Medical image reconstruction method and device based on physical signal prior

Abstract

The invention relates to a medical image reconstruction method and device based on physical signal prior, and the method comprises the steps: building a time dictionary database based on an imaging physical model, and the time dictionary database comprises different tissue characteristic parameter combinations; performing joint decomposition on the time dictionary database and the reduced collection medical data to generate a time base, and mapping the reduced collection medical data to a subspace corresponding to the time base; sparse constraint is applied to the mapped reduced acquisition medical data, and under-sampling information in the reduced acquisition medical data is complemented through iterative optimization; utilizing a machine learning model to convert the optimized reduced acquisition medical data into a tissue parameter quantization graph; and based on the imaging physical model, inverting the tissue parameter quantization graph into a medical image, and iteratively executing the process from applying the sparse constraint to inverting into the medical image on the medical image until the obtained medical image meets a convergence condition. According to the embodiment of the invention, the quality of the generated medical image can be effectively improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technologies, and in particular, to a method and apparatus for medical image reconstruction based on physical signal priors, an electronic device, and a storage medium. Background Art

[0002] Medical imaging technologies, especially Magnetic Resonance Imaging (MRI), play an irreplaceable role in clinical diagnosis and can provide high-resolution soft tissue contrast images. However, the acquisition speed of MRI is relatively slow. Long-time scanning not only affects the comfort of patients but also may introduce motion artifacts, thereby reducing the image quality.

[0003] To shorten the acquisition time, researchers have proposed a series of accelerated reconstruction technologies. These technologies achieve acceleration by reducing the amount of acquired data, but at the same time, they also bring new challenges, such as how to recover high-quality images from undersampled data.

[0004] Existing reconstruction methods have problems of insufficient utilization of various prior constraints. Especially for magnetic resonance images, the quantitative parameter information therein is easily affected by objective factors such as magnetic field inhomogeneity with the change of time and space, resulting in difficulty in guaranteeing the quality of the recovered images. Summary of the Invention

[0005] In view of this, the present disclosure proposes a technical solution for medical image reconstruction based on physical signal priors to improve the quality of generated medical images.

[0006] According to one aspect of the present disclosure, there is provided a method for medical image reconstruction based on physical signal priors, including:

[0007] Constructing a time dictionary library based on an imaging physical model, where the time dictionary library contains different combinations of tissue characteristic parameters;

[0008] Performing joint decomposition on the time dictionary library and the downsampled medical data to generate a time basis, and mapping the downsampled medical data to a subspace corresponding to the time basis;

[0009] Applying a sparse constraint to the mapped downsampled medical data, and complementing the undersampled information in the downsampled medical data through iterative optimization;

[0010] Using a machine learning model to convert the optimized downsampled medical data into a tissue parameter quantization map;

[0011] Based on the imaging physical model, inverting the tissue parameter quantization map into a medical image, and iteratively performing the process of applying the sparse constraint to inverting into a medical image on the medical image until the obtained medical image meets the convergence condition.

[0012] In a possible implementation, constructing the time dictionary library based on the imaging physical model includes:

[0013] Exhaust all parameter combinations based on the quantization range and quantization step of the preset tissue characteristic parameters, and calculate the weighted signals corresponding to each parameter combination through the imaging physical model;

[0014] Arrange the parameter combinations and the corresponding weighted signals in a time series to form a time-signal mapping matrix, thereby forming the time dictionary library.

[0015] In a possible implementation, the medical image is a magnetic resonance image, and the tissue characteristic parameters of the time dictionary library include: T1 value, T2 value, pulse efficiency coefficient α and pulse flip angle θ after B1 field inhomogeneity correction.

[0016] In a possible implementation, jointly decomposing the time dictionary library and the downsampled medical data to generate a time basis includes:

[0017] Perform singular value decomposition on the time dictionary library to extract the first time basis;

[0018] Perform singular value decomposition on the downsampled medical data to extract the second time basis;

[0019] Combine the first time basis and the second time basis in a preset ratio to form a combined time basis.

[0020] In a possible implementation, imposing a sparse constraint on the mapped downsampled medical data and complementing the undersampled information in the downsampled medical data through iterative optimization includes:

[0021] Impose a sparse constraint norm on the mapped downsampled medical data to obtain a sparse matrix after sparse constraint;

[0022] Using the sparse matrix as the initial estimate, iteratively fill the missing part in the matrix through the conjugate gradient method to gradually obtain a complete and clear data matrix.

[0023] In a possible implementation, the machine learning model is a neural network, and using the machine learning model to convert the optimized downsampled medical data into a tissue parameter quantization map includes;

[0024] Input the optimized downsampled medical data into the network to obtain the output T1 and T2 parameter quantization maps.

[0025] In a possible implementation, based on the imaging physical model, inverting the tissue parameter quantization map into a medical image includes:

[0026] Based on the tissue parameter quantization map, calculate the magnetic resonance signal intensity of each pixel through the imaging physical model;

[0027] Map the magnetic resonance signal intensity to the spatial domain through Fourier transform to obtain a medical image.

[0028] According to one aspect of the present disclosure, there is provided a medical image reconstruction device based on physical signal prior, including:

[0029] A dictionary library construction module, configured to construct a time dictionary library based on an imaging physical model, where the time dictionary library contains different combinations of tissue characteristic parameters;

[0030] A mapping module, configured to perform joint decomposition on the time dictionary library and the downsampled medical data to generate a time basis, and map the downsampled medical data to the subspace corresponding to the time basis;

[0031] A sparse constraint module, configured to impose a sparse constraint on the mapped downsampled medical data, and iteratively optimize to complete the undersampled information in the downsampled medical data;

[0032] A machine learning module, configured to use a machine learning model to convert the optimized downsampled medical data into a tissue parameter quantization map;

[0033] An inversion module, configured to invert the tissue parameter quantization map into a medical image based on the imaging physical model; iteratively execute the process of imposing the sparse constraint to inverting into a medical image on the medical image until the obtained medical image meets the convergence condition.

[0034] According to another aspect of the present disclosure, there is provided an electronic device, including a memory, a processor, and a computer program stored on the memory, where the processor executes the computer program to implement the steps of the above method.

