A low-dose CT reconstruction method combining prior images and convolution sparse networks
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI POLYTECHNIC UNIV
- Filing Date
- 2022-08-23
- Publication Date
- 2026-06-23
Smart Images
Figure CN115984394B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computed tomography imaging technology, and more specifically, relates to a low-dose CT reconstruction method that combines prior images and convolutional sparse networks. Background Technology
[0002] Currently, the main imaging modalities for radiological diagnosis are CT and MRI. Compared to MRI, CT has advantages such as faster imaging speed, higher spatial resolution, and lower cost. However, the radiation risks associated with CT have always been a major concern, as the random and cumulative effects of radiation pose significant risks to patient health. The International Commission on Radiation Protection recommends minimizing radiation dose whenever possible where images are available. Meanwhile, although CT has high spatial resolution, its insufficient low-contrast tissue resolution limits its application in the diagnosis of many diseases, such as stroke penumbra quantification, diagnosis of coronary soft plaques, and early tumor diagnosis. Besides spiral diagnostic CT, which is widely used in hospital radiology departments, the rapid development and application of modern diagnostic and treatment technologies increasingly rely on high-quality cone-beam CT (CBCT) imaging to provide real-time patient anatomical information, such as for radiotherapy of tumors, guidance of interventional procedures, three-dimensional angiography, and plastic surgery. This also presents new demands for real-time low-dose CBCT imaging.
[0003] In image-guided radiation therapy (IGRT), repeated scans increase the cumulative radiation dose to patients. Excessive doses can damage non-target organs, causing serious complications or sequelae. During interventional procedures, medical personnel need to use CBCT imaging to accurately locate lesions in real time. The continuous X-ray exposure during the imaging process can cause significant radiation damage to both patients and physicians. Furthermore, digital subtraction angiography (DSA), a crucial technology for non-invasive diagnosis and interventional surgical navigation of vascular diseases, is increasingly widely used. Reducing dose and improving surgical efficiency have become important competitive indicators for DSA systems. In oral CBCT imaging, the X-ray dose to the patient's brain is significant in order to clearly distinguish between teeth and jawbone. In these CBCT applications, relying solely on active protection technologies is insufficient to achieve more significant dose reduction. Effective low-dose algorithms are needed for the imaging chain to reduce the dose from CT scans.
[0004] However, reducing radiation dose and obtaining clear, high-quality images are always contradictory. For specific scanning conditions and examination sites, in order to obtain better image quality and localize smaller lesions, it is generally necessary to increase the radiation dose to enhance the intensity of tissue attenuation signals. On the other hand, unilaterally reducing the dose received by the patient will reduce the total amount of tissue attenuation information in the projection data, leading to a significant increase in noise artifacts in the reconstructed images. The reduction in data volume at some projection angles or detector channels due to dose reduction will also introduce speckle noise and star-shaped artifacts into the reconstructed images, thus affecting clinicians' detection of lesions. "Low Dose" has been a hot research topic in the field of CT imaging for the past 10 years and is also an important indicator for future CT research.
[0005] A search revealed that Chinese patent application No. 2018105861132 discloses a method and apparatus for low-dose CT imaging based on a convolutional residual network. This method first obtains multiple sets of corresponding raw CT projection data under low and normal doses. Second, it establishes a network based on a convolutional residual network (CNN1) in the projection space. This network takes low-dose CT projection data as input and outputs processed data to reduce noise and artifacts in the low-dose projection CT data and improve the signal-to-noise ratio. Subsequently, it reconstructs the projection data into the image space using a FBP based on a Ramp filter kernel, and performs secondary processing in the image space based on a convolutional residual network (CNN2) to further reduce artifacts and noise in the low-dose data. However, this application does not integrate the neural network with the reconstruction; the network processing is performed in stages, making the process complex. Different stages require repeated adjustments and integration, failing to fully utilize information from different spatial data, resulting in limited improvement in imaging performance. Summary of the Invention
[0006] 1. The problem to be solved
[0007] The purpose of this invention is to address the problems of low image quality, low signal-to-noise ratio, and numerous noise artifacts in existing low-dose CT reconstruction methods. This invention provides a low-dose CT reconstruction method combining prior images and a convolutional sparse network (PRCSN). Based on sparse prior theory, this invention uses a convolutional sparse network to correct for errors in the reconstructed image update. By incorporating valuable feature information from the prior image, the accuracy of the network estimation results can be effectively improved, significantly enhancing image quality and reducing noise artifacts.
