Methods, systems, computer equipment, and storage media for removing metal artifacts from medical images

Abstract

This invention provides a method, system, computer device, and storage medium for removing metal artifacts from medical images. The method includes acquiring the original projection data of a metal-containing medical image to be processed and obtaining an initial reconstructed image through a filtered back-projection algorithm; inputting the initial reconstructed image into a topological homeomorphism mapping engine, calculating the differential manifold mapping through an integral velocity vector field to obtain topology-preserving manifold latent space features; extracting a sinusoidal domain mask from the metal implant region based on the original projection data and a physical forward projection model to obtain a damaged sinusoidal sequence; inputting the manifold latent space features and the damaged sinusoidal sequence into a dual-domain joint tensor ring network for high-dimensional tensor decomposition and core tensor iterative completion to obtain dual-domain joint features; constraining the dual-domain joint features based on a physical consistency loss function, and reconstructing the image domain through an inverse homeomorphism mapping operator to obtain a metal-artifact-free medical image. This application can reliably remove metal artifacts.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and computer-aided diagnosis technology, specifically to a method, system, computer device, and storage medium for removing metal artifacts from medical images based on topological homeomorphism and physically constrained tensor ring networks. Background Technology

[0002] Computed tomography (CT) is a routine diagnostic tool in modern medical imaging. When a patient has high-density metal implants (such as dental prostheses, artificial joints, and internal fixation screws for fractures), X-rays undergo attenuation and distortion as they penetrate high atomic number materials. This phenomenon mainly stems from beam hardening (the absorption of low-energy X-ray photons by the metal, leading to an increase in the average energy of the transmitted beam, deviating from the linear assumption of the Beer-Lambert law), photon starvation (the low number of photons passing through the metal, resulting in signal interference from Poisson noise), and scattering effects. These physical effects produce radial bright and dark stripe artifacts and shadow artifacts on the reconstructed CT images, severely obscuring the bone and soft tissue structures around the implant. This not only interferes with the accurate assessment of fracture healing but also reduces the reliability of postoperative rehabilitation evaluation. Therefore, medical image optimization techniques are needed to suppress these artifacts.

[0003] Current metal artifact reduction (MAR) techniques mainly fall into the following categories:

[0004] The first type is based on sine wave interpolation algorithms (such as Normalized Metal Artifact Removal Algorithm, NMAR). This method treats data obscured by metal in the original projection (sine wave) as missing data and fills it in using an interpolation algorithm. Since pure mathematical interpolation does not utilize prior information about the actual anatomical structure, the reconstructed image is prone to generating secondary artifacts at the edges of the implant while removing artifacts in non-metallic areas.

[0005] The second category is based on iterative reconstruction (IR) methods, which iteratively optimize the objective function by introducing regularization terms such as total variation (TV). These methods perform well in terms of physical consistency, but have high computational complexity, are time-consuming, and have limitations in their ability to repair photon starvation.

[0006] The third category is artifact removal methods based on deep learning. Existing deep learning models typically treat MAR (Mass Image Reconstruction) as a mapping problem in the image domain, or use dual-domain networks to perform feature restoration in the sinusoidal image domain and the image domain respectively. However, existing deep learning methods have the following limitations in practical applications: 1) Insufficient topological constraints: Conventional convolutional neural networks (CNNs) lack manifold topological constraints when restoring missing regions, and are prone to generating tissue structures that do not conform to the patient's actual condition based on prior training data (i.e., network illusion), which carries the risk of altering the anatomical topology. 2) Limited spatial correlation extraction: CT sinusoidal image data has non-local spatial geometric correlations. The local receptive field of conventional CNNs limits their ability to extract features from global integral geometric properties, resulting in poor global consistency of sinusoidal image completion. 3) Low integration with physical models: Some deep learning solutions focus on smoothing image visual features without fully integrating the physical equations of X-ray forward and backward projection. This may lead to deviations between the network output image and the physical data measured by the original detector when the image is reprojected onto the sinusoidal image.

