Personalized magnetic resonance imaging method based on artificial intelligence, storage medium and equipment

Abstract

The invention discloses a personalized magnetic resonance imaging method and device based on artificial intelligence, a storage medium and system equipment. The method comprises the following steps: pre-scanning a target object to obtain pre-scanning data including a physiological parameter weighted graph and a magnetic field non-uniformity distribution graph; inputting the pre-scanning data and standard scanning parameters into a trained deep neural network; outputting an individual optimal scanning parameter by the deep neural network through an internal optimization process; personalized magnetic resonance scanning for the subject is performed using the individual optimal scanning parameters. According to the method, a new personalized magnetic resonance imaging normal form is created, artificial intelligence is introduced into a magnetic resonance imaging closed-loop decision, real-time and personalized optimization of the scanning parameters based on individual physiological information is achieved, the problem that in an existing fixed scanning parameter imaging scheme, the image quality is poor due to individual physiological differences is fundamentally solved, and the imaging quality is improved. And the accuracy and clinical diagnosis value of magnetic resonance imaging are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance imaging, and in particular relates to personalized and precise magnetic resonance imaging technology based on real-time optimization of sequence parameters using artificial intelligence. Background Technology

[0002] The clinical application of magnetic resonance imaging (MRI) technology is highly dependent on the setting of scanning parameters. The quality of parameter selection directly determines image contrast, signal-to-noise ratio, and final diagnostic value. However, in routine clinical practice, scanning parameters are usually set fixedly based on population averages or experience, making it difficult to adapt to the complex physiological and pathological differences between individuals, thus limiting the accuracy of the information acquired. For example, Chemical Exchange Saturation Transfer (CEST), as an emerging molecular imaging technique, can non-invasively detect the spatial distribution of low concentrations of metabolites in vivo, and has shown unique advantages and great potential in the diagnosis and treatment of various diseases such as tumors, stroke, and epilepsy. However, the low contrast of the detected signal severely limits its clinical translation.

[0003] Properly setting scanning parameters is a prerequisite for high-contrast CEST imaging, and the optimization of these parameters must be based on the physiological parameters of the imaging subject. Ideally, scanning parameters should be optimized according to the specific physiological conditions of the subject. However, currently widely used scanning parameters are derived from optimization methods based on a fixed set of physiological parameters. This "one-size-fits-all" scanning strategy ignores the significant inter-individual physiological differences brought about by age, race, and especially diseased tissues. This leads to a mismatch between scanning parameter settings and individual physiological states, resulting in poor image contrast, unstable or even failed molecular detection sensitivity, and severely reducing the accuracy of CEST imaging. Summary of the Invention

[0004] The purpose of this invention is to address the problems of existing magnetic resonance imaging (MRI) technology, which, due to fixed scanning parameters, cannot adapt to individual physiological differences, resulting in poor image contrast, low quantitative accuracy, and limited diagnostic value. This invention provides a precise MRI method, device, medium, and system based on artificial intelligence with high clinical and commercial value, to improve the quality and reliability of information acquired by MRI.

[0005] The specific technical solution adopted in this invention is as follows: In a first aspect, the present invention provides a personalized magnetic resonance imaging method based on artificial intelligence, comprising: S1. For the target object to be subjected to magnetic resonance imaging, control the magnetic resonance imaging equipment to perform a pre-scan under the initial scanning parameters to obtain pre-scan data including physiological parameter weighted map and magnetic field inhomogeneity distribution map. S2. The pre-scan data is input into a trained parameter optimization model for personalized optimization of the scan parameters. The parameter optimization model includes a physiological parameter quantification module and a scan parameter optimization module. After the pre-scan data is input, the physiological parameter quantification module first calculates the physiological parameter quantification map based on the physiological parameter weighting map and the magnetic field inhomogeneity distribution map. Then, the scan parameter optimization module performs optimization in the solution space of the scan parameters. During the optimization process, for each feasible solution of the scan parameters, it needs to be compared with the calculated physiological parameter quantification. Figure 1 The simulated magnetic resonance image is obtained by inputting the magnetic resonance signal into the simulator, and the image quality score of the simulated magnetic resonance image is calculated. The optimal solution with the highest image quality score is output as the individual optimal scanning parameters. S3. Send the individual's optimal scanning parameters to the magnetic resonance imaging device to perform a personalized magnetic resonance scan for the target object.

[0006] As a preferred embodiment of the first aspect above, CEST imaging is used for magnetic resonance imaging of the target object, and the optimized scanning parameters are saturation pulse parameters or image acquisition parameters; preferably, the optimized scanning parameters are saturation pulse intensity.

[0007] As a preferred embodiment of the first aspect above, the physiological parameter quantification module employs a deep neural network, with the weighted graph of the physiological parameters and the magnetic field inhomogeneity distribution graph as network inputs, and the physiological parameter quantification graph as network output; preferably, the deep neural network is a neural network based on a self-attention mechanism.

[0008] As a preferred embodiment of the first aspect above, the scanning parameter optimization module includes a magnetic resonance signal simulator and an image quality scoring module. The magnetic resonance signal simulator employs a deep neural network or a Bloch equation simulator, and the image quality scoring module has a built-in image quality evaluation function. Preferably, the deep neural network is a neural network based on a self-attention mechanism. Preferably, the image quality evaluation function is an APTw contrast calculation function between tumor and normal tissue.

