Cardiac MRI (Magnetic Resonance Imaging) quantitative imaging motion correction and parameter graph reconstruction and segmentation method and system

Abstract

The invention provides a heart MRI quantitative imaging motion correction and parameter graph reconstruction and segmentation method and system. The method comprises the following steps: constructing a multi-task model; and constructing a motion-free simulation signal data set of the selected quantity imaging task, and pre-training the single-pixel parameter estimation network and the single-pixel signal simulation network based on the data set. And constructing a real image data set of the selected amount of imaging tasks, fixing network parameters obtained by pre-training, and training a multi-task model to determine values of residual parameters in the model. And performing motion correction, quantitative parameter graph reconstruction and region-of-interest segmentation on a newly acquired image sequence of the selected amount of imaging task through the trained multi-task model, and outputting a reconstructed quantitative parameter graph and a segmentation mask of the quantitative parameter graph. The method is more accurate and quicker, and can realize full-automatic and one-stop post-processing of heart quantitative imaging.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing. Specifically, it relates to a method and system for cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation, and particularly relates to a method for combined motion correction, quantitative parameter map reconstruction and segmentation of cardiac magnetic resonance quantitative imaging based on deep learning. Background Art

[0002] Cardiac magnetic resonance quantitative imaging can measure parameters such as T1, T2, and T2* of myocardial tissue, providing a basis for the diagnosis and grading of related diseases. Compared with traditional contrast-weighted imaging, the parameter values provided by quantitative imaging have clear physical and physiological meanings, and are therefore more objective, accurate and easy to interpret. Typical cardiac magnetic resonance quantitative imaging techniques include: longitudinal relaxation time T1 quantitative imaging, transverse relaxation time T2 quantitative imaging, apparent transverse relaxation time T2* quantitative imaging, and multi-parameter combined quantitative imaging, etc. At present, cardiac magnetic resonance quantitative imaging has been widely applied in clinics and has gradually become the gold standard for non-invasive detection of myocardial tissue characteristics.

[0003] In principle, magnetic resonance quantitative imaging will acquire a series of weighted images at different time points, and then estimate the parameter values at each pixel through pixel-by-pixel curve fitting or dictionary matching, and then form a quantitative parameter map describing the spatial distribution of the parameters. Therefore, only when each frame of the images is spatially aligned with each other, that is, each pixel corresponds to the same tissue at different time points, can the accuracy of parameter estimation be ensured. However, affected by heartbeat and respiratory movements, misregistration often occurs between different frames of images, resulting in a decrease in the accuracy and precision of quantification. For example, some subjects may have arrhythmia, resulting in inaccurate electrocardiogram gating, making the heartbeat motion states between different frames of images different. Another example is that some subjects cannot hold their breath strictly as required, resulting in respiratory movements in the acquired image sequence. Therefore, it is necessary to perform retrospective motion correction, that is, image registration, on the quantitative imaging images before reconstructing the parameter map.