[0007] In summary, how to innovatively propose a method for removing metal artifacts that can extract high-dimensional correlations from sinusoidal graphs, conform to physical forward projection constraints, and possess topology preservation capabilities is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] To address the technical problems of existing metal artifact removal techniques, this invention provides a method, system, computer device, and storage medium for removing metal artifacts from medical images based on topological homeomorphism and physically constrained tensor ring networks. The aim is to improve the shortcomings of existing deep learning MAR technology in terms of topological deformation and physical consistency, and to achieve reliable metal artifact removal.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] On one hand, the present invention provides a method for removing metal artifacts from medical images, comprising:

[0011] The original projection data of the metal-containing medical image to be processed is obtained, and the initial reconstructed image is obtained through a filtered back projection algorithm;

[0012] The initial reconstructed image is input into the topological homeomorphism mapping engine, and the differential manifold mapping is calculated by integrating the velocity vector field to obtain the topology-preserving manifold latent space features.

[0013] Based on the original projection data and the physical forward projection model, a sinusoidal domain mask is extracted from the metal implant area to obtain a damaged sinusoidal sequence.

[0014] The manifold latent space features and the damaged sinusoid sequence are jointly input into a dual-domain joint tensor ring network for high-dimensional tensor decomposition and core tensor iterative completion to obtain dual-domain joint features.

[0015] The joint features of the two domains are constrained based on the physical consistency loss function, and the image domain is reconstructed by the inverse homeomorphism mapping operator to obtain the final medical image with metal artifacts removed.

[0016] Furthermore, the step of acquiring the original projection data of the metal-containing medical image to be processed and obtaining the initial reconstructed image through a filtered back projection algorithm specifically includes: receiving the original X-ray projection signal output by the CT scanner, performing logarithmic transformation, beam hardening correction and photon starvation noise filtering on the original projection signal, and reconstructing the initial image using the inverse Radon transform and filter.

[0017] Furthermore, in the step of inputting the initial reconstructed image into the topological homeomorphism engine, the formula for the topological homeomorphism is expressed as:

[0018]

[0019] in, Indicates continuous time steps Next pixel coordinates Differential homeomorphic deformation field, Indicates at a point in time The non-stationary velocity vector field at that location belongs to the regenerating kernel Hilbert space. , For time-integral infinitesimals; constrain the Jacobian determinant. This ensures the topological invariance of anatomical structures during the mapping process and prevents tissue boundary distortion and folding during the metal removal process.

[0020] Furthermore, in the dual-domain joint tensor ring network, the damaged sine wave sequence is reconstructed into a high-dimensional tensor. Its tensor ring decomposition formula is expressed as: ;in, This represents the value of the element at the corresponding index position in the high-dimensional tensor. The trace operation represents the trace of a matrix. Indicates the first A third-order core tensor For the first Size of the space in one dimension Let be the rank of the tensor ring, and satisfy the closed-loop boundary conditions. This decomposition is used to extract nonlocal low-rank correlation features in the data-missing regions of the sine curve trajectory.

[0021] Furthermore, the dual-domain joint tensor ring network includes a dynamic quantized attention mechanism. This attention mechanism calculates the artifact probability amplitude by mapping classical pixel intensity to Hilbert space. The formula for the attention mechanism is expressed as:

[0022]

[0023] in, This represents the extracted quantized eigenstate vector. For partition normalization constant, The imaginary unit, This represents the phase angle corresponding to the local pixel gradient. It is a non-linear activation function. Let be the projection weight matrix. Indicates the first In the layer network The feature vectors of the hidden nodes To calculate the ground state.

[0024] Furthermore, in the step of constraining the joint features of the two domains based on the physical consistency loss function, the physical consistency loss function... The formula is expressed as:

[0025]

[0026] in, This indicates a data fidelity penalty. Here is the forward projection matrix of the system based on the Beer-Lambert law. The reconstructed attenuation coefficient matrix output by the network. Projected data for the portion without metal obstruction. It is a Poisson diagonal weighted matrix; The curl penalty term for the velocity field is used to maintain the smoothness of the manifold mapping; The nuclear norm of the core tensor; , , and These are all hyperparameters that control the weights of each item.

