Single-scanning T2 and T2*multi-parameter magnetic resonance accurate quantification method and system based on multi-echo separation

Abstract

The invention discloses a single-scanning T2 and T2 * multi-parameter magnetic resonance accurate quantification method and system based on multi-echo separation, and relates to magnetic resonance imaging. Designing a magnetic resonance imaging sequence containing a plurality of echo points; loading to a magnetic resonance simulation platform with performance meeting requirements, and acquiring high-quality simulation data under different layer structures of the simulated brain; a deep neural network model fusing multi-echo information is constructed, and training and verification are completed based on the simulation data set; through end-to-end neural network mapping, rapid joint quantification of tissue relaxation parameters such as T2, T2 * and the like is realized, and meanwhile, the synchronous quantification capability on parameters such as M0, deltaB0, B1 and the like is realized. The density and coverage range of different TE echo sampling are improved by utilizing a multi-echo design, and the fitting accuracy of T2 and T2 * is enhanced; the non-ideal factors such as nonuniformity of a main magnetic field delta B0, nonuniformity of a radio frequency field B1, image distortion and artifacts are modeled and corrected, the scanning time is short, the quantitative precision is high, and the anti-interference capability is high.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging, and in particular to a precise quantitative method and system for single-scan T2 and T2* multi-parameter magnetic resonance imaging based on multiple echo separation. Background Technology

[0002] Magnetic resonance imaging (MRI) is a non-invasive, radiation-free imaging technique widely used in clinical medicine and neuroscience. Traditional MRI images are primarily structural images, offering limited quantitative information on the microscopic properties of tissues. In recent years, quantitative MRI (qMRI) has become a research hotspot, providing spatial distributions of physiological parameters such as T2 relaxation time and T2* relaxation time. This offers more objective and reproducible quantitative indicators for early detection of tissue changes and disease classification. Examples include detecting iron deposition in the brain during neurodegenerative diseases and assessing edema in musculoskeletal disorders.

[0003] However, traditional T2 and T2* quantitative methods rely on multiple scans or pixel-by-pixel fitting based on a single physical model, which has problems such as long scan time, susceptibility to noise interference, and poor quantitative stability. They are difficult to meet the requirements of real-time multi-parameter quantitative imaging, which to some extent restricts their widespread application in clinical settings.

[0004] To address these challenges, several fast T2 quantitative magnetic resonance imaging (MRI) methods based on echo stacking have been proposed, such as sequence design combining echo-plane imaging (EPI) readout strategies (Cai C, Wang C, Zeng Y, et al. Single‐shot T2mapping using overlapping‐echo detachment planar imaging and a deep convolutional neural network[J]. Magnetic Resonance in Medicine,2018, 80(5)). However, most of these methods only support quantitative estimation of a single magnetic resonance parameter (such as T2) and lack support for multi-parameter joint imaging. Furthermore, they typically do not explicitly model magnetic field inhomogeneity during sequence design and data processing, making them prone to image distortion and signal shift under conditions of ΔB0 and field inhomogeneity, thus affecting image reconstruction quality and increasing the uncertainty of quantitative estimation.

[0005] In recent years, multi-parameter quantitative magnetic resonance imaging (MRI) methods based on multiple overlapping-echo detachment (MOLED) have gradually emerged (Ma L, Wu J, YQ, et al. Single-shot multi-parametric mapping based on multiple overlapping-echo detachment (MOLED) imaging[J]. NeuroImage, 2022, 263), becoming an important research direction for improving imaging efficiency and expanding parameter dimensions. This type of method, through precise pulse sequence design, enables the acquisition of echo signals at multiple time points after a single excitation, achieving signal intensity capture at multiple echo times (TE). This allows for the simultaneous estimation of tissue relaxation parameters such as T2 and T2* in a single scan, significantly improving the time efficiency and multi-parameter parallel capability of quantitative magnetic resonance imaging.

[0006] Existing typical MOLED methods usually acquire two sets of k-space data, each encoding four echo points under different TE values, for a total of eight TE-weighted echo points. Due to the small intervals between short TE echoes and the relatively high signal intensity in the short TE acquisition strategy, short TE images exhibit good stability and spatial consistency during registration and reconstruction. However, the signal in the long TE segment suffers from significant attenuation, resulting in a reduced signal-to-noise ratio (SNR). Furthermore, it is susceptible to image blurring, artifact enhancement, and spatial misalignment due to echo overlap and gradient switching. This leads to error accumulation when fitting T2 and T2* in the long TE region, thus affecting the accuracy and consistency of the final quantitative map.

[0007] In addition, although the MOLED method has advantages in saving scanning time and improving parameter coverage, it has a limited number of echo points and insufficient TE distribution density. In particular, it fails to provide sufficient information support in the critical long TE region, which restricts the accurate modeling and fitting of the slow relaxation process of tissue.

[0008] Therefore, the current MOLED method still faces several technical bottlenecks: how to further increase the number of echo points and enhance the TE coverage, especially in the long TE region, to provide richer and higher quality signal information, so as to improve the overall image quality and multi-parameter quantitative accuracy, without significantly extending the scanning time, remains the core problem that needs to be solved in this field. Summary of the Invention

[0009] The primary objective of this invention is to overcome the technical bottlenecks of existing multi-echo quantitative imaging methods, such as uneven TE distribution, low signal-to-noise ratio in long TE segments, and insufficient fitting accuracy. It provides a precise quantitative method for T2 and T2* multi-parameter magnetic resonance imaging based on multi-echo separation in a single scan. This method achieves joint estimation of multiple tissue relaxation parameters in a single magnetic resonance scan, significantly improving the temporal efficiency and spatial consistency of parameter mapping. Simultaneously, this method possesses the ability to effectively correct for various imaging non-ideal factors (including ΔB0, B1 inhomogeneity, EPI distortion, and artifacts), enhancing the adaptability and robustness of quantitative magnetic resonance imaging in complex clinical environments.

[0010] The second objective of this invention is to overcome the technical bottlenecks of existing multiple echo quantitative imaging methods, such as uneven TE distribution, low signal-to-noise ratio in the long TE segment, and insufficient fitting accuracy. This invention provides a single-scan T2 and T2* multi-parameter magnetic resonance imaging (MRI) precise quantitative system based on multiple echo separation. By optimizing pulse sequence design and echo timing strategy, this system continuously acquires three independent sets of k-space data in a single scan. The three echo chains each contain five echo points with different TE weights, resulting in a total of 15 TE values ​​covering short, medium, and long TE ranges. This approach reasonably balances the high signal-to-noise ratio of the short TE segment with the sensitivity requirements of the long TE segment to T2 and T2*, achieving a more complete sampling of the tissue relaxation process.

[0011] To achieve the above-mentioned objectives, the present invention provides the following technical solutions.

[0012] This invention provides a precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance imaging based on multiple echo separation, comprising the following steps:

[0013] 1) Pulse sequence design: In one excitation process of a single scan, multiple radio frequency pulses and one 180° refocusing pulse are applied, and three echo chains are used for acquisition. Each echo chain acquires echo signals with 5 or more different echo times (TE), and a total of 15 or more different TE-weighted magnetic resonance echo signals are acquired to cover T2 and T2* relaxation information from short TE to long TE.

