Limited angle ct reconstruction method, system and device based on wavelet multi-channel scoring

Abstract

The application discloses a limited-angle CT reconstruction method and system based on wavelet multi-channel scoring, and a device, and the method comprises the following steps: acquiring a limited-angle CT noisy image; performing step-by-step denoising on the limited-angle CT noisy image according to a preset time length, and outputting the input data to the next denoising process after performing deep sampling processing on the input data in the image domain and the wavelet domain in each denoising process, and outputting the output data of the last denoising process as a CT reconstruction image. The application analyzes the distribution characteristics of the limited-angle artifacts in different frequency components by introducing wavelet transformation in the image reconstruction process, which is beneficial to separating and removing the artifacts from the image data, ensures the consistency and effectiveness of the details and structures, and improves the quality of the image reconstruction.

Description

Technical Field

[0001] This invention relates to the field of image reconstruction technology, specifically to a finite-angle CT reconstruction method, system, and device based on wavelet multi-channel scoring. Background Technology

[0002] The reconstruction process in computed tomography (CT) aims to improve image quality and reduce radiation exposure. In traditional CT imaging tasks, limited-angle CT has attracted considerable attention due to its unique characteristics of rapid data acquisition and reduced radiation dose. Despite these advantages, the reconstruction process faces significant challenges due to the incompleteness of data acquisition.

[0003] Currently, some scholars have proposed applying the traditional SGM model to finite-angle CT reconstruction tasks. However, finite-angle CT images usually exhibit complex artifacts with directionality and non-uniformity. This is mainly due to the influence of some scanning angles during the scanning acquisition process of finite-angle CT images. The distribution of these complex artifacts hinders the accurate estimation of the probability density function by the SGM model, thereby reducing its reconstruction performance. Summary of the Invention

[0004] This invention provides a limited-angle CT reconstruction method, system, and device based on wavelet multi-channel scoring to solve one or more technical problems existing in the prior art, and at least provide a beneficial option or create conditions.

[0005] Firstly, a finite-angle CT reconstruction method based on wavelet multi-channel scoring is provided, including:

[0006] Acquire finite-angle CT images with added noise;

[0007] The denoised CT image with limited angle is gradually denoised according to a preset time length. In each denoising process, the input data is processed by depth sampling in the image domain and wavelet domain and then output to the next denoising process. The output data of the last denoising process is then used as the CT reconstructed image.

[0008] Furthermore, the step of performing depth sampling processing on the input data in both the image domain and wavelet domain in each denoising process and then outputting it to the next denoising process includes:

[0009] The input data is optimized, and then the optimized input data is sampled once using an image domain-based sampling generator to obtain the first sampled data.

[0010] The first sampled data is optimized, and then wavelet transform is performed on the optimized first sampled data to obtain the first input data group;

[0011] The first input data group is sampled once using a wavelet domain-based sampling generator to obtain the corresponding second sampled data group;

[0012] Perform an inverse wavelet transform on the second sampled data group to obtain the final output data.

[0013] Furthermore, the optimization of the input data includes:

[0014] Obtain the first optimization function, then input the input data into the first optimization function and use the conjugate gradient method to solve for the optimized input data.

[0015] Furthermore, the step of using an image domain-based sampling generator to perform a single sampling of the optimized input data to obtain the first sampled data includes:

[0016] The image domain-based sampling generator includes a first predictor and a first detector;

[0017] The first predictor is used to perform a single sampling of the optimized input data to obtain the first initial sampling data.

[0018] The first initial sampling data is corrected using the first detector, and then data consistency constraints are applied to the corrected first initial sampling data to obtain the first sampling data.

[0019] Furthermore, the optimization of the first sampled data includes:

[0020] By introducing a coupling factor to separate and transform the first optimization function, a second optimization function is obtained;

[0021] After inputting the first sampled data into the second optimization function, the conjugate gradient method is used to solve the problem and obtain the optimized first sampled data.

[0022] Furthermore, the step of using a wavelet-domain-based sampling generator to perform a single sampling of the first input data group to obtain the corresponding second sampled data group includes:

[0023] The wavelet domain-based sampling generator includes a second predictor and a second detector;

[0024] For each first input data in the first input data group, the second predictor is used to perform a single sampling of the first input data to obtain second initial sample data.

[0025] The second detector is used to correct the second initial sampling data, and then data consistency constraints are applied to the corrected second initial sampling data to obtain the corresponding second sampling data.

[0026] Furthermore, the image domain-based sampling generator is obtained in the following way:

[0027] Acquire finite-angle CT training images and a first diffusion model constructed based on stochastic differential equations, wherein the first diffusion model includes a first forward process and a first backward process;

[0028] During the first forward process, noise is gradually injected into the finite-angle CT training image to obtain the first noisy image;

[0029] In the first reverse process, a first scoring network is trained to progressively denoise the first noisy image. The trained first scoring network is an image domain-based sampling generator.

[0030] Furthermore, the wavelet-domain-based sampling generator is obtained in the following way:

[0031] The finite-angle CT training image and a second diffusion model constructed based on stochastic differential equations are obtained. The second diffusion model includes a second forward process and a second backward process.

