Distinguished convolutional dictionary network-based sparse angle CT (Computed Tomography) artifact removal method

Through the differentiated convolutional dictionary network combined with deep learning, the features and artifact features of sparse angle CT images are extracted, and the problem of incomplete artifact removal of sparse angle CT images is solved, and the output of high-quality images is achieved, which is suitable for clinical diagnosis. Reduces radiation exposure and improves diagnosis and treatment effectiveness.

CN119048614APending Publication Date: 2024-11-29ANHUI POLYTECHNIC UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410913004.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-09
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing sparse angle CT image artifact removal is not thorough, the image quality is not high, the feature extraction ability is limited, the encoding and calculation are complex, and the deep learning methods are opaque to process and poor adaptability, making it difficult to directly deploy in clinical practice.

Method used

Using a differentiated convolutional dictionary network method, the image features and artifact features in sparse angle CT images are extracted differently through the combination of convolutional dictionary representation and deep learning, so as to enhance feature learning ability, and gradually remove image prior constraints. Artifact, output high-quality CT images with high contrast, high resolution, less noise, and less artifact.

Benefits of technology

It significantly improves the quality of sparse angle CT images, effectively removes strip and band artifacts, has high contrast in image details, is suitable for clinical diagnosis, reduces radiation exposure to patients, and improves diagnosis and treatment effects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048614A_ABST
    Figure CN119048614A_ABST
Patent Text Reader

Abstract

The invention discloses a sparse angle CT (Computed Tomography) artifact removal method based on a differentiated convolutional dictionary network, and belongs to the technical field of computed tomography. The invention specifically discloses a differentiated convolutional dictionary network. The differentiated convolutional dictionary network comprises an image feature extraction sub-network, an artifact feature extraction sub-network and an image prior sub-network, the image feature extraction sub-network adopts a convolutional dictionary form to extract and express CT image anatomical structure features; the artifact feature extraction sub-network is the same as the image feature extraction sub-network and is used for extracting and expressing sparse angle CT image strip-shaped artifact features; the image prior sub-network adopts a residual network to carry out prior constraint on the CT image; through cascade processing of the three sub-networks, artifacts in the sparse angle CT image are removed, and the image quality is improved. According to the method, artifact removal of the sparse angle CT image can be realized, so that the image after artifact removal is better in quality and less in artifact interference, and clinical diagnosis and analysis are more facilitated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer tomography technology, and more specifically to a sparse angle CT artifact removal method based on a discriminative convolutional dictionary network. Background Art

[0002] Computed Tomography (CT) technology relies on the different degrees of absorption of X-ray beams by the human body to achieve high-precision imaging of various parts of the human body. Its core principles cover the acquisition of digital information, the determination of the X-ray absorption coefficient of each tissue in the scanning plane, and the generation of the final CT grayscale image. In order to reduce the radiation damage of X-rays to patients during CT scanning, two strategies are usually adopted: reducing the tube current and sparse angle scanning. Compared with reducing the tube current, sparse angle scanning has shown more advantages, such as a wider range of applicable scenarios, a small amount of data required, and fast calculation speed. However, this strategy also has challenges in imaging, that is, the lack of projection angles may cause star-shaped artifacts in the reconstructed image, thereby affecting the accuracy of the doctor's diagnosis. Therefore, researchers and engineering designers continue to work hard to optimize the imaging effect under sparse angle scanning through various innovative means, aiming to improve the accuracy and reliability of diagnosis and bring safer and more effective medical services to patients.

[0003] In the field of sparse angle CT imaging, the relevant algorithm strategies usually include the following three categories: 1) projection domain restoration algorithm, which operates directly on the projection data to restore the missing information; 2) image domain post-processing algorithm, which performs post-processing on the basis of the preliminary reconstructed image to optimize the image quality without the help of projection data; 3) iterative reconstruction algorithm, which gradually approaches the ideal reconstruction result through an iterative optimization process. With the rise of large data sets and large sample libraries, deep learning technology has shown excellent application prospects in image processing after sparse angle CT reconstruction. Learning-based image domain post-processing methods are not only easy to implement, but also do not directly rely on the original scan data and have excellent scalability. Therefore, they have become a hot topic of research in recent years. In recent years, although domestic and foreign scholars have proposed many new concepts, new theories and new methods, and have achieved a large number of outstanding processing effects, the relevant imaging effects are still insufficient, the algorithm has poor adaptability, and it is difficult to be directly deployed in clinical practice. Some theories still have some shortcomings: for example, the traditional sparse representation-based learning model method has limited feature extraction capabilities and large encoding calculation complexity; block aggregation artifacts are prone to occur in the feature representation process, and the consistency of pixels between image blocks is ignored. Deep learning needs to be designed through repeated empirical trial and error, and its processing is opaque. The generalization, stability and interpretability of the model are also the main deficiencies faced by the field of deep imaging.

