A low-dose CT reconstruction method based on residual domain iterative optimization network

Abstract

The present invention discloses a low-dose CT reconstruction method based on a residual domain iterative optimization network, which belongs to the field of computer tomography. The low-dose CT reconstruction method of the present invention first establishes a multi-objective optimization function for low-dose CT reconstruction by adding image domain and projection domain residuals; then, the multi-objective optimization function is decomposed and solved, wherein the directly optimized part adopts iterative solution calculation, and the residual domain update part adopts a convolutional sparse coding network to achieve; finally, the iterative reconstruction of low-dose CT is achieved through module cascade. In addition, in order to improve the initial optimization value, the present invention adopts a convolutional sparse coding network in the image domain to obtain the update of the first reconstructed image. The method of the present invention can improve the reconstruction effect, convergence and generalization ability of the network, reduce noise artifacts in the low-dose reconstructed image, and improve the imaging effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer tomography, and more specifically, relates to a low-dose CT reconstruction method based on a residual domain iterative optimization network. Background Art

[0002] Among the many medical imaging devices, X-ray computed tomography (CT) boasts fast scan times and clear images. It has been widely used in the screening and diagnosis of various diseases and is an indispensable medical device in hospitals at all levels. However, the misuse of radiation during CT examinations and outdated equipment have led to a significant increase in radiation doses. According to research by the U.S. National Council on Radiation Protection and Measurement (NCRP), CT is currently the largest source of radiation burden on patients, with over 50% of medical radiation exposure associated with CT clinical use. Furthermore, due to the cumulative effect of X-ray radiation, the duration of harm increases with the number of examinations, posing a significant safety risk to highly sensitive groups such as children, pregnant women, and the elderly. Therefore, the International Commission on Radiological Protection has recommended that X-ray doses be minimized without compromising diagnostic imaging. my country's National Health Commission also issued the "Diagnostic Reference Levels for X-ray Computed Tomography in Adults" in 2018 to standardize and control radiation doses during CT scans.

[0003] However, reducing or shortening X-ray intensity or duration can lead to a decrease or loss of the signal-to-noise ratio (SNR) of the scanning signal, resulting in degraded reconstructed images. For example, speckle noise and star-shaped artifacts can appear in low-dose images, compromising physicians' diagnoses. Without changing the original hardware, there are three main approaches to improving imaging performance during low-dose scanning: First, from the perspective of CT images: Researchers have designed specialized image restoration and processing algorithms to reduce noise and suppress artifacts, resulting in images with high SNR and high contrast. However, the characterization of noise and artifacts in CT images varies significantly across different scanning devices, scanning modes, and reconstruction methods, and the lack of supplementary raw data limits the effectiveness of these approaches. Second, from the perspective of CT projection data: Noise reduction and restoration processing are performed on the raw data or logarithmically transformed projection data to improve projection data consistency, thereby enhancing reconstruction quality. However, due to the high sensitivity of projection data, undercorrection, overcorrection, and decreased data consistency are common during processing. Third, from the perspective of image reconstruction algorithms: In recent years, a large number of iterative reconstruction algorithms have been proposed and have achieved excellent results, particularly statistical iterative reconstruction algorithms based on prior information constraints. However, these algorithms face major challenges: numerous hyperparameters, making adaptive optimization difficult; high algorithmic complexity, requiring repeated iterative calculations; unstable prior information, and poor generalization capabilities, making iterative reconstruction difficult to fully realize its value in clinical applications. While many challenges remain in "low-dose imaging," these will become important indicators for future CT research and a major direction for the development of X-ray imaging.

