Artifact Suppression Method for Sparse Angle CT Images Based on Structure Enhancement and Artifact Estimation

Abstract

The present invention discloses a method for suppressing artifacts in sparse angle CT images based on structure enhancement and artifact estimation, and belongs to the field of computer tomography technology. The present invention comprises: constructing a structure enhancement artifact suppression network for sparse angle CT image processing, which mainly includes two artifact feature extraction subnetworks and structure feature extraction subnetworks with the same network architecture; then constructing an image restoration loss function and a structure loss function and jointly training each subnetwork to form a trained structure enhancement artifact suppression network; finally, the user inputs the sparse angle CT image to be processed into the trained structure enhancement artifact suppression network, and outputs the CT image after artifact suppression. In response to the problem that artifacts in sparse angle CT images are complex and conventional processing easily loses tissue structure information, the present invention adopts a strategy combining artifact extraction with structure enhancement, which can improve the effect of sparse angle CT image processing and obtain high-accuracy CT images containing more tissue structure information.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer tomography, and more particularly, relates to a method for suppressing artifacts in sparse angle CT images based on structure enhancement and artifact estimation. Background Art

[0002] In 1895, German scientist Roentgen first discovered X-rays, which can penetrate objects, pioneering the use of X-rays in medical diagnosis. In 1972, the first computed tomography (CT) device was developed, assisting doctors in clinical diagnosis. Today, CT imaging technology is becoming increasingly sophisticated and widely used, becoming an indispensable part of medical diagnosis and treatment. However, X-rays are partially absorbed by the human body, causing damage to tissues and organs and potentially leading to genetic diseases or cancer.

[0003] Generally, there are two main ways to reduce the X-ray radiation dose in CT scans: reducing the tube current and sparse angle scanning. Reducing the tube current inevitably leads to lower effective photons and increased noise in the projection data (or sinusoidal diagram), thereby reducing the quality of the reconstructed image. Currently, there are a large number of studies on low-tube current CT imaging, and excellent results have been achieved. Sparse angle scanning, as another effective measure to reduce radiation dose, has certain advantages over the tube current reduction method, such as being simple and easy to implement, adaptable to a wide range of scenarios, small amount of data, and fast calculation speed. However, sparse angle scanning will lead to missing projection angles, which will cause the reconstructed image to be full of star-shaped artifacts, affecting the doctor's diagnosis.

[0004] Without changing the original hardware, there are three main approaches to improving imaging performance in sparse-angle scanning. First, from the perspective of CT projection data, processing the original data or logarithmically transformed projection data to restore missing projection data can yield highly consistent original data. However, due to the high sensitivity of projection data, signal supplementation can easily lead to deviations, resulting in erroneous information in the reconstructed image, which limits the effectiveness of this approach. Second, from the perspective of image reconstruction algorithms, statistical iterative reconstruction algorithms based on prior information constraints, for example, offer good sparse-angle reconstruction performance. However, these algorithms suffer from numerous hyperparameters, high algorithmic complexity, unstable prior information, and poor generalization, limiting their practical clinical application. Third, from the perspective of CT image data, image processing techniques can be used to suppress streak artifacts in sparse-angle CT images, resulting in high signal-to-noise ratio and high-contrast images. However, this processing can easily lead to oversmoothing, loss of detail, and decreased contrast. While numerous challenges remain in sparse-angle CT imaging, they will become important indicators for future CT research and a major development direction for X-ray imaging.

[0005] With the improvement of computing power and the generation of large-scale data, data-driven learning algorithms have shown excellent performance in many fields, which has also brought new opportunities for the development of medical image reconstruction algorithms. In the environment of big image data, methods based on deep learning are also the main direction of development of sparse angle CT imaging. For example, in order to improve the suppression and removal of streak artifacts, people have designed a large number of deep neural networks. This type of method mainly constructs a network through sample training and has received widespread attention in the fields of feature recognition, classification and image restoration. However, in the process of low-level data processing, conventional convolutional neural networks have limited feature extraction categories. For example, how to distinguish between tissue structure information and interfering artifact information, and how to ensure that structural information is not lost while suppressing artifact information, are key issues in the process of sparse angle CT image processing. For this reason, the sparse angle CT image artifact suppression method based on structure enhancement and artifact estimation proposed in the present invention can better balance the suppression of artifacts and the retention of tissue details.