[0027] Furthermore, the step of obtaining the final metal artifact-free medical image includes: inputting the features with completed tensor rings into a multi-scale feature decoder; during the decoding process, utilizing the deformation inverse matrix recorded by the topological homeomorphism mapping engine. Perform inverse resampling of the latent space; solve the linear inverse problem with total variation regularization term using the conjugate gradient algorithm, reconstruct and output the medical image with metal artifacts removed.

[0028] On the other hand, the present invention provides a medical image metal artifact removal system, comprising:

[0029] The data acquisition and projection module is used to acquire the original projection data of the metal-containing medical image to be processed, and to obtain the initial reconstructed image through a filtered back projection algorithm;

[0030] The topological homeomorphism mapping module is used to input the initial reconstructed image into the topological homeomorphism mapping engine, calculate the differential manifold mapping through the integral velocity vector field, and obtain the topology-preserving manifold latent space features.

[0031] The sinusoidal mask extraction module is used to extract the sinusoidal domain mask of the metal implant area based on the original projection data and the physical forward projection model to obtain the damaged sinusoidal sequence.

[0032] The tensor ring high-dimensional completion module is used to input the manifold latent space features and the damaged sinusoidal graph sequence into the dual-domain joint tensor ring network to perform high-dimensional tensor decomposition and core tensor iterative completion to obtain dual-domain joint features.

[0033] The physical reconstruction output module is used to constrain the joint features of the two domains based on the physical consistency loss function, and to reconstruct the image domain through the inverse homeomorphism mapping operator to obtain the final metal artifact-free medical image.

[0034] In another aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the above-described method for removing metal artifacts from medical images.

[0035] In another aspect, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the steps in the above-described method for removing metal artifacts from medical images.

[0036] Compared with existing technologies, the medical image metal artifact removal method, system, computer equipment, and storage medium provided by this invention have the following advantages:

[0037] (1) Introducing a topological homeomorphic differential manifold engine: The continuous velocity field integral mapping based on the regenerated kernel Hilbert space is adopted to ensure the positive definiteness of the Jacobian determinant of the mapping transformation, effectively avoiding structural breakage and deformation generated when the network repairs large-area artifacts, and ensuring the topological consistency of the anatomical structure.

[0038] (2) Constructing a closed-loop tensor ring and phase interference mechanism: By utilizing the closed mathematical properties of the tensor ring and adapting to the periodicity of the CT sine wave, the extraction and completion of global projection features were realized; combined with the phase interference principle of the quantum attention mechanism, the features were weighted according to the gradient direction, which improved the sharpness of the tissue edge.

[0039] (3) Combining medical physical projection constraints: Introducing orthographic projection constraints and Poisson weighting based on photon statistics at the network reconstruction end ensures the physical consistency of the results, enabling the CT density values ​​of the reconstructed voxels to have high quantitative analysis accuracy, which can meet the requirements of clinical diagnosis and treatment for data reliability. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the macroscopic process of the medical image metal artifact removal method provided by the present invention.

[0041] Figure 2 This is a diagram of the internal vector flow field integration architecture of the topological homeomorphism mapping engine provided by the present invention.

[0042] Figure 3 This is a schematic diagram of the dual-domain joint tensor ring network structure and the quantized attention interference model provided by the present invention.