[0004] After searching, the application with Chinese patent number 2020113716246 discloses a low-dose cone-beam CT image reconstruction method based on deep learning, including the following steps: transforming the low-dose cone-beam CT raw projection data into multiple projection images; inputting the multiple projection images into a trained projection domain deep convolutional neural network respectively, and using the projection domain deep convolutional neural network to predict the noise distribution in the projection image, and subtract the noise distribution therefrom to output the corresponding high-quality projection image; performing three-dimensional reconstruction on the obtained high-quality projection image to obtain a cone-beam CT image; further including: inputting the cone-beam CT image obtained by three-dimensional reconstruction into a trained image domain deep convolutional neural network, so that the image domain deep convolutional neural network eliminates the noise and artifacts in the cone-beam CT image, and outputs a high-quality cone-beam CT image as the final reconstruction result. Although this method achieves good reconstruction results, it is mainly achieved by combining projection domain processing before the reconstruction stage and image domain processing after the reconstruction stage, focusing on projection domain pre-processing and image domain post-processing.

[0005] For example, Chinese patent application No. 2018107067496 discloses a low-dose CT image decomposition method based on a convolutional neural network, comprising the following steps: Step 1: Reconstruct training images: low-dose CT images and conventional-dose CT images, and subtract the low-dose CT images from the conventional-dose CT images to obtain a noise artifact image; Step 2: Construct a convolutional neural network mapping the low-dose CT images and the noise artifact images Ns; Step 3: Train the constructed convolutional neural network using a certain number of low-dose CT images and the corresponding noise artifact images Ns; Step 4: Process a selected low-dose CT image using the trained convolutional neural network to decompose the anatomical structural components and noise artifact structural components in the selected low-dose CT image. This method primarily operates on images after the reconstruction stage, obtaining high-quality CT images through convolutional neural network decomposition, with a focus on image domain post-processing. Summary of the Invention

[0006] 1. Technical problem to be solved by the invention

[0007] The present invention aims to overcome the above-mentioned shortcomings of existing low-dose CT reconstruction methods and provides a low-dose CT reconstruction method based on a residual domain iterative optimization network. By introducing residuals of image and projection data and performing deep learning and iterative optimization on the residual domain, the present invention improves the network's reconstruction performance, convergence, and generalization capabilities, reduces noise artifacts in low-dose CT reconstructed images, and improves imaging quality.

[0008] 2. Technical solution

[0009] In order to achieve the above object, the technical solution provided by the present invention is:

[0010] A low-dose CT reconstruction method based on a residual domain iterative optimization network of the present invention comprises the following steps:

[0011] Step 1: Establish a multi-objective optimization function for low-dose CT reconstruction based on image and projection data residual domain update, in which a convolutional sparse coding network Ψ(·) is used to accurately estimate and update the image domain residual;

[0012] Step 2: Solve the multi-objective optimization function and implement an iterative optimization network for low-dose CT image reconstruction;

[0013] Step 3: Initialize the first reconstructed CT image to improve the final reconstruction effect;

[0014] Step 4: Train and iterate the network to obtain network model parameters;

[0015] Step 5: Use the trained network to reconstruct low-dose CT images.

[0016] This invention leverages convolutional sparse coding, combining deep neural networks with statistical iterative reconstruction to achieve complementary advantages and improve imaging performance. Furthermore, unlike traditional image priors, this invention incorporates image and projection data residuals. Through deep learning and iterative optimization in the residual domain, it effectively ensures the network's reconstruction quality, convergence, and generalization capabilities.

[0017] Specifically, the multi-objective optimization function for low-dose CT reconstruction based on image and projection data residual domain update established in step 1 is expressed as:

[0018]

[0019] where superscript t is the number of iterations, u is the reconstructed image data, p is the scanned and preprocessed projection data, A is the projection matrix, y is the residual between the projection data p and the reprojection result of the t-th reconstructed image, z is the image domain residual contributed by y, Ψ(·) is the convolutional sparse coding network used to accurately estimate and update the image domain residual z, and μ and ρ are regularization parameters.

[0020] Furthermore, the convolutional sparse coding network Ψ(·) structure in step 1 includes a feature extraction phase, an encoding phase, and a decoding phase. The network structure sequence in the feature extraction phase is: a convolutional layer E and a CReLU activation function. The basic module structure sequence in the encoding phase is: a CReLU activation function, a convolutional layer S. The encoding phase begins and ends with dual convolutional layers G1 and G2. The basic modules are cascaded several times in the middle loop, preferably 10 times, with G1 forming a jump connection with each basic module. The decoding phase consists of a convolutional layer R and a CReLU activation function, which are used to reconstruct the predicted noise signal. Finally, the network output is obtained by adding the initial test signal to the predicted noise signal.