[0004] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0005] 1. Technical problem to be solved by the invention

[0006] The purpose of the present invention is to overcome the problems of unclean artifact removal, low restoration quality, decreased image details and contrast in sparse angle CT images in the prior art, and to provide a sparse angle CT artifact removal method based on a discriminative convolutional dictionary network, referred to as a discriminative convolutional dictionary network (DCDN). The method is to improve the feature information perception, encoding and decoding capabilities by performing discriminative convolutional feature learning on different types of features in sparse angle CT images without changing the existing CT hardware cost, and to obtain rich prior knowledge from a large amount of data. At the same time, deep learning is used to obtain image priors in sample data to jointly promote and improve the artifact removal effect, improve the quality of processed CT images, and output high-contrast, high-resolution, low-noise, and low-artifact high-quality CT images, thereby reducing additional radiation for patients and increasing the benefits of diagnosis and treatment.

[0007] 2. Technical solution

[0008] In order to achieve the above object, the technical solution provided by the present invention is:

[0009] The present invention proposes a method based on a discriminative convolutional dictionary network for artifact removal in sparse angle CT images. The method of the present invention combines the advantages of convolutional dictionary representation and the powerful representation ability of deep learning to establish an interpretable discriminative convolutional dictionary network model to discriminately extract image features and artifact features in sparse angle CT images, enhance the learning ability of different types of features, and at the same time use deep learning to obtain image priors in sample data to jointly promote and improve the artifact removal effect, and improve the quality of processed CT images, so as to output high-contrast, high-resolution, low-noise, and low-artifact high-quality CT images, reduce additional radiation for patients, and increase diagnosis and treatment benefits.

[0010] Specifically, a sparse angle CT artifact removal method based on a discriminative convolutional dictionary network of the present invention comprises the following steps:

[0011] Step 1: Obtain matching conventional CT images and sparse angle CT images to construct a training data set;

[0012] Step 2: Based on the convolutional dictionary sparse representation model, a discriminative convolutional dictionary network is designed for sparse angular CT artifact removal;

[0013] Step 3: Use the training data set to train the discriminative convolutional dictionary network to obtain the parameters of the entire network model;

[0014] Step 4: Input the sparse angle CT image to be processed into the trained discriminative convolutional dictionary network to obtain a high-quality CT image with artifacts removed.

[0015] As a further improvement of the present invention, the basic steps for acquiring data in step 1 are to downsample the projection data of the conventional scan and then reconstruct the data to obtain a sparse angle CT image. The reconstruction method is the same as the conventional CT image reconstruction method. The conventional CT images and the sparse angle CT images in the constructed training data set are matched.

[0016] As a further improvement of the present invention, the discriminative convolutional dictionary network constructed in step 2 includes multiple cascade modules, each module includes an image feature extraction subnetwork, an artifact feature extraction subnetwork and an image prior subnetwork.

[0017] As a further improvement of the present invention, the discriminative convolutional dictionary network established in step 2 is cascaded 20 times in sequence to achieve full extraction of image and artifact features.

[0018] As a further improvement of the present invention, the image feature extraction subnetwork structure in step 2 is: a convolution layer convolves the input image feature map, and then subtracts it from the input image signal, then uses another convolution layer to convolve and add it to the input image feature map, and finally uses batch normalization and ReLU activation function in sequence to output the updated image feature map.

[0019] As a further improvement of the present invention, the structure of the artifact feature extraction subnetwork in step 2 is the same as the structure of the image feature extraction subnetwork, the input is the artifact feature map and the artifact map signal, and the output is the updated artifact feature map.