[0009] As a preferred embodiment of the first aspect, after the scanning parameter optimization module receives the physiological parameter quantitative map calculated by the physiological parameter quantitative module, it searches for the optimal solution in the solution space of the scanning parameters by traversal optimization. The traversal optimization method is as follows: extract all feasible solutions from the solution space of the scanning parameters, then input all feasible solutions in parallel into the magnetic resonance signal simulator to obtain the corresponding simulated magnetic resonance image, and then further obtain the image quality score of the simulated magnetic resonance image corresponding to each feasible solution through the image quality evaluation function. The feasible solution corresponding to the simulated magnetic resonance image with the highest image quality score is taken as the optimal solution, and the scanning parameters corresponding to the optimal solution are taken as the individual optimal scanning parameters.

[0010] As a preferred embodiment of the first aspect, after the scanning parameter optimization module receives the physiological parameter quantitative map calculated by the physiological parameter quantitative module, it searches for the optimal solution in the solution space of the scanning parameters using an iterative optimization method. The iterative optimization method is as follows: a current feasible solution is determined in the solution space of the scanning parameters as the starting point of the iteration, and the optimization operation is performed iteratively. During the optimization process, for each current feasible solution, it is first input into the magnetic resonance signal simulator to obtain the corresponding simulated magnetic resonance image. Then, the image quality score of the simulated magnetic resonance image corresponding to the current feasible solution is obtained through the image quality evaluation function. With the goal of maximizing the image quality score, the gradient descent optimization algorithm is used to calculate the gradient and update to obtain the next feasible solution. When the iteration termination condition is reached, the feasible solution obtained by the last gradient update is taken as the optimal solution, and the scanning parameters corresponding to the optimal solution are taken as the individual's optimal scanning parameters. Preferably, the feasible solution chosen as the starting point of the iteration is the initial scan parameter in S1.

[0011] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, enables the implementation of the artificial intelligence-based personalized magnetic resonance imaging method as described in any of the first aspects above.

[0012] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the artificial intelligence-based personalized magnetic resonance imaging method as described in any of the first aspects above.

[0013] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the AI-based personalized magnetic resonance imaging method as described in any of the first aspects above.

[0014] Fifthly, the present invention provides a magnetic resonance imaging device, which includes a magnetic resonance scanner and a control unit; The control unit stores a computer program, which, when executed, controls the magnetic resonance scanner to implement the artificial intelligence-based personalized magnetic resonance imaging method as described in any of the first aspects above, to complete a personalized magnetic resonance scan of the target object and obtain personalized magnetic resonance imaging data.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Personalized and precise scanning is achieved: This invention breaks through the limitations of traditional fixed parameter protocols, quickly extracts individual physiological information through artificial intelligence algorithms, and optimizes scanning parameters in real time and online, realizing a paradigm shift from "one-size-fits-all" to "tailor-made", enhancing the accuracy of imaging detection and improving the overall diagnostic efficiency of magnetic resonance imaging.

[0016] (2) Ensures the high efficiency of clinical workflow: The present invention can complete parameter optimization within seconds, so that real-time parameter optimization can be seamlessly integrated into the routine clinical scanning process without interrupting the process or significantly increasing the examination time, which is highly practical.

[0017] (3) It has broad applicability and scalability: The technical framework of this invention is not limited to a specific imaging sequence or a specific imaging index. Its core idea can be widely applied to various magnetic resonance imaging techniques such as structure, metabolism and function. It can optimize key sequence parameters and shows great clinical translation potential. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of a personalized magnetic resonance imaging method based on artificial intelligence. Figure 2 This is a diagram of a deep neural network architecture according to an embodiment of the present invention; Figure 3 The following are the optimized results of the water model experiment in the examples. (a) is the T2-weighted image; (b) is the APTw image obtained using different imaging parameters, obtained from left to right by the standard scanning parameters, the optimized scanning parameters based on the method of this invention, and the experimentally determined actual optimal scanning parameters; (c) is a quantitative comparison of the APTw signal contrast between two test tubes obtained using the standard method, the method of this invention, and the experimentally determined actual optimal method; (d) from top to bottom are the APTw images actually acquired under different B1 values ​​and the APTw images predicted under different B1 values ​​based on the method of this invention. Figure 4This example presents the clinical test results of a patient with a high-grade glioma. (a) shows FLAIR and enhanced T1-weighted images, where the yellow curve indicates the tumor region of interest (ROI) and the white curve indicates the normal tissue region of interest (ROI). (b) shows, from left to right, APTw images obtained using standard scanning parameters and the personalized optimized scanning parameters of this invention. (c) shows the signal intensity distribution curves of the tumor-normal tissue interface region under the two schemes. (d) shows, from left to right, the normalized APTw signal obtained based on the method of this invention and the corresponding APTw contrast enhancement effect, where signal normalization is obtained by subtracting the average signal of the normal tissue ROI. Figure 5 This example presents the clinical test results of a patient with a low-grade glioma. (a) shows FLAIR and enhanced T1-weighted images, with the yellow curve indicating the tumor ROI and the white curve indicating the normal tissue ROI. (b) shows, from left to right, APTw images obtained using standard scanning parameters and the personalized optimized scanning parameters of this invention. (c) shows the signal intensity distribution curves of the tumor-normal tissue interface region under the two schemes. (d) shows, from left to right, the normalized APTw signal obtained based on the method of this invention and the corresponding APTw contrast enhancement effect, where signal normalization is obtained by subtracting the average signal of the normal tissue ROI. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in various embodiments of the present invention can be combined accordingly without mutual conflict.