[0006] After searching, the Chinese patent application number 202110331020.7 discloses a sparse angle CT reconstruction method based on wavelet multi-scale convolution feature coding. This method uses the wavelet multi-scale convolution feature coding prior information of high-quality CT images to guide and constrain the iterative reconstruction results, improve the quality of sparse angle CT reconstruction images, and obtain better image results in the embodiments. However, the inventive method belongs to the iterative reconstruction method, which has more parameter adjustments, and the amount of sample data used is limited. The method of the present invention realizes the mapping between end-to-end image data, does not rely on the original projection data of the scan, and has the advantages of high flexibility, few parameters, high efficiency and easy application. Summary of the Invention

[0007] 1. Technical problem to be solved by the invention

[0008] The purpose of the present invention is to overcome the problems of loss of tissue details, many residual artifacts, low contrast and low overall image quality after sparse angle CT image processing in the prior art. It is intended to provide a sparse angle CT image artifact suppression method based on structure enhancement and artifact estimation, and construct a network dedicated to sparse angle CT image artifact suppression, called Structure Enhanced Artifact Suppression Network (SEAS-Net). The network of the present invention is based on a U-shaped structure, and adopts sparse coding and decoding modules of multiple scales to extract different types of feature information, and adopts a feature fusion strategy and a hybrid loss training mode to retain the original tissue structure information as much as possible in improving the artifact suppression performance, the accuracy of the processed image, and the quality of sparse angle CT images, ultimately reducing additional radiation for patients and increasing the benefits of diagnosis and treatment.

[0009] 2. Technical solution

[0010] In order to achieve the above object, the technical solution provided by the present invention is:

[0011] The invention provides a method for suppressing artifacts in sparse angle CT images based on structure enhancement and artifact estimation, comprising the following steps:

[0012] Step 1: Collect high-quality CT image data and convert the collected high-quality CT image u f Use simulation software to obtain sparse angle CT images u s , for the collected high-quality CT images u f Use Gaussian filter to extract edge structure and obtain edge structure image u e , forming a data set {u s ,u f ,u e}, finally the collected data set is randomly divided into training set and test set according to the proportion;

[0013] Step 2: Construct a structure enhancement artifact suppression network for sparse angle CT image processing. The network mainly includes an artifact feature extraction subnetwork and a structure feature extraction subnetwork. Its function is to input the sparse angle CT image u s Generate artifact-suppressed CT image u r ;

[0014] Step 3: Construct an image restoration loss function and a structural loss function and combine them to form a joint loss function to improve the ability of artifact suppression and structural detail preservation;

[0015] Step 4: Using the training set data and the joint loss function to train the structured enhanced artifact suppression network to obtain trained network parameters;

[0016] Step 5: The user inputs the sparse angle CT image in the test set into the trained structure enhancement artifact suppression network, outputs the artifact-suppressed CT image, and compares it with the high-quality CT image.

[0017] Furthermore, the high-quality CT image u in step 1 f and sparse angle CT images u obtained using simulation software s Equal size, matching position, only sparse angle CT image u s The scanning angles are small, and the collected data sets are randomly divided into training sets and test sets in proportion. The number of data sets in the training set is not less than 90% of the total, and the number of data sets in the test set is not more than 10% of the total.

[0018] Furthermore, in step 2, the architecture flow of the structure enhancement artifact suppression network is as follows: first, the artifact feature extraction subnetwork is used to extract artifact feature information; the structural feature extraction subnetwork is designed to extract tissue structure feature information in the sparse angle CT image; then, the artifact feature information is combined with the tissue structure feature information, and the artifact tissue structure component is obtained through convolution operation; finally, through residual jump connection, the original sparse angle CT image and the artifact tissue structure component are added to obtain the artifact suppressed CT image.

[0019] Furthermore, the network architecture of the artifact feature extraction subnetwork in step 2 is the same as that of the structural feature extraction subnetwork.

[0020] Furthermore, the artifact feature extraction subnetwork in step 2 adopts a U-shaped structure, including 4 feature scales, with the left side consisting of a sparse coding module and the right side consisting of a sparse coding module and a decoding module.

[0021] Furthermore, the sparse coding module in step 2 adopts a residual structure, which is a convolutional layer followed by five ReLU operations and a convolutional layer with jump connections; the decoding module adopts a flat structure, which is a convolutional layer, a ReLU operation and a convolutional layer.

