CN117475018A - A CT motion artifact removal method - Google Patents

Abstract

The invention discloses a method for removing CT motion artifacts, which specifically includes the following steps: Step 1, collect CT images with motion artifacts; Step 2, screen out clear CT image slices; Step 3, filter out the CT image slices screened out in Step 2. Artificially simulated motion processes are added to the clear CT images to obtain CT image slices with motion artifacts, and a paired CT image data set is obtained; Step 4, use the paired CT image data set obtained in Step 3 to train the kernel-based UNet network model; Step 5: Use CT image data with motion artifacts to test the kernel-based UNet network model trained in step 4, and output clear CT image data. The method provided by the invention can be used to remove diverse and non-uniform motion artifacts caused by rigid and non-rigid motion in CT images.

Description

A CT motion artifact removal method

Technical field

The invention belongs to the technical field of CT artifact correction and relates to a method for removing CT motion artifacts.

Background technique

In the field of motion artifact removal from medical images, there are currently three main solutions. The first is to shorten the scanning time to reduce the possibility of patient movement during the scanning process; the second is to use external motion tracking equipment to obtain the parameter trajectory of the patient's movement, so that images without motion artifacts can be reconstructed; The third method is to achieve motion artifact correction through motion estimation and motion compensation.

The main methods to shorten the scanning time are to increase the scanning speed of the X-ray source or to use multiple ray sources and detectors to improve the time resolution. Although the scanning time of CT imaging is short, CT imaging artifacts caused by patient movement still often exist and can only be repaired through repeated scanning. The motion data tracking method is to obtain the patient's movement trajectory through an external motion tracking device, and use the motion data to restore the consistency of the projection, and finally reconstruct the image to achieve motion artifact correction. This method has high hardware costs and is difficult to promote and use.

Motion compensation technology is a fuzzy recovery technology that compensates motion through a fuzzy recovery algorithm based on external device gating technology. It is mainly divided into two types: one is based on image registration, and the other is based on image reconstruction or Original data recovery methods. The restoration method based on registration is to first reconstruct the images of each respiratory phase separately, then register the images of other respiratory phases to a reference respiratory phase through the registration algorithm, and finally add the images of each respiratory phase to obtain the image Fuzzy recovery method. For the chest and abdomen with complex organ movements, markers cannot accurately reflect the movement status of organs in the body, and even if elastic transformation is used to optimize the registration technology, there will still be local registration errors. Therefore, this registration-based method is not suitable for these parts. The recovery effect is not good enough. The method based on image reconstruction adds motion information to the imaging system model, or is a respiratory motion recovery method that reorganizes the collected data. Compared with registration-based methods, the use of iterative reconstruction algorithms can add various a priori constraints, and image reconstruction or original data recovery methods can improve the signal-to-noise ratio and contrast of the image.

Contents of the invention

The purpose of the present invention is to provide a method for removing CT motion artifacts, which can be directly applied to reconstructed medical images under conventional (non-gated) scanning conditions, which can reduce the diverse and non-rigid motion artifacts caused by rigid and non-rigid motion in CT images. Uniform motion artifacts achieve better image restoration effects, and compared with image reconstruction algorithms, they run faster and save a lot of computing resources.

The technical solution adopted by the present invention is a CT motion artifact removal method, which specifically includes the following steps:

Step 1, collect CT images with motion artifacts;

Step 2: Screen out clear CT image slices;

Step 3: Add artificial simulated motion processes to the clear CT images screened in step 2 to obtain CT image slices with motion artifacts and obtain paired CT image data sets;

Step 4: Use the paired CT image data set obtained in step 3 to train the kernel-based UNet network model;

Step 5: Use CT image data with motion artifacts to test the kernel-based UNet network model trained in step 4, and output clear CT image data.

The invention is also characterized by:

The specific process of step 1 is:

Step 1.1, use a CT instrument to scan the human body or phantom to obtain the corresponding radiation attenuation signal of a certain area of spiral scanning;

Step 1.2, convert the radiation attenuation signal obtained in step 1.1 into a 360-degree detection signal at each point in the relevant scanning area, and output it as a chord diagram domain image;

Step 1.3: Reconstruct the chordmap domain image obtained in step 1.2 into CT image data through the CT image reconstruction algorithm.

The specific process of step 2 is:

Step 2.1, remove the head and abdomen CT image slices in each CT image sequence one by one through manual observation;

Step 2.2, eliminate CT image slices containing absolute values greater than 3000 voxels;

Step 2.3, remove CT image slices containing metal artifacts in each CT image sequence one by one through manual observation;

Step 2.4, identify the tissue structure of each area in each CT image slice, and eliminate blurred CT image slices with motion artifacts, thereby constructing a clear CT image data set.

