Computed tomography reconstruction method and system based on deep-prior dual-domain network and terminal

Abstract

The computer tomography reconstruction method, system and terminal based on the deep prior dual-domain network of the application, through modeling the deep prior dual-domain consistency such as imaging geometry constraint, image domain and projection domain combination, the low dose computer tomography image is reconstructed. The system not only introduces the physics of imaging geometry into the reconstruction process, but also enhances the model interpretability, compared with the traditional deep learning reconstruction algorithm, not only has more excellent image reconstruction quality, but also has about 4 times acceleration performance and the required data set for effective training is smaller, and can transmit low-level and high-level features, so as to promote the faster convergence of the model, solve the problems such as gradient disappearance and gradient explosion in the traditional deep learning reconstruction algorithm.

Description

Technical Field

[0001] This invention relates to the field of deep learning and computed tomography reconstruction, and in particular to a computed tomography reconstruction method, system and terminal based on deep prior dual-domain networks. Background Technology

[0002] Most existing techniques for deep learning reconstruction focus on refined representations of images in the image or projection domains; however, these models do not incorporate the physics of imaging geometry into the reconstruction process, lacking interpretability—the working principle of deep learning is often referred to as a black box. Traditional deep learning methods often involve large training datasets, requiring the learning of millions to billions of parameters, thus demanding high computational power in system design. Furthermore, some ongoing research on algorithmic derivation has employed residual learning networks such as ResNet, UNet, or simpler networks; however, achieving optimal results may require multiple iterative optimizations, further increasing computational time.

[0003] For example, GE Healthcare recently developed TrueFidelity technology based on deep learning for low-dose computed tomography (LDCT) reconstruction. This network receives low-dose sinograms (raw data) and reconstructs them into low-dose images using its proprietary software. The low-dose images are then compared to real images acquired from high-dose CT scans. A refined representation of the reconstructed image is achieved by optimizing millions of parameters to reduce the difference between the real and low-dose images. The image reconstruction model is repeatedly trained with a large number of training samples until the output image closely resembles the real image. AiCE technology designed by Canon Medical employs a similar process, but differs from the FBP algorithm used in TrueFidelity.

[0004] While these computed tomography (CT) image reconstruction networks exhibit impressive performance, they lack interpretability and require large datasets (thousands of data points) for optimal training. Yan et al. parameterized the proximal operator in the Alternating Direction Multiplier (ADMM) algorithm using CNNs, combining the advantages of model-based and data-driven MRI sparse sampling reconstruction methods. Joas and Ozan unfolded the Primal Dual Hybrid Gradient (PDHG) algorithm to form a feedback deep learning for LDCT reconstruction; however, this method only performs primal dual learning on projection domain data, potentially missing some details inherent in the image domain. Xia et al. proposed an iterative scheme based on manifold and graph integral convolutional networks (MAGIC) to unfold images and low-dimensional manifold spaces. This method utilizes spatial convolution to extract local pixel-level features from images and merges graph convolutions. However, due to its simplicity for approximating spatial filtering or denoising networks, the number of iteration blocks can range from 50 to 100 to achieve good results, significantly increasing the overall computational cost. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a computed tomography reconstruction method, system and terminal based on deep prior dual-domain networks to solve the above-mentioned technical problems in the prior art.

[0006] To achieve the above and other related objectives, the present invention provides a computed tomography reconstruction method based on a depth prior dual-domain network. The method includes: acquiring a low-dose computed tomography image to be reconstructed; and obtaining a corresponding high-quality reconstructed image based on the low-dose computed tomography image using a depth prior dual-domain LDCT reconstruction model that employs imaging geometric constraints and combines the corresponding image domain and projection domain.

[0007] In one embodiment of the present invention, the method for constructing the depth-prior dual-domain LDCT reconstruction model includes: solving the image domain update optimization problem and the projection domain update optimization problem of the original dual objective function based on the ADMM algorithm, respectively, to obtain the corresponding image domain update optimization solution results and projection domain update optimization solution results; parameterizing the image domain update optimization solution results and the projection domain update optimization solution results to obtain the learning parameters for iterating multiple iterative blocks containing convolutional structures in the corresponding training; wherein the convolutional structure adopts the DenseNet structure and deconvolutional layers; and training the depth-prior dual-domain LDCT reconstruction model using training data containing imaging geometric simulation projection data based on each learning parameter.