[0022] Furthermore, the convolution kernel size of the structure enhancement artifact suppression network in step 2 is 3×3, the downsampling layer consists of a convolution layer with a stride of 2, and the upsampling layer is a transposed convolution layer with a stride of 2; after each downsampling, the number of channels of the convolution network doubles, and the initial number of channels is 32.

[0023] Furthermore, the image restoration loss function constructed in step 3 is used to calculate the high-quality CT image u f And the output artifact-suppressed CT image u r The mean square error and structural similarity between: the structural loss function is the edge structure image u e The image u output by the structural feature extraction subnetwork and convolution operation se The mean square error between ; therefore, the joint loss function L is defined r for:

[0024] L r =MSE(u f ,u r )+α(1-SSIM(u f ,u r ))+βMSE(u e ,u se )

[0025] In the above formula, u f For high-quality CT images, u ris the output artifact-suppressed CT image, u se is the image output by the structural feature extraction subnetwork and convolution operation, u e For high-quality CT images u f Extracted edge structure image, MSE(·) is the mean square error calculation function, SSIM(·) is the structure similarity calculation function, α and β are loss weights.

[0026] Furthermore, in step 4, the network model parameters are iteratively updated using a small-batch stochastic gradient descent algorithm. When the change in the joint loss function value before and after the training cycle is within 2%, the training is stopped to obtain the network model parameters. In step 5, the trained network is used to achieve artifact suppression in sparse angle CT images. The specific process is as follows: the data in the test set are put into the trained network one by one, the sparse angle CT image is used as the network input, and the network output is the artifact-suppressed CT image; the artifact-suppressed CT image is numerically compared with the high-quality CT image, and the objective evaluation index is calculated.

[0027] 3. Beneficial effects

[0028] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0029] (1) The sparse angular CT image artifact suppression method based on structure enhancement and artifact estimation of the present invention uses a sparse coding module to implement encoding of features of different scales on the basis of a U-shaped structure, and uses a decoding module to implement decoding of features of different scales. The design of the encoding and decoding modules has the advantages of good interpretability and strong generalization ability.

[0030] (2) The sparse angle CT image artifact suppression method based on structure enhancement and artifact estimation of the present invention comprehensively utilizes two sub-networks to extract artifact features and tissue structure feature information. Finally, through information fusion, it can provide a highly accurate artifact-suppressed CT image. Different loss functions are used in the feature extraction process to promote the learning of each sub-network.

[0031] (3) The present invention's artifact suppression method for sparse angle CT images based on structure enhancement and artifact estimation, and its related experimental results verify that under CT images scanned at different sparse angles, the present invention's method is compared with the artifact suppression method for sparse angle CT reconstruction based on generative adversarial networks (Wasserstein Generative Adversarial Networks with Gradient Penalty, referred to as WGAN-GP) and the sparse angle CT reconstruction method based on wavelet multi-scale convolutional feature coding (Wavelet domain Multi□scale Convolutional sparse coding constrainedReconstruction, referred to as WMCR). The reconstructed image has better visual effect and contrast, and has achieved relatively satisfactory results in artifact suppression and anatomical tissue boundary preservation. It has great advantages in applications such as clinical screening, diagnosis and treatment, and can improve inspection efficiency and reduce radiation damage to patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is an overall flow chart of a method for suppressing artifacts in sparse angle CT images based on structure enhancement and artifact estimation according to an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of the structure of a structure enhancement artifact suppression network according to an embodiment of the present invention;

[0034] Figure 3 The artifact suppression result diagram of the sparse angle CT image (144 scanning angles) in the embodiment (a: high-quality CT image; b: sparse angle CT image; c: WMCR result diagram; d: WGAN-GP result diagram; e: SEAS-Net result diagram);

[0035] Figure 4 A local enlarged view of the artifact suppression result image of the sparse angle CT image (144 scanning angles) in the embodiment (a: high-quality CT image; b: sparse angle CT image; c: WMCR result image; d: WGAN-GP result image; e: SEAS-Net result image);

[0036] Figure 5 The artifact suppression result diagram of the sparse angle CT image (96 scanning angles) in the embodiment (a: high-quality CT image; b: sparse angle CT image; c: WMCR result diagram; d: WGAN-GP result diagram; e: SEAS-Net result diagram);

[0037] Figure 6The following are partial enlarged images of artifact suppression results of sparse angle CT images (96 scanning angles) in the embodiment (a: high-quality CT image; b: sparse angle CT image; c: WMCR result image; d: WGAN-GP result image; e: SEAS-Net result image);

[0038] Figure 7 Schematic diagram of Profile curves of CT images at 144 scanning angles in the embodiment;

[0039] Figure 8 Schematic diagram of Profile curves of CT images at 96 scanning angles in the embodiment. DETAILED DESCRIPTION

[0040] In order to further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings.

