosis, and other specialized fields. However, the artifacts caused by X-ray scattering reduce the ability to resolve the details of anatomical structures in the image. The CT values ​​of reconstructed images that have not been scatter-corrected deviate significantly from the expected results, which seriously affects clinical diagnosis and quantitative measurement. Therefore, CBCT scatter artifact correction has always been a challenging problem that needs to be solved in the industry.

[0003] At present, many scholars at home and abroad have studied the scattering artifacts and correction in CBCT and proposed various methods. Hardware correction methods mainly include collimator correction, filter plate correction, anti-scattering grid correction, air gap correction, etc. Although the classic X-ray collimator, filter plate and other hardware correction methods will increase the original ray ratio, they cannot completely remove the scattering, especially not suitable for CBCT imaging of large field of view (such as chest), which is far from clinical practice. Software correction methods mainly include Monte Carlo simulation method, scattering modeling method, etc. Monte Carlo simulation method obtains scattering information by simulating the motion trajectory of a large number of photons. This method can theoretically calculate the true value of scattering, but it needs to simulate the actual physical process, resulting in large amount of calculation and slow simulation speed. Software and hardware hybrid correction methods mainly include primary ray modulation correction method, scattering correction plate method, etc. However, the use of modulators may introduce additional beam hardening artifacts. The conventional scattering correction plate method requires the irradiated object to be irradiated twice, the scanning time is too long, and the patient's additional radiation exposure is increased.

[0004] Taking advantage of the redundancy of CT scan data, researchers have proposed a method for scatter correction based on different beam blocks (such as full beam blocks, finger cross blocks, and rotating blocks) and through a single scan. However, this type of method can usually only obtain scatter data in a partial area, and it is necessary to combine the scatter estimation model to obtain the scatter data of the entire projection. At the same time, there is a problem of missing object projection in the grating projection area of ​​the blocker. The scatter correction method based on local filtering technology improves imaging performance by using a more flexible data processing framework. It does not require downsampling of projection data or unifying the modulation frequency and amplitude. It can also simulate the effect of beam hardening on modulation to improve the accuracy of scatter estimation. It has broad application prospects in scatter correction based on beam blocks. However, the parameters of the scatter estimation model are usually determined by experience and need to be optimized through continuous testing, which is time-consuming and labor-intensive. In addition, these parameters are usually set to constant values ​​for all projections under the same acquisition, while the photon transmission path and scattering probability at different angles are different, so the accuracy of scatter estimation cannot be guaranteed.

[0005] In view of the defects of the existing CBCT scatter correction technology, the present invention intends to achieve higher CBCT scatter correction accuracy based on a new solution to obtain better CBCT image quality. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a CBCT scatter adaptive correction method based on deep learning.

[0007] To solve the technical problem, the solution of the present invention is:

[0008] A CBCT scatter adaptive correction method based on deep learning is provided, comprising the following steps:

[0009] (1) A beam stopper is placed in front of the radiation source of the CBCT imaging device, and then the object to be measured is scanned to obtain projection data A; after the object to be measured is removed, the scan is performed again to obtain projection data B with the beam stopper retained alone;

[0010] (2) According to different projection positions, the projection is divided into a grating area and a gap area between the gratings; the projection data A and the projection data B are divided according to this rule to obtain the measured scattered data corresponding to the grating area and the measured total projection data corresponding to the gap area;

[0011] (3) inputting the measured data obtained in step (2) into a local filtering scattering estimation model based on deep learning, outputting adaptively optimized parameters, and calculating and obtaining final scattering estimation data; at the same time, retaining the global total projection data obtained by adaptively completing the measured total projection data during the scattering estimation process;

[0012] (4) subtracting the final scatter estimation data from the global total projection data to obtain scatter correction projection data;

[0013] (5) The scatter-corrected projection data is subjected to denoising, and then a reconstruction algorithm is used to obtain the corrected CBCT reconstructed image.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. Based on the application of deep learning networks and the setting of multiple sub-networks corresponding to parameters, the present invention can efficiently and accurately obtain scattering estimation results based on specific angles and internal structures of objects, effectively avoiding the problems of existing scattering correction methods that are time-consuming and labor-intensive when selecting parameters for scattering estimation models and cannot be adaptively determined, and the correction effect of empirically optimized parameters is unstable, thereby improving the accuracy of CBCT scattering correction.