[0020] As a further improvement of the present invention, the image prior subnetwork structure in step 2 is: first, an image feature map output by a convolutional layer convolutional image feature extraction subnetwork and an artifact feature map output by a convolutional layer convolutional artifact feature extraction subnetwork are used to obtain an updated image signal and an artifact map signal, respectively; then, the artifact map signal is subtracted from the sparse angle CT image, and combined with the updated image signal through a Concat layer; finally, the combined information is input into the residual network of three residual blocks and processed to obtain a processed image; the structure of the residual block is a convolutional layer, a batch normalization and ReLU activation function operation, a convolutional layer, a convolutional layer, a batch normalization operation and a jump connection.

[0021] As a further improvement of the present invention, in step 3, the image loss function is constructed by using the mean square error loss of the image signal and the VGG perceptual loss during network training, the artifact loss function is constructed by using the mean square error loss of the artifact image signal, and the network model parameters are iteratively updated by the Adam algorithm.

[0022] As a further improvement of the present invention, in step 3, the weight of the mean square error loss in the image loss is 1, the weight of the VGG perceptual loss is 0.02, and the weights of the image loss and the artifact loss are the same during training; the initial input image signal is a sparse angular CT image, and the initial input artifact map signal, the initial input image feature map, and the initial input artifact feature map are all set to 0.

[0023] 3. Beneficial effects

[0024] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0025] The present invention discloses a discriminative convolutional dictionary network for sparse angle CT artifact removal, the network includes 20 cascade modules, each module includes an image feature extraction subnetwork, an artifact feature extraction subnetwork and an image prior subnetwork; wherein the image feature extraction subnetwork is mainly used for encoding image features and extracting image feature maps, and the structure of the artifact feature extraction subnetwork is the same as that of the image feature extraction subnetwork, and is mainly used for encoding artifact features and extracting artifact feature maps; in addition, by adding an image prior subnetwork, the image features are further constrained to improve the quality of the processed image; the strip and band artifacts in the sparse angle CT image are gradually removed in the form of a cascade, and a high-quality CT image is output. The method of the present invention can realize the artifact removal in the sparse angle CT image without relying on projection data, and realizes the mapping between the low-quality sparse angle CT image and the high-quality CT image through the cascade of the network, so that the image structure details after the artifact removal have high contrast and better image quality, and the result is more conducive to clinical diagnosis and analysis. The experimental results verify that in the artifact removal of sparse angle CT images with 120 scanning angles and 60 scanning angles, the method of the present invention (DCDN) is more effective in removing strip and band artifact information in sparse angle CT images, improving the quality of the original image without affecting the tissue contrast, compared with the existing distinctive feature representation method (Discriminative Feature Representation, referred to as DFR), residual convolutional neural network (ResNet) and adaptive convolutional dictionary network (ACDN), and has no effect on tissue contrast. The CT image after artifact removal has a better visual effect and is more conducive to the diagnosis of clinical physicians. Based on previous research, the method of the present invention starts from the perspective of convolutional sparse coding, draws on the advantages in the fields of dictionary learning and deep learning, and designs a discriminative convolutional dictionary network that is interpretable and can discriminate between image information features and artifact information features. At the same time, it also introduces prior image constraints to achieve more efficient artifact removal and obtain higher quality CT images. The present invention proposes an innovative and practical sparse angle CT artifact removal strategy, which is expected to bring significant improvements to domestic hospital imaging departments and CT manufacturers. Through this strategy, sparse angle imaging technology will be able to be more widely used in actual diagnosis and treatment, reducing unnecessary radiation exposure for patients, while improving the diagnosis and treatment effects, and bringing greater health benefits to patients. Therefore, the present invention has extremely high application value and broad promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1It is a flow chart of the sparse angle CT artifact removal method provided in an embodiment of the present invention;

[0027] Figure 2 A structural diagram of a discriminative convolutional dictionary network provided in an embodiment of the present invention;

[0028] Figure 3 The conventional CT image and the sparse angle CT image of 120 projections used for testing in the embodiment of the present invention (a: conventional CT image; b: sparse angle CT image);

[0029] Figure 4 The results of using different methods to remove artifacts in sparse angle CT images of 120 projections in the embodiments of the present invention (a: DFR; b: ResNet; c: ACDN; d: DCDN);

[0030] Figure 5 Different methods are used in the embodiments of the present invention to remove artifacts in sparse angle CT images of 120 projections, and the error diagram between the obtained result image and the conventional CT image (a: DFR; b: ResNet; c: ACDN; d: DCDN);