The specific process of step 3 is:

Step 3.1, generate motion severity matrix;

Step 3.2, generate a deformation mesh mask based on the motion severity matrix generated in step 1;

Step 3.3: Use the ASTRA reconstruction algorithm to add the artificial motion simulation process to obtain data with elastic deformation motion artifacts.

The specific process of step 3.1 is:

Step 3.1.1, generate a random Gaussian matrix to simulate the motion state of the object or tissue. The matrix size is M*M, the center is at the origin, and the x and y axes correspond to the two matrices;

Step 3.1.2, generate and initialize the displacement matrix, the matrix size is N, where N<M, the x and y axes correspond to two matrices;

Step 3.1.3, randomly generate the center coordinates (C _x ,C _y ) of the displacement matrix;

Step 3.1.4, assign a value to the displacement matrix, which is the submatrix selected by the coordinates (C _x , C _y ) in the Gaussian matrix. The matrix size is N*N;

Step 3.1.5: Multiply the displacement matrix by the scaling coefficient (Dir _x , Dir _y ) to obtain the motion severity matrix. The x and y axes correspond to the two matrices;

Step 3.1.6: Cut the numerical absolute value range of the motion severity matrix to the range of max(Dir _x ,Dir _y ).

The specific process of step 3.2 is:

Step 3.2.1, add the motion severity matrix to the mask to describe the motion status of each voxel in the CT image slice;

Step 3.2.2, use spline interpolation to interpolate the deformed image to obtain the final motion mask.

The specific process of step 3.3 is:

Step 3.3.1, build virtual detector;

Step 3.3.2, simulate the CT acquisition process through the virtual detector constructed in step 3.3.1;

Step 3.3.3, obtain the corresponding chord image domain data through CT scanning;

Step 3.3.4: Obtain the image domain data through the CT image reconstruction algorithm, that is, the CT image slice data corresponding to each point.

The specific process of step 4 is:

Step 4.1, calculate the variational inference of the posterior distribution of the kernel and clear image through Bayesian method derivation;

Step 4.2, derive and calculate the evidence lower bound, and then obtain the loss function corresponding to deep learning training;

Step 4.3: Construct a kernel-based UNet network model, and use the paired data set of CT images with motion artifacts obtained in step 2 to train the constructed kernel-based UNet network model.

The specific process of step 4.3 is:

Step 4.3.1, build the kernel network KNet. The kernel network KNet is constructed using the deep learning network as the baseline network. First, the baseline network is used to extract the preliminary features of the CT image slices with motion artifacts, and the preliminary features are further processed through the linear fully connected layer. Transform it into a one-dimensional feature, and then transform the one-dimensional feature through the softmax() function to obtain a feature whose sum of element values is equal to 1. Finally, convert the feature into a two-dimensional feature and represent the corresponding CT image slice with motion artifacts. Fuzzy kernel;

Step 4.3.2, construct the deblurring network DNet. The deblurring network DNet is constructed using the deep learning network for image restoration as the baseline network.

Step 4.3.3, use the paired CT image data obtained in step 3 as training data, and use the loss function given in step 4.2 to train the kernel network KNet and the deblurring network DNet.

The specific process of step 5 is:

Step 5.1, select the paired CT image data set that was not used for model training in step 3 as the test CT image data set;

Step 5.2, estimate the kernel kernel corresponding to the CT image slice with motion artifacts through the kernel network KNet in the kernel-based UNet network model;

Step 5.3: Based on the CT image slices with motion artifacts and the kernel kernel estimated by the kernel network KNet, a motion artifact-free CT image without artifacts or with clear tissue structure is comprehensively extracted and finally restored through the deblurring network.

The beneficial effects of the present invention are as follows:

(1) The motion artifact simulation method provided by the present invention can obtain more realistic motion artifacts and motion blur effects, and can achieve diversified and non-uniform features of motion artifacts, thereby helping to solve the problem of motion artifacts in real CT image data. Shadow removal problem;

(2) The motion artifact simulation method provided by the present invention does not need to be based on physiological mechanisms such as respiratory motion and cardiac motion. It can efficiently simulate motion artifacts and motion blur characteristics, and then construct a paired CT image data set to solve the problem of CT image data Problems such as difficulty in obtaining and scarcity of paired data of CT images with motion artifacts;

(3) The kernel-based UNet network (KBUNet) provided by the present invention fully extracts motion blur information and can significantly improve the motion artifact removal effect of CT images compared with other methods;

(4) The kernel-based UNet network (KBUNet) provided by the present invention can solve the problem of difficult removal of diverse non-uniform motion artifacts with different motion amplitudes, motion directions, and blur levels, and can handle motion artifacts with different characteristics at the same time. It retains the details of tissue structure in artifact-free areas, improves image clarity, and better optimizes CT image quality, which is beneficial to human tissue identification and lesion identification in relevant image areas, and improves diagnostic results;

(5) The kernel-based UNet network (KBUNet) provided by the present invention can effectively remove motion artifacts and improve the efficiency of CT image diagnosis.