[0032] Perform wavelet transform on the finite-angle CT training images to obtain a finite-angle CT training sub-image group;

[0033] During the second forward process, noise is gradually injected into the finite angle CT training sub-image group to obtain the corresponding second noisy image group;

[0034] In the second reverse process, a second scoring network is trained to progressively denoise the second noisy image group. The trained second scoring network is a wavelet domain-based sampling generator.

[0035] Secondly, a finite-angle CT reconstruction system based on wavelet multi-channel scoring is provided, including:

[0036] The first module is used to acquire finite-angle CT images with added noise.

[0037] The second module is used to perform stepwise denoising on the finite-angle CT noisy image according to a preset time length. In each denoising process, the input data is processed by depth sampling in the image domain and wavelet domain and then output to the next denoising process. The output data of the last denoising process is then used as the CT reconstructed image.

[0038] Thirdly, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the finite-angle CT reconstruction method based on wavelet multi-channel scoring as described in the first aspect.

[0039] The present invention has at least the following beneficial effects: by performing depth sampling processing of the input data in the image domain and wavelet domain in each denoising process included in the image reconstruction task, and introducing wavelet transform to analyze the distribution characteristics of finite angle artifacts in different frequency components, it is beneficial to separate and remove artifacts from the output data of each denoising process, ensure the consistency and effectiveness of detailed structures, and improve the quality of image reconstruction. Attached Figure Description

[0040] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0041] Figure 1 This is a flowchart illustrating a finite-angle CT reconstruction method based on wavelet multi-channel scoring in an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram illustrating the implementation process of stepwise denoising of a finite-angle CT noisy image in an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram illustrating the principle of the first diffusion model based on the image domain in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram illustrating the principle of the second diffusion model based on the wavelet domain in an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram comparing the implementation effects of the WIMD model with other traditional CT reconstruction models in a numerical simulation experiment for 90℃T reconstruction in an embodiment of the present invention.

[0046] Figure 6 This is a schematic diagram comparing the implementation effects of the WIMD model with other traditional CT reconstruction models in an experiment based on sheep lung data, according to an embodiment of the present invention.

[0047] Figure 7 This is a schematic diagram comparing the implementation effects of the WIMD model with other traditional CT reconstruction models in an experiment based on cardiac clinical data, according to an embodiment of the present invention.

[0048] Figure 8 This is a schematic diagram comparing the implementation effects between a WIMD model containing dual-domain sampling and a CT reconstruction model containing single-domain sampling in an ablation experiment according to an embodiment of the present invention.

[0049] Figure 9This is a schematic diagram comparing the PSNR convergence between a WIMD model containing dual-domain sampling and a CT reconstruction model containing single-domain sampling in an ablation experiment according to an embodiment of the present invention.

[0050] Figure 10 This is a schematic diagram of the composition of a finite-angle CT reconstruction system based on wavelet multi-channel scoring in an embodiment of the present invention;

[0051] Figure 11 This is a schematic diagram of the hardware structure of the computer device in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] It should be noted that although functional modules are divided in the system diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first," "second," "third," "fourth," etc., used in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed and are inherent to these processes, methods, products, or apparatuses.

[0054] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a finite-angle CT reconstruction method based on wavelet multi-channel scoring provided by an embodiment of the present invention. The method includes the following:

[0055] Step S110: Obtain a finite-angle CT image with added noise;

[0056] Step S120: Denoise the finite-angle CT image step by step according to a preset time length. In each denoising process, the input data is processed by depth sampling in the image domain and wavelet domain and then output to the next denoising process. The output data of the last denoising process is then used as the CT reconstructed image.

[0057] In this embodiment of the invention, the specific implementation process of step S120 includes the following:

[0058] Step S121: Obtain the preset time length T, which represents the number of denoising processes required to denoise the finite angle CT noisy image;

[0059] Step S122: Obtain the data input to the kth denoising process and record it as the kth input data;

[0060] Step S123: Optimize the kth input data to obtain the optimized kth input data;

[0061] Step S124: Perform a single sampling operation on the optimized k-th input data using an image domain-based sampling generator to obtain the k-th first sampled data;

[0062] Step S125: Optimize the kth first sampled data to obtain the optimized kth first sampled data;

[0063] Step S126: Perform wavelet transform on the optimized kth first sampled data to obtain the kth first input data group;

[0064] Step S127: Perform a single sampling operation on the kth first input data group using a wavelet domain-based sampling generator to obtain the corresponding kth second sample data group;

[0065] Step S128: Perform inverse wavelet transform on the kth second sampled data group to obtain the final kth output data;

[0066] Step S129: Determine whether k+1 is less than or equal to T; if it is less than or equal to T, assign k+1 to k and return to execute step S122 above; if it is greater than T, use the kth output data as the final CT (Computed Tomography) reconstructed image output.

[0067] It should be noted that step S122 above is executed starting from k=1, and the data input to the first denoising process is the finite-angle CT noisy image; the corresponding implementation effect can be seen in [reference needed]. Figure 2 As shown, where x T The finite-angle CT image with added noise is given, and x0 is the CT reconstructed image. Figure 2 The implementation effect of a single denoising process is only given in detail.

