CT image ring artifact data simulation and removal method

Abstract

The invention discloses a deep learning-oriented CT image ring artifact data simulation method, and an interlayer complementarity and central region enhanced ring artifact removal method. In combination with a ring artifact principle and scanning system characteristics, paired ring artifact data and labels are obtained through four steps of numerical model body physical size alignment, quantum noise and electronic noise addition, detector non-uniform response simulation and ring data and acyclic data pairing reconstruction. A double-branch annular artifact removal network is designed. The network uses the data simulation method to obtain a large amount of annular artifact data as a training set, and a pseudo three-dimensional data format is designed as network input in combination with inter-chip artifact information complementarity randomly generated by detector response. The method has good characteristics in the aspects of reducing ring artifacts and recovering image details. Moreover, the data simulation method fully considers the characteristics of the system, the model for simulation data training can be directly applied to unseen real data, and the method has great potential in solving the problem of actual ring artifact removal.

Description

Technical Field

[0001] The invention relates to a CT image annular artifact data simulation and removal method, and belongs to the field of computer image processing. Background Art

[0002] By using photon counting detectors, X-ray computed tomography systems (CT Systems) provide a more efficient way to obtain and reconstruct images. However, there are certain drawbacks in its clinical application, such as Figure 1 As shown in the figure, due to the randomness of the detector response, the stripe artifacts in the projection data eventually appear as concentric ring artifacts in the reconstructed image, which greatly affects the image quality.

[0003] There are two types of ring artifacts: high-frequency ring artifacts and low-frequency ring artifacts. High-frequency ring artifacts are mainly caused by three factors: detector bad pixels, non-uniform response between detector pixel levels, and gaps between detectors. Low-frequency ring artifacts are usually caused by discontinuities between different pixel blocks of the detector.

[0004] There are two traditional methods for removing ring artifacts. One is to decompose the detector response into two material bases that are energy-independent and non-interfering from each other from the perspective of the projection domain material basis, and reconstruct the image based on the two material bases; the other is to extract the ring structure from the reconstructed image domain using various filters to achieve the artifact removal effect. Regardless of which method is used, it requires rich experience and knowledge in image reconstruction and filtering, and the detector responses of each device are inconsistent. Hyperparameters need to be adjusted according to the device, and it cannot be promoted as a universal method. With the continuous development of deep learning technology, it has shown great potential in the field of image restoration. As the most widely used and effective supervised learning method, in order to ensure the high accuracy and generalization of the model, a large amount of labeled data is often required to train the model. Due to the flexibility and variability of the detector response, it is difficult to obtain a large amount of ring artifact data with different characteristics in practice. Therefore, the present invention proposes a CT image ring artifact data simulation method. On the basis of obtaining sufficient training data pairs, a deep learning ring artifact removal method based on inter-slice complementarity and central area enhancement is proposed. Summary of the invention

[0005] Aiming at the problem that the traditional ring removal method has complicated and cumbersome steps, the present invention proposes a method for simulating and removing ring artifacts in CT images. The ring artifact simulation algorithm is used to generate simulated ring artifact data for network training; a dual-branch ring artifact removal network is designed, and a low-quality image with ring artifacts is input, and a high-quality image with ring removal is output, thereby removing the ring artifacts and restoring the details of the original image.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] (1) A ring artifact simulation method based on the ring artifact principle and scanning system characteristics is proposed. The specific steps are as follows:

[0008] Step 1. Numerical phantom resampling and linear transformation to achieve physical size alignment and energy spectrum compensation. To ensure that the numerical representation is consistent with the size and position of the real object, the numerical phantom is first resampled, and the physical size of the numerical phantom is aligned by cropping and padding, taking into full account the actual physical characteristics of the simulation ring. Linear transformation is used to roughly compensate for the energy spectrum differences between different CT systems, see formula (1).

[0009]

[0010] Among them B raw , W raw are the values ​​of water / soft tissue and bone in the original artifact-free images from other systems, B real , W real are the corresponding values ​​in the artifact image from a real CT system.