[0008] 2. Technical Solution
[0009] To solve the above problems, the technical solution adopted by the present invention is as follows:
[0010] The present invention provides a low-dose CT reconstruction method combining prior images and convolutional sparse networks, comprising the following steps:
[0011] Step 1: Obtain high-quality prior images using discriminative feature representation methods;
[0012] Step 2: Design a convolutional sparse network for feature fusion, which includes three basic operations: feature encoding, feature decoding, and prior feature fusion. Prior feature fusion encodes the prior signal and connects it with the input signal features to expand the feature map channels.
[0013] Step 3: Calculate the error between the reconstructed image and the prior image, and estimate the gradient for iterative reconstruction based on the error and the convolutional sparse network;
[0014] Step 4: Fine-tune the reconstructed image using total variation constraints;
[0015] Step 5: Achieve iterative reconstruction of low-dose CT through module cascading;
[0016] Step 6: Train the convolutional sparse network to obtain the network model parameters;
[0017] Step 7: Use the trained network to reconstruct low-dose CT images.
[0018] Furthermore, in step 1, a discriminative feature representation method is used to obtain a high-quality prior image. The specific process is as follows:
[0019] u pr =DFR(u pl (1)
[0020] Where u pr For the prior image, u pl For images planned for CT scans, DFR(·) represents the discriminative feature representation operation.
[0021] Furthermore, the dictionary atom size in the distinctive feature representation is 8×8×4, with 3000 atoms.
[0022] Furthermore, in step 2, the feature encoding operation sequence is a convolutional layer E1, a ReLU activation function, and a convolutional layer E2, and the features of convolutional layer E1 and convolutional layer E2 are added together; the feature decoding is a convolutional layer D1, a ReLU activation function, and a convolutional layer D2.
[0023] Furthermore, in step 2, the convolutional sparse network for feature fusion is designed as a U-shaped structure. The left side is for input signal feature encoding and prior signal feature encoding, and feature fusion is performed on each layer. Each layer includes two feature encodings and one downsampling operation, and a convolution operation is added to the first layer. The right side is mainly for feature decoding. Each layer includes two feature encodings, four feature decodings and one upsampling operation, and outputs the processed signal. The encoding and decoding parts are combined to merge features.
[0024] Furthermore, the downsampling layer consists of a convolutional layer with a stride of 2, and the upsampling layer is a transposed convolutional layer with a stride of 2, for a total of four sampling layers. Each time the downsampling layer is passed, the number of channels in the convolutional network doubles.
[0025] Furthermore, the gradient reconstruction in step 3 is expressed as:
[0026] s t+1 =M(g) t+1 ,m t+1 (2)
[0027] Where the superscript t represents the iteration number, s t+1 To reconstruct the gradient for iteration t+1, g t+1 =FBP(p-Aut), where A is the projection matrix, u t Let p be the image data reconstructed t times, p be the scanned and preprocessed projection data, and FBP(·) be the filtered back projection operation; m t+1 =u t -u pr , where is the error between the reconstructed image at time t+1 and the prior image; M(·) is the convolutional sparse network.
[0028] Furthermore, in step 5, the modules are cascaded 100 times in the entire iterative reconstruction, that is, the total number of iterations T = 100.
[0029] Furthermore, in step 6, the convolutional sparse network is trained as the entire network after module cascading, where the loss function is Loss M Represented as:
[0030] Loss M =MSE(I RD M c (p,u pr ))+αL PR (I RD M c (p,u pr ))+βL NPS (I RD M c (p,u pr (3)
[0031] In the above formula, I RD The high-quality image reconstructed directly from conventional dose CT projection data using a filtered back-projection algorithm is used as the label; p represents the scanned and preprocessed projection data, and u... pr For the prior image, M c (·) represents the entire network after module cascading, and α and β are the loss weights.
[0032] Furthermore, in step 6, the network model parameters are iteratively updated using a mini-batch stochastic gradient descent algorithm. Training stops when the change in the loss function value before and after the training cycle is within 2%, thus obtaining the network model parameters. In step 7, the trained network is used to reconstruct low-dose CT images. The specific process is as follows: the images of the planned CT scan are... Perform discriminative feature representation to obtain high-quality prior images. The low-dose CT projection data to be reconstructed p ldct and high-quality prior images The image is fed into a cascaded iterative reconstruction network, which ultimately outputs the reconstructed image.
