A deep learning-based industrial ct scatter and beam hardening coupling artifact suppression method

Abstract

The application provides a kind of industrial CT scattering and beam hardening coupling artifact suppression method based on deep learning, belongs to the field of industrial CT image processing, for the problem that coupled artifacts cannot be accurately separated and corrected in industrial CT imaging, a beam hardening, scattering coupling artifact model based on real CT image artifact features is constructed, a simulated image and an actual image are mixed to construct a data set by a coupled model representation equation, and artifact suppression and image enhancement are completed by a long-short connection, structure doublet improved neural network, effectively removing multiple types of coupled artifacts and noise, this method is suitable for coupled artifact suppression of any complex structure actual CT image, and the practicability and generality of the method are better.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of industrial CT image processing, and relates to an industrial CT scattering and beam hardening coupled artifact suppression method based on deep learning. BACKGROUND

[0002] As a kind of non-destructive testing, industrial CT uses CT equipment to scan the detected object. The rays are attenuated after passing through the object, and the penetrating rays are received by the detector. Through digital processing, the digital signals (projections) are converted, and then the two-dimensional tomographic images or three-dimensional stereoscopic images that can show the internal structure characteristics, material composition and defect condition of the object are obtained. Its application almost covers all industrial fields. Due to the influence of the physical and mathematical factors existing in the CT system and the surrounding environment, the detector inevitably receives incomplete signals and redundant signals during the actual ray transmission process. The projection image shows artifacts such as beam hardening, scattering and noise. The existence of these unfavorable factors seriously increases the gray unevenness of the reconstructed image, blurs the edges and reduces the contrast, which interferes with the subsequent judgment of the image information and the identification of defects. Therefore, the suppression of artifacts in CT imaging is very necessary for the development of the advantages of industrial CT detection.

[0003] For the artifacts caused by beam hardening or scattering in industrial CT, the current conventional method is to correct a single variable. The artifact correction of scattering can be divided into three categories: hardware obstruction, scattering estimation and software and hardware combination. Hardware suppression reduces the scattering received by the detector through tools, but the effect is limited. The software method describes the scattering through some algorithms and formulas, but the estimation is not accurate enough. The software and hardware combination method compensates for the above shortcomings but still has residual scattering. To solve the problem of beam hardening, the implementation method can be divided into hardware suppression and software correction. The hardware method filters out low-energy rays, but in actual operation, it needs to be used in combination with other methods according to the imaging conditions. The software method mostly needs to be based on prior information or obtained in calibration, and also has certain limitations. With the rise of various neural networks, the effectiveness of the deep learning-based method in removing medical CT metal artifacts has been fully verified, and it has also been gradually applied to the correction of other artifacts. However, the correction object is single, and the inseparability of beam hardening and scattering in industrial CT is not considered systematically. In addition, due to the inability to obtain actual reference images in the industrial field, the data set is missing. In addition, due to the complexity of the environment, there is a big difference between the real image and the simulated image, and the correction effect of the simulated training model directly applied to the actual image is not good, which limits its application. In summary, there is no systematic method to meet the global effective algorithm for the information pollution caused by the coupling of multiple artifacts in industrial CT images. SUMMARY

[0004] In view of the shortcomings of existing technologies, this invention provides a deep learning-based method for suppressing coupling artifacts in industrial CT scattering and beam hardening. Based on the artifact characteristics of actual CT images, a beam hardening and scattering coupling model is established to obtain simulated coupling artifact images. Then, the constructed dataset and the actual dataset are input into a neural network to complete feature learning. By adjusting the parameters and structure of the neural network, the detailed and positional information of coupling artifacts can be fully transmitted. This method can adapt to artifact estimation and coupling artifact suppression for different parts, while simultaneously achieving noise removal and image enhancement.

[0005] The technical solution of this invention is as follows:

[0006] Step 1: Statistically analyze the grayscale of the real projection P1 image containing coupling artifacts, obtain the scattered field distribution and beam hardening curve, obtain the corrected projection, denoted as the reference projection P2, establish a coupling artifact model based on projections P1 and P2, and simplify the model to obtain the coupling artifact model characterization equation.

[0007] Step 2: Construct projection dataset CP1 by pairing projection images P1 and P2, and fit the parameters of the coupled artifact model representation equation based on the mapping relationship of dataset CP1.

[0008] Step 3: Reconstruct projections P1 and P2 respectively, obtain slice images and denote them as actual coupled artifact image S1 and actual reference image S2, and construct actual slice dataset CS1;

[0009] Step 4: Obtain the simulated projection P4 containing coupling artifacts through simulated monoenergetic projection P3, and reconstruct the simulated reference image S3 and the simulated coupling artifact image S4 respectively. Use S3 and S4 to construct the simulated slice dataset CS2.