[0011] Step 2. Add quantum noise and electronic noise. T Projected to the projection domain, see formula (2). Add quantum noise and electronic noise to the projection domain to generate a noisy projection image P N , to more realistically simulate the actual scanning environment, as shown in formula (5).

[0012] P GT =A·V T (2)

[0013]

[0014] V T is a ring-free image, A is the projection matrix of the CT system, P GT is the logarithmically normalized data of the ring-free image projection domain. Q and Q0 are the number of transmitted photons and the number of incident photons, respectively, and E is the electronic noise.

[0015] Step 3. Simulate the non-uniform response of the detector. High-frequency ring artifacts appear as dense concentric circle structures, and low-frequency ring artifacts appear as band-like circular structures with a certain width. High-frequency ring artifacts are mainly caused by three factors: detector dead spots, non-uniform responses between detector pixels, and gaps between detectors. Low-frequency ring artifacts are usually caused by discontinuities between different pixel blocks of the detector. The gaps between detectors and detector dead spots can be corrected using a flat field, respectively. Therefore, this method achieves the effect of generating high-frequency rings and low-frequency rings by simulating the non-uniform responses of the detector pixel level and pixel block level. Determine the random response range of the detector based on the bright and dark field images, and generate a random response matrix P R. The detector inhomogeneity response is introduced by Hadamard.

[0016] P NSI =P N ·P R (4)

[0017] Among them, P NSI The projection image with noise and ring artifacts added. R is the detector random response matrix, P N The projected image is noisy.

[0018] Step 4. Pairing and reconstruction of loop data and non-loop data. Finally, the processed loop data and the non-loop data before processing are paired and reconstructed, as shown in formula (5). That is, the simulated loop data pair is obtained.

[0019] V NRI =FDK(P NSI ),V GT =FDK(P GT ) (5)

[0020] Among them, P NSI P is the projection image with noise and ring artifacts added. GT is the label of the corresponding ring-free and artifact-free projection image. NSI and V GT They are the ring artifact image domain data and corresponding labels after FDK reconstruction.

[0021] (2) A deep learning ring artifact removal algorithm based on inter-slice complementarity and central area enhancement is proposed

[0022] Based on the ring artifact simulation method proposed above, it can overcome the problem that it is difficult to obtain a large amount of ring artifact data with different characteristics in practice. With sufficient training data pairs, this paper designs a dual-branch ring artifact removal to remove ring artifacts.

[0023] Step 1. Design a pseudo-3D input format based on inter-slice complementarity. The detector responses are independent in the head-to-foot direction, i.e., the z-axis direction. Due to the independence of the detector response in the vertical direction, the ring artifacts on different slices are different. Since this method proposes a pseudo-3D data input structure, a Z-axis slice group is taken (5 consecutive slices as a group in this method), and the number of slices is placed in the channel dimension, so that the model can adaptively learn and capture complementary information between adjacent slices, thereby enhancing the restoration of image details and the removal of ring artifacts. Compared with traditional two-dimensional and three-dimensional structures, it retains effective spatial redundant information while improving computational efficiency and reducing model parameters.

[0024] Step 2. Design a dual-branch ring artifact removal network. Use simulated ring artifact images as training data and non-ring artifact data as labels to input into the network. The network structure is as follows: Figure 3 As shown in the figure, the network is a dual-branch structure with the residual U-Net network as the benchmark model. The global ring artifact removal branch learns the global artifact features at multiple scales, and the central area enhancement branch learns the central artifact features at multiple scales. The central area enhancement module of the network design realizes the information fusion between the two branches at multiple scales, ensuring the global ring removal effect while improving the central area ring removal effect. The network is finally restored to the original resolution through the upsampling module. The network training process is as follows: Figure 4 As shown in Figure 2, the loss function is the sum of the MSE losses of the global image and the central area image, see equations (6), (7), and (8). The gradient is calculated through the loss value, and the stochastic gradient descent optimization method AdaW is used to update the network parameters.

[0025] L Total =L GI +L CR (6)

[0026]

[0027] Among them, L Total is the total network loss, L GI is the global branch loss, L CR is the center region branch loss, N is the total number of pixels in the image, and are the prediction results of the global branch and the central branch respectively. and For the corresponding label.