Description of the drawings

Figure 1 is a flow chart of a CT motion artifact removal method according to the present invention;

Figure 2 is a flow chart of collecting and screening CT image data in a CT motion artifact removal method according to the present invention;

Figure 3 is a schematic diagram of motion artifacts in data collected in a CT motion artifact removal method according to the present invention;

Figure 4 is a flow chart of simulated motion artifacts in a CT motion artifact removal method according to the present invention;

Figure 5 is a schematic diagram of the effect of motion artifact simulation using a CT motion artifact removal method of the present invention;

Figure 6 is a schematic structural diagram of the kernel-based UNet network KBUNet designed for a CT motion artifact removal method of the present invention;

Figure 7 is a comparison chart of motion artifact removal results in Example 2 of a CT motion artifact removal method of the present invention;

Figure 8 is a comparison chart of motion artifact removal results in Example 3 of a CT motion artifact removal method of the present invention;

Figure 9 is a comparison chart of motion artifact removal results in Example 4 of a CT motion artifact removal method of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the drawings and specific embodiments.

The present invention is a CT motion artifact removal method that can quickly remove lung motion artifacts while well maintaining the tissue structure of non-artifact areas to solve the diversification caused by rigid and non-rigid motion in CT images. , non-uniformized motion artifacts.

The present invention is a CT motion artifact removal method, which is a deep learning method. The model is named Kernel based UNet (KBUNet) and mainly includes two parts. One part is Kernel Net (KNet). The other part is the deblurring network (DeblurNet, DNet).

The kernel network (KNet) is used to extract and analyze the original CT image with artifacts or unclearness and obtain the blur level mask image of the image, and the output feature map is called the kernel.

The deblurring network (DNet) is used to synthesize the features of the original CT image with artifacts or unclearness and the kernel output by the kernel network (KNet), to comprehensively extract and finally restore the artifact-free, Motion artifact-free CT images with fewer artifacts or clearer tissue structures.

Example 1

The present invention provides a method for removing CT motion artifacts. The flow chart is shown in Figure 1. The flow chart of CT image data collection (step 1) and screening (step 2) provided by the present invention can be referred to Figure 2 of the description.

A CT motion artifact removal method of the present invention specifically includes the following steps:

Step 1: Collect CT images with motion artifacts. Generally, when scanning CT images of the chest and abdomen, the image data obtained contains motion artifacts, because respiratory and cardiac movements are almost inevitable;

Detailed process of step 1:

Step 1.1, use a CT instrument to scan the human body or phantom to obtain the corresponding radiation attenuation signal of a certain area of spiral scanning;

Step 1.2, convert the radiation attenuation signal into a 360-degree detection signal at each point in the relevant scanning area, and output it as a chord diagram domain image;

Step 1.3: Reconstruct the chordmap domain image into common CT image data (DCM format data) through a certain CT image reconstruction algorithm (such as filtered back-projection algorithm, FBP algorithm) to facilitate intuitive analysis of the scanned area.

In this embodiment: the CT image data collected through the method in step 1 includes 26 sequences and 6488 slices, and the size of the CT image slices is 512×512 pixels.

The collected CT image data often contains motion artifacts, please refer to Figure 3 of the manual. Figure 3 contains two CT image slices, with boxes showing the details of each area in the image. It can be seen that the motion artifacts in CT images have complex structures and diverse features. The enlarged picture of the lower left corner of the first CT slice shows a clear artifact-free area, and the other seven enlarged details all have CT motion artifacts of varying degrees, which also illustrates the non-uniformity of motion artifacts in CT images. The degree and effect of movement in different areas are different. The method provided by the present invention can simultaneously process motion artifacts with different characteristics, retain the tissue structure details of artifact-free areas, and improve image clarity.