[0068] More specifically, the implementation process of step S123 above includes, but is not limited to, the following:

[0069] Step S123.1: Obtain the first optimization function, whose corresponding mathematical expression is as follows:

[0070]

[0071] Step S123.2: After inputting the kth input data into the first optimization function, solve it using the existing conjugate gradient method (CG) to obtain the optimized kth input data;

[0072] In the formula, x t-0.5 For the optimized input data, A is the system matrix, where each row contains the coefficients corresponding to a single ray, x t Here, y represents prior information, and η represents the input data. t R1(·) is a balancing coefficient used to finely adjust data consistency, and R1(·) is a regularization term from the image domain.

[0073] As an extension, the origin of the first optimization function is explained as follows:

[0074] First, the mathematical expression for the diffusion model based on the image domain and wavelet domain is proposed as follows:

[0075]

[0076] Removing the constraints from equation (2) above, we obtain:

[0077]

[0078] Equation (3) above clearly emphasizes the intrinsic relationship between data consistency and diffusion model in the image domain and wavelet domain. With the support of data consistency, when solving equation (3) above, it can be decomposed into optimization processes in two different domains, as follows:

[0079]

[0080]

[0081] In the formula, λ t To achieve a balance between solving in the image domain and solving in the wavelet domain, a regularization factor is used, where R2(·) is the regularization term from the wavelet domain, W represents the wavelet transform operation, κ is a pre-given positive constant, and x... t-1 This is for output data.

[0082] First of all, please refer to Figure 3 As shown, the method for obtaining the image domain-based sampling generator mentioned in step S124 above specifically includes the following:

[0083] Step A1: Obtain finite-angle CT training images;

[0084] Step A2: Obtain the first diffusion model based on the image domain. The first diffusion model is constructed using stochastic differential equations, specifically including a first forward process and a first backward process.

[0085] Step A3: Noise is progressively injected into the finite-angle CT training image through the first forward process to obtain the first noisy image. The mathematical expression used in this process is:

[0086] dx=f(x,t)dt+g(t)du; (5)

[0087] Since this invention exemplarily chooses to employ a typical variance-exploding-stochastic differential equation (VE-SDE), that is, limiting f = 0 and... Formula (5) above can be transformed into:

[0088]

[0089] Step A4: In the first reverse process, a first scoring network is trained to progressively denoise the first noisy image. The trained first scoring network is an image domain-based sampling generator. The first scoring network is controlled by parameter θ1, and its first objective function during training is:

[0090]

[0091] In the formula, x represents the input image data, t represents time, f(x,t) is the vector value function, also known as the drift coefficient with respect to the input image data x, g(t) is the scaling function, also known as the diffusion coefficient with respect to the input image data x, and u represents the Wiener process, also known as Brownian motion, which is the integral of Gaussian white noise. The above formula (5) can be understood as follows: after a small time step dt, a Gaussian white noise with mean f(x,t) and variance g(t)du can be added to the input image data x, and σ(t) is a function of noise that monotonically increases with time. Let λ(t) be the first objective function, and λ(t) be the positive weight function. This indicates the expected value, where x0 is the initial input image data, and x... t For the noisy image data at time information t, p t (x t |x0) represents a Gaussian perturbation kernel centered at x0. S is the gradient of the logarithmic probability density.θ1 (x,t) is a score estimator for inverse time SDE in the image domain, which can be trained using a denoising score matching method.

[0092] More specifically, the image domain-based sampling generator specifically includes a first predictor and a first detector, and the implementation process of step S124 includes, but is not limited to, the following:

[0093] Step S124.1: Perform a single sampling operation on the optimized k-th input data through the first predictor to obtain the k-th first initial sampled data. The mathematical expression used in this process is:

[0094]

[0095] Step S124.2: Perform a correction operation on the kth initial sampled data using the first detector. The mathematical expression used in this process is:

[0096]

[0097] Step S124.3: Apply data consistency constraints to the corrected k-th initial sampled data using the Simultaneous Iterative Reconstruction Technique (SIRT) to obtain the k-th initial sampled data. The mathematical expression used in this process is:

[0098] v t-1 =v′ t-1 +QA T P(y-Av′ t-1 (10)

[0099] In the formula, v" t-1 This is the first initial sampled data. The sampling noise scale, To optimize the input data x t-0.5 The noise characteristic value obtained after prediction, z~N(0,1) is the unit noise, v′ t-1 For the first initial sampled data after correction, ∈ t-1 For random noise disturbance, For the first initial sampled data v" t-1 The noise characteristic value v obtained after prediction is t-1 For the first sampled data, Q = diag{1 / ‖a1‖,...,1 / ||a j ||,...,1 / ‖a n ‖} is a diagonal matrix, aj Let n be the j-th column vector of the system matrix A, and n be the number of columns in the system matrix A. P = diag{1 / ‖a1‖,...,1 / ‖a i ‖,...,1 / ‖a m ‖} is a diagonal matrix, a i Let m be the i-th row vector of system matrix A, and let m be the row number of system matrix A. T This is the transpose of the system matrix A.

