Universal imaging physics-driven CT data simulation and CT image noise artifact suppression method

Abstract

The present invention discloses a universal imaging physics-driven CT data simulation and CT image noise artifact suppression method. The present invention calculates the pixel values ​​of bones, fat, muscles, and air in real CT images and real CBCT images respectively, constructs a mapping function, and maps the pixel values ​​of each tissue in the CT image to the pixel values ​​of the corresponding parts in the real CBCT image to obtain simulated CBCT data. Then, the projection of the simulated CBCT data is obtained using the forward projection function, and the projection data of low-dose CBCT is simulated by sparse sampling angles and adding projection domain noise. The projection domain data is then reconstructed using the FDK function, and paired simulated CBCT data that can support supervised learning is obtained in such a universal imaging physics-driven manner. Test results show that this method has good results in reducing artifacts, suppressing noise, and restoring image details on both simulated data and real data.

Description

Technical Field

[0001] The present invention relates to a universal imaging physics-driven CT data simulation and CT image noise artifact suppression method, and belongs to the field of computer image processing. Background Art

[0002] Computed tomography (CT) is an X-ray imaging technique, a core technology in medical imaging, widely used in the diagnosis and assessment of various diseases and conditions. Cone-beam CT, a key development in CT technology, has been widely promoted due to its advantages, including direct reconstruction of three-dimensional images, low X-ray radiation dose, and rapid imaging. However, due to the cone-beam geometry that can cause artifacts and blurring at the edges of the image, and the lower radiation dose that generates noise, CBCT image quality is lower than that of CT images. Therefore, improving CBCT image quality has become an important research area.

[0003] Deep learning is a key technology for improving CBCT image quality. As a data-driven optimization algorithm, deep learning requires a large amount of data to train the algorithm model and improve its effectiveness and robustness. However, due to data privacy concerns, publicly available datasets for improving low-dose CBCT image quality are scarce. Therefore, it is necessary to develop custom simulated CBCT data to test the algorithm's effectiveness, while also using a small amount of real data to enhance its robustness.

[0004] Whether simulated CBCT data is close to real data is crucial for determining whether an algorithm can truly be applied to real data. Furthermore, further improving robustness to real data through real-world data guidance is essential. The former determines whether a deep learning model is capable of improving the image quality of real CBCT data, while the latter focuses on further improving the quality of real CBCT images based on the former. Therefore, it is necessary to explore a CT data simulation method driven by imaging physics and a method for suppressing CT image noise artifacts using unsupervised domain adaptation. Summary of the Invention

[0005] In order to improve the quality of CBCT images, the present invention proposes a universal imaging physics-driven CT data simulation and CT image noise artifact suppression method. The entire process of the method is as follows: Figure 1 As shown in the figure, it can generate simulated CBCT data from real CT data based on the pixel values ​​of each tissue in the real CT image and the pixel values ​​of the corresponding parts in the real CBCT image, and then remove the noise and artifacts on the real data through the CT image noise artifact suppression method based on unsupervised domain adaptation.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: The present invention provides a general imaging physics-driven CT data simulation and CT image noise artifact suppression method, which specifically comprises the following steps:

[0007] Step 1: The real CT data is converted from CT values ​​to pixel values.

[0008] Step 2: The real CBCT projection data is reconstructed using the FDK function. Its parameters include the distance from the ray source to the detector, the distance from the ray source to the human body, the number of detectors, the size of the detector probe, the offset of the detector on the two-dimensional plane, the pixel size of the generated image, and the corresponding projection angle information to obtain the real CBCT image at the pixel value.