[0035] According to another aspect of the present disclosure, there is provided a non-volatile computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0036] According to another aspect of the present disclosure, there is provided a computer program product, including a computer program, or a non-volatile computer-readable storage medium carrying the computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0037] In the embodiments of the present disclosure, through multi-step collaborative optimization, the efficiency and accuracy of medical image reconstruction can be significantly improved.

[0038] First, before the iteration starts, by combining the time dictionary library with the time base of the downsampled medical data and using the physical model to constrain the time information, the theoretical prior and actual dynamic information are fused, correcting the aliasing artifacts and noise caused by downsampling, avoiding the cumulative amplification of errors in the iterative optimization, and reducing the propagation of magnetic field inhomogeneity (such as B1 field error) or downsampling artifacts in the subsequent processes. It not only constrains the rationality of the time dimension information, but also simplifies the data complexity through subspace mapping, providing a high-fidelity low-dimensional representation basis for subsequent optimization.

[0039] Secondly, applying sparse constraints to the mapped downsampled medical data and complementing the undersampled information through iterative optimization can accurately restore high-frequency details and suppress noise and artifacts. The sparse constraint enhances the saliency of key anatomical structures (such as cerebral sulci and gyri, blood vessel edges) by restricting the sum of the absolute values of the data in a specific transform domain, while the iterative optimization of the conjugate gradient method gradually corrects the missing values in the undersampled area, ensuring that the reconstruction result not only meets the fidelity requirements of the actual acquired data but also has high spatial resolution. This process significantly improves the clarity and structural integrity of the image, providing a more reliable visual basis for clinical diagnosis.

[0040] Furthermore, using a machine learning model to convert the optimized downsampled medical data into a tissue parameter quantification map realizes the efficient mapping from low-dimensional subspace data to tissue characteristic parameters (such as T1 and T2 values). The machine learning model quickly extracts the deep features in the data through its end-to-end learning ability, generating a pixel-level quantification parameter distribution map, avoiding the time-consuming operation of traditional point-by-point fitting. At the same time, based on the imaging physical model, the quantification map is inverted into a medical image, and through the iterative execution of the sparse constraint to the inversion process, the physical consistency between the parameters and the image signal is ensured. This closed-loop optimization mechanism not only suppresses the possible overfitting problem of deep learning but also finally outputs a high-quality image that conforms to the magnetic resonance signal equation by repeatedly correcting the parameter errors. Taking the physical model as a hard constraint, forcing the machine learning model to output within a theoretically reasonable range, avoiding unreasonable values output by the deep learning black box. That is to say, the physical model provides clear constraint conditions, making it not easy for the deep learning model to overfit during the training and prediction processes.

[0041] In the embodiments of the present disclosure, time information, spatial information, deep learning prior, and low-rank sparse prior are naturally integrated into a unified framework to form a strongly constrained reconstruction process. In addition, overall, this patent synchronously completes image reconstruction and parameter quantification in a single process through the deep integration of the physical model and data-driven methods, greatly shortening the time-consuming of the traditional two-step method. At the same time, through the joint constraint and iterative optimization in the spatio-temporal dimension, it takes into account the reconstruction speed, image quality, and parameter accuracy, providing an efficient and reliable technical support for the rapid and accurate diagnosis of clinical medical images.

[0042] Other features and aspects of the present disclosure will become apparent from the following detailed description of the exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which are included in and constitute a part of this specification, illustrate exemplary embodiments, features, and aspects of the present disclosure and are used to explain the principles of the present disclosure together with the specification.

[0044] Figure 1 A flowchart showing a medical image reconstruction method based on a prior physical signal according to an embodiment of the present disclosure.

[0045] Figure 2 A block diagram showing a medical image reconstruction device based on a prior physical signal according to an embodiment of the present disclosure.

[0046] Figure 3 A block diagram of a device 1900 for medical image reconstruction based on a prior physical signal shown according to an exemplary embodiment. DETAILED DESCRIPTION

[0047] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0048] As used herein, the terms "comprising," "including," "having," or variations thereof are open-ended and include one or more stated features, wholes, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, wholes, elements, steps, components, functions, or groups thereof.

[0049] When an element is referred to as being "connected," "coupled," "responsive," or variations thereof to another element, it can be directly connected, coupled, or responsive to the other element, or intervening elements may be present.

[0050] Although the terms first, second, third, etc. may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another. Thus, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments.

[0051] As used herein, the term "exemplary" means "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" should not necessarily be construed as being superior to or better than other embodiments.

[0052] In addition, for a better illustration of the present disclosure, numerous specific details are provided in the following detailed implementation manners. Those skilled in the art should understand that the present disclosure can still be implemented without certain specific details. In some instances, methods, means, elements, and circuits well-known to those skilled in the art are not described in detail to highlight the gist of the present disclosure.

[0053] Medical imaging technology, especially magnetic resonance imaging (MRI), has irreplaceable value in disease diagnosis and treatment planning. However, the wide application of MRI is limited by its slow acquisition speed - traditional methods need to strictly follow the Shannon sampling theorem, resulting in too long examination time. Patients are prone to motion artifacts due to body movement, and the contradiction between the demand for high-resolution imaging and clinical efficiency is becoming increasingly prominent. Although related technologies (such as compressive sensing, deep learning) attempt to accelerate data acquisition through undersampling, they often face problems such as insufficient reconstruction quality, loss of details, or poor physical consistency. Especially when quantifying tissue characteristic parameters (such as T1 / T2 relaxation time), error accumulation may directly affect the diagnostic accuracy.

[0054] This solution aims at the above pain points and proposes a physical model-driven medical image joint reconstruction method, which deeply integrates the physical laws of magnetic resonance imaging, low-rank sparse constraints, and deep learning technology. Its core is to embed the prior knowledge of tissue characteristic parameters (T1 / T2) and device errors (such as B1 field inhomogeneity) through a time dictionary library, and combine iterative optimization and end-to-end learning to achieve rapid reconstruction from undersampled data to high-precision images and parameter mappings. This solution is especially suitable for rapid imaging of fine anatomical structures such as the brain and heart, as well as quantitative parameter mapping (such as tumor tissue characteristic analysis). While ensuring image quality, it shortens the reconstruction time to a clinically acceptable range, providing efficient and reliable technical support for precision medicine.