[0041] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0042] The sparse angle CT image artifact suppression method based on structure enhancement and artifact estimation proposed in the present invention can be called the Structure Enhanced Artifact Suppression Network (SEAS-Net) method. It is based on the U-shaped network structure, uses sparse coding and decoding modules of multiple scales to extract different types of feature information, and adopts feature fusion strategy and hybrid loss training mode to retain the original tissue structure information as much as possible in improving the artifact suppression performance, improve the accuracy of the processed image, and improve the quality of sparse angle CT images. Specifically, it mainly includes an artifact feature extraction subnetwork and a structural feature extraction subnetwork with the same network architecture; then constructs an image restoration loss function and a structural loss function and jointly trains each subnetwork to form a trained structure enhancement artifact suppression network; finally, the user inputs the sparse angle CT image to be processed into the trained structure enhancement artifact suppression network, and outputs a CT image with suppressed artifacts. The network of the present invention comprehensively utilizes the extraction of artifact features and tissue structure feature information by the two subnetworks, and finally, through the fusion of information, can provide a CT image after artifact suppression with high accuracy.

[0043] The inventors of this application have also been committed to the research of sparse angle CT image reconstruction and processing technology, and have achieved certain results. For example, the application with application number 201810557240.X discloses a sparse angle CT imaging method based on a convolutional neural network. The application adopts two cascaded neural networks to realize the reconstruction process of sparse angle CT projection data into high-quality CT images, and the output is a CT image, which belongs to the image reconstruction method. The input of the present invention is sparse angle CT image data, and the convolutional neural network with combined structure enhancement and artifact suppression realizes the effective suppression of artifacts based on sparse angle CT images, which belongs to the CT image post-processing or restoration method.

[0044] The present invention will be further described below with reference to specific embodiments.

[0045] Example

[0046] like Figure 1 As shown, the sparse angle CT image artifact suppression method based on structure enhancement and artifact estimation of this embodiment has the following specific steps:

[0047] Step 1: Collect high-quality CT image data and convert the collected high-quality CT image u f Use simulation software to obtain sparse angle CT images u s , for the collected high-quality CT images u f Use Gaussian filter to extract edge structure and obtain edge structure image u e , forming a data set {u s ,u f ,u e}, and finally the collected data set is randomly divided into training set and test set according to proportion.

[0048] Specifically, in the data set, high-quality CT images u f and sparse angle CT images u obtained using simulation software s Equal size, matching position, only sparse angle CT image u s The scanning angles are small, and the collected data sets are randomly divided into training sets and test sets in proportion. The number of data sets in the training set is not less than 90% of the total, and the number of data sets in the test set is not more than 10% of the total.

[0049] Step 2: Construct a structure enhancement artifact suppression network for sparse angle CT image processing. The network mainly includes an artifact feature extraction subnetwork and a structure feature extraction subnetwork. Its function is to input the sparse angle CT image u s Generate artifact-suppressed CT image u r ;

[0050] Specific structure enhancement artifact suppression networks such as Figure 2 As shown in the figure, the network architecture flow is as follows: First, an artifact feature extraction subnetwork is designed to extract artifact feature information, and a structural feature extraction subnetwork is designed to extract tissue structural feature information from sparse angle CT images. The artifact feature extraction subnetwork and the structural feature extraction subnetwork share the same network architecture. Then, the artifact feature information is combined with the tissue structural feature information and the artifact tissue structural component is obtained through a convolution operation. Finally, through residual skip connections, the original sparse angle CT image and the artifact tissue structural component are added to obtain an artifact-suppressed CT image. The artifact feature extraction subnetwork adopts a U-shaped structure, including four feature scales. The left side consists of a sparse coding module, and the right side consists of a sparse coding module and a decoding module. The sparse coding module adopts a residual structure, which is a convolutional layer followed by five ReLU operations and a convolutional layer. The decoding module adopts a straight structure, which is a convolutional layer, a ReLU operation, and a convolutional layer. The convolution kernel size is 3×3. The downsampling layer consists of a convolution layer with a stride of 2, and the upsampling layer is a transposed convolution layer with a stride of 2. After each downsampling, the number of channels of the convolutional network doubles, and the initial number of channels is 32.