[0021] Furthermore, the solution of the multi-objective optimization function in step 2 is realized by decomposition. Formula (1) is decomposed into:

[0022]

[0023]

[0024]

[0025] Furthermore, the iterative optimization network implemented in step 2 is based on the iterative reconstruction module and consists of three parts, namely the solution calculation of equations (2) to (4). The variables in equations (2) to (4) are updated alternately, and the variables y and u can be directly solved. u t+1 =u t+Ψ(z t+1 ), the variable z involves different imaging geometries and its partial derivative is difficult to calculate, so the present invention adopts the method of approximate identity transformation to solve it, and its solution can be expressed as Where the superscript t is the number of iterations, p is the scanned and preprocessed projection data, A is the projection matrix, Ψ(·) is the convolutional sparse coding network, μ and ρ are regularization parameters, and T is the total number of iterations.

[0026] Furthermore, in step 2, the iterative reconstruction modules in the iterative optimization network are cascaded 100 times, that is, the total number of iterations T=100.

[0027] Furthermore, the initialization calculation of the CT reconstruction image in step 3 is as follows:

[0028] u 1 =Φ(FBP(p)) (5)

[0029] Among them, u 1 is the initialized reconstructed image, p is the scanned and preprocessed projection data, FBP(·) is the filtered back projection reconstruction algorithm, and Φ(·) is the convolutional sparse coding network used to initialize the estimation of the reconstructed image. Its network structure is the same as that of Ψ(·).

[0030] Furthermore, in step 4, the network Φ(·) adopts a hybrid loss function Loss of mean square error loss and perceptual loss Φ After training the network Φ(·), the network Ψ(·) is trained. The loss function Loss in Ψ(·) is Ψ Defined by the residual domain:

[0031] Loss Φ =MSE(I RD ,Φ(I FBP ))+αL PR (I RD ,Φ(I FBP )) (6)

[0032] Loss Ψ =MSE(R RD ,Ψ(R D ))+βL PR (R RD ,Ψ(R D )) (7)

[0033] Among them, I RD The high-quality image reconstructed directly from the conventional dose CT projection data using the filtered back-projection algorithm is used as the label; FBP The low-quality image is reconstructed directly from low-dose CT projection data using the filtered back-projection algorithm. PR(·) is the perceptual loss function, R RD =I RD -Φ(I FBP ), R D =FBP(p-AΦ(I FBP )), FBP(·) is the filtered back projection reconstruction algorithm, p is the low-dose projection data; as training samples, A is the projection matrix, α and β are loss weights.

[0034] Furthermore, in step 4, the network model parameters are iteratively updated by a small-batch stochastic gradient descent algorithm to reduce the loss value; when the loss value change before and after the training cycle is within the specified amplitude range, the training is stopped to obtain the network model parameters; the present invention preferably controls the loss value change before and after the training cycle within a range of 2%.

[0035] Furthermore, in step 5, the trained network is used to reconstruct the low-dose CT image. The specific process is as follows: the low-dose CT projection data to be reconstructed p ldct After reconstruction by FBP algorithm, the image is obtained Then initialize the reconstructed image Next, it is sent to the iterative optimization network for reconstruction, and finally the final reconstructed image is output

[0036] 3. Beneficial effects

[0037] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0038] (1) The low-dose CT reconstruction method based on the residual domain iterative optimization network of the present invention can achieve complementary advantages through the feature acquisition capability of the convolutional sparse coding network and combined with the statistical iterative reconstruction framework without increasing the cost of existing CT hardware, serve low-dose CT reconstruction, improve imaging effects, and is expected to reduce missed diagnosis and misdiagnosis rates in low-dose CT imaging, expand low-dose application scenarios, and increase screening and diagnosis benefits. It has a high application and promotion prospect.

