Multi-parameter quantitative method and system based on multi-overlap echo steady-state free procession and multi-task learning network

Abstract

The invention discloses a multi-parameter quantitative method and system based on multi-overlap echo steady-state free procession and a multi-task learning network, and relates to magnetic resonance imaging. The method comprises the following steps: designing and debugging a multi-overlap echo steady-state free precession sequence to obtain a magnetic resonance image with multiple contrast (T1, T2, T2 *) weighting; generating a training sample of the deep neural network; designing a multi-task deep learning network fusing interlayer information; training a network weight by adopting the training sample; and the trained multi-task deep learning network is adopted to map the actually acquired magnetic resonance image, and quantitative images of a plurality of parameters are obtained at the same time. Rich multi-contrast multi-echo information is provided based on multi-overlap echo steady-state free precession, a multi-task deep learning network fuses characteristics of different tasks in different levels, multiple times of scanning are not needed, and rapid and accurate multi-parameter simultaneous quantitative imaging can be achieved in one time of scanning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of magnetic resonance imaging, in particular to a multi-parameter quantitative method and system based on multi-overlapping echo steady-state free precession and multi-task learning network. BACKGROUND

[0002] Magnetic resonance multi-parameter quantitative imaging is a technology that uses magnetic resonance imaging techniques to obtain multiple tissue characteristic parameters to quantitatively characterize the characteristics of biological tissues or lesions. Unlike traditional MRI imaging methods, magnetic resonance multi-parameter quantitative imaging not only provides spatial information of tissue structure, but also quantifies physical parameters of tissue such as relaxation time, diffusion coefficient, and magnetic susceptibility. These parameters can accurately represent the microscopic structure, functional state, and lesion characteristics of the tissue, and are of great significance for disease diagnosis, treatment evaluation, and disease progression monitoring.

[0003] Steady-state free precession (SSFP) is a fast imaging technique that can simultaneously reflect the T1 and T2 relaxation characteristics of tissue and has high signal-to-noise ratio. The multi-overlapping echo steady-state free precession sequence combines the steady-state free precession sequence with the overlapping-echo detachment (OLED) technique to obtain multiple echo signals with different multi-parameter contrasts and phase modulation in one k-space, which can significantly reduce the scanning time and achieve fast and accurate multi-parameter quantitative imaging.

[0004] Multi-task learning is a deep learning method that can simultaneously learn multiple related tasks. When applied to the field of multi-parameter quantitative imaging, it not only introduces implicit regularization to improve the generalization ability of the model, but also learns the texture-related characteristics between tissue parameters, enabling feature sharing and knowledge transfer between different tasks, which can significantly improve imaging efficiency and quantitative accuracy. SUMMARY

[0005] The purpose of the present application is to provide a multi-parameter quantitative method and system based on multi-overlapping echo steady-state free precession and multi-task learning network. The multi-overlapping echo steady-state free precession sequence can quickly acquire multiple overlapping echo images with multiple contrasts in one scan, and the multi-task deep learning network can simultaneously reconstruct M0, T1, T2, T2 * , B0, and B1 multi-parameter quantitative images from the acquired magnetic resonance images.

[0006] To achieve the above-mentioned purpose of the application, the technical scheme adopted by the present application is as follows:

[0007] The present application provides a multi-parameter quantitative method based on multi-overlapping echo steady-state free precession and multi-task learning network, comprising the following steps:

[0008] S1, design a multi-shot steady-state free precession sequence and determine acquisition parameters, the multi-shot steady-state free precession sequence is any magnetic resonance imaging sequence containing a non-equilibrium steady-state free precession sequence to generate multi-shot echoes under multiple radio frequency pulse excitations;

[0009] S2, generate deep neural network training samples; specifically comprising:

[0010] S21, generate a virtual imaging object covering the parameter range of normal and lesion tissues;

[0011] S22, write the multi-shot steady-state free precession sequence on a Bloch simulation platform and simulate data acquisition on the virtual imaging object according to the determined sampling parameters to obtain magnetic resonance imaging signals of the virtual imaging object;

[0012] S23, perform zero padding, inverse Fourier transform, adding Gaussian noise and other operations on the magnetic resonance imaging signals of the virtual imaging object to obtain a series of simulation samples;