Step 2: Manually screen out clear CT image slices without motion artifacts or with slight motion artifacts; the detailed process of step 2:

Since the respiratory and heartbeat movements of the human body are almost inevitable, and some patients are unable to control and hold their breath during medical examinations, many CT image slices scanned by the method in step 1 will contain motion artifacts. In addition, due to the complexity and physical specificity of CT image scanning, other types of artifacts are often encountered, and there are almost no completely clear and clean CT image slices. Therefore, the present invention removes the main artifacts in this step, namely motion artifacts and metal artifacts. The specific process of step 2 is:

Step 2.1, remove the head and abdomen CT image slices in each CT image sequence one by one through manual observation, thereby improving the specificity of the relevant model method in the present invention regarding lung motion artifacts and improving the model effect;

Step 2.2. Since the CT image slices obtained by conventional human tissue scanning have the CT value of each voxel in the range of -1000 to 1000, the corresponding CT slices containing voxels with an absolute value greater than 3000 are automatically eliminated for safety reasons. ;

Step 2.3: Remove CT image slices containing metal artifacts in each CT image sequence one by one through manual observation. Since metal artifacts will significantly affect the CT value distribution of each voxel in the CT slice, and metal artifacts are another common, important, and difficult-to-solve area in CT artifact removal, in view of the fact that this invention only focuses on motion artifacts. Improve and remove, so remove relevant CT image slices with metal artifacts;

Step 2.4, through manual observation and analysis, identify the tissue structure of each area in each CT image slice, and eliminate CT image slices with slightly more motion blur and obvious motion artifacts, and then construct a clear CT image data set.

In this embodiment: the CT image data obtained through the screening in step 2 includes 26 sequences and 2902 slices.

The CT image slice simulation motion artifact flow chart (step 3) provided by the present invention can refer to Figure 4 of the description.

Step 3: Through a specific motion artifact simulation method, artificially simulated motion processes are added to the clear CT images screened in step 2, and then CT image slices with motion artifacts are obtained, and the corresponding paired data sets are obtained. The paired CT image data obtained in this step can be used for unsupervised learning methods; the detailed process of step 3:

Step 3.1, generate motion severity matrix;

Step 3.1.1, generate a random Gaussian matrix (matrix size is M*M, the center is at the origin, and the x and y axes correspond to two matrices). Random matrices from other distributions, or even mixed distributions, can also be used to simulate certain situations. The state of motion of an object or organization;

Step 3.1.2, generate and initialize the displacement matrix displacement (the matrix size is N, where N<M, the x and y axes correspond to two matrices);

Step 3.1.3, randomly generate the center coordinates (C _x ,C _y ) of the displacement matrix displacement;

Step 3.1.4, assign a value to the displacement matrix, which is the submatrix selected for the coordinates (C _x , C _y ) in the Gaussian matrix. The matrix size is N*N;

Step 3.1.5, multiply the displacement matrix by the scaling coefficient (Dir _x , Dir _y ) to obtain the motion severity matrix (the x and y axes correspond to the two matrices);

Step 3.1.6, cut the numerical absolute value range of the motion severity matrix to the range of max(Dir _x ,Dir _y );

Step 3.2, generate the deformation mesh mask, specifically:

Step 3.2.1, add the motion severity matrix to the mask to describe the motion status of each voxel in the CT image slice;

Step 3.2.2, use spline interpolation to interpolate the deformed image to obtain the final motion mask;

Step 3.3, use ASTRA reconstruction algorithm to join the artificial motion simulation process, specifically:

Step 3.3.1, build virtual detector. You can use the third-party library astra in Python and use its creat_proj_geom() and creat_projector() methods to build a virtual detector;

Step 3.3.2, simulate the CT acquisition process, specifically:

Step 3.3.2.1, simulate the object movement process during CT acquisition. Simulate the CT scanning process in the program, and apply motion to the scanned object or tissue according to the deformation grid mask obtained in step 3.2 during scanning, simulating the motion process of the corresponding object, especially the 6-degree-of-freedom non-rigid motion. Simulate movement by scanning while moving and deforming;

Step 3.3.2.2, simulate the CT scanning process: simulate the CT spiral scanning process through the virtual detector constructed in step 3.3.1.

Step 3.4, obtain data with elastic deformation motion artifacts, specifically:

Step 3.4.1, scan to obtain the chord map domain data: obtain the corresponding chord map domain data through CT scanning;

Step 3.4.2, Chordogram domain data reconstruction: Obtain image domain data through CT image reconstruction algorithm (such as filtered back-projection algorithm, FBP algorithm), that is, CT image slice data corresponding to each point.

In this embodiment: the CT image data obtained through the simulation in step 3 includes a set of artifact-free data and a corresponding set of data with motion artifacts, where each set of data contains 130 sequences and 14,510 slices.