[0020] like Figure 1 As shown, as a preferred embodiment of the present invention, a personalized magnetic resonance imaging method based on artificial intelligence is provided, which includes the following steps: S1. For the target object to be subjected to magnetic resonance imaging, control the magnetic resonance imaging equipment to perform a pre-scan under the initial scanning parameters to obtain pre-scan data including physiological parameter weighted maps and magnetic field inhomogeneity distribution maps.

[0021] It should be noted that the target object for magnetic resonance imaging in this invention refers to the individual or patient who needs to undergo magnetic resonance imaging, and the specific object is not limited.

[0022] It should also be noted that the initial scanning parameters used in the above pre-scan can be standard scanning parameters set according to industry standards or expert experience, and the specific optimizable scanning parameters can be adjusted according to actual needs.

[0023] It should also be noted that the magnetic resonance imaging in this invention can preferably be CEST magnetic resonance imaging. The corresponding optimized scanning parameters can be the saturation pulse parameters in the saturation module of the imaging sequence, or the image acquisition parameters in the readout module of the imaging sequence. Specific pulse parameters include pulse intensity, pulse interval, number of pulses, etc.

[0024] It should also be noted that, in the embodiments of the present invention, the above-mentioned physiological parameter weighted map includes the z-spectrum of all voxels as well as the T1-weighted (T1w) map and the T2-weighted (T2w) map, and the above-mentioned magnetic field inhomogeneity distribution map can be a B1 map or a B0 map.

[0025] S2. Input the pre-scan data into the trained parameter optimization model for personalized optimization of the scan parameters. This parameter optimization model includes a physiological parameter quantification module and a scan parameter optimization module. After the pre-scan data is input, the physiological parameter quantification module first calculates the physiological parameter quantification map based on the physiological parameter weighting map and the magnetic field inhomogeneity distribution map. Then, the scan parameter optimization module performs optimization in the solution space of the scan parameters. During the optimization process, for each feasible solution of the scan parameters, it needs to be compared with the calculated physiological parameter quantification map. Figure 1 In the input magnetic resonance signal simulator, the magnetic resonance signal simulator uses the scanning parameters corresponding to the current feasible solution as conditions, calculates the simulated magnetic resonance image based on the physiological parameter quantitative map, and then calculates the image quality score of the simulated magnetic resonance image. The optimal solution with the highest image quality score is output as the individual optimal scanning parameters.

[0026] It should be noted that the solution space of the aforementioned scanning parameters refers to the parameter space composed of the value ranges of all scanning parameters involved in the optimization. All feasible solutions in subsequent optimization algorithms need to be searched within this parameter space. Since the scanning parameters required for magnetic resonance imaging (MRI) are a series of parameter combinations, such as pulse intensity, pulse interval, and pulse count, if some scanning parameters do not participate in the optimization, they can be considered as fixed values. During the optimization process, only the scanning parameters to be optimized need to be sought within the parameter space. After final optimization, the MRI device can combine the optimized optimal scanning parameters with other fixed scanning parameters to achieve MRI. For example, in an embodiment of this invention, the optimized scanning parameter is the saturated pulse intensity (B1 value). At this time, all other CEST scanning parameters are fixed to preset standard values ​​and do not need to participate in optimization. Finally, during personalized imaging, the optimized B1 value can be combined with other fixed CEST scanning parameter standard values ​​to form a complete set of scanning parameters for CEST MRI.

[0027] It should be noted that the quantitative physiological parameter map in this invention can be regarded as an image recording the physiological parameters of each voxel within the imaging domain. For CEST imaging, the physiological parameters of each voxel are preferably the concentration, exchange rate, and relaxation time of CEST exchangeable protons.

[0028] Furthermore, the parameter optimization model of this invention is mainly constructed based on deep neural networks and optimization algorithms. The specific implementation of the two core modules in the parameter optimization model—the physiological parameter quantification module and the scanning parameter optimization module—is described in detail below.

[0029] In the parameter optimization model of this invention, the physiological parameter quantification module employs a deep neural network. Its network input consists of the aforementioned physiological parameter weighted graph and the aforementioned magnetic field inhomogeneity distribution graph, and its network output is the physiological parameter quantification graph. In an embodiment of this invention, the deep neural network is a neural network based on a self-attention mechanism, specifically implemented by cascading a fully connected network after the Transformer module. Although both the input and output of this physiological parameter quantification module are in the form of graphs, the prediction of the physiological parameter quantification value for each voxel within the model is actually independent. Therefore, the physiological parameter quantification module can be considered as performing the prediction of physiological parameter quantification values ​​for all voxels in parallel. That is, based on the physiological parameter weighted value of each voxel (in this embodiment, the physiological parameter weighted value consists of the z-spectrum and T1w and T2w) and the magnetic field inhomogeneity value (in this embodiment, the B1 value is used), the corresponding physiological parameter quantification value for that voxel is predicted (in this embodiment, the concentration, exchange rate, and relaxation time of exchangeable protons in CEST are used), and thus, the physiological parameter quantification values ​​of all voxels constitute a physiological parameter quantification graph covering the imaging domain.