[0055] Figure 1 The flowchart of a medical image reconstruction method based on physical signal prior according to an embodiment of the present disclosure is shown. As Figure 1 shown, the method includes:

[0056] In step S11, a time dictionary library is constructed based on the imaging physical model, and the time dictionary library contains different combinations of tissue characteristic parameters.

[0057] The imaging physical model is used to describe the physical laws of signal generation, acquisition, and reconstruction in medical imaging devices (such as MRI), providing a theoretical basis for image reconstruction. For the magnetic resonance imaging physical model, the imaging physical model can specifically include a magnetic resonance signal expression, which contains pulse sequence parameters (repetition time TR, echo time TE, flip angle θ), B1 field inhomogeneity compensation and other parameters, and is used to map tissue characteristic parameters (T1 / T2) into image signals.

[0058] For the specific imaging physical model, reference can be made to the possible implementation manners provided in this disclosure, which will not be elaborated here.

[0059] The time dictionary library is a signal response library generated by exhaustively listing combinations of tissue characteristic parameters. The tissue characteristic parameters are key parameters that describe the response characteristics of tissues in a magnetic field, and these parameters determine the signal intensity and contrast of tissues in medical images. In a possible implementation manner, the medical image is a magnetic resonance image, and the tissue characteristic parameters of the time dictionary library include: T1 value, T2 value, pulse efficiency coefficient α after B1 field inhomogeneity correction, and pulse flip angle θ.

[0060] T1 is the longitudinal relaxation time. The shorter the T1 value, the faster the tissue recovers longitudinally. T2 is the transverse relaxation time. The shorter the T2 value, the faster the transverse magnetization vector decays. The pulse efficiency coefficient α is a parameter for correcting B1 field inhomogeneity. The pulse flip angle θ is also a parameter for correcting B1 field inhomogeneity.

[0061] In a possible implementation manner, constructing the time dictionary library based on the imaging physical model includes: based on the quantization range and quantization step size of preset tissue characteristic parameters, exhaustively listing all parameter combinations, and calculating the weighted signals corresponding to each parameter combination through the imaging physical model; arranging the parameter combinations and the corresponding weighted signals in a time series to form a time-signal mapping matrix, thereby forming the time dictionary library.

[0062] The value ranges and step sizes of the tissue characteristic parameters (such as T1 value, T2 value, pulse efficiency coefficient α after B1 field inhomogeneity correction, and pulse flip angle θ) can be determined first. The ranges and step sizes of these parameters determine the coverage and accuracy of the time dictionary library.

[0063] By exhaustively listing all possible parameter combinations, a comprehensive parameter set can be generated, ensuring that the time dictionary library can cover all possible tissue characteristics. Through the fine division of the parameter ranges, it is ensured that the dictionary library can accurately reflect the signal characteristics of different tissue types and compensate for equipment hardware errors (such as magnetic field inhomogeneity).

[0064] Then, using the imaging physical model, calculate the weighted signals corresponding to each parameter combination. These signals reflect the signal intensities of different tissues at specific time points.

[0065] Exemplarily, in the case where the medical image is a magnetic resonance image, ignoring the decay effect of T2* (the transverse relaxation time considering the influence of magnetic field inhomogeneity), the longitudinal magnetization vector can be used as the measured signal M θ , in a possible implementation manner, the magnetic resonance formula for generating weighted signals based on tissue characteristic parameters in the imaging physical model is as follows:

[0066]

[0067] Among them, M0 is the initial magnetization intensity, representing the longitudinal magnetization vector intensity in the fully relaxed state (when there is no RF pulse interference), and M before T2IR is the longitudinal magnetization vector before the T2IR is applied in the magnetic resonance sequence. θ is the flip angle of the RF pulse, representing the angle at which the magnetization vector flips from the longitudinal direction to the transverse direction. TR is the time for repeating the θ pulse, and α is the pulse efficiency coefficient used to correct the flip angle deviation caused by the B1 magnetic field inhomogeneity. k is the repetition number index of the RF pulse sequence, used to quantify the dynamic change of the magnetization intensity under the action of multiple pulses.

[0068] E1 = exp(-TR / T1), E2 = exp(-TE prep / T2), E gap = exp(-T gap / T1),

[0069] Among them, TR is the repetition time, that is, the time interval between two consecutive RF pulses. T1 is the longitudinal relaxation time, describing the time constant required for the longitudinal magnetization to recover to the equilibrium state. T2 is the transverse relaxation time, describing the time constant for the transverse magnetization to decay due to spin-spin interaction. TE prep is the Echo Time Preparation, which is the time interval between the T2 preparation pulse and the subsequent IR pulse in the magnetic resonance sequence. T gap is the time interval, referring to the time interval between the T2 preparation and IR pulses (T2IR) and the subsequent standard spoiled gradient-echo (SPGR) sequence after TE prep is used to ensure the stability and accuracy of the signal.

[0070] This model utilizes adiabatic T2 preparation and IR pulses (T2 preparation and IR pulses, T2IR) to generate the contrast of T1 and T2. The duration of TE prep is different in each turbo field-echo (TFE) excitation. After a short time (T gap ), a series of SPGR sequences apply a series of RF pulses with a small flip angle (θ).

[0071] Substituting each parameter combination into the above formula, the corresponding weighted signal can be obtained. Exemplarily, assuming the parameter combination: T1 = 1000ms, T2 = 100ms, α = 0.95, θ = 0.11866 radians, and the pulse sequence parameters TR = 500ms, TE = 10ms. Substituting into the above formula, the corresponding weighted signal can be obtained.

[0072] These signals reflect the signal intensities of different tissues at specific time points., generating signal responses through a physical model to avoid biases caused by empirical assumptions and ensure the physical consistency of the dictionary library.

[0073] Then, arranging each parameter combination and its corresponding weighted signal in a time series forms a matrix. For example, each row of this matrix can represent a parameter combination, and each column can represent the signal intensity at a time point.

[0074] Specific example: Generate a matrix where the rows represent different parameter combinations and the columns represent the signal intensities at different time points. For example:

[0075] Row 1: T1 = 10, T2 = 10, α = 0.9, θ = 0.11866, and the corresponding signal intensity sequence.

[0076] Row 2: T1 = 10, T2 = 10, α = 0.91, θ = 0.11866, and the corresponding signal intensity sequence.

[0077] And so on.