[0100] More specifically, the implementation process of step S125 above includes, but is not limited to, the following:

[0101] Step S125.1: Introduce a coupling factor to separate and transform the first optimization function to obtain the second optimization function. This process is specifically manifested as follows:

[0102] First, by introducing a coupling factor, the above equation (1) is initially transformed to obtain:

[0103]

[0104] Secondly, separating the two variables on the left side of equation (11) above, we obtain the second optimization function as follows:

[0105]

[0106]

[0107] Step S125.2: After inputting the kth first sampled data into the second optimization function, solve it using the existing conjugate gradient method to obtain the optimized kth first sampled data;

[0108] In the formula, x′ t-0.5 For the optimized first sampled data, β t v is the coupling factor. t To process the input data x using the image domain-based sampling generator t The solution results are as follows.

[0109] More specifically, the implementation of step S126 above is as follows: The optimized k-th first sampled data x′ is processed using the existing Discrete Wavelet Transform (DWT) method. t-0.5 The input data is decomposed into four first input data groups containing different frequency components (LL, LH, HL, HH), and then these are combined to form the k-th first input data group, denoted as X. t-0.5Where LL represents the low-frequency component, LH represents the high-frequency component in the horizontal direction, HL represents the high-frequency component in the vertical direction, and HH represents the high-frequency component in the diagonal direction.

[0110] First of all, see Figure 4 As shown, the method for obtaining the wavelet domain-based sampling generator mentioned in step S127 above specifically includes the following:

[0111] Step B1: Obtain the finite-angle CT training image;

[0112] Step B2: Obtain the second diffusion model based on the wavelet domain. The second diffusion model is also constructed using stochastic differential equations, specifically including the second forward process and the second backward process.

[0113] Step B3: Perform wavelet transform on the finite angle CT training image to obtain a finite angle CT training sub-image group. Specifically, the existing discrete wavelet transform method is used to decompose the finite angle CT training image into four finite angle CT training sub-images containing different frequency components (LL, LH, HL, HH), and then form a finite angle CT training sub-image group.

[0114] Step B4: Noise is progressively injected into the finite-angle CT training sub-image group through the first forward process to obtain the corresponding second noisy image group. The mathematical expression used in this process is:

[0115] dx=f(x,t)dt+g(t)du; (5)

[0116] Because this invention exemplarily chooses to employ a typical variance explosion stochastic differential equation, i.e., limiting f = 0 and Formula (5) above can be transformed into:

[0117]

[0118] Step B5: In the first reverse process, a second scoring network is trained to progressively denoise the second noisy image group. The trained second scoring network is a wavelet domain-based sampling generator. The second scoring network is controlled by parameter θ2, and its second objective function during training is:

[0119]

[0120] In the formula, The second objective function is... The initial input is image data containing frequency component c. For noisy image data containing frequency component c, Indicated by A Gaussian perturbation kernel centered at c, where c = LL, LH, HL, HH. The score estimator for inverse-time SDE in the wavelet domain can be trained using a denoised score matching method.

[0121] It should be noted that by training in the wavelet domain, different image patterns of different frequency components can be effectively captured, and the wavelet transform maintains the spatial correspondence with the original image, which can simplify the modeling of the probability density of each wavelet component.

[0122] More specifically, the wavelet domain-based sampling generator specifically includes a second predictor and a second detector, and the implementation process of step S127 includes, but is not limited to, the following:

[0123] First, it should be noted that the k-th first input data group w t-0.5 The dataset contains four first input data points carrying different frequency components. Each of these first input data points only needs to be processed according to steps S127.1 to S127.3 to obtain the corresponding four second sample data points. These are then combined to form the kth second sample data group, denoted as w. t-1 ;

[0124] Step S127.1: Perform a single sampling operation on the first input data using the second predictor to obtain the second initial sampled data. The mathematical expression used in this process is:

[0125]

[0126] Step S127.2: Perform a correction operation on the second initial sampled data using the second detector. The mathematical expression used in this process is:

[0127]

[0128] Step S127.3: Apply data consistency constraints to the corrected second initial sampled data using joint iterative reconstruction techniques to obtain the second sampled data. The mathematical expression used in this process is:

[0129] w c，t-1 =w′ c，t-1 +QA T P(t-Aw′ c，t-1 (17)

[0130] In the formula, w c,t-0.5 The first input data contains the frequency component c. For the first input data w c,t-0.5 The noise characteristic value, w", obtained after prediction isc,t-1 For the second initial sampled data containing frequency component c, w′ c,t-1 This is the corrected second initial sampling data, which includes the frequency component c. For the second initial sampling data w" c,t-1 The noise characteristic value w obtained after prediction c,t-1 This is the second sampled data containing the frequency component c.

[0131] In this embodiment of the invention, by performing depth sampling processing of the input data in the image domain and wavelet domain in each denoising process included in the image reconstruction task, wavelet transform is introduced to analyze the distribution characteristics of finite angle artifacts in different frequency components. This is beneficial for separating and removing artifacts from the output data of each denoising process, ensuring the consistency and effectiveness of the detailed structure, and improving the quality of image reconstruction.

[0132] In the following description, the finite angle CT reconstruction method based on wavelet multi-channel scoring proposed in this invention is referred to as the WIMD model. The reliability of the implementation of the WIMD model is verified through several different experimental methods.