[0033] 3. Beneficial effects
[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0035] (1) The present invention provides a low-dose CT reconstruction method combining prior images and convolutional sparse networks. Based on sparse prior theory, convolutional sparse networks are used to correct the update of reconstruction image errors. By incorporating valuable feature information from prior images, the accuracy of network estimation results can be effectively improved. End-to-end mapping between projection data and image data is realized, and constraint reconstruction based on deep neural networks is truly achieved.
[0036] (2) The low-dose CT reconstruction method of the present invention, which combines prior images and convolutional sparse networks, makes full use of the advantages of prior images of planned scans and convolutional sparse networks. It has the advantages of strong interpretability, high quality of reconstructed images and less noise artifacts. It has great advantages in clinical tumor radiotherapy, which can improve examination efficiency and reduce radiation damage to non-target organs of patients.
[0037] (3) The present invention provides a low-dose CT reconstruction method combining prior images and convolutional sparse networks. Experimental results verify that, under different low-dose CT scan data, the reconstructed images obtained by the present invention have better visual effects and contrast compared with the traditional prior image constrained compress sensing (PICCS) and comprehensive learning enabled adversarial reconstruction (CLEAR) methods. The reconstructed images also achieve satisfactory results in artifact suppression and preservation of anatomical tissue boundaries. Attached Figure Description
[0038] Figure 1 This is a schematic flowchart of a low-dose CT reconstruction method combining prior images and convolutional sparse networks according to the present invention.
[0039] Figure 2 A schematic diagram of the structure of a convolutional sparse network for feature fusion;
[0040] Figure 3 The following are axial reconstruction images of low-dose abdominal projection data in the embodiments (a: reference image; b: prior image; c: FBP reconstruction image; d: PICCS reconstruction image; e: CLEAR reconstruction image; f: PRCSN reconstruction image).
[0041] Figure 4 The following are enlarged views of the axial reconstruction images of the abdominal low-dose projection data in the embodiments (a: reference image; b: prior image; c: FBP reconstruction image; d: PICCS reconstruction image; e: CLEAR reconstruction image; f: PRCSN reconstruction image).
[0042] Figure 5 The noise power spectrum of the axial reconstruction image of the low-dose projection data of the abdomen in the example is shown below (a: noise power spectrum of FBP reconstruction image; b: noise power spectrum of PICCS reconstruction image; c: noise power spectrum of CLEAR reconstruction image; d: noise power spectrum of PRCSN reconstruction image).
[0043] Figure 6 The images shown are axial reconstructions of low-dose lung projection data in the examples (a: reference image; b: prior image; c: FBP reconstruction image; d: PICCS reconstruction image; e: CLEAR reconstruction image; f: PRCSN reconstruction image).
[0044] Figure 7The following are enlarged views of the axial reconstruction images of low-dose lung projection data in the embodiments (a: reference image; b: prior image; c: FBP reconstruction image; d: PICCS reconstruction image; e: CLEAR reconstruction image; f: PRCSN reconstruction image).
[0045] Figure 8 The noise power spectra of the axial reconstruction images of low-dose lung projection data in the examples are shown below (a: noise power spectrum of FBP reconstruction image; b: noise power spectrum of PICCS reconstruction image; c: noise power spectrum of CLEAR reconstruction image; d: noise power spectrum of PRCSN reconstruction image).
[0046] Figure 9 The profile curve of the abdominal image reconstruction in the embodiment;
[0047] Figure 10 The image shown is a profile curve of the lung reconstruction in this embodiment. Detailed Implementation
[0048] The method proposed in this invention is called Prior Image and Convolutional Sparse Network Reconstruction (PRCSN). It uses prior images and a convolutional sparse network to correct errors in the reconstructed image, thereby improving the reconstruction quality of low-dose CT. A feature fusion module is introduced into the convolutional sparse network to incorporate valuable feature information from the prior image into the image to be processed, thus improving the accuracy of the network estimation results. Specifically, firstly, a discriminative feature representation method is used to obtain a high-quality prior image. Secondly, a feature-fusion convolutional sparse network is designed. Next, the error between the reconstructed image and the prior image is calculated, and the gradient is iteratively reconstructed based on the error estimate. Then, the reconstructed image is fine-tuned using total variational constraints. Finally, iterative reconstruction of low-dose CT is achieved through module cascading. This algorithm comprehensively utilizes the advantages of the planned scan's prior image and the convolutional sparse network, offering advantages such as strong interpretability, high-quality reconstructed images, and low noise artifacts. By introducing high-quality prior images, it is expected to improve the reconstructed image quality for certain real-time or multi-channel CT imaging systems.