[0014] 2) Data acquisition and reconstruction: The k-space data of the multiple echo signals are subjected to two-dimensional Fourier transform to reconstruct the corresponding multi-TE magnetic resonance complex image. The real and imaginary parts of the obtained complex image are extracted to form a six-channel magnetic resonance image tensor.

[0015] 3) Deep model quantization: The tensors of the six-channel magnetic resonance images are input into a pre-trained deep neural network model, and the corresponding T2 quantitative map and T2* quantitative map are output through end-to-end mapping; wherein, the deep neural network model is obtained by training on a simulation dataset. During training, the main magnetic field inhomogeneity ΔB0, radio frequency field inhomogeneity B1 and random noise disturbance are added to the simulation image to improve the robustness of the model to magnetic field inhomogeneity and noise.

[0016] In step 1), among the three echo chains: the first echo chain is acquired immediately after the third radio frequency excitation pulse, including short TE echo signals dominated by spin echo (SE) and gradient echo (GRE) signals formed by each excitation; the second echo chain acquires echo signals under medium TE through a 180° refocusing pulse; the third echo chain acquires long TE echo signals after a preset delay, thereby realizing continuous acquisition of signals in different TE segments of short, medium and long. Each echo chain contains T2-weighted and T2*-weighted echo signals. The TE covered by the T2-weighted signal has significant multi-TE and wide-range characteristics, which is consistent with the characteristic that the tissue T2 value is greater than the T2* value.

[0017] Each echo train synchronously contains T2-weighted signals and T2-weighted signals, and the TE covered by the T2-weighted signals has multiple TEs and a wide range, which is suitable for the physical characteristics of tissues with T2 values ​​greater than T2 values.

[0018] The triple echo chain all employs an echo plane imaging (EPI) readout strategy, and all excitation and echo acquisition are completed within a single repetition time (TR).

[0019] The pulse sequence design also includes optimizing and determining key parameters, such as excitation angle, echo interval, shift gradient intensity, imaging field of view, and imaging matrix size.

[0020] Step 2) further includes a step of normalizing the six-channel image tensor, wherein the normalization process is to scale the image intensity values ​​to the [0,1] interval.

[0021] In step 3), the deep neural network model is constructed based on a conventional network structure, such as the U-Net architecture, or the ResNet or Transformer architecture. The deep neural network model achieves rapid joint quantification of T2 and T2* tissue relaxation parameters through end-to-end mapping. No special design is required, and good results can be achieved after reasonable adaptation. The network model input is a six-channel magnetic resonance image tensor, and outputs the T2 quantitative map and T2* quantitative map, as well as the corresponding proton density M0 map, main magnetic field inhomogeneity ΔB0 map and radio frequency field inhomogeneity B1 map.

[0022] Modeling of non-ideal factors also includes the simulation of common artifacts in magnetic resonance imaging. The model uses this modeling to correct image distortion and artifacts, thereby improving the quality of quantitative images.

[0023] In step 3), the training process of the deep neural network model satisfies:

[0024] The loss function uses the pixel-wise mean square error (MSE).

[0025] The Adam optimizer was selected, with an initial learning rate of 0.0001 and a cosine annealing decay strategy.

[0026] Configure L2 weight decay, dropout regularization mechanism and early stopping strategy to prevent model overfitting.

[0027] In step 3), the process of generating the simulation dataset is as follows: the pulse sequence designed in step (1) is loaded into the magnetic resonance imaging simulation platform (SMRI), and the 3D simulated human brain template is imaged and simulated. During the simulation, non-ideal factors such as non-ideal radio frequency excitation contour and non-ideal gradient are introduced to obtain k-space complex data containing multiple echoes.

[0028] In step 3), when quantitatively reconstructing the actual magnetic resonance scan data, the actual k-space data is preprocessed according to the same procedure as in step 2) to ensure that the input image tensor format is consistent with the training samples.

[0029] This invention provides a single-scan T2 and T2* multi-parameter magnetic resonance precise quantitative system based on multiple echo separation, comprising:

[0030] The sequence design module designs the parameters of the multiple overlapping echo magnetic resonance pulse sequence based on the characteristics of the tissue to be tested, and generates a sequence scheme to acquire multiple sets of different TE echo signals in a single scan.

[0031] The simulation data generation module is used to load the sequence scheme into the magnetic resonance simulation platform to simulate the acquisition of multi-echo k-space data containing effects such as non-ideal radio frequency excitation profile, non-uniformity of main magnetic field ΔB0, and non-uniformity of radio frequency field B1.

[0032] The data preprocessing module is used to convert the acquired or simulated k-space data into a complex image in the image domain, separate the real and imaginary parts to form a multi-channel image tensor, and perform normalization and noise addition on the image tensor to obtain standardized network input data.

[0033] The neural network training module is used to train a deep neural network model using the multi-channel image tensor generated by simulation and its corresponding T2 and T2* label maps, and to establish a nonlinear mapping relationship between the echo signal and the T2 and T2* parameters.

[0034] The quantitative image reconstruction module is used to input the multi-channel magnetic resonance image data obtained from actual scanning into a trained deep neural network model and output the T2 quantitative map, T2* quantitative map, proton density M0 map, main magnetic field inhomogeneity ΔB0 map and radio frequency field inhomogeneity B1 map of the imaging object.

[0035] Furthermore, the sequence design module sets up a sequence structure of multiple radio frequency excitation pulses and a single refocusing pulse, and reasonably designs the excitation pulse angle, shift gradient magnitude, and excitation pulse interval to form three sets of echo chains with at least five TE points in a single repetition time. Each echo chain contains echoes of T2 and T2* weighted signals with different TEs, rather than containing only the same echo signal.

[0036] Furthermore, the sequence design module sets up a sequence structure of three radio frequency excitation pulses and one 180° refocusing pulse. By optimizing the excitation pulse angle, the shift gradient magnitude, and the excitation pulse interval, each echo chain contains T2 and T2* weighted signals with different TE values.

[0037] Furthermore, the quantitative image reconstruction module is also used to simultaneously output the distribution images of the M0, ΔB0, and B1 parameters of the imaging object, in order to correct the influence of magnetic field inhomogeneity on imaging.

[0038] Compared with the prior art, the technical effects and outstanding advantages of the present invention are as follows:

[0039] 1. High Acquisition Efficiency: The synergistic optimization of acquisition efficiency and TE coverage breaks through the bottleneck of traditional parameter sampling. This invention achieves the acquisition of 15 or more TE-weighted echo points in a single scan through an innovative sequence structure of 3 excitation pulses + 1 refocusing pulse, avoiding multiple repeated scans. Compared with the traditional MOLED method of 8 echo points, it achieves nearly double the information density. Moreover, the TE distribution uniformly covers the entire range of short (e.g., 10ms), medium, and long (e.g., 180ms) intervals, especially strengthening the signal acquisition capability of the long TE segment. It solves the core problems of insufficient sampling in the long TE region and weak fitting information support in existing technologies. At the same time, it takes into account the high signal-to-noise ratio of the short TE segment and the sensitivity of the long TE segment to the T2 / T2* relaxation characteristics, providing a more complete signal foundation for accurate fitting.