[0078] The finally formed time dictionary library stores all possible parameter combinations and their corresponding signal characteristics, providing a basis for subsequent image reconstruction.

[0079] In the embodiments of the present disclosure, by presetting a wide range of parameter values and step sizes and exhausting all combinations, the time dictionary library can comprehensively cover the characteristics of different tissues. Then, during the image reconstruction process, the parameter combination that best matches the actually acquired data can be found. Moreover, using the weighted signals calculated by the imaging physical model ensures the consistency between the signal characteristics in the time dictionary library and the actual physical process. This helps to reduce artifacts and errors during the reconstruction process and improve the image quality. With the constraint of the physical model and the assistance of the time dictionary library, the entire reconstruction process is more efficient. It reduces the cumbersome iterative optimization steps in traditional methods, shortens the reconstruction time, and the wide coverage of the time dictionary library and the constraint of the physical model enable this method to maintain good performance under different downsampling rates and noise levels, having wide applicability and robustness.

[0080] In step S12, jointly decompose the time dictionary library and the downsampled medical data to generate a time basis, and map the downsampled medical data to the subspace corresponding to the time basis.

[0081] Low-sampling medical data refers to medical image data with a sampling rate lower than the standard requirement, and the complete information needs to be restored through a reconstruction algorithm. Here, a time dictionary library can be used in combination with the low-sampling medical data to generate a low-dimensional feature basis to constrain the time-dimensional information of the data.

[0082] Specifically, the dictionary library and the low-sampling data can be respectively subjected to singular value decomposition (SVD) to obtain their respective basis vectors, and then the basis vectors are combined, and the low-sampling data is mapped to the low-dimensional space corresponding to the time basis. Specifically, the low-sampling data can be right-multiplied by the transpose of the right singular matrix V to map the low-sampling medical data to the low-dimensional space corresponding to the time basis.

[0083] In a possible implementation, the joint decomposition of the time dictionary library and the low-sampling medical data to generate a time basis includes: performing singular value decomposition on the time dictionary library to extract a first time basis; performing singular value decomposition on the low-sampling medical data to extract a second time basis; combining the first time basis and the second time basis in a preset ratio to form a combined time basis.

[0084] Singular value decomposition (SVD) is a matrix decomposition technique used to extract the main features of a matrix. Performing SVD on the time dictionary library can extract a time basis that can represent the main features of the dictionary library. For the convenience of description, it is called the first time basis here.

[0085] Similarly, SVD is used to decompose the low-sampling medical data to extract a time basis that can represent the main features of the actual collected data. For the convenience of description, it is called the second time basis here.

[0086] Then, the time bases extracted from the time dictionary library and the low-sampling medical data are combined in a certain ratio to form the final combined time basis. This combination process can be determined through experimental optimization to improve the accuracy of the time basis.

[0087] In the embodiments of the present disclosure, through singular value decomposition (SVD), the main features of the time dictionary library and the downsampled medical data are extracted, reducing the data dimension and improving the calculation efficiency. Since SVD is utilized, a low-rank constraint can be naturally introduced. By retaining the eigenvectors corresponding to the main singular values, the redundant information of the data is further compressed while the main structural information in the data is retained. The low-rank constraint not only helps to remove noise and artifacts but also enhances the physical consistency of the data, ensuring that the reconstruction result conforms to the actual physical laws. In addition, the time basis of the downsampled medical data can correct the possible errors in the time dictionary library, improving the accuracy of the time basis. By combining the first time basis and the second time basis, while retaining the physical laws, it is possible to flexibly correct device errors and individual differences, enhancing the spatial resolution and time consistency of the reconstructed image, being able to better adapt to different downsampling rates and noise levels, and improving the robustness of the method.

[0088] In step S13, a sparse constraint is imposed on the mapped downsampled medical data, and the undersampled information in the downsampled medical data is complemented through iterative optimization.

[0089] A sparse constraint can be imposed on the mapped downsampled medical data. By utilizing the sparse characteristics of the image, the downsampled medical data is constrained to improve the reconstruction quality. For example, an L1 norm constraint can be imposed on the mapped downsampled data to make it sparser. Specifically, an objective function containing the L1 norm can be constructed, and the conjugate gradient method is used to iteratively correct the values in the missing region.

[0090] In a possible implementation manner, imposing a sparse constraint on the mapped downsampled medical data and complementing the undersampled information in the downsampled medical data through iterative optimization includes: imposing a sparse constraint norm on the mapped downsampled medical data to obtain a sparse matrix after sparse constraint; using the sparse matrix as the initial estimate, and iteratively filling the missing part in the matrix through the conjugate gradient method to gradually obtain a complete and clear data matrix.

[0091] The sparse constraint norm utilizes the sparse characteristics of the signal, that is, the signal has only a few non-zero coefficients in a certain transform domain. Imposing a sparse constraint norm on the downsampled medical data after subspace mapping, that is, minimizing the sum of its absolute values, makes the data sparse in a specific transform domain (such as the gradient domain or the wavelet domain), suppressing noise and unstructured artifacts. Therefore, a sparse constraint norm can be imposed on the mapped downsampled medical data to obtain a sparse matrix after sparse constraint.

[0092] Among them, the sparse constraint norm makes the downsampled medical data exhibit sparsity by minimizing the energy or the number of coefficients of the data in a specific transform domain, thereby enhancing key features and suppressing noise and artifacts. In one example, the sparse constraint norm provided by the present disclosure may be an L1 norm constraint. By minimizing the sum of the absolute values (i.e., the L1 norm) of the data in the transform domain, the data exhibits sparsity in this domain. The L1 norm constraint is widely used in sparse reconstruction, especially in the gradient domain or the wavelet domain, and can effectively retain the edge information of the image while suppressing high-frequency noise. In another example, the sparse constraint norm provided by the present disclosure may also be a total variation norm constraint (Total Variation, TV). By minimizing the L1 norm of the image gradient, the image exhibits sparsity in the gradient domain. The total variation norm constraint is particularly suitable for retaining the edge structure of the image while smoothing the noise in non-edge regions and is widely used in medical image reconstruction and denoising tasks. In another example, the sparse constraint norm provided by the present disclosure may also be a wavelet domain constraint. By minimizing the coefficient energy after wavelet transform, the data exhibits sparsity in the wavelet domain. The wavelet domain constraint can effectively extract the multi-scale features of the image, enhance high-frequency details (such as blood vessels, sulci, etc.), and at the same time suppress noise and artifacts.