[0049] The inventors of this application have also been committed to the research of low-dose CT image reconstruction technology and have achieved certain results. For example, application No. 201810706749.6 discloses a low-dose CT image decomposition method based on convolutional neural network. This application uses a neural network to decompose low-dose CT images to remove noise artifacts, which belongs to the post-processing method. The input of this invention is projection data, which, together with a convolutional sparse network, realizes depth iterative reconstruction based on low-dose CT projection.
[0050] The present invention will be further described below with reference to specific embodiments.
[0051] Example
[0052] like Figure 1 As shown in the figure, the specific steps of the low-dose CT reconstruction method combining prior images and convolutional sparse networks in this embodiment are as follows:
[0053] Step 1: Obtain high-quality prior images using discriminative feature representation methods. The specific process can be written as follows:
[0054] u pr =DFR(u pl (1)
[0055] Where u pr For the prior image, u pl For the images to be scanned in the CT scan, DFR(·) is the discriminative feature representation operation. To ensure the quality of the prior image, the dictionary atom size in the discriminative feature representation is 8×8×4, and the number of atoms is 3000.
[0056] Step 2: Design a convolutional sparse network for feature fusion;
[0057] Specifically, the convolutional sparse network for feature fusion includes three basic operations: feature encoding, feature decoding, and prior feature fusion. The feature encoding operation consists of a convolutional layer E1, a ReLU activation function, and a convolutional layer E2, with the features from convolutional layers E1 and E2 added together. Feature decoding consists of a convolutional layer D1, a ReLU activation function, and a convolutional layer D2. The prior feature fusion operation involves encoding the prior signal and concatenating it with the input signal features to expand the feature map channels. The convolutional sparse network for feature fusion has a U-shaped structure. The left side encodes the input signal and prior signal features, and performs feature fusion on each layer. Each layer includes two feature codes and one downsampling operation, with a convolutional operation added to the first layer. The right side mainly performs feature decoding, with each layer including two feature codes, four feature decoders, and one upsampling operation, outputting the processed signal; the encoding and decoding parts merge features. The downsampling layer consists of a convolutional layer with a stride of 2, and the upsampling layer is a transposed convolutional layer with a stride of 2, performing a total of four sampling layers. Each time the convolutional network passes through a downsampling layer, the number of channels doubles, with an initial number of channels of 32.
[0058] Step 3: Calculate the error between the reconstructed image and the prior image, and estimate the iterative reconstruction gradient based on the error and the convolutional sparse network. The iterative reconstruction gradient can be expressed as:
[0059] s t+1 =M(g) t+1 ,mt+1 (2)
[0060] Where the superscript t represents the iteration number, s t+1 To reconstruct the gradient for iteration t+1, g t+1 =FBP(p-Au) t ), where A is the projection matrix, u t Let p be the image data reconstructed t times, p be the scanned and preprocessed projection data, and FBP(·) be the filtered back projection operation; m t+1 =u t -u pr , where is the error between the reconstructed image at time t+1 and the prior image; M(·) is the convolutional sparse network.
[0061] Step 4: Fine-tune the reconstructed image using total variation constraints, which can be specifically expressed as:
[0062]
[0063] Where R(·) represents the total variation constraint treatment, and λ is the regularization parameter.
[0064] Step 5: Implement iterative reconstruction of low-dose CT through module cascading. The modules are cascaded 100 times in the entire iterative reconstruction, which means the total number of iterations T = 100.
[0065] Step 6: Train the convolutional sparse network to obtain the network model parameters;
[0066] The trained convolutional sparse network is the entire network after module cascading. To preserve the detailed texture features of the reconstructed image, its loss function is Loss. M It can be represented as:
[0067] Loss M =MSE(I RD M c (p,u pr ))+αL PR (I RD M c (p,u pr ))+βL NPS (I RD M c (p,u pr (4)
[0068] Where I RD High-quality images reconstructed directly from conventional dose CT projection data using a filtered back-projection algorithm are used as labels. p represents the scanned and preprocessed projection data, and u... pr For the prior image, M c(·) represents the entire network after module cascading, and α and β are the loss weights. Finally, the network model parameters are iteratively updated using the mini-batch stochastic gradient descent algorithm to reduce the loss value; training stops when the change in the loss function value before and after the training cycle is within 2%, thus obtaining the network model parameters.
[0069] Step 7: Use the trained network to reconstruct low-dose CT images.