[0030] In the parameter optimization model of this invention, the scanning parameter optimization module includes a magnetic resonance signal simulator and an image quality scoring module. The magnetic resonance signal simulator employs a deep neural network or a Bloch equation simulator (which can be implemented using a program, function, or software for solving the Bloch equation). The image quality scoring module has a built-in image quality evaluation function. In embodiments of this invention, the deep neural network is preferably a neural network based on a self-attention mechanism, specifically implemented by cascading a fully connected network after the Transformer module. Similarly, in embodiments of this invention, the image quality evaluation function is preferably an APTw contrast calculation function between tumor and normal tissue; therefore, the image quality score is the APTw contrast between tumor and normal tissue. The scanning parameter optimization module uses an optimization algorithm to schedule the magnetic resonance signal simulator and the image quality scoring module to generate simulated magnetic resonance images and calculate their image quality scores, respectively.

[0031] like Figure 2 The diagram illustrates a specific network structure for a parameter optimization model in an embodiment of the present invention. This parameter optimization model consists of a cascaded physiological parameter quantification module and a parameter optimization model. Both the physiological parameter quantification module and the parameter optimization model employ an N-layer Transformer layer followed by a cascaded fully connected network. The number of Transformer layers N and the number of fully connected layers M in the fully connected network can be adjusted according to actual needs, preferably N=4 and M=1. An activation function, preferably a Sigmoid activation function, can be used to generate the final output after the fully connected layers.

[0032] In this invention, the specific form of the optimization algorithm used in the scan parameter optimization module is not limited; it can be a traversal algorithm, other gradient descent algorithms, genetic algorithms, etc. Two optional optimization algorithms are described below.

[0033] The first optimization algorithm uses a traversal optimization approach. Specifically, after the scanning parameter optimization module receives the physiological parameter quantitative map calculated by the physiological parameter quantitative module, it searches for the optimal solution in the solution space of the scanning parameters using a traversal optimization approach. The traversal optimization method is as follows: extract all feasible solutions from the solution space of the scanning parameters, then input all feasible solutions in parallel into the magnetic resonance signal simulator to obtain the corresponding simulated magnetic resonance images, and then further obtain the image quality score of the simulated magnetic resonance image corresponding to each feasible solution through the image quality evaluation function. The feasible solution corresponding to the simulated magnetic resonance image with the highest image quality score is taken as the optimal solution, and the scanning parameters corresponding to the optimal solution are taken as the individual's optimal scanning parameters.

[0034] This optimized traversal method extracts all feasible solutions at once and then performs a single forward computation in parallel to obtain the image quality scores of all feasible solutions in batches. Theoretically, as long as the density of feasible solutions sampled from the solution space is high enough, the ideal optimal solution can be obtained. However, this method is computationally intensive and is suitable for situations with few parameters to be optimized and a small solution space.

[0035] The second optimization algorithm employs an iterative optimization approach. Specifically, after the scanning parameter optimization module receives the quantitative physiological parameter map calculated by the physiological parameter quantification module, it iteratively searches for the optimal solution in the solution space of the scanning parameters. The iterative optimization method involves determining a current feasible solution in the solution space of the scanning parameters as the starting point for iteration, iteratively performing the optimization operation, and during the optimization process, for each current feasible solution, first inputting it into the magnetic resonance signal simulator to obtain the corresponding simulated magnetic resonance image, and then further obtaining the image quality score of the simulated magnetic resonance image corresponding to the current feasible solution through the image quality evaluation function. With the goal of maximizing the image quality score, the gradient descent optimization algorithm is used to calculate the gradient and update it to obtain the next feasible solution. When the iteration termination condition is met, the feasible solution obtained from the final gradient update is taken as the optimal solution, and the scanning parameters corresponding to the optimal solution are taken as the individual's optimal scanning parameters.

[0036] The gradient descent optimization algorithm described above is existing technology. For ease of understanding, this invention provides an iterative optimization method based on the gradient descent algorithm. This method involves specifying a set of scanning parameters as the starting point in the solution space of the scanning parameters, then gradually updating feasible solutions by calculating gradients, iteratively obtaining image quality scores for different feasible solutions until a termination condition is met. The specific steps are as follows: B1. Using the initial scanning parameters in S1 as the initial feasible solution, i.e. the iteration starting point, the corresponding simulated magnetic resonance image is calculated based on the magnetic resonance signal simulator, and then the image quality score of the simulated magnetic resonance image is calculated through the image quality evaluation function. B2. Using a gradient descent-based optimization algorithm, construct a differentiable optimization computation graph. In the differentiable optimization computation graph, calculate the gradient of the image quality score of the simulated magnetic resonance image with respect to the current scanning parameters. B3. Based on the gradient, iteratively update the scanning parameters along the direction that improves the image quality score to obtain the updated scanning parameters, i.e., the next feasible solution; B4. Repeat steps B2 to B3. In each iteration, use the updated scanning parameters as the new current scanning parameters and recalculate the simulated magnetic resonance image and image quality score until the iteration termination condition is met. Use the scanning parameters obtained in the last iteration as the optimal solution, i.e., the individual optimal scanning parameters mentioned above.

[0037] The above iteration termination condition can be that the update amount of the scan parameters is less than a preset threshold or the maximum number of iterations is reached.

[0038] It should also be noted that although the input and output of this scanning parameter optimization module are in the form of graphs, the prediction of the simulated magnetic resonance signal value for each voxel in the graph is actually independent within the model. Therefore, the scanning parameter optimization module can be regarded as predicting the simulated magnetic resonance signal value for all voxels in parallel. That is, based on the quantitative physiological parameter value of each voxel obtained by the physiological parameter quantification module, the corresponding z-spectrum of that voxel is predicted, and then a simulated magnetic resonance image covering the imaging domain is obtained based on the z-spectrum of all voxels.