[0093] Then, taking the sparse matrix as the initial estimate, an optimization objective function including a data fidelity term and a sparse constraint term is constructed, and the missing part in the matrix is filled iteratively by the conjugate gradient method to minimize the objective function. The conjugate gradient method is an iterative optimization algorithm used to solve large-scale linear systems. Through multiple iterations, the missing part in the matrix is gradually filled, and finally a complete and clear data matrix is obtained.

[0094] The optimized data matrix is restored to the high-dimensional space through the time-base mapping to obtain a complete k-space data matrix, and then a medical image is generated through Fourier transform.

[0095] In the embodiment of the present disclosure, the sparsity of the downsampled medical data is enhanced by the sparse constraint norm, key anatomical structures (such as organ edges) are retained, and at the same time high-frequency noise (such as radial artifacts) is suppressed. Through iterative optimization by the conjugate gradient method, the error in the matrix is gradually reduced, the accuracy of the data is improved, the missing part in the matrix is gradually filled, and finally a complete and clear data matrix is obtained. By combining the sparse constraint and the conjugate gradient method, the calculation efficiency is improved and the reconstruction time is shortened.

[0096] In step S14, the optimized downsampled medical data is converted into a tissue parameter quantization map by using a machine learning model;

[0097] The tissue parameter quantification map is an image that characterizes the spatial distribution of biological tissue characteristic parameters (such as T1 and T2 relaxation times). The optimized downsampled medical data can be converted into a tissue parameter quantification map using a trained machine learning model.

[0098] In a possible implementation, the machine learning model is a neural network. The process of using the machine learning model to convert the optimized downsampled medical data into a tissue parameter quantification map includes: inputting the optimized downsampled medical data into the network to obtain the output T1 and T2 parameter quantification maps.

[0099] In this implementation, the machine learning model is a neural network. Using a neural network as the machine learning model can learn the complex mapping relationship between the downsampled data and the tissue parameter quantification map. Exemplarily, the neural network can be a DenseAttention Unet, which combines dense connections and attention mechanisms and can effectively extract image features and enhance detail information.

[0100] The input of the neural network is the optimized downsampled medical data in step S13, that is, the downsampled medical data after sparse constraint and iterative optimization is used as the input, and the output is the T1 and T2 parameter quantification maps.

[0101] The neural network converts the input downsampled medical data into T1 and T2 parameter quantification maps by learning the features in the training dataset. The two quantification maps respectively represent the quantification values of the T1 and T2 parameters, and each pixel point corresponds to specific T1 and T2 values.

[0102] In the embodiments of the present disclosure, the neural network can learn the complex mapping relationship between the downsampled data and the tissue parameters, improving the accuracy of the conversion. Moreover, it can enhance image details, making the T1 and T2 quantification maps clearer. In addition, the neural network has a certain robustness to noise and artifacts and can maintain good performance under different downsampling rates and noise levels. The fast inference ability of the neural network significantly shortens the reconstruction time, meeting the real-time requirements of clinical applications.

[0103] In step S15, based on the imaging physical model, the tissue parameter quantification map is inverted into a medical image, and the process of applying sparse constraints to the inversion into a medical image is iteratively performed on the medical image until the obtained medical image meets the convergence condition.

[0104] The tissue parameter quantification map (such as the T1 and T2 value distribution map) can be converted into a medical image through the imaging physical model. Exemplarily, using the above magnetic resonance formula, based on the tissue characteristic parameters (T1, T2), the theoretical weighted signal can be generated, and further the medical image can be obtained.

[0105] After obtaining the medical image, the process of applying the sparse constraint to the inversion of the medical image can be iteratively performed on the medical image, that is, steps S13 - S15 are iteratively performed until the obtained medical image meets the convergence condition. The convergence condition can be that an image quality metric (such as SSIM) reaches a certain threshold, or the number of iterations reaches a preset maximum value, or the difference between the medical images of two iterations is less than a set difference threshold.

[0106] If the convergence condition is not met, the medical image is continued to be used as the updated downsampled medical data, a sparse constraint is applied to it, and the undersampled information in the downsampled medical data is complemented through iterative optimization. Then, the optimized downsampled medical data is converted into a tissue parameter quantification map using a machine learning model, and then based on the imaging physical model, the tissue parameter quantification map is inverted into a medical image.

[0107] In the embodiments of the present disclosure, through multi - step collaborative optimization, the efficiency and accuracy of medical image reconstruction can be significantly improved.

[0108] First, before the iteration starts, by combining the time dictionary library with the time base of the downsampled medical data and using the physical model to constrain the time information, the theoretical prior and actual dynamic information are fused, the aliasing artifacts and noise caused by downsampling are corrected, the accumulation and amplification of errors in the iterative optimization are avoided, and the propagation of magnetic field inhomogeneity (such as B1 field error) or downsampling artifacts in subsequent processes is reduced. It not only constrains the rationality of the time - dimension information, but also simplifies the data complexity through subspace mapping, providing a high - fidelity low - dimensional representation basis for subsequent optimization.

[0109] Second, applying a sparse constraint to the mapped downsampled medical data and complementing the undersampled information through iterative optimization can accurately restore high - frequency details and suppress noise and artifacts. The sparse constraint enhances the saliency of key anatomical structures (such as cerebral sulci and gyri, blood vessel edges) by restricting the sum of the absolute values of the data in a specific transform domain, and the iterative optimization of the conjugate gradient method gradually corrects the missing values in the undersampled area, ensuring that the reconstruction result not only meets the fidelity requirements of the actual acquired data but also has a high spatial resolution. This process significantly improves the clarity and structural integrity of the image, providing a more reliable visual basis for clinical diagnosis.