[0004] The main challenge in quantitative imaging motion correction lies in the significant contrast differences between images. Currently, there are mainly two categories of solutions: image feature-based methods and physics model-based methods. Most image feature-based methods utilize the characteristics of the images themselves to design contrast-robust dissimilarity measures. Among them, representative methods include: the mutual information-based registration algorithm, the normalized gradient field-based registration algorithm, and the principal component analysis-based group registration algorithm, etc. In recent years, with the rise of deep learning technology, some learnable dissimilarity measures have also been proposed. For example, a generative adversarial network is constructed, where the discriminator is responsible for predicting the probability that the images are not yet registered, serving as a dissimilarity measure, while the generator achieves registration by minimizing this probability. Another category of image feature-based methods first reduces the contrast differences between images through preprocessing and then uses conventional mean square error or cross-correlation as the dissimilarity measure for calibration. Among them, representative preprocessing methods include: histogram matching, structural characterization map extraction, and deep learning-based image modality conversion, etc. However, image feature-based methods make overly simplistic assumptions about the changes in image contrast and ignore the underlying physical mechanisms. Therefore, they are often inaccurate and not robust enough when dealing with images with drastic contrast changes. In contrast, physics model-based methods use the physical model of signal relaxation to generate simulation images with the same contrast as the real images and without motion, and correct the motion in the former by registering the real images to the simulation images, thus transforming the much more complex contrast registration problem into a relatively simple single-contrast registration problem for solution. Currently, the vast majority of physics model-based methods choose to generate simulation images using curve fitting, so they are not applicable to imaging applications where the signal model is difficult to analyze or fit, such as magnetic resonance fingerprinting and some multi-parameter quantitative imaging. To address this problem, the dictionary matching-based motion correction method uses dictionary matching instead of curve fitting to generate simulation images, thus getting rid of the dependence on the analytical signal model and greatly expanding the applicable range of physics model-based methods. However, both curve fitting and dictionary matching have relatively low computational efficiency. Moreover, existing physics model-based methods usually need to perform image simulation and registration alternately multiple times to converge, resulting in often taking dozens of seconds to process a conventional cardiac quantitative imaging image sequence of (160×160 pixels)×10 frames. In addition, in the case of a large degree of image mismatch, the alternating optimization process is prone to falling into a suboptimal solution, leading to the lack of robustness of such methods for correcting large-amplitude motion. Finally, existing quantitative imaging motion correction methods are not jointly optimized with the segmentation task. Therefore, these methods not only do not utilize the correlation between the registration and segmentation tasks to synergistically improve the performance of the algorithm in the two tasks, but also cannot provide the necessary segmentation results for subsequent quantitative analysis of the target anatomical structure parameter values.

[0005] Therefore, a new technical solution needs to be proposed. Summary of the Invention

[0006] Aiming at the defects in the prior art, the purpose of the present invention is to provide a method and system for cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation.

[0007] According to a method for cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation provided by the present invention, the method includes the following steps:

[0008] Step S1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction and segmentation, with the input being the image sequence of the selected quantitative imaging task collected, and the output being the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map;

[0009] Step S2: Construct a motion-free simulation signal data set for the selected quantitative imaging task, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module and the single-pixel signal simulation network in the simulation map sequence generation module based on this data set;

[0010] Step S3: Construct a real image data set for the selected quantitative imaging task, fix the network parameters obtained by pre-training in Step S2, and train the multi-task model in Step S1 to determine the values of the remaining parameters in this model;

[0011] Step S4: For the newly collected image sequence of the selected quantitative imaging task, perform motion correction, quantitative parameter map reconstruction and segmentation of the region of interest through the trained multi-task model, and output the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map.

[0012] Preferably, the multi-task model in Step S1 includes 6 modules and 2 loss functions; the 6 modules are respectively a reversible motion estimation module, a motion correction module, a quantitative parameter map reconstruction module, a simulation map sequence generation module, a quantitative parameter map segmentation module, and an image mask sequence generation module; the 2 loss functions are a model loss and a segmentation loss.

[0013] Preferably, the definitions of the 6 modules are as follows:

[0014] Reversible motion estimation module: Taking the collected image sequence as the input variable, generating a predicted velocity field sequence through a registration network, and then performing scaling and squaring operations on the velocity field sequence, and finally generating a forward motion field sequence and a backward motion field sequence; the velocity field sequence, the forward motion field sequence and the backward motion field sequence are the outputs of this module;

[0015] Motion correction module: Taking the collected image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, generating a deformed map sequence after motion correction as the output through a deformation operator;

[0016] Quantitative parameter map reconstruction module: Taking the sequence of deformation maps output by the motion correction module as the first input variable, and the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, generating T1- and T2-related quantitative parameter maps corresponding to the sequence of deformation maps as the output through a single-pixel parameter estimation network;

