A high-resolution multi-parameter quantitative method and system for magnetic resonance imaging based on multi-scan overlapping echoes

Abstract

A high-resolution, multi-parameter quantitative magnetic resonance imaging method and system based on multi-scan overlapping echoes relates to magnetic resonance imaging methods. The method comprises: designing a high-resolution, multi-parameter, simultaneous quantitative imaging sequence based on multi-scan overlapping echoes and determining its sampling parameters; using the designed sequence and sampling parameters in a magnetic resonance instrument that meets the sequence and sampling parameters to acquire data from the imaging object and obtain magnetic resonance signals; processing the magnetic resonance signals to obtain images acquired by the sequence; designing a high-fidelity quantitative reconstruction method; and using the method to quantitatively reconstruct the acquired magnetic resonance images to obtain high-resolution, multi-parameter quantitative magnetic resonance images. This method enables high-resolution, multi-parameter simultaneous quantitative magnetic resonance imaging and overcomes the problem of loss of reconstructed details under low signal-to-noise ratio conditions. Experimental results demonstrate that the reconstructed magnetic resonance quantitative parameter images are highly consistent with those obtained by reference methods, exhibit high imaging accuracy, and have broad application prospects.

Description

Technical Field

[0001] The present invention relates to a magnetic resonance imaging method, in particular to a magnetic resonance high-resolution multi-parameter quantitative method and system based on multi-scanning overlapping echoes. Background Art

[0002] Quantitative magnetic resonance imaging can quantitatively measure physiological parameters. Compared with qualitative parameter-weighted imaging, it can provide more objective evaluation criteria, which is of great significance for the evaluation of therapeutic effects and the graded diagnosis of diseases. High-resolution quantitative images can provide better anatomical details and enhance the ability to evaluate tissue structure and identify early pathology. However, traditional quantitative magnetic resonance imaging methods have a long acquisition time and can only quantify one parameter at a time, which greatly limits the application of quantitative magnetic resonance imaging. Based on this, a series of single-scan overlapping echo fast imaging methods have been developed (CN108010100B, CN108663644B, CN110807492B). After one excitation, multiple echoes with single or multiple parameters of different contrast are simultaneously acquired to achieve a single T2 or T2 * parameters or both simultaneously for rapid quantification. However, since the echo train length of a single scan is limited by the transverse relaxation attenuation during signal readout, the imaging resolution of the above invention is low and some detailed structures may be lost. In addition, the above invention uses a reconstruction method based on a convolutional neural network (CNN) to learn the mapping relationship between overlapping echo images and quantitative parameters by minimizing the mean mean squared error (MMSE) between the network output and the quantitative parameters. Therefore, in high-resolution imaging, the smaller the voxel, the lower the signal-to-noise ratio. Due to the small receptive field of CNN, it is difficult for it to capture global features and complex spatial relationships. The reconstruction result is relatively smooth when the signal-to-noise ratio is low. In addition, training using the MMSE method will also reduce the perceptual quality (such as visual blur), especially under low signal-to-noise ratio conditions (Blau Y, Michaeli T. The perception-distortion tradeoff. Proc Cvpr Ieee . 2018;6228-6237). Using multi-scan technology can shorten the readout time of a single scan and improve imaging resolution. Other imaging methods based on multi-scan overlapping echoes (CN113030813B, CN118549870A) can only quantify the T2 parameter and do not consider the impact of changes in signal-to-noise ratio on reconstruction results when the acquisition resolution changes. Therefore, the above invention is not suitable for high-resolution multi-parameter quantitative imaging.

[0003] The Swin Transformer structure is a technology transplanted from the field of natural language processing to the field of computer vision. Thanks to its global self-attention mechanism, it can better model global features and better restore texture details (LiuZ, Lin YT, Cao Yet al. Swin transformer: Hierarchical vision transformerusing shifted windows. 2021 Ieee / Cvf International Conference on Computer Vision (Iccv 2021). 2021;9992-10002). Summary of the Invention

[0004] The purpose of the present invention is to address the problems of loss of reconstruction details, reconstruction quality, and noise robustness under low signal-to-noise ratio conditions in the existing technology. This invention provides a high-resolution, multi-parameter quantitative magnetic resonance imaging method and system based on multi-scan overlapping echoes. This method and system uses multi-scan technology to improve single-scan multi-parameter overlapping echo separation imaging, achieving rapid, high-resolution, multi-parameter quantitative imaging. Furthermore, a high-fidelity quantitative reconstruction method based on the Swin Transformer architecture is designed. By combining the auxiliary task of noise learning with improved MMSE, high-quality quantitative results can still be obtained under low signal-to-noise ratio conditions.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A high-resolution multi-parameter quantitative method for magnetic resonance imaging based on multi-scan overlapping echo acquisition comprises the following steps:

[0007] S1: Design a high-resolution multi-parameter quantitative magnetic resonance imaging sequence based on overlapping echoes, called a multi-scan multi-parameter multiple overlapping echo separation imaging sequence;

[0008] S2: Determine the sampling parameters of the multi-scan multi-parameter multiple overlap echo separation imaging sequence;

[0009] S3: using the designed sequence and sampling parameters in a magnetic resonance instrument that meets the sequence and sampling parameters, collecting data on the imaging object to obtain magnetic resonance signals;

[0010] S4: After rearranging, splicing, zero-filling, and parallel reconstruction of the acquired magnetic resonance signals, a multi-scan, multi-parameter, and multi-overlap echo image is obtained;

[0011] S5: Design a high-fidelity quantitative reconstruction method based on the Swin Transformer structure;

[0012] S6: Obtaining a trained neural network using the high-fidelity quantitative reconstruction method;

[0013] S7: Using the trained neural network to quantitatively reconstruct the acquired magnetic resonance image to obtain a high-resolution multi-parameter magnetic resonance quantitative image.

[0014] Optionally, the designing of a magnetic resonance imaging sequence specifically includes:

[0015] S11: Design a signal excitation module to generate magnetic resonance signals;

[0016] S12: Design a shift gradient module to refocus the generated signals into the same k-space;

[0017] S13: Design a data acquisition module for acquiring magnetic resonance signals;

[0018] Optionally, determining the imaging sequence sampling parameters mainly includes:

[0019] S21: Determine parameters such as RF pulse angle, shape, and application time;

[0020] S22: determining parameters such as the displacement gradient area, direction, and application time;

[0021] S23: determining the number of scans to balance readout time and imaging resolution;

[0022] S24: determining a parallel imaging acceleration factor to speed up acquisition;

[0023] S25: Confirm other acquisition parameters;

[0024] Optionally, the design of a high-fidelity quantitative reconstruction method based on a Swin Transformer structure specifically includes:

[0025] S51: Determine the quantitative reconstruction neural network based on Swin Transformer;

[0026] S52: Design noise learning auxiliary tasks to improve the robustness of neural networks to noise;

[0027] S53: Based on S51 and S52, use the Pytorch deep learning framework to build a neural network training framework and determine hyperparameters such as learning rate, batch size, and number of iterations;

[0028] S54: Generate neural network training samples; including simulated data or actually collected data and their corresponding quantitative parameters.

[0029] A magnetic resonance high-resolution multi-parameter quantitative system based on multi-scan overlapping echo acquisition, comprising:

[0030] The sequence design module is responsible for designing multi-scan, multi-parameter, and multiple overlapping echo separation imaging sequences, including the specific implementation of signal excitation, shift gradient, and data acquisition modules, and passing the design parameters to the sampling parameter determination module;

[0031] The sampling parameter determination module further determines specific sampling parameters such as RF pulse angle, shape, application time, shift gradient area, direction, application time, number of scans, parallel imaging acceleration factor, etc. based on the design parameters, and provides these parameters to the data acquisition module;

[0032] In the magnetic resonance instrument, the data acquisition module collects data on the imaging object according to the given sequence and sampling parameters, obtains magnetic resonance signals, and transmits these signals to the image preprocessing module;

[0033] The image preprocessing module performs preprocessing operations such as rearrangement, splicing, zero filling, and parallel reconstruction on the collected magnetic resonance signals to generate multi-scan, multi-parameter, and multi-overlap echo images, and transmits them to the neural network reconstruction module;

[0034] The neural network reconstruction module uses a high-fidelity quantitative reconstruction neural network based on the Swin Transformer structure to quantitatively reconstruct the preprocessed images to obtain high-resolution multi-parameter magnetic resonance quantitative images. This module also needs to interact with the neural network training module to obtain a trained neural network model.

[0035] The neural network training module is responsible for designing the neural network training framework. It uses the magnetic resonance images with different features and parameters provided by the simulation data generation module as training samples (these training samples also simulate the instability factors in real experiments to improve the robustness of the model) to train the neural network and provide the trained model to the neural network reconstruction module.

[0036] The simulation data generation module is specifically used to generate simulation data required for training;

[0037] The result display module receives the reconstructed images output by the neural network reconstruction module and displays them for user viewing and analysis. These modules work together to implement the complete process from sequence design to data acquisition, preprocessing, quantitative reconstruction, and result display.

[0038] Compared with the prior art, the advantages of the present invention are:

[0039] 1. This invention uses multi-scan overlapping echo technology to acquire magnetic resonance signals through multiple scans, effectively shortening the readout time of a single scan and thus improving imaging resolution. Compared with traditional single-scan methods, this method can capture detailed structures more clearly and avoid information loss.

[0040] 2. This paper introduces the Swin Transformer architecture, an advanced technology ported from natural language processing to computer vision. The Swin Transformer features a global self-attention mechanism, enabling better modeling of global features and restoration of texture details. Compared to traditional convolutional neural networks (CNNs), the Swin Transformer can still produce high-quality quantitative results under low signal-to-noise ratio conditions.