[0016] 2. While realizing adaptive scatter estimation and correction, the present invention can also adaptively complete the object projection information in the blocking plate grating projection area, thereby improving the scatter correction accuracy and reconstructed image quality of CBCT.

[0017] 3. The present invention makes full use of the technical characteristics of deep learning that can extract local and global features from data sets, and establishes a new method for determining parameters of local filtering scattering estimation; at the same time, the introduction of a selector to adaptively train the subnetwork corresponding to each parameter can efficiently and accurately obtain scattering estimation results based on specific angles and internal structures of objects. Therefore, this method avoids the long iterative optimization process of parameter selection in the prior art, as well as the difficulty in adaptively determining parameters and the unstable correction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the structure of the beam stop plate.

[0019] Figure 2 It is a projection diagram of CBCT imaging based on a cross-barrier plate.

[0020] Figure 3 It is a flowchart of CBCT scatter adaptive correction based on deep learning.

[0021] Figure 4 It is a parameter adaptive selection framework for local filtering scattering estimation models based on deep learning networks. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0023] Part I Implementation of the Invention

[0024] The CBCT scatter adaptive correction method based on deep learning of the present invention comprises the following steps:

[0025] 1. Place a beam stop plate in front of the radiation source of the CBCT imaging device, then scan the object to be measured to obtain projection data A; remove the object to be measured and scan again to obtain projection data B with the beam stop plate alone.

[0026] The beam blocking plate used in the present invention can be any one of a full beam blocking plate, a finger crossing blocking plate or a rotating blocking plate. The projection data B plays an important role in scatter estimation, empirical compensation of beam hardening effect, etc. Since the acquisition is only projected on the beam blocking plate, the radiation dose of the object under test will not be increased.

[0027] 2. According to different projection positions, the projection data is divided into a grating area and a gap area between the gratings; the projection data A and the projection data B are divided according to this rule to obtain the measured scattered data corresponding to the grating area and the measured total projection data corresponding to the gap area;

[0028] The projection data segmentation method can be threshold segmentation method, edge segmentation method or deep learning based segmentation method.

[0029] 3. Input the measurement data obtained in the previous step into the local filtering scattering estimation model based on deep learning, output the adaptively optimized parameters, and calculate the final scattering estimation data; at the same time, retain the global total projection data obtained by adaptively completing the measured total projection data during the scattering estimation process; this part is the core innovation content of the present invention.

[0030] (3.1) Different from the conventional practice in the prior art of determining parameters empirically and then repeatedly testing and optimizing, the present invention obtains the parameters of the local filtering scattering estimation model based on a deep reinforcement learning network, thereby achieving adaptive optimization of the model parameters and outputting the calculation results.

[0031] The deep reinforcement learning network specifically refers to a deep Q learning network based on a Q learning framework, which is composed of n sub-networks with the same architecture, each of which is constructed by a convolutional neural network CNN (Convolutional Neural Networks), and n is the number of scattering estimation parameters to be determined; during its operation, the current state refers to the current scattering estimation data, and the action taken refers to the way of adjusting the scattering estimation parameters, which is to increase, decrease or remain unchanged; the input information is the measured scattering data and measured total projection data of CBCT imaging, and the output is the adaptive optimization parameters of the scattering estimation model.

[0032] For example, a deep reinforcement learning network consists of n sub-networks with the same architecture W = {Wi}, i represents η,γ,U,σ,k b , β, etc., a total of n parameters to be adaptively determined. The current state s refers to the current scattering estimation data, and the action a refers to the way to adjust the scattering estimation parameters, which can be increased, decreased, or unchanged. The reward r t By calculating the current state s t and the next state s t+1 The structural similarity index or mean absolute percentage error (MAPE) of the scattering data and the reference scattering data is calculated, and then rewards are given according to the size of the index: if the error at the next moment is less than the error at the current moment, the reward is 1; if the error at the next moment is greater than the error at the current moment, the reward is -1; if they are equal, the reward is 0. The reference scattering data can be obtained through Monte Carlo simulation.