[0010] Step 5: Combine datasets CS1 and CS2 to construct a hybrid dataset CS, and use the improved neural network for calibration training;

[0011] Step 6: Use the trained neural network to correct the image containing artifacts;

[0012] In step 1, the established coupling artifact model is as follows: Where A, B, C, D, and E are parameters, P1 is the coupling artifact projection, P2 is the reference projection, and A, B, C, D, and E are model coefficients. The simplified coupling artifact model characterization equation is y = d1·x 5 +d2·x 4 +d3·x 3 +d4·x 2+d5·x+d6, where the coupled artifact projection P1 is represented by y, the reference projection P2 is represented by x, and d1, d2, d3, d4, d5, and d6 are coefficients obtained by fitting the actual projection dataset CP1;

[0013] In step 5, the improved neural network is achieved by concatenating the image and data channels of the U-Net++ network, while simultaneously fusing attention mechanisms and dilated convolutions.

[0014] The beneficial effects of this invention are: the coupling artifact model provided by this invention more easily reflects the inseparable and coupled effects of beam hardening and scattering in industrial CT images compared to a single model. This invention can address the problem of insufficient actual reference images in dataset creation, overcoming the shortcoming of existing deep learning methods where simulated data training results cannot be directly applied to real data. Furthermore, this method is simple to implement and can effectively utilize the fusion of actual data and constructed simulated data to bridge the gap between the two, resulting in good correction effects.

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method described in this invention;

[0017] Figure 2 The images show the verification results for different actual parts; where (a) and (e) are coupling artifact images, (b) and (f) are images corrected by the ordinary method, (c) and (g) are images corrected by the original U-Net method, and (d) and (h) are images corrected by the method in this paper. Detailed Implementation

[0018] To more clearly illustrate the implementation of this invention, the following will refer to the accompanying drawings. Figure 1 The specific implementation of this invention is described below. The relevant parameters are obtained based on existing industrial cone-beam CT equipment (X-ray source: Comet MXR-451HP / 11, flat panel detector: PerkinElmer XRD 1621AN15 ES). Based on the above-mentioned equipment, the following steps are performed:

[0019] Step 1: Statistically analyze the grayscale of the real projection P1 image containing coupling artifacts to obtain the scattered field distribution and beam hardening curve, and obtain the corrected projection, denoted as the reference projection P2. Establish a coupling artifact model based on projections P1 and P2, and simplify the model to obtain the coupling artifact model characterization equation. The specific derivation process is as follows:

[0020] 1) The simplified law of X-ray transmission in a real CT environment is as follows: Figure 2This can be summarized as follows: the initial ray I0 from the radiation source attenuates as it passes through the component; low-energy rays are absorbed; and a portion of the high-energy rays are scattered as I by the influence of objects and the environment. S Finally, with the remaining transmitted rays I H Together with other factors N, the rays are absorbed by the detector. Therefore, the ray I received by the detector can be approximated as:

[0021] I = I H +I S +N (1);

[0022] 2) The analytical model of the scattered field distribution can be obtained using existing methods, such as equation (2). Generally, the total ray I and the scattered ray I at the detector end are obtained using single-grating scanning technology. S The corresponding data also makes it easy to know that scattered ray I S With transmitted rays I H The relationships exhibit the same characteristic trends, as shown in equation (3), where a1, a2, b1, and b2 are the parameters to be fitted:

[0023]

[0024]

[0025] 3) The beam hardening correction process generally includes obtaining the hardening curve and correcting the multi-energy projection to a single-energy correction line. The hardening curve can be characterized by its relationship with the crossing length (4). According to the empirical multi-energy projection P H The slope of the curve about the traversal path at zero is the monoenergetic projection P. M The slope of the monoenergetic projection is given by equation (5):

[0026]

[0027] P M (l)=c3l (5),

[0028] Therefore, by using the mono-energy projection value corresponding to the multi-energy projection value under the same crossing length, the multi-energy projection is corrected to the mono-energy projection, and an approximate relationship between the mono-energy projection and the multi-energy projection is obtained:

[0029]

[0030] 4) Based on the above relationships and the X-ray attenuation law, the total projection received at the detector end can be represented as the negative logarithm of the received total X-ray intensity and the original intensity. Therefore, the relationship between the total projection at the detector end and the scattered X-rays and hardened transmitted X-rays can be converted into:

[0031]

[0032] To simplify the influence of other factors N on the total projection, it is separated from the logarithmic term and treated as a constant N0.