[0110] Furthermore, a machine learning model is used to convert the optimized downsampled medical data into a tissue parameter quantification map, achieving an efficient mapping from low-dimensional subspace data to tissue characteristic parameters (such as T1 and T2 values). Through its end-to-end learning ability, the machine learning model quickly extracts deep features in the data and generates a pixel-level quantification parameter distribution map, avoiding the time-consuming operation of traditional point-by-point fitting. At the same time, based on the imaging physical model, the quantification map is inverted into a medical image, and by iteratively performing sparse constraints on the inversion process, the physical consistency between the parameters and the image signal is ensured. This closed-loop optimization mechanism not only suppresses the possible overfitting problem in deep learning but also finally outputs high-quality images that conform to the magnetic resonance signal equation by repeatedly correcting parameter errors. Taking the physical model as a hard constraint, the machine learning model is forced to output within a theoretically reasonable range, avoiding unreasonable values output by the deep learning black box. That is to say, the physical model provides clear constraint conditions, making it difficult for the deep learning model to overfit during the training and prediction processes.

[0111] In the embodiments of the present disclosure, time information, spatial information, deep learning priors, and low-rank sparse priors are naturally integrated into a unified framework to form a strongly constrained reconstruction process. Overall, through the deep integration of the physical model and data-driven methods, this patent synchronously completes image reconstruction and parameter quantification in a single process, significantly shortening the time consumption of the traditional two-step method. At the same time, through joint constraints and iterative optimization in the spatio-temporal dimension, it takes into account the reconstruction speed, image quality, and parameter accuracy, providing efficient and reliable technical support for the rapid and accurate diagnosis of clinical medical images.

[0112] In a possible implementation manner, the inverting the tissue parameter quantification map into a medical image based on the imaging physical model includes: calculating the magnetic resonance signal intensity of each pixel through the imaging physical model based on the tissue parameter quantification map; and mapping the magnetic resonance signal intensity to the spatial domain through Fourier transform to obtain a medical image.

[0113] Specific parameters in the tissue parameter quantification map can be input into the imaging physical model. For example, when the tissue parameter quantification map is a distribution map of T1 and T2 values, the T1 and T2 values and preset device parameters (TR, TE, θ, α) can be input into the above magnetic resonance formula to obtain the magnetic resonance signal intensity of the output pixel. Then, the calculated magnetic resonance signal intensity is mapped from the k-space (spatial frequency domain) to the spatial domain through Fourier transform to obtain a medical image.

[0114] In the embodiments of the present disclosure, the tissue parameter quantization map is inverted into a medical image, significantly improving the physical consistency and image quality of the reconstruction result. Based on the tissue parameter quantization map (such as the T1 and T2 value distributions), the magnetic resonance signal intensity of each pixel is calculated through an imaging physical model, directly establishing a theoretical correlation between the tissue characteristic parameters and the signal intensity. The parameters such as T1 and T2 are combined with the device parameters (such as the repetition time TR and the echo time TE) using the magnetic resonance signal equation to ensure that the signal intensity of each pixel strictly conforms to the physical laws, avoiding parameter distortion or non-physiological output (such as a negative T1 value) caused by data deviation in the deep learning model.

[0115] Secondly, the calculated magnetic resonance signal intensity is mapped to the spatial domain through Fourier transform to generate a medical image, realizing an efficient conversion from the parameter space to the image space. Fourier transform converts the theoretical signal from the frequency domain (k-space) to the spatial domain (image domain), simulating the imaging process of a real magnetic resonance device. This step not only preserves the high-frequency details in the signal (such as cerebral sulci and gyri, and microvascular structures), but also forces the image to meet the data fidelity requirements through the embedding of the physical model.

[0116] By integrating the parameter inversion and Fourier transform into an iterative process, an image is generated based on the updated parameters in each iteration, and the parameter values are dynamically adjusted by comparing with the actually acquired data. This closed-loop design not only suppresses the gradual accumulation of errors in the traditional step-by-step method, but also ensures that the direction of each optimization conforms to the theoretical expectation through the hard constraint of the physical model. For example, if the T1 value output by the deep learning model deviates from the physiological range in a certain iteration, the physical inversion will automatically correct the abnormal value through the signal equation and feedback it to the next round of optimization, finally generating a high-quality image that conforms to both the data-driven characteristics and physical consistency. This process improves the reconstruction accuracy while significantly shortening the additional time consumption caused by the separation of reconstruction and parameter fitting in the traditional method, providing a reliable technical basis for clinical real-time imaging.

[0117] The following will describe in detail the specific implementation manner of the method provided by the present disclosure in the field of magnetic resonance.

[0118] First, based on the preset quantization step sizes of the T1 and T2 relaxation time ranges, the pulse efficiency coefficient α, and the pulse flip angle θ, all parameter combinations are exhaustively enumerated to construct a time dictionary library, and the weighted signals corresponding to each combination are calculated through the magnetic resonance signal equation to form a time-signal mapping matrix. This step provides a high-fidelity prior database covering clinical tissue characteristics and device errors for the subsequent process through physically driven signal simulation.

[0119] Subsequently, joint singular value decomposition (SVD) is performed on the time dictionary library and the actually acquired downsampled k-space data. The first time basis representing the theoretical signal pattern and the second time basis reflecting the actual dynamic characteristics are respectively extracted, and a low-dimensional time basis is generated by combining them according to a preset ratio. For example, the first 5 basis vectors from the downsampled data (capturing respiratory motion artifacts) are concatenated with the last 3 basis vectors from the dictionary library (characterizing the T1 / T2 relaxation law) to form an 8-dimensional combined time basis. The downsampled data is mapped to the low-dimensional subspace by right-multiplying the right singular matrix of the combined time basis, significantly reducing the data dimension and suppressing noise interference. For example, the original 256×256 k-space matrix is compressed into a 256×8 subspace representation, laying an efficient computational foundation for subsequent optimization.

[0120] In the low-dimensional subspace, a sparse constraint based on the L1 norm is imposed on the mapped data, and the undersampled information is iteratively completed by the conjugate gradient method. The sparse constraint forces the image to be sparse in the transform domain by restricting the sum of the absolute values of the wavelet transform coefficients. For example, it suppresses Gaussian noise and unstructured artifacts in brain images while preserving the edge details of gray and white matter. During the iterative optimization process, each update combines the data fidelity error and the sparsity evaluation to gradually correct the high-frequency signals in the missing area of the k-space.