[0028] Step 3. Design a post-processing method to process the test data. After training, use simulated data and real data to test the network. The number of slices and the slice step size of the test data are set to 5 and 1 respectively. Continuous slices are taken as the starting slices of each slice group. Each slice is tested multiple times. Finally, the test results of each slice are weighted and averaged to calculate the final result of each slice.

[0029] Experimental results show that the network has good effects on both simulated ring artifact data and real ring artifact data, and can remove ring artifacts while effectively restoring image quality, demonstrating its potential in the actual ring artifact removal problem.

[0030] Compared with the prior art, the advantages of the present invention are as follows: (1) A method for simulating CT image ring artifact data is proposed, which provides data support for the application of supervised learning in the field of ring removal. The method fully considers the principle of ring artifacts and the characteristics of the scanning system. The simulated ring artifact data is of full variety, rich in features, and the same as the actual artifacts. (2) The ring artifact removal effect is significant and has practical potential. The method makes full use of the complementarity of inter-slice information, fully considers the redundant information of the image at both the data design and network design levels, proposes a pseudo-three-dimensional input design and designs a dual-branch ring artifact removal network. The experimental results of simulated data and real data show that the method has good ring removal performance. (3) The image details are restored and the image after the ring artifact removal is guaranteed not to be distorted or lost. The network proposed by the method not only has a good removal effect, but also can well restore the image details and remove noise. It shows good performance, which meets the actual diagnostic needs of doctors. (4) The running speed is fast. The artifact removal algorithm runs on the GPU, the computing requirements are low, and the actual test time is short. (5) It has universality and parameter-free adjustment. This method does not require any parameter adjustment and has a good artifact removal effect for any detector device. (6) The present invention is an automated, end-to-end method that can be integrated into the reconstruction software of a CT scanner. Only by designing a corresponding button, a high-quality image with ring artifacts removed can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the defect imaged for the photon detector

[0032] Figure 2 It is a flow chart of the present invention.

[0033] Figure 3 Figure 2. Network structure diagram for double-branch ring artifact removal.

[0034] Figure 4 Flowchart of network training for dual-branch ring artifact removal

[0035] Figure 5 This is the algorithm test result diagram. DETAILED DESCRIPTION

[0036] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0037] Example: Figure 2 As shown in Figure 2, a method for simulating and removing ring artifacts in CT images is proposed. The entire process of this method is as follows: Figure 2As shown in the figure, the ring artifact simulation algorithm is used to generate simulated ring artifact data for network training; a dual-branch ring artifact removal network is designed, which inputs a low-quality image with ring artifacts and outputs a high-quality image without ring artifacts, thus achieving ring removal while ensuring image details.

[0038] In order to achieve the above object, the technical solution of the present invention is as follows:

[0039] (1) A ring artifact simulation method based on the ring artifact principle and scanning system characteristics is proposed. The specific steps are as follows:

[0040] Step 1. To ensure that the size and position of the numerical representation are consistent with the real object, the numerical phantom is first resampled, and the physical size of the numerical phantom is aligned by cropping and padding, taking into full account the actual physical characteristics of the simulated ring. Linear transformation is used to roughly compensate for the energy spectrum differences between different CT systems, see formula (1).

[0041]

[0042] Among them B raw , W raw are the values ​​of water / soft tissue and bone in the original artifact-free images from other systems, B rea l.W real are the corresponding values ​​in the artifact image from a real CT system.

[0043] Step 3. Add quantum noise and electronic noise. Project the image domain data to the projection domain, add photoelectric noise to the projection data, simulate the scanning state of the real CT system, and obtain the noisy projection image P N .

[0044] P GT =A·V T (2)

[0045]

[0046] V T is a ring-free image, A is the projection matrix of the CT system, P GT is the logarithmically normalized data of the ring-free image projection domain. Q and Q0 are the number of transmitted photons and the number of incident photons, respectively, and E is the electronic noise.