[0040] 2. Significantly improved quantitative accuracy through multi-parameter combined analysis, while considering correction and adaptability: Each echo train in this invention synchronously includes T2-weighted (SE-dominated) and T2-weighted (GRE-dominated) signals, avoiding the attenuation loss caused by the concentration of T2 signals in the long TE segment in traditional methods, and significantly improving the initial signal-to-noise ratio of the T2 signal; combined with an end-to-end deep neural network, it not only achieves synchronous quantification of T2 and T2, but also outputs parameter maps of M0, ΔB0, and B1, and explicitly corrects for non-ideal factors such as magnetic field inhomogeneity and radio frequency field distortion, solving the problems of quantitative error accumulation and poor adaptability to complex clinical environments in existing methods. Experimental verification shows that the Pearson correlation coefficients of T2 and T2* quantitative results reach 0.954 and 0.933, respectively, which are better than the 0.938 and 0.929 of traditional methods.

[0041] 3. Significantly enhanced robustness and imaging stability: The end-to-end mapping of the network improves its adaptability to complex tissue structures and non-ideal magnetic field disturbances. On the one hand, the sequence design adopts the EPI fast readout strategy, and all excitation and echo acquisition are completed within a single TR, effectively reducing motion artifacts and systematic errors. On the other hand, the deep neural network is trained with 3D simulation data containing ΔB0, B1 perturbations and noise modeling, which greatly improves its ability to resist interference from non-ideal factors in actual clinical scanning, avoids the defects of traditional pixel-by-pixel fitting that is easily affected by noise, and ensures the consistency and reliability of quantitative results in different tissue regions (including lesions and edge structures).

[0042] 4. Simple engineering implementation and strong promotional value: This invention requires no modification to mainstream MRI hardware systems. Sequence parameters can be flexibly configured through existing device interfaces. Data preprocessing and network training processes are standardized and compatible with common deep learning platforms (such as PyTorch and TensorFlow). Furthermore, the efficient single-scan acquisition mode meets clinical needs for rapid imaging, avoiding the problems of high patient compliance requirements and high scanning costs associated with multiple scans. This provides a feasible path for the large-scale clinical application of qMRI technology. The standardized process of this invention is compatible with mainstream MRI systems and deep learning platforms, demonstrating good scalability. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the process of the magnetic resonance multi-parameter synchronous quantitative imaging method described in this invention.

[0044] Figure 2 This is a schematic diagram comparing the structure of the present invention with that of a traditional multi-echo sequence. Among them, (a) is a schematic diagram of the multi-parameter magnetic resonance imaging sequence structure of the present invention based on multiple overlapping echoes; (b) is a schematic diagram of the structure of a traditional multi-echo sequence, used to compare the advantages of the present invention in terms of echo superposition strategy and multi-parameter sensitivity from the perspective of pulse sequence design.

[0045] Figure 3 The above is a flowchart of the overall process of the proposed multi-parameter quantitative magnetic resonance imaging system.

[0046] Figure 4 The images show a comparison of the k-space data and corresponding magnetic resonance images of the simulated samples. (a) shows the image data of the simulated sample in k-space and its corresponding magnetic resonance image; (b) shows a comparison of the k-space data and corresponding magnetic resonance images of the simulated sample in the comparative experiment.

[0047] Figure 5 This is a schematic diagram comparing multi-parameter quantitative magnetic resonance imaging (MMRI). (a) shows a multi-parameter MMRI image obtained after quantitative reconstruction of the simulated sample, including important parameters such as T2 and T2*; (b) shows a multi-parameter MMRI image generated from the simulated sample in the comparative experiment, used to verify the advantages of this invention in quantitative accuracy.

[0048] Figure 6 The graphs show the correlation analysis between the quantitative results and the reference values. Specifically, (a) is the correlation analysis between the magnetic resonance quantitative results obtained by the method of this invention and the reference values, used to evaluate the quantitative accuracy of the model; and (b) is the correlation analysis of the quantitative results obtained from the comparative experiment, serving as a basis for performance comparison. Detailed Implementation

[0049] To more comprehensively and clearly illustrate the technical solution of this invention, the principles and implementation paths of this invention will be systematically explained below in conjunction with the accompanying drawings of the embodiments. It should be noted that these embodiments are intended to demonstrate the core ideas and technical advantages of this invention and are not intended to limit the scope of protection of this patent. For those skilled in the art, non-inventive extensions and technical substitutions made under the guidance of this invention, such as introducing preprocessing modules like inversion recovery, fat suppression, or magnetization transfer before the imaging sequence, or performing conventional cyclic adjustments to the sequence structure, should all be considered reasonable extensions of the technical concept of this invention. Their technical effects and core objectives have not been substantially changed, and therefore should still be included within the scope of protection of this invention.

[0050] This invention aims to address the problems of uneven TE distribution of different weighted signals, severe signal attenuation in long TE segments, and insufficient stability of quantitative accuracy in existing multi-echo quantitative magnetic resonance imaging (MRI) methods. It proposes a method and system for quantitative analysis of multiple parameters (T2, T2*) in a single scan based on multiple overlapping echoes. This method acquires wide-coverage, multi-weighted multi-echo signals from short to long TE in a single scan through precise design of the MRI pulse sequence and optimized acquisition strategy. Furthermore, it utilizes an end-to-end deep neural network to achieve rapid and high-precision joint quantification of parameters such as T2 and T2*. Simultaneously, it can correct for the influence of non-ideal factors such as main magnetic field inhomogeneity (ΔB0) and radio frequency field inhomogeneity (B1) on the image, thereby significantly improving the accuracy and robustness of parameter mapping.

[0051] Combination Figure 1 As shown, the multi-parameter quantitative imaging method proposed in this invention mainly includes the following steps:

[0052] S1: Design and construct a multi-parameter magnetic resonance imaging sequence with multiple echo train structure and determine the imaging parameters.

[0053] A magnetic resonance imaging pulse sequence supporting multiple overlapping echo acquisition is designed, providing more reasonable TE coverage for different weighted signals. Three echo chains are continuously acquired in a single excitation scan. Each chain contains several echo signals with different echo times (TE), and each echo chain includes T2 and T2* weighted signals, resulting in a total of more than 15 echo points. The sequence employs a structure of three or more radio frequency excitation pulses and a 180° refocusing pulse, forming a three-segment echo chain acquisition: the first echo chain is acquired immediately after the third excitation, containing spin echo and gradient echo signals at short TE; the second echo chain acquires spin echo and gradient echo signals at medium TE through the refocusing pulse; and the third echo chain acquires spin echo and gradient echo signals at long TE after a longer delay. This design achieves continuous and balanced coverage from short TE to long TE, balancing the high signal-to-noise ratio of the short TE segment with the sensitivity to T2 and T2* signals in the long TE segment.