[0121] The optimized low-dimensional data is converted into a tissue parameter quantification map through a machine learning model (such as DenseAttention UNet), realizing an end-to-end mapping from the subspace to the T1 / T2 value distribution. The network input is 8-dimensional subspace data, and the output is a 256×256 T1 / T2 quantification map, avoiding the time-consuming operation of traditional pixel-by-pixel fitting. Subsequently, based on the magnetic resonance signal equation, the quantification map is inverted into the magnetic resonance weighted signal, and the spatial domain image is generated through Fourier transform. For example, substituting the gray matter T1 = 1000 ms and T2 = 100 ms into the signal equation to calculate the pixel intensity, and then obtaining the T1-weighted image through inverse Fourier transform to ensure that the image contrast strictly conforms to the physical law.

[0122] Finally, the process of performing sparse constraint until inversion into an image is iteratively executed until the convergence condition is met. This closed-loop mechanism can quickly reach convergence, outputting a magnetic resonance image with high resolution, low noise, and accurate parameters. The total time consumption is reduced by more than 90% compared with the traditional method, providing reliable support for clinical real-time diagnosis.

[0123] Next, the comparison process between the present disclosure and the prior art will be described.

[0124] First, an evaluation dataset is constructed based on a fine digital brain phantom, and reference T1 and T2 quantization values are precisely assigned to different brain tissue components (such as gray matter, white matter, cerebrospinal fluid). For example, for gray matter, T1 = 1000 ms and T2 = 100 ms; for white matter, T1 = 800 ms and T2 = 80 ms; for cerebrospinal fluid, T1 = 4000 ms and T2 = 2000 ms. The theoretical signal responses of each tissue are simulated through the magnetic resonance signal equation, and three-dimensional golden radial angle sampling strategy is adopted to generate k-space data. This sampling method covers the k-space through non-uniformly distributed radial trajectories, which can efficiently capture high-frequency information and reduce aliasing artifacts. To approximate the real scanning environment, Gaussian noise with a mean of zero and a standard deviation of 5% is added to the simulated k-space data to simulate the thermal noise and motion artifacts in the actual device. The generated k-space data is a complex-valued frequency-domain signal, which needs to be transformed by Fourier transform into image-domain data as a reference standard.

[0125] It should be noted that although the method for medical image reconstruction based on physical signal prior is introduced above by taking magnetic resonance images as an example, those skilled in the art can understand that the present disclosure should not be limited thereto. In fact, users can flexibly set medical images according to their personal preferences and / or actual application scenarios.

[0126] In addition, the present disclosure also provides an apparatus, an electronic device, a computer-readable storage medium, and a program corresponding to the method. The above can all be used to implement any method provided by the present disclosure. For the corresponding technical solutions and descriptions, please refer to the corresponding records in the method section, which will not be elaborated here.

[0127] Figure 2 The block diagram of a medical image reconstruction apparatus based on physical signal prior according to an embodiment of the present disclosure is shown as Figure 2 shown, the apparatus 20 includes:

[0128] A dictionary library construction module 21, configured to construct a time dictionary library based on an imaging physical model, where the time dictionary library contains different combinations of tissue characteristic parameters;

[0129] A mapping module 22, configured to perform joint decomposition on the time dictionary library and the downsampled medical data to generate a time basis, and map the downsampled medical data to the subspace corresponding to the time basis;

[0130] A sparse constraint module 23, configured to impose a sparse constraint on the mapped downsampled medical data, and iteratively optimize to complete the undersampled information in the downsampled medical data;

[0131] A machine learning module 24, configured to use a machine learning model to convert the optimized downsampled medical data into a tissue parameter quantization map;

[0132] An inversion module 25, configured to invert the tissue parameter quantization map into a medical image based on the imaging physical model; iteratively perform the above process of applying sparse constraints to inverting into a medical image on the medical image until the obtained medical image meets the convergence condition.

[0133] In a possible implementation manner, the dictionary library construction module 21 is configured to:

[0134] Exhaust all parameter combinations based on a preset quantization range and quantization step size of tissue characteristic parameters, and calculate the weighted signals corresponding to each parameter combination through the imaging physical model;

[0135] Arrange the parameter combinations and the corresponding weighted signals in a time series to form a time-signal mapping matrix, thereby forming a time dictionary library.

[0136] In a possible implementation manner, the medical image is a magnetic resonance image, and the tissue characteristic parameters of the time dictionary library include: T1 value, T2 value, pulse efficiency coefficient α after B1 field inhomogeneity correction, and pulse flip angle θ.

[0137] In a possible implementation manner, the mapping module is configured to:

[0138] Perform singular value decomposition on the time dictionary library, and extract the first time basis;

[0139] Perform singular value decomposition on the downsampled medical data, and extract the second time basis;

[0140] Combine the first time basis and the second time basis according to a preset ratio to form a combined time basis.

[0141] In a possible implementation manner, the sparse constraint module is configured to:

[0142] Apply a sparse constraint norm to the mapped downsampled medical data to obtain a sparse matrix after sparse constraint;

[0143] Using the sparse matrix as an initial estimate, iteratively fill in the missing parts in the matrix through the conjugate gradient method to gradually obtain a complete and clear data matrix.

[0144] In a possible implementation manner, the machine learning model is a neural network, and the machine learning module is configured to;

[0145] Input the optimized downsampled medical data into the network to obtain the output T1 and T2 parameter quantization maps.

[0146] In a possible implementation manner, the inversion module is configured to:

[0147] Based on the tissue parameter quantification map, calculate the magnetic resonance signal intensity of each pixel through the imaging physical model;

[0148] Map the magnetic resonance signal intensity to the spatial domain through Fourier transform to obtain a medical image.

[0149] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments. For the sake of brevity, it will not be repeated here.

[0150] The embodiments of the present disclosure also provide an electronic device, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the above method.

[0151] The embodiments of the present disclosure also provide a non-volatile computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0152] The embodiments of the present disclosure also provide a computer program product, including a computer program, or a non-volatile computer-readable storage medium carrying the computer program. When the computer program is executed by a processor, the steps of the above method are implemented.

[0153] Figure 3 It is a block diagram of a device 1900 for medical image reconstruction based on physical signal prior shown according to an exemplary embodiment. For example, the device 1900 can be provided as a server or a terminal device. Referring to Figure 3 , the device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by a memory 1932 for storing instructions executable by the processing component 1922, such as application programs. The application programs stored in the memory 1932 can include one or more modules each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above method.