[0043] Figure 4 This is an example image showing the actual effect of removing metal artifacts provided by the present invention. Detailed Implementation

[0044] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0045] As one specific embodiment, please refer to Figures 1 to 4 As shown, the present invention provides a method for removing metal artifacts from medical images, the method comprising the following steps:

[0046] S1: Obtain the original projection data of the metal-containing medical image to be processed, and obtain the initial reconstructed image through a filtered back projection algorithm;

[0047] S2: Input the initial reconstructed image into the topological homeomorphism mapping engine, calculate the differential manifold mapping through the integral velocity vector field, and obtain the topology-preserving manifold latent space features;

[0048] S3: Based on the original projection data and the physical forward projection model, perform sinusoidal domain mask extraction on the metal implant area to obtain the damaged sinusoidal sequence;

[0049] S4: Input the manifold latent space features and the damaged sinusoidal sequence into a dual-domain joint tensor ring network to perform high-dimensional tensor decomposition and core tensor iterative completion to obtain dual-domain joint features;

[0050] S5: Constrain the joint features of the two domains based on the physical consistency loss function, and reconstruct the image domain through the inverse homeomorphism mapping operator to obtain the final medical image without metal artifacts.

[0051] In implementing the above method, this application introduces topological diffeomorphic mapping within the framework of Lie groups and Lie algebras. By constructing a non-stationary continuous velocity vector field, the transition process from the artifact-containing space to the latent space in the image satisfies the conditions of continuous differentiability and the existence of an inverse mapping. This design ensures that the Jacobian determinant of the mapping process is greater than zero, thereby preventing the breakage and folding of the spatial structure during artifact removal under mathematical constraints, and avoiding the generation of false tissues.

[0052] Furthermore, to address the limitations of traditional neural networks in extracting global information from sine waves, this scheme employs a Tensor Ring (TR) network structure. The medical projection sequence is constructed as a high-dimensional tensor, and tensor ring decomposition extracts long-range nonlocal features through its closed-loop chain structure. Core tensor iterative completion can reduce the number of parameters while also addressing the nonlinear data loss caused by beam hardening.

[0053] Furthermore, this method employs a physical consistency loss function constraint. This requires that the reconstructed image, after forward projection transformation, matches the data from the metal-free region actually captured by the detector, thus ensuring the physical objectivity of the generated image.

[0054] In one embodiment, step S1 specifically includes: receiving the raw X-ray projection signal output by the CT scanner, performing logarithmic transformation, beam hardening correction, and photon starvation noise filtering on the raw projection signal, and reconstructing the initial image using the inverse Radon transform and filters.

[0055] In one implementation, in step S2, where the initial reconstructed image is input into the topological homeomorphism engine, the formula for the topological homeomorphism is expressed as:

[0056]

[0057] in, Indicates continuous time steps Next pixel coordinates Differential homeomorphic deformation field, Indicates at a point in time The non-stationary velocity vector field at that location belongs to the regenerative kernel Hilbert space (RKHS). , For time-integral infinitesimals; constrain the Jacobian determinant. This ensures the topological invariance of anatomical structures during the mapping process and prevents tissue boundary distortion and folding during the metal removal process.

[0058] In the aforementioned implementation process, the topological homeomorphism formula describes a continuous flow model, modeling the artifact removal process as pixel drift within continuous time steps. Due to the velocity vector field... The deformation field after integration in a smooth Hilbert space It possesses differential homeomorphic properties, which allows the geometric adjacency of tissues to be preserved when removing artifact pixel mutations.

[0059] In one implementation, in the dual-domain joint tensor ring network of step S4, the damaged sine wave sequence is reconstructed into a high-dimensional tensor. Its tensor ring decomposition formula is expressed as:

[0060]

[0061] in, This represents the value of the element at the corresponding index position in the high-dimensional tensor. The trace operation represents the trace of a matrix. Indicates the first A third-order core tensor For the first The spatial size of a dimension (such as channel dimension, projection angle dimension, slice thickness dimension). Let be the rank of the tensor ring, and satisfy the closed-loop boundary conditions. This decomposition is used to extract nonlocal low-rank correlation features in the data-missing regions of the sine curve trajectory.

[0062] In the aforementioned implementation process, trace operations are performed. Connecting the first and last core tensors to form a loop structure aligns with the periodic continuous boundary conditions of the projection angles at 0° and 360° in the CT sine wave image. Tensor loop decomposition helps to complete the data features obscured by metal from the high-dimensional latent space.