[0041] 3. The present invention realizes rapid magnetic resonance high-resolution parameters T2 and T2 * Simultaneous quantitative imaging of B1, PD, and ΔB0. Compared to traditional methods that can only quantify one parameter at a time, this method greatly improves imaging efficiency and information volume, providing doctors with a more comprehensive diagnosis basis.

[0042] 4. Due to the use of multi-scan technology and Swin Transformer structure, the quantitative magnetic resonance images obtained by the present invention are highly accurate and have good detail. This is of great significance for evaluating tissue structure and identifying early pathology, helping to improve the accuracy of disease diagnosis.

[0043] 5. This paper incorporates a noise learning auxiliary task into the design of a high-fidelity quantitative reconstruction method based on the Swin Transformer architecture. This design enables the neural network to better learn the noise characteristics during training, thereby maintaining stable reconstruction quality under low signal-to-noise ratio conditions.

[0044] 6. The present invention demonstrates robust reconstruction quality under low signal-to-noise ratio conditions. Even under poor imaging conditions, the present invention can still provide high-quality quantitative MRI images, providing physicians with reliable diagnostic evidence. This technology has broad application prospects and significant clinical value in the field of medical imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of the magnetic resonance high-resolution multi-parameter simultaneous quantitative imaging method of the present invention.

[0046] Figure 2 A flow chart of a high-fidelity quantitative reconstruction method based on the Swin Transformer structure is designed for the present invention.

[0047] Figure 3 This is a multi-scan, multi-parameter, multiple overlapping echo separation imaging sequence diagram and k-space filling trajectory in an embodiment of the present invention.

[0048] Figure 4 This is the neural network model used in the high-fidelity quantitative reconstruction method in the embodiment of the present invention.

[0049] Figure 5This is a high-resolution multi-parameter quantitative image of the human brain obtained by the method of the present invention.

[0050] Figure 6 T2 and T2 obtained by the method of the present invention under different signal-to-noise ratio conditions * Quantitative graph. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work are within the scope of protection of the present invention, including but not limited to: adding common preparation modules such as inversion, saturation, and fat compression before using the sequence and in the intermediate process of cyclic use of the sequence, which should be regarded as the scope of no obvious creative work on the basis of the present invention, that is, within the scope of protection of the present invention.

[0052] like Figure 1 The magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes according to an embodiment of the present invention comprises the following steps:

[0053] S1: Design a high-resolution multi-parameter quantitative magnetic resonance imaging sequence based on overlapping echoes, called a multi-scan multi-parameter multiple overlapping echo separation imaging sequence;

[0054] S11: Design the signal excitation module. Designing the signal excitation module is the first step in the imaging sequence. In this embodiment, multiple radio frequency (RF) pulses are used to generate echo signals of varying contrast. These echo signals provide the parameter weighting information required for subsequent quantitative imaging. To optimize signal excitation, an appropriate flip angle, pulse shape, and application time are selected to ensure signal strength and maximize signal sensitivity to parameters.

[0055] S12: Design a shift gradient module. The shift gradient module is used to refocus individual echo signals into the same k-space. By designing specific shift gradients, each echo signal can be precisely moved to a specific location in k-space. This design not only improves signal acquisition efficiency but also enhances the quality and accuracy of quantitative imaging.

[0056] S13: Design the data acquisition module; the data acquisition module is used to collect magnetic resonance signals during the imaging process. In order to obtain quantitative information of multiple parameters simultaneously, different readout modules (such as SE-EPI and GRE-EPI) are used to obtain T2 and T2 respectively. *The system generates k-space data with equal parameter weighting. Furthermore, multi-scan technology is combined to shorten the sampling echo train length in the pulse sequence through multiple scans to reduce signal relaxation attenuation. Simultaneously, parallel imaging methods are used to reduce scan time, achieving a balance between the minimum signal-to-noise ratio and scan time required for high-resolution multi-parameter quantitative imaging.

[0057] S2: Determine the sampling parameters for the multi-scan, multi-parameter, multiple overlap echo separation imaging sequence. After determining the imaging sequence, further determination of the sampling parameters is required to ensure imaging quality and accuracy. These parameters include the flip angle, shape, and application time of the RF pulse; the area, direction, and application time of the shift gradient; the number of scans; the parallel imaging acceleration factor; and other acquisition parameters (such as the imaging field of view, imaging matrix, acquisition slice thickness, and repetition acquisition time). Optimizing these parameters is crucial for improving imaging quality and parameter quantitative accuracy, including:

[0058] S21: Determine parameters such as RF pulse angle, shape, and application time;

[0059] S22: determining parameters such as the displacement gradient area, direction, and application time;

[0060] S23: Determine the number of scans;

[0061] S24: determining a parallel imaging acceleration factor;

[0062] S25: Confirm other acquisition parameters;

[0063] S3: Use the designed sequence and sampling parameters in a magnetic resonance instrument that meets the sequence and sampling parameters to collect data on the imaging object and obtain magnetic resonance signals; by precisely controlling the excitation pulse, shift gradient and data acquisition process, high-quality magnetic resonance signals can be obtained, providing a reliable data foundation for subsequent processing.