[0017] Simulation map sequence generation module: Taking the quantitative parameter maps output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, and the sequence of deformation maps output by the motion correction module as the third input variable, first generating the simulated signal at each pixel through a single-pixel signal simulation network or an analytical physical model, and then correcting the amplitude of the simulated signal with the actual signal in the sequence of deformation maps to generate a sequence of simulation maps as the output;

[0018] Quantitative parameter map segmentation module: Taking the quantitative parameter maps output by the quantitative parameter map reconstruction module as the input variable, segmenting the region of interest in the parameter maps through a segmentation network to generate a segmentation mask of the parameter maps as the output;

[0019] Image mask sequence generation module: Taking the parameter map mask output by the quantitative parameter map segmentation module as the first input variable, and the sequence of reverse motion fields output by the reversible motion estimation module as the second input variable, generating a sequence of image segmentation masks as the output through a deformation operator;

[0020] The definitions of the two loss functions are as follows:

[0021] Model loss: Taking the sequence of deformation maps output by the motion correction module as the first input variable, and the sequence of simulation maps output by the simulation map sequence generation module as the second input variable;

[0022] Segmentation loss: Taking the sequence of image masks output by the image mask sequence generation module as the first input variable, and the sequence of manually labeled label masks as the second input variable;

[0023] The multi-task model also includes two additional loss functions to improve the accuracy of motion correction, including smoothness loss and low-rank loss:

[0024] Smoothness loss: Taking the sequence of velocity fields output by the reversible motion estimation module as the input variable, and promoting the smooth change of the sequence of velocity fields and the sequence of motion fields generated by it in the spatial dimension by minimizing the spatial regularization term of the sequence of velocity fields. The adopted spatial regularization terms are the first-order diffusion regularization term and the second-order thin plate bending energy;

[0025] Low-rank loss: Taking the sequence of deformation maps output by the motion correction module as the input variable, the robustness of motion correction is improved by maximizing the low-rank property between the images after motion correction; each frame image in the sequence of deformation maps is flattened into a column vector and arranged in chronological order to form a Casorati matrix, and the nuclear norm of this Casorati matrix is calculated as the low-rank loss.

[0026] Preferably, in the step S2, the motion-free simulation signal dataset includes paired parameter value combinations, signal acquisition timestamps, and corresponding time-domain simulation signals; the parameter value combinations are generated by randomly sampling the parameter values; the signal acquisition timestamps are generated by randomly sampling the RR intervals of the heartbeat and adding random noise; the time-domain simulation signals corresponding to the combination of each parameter and the signal acquisition timestamp are generated by a physical model describing the evolution of magnetic resonance signals.

[0027] The single-pixel parameter estimation network takes the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the parameter value combination through a neural network; the loss function of the single-pixel parameter estimation network takes the estimated parameter value combination as the first input variable and the standard parameter value combination as the second input variable.

[0028] The single-pixel signal simulation network takes the parameter value combination as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the simulation signal through a neural network; the loss function of the single-pixel signal simulation network takes the estimated simulation signal as the first input variable and the standard simulation signal as the second input variable.

[0029] Preferably, in the step S3, the real image dataset of the selected quantitative imaging task is obtained by collecting the image data of the subject.

[0030] The present invention also provides a cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation system, and the system includes the following modules:

[0031] Module M1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction and segmentation, with the input being the image sequence of the selected quantitative imaging task, and the output being the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map.

[0032] Module M2: Construct a motion-free simulation signal dataset for the selected quantitative imaging task, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module and the single-pixel signal simulation network in the simulation map sequence generation module based on this dataset.

[0033] Module M3: Construct a real image dataset for the selected quantitative imaging task, fix the network parameters obtained by pre-training in Module M2, and train the multi-task model in Module M1 to determine the values of the remaining parameters in this model.

[0034] Module M4: For the newly acquired image sequence of the selected quantitative imaging task, perform motion correction, reconstruction of quantitative parameter maps, and segmentation of regions of interest through a trained multi-task model, and output the reconstructed quantitative parameter maps and the segmentation masks of the quantitative parameter maps.