[0031] Figure 6 The conventional CT image and 60-projection sparse angle CT image used for testing in the embodiment of the present invention (a: conventional CT image; b: sparse angle CT image);

[0032] Figure 7 The results obtained by using different methods to remove artifacts in sparse angle CT images of 60 projections in the embodiments of the present invention (a: DFR; b: ResNet; c: ACDN; d: DCDN);

[0033] Figure 8 Different methods are used in the embodiments of the present invention to remove artifacts in sparse angle CT images of 60 projections, and the error diagram between the obtained result image and the conventional CT image (a: DFR; b: ResNet; c: ACDN; d: DCDN);

[0034] Fig. 9 and Fig.10 1 is a profile curve comparison diagram of images after artifacts are removed by different methods in the embodiments of the present invention, wherein Fig. 9 There are 60 projection results; Fig.10 There are 120 projection results. DETAILED DESCRIPTION

[0035] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings.

[0036] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0037] The present invention will be further described below in conjunction with the embodiments.

[0038] Example

[0039] Combination Figure 1-Figure 10 As shown, the process of a sparse angle CT artifact removal method based on a discriminative convolutional dictionary network in this embodiment is as follows Figure 1 As shown, the specific steps are as follows:

[0040] Step 1: Obtain matching conventional CT images and sparse angle CT images to construct a training data set;

[0041] Specifically, the basic requirement for constructing a training data set is to downsample the projection data of conventional scans according to the scanning angles. For example, if the projection data of conventional scans has 720 scanning angles, it can be downsampled in an evenly spaced manner, such as 6 times downsampling, to obtain sparse angle CT projection data of 120 scanning angles. The same reconstruction method is used to reconstruct the projection data of conventional scans and sparse angle CT projection data to obtain conventional CT images. and sparse angle CT images This ensures that the conventional CT images are matched with the sparse angle CT images. According to this strategy, a large amount of image data can be collected, among which the conventional CT images is the label image, sparse angle CT image For training images.

[0042] Step 2: Based on the convolutional dictionary sparse representation model, a discriminative convolutional dictionary network is designed for sparse angular CT artifact removal;

[0043] Specifically, we use a discriminative convolutional dictionary representation and an image prior constraint strategy to design an end-to-end discriminative convolutional dictionary network specifically for sparse angle CT image artifact removal. Through training, the network can remove training images (sparse angle CT images). ) in the image, making it close to the label (conventional CT image ). We call this network a discriminative convolutional dictionary network (DCDN for short). Figure 2 As shown in Figure 2, the discriminative convolutional dictionary network includes multiple cascade modules, each of which includes an image feature extraction subnetwork, an artifact feature extraction subnetwork, and an image prior subnetwork.

[0044] The main structure of the image feature extraction subnetwork is: a convolutional layer Convolution input image feature map V n , and then with the input image signal I n Subtract, and then use a convolutional layer Convolution is performed with the input image feature map V n Add, and finally use batch normalization and ReLU activation function in sequence to output the updated image feature map. The mathematical model of the image feature extraction subnetwork can be expressed as:

[0045]

[0046] The superscript n is the nth cascade module, V n is the input image feature map, the initial input image feature map V 0 =0, V n+1 is the updated image feature map, ReLU(g) is the ReLU activation function, BN(g) is the batch normalization operation, I n is the input image signal, the initial input image signal I 0 is a sparse angle CT image, and They are all convolutional layers of the image feature extraction subnetwork in the nth cascade module.

[0047] The structure of the artifact feature extraction subnetwork is the same as that of the image feature extraction subnetwork, and the input is the artifact feature map U n and artifact image signal A n , the output is the updated artifact feature map. The mathematical model of the artifact feature extraction sub-network can be expressed as:

[0048]

[0049] The superscript n is the nth cascade module, U n is the input artifact feature map, the initial input artifact feature map U 0 =0,U n +1 is the updated artifact feature map, ReLU(g) is the ReLU activation function, BN(g) is the batch normalization operation, A nis the input artifact image signal, the initial input artifact image signal A 0 =0, and They are all convolutional layers of the artifact feature extraction subnetwork in the nth cascade module.