[0051] Step 3: Construct an image restoration loss function and a structural loss function and combine them to form a joint loss function to improve the ability of artifact suppression and structural detail preservation;

[0052] Specifically, the constructed image restoration loss function is to calculate the high-quality CT image u f And the output artifact-suppressed CT image u r The mean square error and structural similarity between them; the structural loss function is the edge structure image u e And the image u output by the structural feature extraction subnetwork and convolution operation se The mean square error between ; therefore, the joint loss function L is defined r for:

[0053] L r =MSE(u f ,u r )+α(1-SSIM(u f ,u r ))+βMSE(u e ,u se )

[0054] In the above formula, u f For high-quality CT images, u r is the output artifact-suppressed CT image, u se is the image output by the structural feature extraction subnetwork and convolution operation, u e For high-quality CT images u fFor the extracted edge structure image, MSE(·) is the mean squared error calculation function, SSIM(·) is the structural similarity calculation function, and α and β are loss weights. Therefore, when training the structure enhancement artifact suppression network, it is desirable that the artifact-suppressed CT image be as close to the true high-quality CT image as possible, that is, the smaller the mean squared error value, the better, and the structural similarity value is as large as possible. At the same time, to ensure effective training of the structural feature extraction subnetwork, it is desirable that the structural features extracted by the structural feature extraction subnetwork be similar to the edge structure of the true high-quality CT image, with the mean squared error value as small as possible.

[0055] Step 4: Using the training set data and the joint loss function to train the structured enhanced artifact suppression network to obtain trained network parameters;

[0056] Specifically, the trained structure enhancement artifact suppression network is the entire network after the sub-networks are merged and connected, such as Figure 2 As shown in the figure, in order to protect the detailed texture features of the processed image, the training adopts the mini-batch stochastic gradient descent algorithm to iteratively update the parameters of the structure enhancement artifact suppression network model. When the change of the joint loss function value before and after the training cycle is within 2%, the training can be stopped to obtain the network model parameters;

[0057] Step 5: The user inputs the sparse angle CT image in the test set into the trained structure enhancement artifact suppression network, outputs the artifact-suppressed CT image, and compares it with the high-quality CT image.

[0058] Specifically, artifact suppression in sparse angle CT images is achieved using a trained network. The process is as follows: the data in the test set are put into the trained network one by one, and the sparse angle CT image is used as the network input. The network output is the CT image with artifact suppression; the artifact-suppressed CT image is numerically compared with the high-quality CT image, and the objective evaluation index can be calculated using the high-quality CT image as a reference.

[0059] Effectiveness Evaluation Criteria

[0060] In the embodiment, high-quality CT image data of 10 patients, approximately 6,000 CT images, were first collected. The collected CT image data were all from the same device (Somatom Definition AS+CT). The specific scanning parameters were: tube voltage 100 kVp, tube current 360 mAs, number of detectors 736 × 64, and the size of each detector unit was 1.2856 × 1.0947 mm. 2The distances from the source to the object center and the detector center were 595 mm and 1086 mm, respectively. 1152 projections were scanned with a pitch of 0.6, and high-quality CT images were reconstructed using the conventional FBP reconstruction algorithm. Furthermore, all CT images were simulated using the same scanning parameters (only the number of projections varied) and sparse angle scans were performed and reconstructed to obtain sparse angle CT images. A Gaussian filter was applied to each high-quality CT image to obtain an edge structure image. Two different sparse angle scans (144 and 96, respectively) were used to generate two different sparse angle CT image datasets: the 144-dataset and the 96-dataset. These datasets were then divided into training and test sets with a 9:1 ratio. In training the 144-dataset, the weight parameters α and β of the joint loss function were set to 6.5 and 0.8, respectively; in training the 96-dataset, the weight parameters α and β of the joint loss function were set to 10 and 1, respectively.