[0009] Step 3, such as Figure 1 As shown, based on the pixel values ​​obtained in step 1, the average pixel values ​​of the bone, fat, muscle, and air regions of the real CT image are checked. Based on the pixel values ​​of the bone, fat, muscle, and air regions of the real CBCT image obtained in step 2, the average pixel values ​​of the bone, fat, muscle, and air regions are checked. A function is calculated that maps the pixel values ​​of each part of the CT image to the pixel values ​​of each part of the CBCT image. The pixel values ​​of the real CT image are mapped to the corresponding pixel values ​​of the real CBCT image to obtain a pixel-aligned CT image. Projection data is obtained using a forward projection function. The parameters of the forward projection function include the distance from the radiation source to the detector, the distance from the radiation source to the body, the number of detectors, the size of the detector probe, the offset of the detector on the two-dimensional plane, the pixel size of the generated image, and the corresponding projection angle information. These parameters are kept consistent with the parameters in step 2 to ensure that the obtained projection is closer to the real CBCT projection. The FDK function is then reconstructed with the same parameters as above to obtain a simulated CBCT image.

[0010] Step 4: According to steps 1-3, obtain standard-dose CBCT simulation data and the corresponding low-dose CBCT simulation data. According to step 2, obtain standard-dose CBCT real data and the corresponding low-dose CBCT real data. It should be noted that the first two steps in the acquisition of low-dose CBCT simulation data are consistent with those of standard-dose CBCT simulation data. In step 3, the parameters of the forward projection and FDK function need to be modified. Each parameter is synchronously sampled by one-kth to simulate image artifacts at sparse angles. Poisson noise is added to the CBCT simulation projection after forward projection to simulate the real noise generated by the lower radiation dose. The same is true for the low-dose CBCT real data: one-kth of the projection is sampled and the same Poisson noise as the simulated low-dose is added. On this basis, offline data enhancement is added to randomly generate projection data after shifting in three dimensions and reconstruct paired low-dose CBCT simulation data and standard-dose CBCT simulation data after offline data enhancement.

[0011] Step 5: Based on the paired CBCT simulation data obtained in step 4, the low-dose CBCT simulation image is input into the supervised noise artifact removal backbone network, and the CBCT simulation image after artifacts and noise are removed is output. Figure 3 As shown in the figure, the backbone network for supervised noise artifact removal consists of a downsampling path and an upsampling path, with skip connections in between to transmit image details. We also set the number of channels of feature maps of different resolutions to 32, 64, 128, and 256, increasing the number of channels to minimize GPU memory usage and increase computation speed. The essence of a denoising network is to learn the difference between the input and the label. For this task, we introduce a residual module to transform the network from learning the data itself to learning the noise. Figure 3 The network training process is demonstrated: we first use online data augmentation to expand the data, including random translations and rotations, to prevent overfitting and enhance the network's generalization ability. Next, the image is normalized and fed into a supervised noise artifact removal backbone network. The predicted image quality-improved output is then calculated along with the labels to compute a supervised image quality improvement loss function. The gradient is then calculated from the loss value, and the network parameters are updated using stochastic gradient descent optimization.

[0012] In step 6, based on the low-dose CBCT real data and low-dose CBCT simulation data obtained in step 4, the two data are input into the backbone network respectively, the feature distribution learned by the encoder is extracted, and it is upsampled to the original image size to speed up the operation by occupying the GPU video memory. After obtaining the corresponding feature map, it is input into the domain discriminator. The domain discriminator can extract the deep information of the feature map, perceive the distribution of the input features, and output the distribution of the input feature map. The distribution loss of the domain discriminator is calculated through the unsupervised domain adaptation loss function. Figure 3 Demonstrates the training process of the network.

[0013]

[0014] Among them, I Source represents the simulated CBCT data, I Target represents real CBCT data, Represents the feature extraction output and upsampled to the original image size, Φ Discriminator (·) represents the distribution of the domain discriminator output, Represents the discriminator Φ Discriminator right The gradient of , where:

[0015]

[0016] The unsupervised domain adaptation loss is added to the loss of the supervised noise artifact removal backbone network in step 5. The learning of real data features in unsupervised domain adaptation learning is used to guide the supervised noise artifact removal backbone network to adapt to the features of real data, thereby improving the effectiveness and robustness of the denoising network.