[0064] S4: The acquired magnetic resonance signals are rearranged, spliced, zero-filled, and reconstructed in parallel to obtain a multi-scan, multi-parameter, and multi-overlapping echo image. Specifically, accurate signal alignment and separation must be ensured, and data loss or confusion must be avoided. Subsequently, the data are further processed using a parallel reconstruction method to obtain a high-resolution image.

[0065] S5: Design a high-fidelity quantitative reconstruction method based on the Swin Transformer structure; this method is the core part of the present invention. This method uses deep learning technology and the self-attention mechanism of the Swin Transformer to extract deep features and improve the quality of image reconstruction. At the same time, by designing noise learning auxiliary tasks, the noise characteristics are explicitly understood and separated, and the main task's ability to effectively extract signals is enhanced. This design not only improves the accuracy of reconstruction and the generalization performance of the model, but also significantly improves the performance under low signal-to-noise ratio conditions. Figure 2 , specifically including:

[0066] S51: Determine the quantitative reconstruction neural network based on the Swin Transformer. The neural network structure consists of four modules: shallow feature extraction, deep feature extraction, main task, and auxiliary task. The shallow feature extraction module maps the input image to the shallow feature space through convolutional layers. The deep feature extraction module extracts deep features using Swin Transformer blocks. The main task module obtains quantitative reconstruction parameters through deconvolution layers. The auxiliary task module is used to predict noise.

[0067] S52: Design an auxiliary task for noise learning; this auxiliary task is implemented by sharing the previous feature extraction layer and adding a noise prediction output at the end. This design enables the model to explicitly understand and separate noise characteristics while learning signal features, thereby improving the accuracy and robustness of reconstruction.

[0068] S53: Build a neural network training framework using the Pytorch deep learning framework based on steps S51 and S52 and determine hyperparameters. Optimizing these hyperparameters is crucial for improving the performance and accuracy of the neural network.

[0069] S54: Generate neural network training samples; use simulation methods to generate the training samples required for the deep neural network. By inputting a pulse sequence into the magnetic resonance imaging simulation software and adding corresponding non-ideal terms to simulate real-world conditions, a series of magnetic resonance images with different characteristics and parameters are generated as training samples. These samples not only contain all the characteristics of the object under test but also simulate unstable factors in real experiments (such as excitation pulse angle deviation, shift gradient deviation, and non-uniform field), thereby improving the model's robustness to non-ideal experimental environments.

[0070] S6: Using the high-fidelity quantitative reconstruction method, a trained neural network is obtained. Specifically, the generated training samples are input into the constructed neural network for training. By iteratively training and adjusting the neural network parameters, the loss function is minimized, thereby preserving the deep neural network parameters and ensuring the stability and accuracy of the training process.

[0071] S7: Use the trained neural network to quantitatively reconstruct the collected magnetic resonance images, input the multi-scan multi-parameter multi-overlap echo images into the neural network, and after a series of processing and calculations, obtain high-resolution multi-parameter magnetic resonance quantitative images. These images not only have clear texture structure and high resolution, but also accurately reflect the multiple parameter information of the object to be tested (such as T2, T2 * , B1, PD, ΔB0, etc.).

[0072] Specifically, in step S1, designing a multi-scan multi-parameter multiple overlapping echo separation imaging sequence specifically includes:

[0073] S11: Design signal excitation module;

[0074] Figure 3 (a) is a multi-scan multi-parameter multiple overlapping echo separation imaging sequence diagram proposed by the present invention, wherein α represents the flip angle of the radio frequency pulse, and four main echoes are generated by continuously applying four excitation pulses, which are used to generate signals with different contrast weightings.

[0075] S12: Design the shift gradient module;

[0076] G roi (i = 1, 2, 3, 4) represents the shift gradient of the i-th RF pulse in the frequency encoding direction, G pei (i = 1, 2, 3, 4) represents the shift gradient of the i-th RF pulse in the phase encoding direction, which refocuses the four main echoes to the specified position in k-space.