For a schematic diagram of the motion artifact simulation effect of the present invention (step 3), please refer to Figure 5 of the description. The first row is the original CT image slices, and the second row is the CT image slices with motion artifacts obtained through the motion artifact simulation method provided in step 3. The motion amplitude, direction, and features corresponding to each picture are random, which improves the diversity of data samples and generates paired data sets, which is helpful for the training and application of self-supervised learning methods.

Step 4: Use the paired CT image data set obtained in step 3 to train the kernel-based UNet network (KBUNet), which includes two sub-networks: the kernel network (KNet) and the deblurring network (DNet), and use the Bayesian method to optimize the correlation depth Learn the parameters in the network model; detailed process of step 4:

Step 4.1: Calculate the variational inference of the posterior distribution of the kernel and clear image through Bayesian method derivation, which can then be used for numerical calculation and programming;

Step 4.1.1, assume that the image blur model corresponding to the CT motion artifact generation mechanism is

x＝k*y+n (1)

Among them, x is the clear image, y is the blurred image, k is the blur kernel, n is the noise, * is the two-dimensional convolution operator;

Step 4.1.2, assuming that the real clear image x obeys the distribution z, and the real blur kernel k obeys the distribution h;

Step 4.1.3, assuming that the blur kernel k is a Dirichlet distribution (meaning that the sum of the elements in the blur kernel corresponding to a single blur image is 1):

Among them, the parameters of the Dirichlet distribution are π = (π ₁ , π ₂ ,..., π _M ), α = α ₁ , α ₂ ,..., α _M ), and for m = 1,..., M, there are 0 <π _m <1, α _m >0, and

Step 4.1.4, derive the posterior distribution of the clear image distribution z and the real blur kernel distribution h;

p(z,h|y)∝p(z,h,y)＝p(y|z,h)p(z)p(h) (3)

Step 4.1.5, derive the variational inference of the posterior distribution of the clear image distribution z and the real blur kernel distribution h;

q _Φ (h|y)=Dir(h; ξ(y; Φ)) (5)

Among them, Ψ is the parameter of z derived from h and y. The input is the blur image y and the estimated blur kernel h. The output is the approximate posterior (repaired image) z. It is assumed that the mean value of z is μ _i and the variance is m _i . Correspondingly, Φ is the parameter of the blur kernel h estimated from the blurred image y.

Step 4.2, derive and calculate the evidence lower bound, and then obtain the loss function corresponding to deep learning training, specifically:

Step 4.2.1, derive the evidence lower bound ELBO corresponding to step 4.1.5:

is the loss function, and D _KL is the KL divergence. in:

Among them, ξ _i and k _i are the parameters of Dirichlet distribution (refer to formula 2), ψ ₀ is the Digamma function, μ _i and m _i are the mean and variance of z in formula (4), and ε ₀ is the calculated error.

In step 4.2.2, the corresponding deep learning optimization objective can be derived according to formula (6), as shown in the following formula (9):

Step 4.2.3, obtain the corresponding deep learning loss function, as shown in the following formula (10):

Step 4.3, construct a deep learning model, and use the paired data set of CT images with motion artifacts obtained in step 2 to train the kernel network (KNet) and deblurring network (DNet) proposed in the present invention;

The schematic structural diagram (step 4) of the core-based UNet network KBUNet designed by the present invention can be referred to Figure 6 of the description.

Step 4.3.1, build the kernel network (KNet). The kernel network (KNet) is constructed using a simple shallow deep learning network (such as ResNet, DnCNN) as a baseline network. First, the baseline network is used to extract the preliminary features of CT image slices with motion artifacts. The preliminary features are further processed and transformed into one-dimensional features through the linear fully connected layer. The one-dimensional features are then transformed through the softmax() function to obtain the element value. The feature whose sum is equal to 1 (satisfies the assumption in step 4.1.3 that the blur kernel obeys the Dirichlet distribution) is finally converted into a two-dimensional feature and represents the blur kernel (kernel) corresponding to the CT image slice with motion artifacts. ;

In this embodiment, the related sizes of the fuzzy kernel data and the kernel network (KNet) are 101.

Step 4.3.2, build the deblurring network (DNet). The deblurring network (DNet) is built using deep learning networks for image restoration (such as U-Net, NAFNet, MPRNet, etc.) as a baseline network.