[0039] It should be noted that the above-mentioned parameter optimization model needs to be pre-trained under supervised supervision before being used for actual inference. The training method of the model is existing technology, and only the loss function needs to be reasonably adjusted. In the embodiments of the present invention, the loss function used for training the parameter optimization model can be set as the deviation between the simulated magnetic resonance image and the actual magnetic resonance image corresponding to the same scanning parameters (e.g., the same B1). The deviation can be in the form of the mean square error or average error of the signal values ​​of all voxels in the image (which can be represented by the z-spectrum).

[0040] S3. Send the above-mentioned optimal individual scanning parameters to the magnetic resonance imaging device for performing personalized magnetic resonance scanning of the above-mentioned target object.

[0041] It should be noted that steps S1 to S3 described above can essentially be implemented through computer programs or software modules. Specifically, they can be mounted on a control unit capable of executing computer programs. The resulting individual-optimal scanning parameters can be sent to the magnetic resonance imaging (MRI) device, which then performs MRI on the target object according to these parameters. Because the individual-optimal scanning parameters for MRI are pre-optimized with the goal of maximizing image quality scores, higher quality MRI images can be obtained compared to standard scanning parameters. The resulting MRI images are specifically optimized for the target object and can therefore be considered personalized MRI images.

[0042] Therefore, based on the same inventive concept, the present invention also provides a computer electronic device corresponding to the artificial intelligence-based personalized magnetic resonance imaging method provided in the above embodiments, which includes a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the AI-based personalized magnetic resonance imaging method as described above when executing the computer program.

[0043] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0044] Therefore, based on the same inventive concept, the present invention provides a computer-readable storage medium corresponding to a personalized magnetic resonance imaging method based on artificial intelligence. The storage medium stores a computer program, which, when executed by a processor, can realize the personalized magnetic resonance imaging method based on artificial intelligence as described above.

[0045] Therefore, based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the artificial intelligence-based personalized magnetic resonance imaging method as described above.

[0046] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S3.

[0047] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0048] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0049] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0050] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a magnetic resonance imaging device corresponding to the artificial intelligence-based personalized magnetic resonance imaging method provided in the above embodiments, which includes a magnetic resonance scanner and a control unit; wherein the control unit stores a computer program, which, when executed, controls the magnetic resonance scanner to realize the artificial intelligence-based personalized magnetic resonance imaging method as described in S1~S3 above, completes a personalized magnetic resonance scan of the target object, and obtains personalized magnetic resonance imaging data.

[0051] It should be noted that the magnetic resonance imaging (MRI) device can be any MRI scanner capable of implementing parallel imaging methods. Its structure is existing technology, and mature commercial products can be used; the specific model is not limited. Furthermore, in addition to storing the aforementioned computer program, the control unit of the MRI device should also contain the imaging sequences and other software programs necessary for implementing MRI. This control unit can be a standalone unit or an integrated unit of the MRI scanner. That is, the aforementioned AI-based personalized MRI method can be integrated into the control unit of the MRI device as a data processing program, allowing the MRI scanner to directly output optimized acquisition results without the need for an additional control unit.

[0052] The following describes the personalized magnetic resonance imaging method based on artificial intelligence as described in S1 to S3 above, and combines it with specific embodiments to demonstrate its specific technical effects, so that those skilled in the art can better understand the essence of the present invention.

[0053] Example The personalized magnetic resonance imaging method based on artificial intelligence described in S1-S3 above (hereinafter referred to as the method of this invention for ease of description) will be applied to a specific embodiment to demonstrate its technical effects. The specific framework and process of the method of this invention are as described above, and will not be repeated in full in this embodiment. The following focuses on demonstrating the specific implementation details and technical effects of each step.

[0054] 1. Data Preparation 1.1 Training Data Simulation In this embodiment, the magnetic resonance imaging uses CEST imaging scanning. The scanning parameter to be optimized is the saturation pulse intensity, i.e., the B1 value. All other CEST scanning parameters are standard parameters.

[0055] In this embodiment, B1 = 2μT is used as the initial B1 value in the standard scanning parameters. By constructing and solving the seven-pool Bloch-McConnell equation, z-spectrums under different target B1 values ​​are generated for network training. These z-spectrums cover more than 250 million combinations of tissue parameters and B1 values. The sampling range of the target B1 is 0.5-3.1μT, with a step size of 0.1μT. As mentioned earlier, the parameter optimization model is essentially a parallel or serial prediction based on individual voxels. Therefore, in the final training dataset, each training sample corresponds to the input and output label data of a voxel. The input data includes: the simulated z-spectrum of the current voxel at B1 = 2μT, the corresponding T1, T2, and B1 values ​​(due to the inhomogeneity of the B1 field, the B1 value varies at each voxel, and the B1 value here is not necessarily equal to 2μT). The output label is: the simulated z-spectrum under the target B1. In addition, to increase the robustness and generalization ability of the network, this embodiment introduces perturbations into the data in two ways: first, Gaussian white noise is added to the input with an intensity of 0.5% of the standard deviation of the input signal; second, the B1 value corresponding to the input z spectrum is varied by ±20% to simulate the non-uniformity of the radio frequency transmission field.