[0033] In the training process for parameter selection of the local filtering scattering estimation model, the subnetwork to be run next is determined by a selector to adjust the corresponding parameters, specifically including: after training each subnetwork, calculating the mean absolute percentage error MAPE value between the scattering estimation data and the reference scattering data before and after the parameter update corresponding to the subnetwork, obtaining the MAPE difference before and after each parameter update, and then calculating the average of the MAPE differences obtained multiple times; in the next iteration, training the subnetwork with a larger mean value to ensure that each subnetwork obtains an adaptive number of training times.

[0034] (3.2) If the beam blocking plate used in step 1 is a full beam blocking plate or a cross-finger blocking plate, the projections obtained by a single scan are all blocked projections;

[0035] At this time, the local filtering scattering estimation model is specifically shown as follows:

[0036]

[0037] Among them, s b represents the expected final scattering estimation data; k b is the weight factor obtained through deep learning; s0 represents the original scattering estimation data; ** represents two-dimensional convolution;

[0038] is a Gaussian smoothing function, (k, t) is the position index in the image; σ is the parameter of the Gaussian function, obtained through deep learning;

[0039] f is a binary scattering weighting matrix, which is used to exclude unreliable scattering data, as follows:

[0040]

[0041] Among them, p m Indicates the maximum value of projection data A; is the gradient of the original scattering estimation data; U is the upper limit of the gradient, which is obtained through deep learning.

[0042] The method for obtaining the original scattering estimation data s0 is as follows:

[0043]

[0044] in; is an empirical compensation for the beam hardening effect, used to improve the accuracy of scatter estimation; P m (i, j), I0(i, j) are projection data A and projection data B respectively; η is an empirical parameter for controlling the data correction intensity, which is obtained through deep learning;

[0045] The complementary data representing the measured scattered data and the measured total projection data are obtained by processing the grating area and the gap area of ​​the projection data A and the projection data B, as shown in the following formula:

[0046]

[0047] in is the region segmentation function; Ω h and Ω g They represent the grating area and gap area in the projection respectively, and (i, j) is the position index in the image; is a Gaussian function with parameter γ, which is obtained through deep learning.

[0048] In summary, in the local filtering scattering estimation model based on the full beam blocker or the finger cross blocker, the adaptive optimization parameters include η,γ,U,σ,k b .

[0049] (3.3) If the beam blocking plate used in step 1 is a rotating blocking plate, the projection obtained by a single scan is composed of a blocking projection and an open projection;

[0050] (a) Obtain measured scatter data and measured total projection data by following the operations described in step 2; Obtain the final scatter estimation data s of the blocked projection by following the same calculation method as when using a full beam blocker or a cross-finger blocker (i.e., as shown in step 3.2) b ;

[0051] (b) Based on the final scatter estimation data of the blocked projection, the final scatter estimation data s of the open projection is obtained according to the following formula: p :

[0052]

[0053] in, is a Gaussian function with parameter β, (k, t) is the position index in the image, and β is obtained through deep learning; S b is the scattering estimate data of the occlusion projection.

[0054] In summary, in the local filtering scattering estimation model based on the rotating blocking plate, the adaptively optimized parameters include η,γ,U,σ,k b ,β.

[0055] 4. Subtract the final scatter estimation data from the global total projection data to obtain the scatter correction projection data;

[0056] 5. De-noise the scatter-corrected projection data by filtering or iterative methods; then use a reconstruction algorithm based on filtered back projection, iterative methods, or deep learning networks to obtain the corrected CBCT reconstructed image.

[0057] Part II: A specific application example

[0058] The CBCT scatter adaptive correction method based on deep learning provided by the present invention can be applied to Figure 1 Several beam stoppers are shown in the CBCT imaging environment.