[0054] S2: Three-dimensional multi-echo magnetic resonance simulation sampling based on a simulation platform.

[0055] The designed multiple overlapping echo sequence was loaded into a magnetic resonance imaging (SMRI) simulation platform to perform imaging simulation on a virtual 3D simulated human brain template. During the simulation, non-ideal factors from actual MRI scans (such as non-uniformity of the main magnetic field ΔB0, non-uniformity of the radio frequency field B1, and non-ideal gradient) were introduced to obtain complex k-space data containing the multiple echoes.

[0056] S3: Preprocess the simulation data to construct a training sample set.

[0057] The simulated multi-TE magnetic resonance images are preprocessed. First, the k-space data is reconstructed into image-domain multi-overlapping echo-weighted magnetic resonance images using a two-dimensional Fourier transform, separating the real and imaginary parts. The three echo chain images together form a six-channel image tensor. Then, the multi-channel image data is normalized, and Gaussian noise can be superimposed to simulate noise interference in real-world imaging, thereby enhancing the diversity of training samples and the robustness of the model. Simultaneously, based on the pre-set physical parameters such as organization T2, T2*, M0, ΔB0, and B1 during simulation, corresponding standard quantitative label images are generated as targets for supervised learning.

[0058] S4: Train a deep neural network model.

[0059] An end-to-end deep neural network model based on the U-Net architecture was constructed for multi-parameter quantitative image reconstruction. The preprocessed six-channel image tensor was used as the network input, and the output consisted of T2 and T2* quantitative images (and optional M0, ΔB0, and B1 images). During training, the pixel-wise mean square error (MSE) was used as the loss function, and the network parameters were iteratively optimized using a simulation dataset. Through supervised learning on simulation data, the network learned the nonlinear mapping relationship between multi-echo image signals and quantitative parameters such as T2 and T2*, thereby converting multi-overlapping echo amplitude signals into corresponding quantitative parameter values.

[0060] S5: Quantitative Image Reconstruction and Parameter Estimation.

[0061] The trained neural network was deployed to actual magnetic resonance imaging (MRI) scan data. A two-dimensional Fourier transform and preprocessing consistent with the training phase were performed on the K-space data acquired from the actual scanner to construct a six-channel image tensor input network. Subsequently, forward inference of the deep network was performed to obtain the corresponding T2 and T2* images.

[0062] The following is a detailed description of each of the above steps:

[0063] S1: Design and construct a multi-parameter magnetic resonance imaging sequence with a multiple echo train structure.

[0064] Figure 2 As shown in (a): The imaging sequence structure designed in this invention: This sequence is based on three echo trains, each containing five echo points. Wherein: β i Represents the i-th refocusing pulse; α j This represents the j-th radio frequency excitation pulse, whose flip angle is adjustable, supporting different phases and pulse shapes; δ j G represents the time interval between each pulse or echo point, or between a pulse and an echo point; SS The layer / segment gradient of the layer selection dimension is represented by the application of G simultaneously with the application of the radio frequency excitation pulse.ss G RO1 ~G RO4 G is the shift gradient on the frequency coding dimension; PE1 ~G PE4 The shift gradient is the phase-encoded dimension;

[0065] 1.1 Sequence Design and Excitation Structure Optimization

[0066] Unlike the traditional excitation structure consisting of "4 excitation pulses + 1 refocusing pulse" in the prior art (which typically acquires two k-spaces through two echo chains, where the first echo chain is SE-EPI to acquire T2-weighted information and the second echo chain is GRE-EPI to acquire T2*-weighted information), the method proposed in this invention adopts a novel structural design of 3 excitation pulses + 1 refocusing pulse, which significantly enhances the coverage and information redundancy of the TE distribution.

[0067] 1.2 Multi-echo chain acquisition strategy

[0068] Based on this composite excitation structure, three excitation pulses are applied sequentially, with the third pulse having a larger angle (up to 120°), which not only has an excitation effect but also provides a certain degree of refocusing. Three echo trains are designed to sequentially acquire magnetic resonance signals from different TE points:

[0069] First echo train: Acquired immediately after the third excitation pulse, forming 5 echo points. Two of them are SE echo points symmetrical to the third excitation pulse, mainly weighted by T2; the other three are GRE echo points formed by the three excitation pulses themselves, mainly reflecting T2* attenuation information.

[0070] Second echo train: After the first echo train acquisition is completed, a 180° refocusing pulse is applied without re-excitation to refocus the first five echo trajectories, thereby acquiring a second set of five echo points with T2-dominant weighting. In this stage, the original three GRE echo points simultaneously carry both T2 and T2* weighted components, further enriching the information required for joint estimation.

[0071] The third echo train: after an appropriate time delay, a new set of T2*-weighted echo points are acquired using GRE method to expand the TE coverage, especially enhancing the modeling capability for long TE signals.

[0072] This three-segment acquisition structure achieves a multi-echo acquisition layout with continuous time, complementary characteristics, and reasonable TE distribution, which is significantly better than the traditional dual-chain acquisition strategy.

[0073] 1.3 Timing and Acquisition Mode Design

[0074] Each echo train employs a fast EPI acquisition method to shorten scan time and reduce motion sensitivity. The entire sequence completes all excitation and echo acquisition within one TR (repetition time), effectively improving imaging efficiency and avoiding systematic errors and mismatches caused by multiple excitations. This structure particularly enhances signal acquisition and utilization capabilities in the long TE segment, while maintaining a high signal-to-noise ratio and image stability in the short TE segment, providing balanced and reliable signal support for subsequent quantitative reconstruction.

[0075] The settings for imaging sequence sampling parameters include: the number of radio frequency excitation pulses k; and the values ​​for each α... j and β i Magnitude, phase, and pulse shape; inter-node time interval δ j G SS Area and direction; timing, direction and area of ​​each shift gradient; imaging field of view (FOV), matrix size, parallel factor and other acquisition parameters.

[0076] 1.4 Optimization of Excitation Angle

[0077] To accurately estimate the T2 and T2* images, a constrained optimization framework based on minimizing the weighted Cramér-Rao lower bound (CRLB) is used to determine the optimal excitation flip angle. , and The β flip angle is fixed at 180°. Specifically, for the first two excitation angles... , Optimize, and the last incentive angle The angle is then fixed at 120° to fully utilize the remaining longitudinal magnetization and achieve effective spin refocusing.

[0078]

[0079] in, A trapezoidal weighting function is used to emphasize clinically significant T2 and T2* parameter ranges, specifically T2 (30–200 ms) and T2* (20–150 ms). These parameter ranges are chosen to ensure the optimization process is robust and representative in real-world MRI applications.

[0080]

[0081] in, This represents the relative signal strength of the i-th echo. The first constraint ensures that the signal strength of all echoes is higher than a preset threshold ( ). The threshold was experimentally verified to be 0.05 to ensure a sufficient signal-to-noise ratio (SNR). The second constraint is used to maintain signal strength balance among the echoes to improve overall image quality and quantitative accuracy.