[0039] (2) The present invention provides a low-dose CT reconstruction method based on a residual domain iterative optimization network. By introducing the residuals of image and projection data and performing deep learning and iterative optimization on the image residual domain, the reconstruction effect, convergence and generalization ability of the network can be effectively improved. It also reduces noise artifacts in low-dose reconstructed images, improves tissue contrast, and ultimately reduces additional radiation for users, thereby increasing diagnostic and treatment benefits.

[0040] (3) The present invention provides a low-dose CT reconstruction method based on a residual domain iterative optimization network. The algorithm used has the advantages of strong interpretability, high contrast of reconstructed images, few artifacts, low noise, and strong stability. When applied to diagnosis and treatment, it can improve inspection efficiency and reduce additional radiation damage to patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the process of the low-dose CT reconstruction method based on the residual domain iterative optimization network of the present invention;

[0042] Figure 2 Axial reconstructions of low-dose projection data of the abdomen in the embodiment (a: reference image; b: FBP reconstruction image; c: IR-TV reconstruction image; d: FISTA-Net reconstruction image; e: IR-DRL reconstruction image);

[0043] Figure 3 Sagittal reconstruction images of low-dose projection data of the abdomen in the embodiment (a: reference image; b: FBP reconstruction image; c: IR-TV reconstruction image; d: FISTA-Net reconstruction image; e: IR-DRL reconstruction image);

[0044] Figure 4 Coronal reconstruction images of low-dose projection data of the abdomen in the embodiment (a: reference image; b: FBP reconstruction image; c: IR-TV reconstruction image; d: FISTA-Net reconstruction image; e: IR-DRL reconstruction image);

[0045] Figure 5 Axial reconstructions of low-dose lung projection data in the embodiment (a: reference image; b: FBP reconstruction image; c: IR-TV reconstruction image; d: FISTA-Net reconstruction image; e: IR-DRL reconstruction image);

[0046] Figure 6 Sagittal reconstruction images of low-dose lung projection data in the embodiment (a: reference image; b: FBP reconstruction image; c: IR-TV reconstruction image; d: FISTA-Net reconstruction image; e: IR-DRL reconstruction image);

[0047] Figure 7 Coronal reconstruction images of low-dose lung projection data in the embodiment (a: reference image; b: FBP reconstruction image; c: IR-TV reconstruction image; d: FISTA-Net reconstruction image; e: IR-DRL reconstruction image);

[0048] Figure 8 Profile curves of different reconstruction images of the abdomen in the embodiment;

[0049] Figure 9] are Profile curves of different lung reconstruction images in the embodiment. DETAILED DESCRIPTION

[0050] The present invention discloses a low-dose CT reconstruction method based on a residual domain iterative optimization network, also known as deep residual iterative optimization reconstruction (Iterative Optimization Reconstruction-based Deep Residual-domain Learning, abbreviated as IR-DRL). It starts from the perspective of convolutional sparse coding, combines deep neural networks and statistical iterative reconstruction ideas, thereby achieving complementary advantages and improving imaging effects. At the same time, by learning the residuals of image and projection data, deep learning and iterative optimization in the residual domain are used to promote the reconstruction quality, convergence and generalization ability of the network. Specifically, first, a multi-objective optimization function for image reconstruction is established by adding image domain and projection domain residuals. Then, the multi-objective optimization function is decomposed and solved, wherein the direct optimization calculation part adopts iterative solution calculation, and the update part of the residual domain will be implemented by a convolutional sparse coding network. Finally, iterative reconstruction is implemented in a cascade form. In addition, in order to improve the initial optimization position, a convolutional sparse coding network in the image domain is used to implement the update of the first reconstructed image. That is, the present invention focuses on the reconstruction stage, integrates the neural network into the reconstruction stage, and processes the residual data in the reconstruction process, thereby improving the reconstruction quality.

[0051] Experimental results show that, compared with the traditional total variation constrained reconstruction method (Iterative optimization Reconstruction-based Total Variation, abbreviated as IR-TV) and the fast iterative shrinkage threshold learning network reconstruction method (Learning a Fast Iterative Shrinkage Thresholding Network, abbreviated as FISTA-Net), the method of the present invention (IR-DRL) can effectively suppress the noise artifacts and blurred details caused by ray attenuation in the reconstructed images under different CT scan data, and the reconstructed images have better visual effects and contrast.