[0059] Figure 1 Shown on the left is a conventional full-beam baffle; Figure 1 The middle one shows a finger-cross blocking plate, which is used to obtain blocking projections during CBCT imaging; Figure 1 The right side shows a rotating blocking plate with a blocking area of ​​a cross-finger structure, in which the blocking area 1-1 and the open area 1-2 are cross-distributed, and is used to alternately obtain blocked projection and open projection during CBCT imaging.

[0060] The specific implementation method of this example is as follows: Figure 3 As shown:

[0061] Step 1: Acquire acquisition data: Based on the CBCT imaging system of the beam blocking plate, a CT scanning experiment is performed to obtain projection data A of the blocking plate and the object imaging and projection data B of the blocking plate imaging respectively.

[0062] Step 2: Local area segmentation: Based on the local filtering technology, this example uses the threshold segmentation method to divide the blocking projection into grating areas Ω h and the gap area Ω g The regional segmentation is synchronized to the projection data B to complete the segmentation of the local area. The projection data A and the projection data B are extracted respectively to obtain the measured scattering data corresponding to the grating area and the measured total projection data corresponding to the gap area.

[0063] Step 3: Scattering data estimation: Input the measured scattering data and the measured total projection data into the local filtering scattering estimation model based on deep learning, define the initial values ​​of the parameters to be determined based on experience and input them into the model, output the adaptively optimized scattering estimation parameters, and obtain the final scattering estimation data. At the same time, retain the global total projection data obtained by adaptively completing the measured total projection data during the scattering estimation process.

[0064] In this example, the deep reinforcement learning network specifically refers to a deep Q learning network based on a Q learning framework, which uses CNN as a function approximator to estimate the Q-value function, and minimizes the quadratic loss function to make the action value function close to the optimal to determine the network parameters.

[0065] In this example, the deep reinforcement learning network consists of multiple sub-networks with the same architecture corresponding to the parameters to be determined W = {W i}, i represents the scattering estimation parameters η,γ,U,σ,k b ,β. Since multiple scattering estimation parameters need to be adaptively determined, the present invention introduces a selector in the deep reinforcement learning network to determine the subnetwork to be run next time to adjust the corresponding parameters. The specific process is: after each subnetwork is trained under random sampling for a certain number of times (such as 4 times), the MAPE between the scattering estimation data and the reference scattering data before and after each update of the subnetwork parameters is calculated and subtracted to obtain the MAPE difference before and after each parameter update, and then the average of multiple MAPE differences is calculated (such as calculating the MAPE difference before and after the parameter update obtained by 4 trainings, and then calculating the average of the MAPE difference. If a subnetwork is trained more than 4 times, the data of the last 4 times is taken). In the next iteration, the subnetwork with a larger mean value is trained. Then the latest MAPE difference is added and the earliest MAPE difference is discarded, and then the average is calculated and compared. This ensures that each subnetwork obtains an adaptive number of training times.

[0066] Step 4: Scatter correction: Subtract the final scatter estimation data from the global total projection data of projection data A to obtain scatter correction data.

[0067] Step 5: CT reconstruction: Use filtering or iterative methods to reduce the noise of the scatter correction data; then use filtered back projection, iterative method or deep learning network to reconstruct and obtain a corrected and reconstructed CT image.

[0068] Specifically, when the beam blocking plate is a conventional full-beam blocking plate and a cross-finger blocking plate, the projection obtained by step 1 is a blocking projection. Figure 2 It is a projection diagram based on the imaging of the cross-finger blocking plate, where the blocking projection includes the measured scattering data blocked by the high atomic number grating, i.e., the data corresponding to the grating area 2-1, and the measured total projection data, i.e., the data corresponding to the gap area 2-2.

[0069] When the beam blocking plate is a rotating blocking plate, the projection obtained by step 1 includes a blocking projection and an open projection. The effect of the blocking projection is the same as the conventional full beam blocking plate and the cross-finger blocking plate imaging, but the open projection only includes open data (open total projection data composed of original data and scattered data), and the scattered data is obtained by filtering the scattering estimation of the blocking projection.