[0013] S3, design and train a multi-task deep learning network with the training samples; specifically comprising:

[0014] S31, design a multi-task deep learning network, including a shared feature extraction module, an inter-layer information fusion module, a specific task self-attention module, a cross-task learning module, and a decoder module;

[0015] S32, design an efficient network loss function;

[0016] S33, train the multi-task deep learning network with the simulation samples;

[0017] S4, use the multi-shot steady-state free precession sequence in an advanced magnetic resonance instrument and perform data acquisition on an actual imaging object according to the determined acquisition parameters;

[0018] S5, use the trained multi-task deep learning network to map the magnetic resonance image of the actual imaging object to obtain quantitative images of multiple parameters (M0, T1, T2, T2 * , B0, B1).

[0019] Preferably, in step S1, the multi-shot steady-state free precession sequence is designed and the acquisition parameters are determined, specifically comprising:

[0020] The multi-shot steady-state free precession sequence is based on the steady-state free precession sequence, and equal size shift gradients are applied in the phase encoding and frequency encoding directions, causing imbalance of TR gradients, plus continuous excitation of radio frequency pulses, which can form a series of overlapping echoes in the k-space diagonal direction.

[0021] determining the number of acquisitions n of the echo train; determining the repetition time; determining the echo time of each echo train; determining the size, direction and application time of the shift gradient; determining the imaging field of view, the imaging matrix, the number of acquisition layers, the layer thickness, the interlayer spacing, the bandwidth.

[0022] Preferably, in step S31, a multi-task deep learning network is designed, including a shared feature extraction module, an inter-layer information fusion module, a specific task self-attention module, a cross-task learning module, and a decoder module, specifically including:

[0023] The shared feature extraction module extracts shallow general features between tasks based on a residual convolutional network;

[0024] The inter-layer information fusion module learns the correlation between adjacent image slices based on a cross-attention mechanism to capture three-dimensional image features; the specific task self-attention module learns the representation of a specific task from the shallow general features based on a self-attention mechanism;

[0025] The cross-task learning module is based on a progressive hierarchical extraction structure, which continuously decouples features into specific task representations and general representations shared by all tasks through a dynamic gating mechanism, thereby achieving efficient information exchange and knowledge complementation between tasks; Specifically, a shared module is introduced between tasks to capture general features between different tasks, and then each task extracts features beneficial to itself from the shared layer by layer, which can be expressed as follows:

[0026]

[0027]

[0028] wherein, and are the outputs of the kth shared expert and the specific task expert, respectively; is the weight of the gating structure output of the kth shared expert and the specific task expert; n is a quantitative parameter of the total number of tasks; k is the level of the network;

[0029] Finally, the decoder module outputs a plurality of quantitative images of specific sizes through a series of up-sampling operations.

[0030] Preferably, in step S32, an efficient network loss function is designed, specifically including:

[0031] A multi-task loss dynamic weighting strategy is designed, which adopts a loss weight dynamic adjustment method based on task uncertainty, and introduces a trainable hyperparameter ​To estimate the uncertainty of each task, then to the uncertainty of the task is given smaller weight, uncertainty of the task is given greater weight, to balance the contribution of each task to the model training, the formula is as follows:

[0032]

[0033] Wherein, is the quantitative loss function of different tasks; is a trainable hyperparameter used to estimate the uncertainty of each task; Their values determine whether the output of each task is modeled by Gaussian likelihood or maximum likelihood;

[0034] A multi-modal image joint low-rank constraint method is designed, based on the prior information of quantitative image low-rank, T-SVD tensor decomposition and weighted kernel norm method are used to constrain the joint low-rank of multiple quantitative images output by the network, the formula is as follows:

[0035]

[0036]

[0037]

[0038]

[0039] Wherein, is the tensor singular value decomposition of the three-dimensional tensor , the output is the unit factor , and singular value matrix ; is the singular value matrix obtained after T-SVD tensor decomposition of the three-dimensional tensor; is the singular value of the i-th row and j-th column in the singular value matrix , reflecting the characteristic intensity of the tensor; is the number of rows of the singular value matrix , is the total number of quantitative parameter tasks; is the singular value weight, which is the reciprocal of the singular value of the reference quantitative image after decomposition, is a small value to prevent the denominator from being zero; is the multi-modal image joint low-rank constraint loss function, By assigning different weights to singular values of different sizes, noise and artifacts can be effectively suppressed, while detailed tissue information can be effectively preserved;