[0047] Step 2. Simulate the non-uniform response of the detector. The ring artifacts caused by the gaps between detectors and the bad pixels of the detector can be corrected by flat field correction. Therefore, this method generates high-frequency rings and low-frequency rings by simulating the non-uniform response of the detector at the pixel level and pixel block level. The non-uniform response range of the detector can be determined by the bright and dark fields, and the random response matrix P of the detector is randomly generated. R , add the projected image P NAs in formula (2)(3).

[0048] P NSI =P N ·P R (4)

[0049] Among them, P NSI The projection image with noise and ring artifacts added. R is the detector random response matrix, P N The projected image is noisy.

[0050] Step 4. Pairing and reconstruction of loop data and non-loop data. Finally, the processed loop data and the non-loop data before processing are paired and reconstructed to obtain simulated loop data pairs.

[0051] V NRI =FDK(P NSI ),V GT =FDK(P GT ) (5)

[0052] Among them, P NSI P is the projection image with noise and ring artifacts added. GT is the label of the corresponding ring-free and artifact-free projection image. NSI and V GT They are the ring artifact image domain data and corresponding labels after FDK reconstruction.

[0053] (2) A deep learning ring artifact removal algorithm based on inter-slice complementarity and central area enhancement is proposed

[0054] Based on the ring artifact simulation method proposed above, it is possible to overcome the problem that it is difficult to obtain a large amount of ring artifact data with different characteristics in practice. With sufficient training data pairs, this paper designs a dual-branch ring artifact removal network to remove ring artifacts.

[0055] Step 1. Design a pseudo-3D input format based on inter-slice complementarity. The detector responses are independent in the head-to-foot direction, i.e., the z-axis direction, and the detector responses are random. This method proposes a pseudo-3D pseudo-data input structure, takes a Z-axis slice group (in this example, 5 slices per group), and places the number of slices into the channel dimension.

[0056] Step 2. Design a dual-branch ring artifact removal network. Use simulated ring artifact images as training data and non-ring artifact data as labels to input into the network. The network structure is as follows: Figure 3As shown in the figure, the dual-branch ring artifact removal network is a dual-branch structure with the residual U-Net network as the benchmark model. The global ring artifact removal branch learns the global artifact features at multiple scales, and the central area enhancement branch learns the central artifact features at multiple scales. The central area enhancement module of the network design promotes the information fusion between the two branches at multiple scales, ensuring the global ring removal effect while improving the central area ring removal effect, and finally restoring the original resolution through the upsampling module. The network training process is as follows: Figure 4 As shown in Figure 2, the loss function is the sum of the MSE losses of the global image and the central area image. See equations (6), (7), and (8). The gradient is calculated through the loss value, and the stochastic gradient descent optimization method AdaW is used to update the network parameters.

[0057] L Total =L GI +L CR (6)

[0058]

[0059] Step 3. Design a post-processing method to process the test data. After the training is completed, the network is tested using simulated data and real data. Continuous slices are taken as the starting slices of each slice group. Each slice is tested multiple times. Finally, the test results of each slice are weighted and averaged to calculate the final result of each slice.

[0060] Experimental results show that the network has good effects on both simulated and real ring artifact data, and can effectively restore image quality while removing ring artifacts, proving its potential in solving practical ring artifact problems.

[0061] The present invention provides a method for simulating and removing ring artifact data in CT images. The ring removal test results are as follows: Figure 5 shown.

[0062] Effect evaluation:

[0063] The present invention discloses a method for simulating and removing annular artifact data in CT images. The test results of the method are as follows: Figure 5 As shown. The figure shows two groups of test images, including the simulation and removal effect of ring artifacts under ring-free data and the removal effect of ring artifacts under real data. The test was tested in the integrated software. The test process is: import the ring artifact data into the software, then click the "Remove Ring" button to get a high-quality image, and click "Save" to save the de-ringed image to a file. The algorithm runs fast, and the effect of removing ring artifacts and retaining image features is significant. It shows that this method can quickly and significantly remove ring artifacts, greatly enhance image quality, and facilitate doctors to make diagnosis and treatment decisions.