[0017] Compared with the existing technology, the advantages of the present invention are as follows: (1) By collecting CT data of real human chests and applying offline data enhancement technology, this method can generate data that is very close to the real CBCT dataset in terms of tissue pixel value range, detail contrast and noise artifacts, thereby achieving accurate simulation effects. This means that in the field of medical imaging, we can obtain more realistic and reliable simulation data, which is helpful for the research and diagnosis of medical imaging; (2) The scheme uses residual learning and multi-scale feature learning methods to effectively remove noise artifacts in CT images while keeping the details very close to the standard dose. This means that even in a noisy environment, we can maintain high image quality, which helps doctors accurately diagnose patients; (3) Through the unsupervised domain transfer method, the model's transfer learning ability from simulation data to real data is improved, thereby enhancing the model's generalization performance. This means that our model can better adapt to real data and can show good stability even when there are certain differences between real data and simulation data; (4) The scheme runs on a GPU, so in actual testing, it only takes 190 milliseconds to divide the positioning line for a single positioning image, which has a fast computing speed. This means that our method not only improves in quality, but also has obvious advantages in computational efficiency, which can process medical imaging data more quickly and provide more efficient support for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 、 Figure 2 Flowchart of the present invention.

[0019] Figure 3 This is the network training graph.

[0020] Figure 4 The test result graph is shown in Figure 2. DETAILED DESCRIPTION

[0021] The present invention is further illustrated below with reference to specific examples. It should be understood that these examples 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, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0022] Example: Figure 1 、 Figure 2As shown, the present invention provides a universal imaging physics-driven CT data simulation and CT image noise artifact suppression method, and the final test results are shown in Figure 4 The specific steps are as follows:

[0023] Step 1: The real CT data is converted from CT values ​​to pixel values.

[0024] Step 2: The real CBCT projection data is reconstructed using the FDK function. Its parameters include the distance from the ray source to the detector, the distance from the ray source to the human body, the number of detectors, the size of the detector probe, the offset of the detector on the two-dimensional plane, the pixel size of the generated image, and the corresponding projection angle information to obtain the real CBCT image at the pixel value.

[0025] Step 3, such as Figure 1 As shown, based on the pixel values ​​obtained in step 1, the average pixel values ​​of the bone, fat, muscle, and air regions of the real CT image are checked. Based on the pixel values ​​of the bone, fat, muscle, and air regions of the real CBCT image obtained in step 2, the average pixel values ​​of the bone, fat, muscle, and air regions are checked. A function is calculated that maps the pixel values ​​of each part of the CT image to the pixel values ​​of each part of the CBCT image. The pixel values ​​of the real CT image are mapped to the corresponding pixel values ​​of the real CBCT image to obtain a pixel-aligned CT image. Projection data is obtained using a forward projection function. The parameters of the forward projection function include the distance from the radiation source to the detector, the distance from the radiation source to the body, the number of detectors, the size of the detector probe, the offset of the detector on the two-dimensional plane, the pixel size of the generated image, and the corresponding projection angle information. These parameters are kept consistent with the parameters in step 2 to ensure that the obtained projection is closer to the real CBCT projection. The FDK function is then reconstructed with the same parameters as above to obtain a simulated CBCT image.

[0026] Step 4: According to steps 1-3, obtain standard-dose CBCT simulation data and the corresponding low-dose CBCT simulation data. According to step 2, obtain standard-dose CBCT real data and the corresponding low-dose CBCT real data. It should be noted that the first two steps in the acquisition of low-dose CBCT simulation data are consistent with those of standard-dose CBCT simulation data. In step 3, the parameters of the forward projection and FDK function need to be modified. Each parameter is synchronously sampled by one-kth to simulate image artifacts at sparse angles. Poisson noise is added to the CBCT simulation projection after forward projection to simulate the real noise generated by the lower radiation dose. The same is true for the low-dose CBCT real data: one-kth of the projection is sampled and the same Poisson noise as the simulated low-dose is added. On this basis, offline data enhancement is added to randomly generate projection data after shifting in three dimensions and reconstruct paired low-dose CBCT simulation data and standard-dose CBCT simulation data after offline data enhancement.