[0040] Finally, the loss function of the multi-task network is expressed as follows:

[0041]

[0042] wherein, is the quantitative loss function for different tasks, is the trainable hyper-parameter for estimating the uncertainty of each task, is the likelihood modeling type identifier, whose value determines whether the output of each task is modeled by Gaussian likelihood or maximum likelihood, and the value of 1 corresponds to Gaussian likelihood modeling, and the value of 2 corresponds to maximum likelihood modeling; is the multi-modal image joint low-rank constraint loss term.

[0043] The application also provides a multi-parameter simultaneous quantitative system based on the SSFP-MOLED and the 2.5D multi-task model, comprising:

[0044] a sequence design module, used for designing a multi-slice echo steady free precession sequence with multiple contrasts and determining specific acquisition parameters; the multi-slice echo steady free precession sequence is any magnetic resonance imaging sequence containing a non-equilibrium steady free precession sequence to generate multi-slice echoes under multiple radio frequency pulse excitations;

[0045] a data acquisition module, used for performing 3D mode scanning on a to-be-measured object by using the multi-slice echo steady free precession sequence under the set acquisition parameters, and obtaining magnetic resonance imaging signal data of the to-be-measured object;

[0046] a data simulation module, used for compiling the multi-slice echo steady free precession sequence on a simulation platform, performing simulated data acquisition on a virtual imaging object according to the set sampling parameters, and then performing zero padding, domain transformation, noise adding and other processing on the acquired k-space data to obtain a series of simulation samples for deep learning network training;

[0047] a multi-task learning network training module, used for designing a multi-task deep learning network fusing interlayer information, comprising a shared feature extraction module, an interlayer information fusion module, a specific task self-attention module, a cross-task learning module and a decoder module; and designing a loss function comprising multi-task loss dynamic weighting and multi-modal image joint low-rank constraint, used for balancing the learning speed of each task, relieving gradient conflicts and improving quantitative map quality;

[0048] a multi-parameter quantitative image reconstruction module, used for mapping the magnetic resonance image of an actual imaging object by using the trained multi-task deep learning network to obtain quantitative images of multiple parameters (M0, T1, T2, T2 * , B0 and B1).

[0049] Compared with the prior art, the application has the following beneficial effects:

[0050] 1. The multi-echo steady-state free precession sequence is designed in the application, and a multi-echo image with high signal-to-noise ratio and multiple contrasts can be obtained.

[0051] 2. The application can reconstruct quantitative images of M0, T1, T2, T2 * , B0 and B1 from a single scan of a magnetic resonance image without multiple acquisitions.

[0052] 3. The application can simultaneously output multiple quantitative images by designing a multi-task deep learning network to jointly reconstruct multiple quantitative parameters without training a separate network for each parameter quantitative task.

[0053] 4. The multi-task deep learning network designed in the application is based on a progressive hierarchical extraction structure, which continuously decouples features into specific task representations and general representations shared by all tasks through a dynamic gating mechanism, thereby achieving efficient information exchange and knowledge complementation between tasks. Not only does it improve the generalization ability of the model, but also improves the accuracy of parameter quantification. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The flowchart of the magnetic resonance multi-parameter quantitative imaging method of the embodiment of the application;

[0055] Figure 2 The multi-echo steady-state free precession sequence diagram of the embodiment of the application;

[0056] Figure 3 The multi-task deep learning network structure diagram capable of fusing interlayer information in the embodiment of the application;

[0057] Figure 4 The multi-echo image acquired by the multi-echo steady-state free precession sequence in the embodiment of the application; wherein the upper row is a k-space image, and the lower row is a weighted image;

[0058] Figure 5 The T1, T2, T2 * , B0 and B1 quantitative images reconstructed in the embodiment of the application and the quantitative images obtained by fitting the traditional method and the quantitative images reconstructed by the single-task U-Net network are compared; wherein Reference in (a) is the quantitative image obtained by fitting the traditional method; Ours is the quantitative image reconstructed by the multi-echo free precession sequence and the multi-task learning network in the application; (b) is the T1, T2, T2 * image reconstructed by the traditional method, the single-task U-Net network and the multi-task deep learning network, and the local magnification detail comparison. In Reference, the traditional acquisition sequence used for each parameter is respectively: : IR-TSE sequence; : SE sequence; : GRE sequence; : TSE sequence;