[0056] 1.2 Water Model Preparation This embodiment prepared a bovine serum albumin (BSA) aqueous model (pH = 7.0): a 15 cm diameter plastic bottle was used to hold phosphate buffer solution, and two 3 cm diameter test tubes were placed inside, one containing 8% BSA aqueous solution and the other containing 4% BSA aqueous solution. MnCl2 and agarose were added to ensure that the T1 and T2 values ​​of the BSA aqueous model were within the physiologically relevant range for humans.

[0057] 1.3 MRI Data Acquisition To demonstrate the actual performance of this method in clinical settings, this embodiment uses BSA water phantoms and brain tumor patients as validation subjects to conduct CEST-MRI scanning experiments.

[0058] Specifically, a 3-Tesla Siemens scanner (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany) with 20 head coils was used to scan water phantoms and the brains of seven patients with brain tumors. The 2D fast spin echo (TSE) CEST imaging sequence was used in the scanning experiments. The specific acquisition parameters used during pre-acquisition were: saturation pulse duration of 1.0 s, intensity of 2 μT, flip angle FA = 90°; echo time (TE) = 6.7 ms; repetition time (TR) = 3 s; field of view (FOV) = 212 × 186 mm. 2 Resolution = 2.2 × 2.2 mm 2 Slice thickness = 5 mm; acquisition turbine coefficient = 96. A total of 54 frequency-shifted frames were acquired, including unsaturated frames S0 and saturated frames saturated at frequencies of 0, ±0.25, ±0.5, ±0.75, ±1, ±1.5, ±2 (2), ±2.5 (2), ±3 (2), ±3.25 (2), ±3.5 (6), ±3.75 (2), ±4 (2), ±4.5, ±5, ±6 ppm (the numbers in parentheses represent the number of times the corresponding frequency frame was repeatedly acquired). For the optimized personalized scan, the same CEST imaging sequence was used, but the saturation intensity was optimized using the parameters of this invention, and 9 frequency-shifted frames were acquired, including unsaturated frames S0 and saturated frames saturated at frequencies of ±3, ±3.5 (2), ±4 ppm. To verify the accuracy of the optimization method of the present invention, multiple sets of response z-spectrum data of the water model were measured at room temperature as the z-spectrum changed with the B1 field strength (1.5~4.5 μT, step size 0.5 μT).

[0059] In addition, during the pre-scan, after completing CEST-MRI data acquisition, B0, B1, T1, and T2 maps were acquired at the same slice location, using the same field of view and matrix size. Specifically, B0 and B1 maps were acquired using a dual-echo gradient echo sequence (echo time = 4.92 and 9.84 ms) and a pre-modulated RF pulse sequence, respectively, with B0 used to correct the original z-spectrum. T1 and T2 maps were acquired using an inversion recovery sequence (inversion time = 150–1300 ms) and a multi-echo spin echo sequence (echo time = 12–192 ms), respectively. To correct for B0 field inhomogeneities, this embodiment calculated the B0 spectrum using the Water Saturation Shift Referencing (WASR) method. The WASSR sequence used had a TR of 2s and a saturation pulse intensity of 0.5μT. A total of 26 frequency points were acquired, which were equally spaced between -1.5 and 1.5 ppm. Other parameters were consistent with the CEST imaging sequence mentioned above.

[0060] 2. Parameter optimization model building and training In this embodiment, the parameter optimization model is built using the deep learning framework PyTorch, as shown below. Figure 2 The deep neural network shown consists of a physiological parameter quantification module and a scan parameter optimization module. Both the physiological parameter quantification module and the scan parameter optimization module utilize a Transformer-based neural network. Specifically, the Transformer-based neural network comprises three Transformer layers and one fully connected layer, with each Transformer layer having four attention heads. A Sigmoid activation function is used after the fully connected layer to ensure that the network output remains within a physiologically reasonable range.

[0061] In this embodiment, the input to the physiological parameter quantification module is a weighted physiological parameter map (z-spectrum, T1 map, and T2 map of all voxels at B1 = 2 μT) and a magnetic field inhomogeneity distribution map (B1 map). The output is an encoded physiological parameter quantification map (containing the concentration, exchange rate, and relaxation time of CEST exchangeable protons for all voxels). The scanning parameter optimization module employs an iterative optimization method, uniformly sampling within the target B1 sampling range of 0.5-3.1 μT at a step size of 0.1 μT. The target B1 value obtained from each sampling is used as a conditional parameter, which is then used in conjunction with the physiological parameter quantification. Figure 1 In a parallel input magnetic resonance signal simulator, simulated magnetic resonance images corresponding to different target B1 values ​​are obtained.

[0062] The entire deep neural network framework is trained end-to-end using a simulation-generated dataset. During training, the Adam optimizer is used to minimize the loss function value shown in Equation (1) below on the set of network parameters. Optimization was performed to obtain relatively optimal network parameters. The neural network was iterated for 50 epochs on a single NVIDIA RTX 3090 Ti GPU. During training, the validation set was used to test the generalization ability of the current model and determine whether the network had reached the convergence condition.

[0063] In this embodiment, the loss function used for training the parameter optimization model is: (1) in and These represent the mappings corresponding to the physiological parameter quantification module and the scan parameter optimization module, respectively. and These represent the sets of learnable parameters in these two modules, Given the z-spectral vector with input B1 = 2μT, Represents other input parameter plots (B1, T1, and T2 plots). Representing target B1, Let z be the target z-spectral vector corresponding to target B1.