[0063] In one implementation, the dual-domain joint tensor ring network in step S4 includes a dynamically quantized attention mechanism. This attention mechanism calculates the artifact probability amplitude by mapping classical pixel intensity to Hilbert space. The formula for the attention mechanism is expressed as:

[0064]

[0065] in, This represents the extracted quantized eigenstate vector. For partition normalization constant, The imaginary unit, This represents the phase angle corresponding to the local pixel gradient. It is a non-linear activation function. Let be the projection weight matrix. Indicates the first In the layer network The feature vectors of the hidden nodes To calculate the ground state.

[0066] In the aforementioned implementation process, considering the variation characteristics of pixel gradients in the artifact region, the modal information of pixel intensity and the direction information of gradient are mapped to the amplitude and phase in complex space. Based on the principle of phase interference, the features of tissue boundaries with consistent orientation are enhanced, while the features of divergent artifact noise are attenuated, thereby improving the network's ability to distinguish edge details.

[0067] In one implementation, in step S5, the step of constraining the joint features of the two domains based on the physical consistency loss function, the physical consistency loss function... The formula is expressed as:

[0068]

[0069] in, This indicates a data fidelity penalty. Here is the forward projection matrix of the system based on the Beer-Lambert law. The reconstructed attenuation coefficient matrix output by the network. Projected data for the portion without metal obstruction. It is a Poisson diagonal weighted matrix; The curl penalty term for the velocity field is used to maintain the smoothness of the manifold mapping; The nuclear norm of the core tensor; , , and These are all hyperparameters that control the weights of each item.

[0070] In the aforementioned implementation process, the data fidelity penalty term is combined with statistical weighting. The three elements together constitute a joint optimization framework: ensuring the weight of non-metallic region data in the reconstruction; the curl penalty term constrains the fluid mapping process to generate smooth deformation; and the kernel norm term imposes low-rank constraints to extract the underlying features of the background tissue.

[0071] In one implementation, step S5, obtaining the final metal artifact-free medical image, includes: inputting the features completed by tensor ring completion into a multi-scale feature decoder; during the decoding process, utilizing the deformation inverse matrix recorded by the topological homeomorphism mapping engine. Perform inverse resampling of the latent space; solve the linear inverse problem with total variation regularization term using the conjugate gradient algorithm, reconstruct and output the medical image with metal artifacts removed.

[0072] To better understand the medical image metal artifact removal method provided by this invention, the following examples further illustrate the method. The method of this application includes the following steps:

[0073] S100: Data Preprocessing and Physical Beam Forward Projection

[0074] For example, acquiring raw data generated by a clinical CT scanner. The original transmitted photon count intensity Converted to attenuated projected values ​​based on the modified Beer-Lambert law: ;in, The initial intensity of the incident X-rays. This represents the Poisson-Gaussian mixed noise term. Subsequently, the filtered back projection (FBP) algorithm is used, combining a high-pass filter with the Radon inverse transform operator. Reconstruct the initial image containing artifacts. In the image domain, the metal implant mask is extracted through thresholding segmentation and morphological operations. The metal mask is projected onto the system's forward projection matrix. Reprojecting back into the sinusoidal domain yields the sinusoidal metal mask trajectory. , The marked area is the sequence of damaged sine curves.

[0075] S200: Topological homeomorphism at the manifold geometry level

[0076] For example, this application defines the repair process as a continuous fluid deformation process. Given an initial artifact image as... Construct a time-evolving velocity vector field differential equation: , ;in, Describes from time arrive The deformation process, This is an identity mapping. Solving the differential equation by integration yields... To ensure the preservation of the topology, the velocity field... An inner product norm penalty term based on the reproducing kernel Hilbert space (RKHS) is applied. The resulting mapping Jacobian matrix is... According to the Liouville formula, the following conditions are met: Since the range of the exponential function is greater than zero, this method ensures that the Jacobian determinant of the spatial mapping is always positive, keeping the generating network on a differential homeomorphic manifold and thus avoiding changes to the connectivity of the anatomical structure during the removal of metal artifacts. After performing this step, the latent space features of the manifold are obtained. .