[0008] In one embodiment of the present invention, the parameterization of the image domain update optimization solution result and the projection domain update optimization solution result to obtain the learning parameters for iterating multiple iterative blocks of the corresponding training containing convolutional structures includes: obtaining the image domain update optimization parameters and projection domain update optimization parameters corresponding to each iteration based on the image domain update optimization solution result and the projection domain update optimization solution result; and obtaining the learning parameters for iterating multiple times for each iterative block of the corresponding training containing convolutional structures based on the image domain update optimization parameters and projection domain update optimization parameters of each iteration.

[0009] In one embodiment of the present invention, the learning parameters of each iterative block are obtained by the inverse operator, threshold operator, and update layer obtained from the image domain update optimization parameters for the corresponding iteration number, and by the inverse operator, threshold operator, and update layer obtained from the projection domain update optimization parameters.

[0010] In one embodiment of the present invention, the imaging geometric simulation projection data includes: geometric parameters for simulating projection data; wherein the geometric parameters include: detector length, distance between detector and X-ray source, focal point of X-ray source, distance between detector and rotation center, object to be simulated, and radius of phantom.

[0011] In one embodiment of the present invention, each iteration block is performed four times.

[0012] In one embodiment of the present invention, the depth-prior dual-domain LDCT reconstruction model is further constrained by a real image, and the depth-prior dual-domain LDCT reconstruction model includes: an iterative block containing a convolutional structure of multiple connected layers; wherein the convolutional structure includes: a deformable layer and a reconstruction layer; and wherein the deformable layer includes: a DenseNet block of multiple connected layers; the reconstruction layer includes: a deconvolutional layer of multiple connected layers.

[0013] In one embodiment of the present invention, the depth-prior dual-domain LDCT reconstruction model uses Gaussian noise to perform random initialization of the depth prior.

[0014] To achieve the above and other related objectives, the present invention provides a computed tomography reconstruction system based on a depth prior dual-domain network. The system includes: a data acquisition module for acquiring a low-dose computed tomography image to be reconstructed; and a reconstruction module connected to the data acquisition module for obtaining a corresponding high-quality reconstructed image based on the low-dose computed tomography image using a depth prior dual-domain LDCT reconstruction model that employs imaging geometric constraints and combines the corresponding image domain and projection domain.

[0015] To achieve the above and other related objectives, the present invention provides a computed tomography reconstruction terminal based on a deep prior dual-domain network, comprising: one or more memories and one or more processors; the one or more memories are used to store a computer program; the one or more processors are connected to the memories and are used to run the computer program to execute the computed tomography reconstruction method based on a deep prior dual-domain network.

[0016] As described above, this invention is a computed tomography reconstruction method, system, and terminal based on a deep prior dual-domain network, which has the following beneficial effects: This invention reconstructs low-dose computed tomography images by employing deep prior dual-domain LDCT remodeling with imaging geometric constraints and corresponding image and projection domains combined. This not only incorporates the physics of imaging geometry into the reconstruction process, making it interpretable, but also accelerates the process by about 4 times compared to traditional deep learning algorithms and requires a smaller dataset for effective training. Furthermore, it can transfer low-level and high-level features, promoting faster model convergence and solving the gradient vanishing and gradient exploding problems, thereby reducing the number of iteration blocks. Attached Figure Description

[0017] Figure 1 The diagram shown is a flowchart of a computed tomography reconstruction method based on a deep prior dual-domain network according to an embodiment of the present invention.

[0018] Figure 2 The diagram shows a flowchart of the optimization solution using the ADMM algorithm in one embodiment of the present invention.

[0019] Figure 3 The diagram shown is a schematic diagram of the expansion algorithm of the analytical equation in one embodiment of the present invention.

[0020] Figure 4 The diagram shows geometric parameters in one embodiment of the present invention.

[0021] Figure 5 The diagram shown is a structural schematic of the DPDudoNet framework in one embodiment of the present invention.

[0022] Figure 6 The diagram shown is a schematic representation of a computed tomography reconstruction system based on a deep prior dual-domain network according to an embodiment of the present invention.