[0077] S13: Design data acquisition module;

[0078] In order to obtain quantitative information of multiple parameters simultaneously, SE-EPI and GRE-EPI readout modules were used to obtain k-space weighted by different parameters of interest, respectively. The former was mainly weighted by T2 signal, while the latter was mainly weighted by T2 *Signal weighting is the primary approach. Considering that single-scan multi-parameter, multi-overlap echo imaging is limited by lateral relaxation attenuation and cannot produce high-resolution images, the present invention incorporates a multi-scan approach. This shortens the sampling echo train length in the pulse sequence through multiple scans, reducing signal relaxation attenuation. This, combined with a parallel imaging approach, reduces the increased scan time associated with multi-scan approaches, achieving a balance between the minimum signal-to-noise ratio and scan time required for high-resolution, multi-parameter quantitative imaging. The entire sequence is repeated N times to cover the entire k-space. Figure 3 (b) shows the arrangement of N scans in k-space, where R is the parallel imaging acceleration factor. Figure 3 (c) shows the k-space and weighted images acquired by a multi-scan multi-parameter multiple overlap echo separation imaging sequence.

[0079] Specifically, in step S2, determining the sampling parameters of the multi-scan multi-parameter multiple overlap echo separation imaging sequence specifically includes:

[0080] S21: Determine the magnitude of each RF pulse flip angle α, the pulse shape, and the interval Δ between each RF excitation pulse i (i = 1, 2, 3, 4);

[0081] S22: Determine each shift gradient G roi , G pei the area, direction and application time of the

[0082] S23: Determine the number of repeated scans N shot ;

[0083] S24: determining a parallel imaging acceleration factor R;

[0084] After the above parameters are determined, the position and weighting of each overlapping echo in k-space are determined;

[0085] S25: Determine other acquisition parameters such as imaging field of view, imaging matrix, acquisition slice thickness, and repeated acquisition time;

[0086] Specifically, in S4, the magnetic resonance signal is collected according to Figure 3 After rearrangement, splicing, zero filling and parallel reconstruction in the manner of (b), a multi-scan multi-parameter multi-overlap echo image is obtained;

[0087] Specifically, in S5, the high-fidelity quantitative reconstruction method based on the Swin Transformer structure is as follows:

[0088] S51: Determine the quantitative reconstruction neural network based on Swin Transformer;

[0089] The quantitative reconstruction neural network comprises four modules: shallow feature extraction, deep feature extraction, main task, and auxiliary task. Shallow feature extraction is performed by two cascaded convolutional layers. The first convolutional layer maps the input image to a shallow feature space. The second convolutional layer further extracts and downsamples the shallow features, improving the efficiency of deep feature extraction and alleviating computational pressure. Deep feature extraction consists of multiple Swin Transformer blocks, leveraging the high-performance self-attention mechanism of the Transformer to achieve deep feature extraction and improve image reconstruction quality. The Swin Transformer block consists of multiple Swin Transformer layers, connected head-to-tail using residuals. The Swin Transformer layer first passes the input features through a normalization layer (LayerNorm), then through a multi-head self-attention module, introducing residuals at the end of the multi-head self-attention, then through a normalization layer (LayerNorm), and finally through a multi-layer perceptron (MLP). The main task module performs upsampling through a deconvolutional layer to obtain quantitative reconstruction parameters; the auxiliary task module obtains added noise through another deconvolutional layer.

[0090] S52: Design auxiliary tasks for noise learning;

[0091] The auxiliary task of noise learning is implemented by sharing the previous feature extraction layer and adding a noise prediction output at the end. Specifically, the main task network and the auxiliary task network share the same feature extraction structure to extract common features from the input data. After the feature extraction is completed, an output module for predicting noise is attached to estimate the noise component in the input data. Through this design, the model explicitly understands and separates the noise characteristics while learning the signal features, thereby enhancing the main task's ability to effectively extract the signal, significantly improving the reconstruction accuracy and the generalization performance of the model, especially under low signal-to-noise ratio conditions. Such a design can be expressed as:

[0092]

[0093] Among them, L total Represents the loss of the entire model, L main represents the loss of the main task, L noise represents the noise estimation loss, and λ is the weight coefficient used to balance the learning of the main task and the auxiliary task.

[0094] S53: Building a neural network training framework using the Pytorch deep learning framework according to steps S51 and S52, and determining hyperparameters;

[0095] Determine the number of input and output channels of the neural network and the number of Swin Transformer blocks; determine the embedding dimension of the Swin Transformer structure, the number of attention heads and the sliding window size; determine the auxiliary task weight λ; determine the optimizer and learning rate for training the deep neural network;

[0096] S54: Generate neural network training samples;

[0097] Based on the determined sequence and parameters, training samples for a deep neural network are generated. The present invention takes into account that a neural network requires a large number of training samples, while real training samples are difficult to obtain. Therefore, the present invention uses a simulation method to generate the training samples required for a deep neural network. The specific steps are as follows:

[0098] Inputting the pulse sequence into magnetic resonance imaging simulation software and adding corresponding non-ideal terms based on the non-idealities of the actual experiment to simulate the actual situation as much as possible;

[0099] A set number of random templates are generated based on the characteristics of the object to be tested (or templates are generated using an MRI dataset with the characteristics of the object to be tested). The complexity of the templates should match the complexity of the object to be tested and should be able to contain all the characteristics of the experimental sample. The number of templates needs to be sufficient to ensure good reconstruction quality.