[0154] The device 1900 may further include a power supply component 1926 configured to perform power management of the device 1900, a wired or wireless network interface 1950 configured to connect the device 1900 to a network, and an input / output interface 1958 (I / O interface). The device 1900 can operate based on an operating system stored in the memory 1932, such as Windows Server TM , MacOS X TM , Unix TM , Linux TM , FreeBSDTM or the like.

[0155] In an exemplary embodiment, a non - volatile computer - readable storage medium is further provided, such as a memory 1932 including computer program instructions, and the above - mentioned computer program instructions can be executed by a processing component 1922 of the device 1900 to complete the above - mentioned method.

[0156] A computer - readable storage medium can be a tangible device that can hold and store programs / instructions used by an instruction - execution device. A computer - readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (non - exhaustive list) of the computer - readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read - only memory (ROM), an erasable programmable read - only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read - only memory (CD - ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punched card or raised structures in grooves storing instructions thereon, and any suitable combination of the above. The computer - readable storage medium used herein is not construed as an instantaneous signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagated through a waveguide or other transmission medium (e.g., optical pulses through an optical fiber cable), or electrical signals transmitted through wires.

[0157] The computer programs (or computer - readable program instructions) described herein can be downloaded from a computer - readable storage medium to various computing / processing devices, or downloaded to an external computer or external storage device through a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter or network interface in each computing / processing device receives the computer - readable program instructions from the network and forwards the computer - readable program instructions for storage in the computer - readable storage media in each computing / processing device.

[0158] A computer program (or computer program instructions) for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state-setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages - such as Smalltalk, C++, etc., and conventional procedural programming languages - such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on a user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network - including a local area network (LAN) or a wide area network (WAN) - or, alternatively, may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, by using the state information of the computer-readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit may execute the computer-readable program instructions to implement various aspects of the present disclosure.

[0159] Aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer-readable program instructions.

[0160] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that the instructions, when executed by the processor of the computer or other programmable data processing apparatus, create a means for implementing the functions / acts specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium, which causes a computer, a programmable data processing apparatus, and / or other devices to operate in a particular manner, such that the computer-readable medium storing the instructions includes a manufacture comprising instructions for implementing various aspects of the functions / acts specified in one or more blocks of the flowchart and / or block diagram.

[0161] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device, causing a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process such that the instructions executed on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in one or more boxes of the flowchart and / or block diagram.

[0162] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of code, or a portion of an instruction, which contains one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two consecutive blocks may in fact be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block of the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or acts, or by a combination of dedicated hardware and computer instructions.

[0163] The embodiments of the present disclosure have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or improvements made to the technology in the marketplace, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. A medical image reconstruction method based on prior physical signals, characterized in that, Comprising: Constructing a time dictionary library based on an imaging physical model, the time dictionary library containing different combinations of tissue characteristic parameters; Jointly decomposing the time dictionary library and the downsampled medical data to generate a time basis, and mapping the downsampled medical data to the subspace corresponding to the time basis; Applying a sparse constraint to the mapped downsampled medical data, and iteratively optimizing to complete the undersampled information in the downsampled medical data; Using a machine learning model to convert the optimized downsampled medical data into a tissue parameter quantification map; Based on the imaging physical model, inverting the tissue parameter quantification map into a medical image, and iteratively performing the process of applying the sparse constraint to inverting into a medical image to the medical image until the obtained medical image meets the convergence condition.

2. The method according to claim 1, characterized in that, The constructing the time dictionary library based on the imaging physical model includes: Based on the quantization range and quantization step of the preset tissue characteristic parameters, exhausting all parameter combinations, and calculating the weighted signals corresponding to each parameter combination through the imaging physical model; Arranging the parameter combinations and the corresponding weighted signals in a time series to form a time-signal mapping matrix, thereby forming a time dictionary library.

3. The method according to claim 1, wherein The medical image is a magnetic resonance image, and the tissue characteristic parameters of the time dictionary library include: T1 value, T2 value, pulse efficiency coefficient α and pulse flip angle θ after B1 field inhomogeneity correction.

4. The method according to claim 1, characterized in that The jointly decomposing the time dictionary library and the downsampled medical data to generate a time basis includes: Performing singular value decomposition on the time dictionary library to extract a first time basis; Performing singular value decomposition on the downsampled medical data to extract a second time basis; Combining the first time basis and the second time basis in a preset ratio to form a combined time basis.

5. The method according to claim 1, wherein The applying a sparse constraint to the mapped downsampled medical data and iteratively optimizing to complete the undersampled information in the downsampled medical data includes: Applying a sparse constraint norm to the mapped downsampled medical data to obtain a sparse matrix after sparse constraint; Using the sparse matrix as an initial estimate, and iteratively filling the missing part in the matrix by the conjugate gradient method to gradually obtain a complete and clear data matrix.

6. The method according to claim 1, wherein The machine learning model is a neural network, and the using the machine learning model to convert the optimized downsampled medical data into a tissue parameter quantification map includes; Inputting the optimized downsampled medical data into the network to obtain the output T1 and T2 parameter quantification maps.

7. The method according to claim 1, characterized in that The inverting the tissue parameter quantification map into a medical image based on the imaging physical model includes: Based on the tissue parameter quantification map, calculating the magnetic resonance signal intensity of each pixel through the imaging physical model; Mapping the magnetic resonance signal intensity to the spatial domain through Fourier transform to obtain a medical image.

8. A medical image reconstruction device based on the prior of physical signals, characterized in that, Comprising: A dictionary library construction module for constructing a time dictionary library based on an imaging physical model, the time dictionary library containing different combinations of tissue characteristic parameters; A mapping module for jointly decomposing the time dictionary library and the downsampled medical data to generate a time basis, and mapping the downsampled medical data to the subspace corresponding to the time basis; A sparse constraint module, which is used to impose a sparse constraint on the downsampled medical data after mapping, and iteratively optimize to complete the undersampled information in the downsampled medical data; A machine learning module, which is used to convert the optimized downsampled medical data into a tissue parameter quantification map by using a machine learning model; An inversion module, which is used to invert the tissue parameter quantification map into a medical image based on the imaging physical model; Iteratively execute the process of imposing the sparse constraint to inverting into a medical image on the medical image until the obtained medical image meets the convergence condition.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 7.

Images