[0133] 1. Numerical simulation experiment

[0134] (1) Selection of comparison models: Five traditional CT reconstruction models were selected for experimental comparison and analysis, namely FBP (Filter Back Projection) model, FISTA (Fast Iterative Shrinkage-Thresholding Algorithm) model, DPS (Deep Probability Subsampling) model, FBPConvNet model and DOLCE model.

[0135] (2) Dataset creation: The WIMD model proposed in this invention was trained and evaluated using the simulation dataset of the AAPM (American Association of Medical Physicists) 2016 CT Low Dose Challenge; the iso-angle fan beam projection detector was used as an experimental auxiliary device, and the distance from the rotation center to the light source and the detector was set to 512 mm. The finite angle scanning process required to be applied included 180 viewing angles evenly distributed within a 90° range; the simulation dataset was scanned using the iso-angle fan beam projection detector to generate the final simulation dataset containing only 90°T images.

[0136] (3) Experimental environment setup: A high-performance computing system equipped with an NVIDIA RTX A6000 graphics processing unit with 48GB of memory.

[0137] (4) Analysis of experimental results:

[0138] See Figure 5 As shown, it mainly provides CT reconstruction results corresponding to two different slice images. The first row of images is the CT reconstruction result corresponding to the first slice image, the second row of images is the difference effect between the first row of images and the first slice image, the third row of images is the CT reconstruction result corresponding to the second slice image, and the fourth row of images is the difference effect between the third row of images and the second slice image.

[0139] from Figure 5 It is evident that the CT reconstruction results obtained by the FBP and FISTA models after implementation exhibit significant image distortion, damaging subtle differences and unique features in the image. The CT reconstruction results obtained by the FBPConvNet model after implementation show obvious artifacts. After extracting and enlarging the region of interest (i.e., the rectangular box region), it can be seen that the CT reconstruction results obtained by the DPS and DOLCE models after implementation fail to faithfully capture details. In contrast, the WIMD model proposed in this invention best restores complex structures and subtle image details in the CT reconstruction results after implementation, which can be further verified by the relevant difference effect diagram.

[0140] To further quantify and demonstrate the advantages of the WIMD model proposed in this invention, two evaluation parameters are introduced here: the PSNR (Peak Signal-to-Noise Ratio) index and the SSIM (Structural Similarity Index) index. The higher the values ​​of the PSNR index and / or the SSIM index, the higher the similarity between the two images used for comparison.

[0141] Figure 5 The first row of images contains six CT reconstructed images corresponding to the first slice image. The PSNR and SSIM values ​​between each CT reconstructed image and the first slice image are calculated. Figure 5 The third row of images contains six CT reconstructed images corresponding to the second slice image. The PSNR and SSIM values ​​between each CT reconstructed image and the second slice image are calculated. The specific calculation results are shown in Table 1. Through numerical calculation, it is verified again that the reconstructed image obtained after the implementation of the WIMD model proposed in this invention is closest to the real slice image.

[0142] Table 1 shows the quantification results of PSNR and SSIM in the numerical simulation experiments.

[0143]

[0144] 2. Experiments based on sheep lung data

[0145] (1) Selection of comparison models: Five traditional CT reconstruction models were selected for experimental comparison and analysis, namely FBP model, FISTA model, DPS model, FBPConvNet model and DOLCE model.

[0146] (2) Dataset creation: The WIMD model trained directly on the simulated dataset of the AAPM 2016 CT low-dose challenge was used; the lungs of anesthetized sheep were scanned by Somatom Sensation 64-slice CT scanner, and a field of view (FOV) with a radius of 25.05 cm was defined. The first sheep lung slice image in the [0°, 90°] angle range and the second sheep lung slice image in the [0°, 60°] angle range were further obtained.

[0147] (3) Experimental environment setup: A high-performance computing system equipped with an NVIDIA RTX A6000 graphics processing unit with 48GB of memory.

[0148] (4) Analysis of experimental results:

[0149] See Figure 6 As shown, the first row of images represents the CT reconstruction results corresponding to the first sheep lung slice image, and the second row of images represents the CT reconstruction results corresponding to the second sheep lung slice image; from Figure 6 It is evident that the CT reconstruction results obtained by the FBP and FISTA models after implementation exhibit significant image artifacts, making it difficult to discern image details and features. The FBPConvNet model, after implementation, misses image details and introduces obvious new artifacts. Furthermore, by extracting and enlarging the region of interest (i.e., the rectangular bounding box region), it can be seen that the CT reconstruction results obtained by the DPS and DOLCE models after implementation fail to faithfully capture details. In contrast, the WIMD model proposed in this invention, after implementation, best restores complex structures and subtle image details.

[0150] To further quantify and demonstrate the advantages of the WIMD model proposed in this invention, two evaluation parameters, PSNR and SSIM, are introduced here. Figure 6 The first row of images contains six CT reconstructed images corresponding to the first sheep lung slice image. The PSNR and SSIM values ​​between each CT reconstructed image and the first sheep lung slice image are calculated. Figure 6The second row of images contains six CT reconstructed images corresponding to the second sheep lung slice image. The PSNR and SSIM values ​​between each CT reconstructed image and the second sheep lung slice image are calculated. The specific calculation results are shown in Table 2. Through numerical calculation, it is verified again that the reconstructed image obtained after the implementation of the WIMD model proposed in this invention is closest to the real slice image.