[0050] The main structure of the image prior subnetwork is: First, a convolutional layer Image feature map V output by the convolutional image feature extraction subnetwork n+1 , a convolutional layer The artifact feature map U output by the convolutional artifact feature extraction subnetwork n+1 , respectively obtain the image signals of the intermediate steps and artifact image signal A n+1 ; Then, the artifact map signal is subtracted from the sparse angle CT image and combined with the image signal of the intermediate step through the Concat layer Combine; finally, the combined information is input into the residual network of the three residual blocks and processed to obtain the processed image. The structure of the residual block is a convolutional layer S r,1 , a batch normalization and ReLU activation function operation, a convolutional layer S r,2 , a convolutional layer S r,3 , a batch normalization operation and a skip connection. The mathematical model of the image prior subnetwork can be expressed as:

[0051]

[0052] α o =α i -BN(S r,3 (S r,2 (ReLU(BN(S r,1 (α i ))))))(4)

[0053] The superscript n is the nth cascade module. n (g) is the Concat layer, which connects features of different scales. ResNet n (g) is the residual network, I n+1 is the final updated image signal, A n+1 is the updated artifact image signal. i is the residual block input signal, α o is the residual block output signal, S r,1 , S r,2 and S r,3 They are all convolutional layers in the residual block, ReLU(g) is the ReLU activation function, and BN(g) is the batch normalization operation.

[0054] Finally, 20 identical modules are cascaded sequentially to achieve full extraction of image and artifact features, and the updated image signal is finally output as I 20 , that is, the processed high-quality CT image.

[0055] Step 3: Use the training data set to train the discriminative convolutional dictionary network to obtain the parameters of the entire network model;

[0056] Specifically, the training data, i.e., sparse angular CT images Input into the neural network, output high-quality CT image I after training data artifacts are removed 20 , and conventional CT images As the reference object, the image loss function is constructed using mean square error loss and VGG perceptual loss, the artifact loss function is constructed using artifact signal mean square error loss, and the network model parameters are iteratively updated through the Adam algorithm. Image loss function Loss I And artifact loss function Loss A Defined as:

[0057]

[0058] Where MSE(g) is the mean square error function, P vgg (g) is the perception error function, V gg (g) is the feature extraction operation of the VGG19 network, λ1 and λ2 are the mean square error loss weight and VGG perceptual loss, where the mean square error loss weight is 1 and the VGG perceptual loss weight is 0.02. The weights of image loss and artifact loss are the same during training. The initial input image signal I 0 is a sparse angle CT image, the initial input artifact image signal A 0 , the initial input image feature map V 0 and the initial input artifact feature map U 0 All are set to 0. Finally, through periodic training, the convolutional layer parameters of the DCDN network are obtained.

[0059] Step 4: Input the sparse angle CT image to be processed into the trained discriminative convolutional dictionary network to obtain a high-quality CT image with artifacts removed.

[0060] Specifically, the sparse angle CT images that need to be processed Input the DCDN network with the convolutional layer parameters obtained, and output a high-quality CT image with artifacts removed.

[0061] Effectiveness Evaluation Criteria

[0062] In order to verify the effect of the method of the present invention, the following comparative experiment was conducted: The experiment selected CT data scanned by the uCT-760 scanning device from United Imaging Medical for training and testing. All data are from clinical patient scans, including conventional CT data from 14 anonymous patients. The conventional CT image in this data set is the label, which is an image reconstructed from 1200 projection data (all using the FBP reconstruction algorithm). At the same time, projection data of 60 and 120 cases are equally extracted from the projection data, and sparse angle CT images are reconstructed. The scanning geometry of data acquisition is: cone-beam spiral scanning; the distance from the source to the detector and the isocenter are 950mm and 570mm respectively; the detector has 936×80 units, and the size of each unit is 1.548×1.405mm 2 The reconstructed image size is 512×512, and the reconstruction layer thickness is 1 mm.

[0063] Figure 3 are conventional CT images and 120-projection sparse angle CT images used for testing in the embodiment of the present invention, Figure 6 The conventional CT image and 60 projected sparse angle CT images used for testing in the embodiment of the present invention include the conventional CT image and the corresponding matching sparse angle CT images. To facilitate the display of image details, the display window width of the CT image is set to 400HU (Housfield Un its, HU), and the window level is set to 50HU. Figure 4 and Figure 7 To correspond to the results after artifact removal of sparse angle CT images, the comparison methods include the traditional discriminative dictionary learning DFR method, the ResNet network method based on deep learning, and the ACDN method based on adaptive convolutional dictionary network, which are used to compare the artifact removal effects of different methods. Figure 5 and Figure 8 This is the error map between the sparse angle CT image after artifact removal and the conventional CT image. To facilitate detail display, the error map display window width is set to 50HU and the window level is 25HU. The error map can be used to compare the effect of artifact removal and the difference between the processed result and the conventional CT image.