[0061] In the accompanying drawings, all CT image display windows are abdominal windows, i.e., [-250HU, 150HU] (Housfield Units, HU). By comparing the visual effects of the processed CT images, it can be seen that the CT images processed by the SEAS-Net network of the present invention are superior to those of the WMCR and WGAN-GP methods. The strip artifacts caused by sparse angle scanning are basically suppressed, the detailed structure is revealed, and the image quality is high, especially the local magnification, the details of which are clearer; by comparing the selected regions of interest, such as the structural details of the image, contrast, anatomical tissue texture, and residual intensity of noise artifacts, it is believed that the present invention can effectively improve the quality of CT images. By comparing the profile curves of the selected regions of interest, it can be found that the CT value of the CT image processed by the present invention is closer to that of the high-quality CT image, and the deviation of its intensity is minimal. Furthermore, the results of the implementation case are quantitatively compared using reference evaluation indicators, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The calculation methods of PSNR and SSIM are as follows:

[0062]

[0063]

[0064] where u r To suppress artifacts in CT images, u f is a high-quality CT image used for simulation, N is the total number of pixels in a single image, and L is u f The maximum pixel value of μ T and μ rRepresents CT image u r and u f The average value of the total pixel CT value; σ T and σ r Represents CT image u r and u f The standard deviation of the total pixel CT value, σ Tr is the CT image u r and u f The covariance of , constants C1 and C2 are default constants.

[0065] Processing result evaluation

[0066] The sparse angle CT image results after processing by different methods are shown in the figure below. Figure 3 and Figure 5 As shown. It can be seen from the figure that the original sparse angle CT image contains many artifacts, which almost cover the details of the entire CT image and are difficult to identify. After being processed by the WMCR method, the WGAN-GP method and the SEAS-Net method of the present invention, the image has been significantly improved, most of the artifacts have been removed, especially the artifacts with larger intensity have also been well suppressed, and the original CT tissue structure information has been revealed, which is more conducive to the physician's observation of details. In the 144-data test, by comparing the high-quality CT image and the WGAN-GP result image, it can be found that the resolution of the WGAN-GP image is lower, and the tissue structure in some areas is blurred, such as the blurred phenomenon in the liver vein tissue area in the figure, and the artifact phenomenon at the lung ribs; by comparing the reference image and the WGAN-WMCR image, it can be found that some artifacts still remain in the WMCR result image, and the artifacts with larger intensity have not been completely removed, such as the area near the ribs; and by comparison, it can be seen that the method of the present invention is superior to the comparison method in terms of artifact suppression, tissue structure destruction, and detail retention. By observing and comparing the local enlarged image ( Figure 4 and Figure 6 ), it can be seen that the CT image processed by the present invention can better retain fine tissue details, and the image visual texture is closer to a high-quality CT image.

[0067] High-quality CT images were used as reference images for quantitative evaluation to calculate the PSNR and SSIM of CT images processed by different methods, in order to quantitatively analyze the image quality of CT after artifact suppression. The results are shown in Table 1 below. The quantization values in the table are the mean and variance of the test set image quantization, listed in the form of mean ± variance. As can be seen from Table 1, under sparse angle scanning conditions, the PSNR and SSIM of CT images decreased significantly. The fewer the scanning angles, the lower the average PSNR and SSIM. The average PSNR and SSIM of CT images processed by the SEAS-Net method of the present invention were higher than those of the comparison method in the 144-dataset and 96-dataset, and the changes in the quantization values in the data sets were smaller. Through the specific CT value change curve (region of interest profile curve), the real signal intensity changes of different images can be more intuitively discovered. Figure 7 and Figure 8 is a profile curve diagram of the region of interest. It can be seen from the figure that the CT image processed by the present invention is closer to the high-quality CT image, and its intensity deviation is smaller.

[0068] Table 1

[0069]

[0070] In summary, the method of the present invention can effectively suppress artifacts in sparse angle CT images to obtain higher quality CT images, and has certain application prospects.

[0071] The above is a schematic description of the present invention and its embodiments. This description is not restrictive and is only one embodiment of the present invention. Therefore, if a person skilled in the art is inspired by this description and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without creatively designing them, they shall fall within the scope of protection of the present invention.