[0023] Figure 7 The diagram shown is a schematic representation of a computed tomography reconstruction terminal based on a deep prior dual-domain network according to an embodiment of the present invention. Detailed Implementation

[0024] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0025] It should be noted that in the following description, reference is made to the accompanying drawings, which illustrate several embodiments of the invention. It should be understood that other embodiments may also be used, and changes in mechanical composition, structure, electrical system, and operation may be made without departing from the spirit and scope of the invention. The following detailed description should not be considered limiting, and the scope of the embodiments of the invention is defined only by the claims of the published patents. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. Spatially related terms, such as “upper,” “lower,” “left,” “right,” “below,” “below,” “lower part,” “above,” “upper part,” etc., may be used herein to illustrate the relationship between one element or feature shown in the figures and another element or feature.

[0026] Throughout this specification, when it is said that a part is "connected" to another part, this includes not only "direct connection" but also "indirect connection" by placing other elements in between. Furthermore, when it is said that a part "includes" a certain constituent element, unless otherwise stated otherwise, this does not exclude other constituent elements, but rather means that other constituent elements may also be included.

[0027] The terms "first," "second," and "third," etc., used herein are for the purpose of describing various parts, components, regions, layers, and / or segments, but are not limiting. These terms are used only to distinguish one part, component, region, layer, or segment from others. Therefore, the "first part," "component," "region," "layer," or "segment" described below may refer to a "second part," "component," "region," "layer," or "segment" without departing from the scope of this invention.

[0028] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are interpreted as inclusive, or mean any one or any combination thereof. Thus, “A, B, or C” or “A, B, and / or C” means “any one of: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition arise only when combinations of elements, functions, or operations are inherently mutually exclusive in some manner.

[0029] This invention provides a computed tomography (CT) reconstruction method, system, and terminal based on a deep prior dual-domain network. By employing imaging geometric constraints and combining the corresponding image domain and projection domain in deep prior dual-domain LDCT remodeling, low-dose CT images are reconstructed. This not only incorporates the physics of imaging geometry into the reconstruction process, making it interpretable, but also accelerates the process by approximately 4 times compared to traditional deep learning algorithms, requires a smaller dataset for effective training, and can transfer low-level and high-level features, promoting better convergence. It solves the gradient vanishing and gradient exploding problems, thereby reducing the number of iteration blocks.

[0030] The present invention will now be described in detail with reference to the accompanying drawings, so that those skilled in the art can readily implement it. The present invention can be embodied in many different forms and is not limited to the embodiments described herein.

[0031] like Figure 1 This diagram illustrates a flowchart of a computed tomography reconstruction method based on a deep prior dual-domain network, as described in an embodiment of the present invention.

[0032] The method includes:

[0033] Step S11: Acquire the low-dose computed tomography image to be reconstructed.

[0034] Optionally, the computed tomography images can be preprocessed to a set spatial resolution; they can also be subsequently processed with data augmentation, such as symmetric inversion or random rotation, to achieve data expansion.

[0035] Step S12: Based on the depth prior dual-domain LDCT reconstruction model that employs imaging geometric constraints and combines the corresponding image domain and projection domain, obtain the corresponding reconstructed image according to the low-dose computed tomography image.

[0036] Optionally, the method for constructing the depth-prior dual-domain LDCT reconstruction model includes:

[0037] Based on the ADMM algorithm, the image domain update optimization problem and the projection domain update optimization problem of the original dual objective function are solved respectively to obtain the corresponding image domain update optimization solution and projection domain update optimization solution.

[0038] The image domain update optimization solution and the projection domain update optimization solution are parameterized to obtain the learning parameters for iterative training of multiple iterative blocks containing convolutional structures; wherein, the convolutional structure adopts the DenseNet structure and deconvolutional layers.

[0039] Based on the learning parameters, a depth-prior dual-domain LDCT reconstruction model is trained using training data containing imaging geometric simulation projection data. Specifically, the learning parameters for each iteration block are used to constrain the corresponding iteration blocks in the reconstruction model.

[0040] Optional, such as Figure 2 As shown, the original dual objective function includes:

[0041]

[0042] The image domain update optimization solution obtained by using the ADMM algorithm to solve the original dual objective function includes:

[0043]

[0044]

[0045]

[0046] The optimization results obtained by solving the projection domain update problem of the original dual objective function using the ADMM algorithm include:

[0047]

[0048]

[0049]

[0050] Where μ1 and μ2 are image regularization parameters, It is the inverse operator. For thresholding operators, use CNN operators as... An approximate calculation is performed, with the following trainable parameters: α, γ, μ1, μ2.