[0035] Preferably, the multi-task model in the module M1 includes 6 modules and 2 loss functions; the 6 modules are respectively a reversible motion estimation module, a motion correction module, a quantitative parameter map reconstruction module, a simulated image sequence generation module, a quantitative parameter map segmentation module, and an image mask sequence generation module; the 2 loss functions are a model loss and a segmentation loss.

[0036] Preferably, the definitions of the 6 modules are as follows:

[0037] Reversible motion estimation module: Taking the acquired image sequence as the input variable, generating a predicted velocity field sequence through a registration network, then performing scaling and squaring operations on the velocity field sequence, and finally generating a forward motion field sequence and a backward motion field sequence; the velocity field sequence, the forward motion field sequence, and the backward motion field sequence are the outputs of this module;

[0038] Motion correction module: Taking the acquired image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, generating a deformed image sequence after motion correction as the output through a deformation operator;

[0039] Quantitative parameter map reconstruction module: Taking the deformed image sequence output by the motion correction module as the first input variable and the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, generating T1 and T2 related quantitative parameter maps corresponding to the deformed image sequence through a single-pixel parameter estimation network as the output;

[0040] Simulated image sequence generation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, and the deformed image sequence output by the motion correction module as the third input variable, first generating the simulated signal at each pixel through a single-pixel signal simulation network or an analytical physical model, and then correcting the amplitude of the simulated signal with the actual signal in the deformed image sequence to generate a simulated image sequence as the output;

[0041] Quantitative parameter map segmentation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the input variable, segmenting the region of interest in the parameter map through a segmentation network, and generating the segmentation mask of the parameter map as the output;

[0042] Image mask sequence generation module: Using the parameter map mask output by the quantitative parameter map segmentation module as the first input variable, and the reverse motion field sequence output by the reversible motion estimation module as the second input variable, an image segmentation mask sequence is generated as the output through a deformation operator;

[0043] The definitions of the two loss functions are as follows:

[0044] Model loss: Using the deformation map sequence output by the motion correction module as the first input variable, and the simulation map sequence output by the simulation map sequence generation module as the second input variable;

[0045] Segmentation loss: Using the image mask sequence output by the image mask sequence generation module as the first input variable, and the manually labeled label mask sequence as the second input variable;

[0046] The multi-task model also includes two additional loss functions to improve the accuracy of motion correction, including smoothness loss and low-rank loss:

[0047] Smoothness loss: Using the velocity field sequence output by the reversible motion estimation module as the input variable, by minimizing the spatial regularization term of the velocity field sequence, it promotes the smooth change of the velocity field sequence and the motion field sequence generated by it in the spatial dimension. The spatial regularization terms adopted are the first-order diffusion regularization term and the second-order thin plate bending energy;

[0048] Low-rank loss: Using the deformation map sequence output by the motion correction module as the input variable, by maximizing the low-rank property between the images after motion correction to improve the robustness of motion correction; Flatten each frame image in the deformation map sequence into a column vector and arrange them in chronological order to form a Casorati matrix, and calculate the nuclear norm of this Casorati matrix as the low-rank loss.

[0049] Preferably, in the module M2, the motionless simulation signal data set includes paired parameter value combinations, signal acquisition timestamps, and corresponding time-domain simulation signals; The parameter value combinations are generated by randomly sampling the parameter values; The signal acquisition timestamps are generated by randomly sampling the RR intervals of the heartbeat and adding random noise; The time-domain simulation signals corresponding to the combinations of each parameter and the signal acquisition timestamp are generated by a physical model describing the evolution of magnetic resonance signals;

[0050] The single-pixel parameter estimation network uses the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the parameter value combination through a neural network; The loss function of the single-pixel parameter estimation network uses the estimated parameter value combination as the first input variable and the standard parameter value combination as the second input variable;

[0051] The single-pixel signal simulation network takes the parameter value combination as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the simulation signal through a neural network; the loss function of the single-pixel signal simulation network takes the estimated simulation signal as the first input variable and the standard simulation signal as the second input variable.