[0100] The template is simulated and sampled using simulation software to obtain the template's magnetic resonance signal. During the simulated sampling process, unstable factors are added to take into account changes in the actual experimental environment to improve the robustness of the network model to non-ideal experimental environments. The unstable factors include excitation pulse angle deviation, shift gradient deviation, and non-uniform field.

[0101] The magnetic resonance signals of the templates are rearranged into two-dimensional k-space signals, and then two-dimensional Fourier transform is performed to obtain magnetic resonance images of each template. The magnetic resonance images are normalized and random Gaussian noise is added to the real and imaginary parts respectively.

[0102] The corresponding quantitative parameters in the template and the added noise are used as the corresponding labels for network training;

[0103] The normalized magnetic resonance images of each template and the corresponding labels constitute a training sample, and all simulation training samples are divided according to proportion to form a simulation sample training set and a verification set.

[0104] Specifically, in step S6, the neural network is trained using the training samples as follows:

[0105] The training sample set is input into the deep neural network in batches for iterative training. Each time the network is trained, the value of the loss function is calculated. The parameters of the neural network are adjusted by the gradient descent method based on this value and the loss function to minimize the value of the loss function and save the deep neural network parameters. During training, the deep neural network achieves accurate quantitative reconstruction of the parameters by minimizing the error between the output and the quantitative parameters, while minimizing the error between the output and the added noise, reducing the interference of the noise, improving the gradient update direction, and improving the performance under low signal-to-noise ratio conditions and the overall performance and robustness.

[0106] Specifically, in step S7, the trained neural network is used to quantitatively reconstruct the acquired magnetic resonance image to obtain a high-resolution multi-parameter magnetic resonance quantitative image, as follows:

[0107] The multi-scan multi-parameter multi-overlap echo image obtained in step S4 is input into the trained neural network obtained in step S6 to reconstruct a high-resolution multi-parameter magnetic resonance quantitative image.

[0108] A specific embodiment is given below.

[0109] Step 1: Design a multi-scan, multi-parameter, multiple overlapping echo separation imaging sequence and determine the sequence sampling parameters.

[0110] MRI sequence design Figure 3 As shown in (a) in the figure, the excitation pulses α are all Sinc pulses with a flip angle of 30°, and the refocusing pulses β are Sinc pulses with a flip angle of 180°.

[0111] The time between each excitation pulse is Δ1 = Δ2 = Δ3 = 8.61ms, Δ4 = 7.78ms;

[0112] Repeat scan times N shot = 4;

[0113] Parallel imaging acceleration factor R = 2;

[0114] Assume that the readout module has a readout gradient area A in the frequency encoding direction during sampling ro ; Assume that the area of ​​the phase encoding direction of the readout module during sampling is A pe . A roi and A pei represent the shift gradient G applied after the i-th excitation pulse roi and G pei The relative areas of each shift gradient are: A roi / A ro = 1 / 4, 1 / 4, 1 / 4, 1 / 8, G roi / G ro= -44 / 256, 22 / 256, -44 / 256, -88 / 256.

[0115] The imaging field of view (FOV) is 22 cm × 22 cm;

[0116] The imaging matrix size is 256 × 248;

[0117] The collection layer thickness is 3 mm;

[0118] The repeated acquisition time was 5 s.

[0119] Step 2: Use the designed sequence and sampling parameters in the magnetic resonance instrument that meets the sequence and sampling parameters to collect data on the imaging object and obtain magnetic resonance signals. Figure 3 After parallel reconstruction, we can obtain Figure 3 The multi-scan, multi-parameter, multiple overlapping echo image shown in (c) contains multiple overlapping echoes in its k-space and exhibits strip-like interference fringes in the image domain.

[0120] Step 3: Design a high-fidelity quantitative reconstruction method based on the Swin Transformer structure.

[0121] Determine the number of input and output channels of the neural network and the number of Swin Transformer blocks. Figure 4 As shown, the sequence described in this embodiment collects 2 k-space data, corresponding to 2 magnetic resonance images, and uses the real and imaginary parts of each magnetic resonance image as the input of the deep neural network, so the number of input channels of the deep neural network is 4. In this embodiment, T2, T2 * , PD, B1, and ΔB0, so the number of output channels of the main task module of the neural network is 5. The auxiliary task module outputs the estimated noise, so the number of output channels of the auxiliary task module is 1. The number of Swin Transformer blocks is 6, the embedding dimension is 96, the number of attention heads is 4, and the sliding window size is 8. A modified L1 norm is used as the loss function for the main task, and the mean squared error is used as the loss function for the auxiliary task, as follows:

[0122]

[0123] Among them, P n Indicates target parameter maps (such as T2 and T2 * ), S n is the image acquired by the magnetic resonance sequence, F(S n ;θ) represents the output of the network, n represents the nth sample in the batch size N, N p(Sn;θ) is the noise of auxiliary task prediction, N in represents the actual noise added, ϵ is a small constant used to improve the L1 norm to ensure numerical stability, and θ is the neural network parameter. The weight coefficient of the auxiliary task is λ = 0.15.