[0059] Figure 6 The structure diagram of a magnetic resonance multi-parameter quantitative imaging system is disclosed in the embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the present application clearer and more explicit, the specific embodiments of the present application will be described in detail below with reference to the drawings. It should be noted that the following described embodiments are only representative embodiments of the present application, which are intended to help understand the core idea of the present application, and are not a limitation on the scope of protection. Based on the disclosed embodiments of the present application, any person skilled in the art can obtain all technical solutions through logical deduction, technical superposition or structural transformation without departing from the concept of the present application, which are all considered to fall within the protection scope of the claims of the present application.

[0061] Referring to Figure 1 The embodiment discloses a flowchart of a multi-parameter quantitative method based on a multi-overlapping echo steady-state free precession and a multi-task learning network, and specifically includes:

[0062] S1, a multi-overlapping echo steady-state free precession sequence is designed and acquisition parameters are determined, and the multi-overlapping echo steady-state free precession sequence is any magnetic resonance imaging sequence containing a non-equilibrium steady-state free precession sequence to generate multi-overlapping echoes under multiple radio frequency pulse excitations;

[0063] S2, a deep neural network training sample is generated; specifically including:

[0064] S21, a virtual imaging object is generated;

[0065] S22, the multi-overlapping echo steady-state free precession sequence is programmed on a Bloch simulation platform, and simulated data acquisition is performed on the virtual imaging object according to the determined sampling parameters, to obtain a magnetic resonance imaging signal of the virtual imaging object;

[0066] S23, the magnetic resonance imaging signal of the virtual imaging object is subjected to zero padding, inverse Fourier transform, addition of Gaussian noise and other operations, to obtain a series of simulation samples;

[0067] S3, a multi-task deep learning network is designed and trained with the training sample; specifically including:

[0068] S31, a multi-task deep learning network is designed, including a feature extraction module, an inter-layer information fusion module, a specific task self-attention module, a cross-task learning module and a decoder module;

[0069] S32, an efficient network loss function is designed;

[0070] S33 training a multi-task deep learning network with the simulation samples;

[0071] S4 acquiring data of an actual imaging object in an advanced magnetic resonance instrument using the multi-shot multi-echo steady state free precession sequence and according to determined acquisition parameters.

[0072] S5 mapping the magnetic resonance image of the actual imaging object using the trained multi-task deep learning network to obtain quantitative images of multiple parameters (M0, T1, T2, T2 * , B0, B1).

[0073] Specifically, in S1, the multi-shot multi-echo steady state free precession sequence is designed and the acquisition parameters are determined, specifically including:

[0074] The multi-shot multi-echo steady state free precession sequence is based on the steady state free precession sequence, and equal size shift gradients are applied in the phase encoding and frequency encoding directions, causing imbalance of the TR gradient, and the continuous excitation of the radio frequency pulse can form a series of overlapping echoes in the k-space diagonal direction.

[0075] The number of echo chains n is determined, the repetition time is determined, the echo time of each echo chain is determined, the size, interval, phase and pulse shape of the excitation pulse flip angle are determined, the size, direction and application time of the shift gradient are determined, and the imaging field of view, imaging matrix, number of acquisition layers, layer thickness, layer interval, and bandwidth are determined.

[0076] Specifically, in S31, the multi-task deep learning network is designed, including a shared feature extraction module, an inter-layer information fusion module, a specific task self-attention module, a cross-task learning module, and a decoder module, specifically including:

[0077] The shared feature extraction module extracts shallow general features between tasks based on residual convolution; the inter-layer information fusion module learns the correlation between adjacent image slices based on the cross-attention mechanism to capture three-dimensional image features; the specific task self-attention module learns the representation of the specific task from the shallow general features based on the self-attention mechanism; the cross-task learning module is based on a progressive hierarchical extraction structure, and continuously decouples the features into specific task representations and general representations shared by all tasks through a dynamic gating mechanism, thereby realizing efficient information exchange and knowledge complementation between tasks. Finally, the decoder module outputs multiple quantitative images of specific sizes through a series of upsampling operations and jump connections.