[0052] Preferably, in the module M3, the real image dataset of the selected quantitative imaging task is obtained by collecting the image data of the subject.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The present invention constructs a simulation image with the same contrast as the real image and without motion by using the physical model of the evolution of magnetic resonance signals, thereby transforming the much more complex multi-contrast registration problem into a relatively simple single-contrast registration problem to solve. Therefore, it is superior to the existing image feature-based methods in terms of accuracy and robustness.

[0055] 2. The present invention uses a pre-trained parameter estimation network and a signal simulation network to generate simulation images, rather than curve fitting or dictionary matching in the existing physical model-based methods. Therefore, it not only does not depend on an analytical signal model, but also greatly shortens the calculation time of image simulation; at the same time, this method uses a deep neural network to learn the mapping from the image to the motion field, thereby avoiding the alternating optimization process in the existing methods and increasing the processing speed by hundreds of times.

[0056] 3. The present invention adopts a multi-task framework of joint motion correction, quantitative parameter map reconstruction, and segmentation, which not only enables it to synergistically optimize the algorithm modules of each task by using the correlation between different tasks during the training stage, thereby further improving accuracy and robustness, but also enables it to perform fully automatic and one-stop post-processing on newly acquired images during the prediction stage, providing a basis for subsequent quantitative analysis. Description of the Drawings

[0057] By reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0058] Figure 1 It is the implementation flowchart of the method for joint motion correction, quantitative parameter map reconstruction, and segmentation of cardiac magnetic resonance quantitative imaging based on deep learning of the present invention;

[0059] Figure 2 It is the algorithm schematic diagram of the method for joint motion correction, quantitative parameter map reconstruction, and segmentation of cardiac magnetic resonance quantitative imaging based on deep learning of the present invention;

[0060] Figure 3Schematic diagram of the results of the free-breathing cardiac magnetic resonance T1 quantitative imaging combined with motion correction, quantitative parameter map reconstruction and segmentation method based on deep learning according to an embodiment of the present invention;

[0061] Figure 4 Schematic diagram of the results of the free-breathing cardiac magnetic resonance T1 and T2 quantitative imaging combined with motion correction, quantitative parameter map reconstruction and segmentation method based on deep learning according to an embodiment of the present invention. Detailed implementation manners

[0062] The present invention will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0063] Example 1:

[0064] Referring to Figure 1 and Figure 2 According to a method for motion correction, parameter map reconstruction and segmentation of cardiac MRI quantitative imaging provided by the present invention, the method includes the following steps:

[0065] Step S1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction and segmentation, with the input being the image sequence of the selected quantitative imaging task collected, and the output being the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map; preferably, the multi-task model includes 6 modules and 2 loss functions; the 6 modules are respectively a reversible motion estimation module, a motion correction module, a quantitative parameter map reconstruction module, a simulation map sequence generation module, a quantitative parameter map segmentation module, and an image mask sequence generation module; the 2 loss functions are a model loss and a segmentation loss.

[0066] Reversible motion estimation module: Taking the collected image sequence as the input variable, generating a predicted velocity field sequence through a registration network, then performing scaling and squaring operations on the velocity field sequence, and finally generating a forward motion field sequence and a reverse motion field sequence; the velocity field sequence, the forward motion field sequence and the reverse motion field sequence are the outputs of this module;

[0067] Motion correction module: Taking the collected image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, generating a motion-corrected deformation map sequence as the output through a deformation operator;

[0068] Quantitative parameter map reconstruction module: Using the deformation map sequence output by the motion correction module as the first input variable and the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, it generates T1- and T2-related quantitative parameter maps corresponding to the deformation map sequence as the output through a single-pixel parameter estimation network;

[0069] Simulation map sequence generation module: Using the quantitative parameter maps output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, and the deformation map sequence output by the motion correction module as the third input variable, it first generates the simulated signal at each pixel through a single-pixel signal simulation network or an analytical physical model, and then corrects the amplitude of the simulated signal with reference to the actual signal in the deformation map sequence to generate the simulation map sequence as the output;