[0077] S300: Tensor Ring Decomposition and High-Dimensional Feature Compensation in the Sinusoidal Domain

[0078] Reconstructing clinical CT data into higher-order tensors (These represent detector channel element, detector row number, rotation projection angle, and scan layer thickness, respectively).

[0079] The sinusoidal graph to be repaired and its corresponding manifold latent space feature tensors are stacked and input into a tensor loop network. The end-to-end structure of the tensor loop adapts to the periodicity of the physical rotation of the CT scan. The core tensor is optimized... Establish a complete objective functional:

[0080]

[0081] in, For Hadamard's element-wise product, The nuclear norm is used. This model utilizes the joint distribution between different projection angles and scanning layers for data completion.

[0082] Introducing a dynamic quantized attention mechanism into tensor networks to control pixel intensity Mapping to the complex ground state: Introducing phase angle Encodes the gradient direction field. Tissue edge features in the same direction produce constructive interference and are preserved, while divergent artifact noise is attenuated due to phase cancellation, improving the edge sharpness of artifact removal.

[0083] S400: Bi-domain Joint Consistency Constraints Based on Physics Inverse Problem Solving

[0084] In the output layer, image reconstruction is defined as a physics inverse problem solution process. This utilizes a physics consistency loss function. Optimization was performed, including a data fidelity penalty item. Using Poisson diagonal weighted matrix .matrix The weight distribution is based on the variance of the photon count received by the detector unit, which increases the proportion of high signal-to-noise ratio region data in the reconstruction.

[0085] Solving linear inverse problems with total variation regularization using the conjugate gradient (CG) optimizer: During the solution process, the deformation inverse matrix provided by the topology engine is used. Inverse spatial resampling is performed to output a final image with physical fidelity.

[0086] In a clinical validation example: For patient scans with bilateral high-density titanium alloy hip joint prostheses, traditional network models, when removing stripe artifacts, are prone to generating irregular edges or tomographic misalignments within the pelvic cavity due to a lack of topological constraints. After processing with the metal artifact removal method based on topological homeomorphism and tensor ring networks provided in this application, the tensor ring network recovers the missing sinusoidal data using the correlation between angles; simultaneously, topological homeomorphism ensures that the Jacobian determinant of the mapping process is greater than zero, maintaining the continuity of the anatomical structures of the rectum and bladder wall. The output reconstructed image shows no topological breaks, and the physical fidelity projection error is significantly reduced compared to the comparative model.

[0087] Therefore, those skilled in the art should understand that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows, unless explicitly stated herein, i.e., there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily need to be completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages does not necessarily have to be sequential, but can be executed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0088] As another specific embodiment, the present invention provides a medical image metal artifact removal system, comprising:

[0089] The data acquisition and projection module is used to acquire the original projection data of the metal-containing medical image to be processed, and to obtain the initial reconstructed image through a filtered back projection algorithm;

[0090] The topological homeomorphism mapping module is used to input the initial reconstructed image into the topological homeomorphism mapping engine, calculate the differential manifold mapping through the integral velocity vector field, and obtain the topology-preserving manifold latent space features.

[0091] The sinusoidal mask extraction module is used to extract the sinusoidal domain mask of the metal implant area based on the original projection data and the physical forward projection model to obtain the damaged sinusoidal sequence.

[0092] The tensor ring high-dimensional completion module is used to input the manifold latent space features and the damaged sinusoidal graph sequence into the dual-domain joint tensor ring network to perform high-dimensional tensor decomposition and core tensor iterative completion to obtain dual-domain joint features.

[0093] The physical reconstruction output module is used to constrain the joint features of the two domains based on the physical consistency loss function, and to reconstruct the image domain through the inverse homeomorphism mapping operator to obtain the final metal artifact-free medical image.

[0094] For other specific limitations regarding medical image metal artifact removal systems, please refer to the description of medical image metal artifact removal methods above, which will not be repeated here.