[0033] 5) Substituting equations (6) and (3) into the coupling artifact model expression (7), the final coupling artifact model is equation (8). Other parameters A, B, C, D, and E are used to replace the constant terms. The total projection P at the detector end is the coupling artifact projection P1, and the single-energy projection P M For reference projection P2, equation (8) can be changed to equation (9):

[0034]

[0035]

[0036] This formula represents the relationship between the reference projection P2 and the coupled artifact projection P1 in the actual situation. As can be seen from the derivation process, parameters C and D are parameters related to hardening, while parameters A and B are parameters related to scattering and hardening. Scattering and hardening cannot be directly separated, which reflects the true inseparability and mutual coupling of scattering and hardening.

[0037] 6) The above steps are the process of establishing the coupling artifact model. However, existing CT scanning equipment can directly correct the projection through hardware devices and integrated systems, and can obtain the corrected projection P2 and the coupling artifact projection P1. Thus, the parameters in equation (9) can be directly fitted. In order to optimize the fitting difficulty, let P1 be y and P2 be x. The coupling artifact model is simplified by constraints and parameter expressions to obtain the analytical solution of y with respect to x, which is denoted as the coupling artifact model characterization equation:

[0038] y = d1x 5 +d2x 4 +d3x 3 +d4x 2 +d5x+d6 (10),

[0039] In the formula, d1, d2, d3, d4, d5, and d6 can be obtained by fitting the actual projection dataset CP1 of different models.

[0040] Step 2: Construct projection dataset CP1 by pairing projection images P1 and P2, and fit the parameters of the coupled artifact model representation equation based on the mapping relationship of dataset CP1.

[0041] Taking a cylindrical stainless steel 3D printed part as an example, the scanning parameters are: X-ray source voltage 450kV and current 1.5mA, geometric parameters are: distance from X-ray source to detector 1200mm, distance from X-ray source to rotation center 870mm, sampling range, sampling mode and number of sampling frames are 360 ​​uniform sampling frames from 0-360°. Since the existing equipment can directly obtain the correction projection P2 through hardware and software integration system, the obtained coupling artifact projection P1 is directly paired with the reference projection P2 to construct the projection dataset CP1. With 30° interval, the corresponding projection grayscale data under different scanning angles are extracted. The grayscale value of projection P2 is used as the independent variable x, and the grayscale value of the corresponding position of P1 is used as the dependent variable y to obtain the mapping relationship of the projection values, that is, the parameters in equation (10) are obtained: d1=0, d2=0.0001, d3=-0.0019, d4=-0.0038, d5=0.3410, d6=0.0035.

[0042] Step 3: Reconstruct projections P1 and P2 respectively, obtain slice images and denote them as actual coupled artifact image S1 and actual reference image S2, and construct actual slice dataset CS1;

[0043] The reconstruction parameters used were: FDK algorithm, projection resolution of 512×512, and reconstruction pixel size of 0.1393mm. The projections P1 and P2 obtained in step 1 were reconstructed to obtain the actual coupled artifact image S1 and the actual reference image S2, and the actual dataset CS1 was constructed.

[0044] Step 4: Obtain the simulated projection P4 containing coupling artifacts through simulated monoenergetic projection P3, and reconstruct the simulated reference image S3 and the simulated coupling artifact image S4 respectively. Use S3 and S4 to construct the simulated slice dataset CS2.

[0045] The monoenergetic projection P3 is generated by simulation software. The slice image is reconstructed from the projection using a reconstruction algorithm, which is the simulation reference image S3. The gray values ​​of the monoenergetic projection are extracted. The simplified characterization equation constructed by the fitting parameters in step 2 is used to obtain the simulation coupled artifact projection P4. The slice image S4 is reconstructed. The simulation dataset CS2 is constructed from images S3 and S4.

[0046] Step 5: Combine datasets CS1 and CS2 to construct a hybrid dataset CS. Then, use the improved neural network for calibration training. The specific improvements and training parameters of the network are as follows:

[0047] 1) Design a feature extraction layer and a feature enhancement layer for the mixed training set CS to extract features from the input image, and use a multi-scale fusion method to establish a one-to-one mapping;

[0048] 2) The U-Net++ structure is adopted, which learns and fuses features of different depths by combining long and short connections, while passing positional information and supplementing detailed information to improve learning accuracy. In addition, the flexible network structure combined with deep supervision can perform model pruning, which can significantly reduce the number of parameters in the deep network with a huge number of parameters within an acceptable accuracy range.

[0049] 3) A dual-sequential network structure was designed. The highest standard is still the high-quality simulation reference image, which undertakes the corresponding tasks and can better clarify the learning objectives: The first sequential module mainly maps the overall position information and some detail information, and removes coupling artifacts as a whole. At this time, the learning objective is mainly the actual corrected image, learning the image effect after correction by traditional methods; The main task of the second sequential module is to complete the transition from the actual corrected image to the simulation reference image level, controlling the detail information while enhancing the image, and learning the effect similar to the simulation reference image.