Step 4.3.2.1, design the basic architecture of the deblurring network (DNet). Add the blur kernel (kernel) obtained in step 4.3.1 to each layer of the benchmark network, and fuse the feature maps of each layer obtained by the benchmark network with the blur kernel (kernel), which is equivalent to introducing motion blur in the image repair process. Information, thereby obtaining clearer CT image slices with artifacts removed or weakened;

In this embodiment, the deblurred network (DNet) benchmark network adopts MPRNet, in which the UNet module adopts a depth of 6 layers.

Step 4.3.2.2, design the fusion method of feature maps and fuzzy kernel of each layer of the benchmark network. Merge the fuzzy kernel (kernel) with the current layer feature map (feature), and obtain feature I through convolution and activation function operations. Convolve the fuzzy kernel (kernel) with feature I to obtain feature II. Convolve feature I The operation is performed to obtain feature III, and feature II and feature III are merged and superimposed to obtain the final fusion feature;

In this embodiment, each UNet module and ORSNet module in the deblurring network (DNet) are integrated into the fuzzy kernel (kernel) data output by the kernel network (KNet).

Step 4.3.3, train the kernel network (KNet) and deblurring network (DNet), specifically:

Step 4.3.3.1, training method. Use the paired CT image data obtained in step 3 as training data, and use the loss function given in step 4.2 to train the kernel network (KNet) and deblurring network (DNet).

Step 4.3.3.2, training details. The kernel network (KNet) and the deblurred network (DNet) can be trained separately, that is, when training one network, the parameters of the other network are frozen, the connection between the models and the feature fusion are interrupted, and the paired CT is used to affect the slice data and the blur kernel obtained by simulation. (kernel) train two networks separately, and then train all parameters as a whole after the training effect is stable. You can also directly train the two models as a whole. The training process can be ended after the model training indicator data is stable;

In this example, epochs is set to 100 and batch size is set to 8. The 14510 pairs of matching CT image slices simulated in step 3 are divided into a training set (11608 pairs of CT image slices) and a test set (2902 pairs of CT image slices) in a ratio of 4:1, and the training set (11608 pairs of CT image slices) is used slice) training. During the training process, gradient clipping is added through norm clipping and absolute value clipping to avoid model parameter explosion.

Step 5: Use CT image data with motion artifacts to test and evaluate the relevant model performance of the CT motion artifact removal method in the present invention, and output corresponding clearer CT image data without motion artifacts or with reduced motion artifacts. Achieve the effect of quickly and efficiently removing motion artifacts.

Detailed process of step 5:

Step 5.1, select the CT image data set for testing. The paired CT image data set obtained in step 3 and not used for model training can be used to calculate image indicators and measure the model deblurring effect. Real unpaired blur data with motion artifacts can also be used to efficiently remove motion artifacts and evaluate the blur removal effect;

In this embodiment, the test set selected in advance (2902 pairs of CT image slices) is used to evaluate the image effect and calculate related indicators. The main indicators include MSE, PSNR, and SSIM. Finally, on the test set, MSE, PSNR, and SSIM reached 2.83×10 ^-4 , 36.24, and 0.96 respectively.

In this embodiment, the real belt motion artifacts eliminated in step 2.4 are also used for evaluation testing.

Step 5.2, estimate the kernel corresponding to the CT image slice with motion artifacts through the kernel network (KNet);

Step 5.3: Based on the CT image slices with motion artifacts and the kernel estimated by the kernel network (KNet), the artifact-free and less artifact-free images are comprehensively extracted and finally restored through the deblurring network (DeblurNet, DNet). Or CT images with clearer tissue structure and motion artifact removal.

Example 2

Please refer to Figure 7 of the manual. The first picture in the first row is the original CT image slice without motion artifacts in step 2; the first picture in the second row is the paired CT image slice with motion artifacts simulated in step 3; the first picture in the first row is The two pictures are the artifact removal effects obtained through DnCNN model training; the third picture in the first row is the artifact removal effect obtained through UNet model training; the fourth picture in the first row is the artifact removal effect obtained through MPRNet model training Artifact removal effect picture; the second picture in the second row is the artifact removal effect picture obtained through DeepDeblurNet model training; the third picture in the second row is the artifact removal effect picture obtained through NAFNet model training; the second row The four pictures are the artifact removal effect pictures obtained through the diffusion model method training; the fifth picture in the first row is the artifact removal effect picture obtained through the KBUNet model training provided by the present invention, but the data set is trained using block data. That is, each slice data in the original 512×512 pixel size CT image slice data set in step 4.3.3.2 is split into 64×64 pixel size CT image slices; the fifth picture in the second row is provided by the present invention Artifact removal effect diagram obtained by KBUNet model training, where the data set used is the original 512×512 pixel size CT image slice in step 4.3.3.2.