[0095] Therefore, those skilled in the art should understand that each functional module in the above-mentioned medical image metal artifact removal system can be implemented in whole or in part through software, hardware, or a combination thereof, and each of the above-mentioned functional modules can be embedded in the processor of the computer device in hardware form or independent of the processor, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each of the above modules.

[0096] In one embodiment, a computer device is provided, which may be a terminal, the internal structure of which is known to those skilled in the art. The computer device includes a processor, memory, network interface, display, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs, and the internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, it implements a method for removing metal artifacts from medical images. The display of the computer device may be a liquid crystal display (LCD) or an e-ink display. The input device may be a touch layer covering the display, buttons or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0097] As another specific embodiment, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the above-described method for removing metal artifacts from medical images.

[0098] For other specific limitations regarding computer equipment, please refer to the description of the medical image metal artifact removal method above, which will not be repeated here.

[0099] As another specific embodiment, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the steps in the above-described method for removing metal artifacts from medical images.

[0100] For other specific limitations regarding computer-readable storage media, please refer to the description of methods for removing metal artifacts from medical images above, which will not be repeated here.

[0101] It will be understood by those skilled in the art that any reference to memory, storage medium, or database in the above embodiments of this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be obtained in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), synchronous link DRAM (SLDRAM), memory bus direct RAM (RDRAM), enhanced SDRAM (ESDRAM), and direct memory bus dynamic RAM (DRDRAM), etc.

[0102] Compared with existing technologies, the medical image metal artifact removal method, system, computer equipment, and storage medium provided by this invention have the following advantages:

[0103] (1) Introducing a topological homeomorphic differential manifold engine: The continuous velocity field integral mapping based on the regenerated kernel Hilbert space is adopted to ensure the positive definiteness of the Jacobian determinant of the mapping transformation, effectively avoiding structural breakage and deformation generated when the network repairs large-area artifacts, and ensuring the topological consistency of the anatomical structure.

[0104] (2) Constructing a closed-loop tensor ring and phase interference mechanism: By utilizing the closed mathematical properties of the tensor ring and adapting to the periodicity of the CT sine wave, the extraction and completion of global projection features were realized; combined with the phase interference principle of the quantum attention mechanism, the features were weighted according to the gradient direction, which improved the sharpness of the tissue edge.

[0105] (3) Combining medical physical projection constraints: Introducing orthographic projection constraints and Poisson weighting based on photon statistics at the network reconstruction end ensures the physical consistency of the results, enabling the CT density values ​​of the reconstructed voxels to have high quantitative analysis accuracy, which can meet the requirements of clinical diagnosis and treatment for data reliability.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for removing metal artifacts from medical images, characterized in that, include: The original projection data of the metal-containing medical image to be processed is obtained, and the initial reconstructed image is obtained through a filtered back projection algorithm; The initial reconstructed image is input into the topological homeomorphism mapping engine, and the differential manifold mapping is calculated by integrating the velocity vector field to obtain the topology-preserving manifold latent space features. Based on the original projection data and the physical forward projection model, a sinusoidal domain mask is extracted from the metal implant area to obtain a damaged sinusoidal sequence. The manifold latent space features and the damaged sinusoid sequence are jointly input into a dual-domain joint tensor ring network for high-dimensional tensor decomposition and core tensor iterative completion to obtain dual-domain joint features. The joint features of the two domains are constrained based on the physical consistency loss function, and the image domain is reconstructed by the inverse homeomorphism mapping operator to obtain the final medical image with metal artifacts removed.

2. The method for removing metal artifacts from medical images according to claim 1, characterized in that, The steps of acquiring the original projection data of the metal-containing medical image to be processed and obtaining the initial reconstructed image through a filtered back projection algorithm specifically include: receiving the original X-ray projection signal output by the CT scanner, performing logarithmic transformation, beam hardening correction and photon starvation noise filtering on the original projection signal, and reconstructing the initial image using the inverse Radon transform and filters.