[0052] The present invention will be further described below with reference to specific embodiments.

[0053] Example

[0054] Combine Figure 1 In this embodiment, a low-dose CT reconstruction method based on a residual domain iterative optimization network is described, and the specific steps are as follows:

[0055] Step 1: Establish a multi-objective optimization function for low-dose CT reconstruction based on image and projection data residual domain update. The objective optimization function can be expressed as:

[0056]

[0057] Wherein, the superscript t is the number of iterations, u is the reconstructed image data, p is the scanned and preprocessed projection data, A is the projection matrix, y is the residual between the projection data p and the reprojection result of the t-th reconstructed image, z is the image domain residual contributed by y, Ψ(·) is the convolutional sparse coding network used to accurately estimate and update the image domain residual z, and μ and ρ are regularization parameters, which will be selected empirically based on the specific data.

[0058] The convolutional sparse coding network Ψ(·) is used to accurately estimate and update the image residual. Its network structure includes a feature extraction stage, an encoding stage, and a decoding stage. The network structure sequence in the feature extraction stage is: a convolutional layer E and a CReLU activation function. The basic module structure sequence in the encoding stage is: a CReLU activation function and a convolutional layer S. The encoding stage uses dual convolutional layers G1 and G2 at the beginning and end. The basic modules are cascaded multiple times in the intermediate loop (specifically 10 times in this embodiment), and G1 forms a skip connection with each basic module. The decoding stage consists of a convolutional layer R and a CReLU activation function to reconstruct the predicted noise signal. Finally, the network output is obtained by adding the initial test signal to the predicted noise signal. In this embodiment, the CReLU activation function can be expressed as: CReLU(x,θ)=ReLU(x-θ)-ReLU(-x-θ), where x and θ are the input data and ReLU(·) is the ReLU activation function.

[0059] This embodiment uses a convolutional sparse coding network to accurately estimate and update image residuals and optimizes the network structure, effectively improving low-dose CT imaging. By learning image and projection data residuals, deep learning and iterative optimization in the residual domain can improve the network's reconstruction quality, convergence, and generalization capabilities. Specifically, in this embodiment, the convolution kernel size is 3×3, the number of convolution kernel channels is set to 32 during the feature extraction phase, and 64 for all other phases.

[0060] Step 2: Decompose and solve the multi-objective optimization function and implement an iterative optimization network for low-dose CT image reconstruction;

[0061] Specifically, after obtaining the reconstruction multi-objective optimization function, it is decomposed into the following three formulas:

[0062]

[0063]

[0064]

[0065] The iterative reconstruction module in the iterative optimization network consists of three parts, namely the solution calculation of formula (2) to formula (4). The variables in formula (2) to formula (4) are updated alternately, and the variables y and u can be directly solved. u t+1 =u t +Ψ(z t+1 ), the variable z involves different imaging geometries and it is difficult to calculate its partial derivative, so the approximate identity transformation is used to solve it, and the solution can be expressed as Where the superscript t represents the number of iterations, p represents the scanned and preprocessed projection data, A represents the projection matrix, Ψ(·) represents the convolutional sparse coding network, and μ and ρ represent regularization parameters, which are selected empirically based on the specific data. Finally, an iterative optimization network is established by cascading iterative reconstruction modules multiple times to achieve iterative reconstruction of the CT projection data. In this embodiment, the total number of iterations, T, is 100.

[0066] Step 3: Initialize the first reconstructed CT image to improve the final reconstruction effect. The initialization calculation of the first reconstructed CT image is as follows:

[0067] u 1 =Φ(FBP(p)) (5)

[0068] Among them, u 1 is the initialized reconstructed image, p is the scanned and preprocessed projection data, FBP(·) is the filtered back projection (FBP) analytical reconstruction algorithm. In this embodiment, a ramp filter function is used. Φ(·) is a convolutional sparse coding network, and the network structure of Φ(·) is the same as that of Ψ(·).