[0078] Specifically, in S32, an efficient network loss function is designed, specifically including:

[0079] We design a multi-task loss dynamic weighting strategy, employing a loss weight dynamic adjustment method based on task uncertainty, by introducing trainable hyperparameters. To estimate the uncertainty of each task, tasks with high uncertainty are assigned smaller weights, and tasks with low uncertainty are assigned larger weights, in order to balance the contribution of each task to the model training. The formula is as follows:

[0080]

[0081] in, Quantitative loss functions for different tasks; These are trainable hyperparameters used to estimate the uncertainty for each task; Their values ​​determine whether the output of each task is modeled using Gaussian likelihood or maximum likelihood.

[0082] A multimodal image joint low-rank constraint method is designed. Based on the prior information of low-rank in quantitative images, T-SVD tensor decomposition and weighted nuclear norm method are used to jointly constrain the low-rank of multiple quantitative images output by the network. The formula is as follows:

[0083]

[0084]

[0085]

[0086]

[0087] in, A three-dimensional tensor composed of multiple quantitative images output by the network; The singular value matrix is ​​obtained after the three-dimensional tensor is decomposed by T-SVD tensor decomposition. Singular value weights are the singular values ​​obtained after decomposing a reference quantitative image. The reciprocal, To prevent the denominator from being zero; By assigning different weights to singular values ​​of different sizes, noise and artifacts can be effectively suppressed while effectively preserving detailed tissue information.

[0088] Finally, the loss function formula for the multi-task network is expressed as follows:

[0089]

[0090] The following detailed description of the process of a multi-parameter quantitative method based on multi-overlapping echo steady-state free precession and multi-task learning network will be provided through a specific embodiment, including the following steps.

[0091] Step 1: design a multi-shot steady-state free precession sequence and determine the acquisition parameters;

[0092] Step 2: generate training samples for the deep neural network;

[0093] Step 3: design and train a multi-task deep learning network with the simulation samples;

[0094] Step 4: use the multi-shot steady-state free precession sequence in an advanced magnetic resonance instrument and collect data from an actual imaging object according to the determined acquisition parameters.

[0095] Step 5: use the trained multi-task deep learning network to map the images of the actual imaging object to obtain quantitative images of multiple parameters (M0, T1, T2, T2 * , B0, B1).

[0096] In step 1, the design of the multi-shot steady-state free precession sequence refers to Figure 2 . The acquisition parameters of the multi-shot steady-state free precession sequence are as follows: the radio frequency excitation pulse is a Sinc pulse with an initial flip angle of , an interval of ; , , , and , , , are the shift gradients applied in the frequency encoding dimension and the phase encoding dimension, with relative areas of , , , ; the repetition time TR and the echo times TE1, TE2, TE3, TE4, TE5 are 18 ms, 3 ms, 6 ms, 9 ms, 12 ms, and 15 ms, respectively; the imaging field of view is ; the sampling matrix is , and the imaging matrix is ; the number of acquisition layers is 65; the layer thickness is 2 mm; the interlayer interval is 2 mm; and the bandwidth is .

[0097] In step 2, the magnetic resonance imaging signals of the virtual imaging object are zero-padded from to , and after inverse Fourier transform, about 25 dB of Gaussian noise is added in the image domain to serve as network training samples.

[0098] In step 3, the multi-task deep learning network structure refers to Figure 3In the shared feature extraction module, the number of channels in the output feature maps of each convolutional layer are 256, 512, 1024, and 2048, respectively. In the task-specific self-attention module, each task extracts specific representations from shallow general features, reducing the dimensionality of the feature maps to 512. In the cross-task learning module, a shared module is introduced between tasks to capture general features across different tasks. Then, each task extracts features that are advantageous to itself layer by layer from the shared layer. The transfer of all features is based on the cross-attention mechanism of the gating structure to obtain the weights. The outputs of the shared module and the task-specific module are expressed by the following formulas:

[0099]

[0100]

[0101]

[0102]