[0082] (0°~90°) and The range of values ​​(0°~90°) was set based on preliminary assessments and physiological factors, aiming to achieve a balance between magnetization utilization efficiency and signal stability. Ultimately, the optimal parameters were determined using the CRLB optimization method: =45°, = 45°, all located in the middle of their optimal value range.

[0083] S2: Three-dimensional multi-echo magnetic resonance simulation sampling based on a simulation platform

[0084] To overcome the problems of high cost, uncontrollable quality, and limited structural coverage in acquiring real magnetic resonance imaging (MRI) samples, this invention introduces three-dimensional (3D) simulation data to replace traditional two-dimensional (2D) simulation data. 3D simulation fully considers the pulse contour features and non-ideal factors in actual scanning, ensuring that the generated data possesses both structural realism and physical consistency, significantly improving the effectiveness and robustness of deep network training. The specific implementation process includes the following steps:

[0085] 2.1 Explanation of the Construction and Source of Two-Dimensional Templates

[0086] Two-dimensional image templates with structural differences were constructed to simulate real magnetic resonance imaging (MRI) scans. These templates can be obtained in two ways: firstly, by generating random structures based on statistical models and parameter distribution rules; and secondly, by directly extracting two-dimensional slice images from actual MRI datasets.

[0087] 2.2 Construction of 3D Simulation Model

[0088] Although the template is a two-dimensional image, the three-dimensional simulation, by introducing shape pulse files from the magnetic resonance system, controls the excitation layer thickness and excitation position, enabling the template to be scanned layer by layer in three-dimensional space. The simulation platform treats the two-dimensional template as a series of slices with thickness and sequentially applies magnetic resonance excitation and acquisition processes to obtain a three-dimensional weighted magnetic resonance signal response that simulates the actual scanning process.

[0089] 2.3 Simulation of Sequence Loading and Echo Chain Acquisition

[0090] The triple echo chain structure sequence designed in this invention was loaded onto a magnetic resonance simulation platform (such as SMRI), and a complete scan of each template was performed. The entire sampling process simulated the acquisition of echo points with different TE weights in a single scan, covering multiple signal attenuation stages from short TE to long TE, realizing the acquisition of echo signals at multiple time points in a single acquisition. The simulation results fully reflect the superposition effect of the excitation pulse on different slice layers in the experiment.

[0091] S3: Preprocess the simulation data to construct a training sample set.

[0092] To ensure the quality, robustness, and diversity of deep neural network training data, this invention designs a complete data preprocessing workflow after obtaining the initial simulation images. This workflow optimizes the training sample set from multiple dimensions, including data structure transformation, noise modeling, patch extraction, and label standardization. Specifically, it includes the following aspects:

[0093] 3.1 Image Domain Transformation and Channel Construction

[0094] The simulated K-space signals are subjected to two-dimensional inverse Fourier transforms to obtain complex images in the image domain. Each set of images is split into two channels, real and imaginary, which constitute the input to the deep learning network.

[0095] 3.2 Quantitative Label Generation and Registration

[0096] Based on the original parameters (such as T2 and T2*) in the simulation template, corresponding quantitative label images are automatically generated as the target output of the network's supervised learning. If there are slight displacements or scale errors between different TE-weighted images, rigid or non-rigid registration between the images is required to ensure strict alignment of the labels with the input data in spatial location.

[0097] 3.3 Noise Simulation and Robustness Enhancement

[0098] To enhance the network's ability to cope with signal-to-noise ratio changes in actual MRI acquisition, this invention actively adds various Gaussian or Rayleigh noise models to the simulation data in the image domain. This simulates the image quality degradation caused by signal fluctuations, coil interference, or external environmental noise during actual scanning, thereby improving the robustness of the model in clinical applications.

[0099] 3.4 Normalization Process

[0100] All input images are normalized to ensure uniform data distribution across channels and prevent model convergence from being affected by differences in dynamic range.

[0101] S4: Train a deep neural network model.

[0102] After constructing the training samples, the process proceeds to the deep neural network training stage proposed in this invention. This stage aims to learn the nonlinear mapping relationship between different TE-weighted images and target quantitative parameters (T2 and T2*, etc.) in the simulation data, achieving high-precision reconstruction from multi-channel images to quantitative maps. The training process mainly includes the following technical steps:

[0103] First, design the overall structure of the deep neural network, clarify the type, number and connection method of each network layer, and ensure that the network has sufficient expressive power to cope with complex multi-parameter reconstruction tasks;

[0104] Secondly, based on the data characteristics, the number of input channels and output channels of the network are determined to achieve functional matching from multi-channel complex image input to target quantitative parameter map output;

[0105] Then, select an appropriate loss function and optimization strategy as the basis for parameter adjustment and performance optimization during model training;

[0106] During network training, the constructed training sample set is input into the neural network in batches, the backpropagation algorithm is executed, the network parameters are automatically adjusted to reduce the loss function value, and the network performance is continuously optimized in the iteration.

[0107] As training progresses, the network convergence is continuously monitored. Training is considered complete when the loss function stabilizes and reaches the expected performance standard.

[0108] Finally, the parameters of the trained neural network model are saved to provide core support for the subsequent inference and reconstruction stages of multi-parameter quantitative magnetic resonance images.

[0109] S5: Quantitative Image Reconstruction and Parameter Estimation.

[0110] After training the neural network model, the trained model can be applied to the quantitative image reconstruction stage of actual magnetic resonance data. The goal of this step is to obtain corresponding quantitative parameter images such as T2 and T2* by inputting the actual acquired image domain data, achieving efficient and accurate multi-parameter reconstruction. The process is as follows:

[0111] First, the raw k-space data obtained during the scanning process is received and processed, and then converted to the image domain to generate a complex image containing multiple TE weighting information.

[0112] Secondly, the converted image data is subjected to necessary normalization and complex decomposition to ensure that its format matches the network input requirements, including separation of real and imaginary parts and scale adjustment.

[0113] Then, the preprocessed image is input into the trained neural network model by channel to perform forward inference operations;

[0114] The neural network automatically outputs the corresponding quantitative parameter image results such as T2 and T2* based on the mapping relationship it has learned;

[0115] Finally, the network output is post-processed (such as limiting the value range, image restoration, coordinate alignment, etc.) to obtain a multi-parameter quantitative image with physical meaning and spatial consistency.

[0116] This reconstruction process is characterized by speed, automation, and high precision, and is significantly superior to traditional multi-sequence, multi-scan quantitative imaging schemes. It reduces acquisition time while improving the quantitative accuracy and spatial consistency of images.

[0117] like Figure 2 As shown in (a) above, this is a schematic diagram of the multiple overlapping echo sequence structure proposed in this invention; Figure 2 (b) shows a schematic diagram of a traditional multiple echo sequence structure used for comparative experiments. Analysis of the key technical parameters of the two sequences clearly demonstrates the significant advantages of this invention in signal acquisition efficiency, parameter coverage, and signal-to-noise ratio.