[0027] Step 5: Based on the paired CBCT simulation data obtained in step 4, the low-dose CBCT simulation image is input into the supervised noise artifact removal backbone network, and the CBCT simulation image after artifacts and noise are removed is output. Figure 3 As shown in the figure, the backbone network for supervised noise artifact removal consists of a downsampling path and an upsampling path, with skip connections in between to transmit image details. We also set the number of channels of feature maps of different resolutions to 32, 64, 128, and 256, increasing the number of channels to minimize GPU memory usage and increase computation speed. The essence of a denoising network is to learn the difference between the input and the label. For this task, we introduce a residual module to transform the network from learning the data itself to learning the noise. Figure 3 The network training process is demonstrated: we first use online data augmentation to expand the data, including random translations and rotations, to prevent overfitting and enhance the network's generalization ability. Next, the image is normalized and fed into a supervised noise artifact removal backbone network. The predicted image quality-improved output is then calculated along with the labels to compute a supervised image quality improvement loss function. The gradient is then calculated from the loss value, and the network parameters are updated using stochastic gradient descent optimization.

[0028] In step 6, based on the low-dose CBCT real data and low-dose CBCT simulation data obtained in step 4, the two data are input into the backbone network respectively, the feature distribution learned by the encoder is extracted, and it is upsampled to the original image size to speed up the operation by occupying the GPU video memory. After obtaining the corresponding feature map, it is input into the domain discriminator. The domain discriminator can extract the deep information of the feature map, perceive the distribution of the input features, and output the distribution of the input feature map. The distribution loss of the domain discriminator is calculated through the unsupervised domain adaptation loss function. Figure 3 Demonstrates the training process of the network.

[0029]

[0030] Among them, I Source represents the simulated CBCT data, I Target represents real CBCT data, Represents the feature extraction output and upsampled to the original image size, Φ Discriminator (·) represents the distribution result of the domain discriminator output.

[0031] The unsupervised domain adaptation loss is added to the loss of the supervised noise artifact removal backbone network in step 5. The learning of real data features in unsupervised domain adaptation learning is used to guide the supervised noise artifact removal backbone network to adapt to the features of real data, thereby improving the effectiveness and robustness of the denoising network.

[0032] Effect evaluation:

[0033] The present invention discloses a general imaging physics-driven CT data simulation and CT image noise artifact suppression method. The test results of this method are as follows Figure 4 The figure shows three sets of test images, including real CBCT images of the human chest at different locations. The test was conducted using the CMD command line in a Windows system. The test process was as follows: first, enter condaactivate environment to activate the environment configuration for running the code, then enter python data_simulated.py to simulate the data, and finally enter python test.py to test the data artifact noise removal. The algorithm runs in an average of approximately 190ms, which is fast. The simulation closely matches the real data, and the noise artifact removal is relatively accurate. This demonstrates that this method can simulate data quickly and achieve high quality, and remove noise artifacts.