[0064] Visual Effects Evaluation

[0065] The visual effect is mainly evaluated qualitatively from the aspects of uniformity, tissue texture, detail contrast, artifact intensity and distribution of the processed image. Figure 4 and Figure 7It can be seen that the artifact removal method of the present invention can obtain higher quality CT images. The visual effects of the processed images are better than those of the DFR, ResNet and ACDN methods. The image details and contrast are better preserved, and the image texture is more natural without smoothing. By comparing the experimental results images Figure 5 and Figure 8 It can be seen that after removing artifacts, the sparse angle CT image is closer to the conventional CT image, and the difference between the processed result and the conventional CT image is smaller. In the comparison method, after DFR processing, there are still some artifacts remaining, and some high-intensity artifacts are not completely removed; after ResNet network processing, all artifacts are removed, but after processing, there is an over-smoothing phenomenon, and the details of small tissue structures are lost, and the contrast is reduced, which affects the diagnosis; after ACDN method processing, some artifacts are also left, and from the difference map, it can be seen that the edges of some tissue structures are destroyed, and there are certain differences.

[0066] Quantitative evaluation

[0067] For the experimental results, quantitative comparison with reference is also used. For example, Peak Signal to Noise Ratio (PSNR) is the ratio between the maximum power of the signal and the noise power; Structural Similarity Index Measure (SSIM) examines the similarity of structural details between the contrasting images. Frechet perception distance (FID) is used to evaluate the fidelity of the processed CT image. The pre-trained Intelligence v3 model is usually used to calculate the FID score. The experiment will further use three quantitative indicators, PSNR, SSIM and FID, to verify the effectiveness of the method of the present invention. The calculation methods of PSNR and SSIM are as follows:

[0068]

[0069]

[0070] where x T is the image after artifact removal, x r is used for conventional CT images, i,j is the total number of image pixels; H max For x r The maximum value of μ xT and μ xr Represents the CT image x T and x r The average value of the total pixel CT value; σ xT and σxr Represents the CT image x T and x r The standard deviation of the total pixel CT value, σ xTr is the CT image x T With x r The covariance of the constant C1=(0.01×H max ) 2 , C2=(0.03×H max ) 2 Taking the conventional CT image as the reference image, the PSNR, SSIM and FID values ​​of the CT image after artifact removal by the DFR, ResNet, ACDN and DCDN methods can be calculated, and the results are shown in Table 1.

[0071] Table 1

[0072]

[0073] It can be seen from Table 1 that in the sparse angle CT image artifact removal experiments with 120 projection data and 60 projection data, the DCDN method of the present invention can obtain higher PSNR and SSIM values, and lower FID, which also means that after artifact removal, the CT image is closer to the conventional CT image. Fig. 9 and Fig.10 To quantitatively compare the fluctuation of HU values ​​at the same position of CT images after removing artifacts, it can be found that the fluctuations of DFR and ACDN are more dramatic, that is, there are more artifacts, but the fluctuation amplitude of ACDN is lower than that of DFR; the HU line of ResNet is smoother and the intensity fluctuation range is small, that is, there is less noise artifact interference, but the changes in tissue components are too gentle, which also means a decrease in image contrast; the fluctuation of DCDN is weaker than that of DFR and ACDN, but it can also remain sharp at the tissue changes, and the image contrast is higher than ResNet, which also shows that the visual effect of CT images after the artifacts are removed by the method of the present invention is better. After detailed experimental verification, the method of the present invention has shown significant effects, and can accurately eliminate the artifact components in sparse angle CT images, thereby significantly improving the quality of the original image. It is worth mentioning that while achieving artifact elimination, the method retains the tissue anatomical structure information in the image intact, ensuring the accuracy and integrity of the anatomical details. The processed image not only has excellent visual effects, but also has significantly improved contrast, which can provide doctors with clearer and more accurate diagnostic basis. By improving image quality and diagnostic accuracy, the method of the present invention will further promote the improvement of diagnosis and treatment benefits and bring more social value to the medical industry. Therefore, its potential application and promotion prospects are broad.