[0051] Optionally, the parameterization of the image domain update optimization solution and the projection domain update optimization solution to obtain the learning parameters for iterating through multiple iterative blocks containing convolutional structures in the corresponding training includes:

[0052] Based on the image domain update optimization solution results and the projection domain update optimization solution results, the corresponding image domain update optimization parameters and projection domain update optimization parameters for each iteration are obtained;

[0053] Based on the image domain update optimization parameters and the projection domain update optimization parameters of each iteration, the learning parameters of each iteration block containing the convolutional structure are obtained after multiple iterations.

[0054] Preferably, each iteration block undergoes four iterations.

[0055] In a preferred embodiment, the learning parameters for each iteration block are obtained by the inverse operator, threshold operator, and update layer obtained from the image domain update optimization parameters for the corresponding iteration number, and by the inverse operator, threshold operator, and update layer obtained from the projection domain update optimization parameters; specifically, as follows Figure 3 As shown, it gives Figure 1 A schematic diagram of the expansion algorithm for analytical equations. and These are CNN operators, representing the negation operator, thresholding operator, and layer update operator.

[0056] In one embodiment, the imaging geometric simulation projection data includes: geometric parameters for simulating the projection data; wherein, the geometric parameters include: detector length, distance between the detector and the X-ray source, focal point of the X-ray source, distance between the detector and the rotation center, the object to be simulated, and the radius of the phantom. Specifically, training data is prepared by simulating projection data using predefined geometric figures; such as... Figure 4 As shown, the geometric parameters include: D = detector length; DSD = distance between detector and X-ray source; F = focal point of X-ray source; DOR = distance between detector and rotation center; O = object to be simulated; r = radius of phantom.

[0057] In one embodiment, the network is trained using predefined parameters, and then the results are evaluated using a common image quality metric.

[0058] In one embodiment, the depth-prior dual-domain LDCT reconstruction model is further constrained by real images, and the depth-prior dual-domain LDCT reconstruction model includes:

[0059] Multiple connected iterative blocks containing convolutional structures; wherein the convolutional structures include: deformable layers and reconstructed layers;

[0060] Furthermore, the deformation layer comprises multiple connected DenseNet blocks; the reconstruction layer comprises multiple connected deconvolutional layers. This framework combines projection and image priors with imaging physics in a semi-supervised manner. The threshold (nearest neighbor) operator parameters are updated in each epoch through end-to-end training based on a defined loss function.

[0061] In one embodiment, the depth-prior dual-domain LDCT reconstruction model uses Gaussian noise to perform random initialization of the depth prior.

[0062] To better illustrate the above-mentioned computed tomography reconstruction method based on deep prior dual-domain networks, the present invention provides the following specific embodiments.

[0063] Example 1: A computed tomography reconstruction method based on deep prior dual-domain networks. For example... Figure 5 The diagram shown is a schematic of the DPDudoNet framework used.

[0064] This implementation works in four stages:

[0065] 1) First stage: Solving the dual-domain optimization problem. Obtaining the solution to the optimization problem for the unfolded network of LDCT, with the inverse operator and threshold operator approximated using appropriate CNN operators.

[0066] 2) Second stage: Using DenseNet with deconvolutional layers, an unfolding algorithm is designed to approximate the solution. This involves obtaining the learned... and The parameters are CNN operators used to represent the inverse operator, threshold operator, and update layer.

[0067] 3) Third stage: Prepare training data by simulating projection data using predefined geometry. Geometric parameters used to simulate projection data: D = detector length; DSD = distance between detector and X-ray source; F = focal point of X-ray source; DOR = distance between detector and rotation center; O = object to be simulated; r = radius of phantom.

[0068] 4) Fourth stage: Train the network using predefined parameters, and then evaluate the results using a common image quality metric. For example... Figure 5 As shown, this framework combines projection and image priors with imaging physics in a semi-supervised manner. The threshold (nearest neighbor) operator parameters are updated in each epoch through end-to-end training based on a defined loss function. N is the number of iterations corresponding to each block (N=4). It uses Gaussian noise to perform random initialization of depth priors.