[0034] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person familiar with the technology can understand and think of any changes or replacements within the technical scope disclosed by the present invention, which should be included in the scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A general imaging physics-driven CT data simulation and CT image noise artifact suppression method, characterized in that: The specific steps are as follows: Step 1: Real CT image I CT Convert CT value to pixel value. Step 2: The real CBCT projection data p CBCT The real CBCT image is reconstructed by FDK function. Step 3: The real CBCT image I under the pixel value obtained in step 2 CBCT , check the pixel values ​​corresponding to its bones, fat, muscles and air, and the real CT image I obtained in step 1 CT , check the pixel values ​​corresponding to the bone, fat, muscle and air, and calculate the function f that maps the CT image to the CBCT image CT_to_CBCT , according to the obtained mapping function f CT_to_CBCT , the real CT data I CT The pixel values ​​are mapped to the real CBCT data I CBCT The corresponding pixel value is obtained by using the mapped CT data I CT_simlated , the projection data p is obtained through the forward projection function cbct_simulated , and then reconstructed by FDK function to obtain the simulated CBCT data I CBCT_simulated , Step 4: According to steps 1-3, obtain the standard dose CBCT simulation data I CBCT_simulated_nd and the corresponding low-dose CBCT simulation data I CBCT_simulated_ld , Step 5: According to the paired CBCT simulation data obtained in step 3, input the low-dose CBCT simulation data I CBCT_simulated_ld , calculated by denoising network and standard dose CBCT simulation data I CBCT_simulated_nd The L1 loss between them is used to remove noise and artifacts on low-dose data by reducing the loss through network learning. Step 6: Introducing low-dose CBCT real data I CBCT_ld , and the input low-dose CBCT simulation data of the backbone network I CBCT_simulated_ld We conduct unsupervised domain adaptation learning together, use the domain discriminator, and use the WGAN_GP loss function to calculate the distribution loss of the two different domain features. This is added to the loss of the supervised noise artifact removal backbone network to guide the network to learn the characteristics and data distribution of the real data domain. The network with the final training weights is input into the low-dose CBCT real data I CBCT_ld , output I after removing artifact noise CBCT_output ; The specific implementation of unsupervised domain adaptation in step 6 is as follows: D target =Φ Discriminator (feature target ) D source =Φ Discriminator (feature source ) Among them, I source , I target Represents low-dose CBCT simulation data and low-dose CBCT real data, It means extracting the corresponding features and upsampling them to the original image size, Φ Discriminator It represents the domain discriminator, which is implemented by multi-layer convolution and is used to output the distribution of input features, D target , d source It represents the real feature distribution and simulated feature distribution output by the domain discriminator respectively, and finally the unsupervised domain adaptation loss function is used to calculate the loss of unsupervised domain adaptation. in, and represents the expected distribution of source and target domain features, λ gp represents the weight of the gradient guidance part, Represents the discriminator Φ Discriminator right The gradient of , where: 。 2. The general imaging physics-driven CT data simulation and CT image noise artifact suppression method according to claim 1, characterized in that: The real CT data in step 1 comes from the public CT dataset. The CT data is converted from HU value to mu value.

3. The general imaging physics-driven CT data simulation and CT image noise artifact suppression method according to claim 1, characterized in that: In step 2, the real CBCT is reconstructed according to the given reconstruction parameters using the FDK function of the third-party library TIGRE.

4. The general imaging physics-driven CT data simulation and CT image noise artifact suppression method according to claim 1, characterized in that: In step 3, check the pixel values. Open the image using ImageJ software, zoom in on the bone, fat, muscle, and air areas, select and calculate their corresponding average values ​​to obtain the corresponding pixel values ​​of each tissue.

5. The general imaging physics-driven CT data simulation and CT image noise artifact suppression method according to claim 1, characterized in that: The parameters used in the forward projection function mentioned in step 3 are consistent with the reconstruction parameters of the real CBCT data in step 2.

6. The general imaging physics-driven CT data simulation and CT image noise artifact suppression method according to claim 1, characterized in that: The parameters used by the FDK function in step 3 are consistent with the reconstruction parameters of the real CBCT data in step 2.

7. The general imaging physics-driven CT data simulation and CT image noise artifact suppression method according to claim 1, characterized in that: Step 4 Low-dose CBCT simulation data I CBCT_simulated_ld Before performing FDK function reconstruction, one-kth of the projections is selected and a certain amount of Poisson noise is added to the projection data. Finally, the FDK function is used to reconstruct the simulated low-dose CBCT data.

Images