Example 3

Please refer to Figure 8 of the manual. The first picture in the first row is the original CT image slice without motion artifacts in step 2; the first picture in the second row is the paired CT image slice with motion artifacts simulated in step 3; the first picture in the first row is The two pictures are the artifact removal effects obtained through DnCNN model training; the third picture in the first row is the artifact removal effect obtained through UNet model training; the fourth picture in the first row is the artifact removal effect obtained through MPRNet model training Artifact removal effect picture; the second picture in the second row is the artifact removal effect picture obtained through DeepDeblurNet model training; the third picture in the second row is the artifact removal effect picture obtained through NAFNet model training; the second row The four pictures are the artifact removal effect pictures obtained through the diffusion model method training; the fifth picture in the first row is the artifact removal effect picture obtained through the KBUNet model training provided by the present invention, but the data set is trained using block data. That is, each slice data in the original 512×512 pixel size CT image slice data set in step 4.3.3.2 is split into 64×64 pixel size CT image slices; the fifth picture in the second row is provided by the present invention Artifact removal effect diagram obtained by KBUNet model training, where the data set used is the original 512×512 pixel size CT image slice in step 4.3.3.2.

Example 4

Please refer to Figure 9 of the manual. The first picture in the first row is the original CT image slice without motion artifacts in step 2; the first picture in the second row is the paired CT image slice with motion artifacts simulated in step 3; the first picture in the first row is The two pictures are the artifact removal effects obtained through DnCNN model training; the third picture in the first row is the artifact removal effect obtained through UNet model training; the fourth picture in the first row is the artifact removal effect obtained through MPRNet model training Artifact removal effect picture; the second picture in the second row is the artifact removal effect picture obtained through DeepDeblurNet model training; the third picture in the second row is the artifact removal effect picture obtained through NAFNet model training; the second row The four pictures are the artifact removal effect pictures obtained through the diffusion model method training; the fifth picture in the first row is the artifact removal effect picture obtained through the KBUNet model training provided by the present invention, but the data set is trained using block data. That is, each slice data in the original 512×512 pixel size CT image slice data set in step 4.3.3.2 is split into 64×64 pixel size CT image slices; the fifth picture in the second row is provided by the present invention Artifact removal effect diagram obtained by KBUNet model training, where the data set used is the original 512×512 pixel size CT image slice in step 4.3.3.2.

Figures 7, 8, and 9 of the description corresponding to Embodiments 2 to 4 show the motion artifact removal effect of CT image slices of different parts of the chest. It can be seen that the kernel-based UNet network (KBUNet) provided by the present invention fully extracts motion blur information. Compared with other methods, it can significantly improve the effect of removing motion artifacts in CT images; it can solve different motion amplitudes, motion directions, and blur degrees. It solves the problem of difficult removal of diverse non-uniform motion artifacts. It can handle motion artifacts with different characteristics at the same time, retain the structural details of artifact-free areas, improve image clarity, and better optimize CT image quality, which is beneficial to related images. It can identify regional human tissues and identify lesions to improve the diagnostic effect; it can also effectively remove motion artifacts and improve the efficiency of CT imaging diagnosis.

Through the above method, the present invention discloses a CT motion artifact removal method, which can efficiently remove motion artifacts, improve the artifact removal effect, and improve CT image quality to meet actual use needs.

The method provided by the invention is an image restoration method, which belongs to the image-based (post-) reconstruction algorithm and is directly applied to medical images reconstructed under conventional (non-gated) scanning conditions. It does not require hardware assistance and image registration, and is fast. , can obtain better image restoration effects and save a lot of computing resources.

Claims (10)

1. The CT motion artifact removal method is characterized by comprising the following steps of:

step 1, acquiring CT images with motion artifacts;

step 2, screening out clear CT image slices;

step 3, adding an artificial simulation motion process to the clear CT images screened in the step 2, so as to obtain CT image slices with motion artifacts, and obtaining paired CT image data sets;

step 4, training a core-based UNet network model by using the paired CT image data set obtained in the step 3;

and 5, testing the core-based UNet network model trained in the step 4 by using CT image data with motion artifacts, and outputting clear CT image data.

2. The method for removing CT motion artifacts according to claim 1, wherein the specific procedure of step 1 is as follows:

step 1.1, scanning a human body or a phantom by using a CT instrument to obtain a corresponding radiation attenuation signal of helical scanning of a certain area;

step 1.2, converting the radiation attenuation signals obtained in the step 1.1 into 360-degree detection signals of each point in a relevant scanning area, and outputting the detection signals as chord domain images;

and step 1.3, reconstructing the chord domain image obtained in the step 1.2 into CT image data through a CT image reconstruction algorithm.