[0069] Inspired by the work of Yan et al. and Joas and Ozan, this method proposes a novel prior-based dual-domain reconstruction model for LDCT reconstruction, which is easier to interpret and generalize. This method embeds the inherent imaging geometry constraints into the joint learning process between the spatial and cubemap domains through dual-domain updates. and The inverse problem approximates the coupling of a DenseNet CNN with a deconvolutional network. Deep priors are integrated by randomly initializing CNNs with Gaussian noise to capture low-level information, even with limited training data and time. This method contributes in four ways:

[0070] 1) A dual-domain model combining the sinogram and image domains was developed, and then the ADMM algorithm with the primal dual objective function was used to solve the problem. This helps to integrate information into the sinogram and image domains through data consistency and the training update process.

[0071] 2) Unfold this model into a network and approximate it with a deconvolutional layer, similar to DenseNet. This process provides better interpretability for the network because each block can be viewed as an iterative process of the ADMM algorithm.

[0072] 3) The unexpanded network can be effectively trained on smaller datasets, which are about 4 times smaller than those of traditional deep learning algorithms, because the network inherits this generalization property of the general ADMM algorithm.

[0073] 4) This method approximates a CNN using DenseNet with deconvolutional layers and exhibits better performance than other residual networks such as ResNet and UNet. This is because there are dense connections between different layers in DenseNet, allowing the transfer of low-level and high-level features, thus promoting better convergence and mitigating the problems of uplighting and exploding gradients, thereby reducing the number of iteration blocks (N=4). This method also achieves better performance by incorporating depth priors into the network through random initialization of convolutional and non-convolutional layers.

[0074] Similar to the principles of the above embodiments, the present invention provides a computed tomography reconstruction system based on a deep prior dual-domain network.

[0075] The following specific embodiments are provided in conjunction with the accompanying drawings:

[0076] like Figure 6 This diagram illustrates the structure of a computed tomography reconstruction system based on a deep prior dual-domain network, according to an embodiment of the present invention.

[0077] The system includes:

[0078] Data acquisition module 61 is used to acquire low-dose computed tomography images to be reconstructed;

[0079] The reconstruction module 62, connected to the data acquisition module 61, is used to obtain the corresponding reconstructed image based on the low-dose computed tomography image, using a depth-prior dual-domain LDCT reconstruction model that employs imaging geometric constraints and combines the corresponding image domain and projection domain.

[0080] It should be noted that, as should be understood Figure 6 The division of modules in the system embodiment is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these units can be implemented entirely in software through processing element calls; they can be implemented entirely in hardware; or some units can be implemented by processing element calls to software, while others are implemented in hardware.

[0081] Since the implementation principle of the computed tomography reconstruction system based on deep prior dual-domain networks has been described in the foregoing embodiments, it will not be repeated here.

[0082] Optionally, the method for constructing the depth-prior dual-domain LDCT reconstruction model includes: based on the ADMM algorithm, solving the image domain update optimization problem and the projection domain update problem of the original dual objective function respectively to obtain the corresponding image domain update optimization solution results and projection domain update optimization solution results; parameterizing the image domain update optimization solution results and the projection domain update optimization solution results to obtain the learning parameters for iteratively training multiple iterative blocks containing convolutional structures; wherein, the convolutional structure adopts the DenseNet structure and deconvolutional layers; and training the depth-prior dual-domain LDCT reconstruction model using training data containing imaging geometric simulation projection data based on each learning parameter.

[0083] Optionally, the parameterization of the image domain update optimization solution and the projection domain update optimization solution to obtain the learning parameters for iterating multiple iterative blocks of the corresponding training containing convolutional structures includes: obtaining the image domain update optimization parameters and projection domain update optimization parameters for each iteration based on the image domain update optimization solution and the projection domain update optimization solution; and obtaining the learning parameters for iterating multiple times for each iterative block of the corresponding training containing convolutional structures based on the image domain update optimization parameters and projection domain update optimization parameters for each iteration.

[0084] Optionally, the learning parameters for each iteration block are obtained by the inverse operator, threshold operator, and update layer obtained from the image domain update optimization parameters for the corresponding iteration number, as well as by the inverse operator, threshold operator, and update layer obtained from the projection domain update optimization parameters.