[0070] The specific process is as follows: Images from the planned CT scan... Perform discriminative feature representation to obtain high-quality prior images. The low-dose CT projection data to be reconstructed p ldct and high-quality prior images The image is fed into a cascaded iterative reconstruction network, which ultimately outputs the reconstructed image.
[0071] Performance Evaluation Criteria
[0072] In this embodiment, high-quality abdominal and lung CT image data under respiratory gating were used for simulation. Poisson noise was added to the projected data to obtain scan data at approximately 0.25 times the conventional dose. In this embodiment, the first phase data was used as the reconstruction object, and the fifth phase data was used as the prior image, followed by relevant reconstruction and analysis. The parameters used in the low-dose scan simulation were: tube voltage of 100 kVp, tube current of 85 mAs, detector size of 736 × 64, and each detector element size of 1.2856 × 1.0947 mm. 2 The distances from the X-ray source to the center of the object and the center of the detector are 60cm and 110cm, respectively. Projection data from 1200 angles are collected under a circular scanning trajectory. In the PRCSN algorithm, the regularization parameters λ, α, and β are 100, 0.01, and 0.001, respectively. The PICCS and CLEAR algorithms will be manually adjusted to their optimal effects according to their references.
[0073] In the accompanying figures, the CT image shows a window width of 400 HU (Housfield Units, HU) and a window level of 50 HU. The reference image is an FBP reconstructed image under conventional dose. By comparing the visual effects of the reconstructed images, it can be seen that the PRCSN reconstruction algorithm of this invention is superior to the PICCS and CLEAR reconstruction algorithms, exhibiting better image quality, especially in magnified local images where details are clearer. Comparing the noise power spectrum of the reconstructed images reveals that the noise intensity of the reconstructed image from this invention is lower. Comparing the profile curves of the selected region of interest shows that the CT values of the reconstructed image from this invention are closer to the reference image, with minimal intensity deviation. Furthermore, the results of the implementation case are quantitatively compared using reference evaluation metrics: Root Mean Square Error (RSME) and Mean Structural Similarity Index (M-SSIM). The calculation methods for RSME and M-SSIM are as follows:
[0074]
[0075]
[0076] Where x T For the last updated image to be reconstructed, x r For high-quality reference images used in simulation, N is the total number of image pixels and S is the number of CT image layers; and Let x represent the i-th layer CT image respectively. T and x r The average CT value of the total pixels; and Let x represent the i-th layer CT image respectively. T and x r The standard deviation of the total pixel CT value, For the CT image of the i-th layer x T With x r The covariance, constants C1 and C2 are default constants.
[0077] Reconstruction Result Evaluation
[0078] By observing and comparing the axial reconstruction results of different layers ( Figure 3 and Figure 6As can be seen, under low-dose scanning conditions, images obtained by the traditional FBP algorithm are almost unusable for direct diagnosis, containing a large amount of high-intensity noise and streak artifacts. Compared to FBP, PICCS reconstructed images show significant improvement, with noise being well suppressed. However, this also destroys the detailed information of the original tissue region, resulting in oversmoothing and blurring of some tissue edges. The CLEAR algorithm combines the advantages of the projection domain and the image domain, achieving better reconstruction results. While suppressing noise and artifacts, it also achieves good contrast and preserves image details to some extent. However, it still has shortcomings, such as the blurring of the liver vein tissue region and artifacts at the lung ribs, as shown in the image. PRCSN reconstructed CT images have less noise and artifacts, while possessing better tissue differentiation capabilities and effectively preserving the edges of anatomical tissue structures. By observing and comparing the magnified local images (…), further improvements can be made. Figure 4 and Figure 7 Compared with other algorithms, the method of this invention can better preserve fine tissue details in the reconstructed CT images, and the visual texture of the images is closer to that of FBP reconstructed images under conventional doses. This is achieved by observing and comparing the noise power spectrum (…). Figure 5 and Figure 8 The method of the present invention reconstructs images with less noise artifacts.
[0079] Using conventional dose FBP reconstructed images as references, the RSME and M-SSIM values of reconstructed images at different data points were calculated. The M-SSIM value was calculated as the average of three adjacent images. RSME and M-SSIM values were used as quantitative indicators to evaluate the reconstructed images and quantitatively analyze their quality. The results are shown in Table 1. Table 1 shows that under low-dose scanning conditions, the abdominal and lung CT images reconstructed using the PRCSN method have lower RSME and higher M-SSIM values. Through comparison... Figure 9 , Figure 10 By selecting the region of interest profile curve, it can be found that the CT value of the reconstructed image of the present invention is closer to that of the reference image, and its intensity deviation is minimal.