Claims

1. A sparse angle CT image artifact suppression method based on structure enhancement and artifact estimation, characterized in that: The following steps are involved: Step 1: Collect high-quality CT image data and collect high-quality CT images Using simulation software to obtain sparse angle CT images , for the collected high-quality CT images Use filters to extract edge structures and obtain edge structure images , forming a data set ,Finally, the collected data set is randomly divided into training set and test set; Step 2: Construct a structure enhancement artifact suppression network for sparse angle CT image processing. The network includes an artifact feature extraction subnetwork and a structure feature extraction subnetwork. Its function is to input sparse angle CT images. Generate artifact-suppressed CT images ; The architecture of the structure-enhanced artifact suppression network is as follows: First, an artifact feature extraction subnetwork is designed to extract artifact feature information; a structural feature extraction subnetwork is designed to extract tissue structural feature information from sparse-angle CT images; then, the artifact feature information is combined with the tissue structural feature information, and the artifact tissue structural component is obtained through a convolution operation; finally, the original sparse-angle CT image and the artifact tissue structural component are added through residual skip connections to obtain the artifact-suppressed CT image. Step 3: Construct an image restoration loss function and a structural loss function and combine them to form a joint loss function to improve the ability of artifact suppression and structural detail preservation; Constructing image restoration loss function to calculate high-quality CT images And the output artifact-suppressed CT image The mean square error and structural similarity between them; the structural loss function is the edge structure image The image output by the structural feature extraction subnetwork and convolution operation The mean square error between ; therefore, the joint loss function is defined as for: In the above formula, For high-quality CT images, is the output CT image with suppressed artifacts, The image output by the structural feature extraction subnetwork and convolution operation, For high-quality CT images Extracted edge structure image, is the mean square error calculation function, is the structural similarity calculation function, and is the loss weight; Step 4: Use the training set data and the joint loss function to train the structure-enhanced artifact suppression network to obtain the trained network parameters; Step 5: The user inputs the sparse angle CT image in the test set into the trained structure enhancement artifact suppression network, outputs the artifact-suppressed CT image, and compares it with the high-quality CT image.

2. The method for artifact suppression in sparse angle CT images based on structure enhancement and artifact estimation according to claim 1, characterized in that: High-quality CT images from step 1 and sparse angle CT images Equal in size, matching in position, only sparse angle CT images The scanning angle is small; finally, the collected data groups are randomly divided into training sets and test sets according to proportion, the number of data groups in the training set is not less than 90% of the total, and the number of data groups in the test set is not more than 10% of the total.

3. The method for artifact suppression in sparse angle CT images based on structure enhancement and artifact estimation according to claim 1, characterized in that: The artifact feature extraction subnetwork in step 2 has the same network architecture as the structural feature extraction subnetwork.

4. The method for artifact suppression in sparse angle CT images based on structure enhancement and artifact estimation according to claim 3, characterized in that: The artifact feature extraction subnetwork in step 2 adopts a U-shaped structure, including four feature scales. The left side is composed of a sparse coding module, and the right side is composed of a sparse coding module and a decoding module.

5. The method for artifact suppression in sparse angle CT images based on structure enhancement and artifact estimation according to claim 4, characterized in that: In step 2, the sparse coding module adopts a residual structure, which is a convolutional layer followed by five ReLU operations and a convolutional layer with jump connections; the decoding module adopts a flat structure, which is a convolutional layer, a ReLU operation and a convolutional layer.

6. The method for artifact suppression in sparse angle CT images based on structure enhancement and artifact estimation according to claim 1, characterized in that: In step 2, the convolution kernel size in the structure enhancement artifact suppression network is 3×3, the downsampling layer consists of a convolution layer with a stride of 2, and the upsampling layer is a transposed convolution layer with a stride of 2; after each downsampling, the number of channels of the convolution network doubles, and the initial number of channels is 32.

7. The method for artifact suppression in sparse angle CT images based on structure enhancement and artifact estimation according to claim 1, characterized in that: In step 4, the network model parameters are iteratively updated using the mini-batch stochastic gradient descent algorithm. When the change in the joint loss function value before and after the training cycle is within 2%, the training is stopped to obtain the network model parameters.

8. The method for artifact suppression in sparse angle CT images based on structure enhancement and artifact estimation according to claim 7, characterized in that: In step 5, the trained network is used to achieve artifact suppression in sparse angle CT images. The specific process is as follows: the data in the test set are put into the trained network one by one, and the sparse angle CT image is used as the network input. The network output is the CT image with suppressed artifacts; the artifact-suppressed CT image is numerically compared with the high-quality CT image, and the objective evaluation index is calculated.

Images