[0085] Optionally, the imaging geometric simulation projection data includes: geometric parameters for simulating the projection data; wherein the geometric parameters include: detector length, distance between the detector and the X-ray source, focal point of the X-ray source, distance between the detector and the rotation center, the object to be simulated, and the radius of the phantom.

[0086] Optionally, each iteration block can be iterated four times.

[0087] Optionally, the depth-prior dual-domain LDCT reconstruction model is further constrained by a real ground image, and the depth-prior dual-domain LDCT reconstruction model includes: multiple connected iterative blocks containing convolutional structures; wherein the convolutional structure includes: a deformable layer and a reconstruction layer; and wherein the deformable layer includes: multiple connected DenseNet blocks; the reconstruction layer includes: multiple connected deconvolutional layers.

[0088] Optionally, the depth-prior dual-domain LDCT reconstruction model uses Gaussian noise to enforce random initialization of the depth prior.

[0089] like Figure 7 This is a schematic diagram of the structure of a computed tomography reconstruction terminal 10 based on a deep prior dual-domain network according to an embodiment of the present invention.

[0090] The computed tomography reconstruction terminal 70 based on a deep prior dual-domain network includes: a memory 71 and a processor 72. The memory 71 stores computer programs; the processor 72 runs the computer programs to implement, for example,... Figure 1 The aforementioned computed tomography reconstruction method based on deep prior dual-domain networks.

[0091] Optionally, the number of memories 71 can be one or more, and the number of processors 72 can be one or more. Figure 7 Each example is taken as an instance.

[0092] Optionally, the processor 72 in the computed tomography reconstruction terminal 70 based on a deep prior dual-domain network will proceed as follows: Figure 1 The steps described involve loading one or more instructions corresponding to the process of an application into memory 71, and having the processor 72 run the application stored in the first memory 71, thereby achieving the following: Figure 1 The various functions in the computed tomography reconstruction method based on deep prior dual-domain networks.

[0093] Optionally, the memory 71 may include, but is not limited to, high-speed random access memory and non-volatile memory. For example, one or more disk storage devices, flash memory devices, or other non-volatile solid-state storage devices; the processor 72 may include, but is not limited to, a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0094] Optionally, the processor 72 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0095] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed, implements as follows: Figure 1 The illustrated method is a computed tomography reconstruction method based on a deep prior dual-domain network. The computer-readable storage medium may include, but is not limited to, floppy disks, optical disks, CD-ROMs (Read-Only Optical Disk Memory), magneto-optical disks, ROMs (Read-Only Memory), RAMs (Random Access Memory), EPROMs (Erasable Programmable Read-Only Memory), EEPROMs (Electrically Erasable Programmable Read-Only Memory), magnetic cards or optical cards, flash memory, or other types of media / machine-readable media suitable for storing machine-executable instructions. The computer-readable storage medium may be a product not connected to a computer device or a component used in a computer device.

[0096] In summary, the computed tomography reconstruction system based on a deep prior dual-domain network of the present invention reconstructs low-dose computed tomography images by employing deep prior dual-domain LDCT remodeling with imaging geometric constraints and corresponding image and projection domains combined. This not only incorporates imaging geometric constraints into the reconstruction process but also provides interpretability. It is approximately four times faster than traditional deep learning algorithms and requires less effective training data. Furthermore, it can transfer both low-level and high-level features, promoting better convergence. It also solves the gradient vanishing and gradient exploding problems in traditional deep learning reconstruction algorithms, thereby reducing the number of iteration blocks. Therefore, the present invention effectively overcomes the various shortcomings of existing technologies and has high industrial application value.