3. The method for removing CT motion artifacts according to claim 2, wherein the specific process of step 2 is as follows:

step 2.1, removing head and abdomen CT image slices in each CT image sequence one by one through manual observation;

step 2.2, eliminating CT image slices containing voxels with absolute values greater than 3000;

step 2.3, removing CT image slices containing metal artifacts in each CT image sequence one by one;

and 2.4, judging the tissue structure of each region in each CT image slice, and removing the blurred CT image slices with motion artifacts so as to construct a clear CT image data set.

4. The method for removing CT motion artifacts according to claim 1, wherein the specific procedure in step 3 is as follows:

step 3.1, generating a motion severity matrix;

step 3.2, generating a deformation grid mask based on the motion severity matrix generated in the step 1;

and 3.3, adding an artificial motion simulation process by using an ASTRA reconstruction algorithm to obtain data with elastic deformation motion artifacts.

5. The method of claim 4, wherein the specific process of step 3.1 is as follows:

step 3.1.1, generating a random Gaussian matrix to simulate the motion state of an object or tissue, wherein the matrix is M, the center is at the origin, and the x and y axes correspond to the two matrices;

step 3.1.2, generating and initializing a displacement matrix displacement, wherein the size of the matrix is N, and N < M, and x and y axes correspond to two matrices;

step 3.1.3, randomly generating the center coordinates (C _x ,C _y )；

Step 3.1.4, assigning a displacement matrix displacement, which is the coordinates (C _x ,C _y ) The size of the selected submatrix is N;

step 3.1.5 multiplying the displacement matrix displacement by the scaling factor (Dir _x ,Dir _y ) Obtaining a motion severity matrix, wherein the x-axis and the y-axis correspond to the two matrices;

step 3.1.6, clipping the absolute value range of the motion severity matrix to max (Dir _x ,Dir _y ) Within the range.

6. The method according to claim 5, wherein the specific process of step 3.2 is:

step 3.2.1, adding a motion severity matrix into a mask for describing the motion condition of each voxel in the CT image slice;

and 3.2.2, interpolating the deformed image by using spline interpolation to obtain a final motion mask.

7. The method of claim 5, wherein the specific process of step 3.3 is:

step 3.3.1, constructing a virtual detector;

step 3.3.2, simulating a CT acquisition process through the virtual detector constructed in the step 3.3.1;

step 3.3.3, obtaining corresponding chord domain data through CT scanning;

and 3.3.4, obtaining image domain data, namely CT image slice data corresponding to each point position, through a CT image reconstruction algorithm.

8. The method for removing CT motion artifacts according to claim 1, wherein the specific procedure of step 4 is as follows:

step 4.1, deducing and calculating kernel and posterior distribution variation inference of the clear image through a Bayesian method;

step 4.2, deducing and calculating a lower bound of evidence, and further obtaining a loss function corresponding to the deep learning training;

and 4.3, constructing a core-based UNet network model, and training the constructed core-based UNet network model by using the CT image pairing data set with the motion artifact obtained in the step 2.

9. The method of claim 8, wherein the specific process of step 4.3 is:

step 4.3.1, constructing a kernel network KNet, constructing the kernel network KNet by using a deep learning network as a reference network, firstly extracting preliminary features of CT image slices with motion artifacts through the reference network, further processing and converting the preliminary features into one-dimensional features through a linear full-connection layer, then converting the one-dimensional features into features with the sum of element values equal to 1 through softmax () function, and finally converting the features into two-dimensional features and fuzzy kernel corresponding to the CT image slices with the motion artifacts;

step 4.3.2, constructing a deblurring network DNet, wherein the deblurring network DNet is constructed by using a deep learning network for image recovery as a reference network;

and 4.3.3, training the kernel network KNet and the deblurring network DNet by using the loss function given in the step 4.2 by using the paired CT image data obtained in the step 3 as training data.

10. The method of claim 9, wherein the specific process of step 5 is:

step 5.1, selecting the paired CT image data set which is not used for model training in the step 3 as a CT image data set for testing;

step 5.2, estimating a kernel corresponding to the CT image slice with the motion artifact through a kernel network KNet in the kernel-based UNet network model;

and 5.3, comprehensively extracting and finally recovering the artifact-free or clear-tissue-structure motion artifact-free CT image through a deblurring network according to the kernel estimated by the motion artifact-free CT image slice and the kernel network KNet.