[0118] (1) Analysis of the number of echo points and TE distribution

[0119] The sequence described in this invention collects more echo points under each excitation, specifically using three or more echo chains for sampling. Compared to the layout of only two echo chains and four echo points in the comparison sequence, it achieves denser and more continuous TE sampling in both spatial and temporal dimensions.

[0120] This design allows for a wider distribution of TE, covering multiple time scales from short TE (e.g., 10 ms) to long TE (e.g., 180 ms), while also ensuring a more uniform distribution. This facilitates joint fitting of multiple parameters of T2 and T2*, improving the robustness of model training and inversion.

[0121] (2) Advantages of signal-to-noise ratio analysis and signal acquisition strategies

[0122] exist Figure 2 In the comparison sequence shown in (b), the T2* weighted echo is mainly concentrated in the second echo chain, that is, it is acquired at a later time after multiple T2 weighted attenuations. At this time, the tissue signal intensity drops significantly, resulting in a low signal-to-noise ratio of the T2* signal, which is easily affected by background noise and system instability, thus reducing the quantitative accuracy.

[0123] In contrast, the present invention Figure 2The sequence shown in (a) achieves synchronous coverage of T2-weighted and T2*-weighted signals in the first echo train, with some T2*-weighted echo points even acquired before the T2-weighted echo points. This structural design significantly reduces the secondary loss of the T2* signal caused by T2 attenuation, while improving the initial signal-to-noise ratio of the T2* signal itself, resulting in higher quantization reliability and image clarity. Furthermore, since the T2-weighted and T2*-weighted signals are acquired in a coordinated manner within the same echo train, sequence structure splits and timing waste are avoided, improving overall acquisition efficiency and sequence stability.

[0124] (3) Coverage advantages of T2 signal acquisition

[0125] The echo time layout designed in this invention covers a wider range in the T2 direction, encompassing tissue characteristics from short T2 to long T2, while avoiding the fitting instability problem caused by sparse or clustered echo points in traditional sequences. The higher density TE coverage also improves the sampling capability for long-tailed T2 signals, enhancing the quantitative discrimination performance for complex structures such as lesion areas and tissue edges.

[0126] In summary, Figure 2 The sequence of the present invention shown in (a) is compared to Figure 2 The traditional sequence (b) in the middle has a series of technical advantages, such as more echo points, more scientific and reasonable TE distribution, optimized T2 and T2* acquisition mechanism, and higher signal-to-noise ratio performance, which can significantly improve the model fitting accuracy and final imaging quality in quantitative magnetic resonance imaging.

[0127] Figure 3 This diagram illustrates the overall flowchart of the single-scan T2 and T2* multi-parameter magnetic resonance precise quantitative system based on multiple echo separation according to the present invention. This system, based on multiple overlapping echo technology, aims to achieve synchronous high-precision quantitative analysis of parameters such as T2 and T2* in a single scan. The overall system structure includes the following functional modules:

[0128] The sequence design module is responsible for designing magnetic resonance pulse sequences with multiple overlapping echo structures. It provides a flexible parameter configuration interface, supporting the adjustment of key parameters such as excitation angle, echo interval, gradient waveform, and TE distribution to generate single-scan acquisition schemes with rich TE weighting characteristics, meeting the requirements of multi-parameter synchronous quantitative imaging.

[0129] The simulation data generation module loads the designed pulse sequence into the magnetic resonance imaging simulation platform. By setting physiological parameters of different tissues (including T2, T2*, M0, ΔB0, B1, etc.) and introducing non-ideal factors such as magnetic field inhomogeneity in ΔB0 and B1, it improves the realism and generalization ability of the simulation data. After simulating the real acquisition process, it outputs k-space complex signal data containing multiple echo information. The simulation data generation module performs a two-dimensional Fourier transform on the k-space data to convert it into an image domain signal. Then, it separates the real and imaginary parts to form six channels and performs normalization processing. At the same time, it introduces appropriate noise simulation to enhance data diversity and network robustness. Finally, it generates a standardized multi-channel image tensor as the network input. Meanwhile, based on the above-mentioned tissue physical parameters preset in the simulation process, it generates corresponding standardized quantitative parameter images as reference labels in the supervised training process of the deep neural network.

[0130] The data preprocessing module performs a two-dimensional Fourier transform on the k-space data obtained from the simulation, converting it into image domain data. It also performs real and imaginary part channel separation and normalization on the complex image. To enhance the diversity and robustness of the training data, this module can also add a noise simulation process, ultimately outputting a standardized multi-channel image tensor as input for subsequent network training.

[0131] The neural network construction and training module is responsible for constructing the deep neural network structure required for multi-parameter quantitative image reconstruction. It explicitly accepts six-channel image input and outputs T2 and T2* quantitative maps (and optional M0, ΔB0, and B1 images), defining the type and connection methods of each layer. Supervised learning is performed using the standardized training samples to establish a nonlinear mapping relationship from image signals to T2 and T2* quantitative maps. During training, a loss function guides parameter optimization, continuously improving model parameters to enhance reconstruction accuracy and stability.

[0132] The quantitative image reconstruction module deploys the trained neural network onto actual magnetic resonance imaging (MRI) scan data. It performs a two-dimensional Fourier transform and preprocessing consistent with the training phase on the K-space data acquired from the actual scanner, constructing a six-channel image tensor input network. Subsequently, forward inference of the deep network is performed to obtain the corresponding T2 and T2* quantitative images. The reconstruction results can be used for further medical image analysis or clinical diagnosis, and support quantitative error assessment and model generalization performance validation.

[0133] The following is a specific implementation example. This method, based on a specially designed magnetic resonance pulse sequence, achieves high-precision T2 and T2* parameter estimation through three-dimensional simulation sample generation, high-precision neural network training, and quantitative image reconstruction. The overall process includes the following steps:

[0134] Step 1: Design of Multiple Overlapping Echo Pulse Sequences

[0135] This step is designed to achieve simultaneous quantitative T2 and T2* magnetic resonance imaging pulse sequences in a single scan. Its structural diagram is shown below. Figure 2 As shown in (a) above. The key parameter configurations are as follows:

[0136] Radio frequency excitation pulse: Sinc pulse, with the flip angle set to α1=α2=45°;

[0137] Dual-function pulse (excitation and focusing): Sinc pulse, flip angle α3 = 120°;

[0138] Refocusing pulse: Sinc pulse, flip angle β1 = 180°;

[0139] Timing interval parameter (unit: ms):

[0140] δ1=13.4345, δ2=11.9564, δ3=4.478175, δ4=14.45635, δ5=6;

[0141] Imaging field of view (FOV): 220 mm × 220 mm;

[0142] Imaging matrix size: 156 × 156;

[0143] Shift gradient settings:

[0144] Frequency coding direction: G ro1 =-16 / 128, G ro2 =-32 / 128, G ro3 =64 / 128, G ro4 =-1

[0145] Phase encoding direction: G pe1 =48 / 128, G pe2 =-32 / 128, G pe3 =-64 / 128, G pe4 =-1;

[0146] Layer selection gradient (G) SS ): Applied during radio frequency excitation to achieve layer thickness control and slice positioning.