[0069] Step 4: Train and iterate to optimize the network and obtain network model parameters

[0070] The network Φ(·) adopts a hybrid loss function of mean square error loss and perceptual loss Loss Φ After training the network Φ(·), the network Ψ(·) is trained. The loss function Loss in Ψ(·) is Ψ Defined by the residual domain:

[0071] Loss Φ =MSE(I RD ,Φ(I FBP ))+αL PR (I RD ,Φ(I FBP )) (6)

[0072] LossΨ =MSE(R RD ,Ψ(R D ))+βL PR (R RD ,Ψ(R D )) (7)

[0073] Among them, I RD The high-quality image reconstructed directly from the conventional dose CT projection data using the filtered back-projection algorithm is used as the label. FBP The low-quality image is reconstructed directly from low-dose CT projection data using the filtered back-projection algorithm. PR (·) is the perceptual loss function, R RD =I RD -Φ(I FBP ), R D =FBP(p-AΦ(I FBP ), FBP(·) is the filtered back projection algorithm, p is the low-dose projection data. As training samples, A is the projection matrix, and α and β are loss weights.

[0074] Finally, the network model parameters are iteratively updated through the mini-batch stochastic gradient descent algorithm (the algorithm parameters use the default values) to reduce the loss value; when the loss value before and after the training cycle changes within 2%, the training is stopped and the network model parameters are obtained.

[0075] Step 5: Use the trained network to reconstruct low-dose CT images. The specific process is as follows: the low-dose CT projection data to be reconstructed p ldct After reconstruction by FBP algorithm, the image is obtained Then the initial reconstructed image Next, it is sent to the iterative optimization network for reconstruction, and finally the final reconstructed image is output

[0076] Effectiveness Evaluation Criteria

[0077] In this example, high-quality abdominal and lung CT image data were used for simulation. Poisson noise was added to the projection data to obtain scan data at approximately 0.25 times the conventional dose, which was then reconstructed and compared. The scanning parameters used in the computer simulation were: a detector size of 960×400, with each unit measuring 0.78mm×0.78mm in physical size. During scanning, the distances from the X-ray source to the object center and the detector center were 50cm and 100cm, respectively. Projection data was collected from 720 angles along a circular scanning trajectory. For IR-DRL reconstruction, the regularization parameters μ, ρ, α, and β for the abdominal projection data were 0.01, 0.1, 0.001, and 0.001, respectively, and for the lung data, 0.015, 0.1, 0.002, and 0.001, respectively. The IR-TV and FISTA-Net reconstruction algorithms were manually tuned to optimal results according to their references.

[0078] In the accompanying figure, the CT image display window is [-250 150] HU (Housfield Units, HU), and the reference image is an FBP reconstructed image at a conventional dose. The reconstruction results are compared through image visual effects. As can be seen from the figure, the IR-DRL reconstruction algorithm of the embodiment is superior to the IR-TV and FISTA-Net reconstruction algorithms, and the image quality is better. By comparing the profile curves of the selected region of interest, the reconstruction results of different algorithms can be carefully observed. Compared with the reference image, the deviation of the CT value intensity of the IR-DRL reconstructed image obtained by this embodiment is the smallest (as shown in Figure 2). Figure 8 and Figure 9 White line segments mark the area).

[0079] Furthermore, the results of this embodiment are quantitatively compared using reference evaluation indicators, Peak Signal to Noise Ratio (PSNR), Mean Structural Similarity Index Measure (M-SSIM), and Contrast-to-Noise Ratio (CNR). The calculation methods for PSNR, M-SSIM, and CNR are as follows:

[0080]

[0081]

[0082]

[0083] Among them, x T is the last updated reconstructed image, x ris a high-quality reference image used for simulation, N is the total number of image pixels, S is the number of CT image layers; H max is x r The maximum value of and Represents the i-th layer CT image x T and x r The average value of the total pixel CT value; and Represents the i-th layer CT image x T and x r The standard deviation of the total pixel CT value, is the i-th layer CT image x T with x r The covariance of the constant C1=(0.01×H max ) 2 , C2=(0.03×H max ) 2 . and Represents the region of interest x of the i-th layer CT image ROI and background area x BG The average CT value, and Represents the region of interest x of the i-th layer CT image ROI and background area x BG Standard deviation of CT values.