[0070] Quantitative parameter map segmentation module: Using the quantitative parameter maps output by the quantitative parameter map reconstruction module as the input variable, it segments the region of interest in the parameter maps through a segmentation network to generate the segmentation mask of the parameter maps as the output;

[0071] Image mask sequence generation module: Using the parameter map mask output by the quantitative parameter map segmentation module as the first input variable and the reverse motion field sequence output by the reversible motion estimation module as the second input variable, it generates the image segmentation mask sequence as the output through a deformation operator;

[0072] Model loss: Using the deformation map sequence output by the motion correction module as the first input variable and the simulation map sequence output by the simulation map sequence generation module as the second input variable;

[0073] Segmentation loss: Using the image mask sequence output by the image mask sequence generation module as the first input variable and the manually annotated label mask sequence as the second input variable;

[0074] The multi-task model further includes 2 additional loss functions to improve the accuracy of motion correction, including smoothness loss and low-rank loss:

[0075] Smoothness loss: Using the velocity field sequence output by the reversible motion estimation module as the input variable, it promotes the smooth change of the velocity field sequence and the motion field sequence generated by it in the spatial dimension by minimizing the spatial regularization term of the velocity field sequence. The spatial regularization term adopted is the first-order diffusion regularization term and the second-order thin plate bending energy;

[0076] Low-rank loss: Using the deformation map sequence output by the motion correction module as the input variable, it improves the robustness of motion correction by maximizing the low-rank property between the images after motion correction; flattens each frame image in the deformation map sequence into a column vector and arranges them in chronological order to form a Casorati matrix, and calculates the nuclear norm of this Casorati matrix as the low-rank loss.

[0077] Step S2: Construct a motionless simulation signal dataset for the selected quantitative imaging task, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module and the single-pixel signal simulation network in the simulated image sequence generation module based on this dataset; the motionless simulation signal dataset includes paired parameter value combinations, signal acquisition timestamps, and corresponding time-domain simulation signals; the parameter value combinations are generated by randomly sampling the parameter values; the signal acquisition timestamps are generated by randomly sampling the RR intervals of the heartbeat and adding random noise; the time-domain simulation signals corresponding to the combination of each parameter and the signal acquisition timestamp are generated by a physical model describing the evolution of magnetic resonance signals.

[0078] The single-pixel parameter estimation network takes the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the parameter value combination through a neural network; the loss function of the single-pixel parameter estimation network takes the estimated parameter value combination as the first input variable and the standard parameter value combination as the second input variable.

[0079] The single-pixel signal simulation network takes the parameter value combination as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the simulation signal through a neural network; the loss function of the single-pixel signal simulation network takes the estimated simulation signal as the first input variable and the standard simulation signal as the second input variable.

[0080] Step S3: Construct a real image dataset for the selected quantitative imaging task, fix the network parameters obtained by pre-training in Step S2, and train the multi-task model in Step S1 to determine the values of the remaining parameters in the model; the real image dataset for the selected quantitative imaging task is obtained by collecting the image data of the subject.

[0081] Step S4: For the newly acquired image sequence of the selected quantitative imaging task, perform motion correction, quantitative parameter map reconstruction, and segmentation of the region of interest through the trained multi-task model, and output the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map.

[0082] The present invention also provides a cardiac MRI quantitative imaging motion correction, parameter map reconstruction, and segmentation system, and the cardiac MRI quantitative imaging motion correction, parameter map reconstruction, and segmentation system can be implemented by executing the process steps of the cardiac MRI quantitative imaging motion correction, parameter map reconstruction, and segmentation method, that is, those skilled in the art can understand the cardiac MRI quantitative imaging motion correction, parameter map reconstruction, and segmentation method as the preferred implementation manner of the cardiac MRI quantitative imaging motion correction, parameter map reconstruction, and segmentation system.