[0147] Step 2: Simulation generation of training samples

[0148] 2.1 Simulation Environment Construction

[0149] The magnetic resonance pulse sequence designed above was imported into the magnetic resonance imaging simulation platform SMRI, a custom-shaped pulse waveform file was loaded, and non-ideal factors in actual imaging, such as excitation angle error, gradient mismatch, field strength inhomogeneity (ΔB0 / B1 perturbation) and system noise, were considered to construct a simulation environment that approximates the real imaging conditions.

[0150] 2.2 Template Generation and Modeling

[0151] Two-dimensional image templates were constructed, sourced from randomly generated tissues or publicly available MRI datasets, ensuring sufficient structural complexity to cover common anatomical structures and lesion features. To simulate the sampling effect of three-dimensional simulation, shape pulse files and a slice scanning mechanism were introduced. Specifically, based on the two-dimensional template, a magnetic resonance excitation of a certain slice thickness was simulated. Within this slice thickness, a central slice and four slices above and below it (a total of nine slices) were selected, and virtual scans were performed on each slice with different excitation parameters. The K-space complex signals of the nine slices were then weighted and superimposed to achieve effective simulation of the three-dimensional imaging effect.

[0152] 2.3K Spatial Reconstruction and Tag Generation

[0153] Two-dimensional K-space data is formed by rearranging the complex echo signals, and an image domain image is obtained by performing a two-dimensional inverse Fourier transform. Standardized label maps (T2, T2*, M0, ΔB0, B1) are generated by combining the physical properties of the template.

[0154] Step 3: Data Preprocessing

[0155] To adapt to the input and output characteristics of the neural network and enhance the robustness of network training, the following preprocessing operations are performed on the simulation-generated data:

[0156] Complex image channel splitting: Each group of complex images is split into real and imaginary part channels, forming a total of 6 input channels;

[0157] Data format adjustment: Unify image dimensions to [C, H, W], where C=6;

[0158] Normalization: Perform maximum and minimum value normalization on each input channel to scale the intensity values ​​to [0,1].

[0159] Noise enhancement: Gaussian noise is added to simulate the uncertainty in the data acquisition process, thereby improving the network's generalization ability;

[0160] Training sample construction: The normalized magnetic resonance images and the five-channel label images are combined to form a one-to-one corresponding training sample;

[0161] Dataset partitioning: The training set and validation set are divided in an 80%:20% ratio.

[0162] Step 4: Deep Neural Network Construction and Training

[0163] 4.1 Network Structure Definition

[0164] Deep convolutional networks built on the basis of U-Net have good image feature extraction capabilities and multi-scale context integration capabilities. Other structures (such as ResNet, Transformer, etc.) can also be used as alternatives.

[0165] 4.2 Input / output format:

[0166] Input: A six-channel magnetic resonance imaging tensor, derived from three echo train images (six channels in total, including real and imaginary parts).

[0167] Output: Five-channel quantitative graphs, namely T2, T2*, M0, ΔB0, and B1.

[0168] 4.3 Loss Function Design

[0169] The pixel-wise mean square error (MSE) loss function is used. ,in M represents the network mapping function; M represents the batch size; j represents the j-th sample input data; The label data is represented by W and b, which represent the network weights and biases.

[0170] 4.4 Training Strategies

[0171] Optimizer: Adam optimizer is used;

[0172] Initial learning rate: 0.0001;

[0173] Learning rate decay strategy: Cosine annealing decay is adopted to avoid getting trapped in local optima and help improve convergence speed and accuracy. Every 30 epochs is a decay cycle, and the learning rate is reset when restarting. The minimum learning rate is set to 0.000001.

[0174] Batch size: Set to 8, which can be dynamically adjusted according to GPU memory to ensure training stability and reliable statistical features within the batch;

[0175] Regularization mechanism: L2 weight decay is set to 0.00001 to prevent model overfitting, and dropout is set to 0.2;

[0176] Early stopping strategy: If the validation set loss does not decrease significantly within 10 consecutive epochs, training is terminated early to prevent overfitting and retain the network weights that perform best on the validation set.

[0177] Termination condition: Save the final model after the validation set loss is stable or the early stopping criterion is met.

[0178] Step 5: Quantitative Image Reconstruction

[0179] The magnetic resonance imaging sequence is used to acquire data of the imaging object to obtain the k-space data of the imaging object;

[0180] A two-dimensional Fourier transform is performed on the k-space data of the imaging object to obtain the magnetic resonance image of the imaging object.

[0181] Figure 4 (a) in the figure represents the k-space data and magnetic resonance image of the human brain simulation sample in this invention; each k-space contains several overlapping echoes, and the positions of each overlapping echo are determined by... Figure 4 The shift gradients in the magnetic resonance imaging sequence shown are determined by these gradients. Figure 4 As shown, the magnetic resonance image corresponding to k-space containing overlapping echoes is a striped, aliased image. A comparative experiment was conducted using existing multi-parameter quantitative magnetic resonance imaging methods based on multiple overlapping echoes. Figure 4 (b) in the comparison experiment shows the k-space data and magnetic resonance images of the simulated human brain sample.

[0182] The magnetic resonance images of the simulated human brain sample are input into the deep neural network trained in step 3 for reconstruction, resulting in... Figure 5 The multi-parameter quantitative image of magnetic resonance is shown in (a). Figure 5 Image (a) shows the five quantitative magnetic resonance imaging (MRI) parameters T2, T2*, M0, ΔB0, and B1 of the simulated human brain sample. The images in the first row are the MRI quantitative parameter images reconstructed by the deep neural network, the images in the second row are the corresponding label images, and the images in the third row are the absolute error values ​​between the reconstruction results of the deep neural network and the corresponding label images. Figure 5 Figure (b) shows the reconstructed images of five quantitative magnetic resonance parameters (T2, T2*, M0, ΔB0, and B1) from the comparative experiment. A correlation analysis was performed on the two methods, and their Pearson correlation coefficient (PCC) was calculated. Figure 6 As shown, the correlation slopes of the reconstructed T2 and T2* values ​​obtained by this invention are close to 1, at 0.9113 (T2) and 0.8714 (T2*), respectively, indicating high fitting accuracy. The Pearson correlation coefficients of this invention on the T2 and T2* quantitative parameter images are 0.954 and 0.933, respectively, both higher than the 0.938 and 0.929 of the comparative experiment, thus demonstrating that the quantitative parameter image values ​​of this invention are more accurate and have higher imaging precision.

[0183] The k-space data acquired in this invention is transformed to the image domain after Fourier transform. The complex real and imaginary parts of three sets of images are separated and extracted, and input as 6-channel image tensors into the neural network model. Simultaneously, based on pre-set physical parameters such as organization T2, T2*, M0, ΔB0, and B1 during simulation, corresponding standard quantitative label images are generated as targets for supervised learning. This improves the expressive power of the network learning while maintaining information integrity. Compared to traditional acquisition schemes that only cover a small amount of TE, the design of this invention makes the TE distribution more uniform along the time axis and provides more information, especially significantly enhancing the sampling density and signal preservation ability in long TE regions, which helps improve fitting stability and quantitative accuracy.