[0064] It should be noted that in practical applications, in addition to the magnetic resonance signal simulator, the scanning parameter optimization module also needs to have an optimization algorithm. The optimization algorithm needs to combine the image quality evaluation function built into the image quality scoring module to calculate the image quality score of the simulated magnetic resonance image obtained for different target B1 values, and use the highest image quality score as the selection target to obtain the optimal target B1 value. However, since the optimization algorithm itself does not have learnable parameters, it is not necessary to consider the optimization algorithm during the model training stage. That is, the scanning parameter optimization module directly calls the magnetic resonance signal simulator to calculate the simulated magnetic resonance image for all feasible target B1 values, and then optimizes the entire deep neural network using the loss function shown in the above formula (1).

[0065] In embodiments of the present invention, the image quality assessment function is an APTw contrast calculation function between tumor and normal tissue. Specifically, the CEST effect, using APTw images as an example, is quantified through magnetization transfer rate asymmetry analysis. APTw is defined as: (2) Therefore, the APTw contrast ratio between tumor and normal tissue is calculated as follows: APTC = [APTw] tumor – [APTw] normal , (3) Among them [APTw] tumor and [APTw] normal These represent the average APTw values ​​within the manually delineated tumor and contralateral normal tissue ROI, respectively.

[0066] Based on the above model framework and training method, once the parameter optimization model is trained, the deep neural network can output the corresponding target z-spectrum for any specified target saturation pulse intensity. During actual clinical use, the network's scanning parameter optimization module adopts an traversal optimization method to generate a target z-spectrum set for all candidate scanning parameters at once. The simulated magnetic resonance image can be obtained based on the target z-spectrum set. Then, according to the calculation formula of APTw contrast APTC shown in the above formula (2), the scanning parameter with the highest APTC can be defined as the individual's optimal scanning parameter for subsequent personalized magnetic resonance imaging.

[0067] 3. Personalized Imaging Process Step 1: Pre-scan. During this step, a standard z-spectrum is acquired using B1 = 2 μT, along with quantitative B0, B1, T1w, and T2w plots. This step takes approximately 5 minutes.

[0068] Step 2: Data Processing. In the data processing stage, based on the original standard z-spectrum and B0 map, B0 correction of the z-spectrum is performed to obtain the corrected z-spectrum as the model input. Then, the T1w and T2w maps are fitted to obtain the T1 and T2 maps. The ROI of the tumor and the contralateral normal tissue is manually delineated on the T1 map, which takes about 5 minutes.

[0069] Step 3: Personalized Parameter Optimization. Based on the processed data, the corrected z-spectrum, T1 map, and T2 map are used as physiological parameter weighted maps, and the B1 map is used as a magnetic field inhomogeneity distribution map. These are input into the trained parameter optimization model to optimize the saturation pulse intensity of CEST imaging (i.e., the aforementioned target B1 value) to obtain the highest B1 value for APTC. This step takes approximately 10 seconds.

[0070] Step 4: Personalized MRI. Using optimized saturation intensity values, a personalized and precise CEST acquisition is performed on the current patient. This step takes approximately 30 seconds.

[0071] This invention employs a parallel coordination strategy to improve clinical efficiency. While the workstation is processing data and optimizing parameters, the MRI scanner remains operational and can continue to perform routine clinical sequences (such as T1-weighted and T2-weighted scans), thus minimizing additional examination time. Each patient requires only about 5 minutes of additional scanning time.

[0072] 4. Data Post-processing and Analysis In this embodiment, since the CEST signal amplitude varies significantly at different B1 levels, to facilitate visual comparison between the optimized and standard schemes, the APTw graph is displayed using an automatic scaling strategy for the color bar. The lower and upper limits of the display window relative to [APTw]normal are set as follows: (lower bound, upper bound) = ([APTw] normal - 6, [APTw] tumor + 4) (4) 5. Results Analysis Figure 3 Using a BSA water model as an example, the accuracy and effectiveness of the scanning parameter optimization method of this invention are demonstrated. The results show that the B1 value optimized by the method of this invention is highly consistent with the experimentally measured optimal value, and achieves a very close contrast ratio (2.68% vs. 2.73%). Compared to standard scanning parameters, the scanning scheme optimized by the method of this invention improves the contrast ratio by 44.07%, reaching 94% of the maximum possible contrast ratio. Visually, compared to the standard acquisition scheme, the signal distinction between the two test tubes is clearer in the APTw image optimized by the method of this invention.

[0073] Figure 4 Using a case of a high-grade glioma patient as an example, the optimization effect of the method of the present invention is demonstrated. In the individual optimal scanning parameters obtained by the method of the present invention, the saturation pulse intensity B1 is 0.9 μT. The resulting APTw image shows a clearer tumor boundary and highly matches the enhancement area of ​​the anatomical T1-weighted enhanced image. Figure 4 (ab); The images acquired based on the method of this invention also reveal more refined intratumoral heterogeneity, fully displaying the ring-like enhancement structure (ab); Figure 4 (See the red arrow in b), and the APTw contrast value is 44.76% higher than the standard scan parameters.