[0050] 4) By constructing a simulated coupled artifact image that is closer to the actual image, the gap between the actual image and the simulated image is reduced. The coupled artifact images S1 and S4 are used as network inputs, the reference images S2 and S3 are used as the final outputs, and the simulated reference image S3 is used as the highest label. This not only standardizes the learning results in each module, but also serves as the core of the image channel that connects the first and second modules.

[0051] 5) Since this paper uses a two-step sequential network, but the parameters passed are quite large, reducing useless information during training is beneficial to improving the network training speed. A specific location attention module that integrates attention mechanism and dilated convolution is designed to optimize the network.

[0052] 6) Set the number of iterations per round to 2000, with 30 iterations, and monitor the loss value. During network training, augmentation data is generated at a batch rate of 2. Specific operations include image rotation by 90°, 180°, and 270°, scaling up and down using scaling factors of 0.5 and 2, and translation in the horizontal and vertical directions.

[0053] Step 6: Use the trained neural network to correct the image containing artifacts.

[0054] In this embodiment, a method for suppressing artifacts in industrial CT scattering and beam hardening coupling based on deep learning is characterized by:

[0055] 1) Based on prior knowledge of beam hardening and scattering in actual industrial CT, a coupled artifact model is constructed to allow the simulated image to restore the complex information of the real image. This model is then input into the network along with the actual image to solve the problem of insufficient actual data.

[0056] 2) By improving the dual-linked optimized U-shaped network, multiple types of artifacts and noise problems can be solved simultaneously. The information transmitted by each link network is more accurate and effective, and they compensate for each other to maximize their functions.

[0057] 3) The network is fed with both simulated and real data, with the simulated ideal image as the core target of the entire learning process. This overcomes the barrier of real images and enables the network to learn at the level of the reference image.

[0058] Figure 2 By comparing the method of the present invention with other methods for actual images of different parts, it can be seen that the method of the present invention can not only effectively remove coupling artifacts, but also suppress noise and enhance the image, thereby significantly improving the image contrast and clarity.

Claims

1. A deep learning based industrial CT scatter and beam hardening coupling artifact suppression method, characterized in that The method comprises the following steps: Step 1: Statistics is performed on the gray scale of a real projection P1 image containing coupling artifacts, a scatter field distribution and a beam hardening curve are obtained, a corrected projection is obtained, denoted as a reference projection P2, a coupling artifact model is established based on the projections P1 and P2, the coupling artifact model is simplified to obtain a coupling artifact model representation equation; Step 2: The projection images P1 and P2 are constructed into a projection data set CP1 in a paired manner, and the parameters of the coupling artifact model representation equation are fitted according to the mapping relationship of the data set CP1; Step 3: The projections P1 and P2 are respectively reconstructed to obtain slice images denoted as actual coupling artifact images S1 and actual reference images S2, and an actual slice data set CS1 is constructed; Step 4: A simulation single-energy projection P3 is obtained through simulation, a simulation projection P4 containing coupling artifacts is obtained, simulation reference images S3 and simulation coupling artifact images S4 are respectively reconstructed, and a simulation slice data set CS2 is constructed using S3 and S4; Step 5: The data sets CS1 and CS2 are mixed to construct a mixed data set CS, and the improved neural network is used for correction training; Step 6: The trained neural network is used to correct the image containing artifacts. In the step 1, the established coupling artifact model is Wherein A, B, C, D, E are parameters, P1 is the coupling artifact projection, P2 is the reference projection, A, B, C, D, E are model coefficients, and the simplified coupling artifact model representation equation is y=d1·x 5 +d2·x 4 +d3·x 3 +d4·x 2 +d5·x+d6, wherein the coupling artifact projection P1 is represented by y, the reference projection P2 is represented by x, and d1, d2, d3, d4, d5, d6 are coefficients obtained by fitting the actual projection data set CP1. In the step 5, the improved neural network realizes the series connection of the image channel and the data channel of the U-Net++ network, and simultaneously realizes the fusion of the attention mechanism and the hollow convolution.

2. The deep learning based industrial CT scatter and beam hardening coupling artifact suppression method of claim 1, wherein: The coupling artifact model representation equation established in the step 1 can obtain the coupling artifact projection P1 according to the reference projection P2, and the data set is expanded through the coupling artifact model representation equation.

3. The deep learning based industrial CT scatter and beam hardening coupling artifact suppression method of claim 1, wherein: In the step 5, the improved neural network model is trained by using the mixed data set of the simulation data and the real data.

Images