[0070] Figure 4 Shown is the training framework of the deep learning based local filtering scattering estimation model.

[0071] This example is based on finger-crossing barrier imaging and the corresponding five scattering estimation parameters are determined by network adaptation. Its deep reinforcement learning network is based on CNN and implemented using the Keras network framework. Therefore, n = 5 CNN sub-networks with the same architecture are formed. σ ,W γ ,W U ,W η ,W kb}. The specific implementation process is as follows:

[0072] (1) Input the measured scattering data and the measured total projection data of projection data A and projection data B, and define the parameters to be determined η, γ, U, σ, k b , the initial value of ω γ ,ω σ function.

[0073] (2) Based on ω γ The function processes the measured scattering data and the measured total projection data of step 1 respectively to obtain the global scattering data of projection data A and B and global total projection data Based on the obtained global data and the projection data A(P m ), projection data B(I0) and correction parameter η are calculated to obtain the original scattering estimation data s0.

[0074] (3) Based on ω σ The function obtains the scattering estimation data s b , that is, obtaining the initial scattering estimation data.

[0075] (4) Train each CNN sub-network in the deep reinforcement learning network under random sampling, and obtain new parameters after reaching a certain number of times. In this example, it is defined as reaching 4 times, and each parameter is updated 4 times.

[0076] (5) According to steps (2) and (3), the scattering estimation data before and after each parameter update is calculated, and each sub-network obtains scattering estimation data for four parameter updates. The MAPE between each scattering estimation data and the reference scattering data is calculated and subtracted from the MAPE before the current update to obtain the MAPE difference. Then, the average of the four MAPE differences of each sub-network is calculated as the Loss. In this example, there are a total of five parameters to be determined, and five Losses are obtained (Loss1-Loss5). In the next iteration, the sub-network with the larger Loss value will be trained. The reference scattering data is obtained by Monte Carlo simulation.

[0077] (6) After the latest sub-network parameter update, the method described in step 5 is used to obtain the latest MAPE difference, and then the latest MAPE difference is added and the earliest MAPE difference is discarded. The mean of the MAPE difference is calculated as the new Loss, and then compared with other sub-networks. In the next iteration, the sub-network with a larger Loss will be trained. This ensures that each sub-network obtains an adaptive number of training times.

[0078] The iteration stopping condition of this example is to reach 10000 iterations.

[0079] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A CBCT scatter adaptive correction method based on deep learning, characterized in that: The following steps are involved: (1) A beam blocking plate is placed in front of the radiation source of the CBCT imaging device, and then the object to be measured is scanned to obtain projection data A; After removing the object to be measured, scanning is performed again to obtain projection data B with the beam blocking plate retained alone; (2) According to different projection positions, the projection is divided into a grating area and a gap area between the gratings; the projection data A and the projection data B are divided according to this rule to obtain the measured scattered data corresponding to the grating area and the measured total projection data corresponding to the gap area; (3) inputting the measured data obtained in step (2) into a local filtering scattering estimation model based on deep learning, outputting adaptively optimized parameters, and calculating and obtaining final scattering estimation data; at the same time, retaining the global total projection data obtained by adaptively completing the measured total projection data during the scattering estimation process; (4) subtracting the final scatter estimation data from the global total projection data to obtain scatter correction projection data; (5) The scatter-corrected projection data is subjected to denoising, and then a reconstruction algorithm is used to obtain the corrected CBCT reconstructed image.

2. The method according to claim 1, characterized in that In the step (1), the beam blocking plate is a full beam blocking plate, a cross-finger blocking plate or a rotating blocking plate.

3. The method according to claim 1, characterized in that In the step (2), the method for segmenting the projection data is any one of the following: a threshold segmentation method, an edge segmentation method, or a deep learning-based segmentation method.