[0074] for Figure 5 The example shown is a low-grade glioma patient. The standard scanning protocol, due to the "one-size-fits-all" acquisition (i.e., the saturation pulse intensity B1 is completely fixed), suffers from sensitivity limitations and fails to provide discernible tumor-normal tissue contrast. The method of this invention, optimized with a customized scheme using the individual's optimal scanning parameter B1=0.7μT, overcomes this deficiency and significantly improves lesion visibility. Figure 5 (b), achieving a 344.37% contrast improvement ( Figure 5 (d); The standardized tumor-normal tissue interface signal intensity curve shows that the contrast enhancement is mainly due to the increase in tumor signal, while the normal tissue signal remains basically stable. Figure 5 (c and d).

[0075] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A personalized magnetic resonance imaging method based on artificial intelligence, characterized in that, include: S1. For the target object to be subjected to magnetic resonance imaging, control the magnetic resonance imaging equipment to perform a pre-scan under the initial scanning parameters to obtain pre-scan data including physiological parameter weighted map and magnetic field inhomogeneity distribution map. S2. Input the pre-scan data into the trained parameter optimization model for personalized optimization of scan parameters. The parameter optimization model includes a physiological parameter quantification module and a scan parameter optimization module. After the pre-scan data is input, the physiological parameter quantification module first calculates the physiological parameter quantification map based on the physiological parameter weighting map and the magnetic field inhomogeneity distribution map. Then, the scan parameter optimization module optimizes the solution in the solution space of the scan parameters. During the optimization process, for each feasible solution of the scan parameters, it needs to be input together with the calculated physiological parameter quantification map into the magnetic resonance signal simulator to obtain a simulated magnetic resonance image. The image quality score of the simulated magnetic resonance image is calculated, and the optimal solution with the highest image quality score is output as the individual optimal scan parameter. S3. Send the individual's optimal scanning parameters to the magnetic resonance imaging device to perform a personalized magnetic resonance scan for the target object.

2. The personalized magnetic resonance imaging method based on artificial intelligence as described in claim 1, characterized in that, The magnetic resonance imaging of the target object uses CEST imaging, and the optimized scanning parameters are either saturation pulse parameters or image acquisition parameters; preferably, the optimized scanning parameters are saturation pulse intensity.

3. The personalized magnetic resonance imaging method based on artificial intelligence as described in claim 1, characterized in that, The physiological parameter quantification module employs a deep neural network, with the weighted graph of the physiological parameters and the magnetic field non-uniformity distribution graph as its network inputs, and the physiological parameter quantification graph as its network output; preferably, the deep neural network is a neural network based on a self-attention mechanism.

4. The personalized magnetic resonance imaging method based on artificial intelligence as described in claim 1, characterized in that, The scanning parameter optimization module includes a magnetic resonance signal simulator and an image quality scoring module. The magnetic resonance signal simulator uses a deep neural network or a Bloch equation simulator, and the image quality scoring module has a built-in image quality evaluation function. Preferably, the deep neural network is a neural network based on a self-attention mechanism. Preferably, the image quality evaluation function is an APTw contrast calculation function between tumor and normal tissue.

5. The personalized magnetic resonance imaging method based on artificial intelligence as described in claim 1, characterized in that, When the scanning parameter optimization module receives the physiological parameter quantitative map calculated by the physiological parameter quantitative module, it searches for the optimal solution in the solution space of the scanning parameters by traversing the optimization method. The traversal optimization method is as follows: extract all feasible solutions from the solution space of the scanning parameters, then input all feasible solutions in parallel into the magnetic resonance signal simulator to obtain the corresponding simulated magnetic resonance image, and then further obtain the image quality score of the simulated magnetic resonance image corresponding to each feasible solution through the image quality evaluation function. The feasible solution corresponding to the simulated magnetic resonance image with the highest image quality score is taken as the optimal solution, and the scanning parameters corresponding to the optimal solution are taken as the individual's optimal scanning parameters.

6. The personalized magnetic resonance imaging method based on artificial intelligence as described in claim 1, characterized in that, After the scanning parameter optimization module receives the physiological parameter quantitative map calculated by the physiological parameter quantitative module, it searches for the optimal solution in the solution space of the scanning parameters using an iterative optimization method. The iterative optimization method is as follows: a current feasible solution is determined in the solution space of the scanning parameters as the iteration starting point, and the optimization operation is performed iteratively. During the optimization process, for each current feasible solution, it is first input into the magnetic resonance signal simulator to obtain the corresponding simulated magnetic resonance image. Then, the image quality score of the simulated magnetic resonance image corresponding to the current feasible solution is obtained through the image quality evaluation function. With the goal of maximizing the image quality score, the gradient descent optimization algorithm is used to calculate the gradient and update to obtain the next feasible solution. When the iteration termination condition is reached, the feasible solution obtained by the last gradient update is taken as the optimal solution, and the scanning parameters corresponding to the optimal solution are taken as the optimal scanning parameters of the individual. Preferably, the feasible solution chosen as the starting point of the iteration is the initial scan parameter in S1.

7. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they can realize the personalized magnetic resonance imaging method based on artificial intelligence as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the personalized magnetic resonance imaging method based on artificial intelligence as described in any one of claims 1 to 6.

9. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the personalized magnetic resonance imaging method based on artificial intelligence as described in any one of claims 1 to 6 when executing the computer program.

10. A magnetic resonance imaging device, characterized in that, Includes a magnetic resonance scanner and a control unit; The control unit stores a computer program, which, when executed, controls the magnetic resonance scanner to implement the personalized magnetic resonance imaging method based on artificial intelligence as described in any one of claims 1 to 6, thereby completing a personalized magnetic resonance scan of the target object and obtaining personalized magnetic resonance imaging data.

Images