[0184] This invention designs a magnetic resonance imaging (MRI) sequence with multiple echo points; optimizes and determines key parameters such as excitation angle, inter-echo time interval, shift gradient intensity, imaging field of view, and imaging matrix size; loads the designed sequence onto an MRI simulation platform with sufficient performance to acquire high-quality simulation data under different brain layers; constructs a deep neural network model that integrates multi-echo information, and completes training and validation based on the simulation dataset; achieves rapid joint quantification of tissue relaxation parameters such as T2 and T2* through end-to-end neural network mapping, while also possessing the ability to simultaneously quantify parameters such as M0, ΔB0, and B1. The multi-echo design improves the sampling density and coverage of different TE echoes, enhancing the accuracy of T2 and T2* fitting; simultaneously, it can model and correct common non-ideal factors in MRI such as main magnetic field (ΔB0) inhomogeneity, radio frequency field (B1) inhomogeneity, image distortion, and artifacts, effectively improving quantitative image quality and reconstruction robustness. This invention constructs a high-precision quantitative imaging framework with balanced temporal distribution, enhanced parameter sensitivity, and an end-to-end learnable structure through the rational scheduling of multiple TE echo points and the coordinated design of short and long TEs. While maintaining acquisition efficiency, it significantly improves the quantitative quality and robustness of multiple parameters such as T2 and T2*. This invention offers advantages such as short scan time, high quantitative accuracy, and strong anti-interference capability, making it suitable for high-precision quantitative brain magnetic resonance imaging and multi-parameter analysis scenarios.

[0185] The above embodiments are merely preferred embodiments of the present invention and should not be considered as limiting the scope of the present invention. All equivalent variations and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance imaging based on multiple echo separation, characterized in that, Includes the following steps: S1. Pulse sequence design: During one excitation process of a single scan, three radio frequency excitation pulses and one 180° refocusing pulse are applied to form three echo chains for acquisition. Each echo chain contains echo signals with five different echo times (TE), resulting in a total of 15 magnetic resonance echo signals covering the short, medium, and long TE ranges. Each echo chain also synchronously contains T2-weighted signals and T2*-weighted signals. S2. Data Acquisition and Reconstruction: Perform two-dimensional Fourier transform on the k-space data of the multiple echo signals to reconstruct the multiTE magnetic resonance complex image, separate the real and imaginary parts of the complex image to form a six-channel magnetic resonance image tensor, and normalize the image tensor. S3. Deep model quantization: The tensors of the six-channel magnetic resonance images are input into a pre-trained deep neural network model, and the T2 quantitative map, T2* quantitative map, and optional proton density M0 map, main magnetic field inhomogeneity ΔB0 map, and radio frequency field inhomogeneity B1 map are output through end-to-end mapping; the deep neural network model is obtained by training on a simulation dataset containing ΔB0, B1 perturbations and noise modeling.

2. The precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance based on multiple echo separation according to claim 1, characterized in that, In step S1: the first echo chain is acquired after the third RF excitation pulse, including the spin echo (SE) signal and gradient echo (GRE) signal under short TE; the second echo chain is acquired through a 180° refocusing pulse, including the spin echo (SE) signal and gradient echo (GRE) signal under medium TE; the third echo chain is acquired after a preset delay, including the gradient echo (GRE) signal under long TE.

3. The precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance based on multiple echo separation according to claim 1, characterized in that, In step S1, the key parameters of the pulse sequence satisfy the following: the flip angle of the radio frequency excitation pulse is α1=α2=45°, α3=120°, and the flip angle of the refocusing pulse is 180°; all three echo chains adopt the echo plane imaging EPI readout strategy, and all excitation and echo acquisition are completed within a single repetition time TR.

4. The precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance based on multiple echo separation according to claim 1, characterized in that, In step S1, the pulse sequence design also includes optimizing and determining key parameters, including excitation angle, inter-echo time interval, shift gradient intensity, imaging field of view, and imaging matrix size.

5. The precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance based on multiple echo separation according to claim 1, characterized in that, In step S3, the deep neural network model is built based on the U-Net architecture, or the ResNet or Transformer architecture can be selected. The deep neural network model realizes the rapid joint quantification of T2 and T2* tissue relaxation parameters through end-to-end mapping. The network model input is a six-channel magnetic resonance image tensor, and outputs the T2 quantification map and T2* quantification map, as well as the corresponding proton density M0 map, main magnetic field inhomogeneity ΔB0 map and radio frequency field inhomogeneity B1 map.

6. The precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance based on multiple echo separation according to claim 1, characterized in that, In step S3, the training process of the deep neural network model satisfies the following: the loss function uses the pixel-wise mean square error (MSE), and the formula is: in, M represents the network mapping function; M represents the batch size; j represents the j-th sample input data; The label data is represented by W and b, which represent the network weights and biases. The Adam optimizer is selected with an initial learning rate of 0.0001 and a cosine annealing decay strategy. L2 weight decay, dropout regularization, and early stopping strategy are configured.

7. The precise quantitative method for single-scan T2 and T2* multi-parameter magnetic resonance based on multiple echo separation according to claim 1, characterized in that, In step S3, the process of generating the simulation dataset is as follows: the pulse sequence designed in step S1 is loaded into the magnetic resonance imaging simulation platform, and the 3D simulated human brain template is subjected to imaging simulation. During the simulation, non-ideal factors such as non-ideal radio frequency excitation profile and non-ideal gradient are introduced to obtain k-space complex data containing multiple echoes.

8. A single-scan T2 and T2* multi-parameter magnetic resonance precise quantitative system based on multiple echo separation, characterized in that, include: Sequence Design Module: Used to design a multi-overlapping echo pulse sequence containing 3 RF excitation pulses and 1 180° refocusing pulse, and to configure the acquisition parameters of the three echo chains so that each echo chain contains 5 T2 and T2* weighted echo signals with different TEs; Simulation data generation module: used to load pulse sequences into the magnetic resonance simulation platform, simulate multi-echo k-space complex data with ΔB0 and B1 disturbances, and generate corresponding T2, T2*, M0, ΔB0, and B1 standard quantitative labels; Data preprocessing module: used to convert k-space data into complex images in the image domain, separate the real and imaginary parts to form a six-channel image tensor, and perform normalization and noise addition processing; Neural network training module: used to train a deep neural network model using a six-channel image tensor and standard quantitative labels, and to establish a nonlinear mapping relationship between echo signals and multi-parameter quantitative maps; Quantitative Image Reconstruction Module: This module preprocesses the actual scanned k-space data into a six-channel image tensor, inputs it into the trained model, and outputs T2 quantitative image, T2* quantitative image, and optional M0, ΔB0, and B1 images.

9. The single-scan T2 and T2* multi-parameter magnetic resonance precise quantitative system based on multiple echo separation according to claim 8, characterized in that, The sequence design module optimizes the flip angle of the radio frequency excitation pulse through a weighted Cramér-Rao lower bound minimization framework, so that α1, α2∈[0°,90°], α3=120°.

Images