4. The method according to claim 1, characterized in that If the beam blocking plate used in step (1) is a full beam blocking plate or a cross-finger blocking plate, the projections obtained by a single scan are all blocked projections; At this time, the local filtering scattering estimation model is specifically shown as follows: Among them, s b represents the expected final scattering estimation data; k b is the weight factor obtained through deep learning; s0 represents the original scattering estimation data; ** represents two-dimensional convolution; is a Gaussian smoothing function, (k, t) is the position index in the image; σ is the parameter of the Gaussian function, obtained through deep learning; f is a binary scattering weighting matrix, which is used to exclude unreliable scattering data, as follows: Γ={0<s0<p m ,|▽s0|<U}; Among them, p m represents the maximum value of the projection data A; ▽s0 is the gradient of the original scattering estimation data; U is the upper limit of the gradient, which is obtained through deep learning; The method for obtaining the original scattering estimation data s0 is as follows: in; is an empirical compensation for the beam hardening effect, used to improve the accuracy of scatter estimation; P m (i, j), I0(i, j) are projection data A and projection data B respectively; η is an empirical parameter for controlling the data correction intensity, which is obtained through deep learning; The complementary data representing the measured scattered data and the measured total projection data are obtained by processing the grating area and the gap area of ​​the projection data A and the projection data B, as shown in the following formula: in is the region segmentation function; Ω h and Ω g They represent the grating area and gap area in the projection respectively, and (i, j) is the position index in the image; is a Gaussian function with parameter γ, which is obtained through deep learning; In summary, in the local filtering scattering estimation model based on the full beam blocker or the finger cross blocker, the adaptive optimization parameters include η,γ,U,σ,k b .

5. The method according to claim 1, characterized in that If the beam blocking plate used in step (1) is a rotating blocking plate, the projection obtained by a single scan is composed of a blocking projection and an open projection; (a) According to the operation content described in step (2), the measured scattering data and the measured total projection data are obtained, and the final scattering estimation data s of the blocked projection is obtained in the same calculation method as when using a full beam blocker or a cross-finger blocker b ; (b) Based on the final scatter estimation data of the blocked projection, the final scatter estimation data s of the open projection is obtained according to the following formula: p : in, is a Gaussian function with parameter β, (k, t) is the position index in the image, and β is obtained through deep learning; s b is the final scattering estimation data of the occlusion projection; In summary, in the local filtering scattering estimation model based on the rotating blocking plate, the adaptively optimized parameters include η,γ,U,σ,k b ,β.

6. The method according to claim 1, characterized in that The local filtering scattering estimation model based on deep learning in step (3) specifically refers to using a deep reinforcement learning network to obtain parameters of the local filtering scattering estimation model, and realize parameter adaptive optimization and calculation result output; The deep reinforcement learning network specifically refers to a deep Q learning network based on a Q learning framework, which uses CNN as a function approximator to estimate the Q value function, and makes the action value function close to the optimal by minimizing the quadratic loss function to determine the network parameters.

7. The method according to claim 6, characterized in that The deep reinforcement learning network is composed of n sub-networks with the same architecture, each of which is constructed by a convolutional neural network (CNN), and n is the number of scattering estimation parameters to be determined; during its operation, the current state refers to the current scattering estimation data, and the action taken refers to the way of adjusting the scattering estimation parameters, which is to increase, decrease or remain unchanged; The input information is the measured scatter data and measured total projection data of CBCT imaging, and the output is the adaptively optimized parameters of the scatter estimation model.

8. The method according to claim 6, characterized in that In the training process of selecting parameters for the local filter scattering estimation model, the selector is used to determine the subnetwork to be run next time to adjust the corresponding parameters, including: After training each sub-network, the mean absolute percentage error (MAPE) between the scattering estimation data and the reference scattering data before and after the parameter update corresponding to the sub-network is calculated to obtain the MAPE difference before and after each parameter update, and then the average of the MAPE differences obtained multiple times is calculated; in the next iteration, the sub-network with a larger mean is trained to ensure that each sub-network obtains an adaptive number of training times.

9. The method according to claim 1, characterized in that: In the step (5), the scatter correction data is subjected to noise reduction processing by a filtering method or an iterative method; and the CBCT image is reconstructed by a filtered back projection method, an iterative method or a deep learning network.