[0084] Evaluation of reconstruction results

[0085] By comparing the axial reconstruction results (such as Figure 2 and Figure 5 As shown in the figure), it was found that the low-dose FBP reconstructed image was seriously disturbed by noise and stripe artifacts, and the contrast between different tissues was low. The IR-TV reconstructed image contained less noise, but was blurred at the edges of the liver and stomach tissues. In the FISTA-Net reconstructed CT image, the liver vein tissue area was somewhat blurred. In the IR-DRL reconstructed CT image, there was less noise and artifacts, and at the same time, it had better tissue differentiation ability, and could well maintain the edges of the anatomical tissue structure. The vascular and cyst areas were all well displayed. By observing and comparing the sagittal and coronal images (as shown in the figure), the IR-DRL reconstructed image had less noise and artifacts, and had better tissue differentiation ability. It could well maintain the edges of the anatomical tissue structure, and the vascular and cyst areas were all well displayed. Figure 3 、 Figure 4 、 Figure 6 and Figure 7 As shown, it can be seen that under low-dose scanning conditions, compared with other algorithms, the CT image reconstructed by this embodiment can better preserve fine tissue details, and the image visual texture is closer to the FBP reconstructed image under conventional doses.

[0086] The conventional dose FBP reconstructed image is used as the reference image to calculate the PSNR, M-SSIM and CNR values ​​of the reconstructed images with different data. The M-SSIM and CNR values ​​are calculated as the average of three adjacent images. Figure 2 (a) and Figure 5 The black frame area in the image (a) is the selected region of interest, and the white frame area is the selected background area. PSNR, M-SSIM and CNR values ​​are used as quantitative indicators to evaluate the reconstructed image to quantitatively analyze the quality of the reconstructed image. The results are shown in Table 1 below. As can be seen from Table 1, in the reconstruction of low-dose abdominal and lung CT images, the quantitative indicators of the FBP reconstructed image are the worst, and the reconstruction results of IR-TV and FISTA-Net have improved to a certain extent. The IR-DRL method of this embodiment can obtain higher M-SSIM and PSNR values. Figure 8 、 Figure 9 It can be seen that in the selected pixels ( Figure 8 and Figure 9 The white line segments mark the area in the middle image, and Ref is the reference image. In the IR-DRL reconstructed image, the pixel value jumps at the tissue boundary are more obvious, the tissue boundary is sharper, and the curve trend is closer to the reference image.

[0087] Table 1 Evaluation results of reconstructed images

[0088]

[0089]

[0090] It can be seen from the above experiments that the method of the present invention can obtain higher quality low-dose CT reconstructed images, and has certain application prospects.

Claims

1. A low-dose CT reconstruction method based on residual domain iterative optimization network, characterized in that: The following steps are involved: Step 1: Establish a low-dose CT reconstruction multi-objective optimization function based on image and projection data residual domain update. The established low-dose CT reconstruction multi-objective optimization function is expressed as: Where: superscript t is the number of iterations, u is the reconstructed image data, p is the scanned and preprocessed projection data, A is the projection matrix, y is the residual between the projection data p and the reprojection result of the t-th reconstructed image, z is the image domain residual contributed by y, Ψ(·) is the convolutional sparse coding network used to accurately estimate and update the image domain residual z, μ and ρ are regularization parameters; Among them, the convolutional sparse coding network Ψ(·) is used to accurately estimate and update the image domain residual; Step 2: solving the multi-objective optimization function and implementing an iterative optimization network for low-dose CT image reconstruction; Step 3: Initialize the first reconstructed CT image; Step 4: Train and iterate the network to obtain network model parameters; Step 5: Use the trained network to reconstruct low-dose CT images.