[0083] Example 2:

[0084] The present invention also provides a cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation system, which includes the following modules:

[0085] Module M1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction and segmentation. The input is the image sequence of the selected quantitative imaging task collected, and the output is the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map. The multi-task model includes 6 modules and 2 loss functions. The 6 modules are respectively a reversible motion estimation module, a motion correction module, a quantitative parameter map reconstruction module, a simulated image sequence generation module, a quantitative parameter map segmentation module, and an image mask sequence generation module. The 2 loss functions are a model loss and a segmentation loss.

[0086] Reversible motion estimation module: Taking the collected image sequence as the input variable, generating a predicted velocity field sequence through a registration network, then performing scaling and squaring operations on the velocity field sequence, and finally generating a forward motion field sequence and a backward motion field sequence. The velocity field sequence, the forward motion field sequence and the backward motion field sequence are the outputs of this module.

[0087] Motion correction module: Taking the collected image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, generating a deformed image sequence after motion correction as the output through a deformation operator.

[0088] Quantitative parameter map reconstruction module: Taking the deformed image sequence output by the motion correction module as the first input variable and the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, generating T1 and T2 related quantitative parameter maps corresponding to the deformed image sequence through a single-pixel parameter estimation network as the output.

[0089] Simulated image sequence generation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, and the deformed image sequence output by the motion correction module as the third input variable. First, generating the simulated signal at each pixel through a single-pixel signal simulation network or an analytical physical model, and then correcting the amplitude of the simulated signal with the actual signal in the deformed image sequence to generate a simulated image sequence as the output.

[0090] Quantitative parameter map segmentation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the input variable, segmenting the region of interest in the parameter map through a segmentation network, and generating the segmentation mask of the parameter map as the output.

[0091] Image mask sequence generation module: Taking the parameter map mask output by the quantitative parameter map segmentation module as the first input variable and the backward motion field sequence output by the reversible motion estimation module as the second input variable, generating an image segmentation mask sequence as the output through a deformation operator.

[0092] Model loss: The first input variable is the sequence of deformation maps output by the motion correction module, and the second input variable is the sequence of simulated images output by the simulated image sequence generation module;

[0093] Segmentation loss: The first input variable is the sequence of image masks output by the image mask sequence generation module, and the second input variable is the sequence of manually annotated label masks;

[0094] The multi-task model also includes 2 additional loss functions to improve the accuracy of motion correction, including smoothness loss and low-rank loss:

[0095] Smoothness loss: The input variable is the sequence of velocity fields output by the reversible motion estimation module. By minimizing the spatial regularization term of the velocity field sequence, it promotes the smooth change of the velocity field sequence and the motion field sequence generated by it in the spatial dimension. The spatial regularization terms adopted are the first-order diffusion regularization term and the second-order thin plate bending energy;

[0096] Low-rank loss: The input variable is the sequence of deformation maps output by the motion correction module. By maximizing the low-rank property between the images after motion correction, the robustness of motion correction is improved; each frame image in the deformation map sequence is flattened into a column vector and arranged in chronological order to form a Casorati matrix, and the nuclear norm of this Casorati matrix is calculated as the low-rank loss.

[0097] Module M2: Construct a dataset of motion-free simulation signals for the selected quantitative imaging tasks, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module and the single-pixel signal simulation network in the simulated image sequence generation module based on this dataset; the dataset of motion-free simulation signals includes paired parameter value combinations, signal acquisition timestamps, and corresponding time-domain simulation signals; the parameter value combinations are generated by randomly sampling the parameter values; the signal acquisition timestamps are generated by randomly sampling the RR intervals of the heartbeat and adding random noise; the time-domain simulation signals corresponding to the combination of each parameter and the signal acquisition timestamp are generated by a physical model describing the evolution of magnetic resonance signals;

[0098] The single-pixel parameter estimation network takes the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the parameter value combination through a neural network; the loss function of the single-pixel parameter estimation network takes the estimated parameter value combination as the first input variable and the standard parameter value combination as the second input variable;

[0099] The single-pixel signal simulation network takes the parameter value combination as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the simulation signal through a neural network; the loss function of the single-pixel signal simulation network takes the estimated simulation signal as the first input variable and the standard simulation signal as the second input variable.