[0151] Table 2 shows the quantitative results of PSNR and SSIM in the experiment based on sheep lung data.

[0152]

[0153]

[0154] 3. Experiments based on cardiac clinical data

[0155] (1) Selection of comparison models: Five traditional CT reconstruction models were selected for experimental comparison and analysis, namely FBP model, FISTA model, DPS model, FBPConvNet model and DOLCE model.

[0156] (2) Dataset creation: The WIMD model trained directly on the simulated dataset of the AAPM 2016 CT low-dose challenge was used; the GE Discovery CT750 HD scanner was used to perform coronary artery CT scans on clinical patients in spiral mode, and a scan trajectory with a radius of 53.85 cm was defined. The first coronary artery slice image in the [0°, 90°] angle range and the second coronary artery slice image in the [0°, 60°] angle range were further obtained.

[0157] (3) Experimental environment setup: A high-performance computing system equipped with an NVIDIA RTX A6000 graphics processing unit with 48GB of memory.

[0158] (4) Analysis of experimental results:

[0159] See Figure 7 As shown, the first row of images represents the CT reconstruction results corresponding to the first coronary artery slice image, and the second row of images represents the CT reconstruction results corresponding to the second coronary artery slice image; from Figure 7It is evident that the CT reconstruction results obtained by the FBP and FISTA models after implementation exhibit significant image distortion, while the CT reconstruction results obtained by the FBPConvNet model miss image details. Furthermore, by extracting and enlarging the region of interest (i.e., the rectangular box region), it can be seen that the CT reconstruction results obtained by the DPS and DOLCE models after implementation fail to faithfully capture details. In contrast, the WIMD model proposed in this invention best restores complex structures and subtle image details.

[0160] To further quantify and demonstrate the advantages of the WIMD model proposed in this invention, two evaluation parameters, PSNR and SSIM, are introduced here. Figure 7 The first row of images contains six CT reconstructed images corresponding to the first coronary artery slice image. The PSNR and SSIM values ​​between each CT reconstructed image and the first coronary artery slice image are calculated. Figure 7 The second row of images contains six CT reconstructed images corresponding to the second coronary artery slice image. The PSNR and SSIM values ​​between each CT reconstructed image and the second coronary artery slice image are calculated. The specific calculation results are shown in Table 3. Through numerical calculation, it is verified again that the reconstructed image obtained after the implementation of the WIMD model proposed in this invention is closest to the real slice image.

[0161] Table 3. Quantitative results of PSNR and SSIM in experiments based on cardiac clinical data.

[0162]

[0163] 4. Ablation test

[0164] (1) Selection of comparison models: Two traditional CT reconstruction models were selected for experimental comparison and analysis, namely the first VE-SDE model that performs reconstruction tasks in the wavelet domain and the second VE-SDE model that performs reconstruction tasks in the image domain.

[0165] (2) Dataset creation: The WIMD model trained directly on the simulated dataset of the AAPM 2016 CT low-dose challenge was used; and a CT slice image was randomly selected.

[0166] (3) Experimental environment setup: A high-performance computing system equipped with an NVIDIA RTX A6000 graphics processing unit with 48GB of memory.

[0167] (4) Analysis of experimental results:

[0168] from Figure 8 It is evident that although the first VE-SDE model produces only a few artifacts in the CT reconstruction results after implementation, inconsistencies in some detailed structures can be observed after extracting and magnifying the region of interest (i.e., the rectangular region). These detailed structures are crucial in clinical practice, and distorted structures may lead to misdiagnosis. The second VE-SDE model tends to over-smooth the CT reconstruction results after implementation, and some detailed textures can be missed after extracting and magnifying the region of interest (i.e., the rectangular region). In contrast, the WIMD model proposed in this invention produces only a few artifacts in the CT reconstruction results after implementation and ensures consistency in detailed structures.

[0169] To further quantify and demonstrate the advantages of the WIMD model proposed in this invention, two evaluation parameters, PSNR and SSIM, are introduced here. Figure 8 The three CT reconstructed images corresponding to the CT slice image are used to calculate the PSNR and SSIM values ​​between each CT reconstructed image and the CT slice image. The specific calculation results are shown in Table 4. Through numerical calculation, it is verified again that the reconstructed image obtained after the implementation of the WIMD model proposed in this invention is closest to the real slice image.

[0170] Table 4. Quantitative results of PSNR and SSIM in ablation experiments.

[0171]

[0172] To further demonstrate the advantages of the WIMD model proposed in this invention, the PSNR corresponding to different sampling steps is calculated and convergence analysis is performed during the reverse reconstruction processes of the first VE-SDE model, the second VE-SDE model, and the WIMD model. (See [link to relevant documentation]). Figure 9 As shown, in the first 1000 sampling steps, the PSNR of the three models increases at roughly the same rate, while in the last 1000 sampling steps, the WIMD model increases significantly at a faster rate, demonstrating better reconstruction performance.

[0173] Please refer to Figure 10 , Figure 10 This invention provides a finite-angle CT reconstruction system based on wavelet multi-channel scoring, the system comprising:

[0174] The first module 210 is used to acquire finite-angle CT images with added noise.