2. The low-dose CT reconstruction method based on residual domain iterative optimization network according to claim 1, characterized in that: The structure of the convolutional sparse coding network Ψ(·) includes a feature extraction stage, an encoding stage, and a decoding stage. The network structure sequence in the feature extraction stage is: a convolutional layer E and a CReLU activation function; the structure sequence of the basic modules in the encoding stage is: a CReLU activation function and a convolutional layer S; the encoding stage uses dual convolutional layers G1 and G2 at the beginning and end, and the basic modules are cascaded several times in the middle cycle, and G1 forms a jump connection with each basic module; the decoding stage consists of a convolutional layer R and a CReLU activation function, which are used to reconstruct the predicted noise signal; finally, the network output is obtained by adding the initial test signal to the predicted noise signal.

3. The low-dose CT reconstruction method based on residual domain iterative optimization network according to claim 1, characterized in that: In step 2, the multi-objective optimization function is solved by decomposition, which is expressed as follows:

4. The low-dose CT reconstruction method based on residual domain iterative optimization network according to claim 3, characterized in that: In step 2, the iterative optimization network uses the iterative reconstruction module as the basic unit, which includes three parts: the solution calculation of equations (2) to (4). The variables in equations (2) to (4) are updated alternately. The solution results of variables y, u, and z are as follows: u t+1 =u t +Ψ(z t+1 ) In the above formula, the superscript t is the number of iterations, p is the scanned and preprocessed projection data, A is the projection matrix, Ψ(·) is the convolutional sparse coding network, μ and ρ are regularization parameters, and T is the total number of iterations.

5. The low-dose CT reconstruction method based on residual domain iterative optimization network according to claim 4, characterized in that: In the iterative optimization network of step 2, the iterative reconstruction modules are cascaded 100 times, that is, the total number of iterations T = 100.

6. A low-dose CT reconstruction method based on a residual domain iterative optimization network according to any one of claims 1 to 5, characterized in that: The calculation of the initialization of the first reconstructed CT image in step 3 is as follows: you 1 =Φ(FBP(p)) (5) Among them, u 1 is the initialized reconstructed image, p is the scanned and preprocessed projection data, FBP(·) is the filtered back projection analytical reconstruction algorithm, Φ(·) is the convolutional sparse coding network used to initialize the estimation of the reconstructed image, and its network structure is the same as Ψ(·).

7. The low-dose CT reconstruction method based on residual domain iterative optimization network according to claim 6, characterized in that: In step 4, the network Φ(·) adopts a hybrid loss function Loss of mean square error loss and perceptual loss Φ After training the network Φ(·), the network Ψ(·) is trained. The loss function Loss in Ψ(·) is Φ Defined by the residual domain: Loss Φ =MSE(I RD ,Φ(I FBP ))+αL PR (I RD ,Φ(I FBP )) (6) Loss Ψ =MSE(R RD ,Ψ(R D ))+βL PR (R RD ,Ψ(R D )) (7) Loss Φ Among them, I RD The high-quality image reconstructed directly from the conventional dose CT projection data using the filtered back-projection algorithm is used as the label; FBP The low-quality image is reconstructed directly from low-dose CT projection data using the filtered back-projection algorithm. PR (·) is the perceptual loss function, R RD =I RD -Φ(I FBP ), R D =FBP(p-AΦ(I FBP )), FBP(·) is the filtered back projection reconstruction algorithm, p is the low-dose projection data; as training samples, A is the projection matrix, α and β are loss weights.

8. The low-dose CT reconstruction method based on residual domain iterative optimization network according to claim 7, characterized in that: In step 4, the network model parameters are iteratively updated through the small batch stochastic gradient descent algorithm. When the loss value before and after the training cycle changes within the specified amplitude range, the training is stopped to obtain the network model parameters.

9. A low-dose CT reconstruction method based on residual domain iterative optimization network according to any one of claims 1 to 5, characterized in that: In step 5, the trained network is used to reconstruct low-dose CT images. The specific process is as follows: the low-dose CT projection data to be reconstructed p ldct After reconstruction by FBP algorithm, the image is obtained Then initialize the reconstructed image Next, it is sent to the iterative optimization network for reconstruction, and finally the final reconstructed image is output

Images