[0074] The above is a schematic description of the present invention and its implementation mode, which is not restrictive and is only one of the implementation modes of the present invention, and is not actually limited thereto. Therefore, if a person skilled in the art is inspired by it and, without departing from the purpose of the invention, designs a structure and an implementation mode similar to the technical solution without creativity, they shall all fall within the protection scope of the present invention.

Claims

1. A sparse angular CT artifact removal method based on a discriminative convolutional dictionary network, characterized in that: The following steps are involved: Step 1: Obtain matching conventional CT images and sparse angle CT images to construct a training data set; Step 2: Based on the convolutional dictionary sparse representation model, a discriminative convolutional dictionary network is designed for sparse angular CT artifact removal; Step 3: Use the training data set to train the discriminative convolutional dictionary network to obtain the parameters of the entire network model; Step 4: Input the sparse angle CT image to be processed into the trained discriminative convolutional dictionary network to obtain a high-quality CT image with artifacts removed.

2. The method for removing sparse angular CT artifacts based on a discriminative convolutional dictionary network according to claim 1, characterized in that: The basic steps for acquiring data in step 1 are to downsample the projection data of the conventional scan and then reconstruct the sparse angle CT image. The reconstruction method is the same as that of the conventional CT image. The conventional CT image and the sparse angle CT image in the constructed training data set are matched.

3. The method for removing sparse angular CT artifacts based on a discriminative convolutional dictionary network according to claim 1, characterized in that: The discriminative convolutional dictionary network designed in step 2 includes multiple cascade modules, each of which includes an image feature extraction subnetwork, an artifact feature extraction subnetwork and an image prior subnetwork.

4. The low-dose CT image restoration method based on a multi-scale convolutional coding network according to claim 3, characterized in that: The discriminative convolutional dictionary network established in step 2 cascades 20 identical modules in sequence to fully extract image and artifact features.

5. The method for removing sparse angular CT artifacts based on a discriminative convolutional dictionary network according to claim 3, characterized in that: The image feature extraction subnetwork structure in step 2 is: a convolutional layer convolves the input image feature map, then subtracts it from the input image signal, then uses another convolutional layer to convolve and add it to the input image feature map, and finally uses batch normalization and ReLU activation function in sequence to output the updated image feature map.

6. The method for removing sparse angular CT artifacts based on a discriminative convolutional dictionary network according to claim 5, characterized in that: The structure of the artifact feature extraction subnetwork in step 2 is the same as that of the image feature extraction subnetwork. The input is the artifact feature map and the artifact map signal, and the output is the updated artifact feature map.

7. The method for removing sparse angular CT artifacts based on a discriminative convolutional dictionary network according to claim 6, characterized in that: The image prior subnetwork structure in step 2 is as follows: first, an image feature map output by a convolutional layer convolutional image feature extraction subnetwork and an artifact feature map output by a convolutional layer convolutional artifact feature extraction subnetwork are used to obtain updated image signals and artifact map signals, respectively; then, the artifact map signal is subtracted from the sparse angle CT image, and combined with the updated image signal through the Concat layer; finally, the combined information is input into the residual network of three residual blocks and processed to obtain the processed image; the structure of the residual block is a convolutional layer, a batch normalization and ReLU activation function operation, a convolutional layer, a convolutional layer, a batch normalization operation and a jump connection.

8. The method for removing sparse angular CT artifacts based on a discriminative convolutional dictionary network according to claim 1, characterized in that: In step 3, during network training, the image loss function is constructed by using the mean square error loss of image signals and the VGG perceptual loss, the artifact loss function is constructed by using the mean square error loss of artifact image signals, and the network model parameters are iteratively updated through the Adam algorithm.

9. The method for removing sparse angular CT artifacts based on a discriminative convolutional dictionary network according to claim 8, characterized in that: In step 3, the mean square error loss weight in the image loss is 1, the VGG perceptual loss weight is 0.02, and the weights of the image loss and the artifact loss are the same during training; the initial input image signal is a sparse angular CT image, and the initial input artifact map signal, the initial input image feature map, and the initial input artifact feature map are all set to 0.

Citation Information

Cited By

  • Urinary calculus CT image enhancement processing method

    CN120672599A