[0097] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A computed tomography reconstruction method based on a deep-prior dual-domain network, characterized in that, The method comprises: acquiring a low-dose computed tomography image to be reconstructed; obtaining a corresponding reconstructed image from the low-dose computed tomography image based on a deep-prior dual-domain LDCT reconstruction model combining an imaging geometry constraint and corresponding image domain and projection domain; wherein the deep-prior dual-domain LDCT reconstruction model construction method comprises: based on an ADMM algorithm, solving an image domain update optimization problem and a projection domain update optimization problem of an original dual objective function respectively to obtain corresponding image domain update optimization solving results and projection domain update optimization solving results; parameterizing the image domain update optimization solving results and the projection domain update optimization solving results to obtain learning parameters corresponding to training a plurality of iterative blocks containing a convolution structure for iteration; wherein the convolution structure adopts a DenseNet structure and a deconvolution layer; based on each learning parameter, training a deep-prior dual-domain LDCT reconstruction model using training data containing imaging geometry simulation projection data; the parameterization of the image domain update optimization solving results and the projection domain update optimization solving results to obtain learning parameters corresponding to training a plurality of iterative blocks containing a convolution structure for iteration comprises: obtaining image domain update optimization parameters and projection domain update optimization parameters corresponding to each iteration based on the image domain update optimization solving results and the projection domain update optimization solving results; based on the image domain update optimization parameters and the projection domain update optimization parameters of each iteration, obtaining learning parameters corresponding to training each iterative block containing a convolution structure for multiple iterations. 2.The computed tomography reconstruction method based on deep-prior dual-domain network according to claim 1, wherein, The learning parameters of each iterative block are obtained by an inverse operator, a threshold operator and an update map layer obtained from the image domain update optimization parameters corresponding to the number of iterations, and an inverse operator, a threshold operator and an update map layer obtained from the projection domain update optimization parameters. 3.The computer tomography reconstruction method based on deep-prior dual-domain network according to claim 1, characterized in that, The imaging geometry simulation projection data comprises: geometric parameters for simulating projection data; wherein the geometric parameters comprise: detector length, distance between detector and X-ray source, focus of X-ray source, distance between detector and rotation center, object to be simulated, and radius of phantom. 4.The computer tomography reconstruction method based on deep-prior dual-domain network according to claim 1, characterized in that, Each iterative block performs four iterations. 5.The computer tomography reconstruction method based on deep-prior dual-domain network according to claim 1, wherein, The deep-prior dual-domain LDCT reconstruction model is further constrained by a real image, and the deep-prior dual-domain LDCT reconstruction model comprises: a plurality of iterative blocks containing a convolution structure of a plurality of connection layers; wherein the convolution structure comprises: a deformation layer and a reconstruction layer; and wherein the deformation layer comprises: a DenseNet block of a plurality of connection layers; and the reconstruction layer comprises: a deconvolution layer of a plurality of connection layers. 6.The computer tomography reconstruction method based on deep-prior dual-domain network according to claim 1, wherein, The deep-prior dual-domain LDCT reconstruction model uses Gaussian noise to perform random initialization of the deep-prior.

7. A computed tomography reconstruction system based on a deep prior dual-domain network, characterized in that, The system comprises: a data acquisition module for acquiring a low-dose computed tomography image to be reconstructed; a reconstruction module connected to the data acquisition module for obtaining a corresponding reconstructed image from the low-dose computed tomography image based on a deep-prior dual-domain LDCT reconstruction model combining an imaging geometry constraint and corresponding image domain and projection domain; The deep-prior bi-domain LDCT reconstruction model construction method comprises: based on an ADMM algorithm, solving an image domain update optimization problem and a projection domain update optimization problem of an original dual objective function respectively to obtain corresponding image domain update optimization solving results and projection domain update optimization solving results; parameterizing the image domain update optimization solving results and the projection domain update optimization solving results to obtain learning parameters corresponding to training a plurality of iterative blocks containing a convolution structure to iterate; wherein the convolution structure adopts a DenseNet structure and a deconvolution layer; based on each learning parameter, training a deep-prior bi-domain LDCT reconstruction model by using training data containing imaging geometry simulation projection data; The parameterizing the image domain update optimization solving results and the projection domain update optimization solving results to obtain learning parameters corresponding to training a plurality of iterative blocks containing a convolution structure to iterate comprises: obtaining image domain update optimization parameters and projection domain update optimization parameters corresponding to each iteration based on the image domain update optimization solving results and the projection domain update optimization solving results; obtaining learning parameters corresponding to training each iterative block containing a convolution structure to iterate multiple times based on the image domain update optimization parameters and the projection domain update optimization parameters of each iteration.

8. A computer tomography reconstruction terminal based on a deep-prior dual-domain network, characterized in that, Comprise: one or more memories and one or more processors; The one or more memories are used to store a computer program; The one or more processors, connected to the memory, are used to run the computer program to execute the method according to any one of claims 1 to 6.

Images