[0124] The Adam optimizer was used to adjust the learning rate using exponential decay: the initial learning rate was set to 0.0001, the number of iterations was set to decay once every 20,000 times, and the decay rate was set to 0.8.

[0125] To generate neural network training samples, pulse sequences are imported into magnetic resonance imaging simulation software. Correction terms are added based on non-ideal factors encountered in real-world experiments to closely replicate actual experimental conditions. A series of random templates are generated based on the characteristics of the object under test. The complexity of the templates should match the characteristics of the object under test and encompass all features of the experimental sample. A sufficient number of templates is required to ensure good reconstruction quality. The templates are then sampled using simulation software to obtain corresponding magnetic resonance signals. During the sampling process, factors such as excitation pulse angle deviation, gradient deviation, and noise are incorporated to account for variations in the real-world experimental environment, enhancing the robustness of the model under non-ideal experimental conditions. The template magnetic resonance echo signals are reordered into two-dimensional k-space signals. After a two-dimensional Fourier transform, magnetic resonance images for each template are generated. These images are then normalized and noise is added. Based on the characteristics of the parameters, the corresponding quantitative parameters in the templates are normalized to a range suitable for network learning, generating the corresponding labeled data. Finally, the normalized template magnetic resonance images and corresponding labels are combined to form training samples. All simulated samples are then divided proportionally into training and validation sets.

[0126] The training sample set is randomly divided into two parts: 80% for training and 20% for validation.

[0127] When training a deep neural network, the training sample set is input into the deep neural network in batches for iterative training. Each time the network is trained, the value of the loss function is calculated, and the parameter value of the neural network is automatically adjusted according to the value to reduce the value of the loss function. The above training is repeated until the value of the loss function tends to be stable, and the deep neural network parameter θ is saved.

[0128] Step 4: quantitatively reconstruct the acquired magnetic resonance images using the trained neural network to obtain high-resolution multi-parameter magnetic resonance quantitative images.

[0129] The real human brain image collected in step 2 is input into the neural network trained in step 3 for reconstruction, generating Figure 4 The multi-parameter quantitative magnetic resonance images are shown. Figure 5 Showing the real human brain's T2, T2 *Figure 2 shows five quantitative MRI parameter images: PD, B1, and ΔB0. The image on the left is a quantitative reference image acquired using a reference method, while the image on the right is a quantitative parameter image generated using the high-fidelity reconstruction method of the present invention. As can be seen, the quantitative MRI parameter images generated by the reconstruction method of the present invention are highly consistent with those obtained using the reference method, exhibiting high imaging accuracy and clear texture structures.

[0130] Figure 6 The results of reconstructing the simulated human brain samples in the validation set by inputting them into the neural network trained in step 3 are shown, where different degrees of Gaussian noise are added to simulate various signal-to-noise ratio conditions. The leftmost image in the first row is the real image of T2, and the right side is the T2 image generated by the reconstruction method of the present invention under different signal-to-noise ratio conditions; the second row is the absolute error image between the reconstruction result and the corresponding label image, magnified 10 times for easier observation. The third and fourth rows are T2 * The reconstruction and error images show that the high-fidelity reconstruction method based on the Swin Transformer exhibits high accuracy and detail preservation under various signal-to-noise ratio conditions, and can achieve good reconstruction results under low signal-to-noise ratio conditions.

[0131] By comparing the imaging results of the reference method and the method of the present invention, it was found that the magnetic resonance quantitative parameter images generated by the reconstruction method of the present invention were highly consistent with the results of the reference method, with high imaging accuracy and clear texture structure. In addition, the results of reconstruction under different signal-to-noise ratio conditions also showed that the method of the present invention exhibited high accuracy and detail preservation capabilities under various conditions. The present invention combines multi-scanning technology, multi-parameter imaging and a high-fidelity quantitative reconstruction method based on Swin Transformer to achieve high-resolution, multi-parameter, low-noise MRI imaging. Experimental results show that the present invention is superior to traditional methods in terms of imaging quality and parameter quantitative accuracy.