[0080] Table 1
[0081]
[0082] Therefore, in summary, the method of the present invention can obtain higher quality low-dose CT reconstruction images and has certain application prospects.
Claims
1. A low-dose CT reconstruction method combining prior images and convolutional sparse networks, characterized in that, Includes the following steps: Step 1: Obtain high-quality prior images using discriminative feature representation methods. The specific process is as follows: （1） in For prior images, For images of the planned CT scan, Operations are performed to represent distinctive features; Step 2: Design a convolutional sparse network for feature fusion, which includes three basic operations: feature encoding, feature decoding, and prior feature fusion. Prior feature fusion encodes the prior signal and connects it with the input signal features to expand the feature map channels. Step 3: Calculate the error between the reconstructed image and the prior image, and estimate the iterative reconstruction gradient based on the error and the convolutional sparse network; the iterative reconstruction gradient is expressed as: （2） superscript t For the number of iterations, for t Reconstruct the gradient in +1 iterations. , A For the projection matrix, for t Secondary reconstructed image data, For the scanned and preprocessed projection data, This is a filtered back projection operation; ,for t +1 error between the reconstructed image and the prior image; For convolutional sparse networks; Step 4: Fine-tune the reconstructed image using total variation constraints; Step 5: Achieve iterative reconstruction of low-dose CT through module cascading; Step 6: Train the convolutional sparse network to obtain the network model parameters; Step 7: Use the trained network to reconstruct low-dose CT images.
2. The low-dose CT reconstruction method combining prior images and convolutional sparse networks according to claim 1, characterized in that: The dictionary in the distinctive feature representation has an atom size of 8×8×4 and a number of 3000 atoms.
3. The low-dose CT reconstruction method combining prior images and convolutional sparse networks according to claim 1, characterized in that: In step 2, the feature encoding operation is performed in the order of one convolutional layer. A ReLU activation function and a convolutional layer Convolutional layer With convolutional layers Add the characteristics together; Feature decoding is performed using a convolutional layer. A ReLU activation function and a convolutional layer .
4. The low-dose CT reconstruction method combining prior images and convolutional sparse networks according to claim 3, characterized in that: In step 2, the convolutional sparse network for feature fusion is designed as a U-shaped structure. The left side is for input signal feature encoding and prior signal feature encoding, and feature fusion is performed on each layer. Each layer includes two feature encodings and one downsampling operation, and a convolution operation is added to the first layer. The right side is mainly for feature decoding. Each layer includes two feature encodings, four feature decodings and one upsampling operation, and outputs the processed signal. The encoding and decoding parts are combined to merge features.
5. The low-dose CT reconstruction method combining prior images and convolutional sparse networks according to claim 4, characterized in that: The downsampling layer consists of a convolutional layer with a stride of 2, and the upsampling layer is a transposed convolutional layer with a stride of 2. A total of four sampling layers are performed. The number of channels in the convolutional network doubles each time a downsampling layer is passed.
6. A low-dose CT reconstruction method based on combined prior images and convolutional sparse networks according to any one of claims 1-5, characterized in that: In step 5, the modules are cascaded 100 times during the entire iterative reconstruction, which is the total number of iterations. T =100.
7. A low-dose CT reconstruction method based on combined prior images and convolutional sparse networks according to any one of claims 1-5, characterized in that: In step 6, the convolutional sparse network is trained as the entire network after module cascading, where the loss function is... Represented as: （3） In the above formula, High-quality images reconstructed directly from conventional dose CT projection data using a filtered back-projection algorithm are used as labels. For the scanned and preprocessed projection data, For prior images, For the entire network after module cascading, and This is the loss weight.
8. The low-dose CT reconstruction method combining prior images and convolutional sparse networks according to claim 7, characterized in that: In step 6, the network model parameters are iteratively updated using the mini-batch stochastic gradient descent algorithm. Training stops when the change in the loss function value before and after the training cycle is within 2%, thus obtaining the network model parameters. In step 7, the trained network is used to reconstruct low-dose CT images. The specific process is as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Perform discriminative feature representation to obtain high-quality prior images. ; low-dose CT projection data to be reconstructed and high-quality prior images The image is fed into a cascaded iterative reconstruction network, which ultimately outputs the reconstructed image. .