[0175] The second module 220 is used to perform step-by-step denoising on the finite-angle CT noisy image according to a preset time length. In each denoising process, the input data is processed by depth sampling in the image domain and wavelet domain and then output to the next denoising process. The output data of the last denoising process is then used as the CT reconstructed image.

[0176] The content of the above method embodiments is applicable to this system embodiment. The functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are the same as those in the above method embodiments. Therefore, they will not be repeated here.

[0177] Furthermore, embodiments of the present invention also provide a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements a finite-angle CT reconstruction method based on wavelet multi-channel scoring as described in the above embodiments. The computer-readable storage medium includes, but is not limited to, any type of disk (including floppy disk, hard disk, optical disk, CD-ROM, and magneto-optical disk), ROM (Read-Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic cards, or optical cards. In other words, the storage device includes any medium on which a device (e.g., a computer, mobile phone, etc.) stores or transmits information in a readable form, and can be a read-only memory, a disk, or an optical disk, etc.

[0178] also, Figure 11 This is a schematic diagram of the hardware structure of a computer device provided in an embodiment of the present invention. The computer device includes components such as a processor 320, a memory 330, an input unit 340, and a display unit 350. Those skilled in the art will understand that... Figure 11The illustrated device structure is not intended to limit all devices and may include more or fewer components than shown, or combine certain components. The memory 330 can be used to store the computer program 310 and various functional modules. The processor 320 runs the computer program 310 stored in the memory 330, thereby performing various functional applications and data processing of the device. The memory can be internal memory or external memory, or include both internal and external memory. Internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory. External memory may include hard disks, floppy disks, USB flash drives, magnetic tapes, etc. The memory 330 disclosed in the embodiments of this invention includes, but is not limited to, these types of memory. The memory 330 disclosed in the embodiments of this invention is only an example and not a limitation.

[0179] Input unit 340 is used to receive signal input and user-input keywords. Input unit 340 may include a touch panel and other input devices. The touch panel can collect user touch operations on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near the touch panel) and drive the corresponding connection device according to a pre-set program; other input devices may include, but are not limited to, one or more of physical keyboards, function keys (such as play control buttons, power buttons, etc.), trackballs, mice, joysticks, etc. Display unit 350 can be used to display user-input information or information provided to the user, as well as various menus of the terminal device. Display unit 350 may be in the form of a liquid crystal display, organic light-emitting diode, etc. Processor 320 is the control center of the terminal device, connecting various parts of the entire device through various interfaces and lines, performing various functions and processing data by running or executing software programs and / or modules stored in memory 330, and calling data stored in memory 330.

[0180] As one embodiment, the computer device includes a processor 320, a memory 330, and a computer program 310, wherein the computer program 310 is stored in the memory 330 and configured to be executed by the processor 320, and the computer program 310 is configured to perform a finite-angle CT reconstruction method based on wavelet multi-channel scoring in the above embodiment.

[0181] Although the description of this application has been quite detailed and particularly focused on several of the described embodiments, it is not intended to limit itself to any of these details or embodiments or any particular embodiment. Rather, it should be considered as effectively covering the intended scope of this application by referring to the appended claims and taking into account the prior art, which provides for a broad possible interpretation of these claims. Furthermore, the foregoing description of this application with respect to embodiments foreseeable by the inventors is intended to provide a useful description, and non-substantial modifications to this application that have not yet been foreseen may still represent equivalent modifications.

Claims

1. A finite-angle CT reconstruction method based on wavelet multi-channel scoring, characterized in that, include: Acquire finite-angle CT images with added noise; The denoised CT image with limited angle is gradually denoised according to a preset time length. In each denoising process, the input data is processed by depth sampling in the image domain and wavelet domain and then output to the next denoising process. The output data of the last denoising process is then used as the CT reconstructed image. The step of performing depth sampling processing on the input data in both the image domain and wavelet domain during each denoising process and then outputting it to the next denoising process includes: The input data is optimized, and then the optimized input data is sampled once using an image domain-based sampling generator to obtain the first sampled data. The first sampled data is optimized, and then wavelet transform is performed on the optimized first sampled data to obtain the first input data group; The first input data group is sampled once using a wavelet domain-based sampling generator to obtain the corresponding second sampled data group; Perform inverse wavelet transform on the second sampled data group to obtain the final output data; The image-domain-based sampling generator includes a first predictor and a first detector. The step of using the image-domain-based sampling generator to perform a single sampling of the optimized input data to obtain the first sampled data includes: The first predictor is used to perform a single sampling of the optimized input data to obtain the first initial sampled data. The mathematical expression used in this process is as follows: ； In the formula, This is the first initial sampled data. For input data, For the optimized input data, The sampling noise scale, The noise feature values ​​carried by the optimized input data, Unit noise; The mathematical expression used in correcting the first initial sampled data using the first detector is as follows: ； In the formula, The first initial sampled data after correction. For random noise disturbance, The noise characteristic value carried by the first initial sampled data; Data consistency constraints are applied to the corrected initial sampled data to obtain the first sampled data; The wavelet-domain-based sampling generator includes a second predictor and a second detector. The step of using the wavelet-domain-based sampling generator to perform a single sampling of the first input data set to obtain the corresponding second sampled data set includes: For each first input data in the first input data group, the second predictor performs a single sampling of the first input data to obtain second initial sample data. The mathematical expression used in this process is as follows: ； In the formula, The first input data contains the frequency component c. This is the second initial sampling data containing the frequency component c. The noise characteristics carried by the first input data containing the frequency component c; The second detector is used to correct the second initial sampled data. The mathematical expression used in this process is as follows: ； In the formula, This is the corrected second initial sampling data, which includes the frequency component c. The noise characteristic value carried by the second initial sampled data containing the frequency component c; Data consistency constraints are applied to the corrected second initial sampled data to obtain the corresponding second sampled data.