[0132] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A high-resolution multi-parameter quantitative method for magnetic resonance imaging based on multi-scan overlapping echoes, characterized in that The following steps are involved: S1: Design a high-resolution, multi-parameter quantitative magnetic resonance imaging sequence based on overlapping echoes, called a multi-scan, multi-parameter, multiple overlapping echo separation imaging sequence, including the design of a signal excitation module, a shift gradient module, and a data acquisition module; S2: Determine the sampling parameters of the multi-scan multi-parameter multiple overlap echo separation imaging sequence; S3: using the designed sequence and sampling parameters in a magnetic resonance instrument that meets the sequence and sampling parameters, collecting data on the imaging object to obtain magnetic resonance signals; S4: After rearranging, splicing, zero-filling, and parallel reconstruction of the acquired magnetic resonance signals, a multi-scan, multi-parameter, and multi-overlap echo image is obtained; S5: Design a high-fidelity quantitative reconstruction method based on the Swin Transformer structure. This method includes determining the quantitative reconstruction neural network based on the Swin Transformer, designing noise learning auxiliary tasks, building a neural network training framework using the Pytorch deep learning framework and determining hyperparameters, and generating neural network training samples. S6: training the neural network using the high-fidelity quantitative reconstruction method designed in step S5 to obtain a trained neural network; S7: Use the neural network trained in step S6 to quantitatively reconstruct the acquired magnetic resonance image, input the multi-scan multi-parameter multi-overlap echo image into the neural network, and obtain a high-resolution multi-parameter magnetic resonance quantitative image.

2. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S1 , the design of the signal excitation module includes using multiple radio frequency pulses to generate echo signals with different contrasts.

3. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S1 , the design of the shift gradient module includes designing a specific shift gradient so that each echo signal is refocused into the same k-space and accurately moved to a specified position.

4. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S1, the design of the data acquisition module includes adopting different readout modules to acquire k-space data with different parameter weightings, and combining multi-scan technology and parallel imaging methods to shorten the sampling echo chain length in the pulse sequence and reduce the scanning time.

5. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S2, the determining of sampling parameters of a multi-scan multi-parameter multiple overlapping echo separation imaging sequence includes: S21: Determine the RF pulse angle, shape and application time; S22: determining the displacement gradient area, direction and application time; S23: determining the number of scans to balance readout time and imaging resolution; S24: determining a parallel imaging acceleration factor to speed up acquisition; S25: Confirm other acquisition parameters.

6. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S5, the quantitative reconstruction neural network includes a shallow feature extraction module, a deep feature extraction module, a main task module and an auxiliary task module, wherein the deep feature extraction module uses the Swin Transformer block to extract deep features; the noise learning auxiliary task is implemented by sharing the previous feature extraction layer and adding a noise prediction output at the end to improve the reconstruction accuracy and the generalization performance of the model.

7. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S5, generating neural network training samples includes using a simulation method to generate magnetic resonance images with different features and parameters as training samples, and simulating unstable factors in real experiments to improve the robustness of the model to non-ideal experimental environments.

8. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S6, the neural network training includes inputting the training sample set into the deep neural network in batches for iterative training, and adjusting the parameters of the neural network to minimize the value of the loss function.

9. A magnetic resonance high-resolution multi-parameter quantitative method based on multi-scan overlapping echoes as claimed in claim 1, characterized in that In step S7, the high-resolution multi-parameter MRI quantitative image includes parameters T2, T2 * , B1, PD, and ΔB0 images, these quantitative parameter maps can be acquired simultaneously and have clear texture structure and noise robustness.

10. A high-resolution multi-parameter quantitative magnetic resonance system based on multi-scan overlapping echo acquisition, characterized in that include: The sequence design module is responsible for designing multi-scan, multi-parameter, and multiple overlapping echo separation imaging sequences, including the specific implementation of signal excitation, shift gradient, and data acquisition modules, and passing the design parameters to the sampling parameter determination module; The sampling parameter determination module further determines the specific sampling parameters of the RF pulse angle, shape, application time, shift gradient area, direction, application time, number of scans, and parallel imaging acceleration factor based on the design parameters, and provides these parameters to the data acquisition module; In the magnetic resonance instrument, the data acquisition module collects data on the imaging object according to the given sequence and sampling parameters, obtains magnetic resonance signals, and transmits these signals to the image preprocessing module; The image preprocessing module performs rearrangement, splicing, zero filling, and parallel reconstruction preprocessing operations on the collected magnetic resonance signals to generate multi-scan, multi-parameter, and multi-overlap echo images, and transmits them to the neural network reconstruction module; The neural network reconstruction module uses a high-fidelity quantitative reconstruction neural network based on the Swin Transformer structure to quantitatively reconstruct the preprocessed images to obtain high-resolution multi-parameter magnetic resonance quantitative images. This module also needs to interact with the neural network training module to obtain a trained neural network model. The neural network training module is responsible for designing the neural network training framework, using the magnetic resonance images with different characteristics and parameters provided by the simulation data generation module as training samples to train the neural network, and providing the trained model to the neural network reconstruction module; The simulation data generation module is specifically used to generate simulation data required for training; The result display module receives the reconstructed image output by the neural network reconstruction module and displays it for users to view and analyze; These modules work together to implement the complete process from sequence design to data acquisition, preprocessing, quantitative reconstruction and result presentation.

Images