3. The method for removing metal artifacts from medical images according to claim 1, characterized in that, In the step of inputting the initial reconstructed image into the topological homeomorphism engine, the formula for topological homeomorphism is expressed as: in, Indicates continuous time steps Next pixel coordinates Differential homeomorphic deformation field, Indicates a point in time The non-stationary velocity vector field at that location belongs to the regenerating kernel Hilbert space. , For time-integral infinitesimals; constrain the Jacobian determinant. This ensures the topological invariance of anatomical structures during the mapping process and prevents tissue boundary distortion and folding during the metal removal process.

4. The method for removing metal artifacts from medical images according to claim 1, characterized in that, In the dual-domain joint tensor ring network, the damaged sine wave sequence is reconstructed into a high-dimensional tensor. Its tensor ring decomposition formula is expressed as: ;in, This represents the value of the element at the corresponding index position in the high-dimensional tensor. The trace operation represents the trace of a matrix. Indicates the first A third-order core tensor For the first Size of the space in one dimension Let be the rank of the tensor ring, and satisfy the closed-loop boundary conditions. This decomposition is used to extract nonlocal low-rank correlation features in the data-missing regions of the sine curve trajectory.

5. The method for removing metal artifacts from medical images according to claim 1, characterized in that, The dual-domain joint tensor ring network includes a dynamic quantized attention mechanism, which calculates artifact probability amplitude by mapping classical pixel intensity to Hilbert space. The formula for the attention mechanism is as follows: in, This represents the extracted quantized eigenstate vector. For partition normalization constant, The imaginary unit, This represents the phase angle corresponding to the local pixel gradient. It is a non-linear activation function. Let be the projection weight matrix. Indicates the first In the layer network The feature vectors of the hidden nodes To calculate the ground state.

6. The method for removing metal artifacts from medical images according to claim 1, characterized in that, In the step of constraining the joint features of the two domains based on the physical consistency loss function, the physical consistency loss function... The formula is expressed as: in, This indicates a data fidelity penalty. Here is the forward projection matrix of the system based on the Beer-Lambert law. The reconstruction attenuation coefficient matrix is ​​the output of the network. Projected data for the portion without metal obstruction. It is a Poisson diagonal weighted matrix; The curl penalty term for the velocity field is used to maintain the smoothness of the manifold mapping; The nuclear norm of the core tensor; , , and These are all hyperparameters that control the weights of each item.

7. The method for removing metal artifacts from medical images according to claim 1, characterized in that, The steps for obtaining the final metal artifact-free medical image include: inputting the features with completed tensor ring completion into a multi-scale feature decoder; and during the decoding process, utilizing the deformation inverse matrix recorded by the topological homeomorphism mapping engine. Perform inverse resampling of the latent space; solve the linear inverse problem with total variation regularization term using the conjugate gradient algorithm, reconstruct and output the medical image with metal artifacts removed.

8. A medical image metal artifact removal system, characterized in that, include: The data acquisition and projection module is used to acquire the original projection data of the metal-containing medical image to be processed, and to obtain the initial reconstructed image through a filtered back projection algorithm; The topological homeomorphism mapping module is used to input the initial reconstructed image into the topological homeomorphism mapping engine, calculate the differential manifold mapping through the integral velocity vector field, and obtain the topology-preserving manifold latent space features. The sinusoidal map mask extraction module is used to extract the sinusoidal map domain mask of the metal implant area based on the original projection data and the physical forward projection model, so as to obtain the damaged sinusoidal map sequence. The tensor ring high-dimensional completion module is used to input the manifold latent space features and the damaged sinusoidal graph sequence into the dual-domain joint tensor ring network to perform high-dimensional tensor decomposition and core tensor iterative completion to obtain dual-domain joint features. The physical reconstruction output module is used to constrain the joint features of the two domains based on the physical consistency loss function, and to reconstruct the image domain through the inverse homeomorphism mapping operator to obtain the final metal artifact-free medical image.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps in the medical image metal artifact removal method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing instructions, characterized in that, When the instructions are executed on a computer, the computer performs the steps of the medical image removal method for removing metal artifacts as described in any one of claims 1 to 7.

Images