2. The finite-angle CT reconstruction method based on wavelet multi-channel scoring according to claim 1, characterized in that, The optimization of the input data includes: Obtain the first optimization function, then input the input data into the first optimization function and use the conjugate gradient method to solve for the optimized input data.

3. The finite-angle CT reconstruction method based on wavelet multi-channel scoring according to claim 2, characterized in that, The optimization of the first sampled data includes: By introducing a coupling factor to separate and transform the first optimization function, a second optimization function is obtained; After inputting the first sampled data into the second optimization function, the conjugate gradient method is used to solve the problem and obtain the optimized first sampled data.

4. The finite-angle CT reconstruction method based on wavelet multi-channel scoring according to claim 1, characterized in that, The image domain-based sampling generator is obtained in the following way: Acquire finite-angle CT training images and a first diffusion model constructed based on stochastic differential equations, wherein the first diffusion model includes a first forward process and a first backward process; During the first forward process, noise is gradually injected into the finite-angle CT training image to obtain the first noisy image; In the first reverse process, a first scoring network is trained to progressively denoise the first noisy image. The trained first scoring network is an image domain-based sampling generator.

5. The finite-angle CT reconstruction method based on wavelet multi-channel scoring according to claim 4, characterized in that, The wavelet domain-based sampling generator is obtained in the following way: The finite-angle CT training image and a second diffusion model constructed based on stochastic differential equations are obtained. The second diffusion model includes a second forward process and a second backward process. Perform wavelet transform on the finite-angle CT training images to obtain a finite-angle CT training sub-image group; During the second forward process, noise is gradually injected into the finite angle CT training sub-image group to obtain the corresponding second noisy image group; In the second reverse process, a second scoring network is trained to progressively denoise the second noisy image group. The trained second scoring network is a wavelet domain-based sampling generator.

6. A finite-angle CT reconstruction system based on wavelet multi-channel scoring, characterized in that, include: The first module is used to acquire finite-angle CT images with added noise. The second module is used to perform stepwise denoising on the finite-angle CT noisy image according to a preset time length. In each denoising process, the input data is processed by depth sampling in the image domain and wavelet domain and then output to the next denoising process. The output data of the last denoising process is then used as the CT reconstructed image. The step of performing depth sampling processing on the input data in both the image domain and wavelet domain during each denoising process and then outputting it to the next denoising process includes: The input data is optimized, and then the optimized input data is sampled once using an image domain-based sampling generator to obtain the first sampled data. The first sampled data is optimized, and then wavelet transform is performed on the optimized first sampled data to obtain the first input data group; The first input data group is sampled once using a wavelet domain-based sampling generator to obtain the corresponding second sampled data group; Perform inverse wavelet transform on the second sampled data group to obtain the final output data; The image-domain-based sampling generator includes a first predictor and a first detector. The step of using the image-domain-based sampling generator to perform a single sampling of the optimized input data to obtain the first sampled data includes: The first predictor is used to perform a single sampling of the optimized input data to obtain the first initial sampled data. The mathematical expression used in this process is as follows: ； In the formula, This is the first initial sampled data. For input data, For the optimized input data, The sampling noise scale, The noise feature values ​​carried by the optimized input data, Unit noise; The mathematical expression used in correcting the first initial sampled data using the first detector is as follows: ； In the formula, The first initial sampled data after correction. For random noise disturbance, The noise characteristic value carried by the first initial sampled data; Data consistency constraints are applied to the corrected initial sampled data to obtain the first sampled data; The wavelet-domain-based sampling generator includes a second predictor and a second detector. The step of using the wavelet-domain-based sampling generator to perform a single sampling of the first input data set to obtain the corresponding second sampled data set includes: For each first input data in the first input data group, the second predictor performs a single sampling of the first input data to obtain second initial sample data. The mathematical expression used in this process is as follows: ； In the formula, The first input data contains the frequency component c. This is the second initial sampling data containing the frequency component c. The noise characteristics carried by the first input data containing the frequency component c; The second detector is used to correct the second initial sampled data. The mathematical expression used in this process is as follows: ； In the formula, This is the corrected second initial sampling data, which includes the frequency component c. The noise characteristic value carried by the second initial sampled data containing the frequency component c; Data consistency constraints are applied to the corrected second initial sampled data to obtain the corresponding second sampled data.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, The processor executes the computer program to implement the finite-angle CT reconstruction method based on wavelet multi-channel scoring as described in any one of claims 1 to 5.

Images