[0064] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technology can understand and think of any changes or substitutions within the technical scope disclosed by the present invention, which should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for simulating and removing ring artifact data in CT images, characterized in that: The method comprises the following steps: Step 1: Physical size alignment and energy spectrum difference compensation of the ring-free images are performed to eliminate the geometric and numerical differences between different CT systems. Step 2: Project the image aligned in step 1 to the projection domain, add quantum noise and electronic noise to more realistically simulate the actual scanning system. Step 3, generate a non-uniform response coefficient matrix, and use the point-by-point Hadamard product to add to the data processed in step 2 to simulate the non-uniform response of the detector, generating low-frequency and high-frequency rings in the image domain. Step 4, pair the ring data processed in step 3 with the acyclic data processed in step 1, and reconstruct them in the image domain to obtain a pair of ring image and acyclic image data. Step 5: Convert the simulated ring data pair obtained in step 4 into pseudo three-dimensional input data and input it into the network for training. Step 6: Input the real ring artifact data into the network for de-ringing to obtain a high-quality ring-free image after de-ringing.

2. The method for simulating and removing ring artifact data based on CT images according to claim 1, characterized in that: The physical size alignment and energy spectrum difference compensation of the numerical phantom mentioned in step 1, that is, the size alignment between the numerical phantom and the actual object is achieved through cropping, filling, etc., and the energy spectrum inconsistency is compensated by linear transformation through the following formula to ensure that the simulation loop has the actual system characteristics and has the same system characteristics. Among them B raw , W raw are the values ​​of water / soft tissue and bone in the original artifact-free images from other systems, B real , W real are the corresponding values ​​in the artifact image from a real CT system.

3. The method for simulating and removing ring artifact data in CT images according to claim 1, characterized in that: As proposed in step 2, Gaussian noise and Poisson noise are added to the projection data, and the projection data and noise data are obtained using the following formula, and the noise is added to the projection data to obtain the noise projection map P N , P GT =A·V T Among them, V T is a ring-free image, A is the projection matrix of the CT system, P GT is the projection domain data of the ring-free image after logarithmic processing, Q and Q0 are the number of transmitted photons and incident photons, respectively, and E is the electronic noise.

4. The method for simulating and removing ring artifact data in CT images according to claim 1, characterized in that: The simulated detector non-uniform response proposed in step 3, the flat field projection diagram can realize the correction of detector bad points and detector gaps. After determining the detector fluctuation response range using flat field and dark field, the point-by-point Hadamard product is used to P N Add non-uniform response P R , P NSI =P N ·P R Among them, P NSI V is the projection image with added noise and ring artifacts. R is the detector random response matrix, P N The projected image is noisy.

5. The method for simulating and removing ring artifact data in CT images according to claim 1, characterized in that: The pseudo-3D input data format proposed in step 5 enables the model to adaptively learn and capture the complementary information between adjacent slices. The specific approach is to take 5 consecutive slices in the depth direction as a slice group and place the number of slices into the channel dimension of the network input.

6. The method for simulating and removing ring artifact data in CT images according to claim 1, characterized in that: The network training process proposed in step 5 uses the simulated ring artifact data generated in steps 1-4 as training data and the corresponding acyclic data as labels, which are input into the network in a pseudo-three-dimensional input format. The network uses a dual-branch structure to extract global ring artifact features and central area ring artifact features at different scales, respectively, and uses a central area enhancement module to fuse the dual-branch information to improve the central area artifact removal effect. The loss function is shown in the following formula. The gradient is calculated by the loss value, and the network parameters are updated using the stochastic gradient descent optimization method AdaW. L Total =L GI +L CR Among them, L Total is the total network loss, L GI is the global branch loss, L CR is the center region branch loss, N is the total number of pixels in the image, and are the prediction results of the global branch and the central branch respectively. and For the corresponding label.

7. The method for simulating and removing ring artifact data in CT images according to claim 1, characterized in that: The network test process proposed in step 6 uses a post-processing method to deal with the inter-slice image jump problem generated by the test data. The specific approach is that the test data takes continuous slices as the starting slices of each slice group. Each slice has multiple tests. Finally, the test results of each slice are weighted and averaged to calculate the final result of each slice.

Images