[0103] In this cross-task learning module, For the first The output of the layer shared feature module is used to capture common features across different tasks. Then it is the first Shared features of layers; It is the first The output of the i-th specific task module in the layer corresponds to the feature representation of a single parameter task such as M0, T1, etc. For the first The output of the i-th specific task module in the layer; Represents the total number of tasks as a quantitative parameter; It is a multi-head cross-attention module that enables cross-task information interaction between shared features and specific task features; for Modules are used to stabilize feature distributions; It is a multilayer perceptron module used for feature nonlinear transformation; and These are the shared feature module and the specific task module. and The processed intermediate features. This module achieves knowledge transfer through bidirectional interaction between shared features and task-specific features: when shared features are updated, the shared features of the previous layer interact with all task-specific features of the previous layer. After interaction, then Enhanced When a specific task feature is updated, the i-th specific task feature shares features with the previous layer. After interaction, then Enhanced , fully utilizing the correlation between multiple parameters of magnetic resonance, achieving high-precision quantification of multiple parameters simultaneously in one scan, and balancing task specificity and information complementarity.

[0104] In step 3, in the multi-task deep learning network, in order to avoid negative transfer, M0, T1, T2, T2 * The tasks are divided into a group for common training; the B0 and B1 tasks are branched separately to obtain the final output through a traditional U-Net network;

[0105] In step 3, the training set and the validation set respectively contain 4000 and 500 training samples, each training sample contains the real part and the imaginary part data of 5 echo chains collected once, a total of 10 channels; the model training adopts the Adam optimizer, the initial learning rate is 0.0001, and the decay coefficient is 0.8; in the loss function, All take 2, adopt maximum likelihood modeling, M0, T1, T2, T2 * The loss function of the task is the structural similarity index, and the loss functions of the B0 and B1 tasks are mean square errors.

[0106] In step 4, Figure 4 The k-space image and the weighted image obtained by using the multi-echo steady-state free precession sequence to collect actual human brain data. Figure 4 The upper row in the figure is k-space data (kspce1-kspce5) collected by the multi-echo steady-state free precession sequence, Figure 4 The lower row in the figure is the corresponding image domain weighted image (image1-image5). The distribution of multiple echoes in the upper row k-space (the position and intensity of different brightness points) reflects the characteristics of multi-echo; the clear human brain structure (such as brain gray matter, white matter, ventricle, etc.) in the lower row image domain proves that the sequence has high signal-to-noise ratio and contrast discrimination ability. These images directly verify the effectiveness of the sequence design of the application, and multiple contrast echo signals can be obtained simultaneously in one scan.

[0107] In step 5, Figure 5 In (a) of the figure, Reference is the quantitative image fitted by the traditional method; Ours is the quantitative image reconstructed by the multi-echo free precession sequence and the multi-task learning network of the application; it can be seen that the T1, T2, T2 * The images not only have clear human brain texture structure, but also have very accurate quantitative measurement values; on the B0 and B1 images, the application is also very consistent with the distribution of the inhomogeneous field measured by the gold standard method. Figure 5 In (b) of the figure, the T1, T2, T2 *The local enlarged detail comparison of the images shows that, compared with the U-Net network, the multi-task deep learning network reconstructs more abundant texture structures on the T1 images, and obtains clearer quantitative results on the T2, T2 * images, which shows the advantages of the multi-task learning network in quantitative accuracy and robustness.

[0108] Referring to Figure 6 the drawings, the embodiment also discloses a multi-parameter quantitative system based on the multi-echo steady-state free precession and the multi-task learning network, which comprises:

[0109] A sequence design module 61 is used for designing a multi-echo steady-state free precession sequence with multiple contrasts and determining specific acquisition parameters; the multi-echo steady-state free precession sequence is any magnetic resonance imaging sequence containing a non-equilibrium steady-state free precession sequence to generate multiple echoes under multiple radio frequency pulse excitations.

[0110] A data acquisition module 62 is used for performing 3D mode scanning on a to-be-measured object by using the multi-echo steady-state free precession sequence under the set acquisition parameters, so as to obtain magnetic resonance imaging signal data of the to-be-measured object.

[0111] A data simulation module 63 is used for writing the multi-echo steady-state free precession sequence on a simulation platform, performing simulated data acquisition on a virtual imaging object according to the set sampling parameters, and then performing zero padding, domain transformation, noise adding and other processing on the acquired k-space data, so as to obtain a series of simulation samples for deep learning network training.

[0112] A multi-task learning network training module 64 is used for designing a multi-task deep learning network fusing interlayer information, which comprises a shared feature extraction module, an interlayer information fusion module, a specific task self-attention module, a cross-task learning module and a decoder module; a loss function comprising multi-task loss dynamic weighting and multi-modal image joint low-rank constraint is designed, which is used for balancing the learning speed of each task, relieving the gradient conflict and improving the quality of quantitative images.

[0113] A multi-parameter quantitative image reconstruction module 65 is used for mapping the magnetic resonance images of an actual imaging object by using the trained multi-task deep learning network, so as to obtain quantitative images of multiple parameters (M0, T1, T2, T2 * , B0 and B1).

[0114] The principles and operation modes of the present application are described through the above specific implementation cases, which aims to provide a clear understanding framework to help grasp the core idea and operation points of the present application. However, it should be clear that these cases should not be regarded as a limitation on the application range of the present application, and all kinds of improvements and innovations based on the core idea of the present application should be included in the protection scope of the present application.

Claims

1. A multi-parameter quantitative method based on multi-overlapping echo steady-state free precession and multi-task learning network, characterized in that, include: S1. Design a multi-overlapping echo steady-state free precession sequence and determine the acquisition parameters. The multi-overlapping echo steady-state free precession sequence is an arbitrary magnetic resonance imaging sequence that includes a non-equilibrium steady-state free precession sequence to generate multi-overlapping echoes under multiple excitations by radio frequency pulses. S2. Generate deep neural network training samples; specifically including: S21. Generate virtual imaging objects covering the parameter range of normal and diseased tissues; S22. Write the multi-overlapping echo steady-state free precession sequence on the Bloch simulation platform and perform simulated data acquisition on the virtual imaging object according to the determined sampling parameters to obtain the magnetic resonance imaging signal of the virtual imaging object. S23. The magnetic resonance imaging signal of the virtual imaging object is zero-filled, inverse Fourier transformed, and Gaussian noise is added to obtain a series of simulation samples; S3. Design and train a multi-task deep learning network using training samples; specifically including: S31. Design a multi-task deep learning network, including a shared feature extraction module, an inter-layer information fusion module, a task-specific self-attention module, a cross-task learning module, and a decoder module; S32. Design an efficient network loss function; S33. Train a multi-task deep learning network using simulated samples; S4. In an advanced magnetic resonance imaging system, the aforementioned multi-overlapping echo steady-state free precession sequence is used to acquire data from the actual imaging object according to the determined acquisition parameters. S5. The trained multi-task deep learning network is used to map the magnetic resonance image of the actual imaging object to obtain multiple parameters M0, T1, T2, T3. * Quantitative images of B0 and B1.

2. The multi-parameter quantitative method based on multi-overlapping echo steady-state free precession and multi-task learning network according to claim 1, characterized in that, In step S1, the multi-overlapping echo steady-state free precession sequence is designed and the acquisition parameters are determined, specifically including: Based on the steady-state free precession sequence, the multi-overlapping echo steady-state free precession sequence applies an equal shift gradient in the phase encoding and frequency encoding directions, causing an imbalance in the gradient within the TR. In addition, with the continuous excitation of radio frequency pulses, a series of overlapping echoes can be formed in the diagonal direction of k-space. Determine the number of echo trains, n; determine the repetition time; determine the echo time of each echo train; determine the excitation pulse flip angle. The size, spacing, phase, and pulse shape of the pulse are determined; the magnitude, direction, and application time of the shift gradient are determined; and the imaging field of view, imaging matrix, number of acquisition layers, layer thickness, interlayer spacing, and bandwidth are determined.

3. The multi-parameter quantitative method based on multi-overlapping echo steady-state free precession and multi-task learning network according to claim 1, characterized in that, In step S31, a multi-task deep learning network is designed, including a shared feature extraction module, an inter-layer information fusion module, a task-specific self-attention module, a cross-task learning module, and a decoder module, specifically including: The shared feature extraction module extracts shallow common features between tasks based on residual convolution; the inter-layer information fusion module learns the correlation between adjacent image slices based on the cross-attention mechanism to capture 3D image features; the task-specific self-attention module learns the representation of the specific task from the shallow common features based on the self-attention mechanism. The cross-task learning module is based on a progressive hierarchical extraction structure. Through a dynamic gating mechanism, it continuously decouples features into task-specific representations and a common representation shared by all tasks, thereby achieving efficient information exchange and knowledge complementarity between tasks. Specifically, a shared module is introduced among tasks to capture common features across different tasks. Then, each task extracts features beneficial to itself layer by layer from the shared layer, as expressed in the following formula: in, and These are the outputs of the shared expert at layer k and the task-specific expert, respectively. The weights output by the gating structure of shared experts and task-specific experts at layer k; Finally, the decoder module outputs multiple quantitative images of a specific size through skip connections and a series of upsampling operations.

4. The multi-parameter quantitative method based on multi-overlapping echo steady-state free precession and multi-task learning network according to claim 1, characterized in that, In step S32, an efficient network loss function is designed, specifically including: We design a multi-task loss dynamic weighting strategy, employing a loss weight dynamic adjustment method based on task uncertainty, by introducing trainable hyperparameters. To estimate the uncertainty of each task, tasks with high uncertainty are assigned smaller weights, and tasks with low uncertainty are assigned larger weights, in order to balance the contribution of each task to the model training. The formula is as follows: in, Quantitative loss functions for different tasks; These are trainable hyperparameters used to estimate the uncertainty for each task; Their values ​​determine whether the output of each task is modeled using Gaussian likelihood or maximum likelihood. A multimodal image joint low-rank constraint method is designed. Based on the prior information of low-rank in quantitative images, T-SVD tensor decomposition and weighted nuclear norm method are used to jointly constrain the low-rank of multiple quantitative images output by the network. The formula is as follows: in, For three-dimensional tensors Perform tensor singular value decomposition and output unitary factors. , and singular value matrix ; The singular value matrix is ​​obtained after the three-dimensional tensor is decomposed by T-SVD tensor decomposition. Singular value matrix The singular value in the i-th row and j-th column reflects the tensor feature strength; Singular value matrix the number of rows, The total number of tasks is a quantitative parameter. Singular value weights are the singular values ​​obtained after decomposing a reference quantitative image. The reciprocal, To prevent the denominator from being zero; For multimodal image joint low-rank constraint loss function, By assigning different weights to singular values ​​of different sizes, noise and artifacts are suppressed while effectively preserving detailed tissue information. Finally, the loss function formula for the multi-task network is expressed as follows: in, For different tasks, the quantitative loss function is... Trainable hyperparameters are used to estimate the uncertainty for each task. This is an identifier for the likelihood modeling type. Its value determines whether the output of each task uses Gaussian likelihood modeling or maximum likelihood modeling. A value of 1 corresponds to Gaussian likelihood modeling, and a value of 2 corresponds to maximum likelihood modeling. This is the joint low-rank constraint loss term for multimodal images.

5. A multi-parameter quantitative system based on multi-overlapping echo steady-state free precession and multi-task learning network, characterized in that, include: The sequence design module is used to design multi-contrast, multi-overlapping echo steady-state free precession sequences and determine specific acquisition parameters. The multi-overlapping echo steady-state free precession sequence is an arbitrary magnetic resonance imaging sequence that includes a non-equilibrium steady-state free precession sequence that generates multi-overlapping echoes under multiple excitations by radio frequency pulses. The data acquisition module, under the set acquisition parameters, uses the multi-overlapping echo steady-state free precession sequence to perform 3D mode scanning on the object under test, and obtains the magnetic resonance imaging signal data of the object under test. The data simulation module writes the multi-overlapping echo steady-state free precession sequence on the simulation platform and performs simulated data acquisition on the virtual imaging object according to the set sampling parameters. Then, it performs zero-filling, domain transformation, and noise addition processing on the acquired k-space data to obtain a series of simulation samples for deep learning network training. The multi-task learning network training module designs a multi-task deep learning network that integrates inter-layer information, including a shared feature extraction module, an inter-layer information fusion module, a task-specific self-attention module, a cross-task learning module, and a decoder module; a loss function that includes dynamic weighting of multi-task loss and joint low-rank constraint of multimodal image is designed to balance the learning speed of each task, alleviate gradient conflict, and improve the quality of quantitative maps. The multi-parameter quantitative image reconstruction module maps the trained multi-task deep learning network to the magnetic resonance image of the actual imaging object, obtaining multiple parameters M0, T1, T2, and T3. * Quantitative images of B0 and B1.

Images