[0100] Module M3: Construct a real image dataset for the selected quantitative imaging task, fix the network parameters pre-trained in Module M2, and train the multi-task model in Module M1 to determine the values of the remaining parameters in the model; the real image dataset for the selected quantitative imaging task is obtained by collecting the image data of the subject.

[0101] Module M4: For the newly acquired image sequence of the selected quantitative imaging task, perform motion correction, quantitative parameter map reconstruction, and segmentation of the region of interest through the trained multi-task model, and output the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map.

[0102] Example 3:

[0103] The following will introduce in detail an embodiment of the present invention applied to free-breathing cardiac magnetic resonance T1 quantitative imaging based on the STONE (Slice-Interleaved T1) sequence in conjunction with the accompanying drawings. Figure 1 This is the implementation flowchart of the present invention, Figure 2 This is the algorithm schematic diagram of the present invention. Specifically, this embodiment includes the following steps:

[0104] Step S1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction, and segmentation, with the input being the acquired STONE image sequence and the output being the reconstructed T1 map and the segmentation mask of the T1 map. This model includes 6 modules and 4 loss functions. The 6 modules are respectively a reversible motion estimation module, a motion correction module, a quantitative parameter map reconstruction module, a simulation map sequence generation module, a quantitative parameter map segmentation module, and an image mask sequence generation module; the 4 loss functions are a model loss, a segmentation loss, a smoothing loss, and a low-rank loss.

[0105] Among them, in Step S1, the definitions of the 6 modules are:

[0106] Reversible motion estimation module: The reversible motion estimation module takes the acquired STONE image sequence as the input variable, generates a predicted velocity field sequence through a registration network, then performs scaling and squaring operations on the velocity field sequence, and finally generates a forward motion field sequence and a backward motion field sequence. The velocity field sequence, the forward motion field sequence, and the backward motion field sequence are the outputs of this module. Among them, the architecture of the registration network is a three-dimensional U-Net.

[0107] Motion correction module: The motion correction module takes the acquired image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, and generates a sequence of deformed maps after motion correction as the output through a deformation operator.

[0108] Quantitative parameter map reconstruction module: The quantitative parameter map reconstruction module takes the sequence of deformed maps output by the motion correction module as the first input variable and the acquisition timestamps (inversion times) of each image in the STONE sequence as the second input variable, and generates the T1 map corresponding to the sequence of deformed maps as the output through a single-pixel parameter estimation network. Among them, the architecture of the single-pixel parameter estimation network is a multi-layer perceptron mechanism.

[0109] Simulated image sequence generation module: The simulated image sequence generation module takes the T1 map output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in the STONE sequence as the second input variable, and the sequence of deformed maps output by the motion correction module as the third input variable, generates the simulated signal at each pixel through the signal relaxation model of the STONE sequence, and then corrects the amplitude of the simulated signal with the actual signal in the sequence of deformed maps to generate a sequence of simulated images as the output. Among them, the signal relaxation model of the STONE sequence is:

[0110] M z (TI) = M0|1 - 2e -TI / T1 | (1)

[0111] Among them, M z is the longitudinal magnetization vector signal, M0 is the longitudinal magnetization vector signal at full recovery, TI is the inversion time, and T1 is the longitudinal relaxation time.

[0112] Quantitative parameter map segmentation module: The quantitative parameter map segmentation module takes the T1 map output by the quantitative parameter map reconstruction module as the input variable, and segments the region of interest in the T1 map through a segmentation network to generate a segmentation mask of the T1 map as the output. Among them, the architecture of the segmentation network is a 2D U-Net, and the 3 channels of the segmentation mask correspond to the left ventricular myocardium, the left ventricular blood pool, and the background respectively.

[0113] Image mask sequence generation module: The image mask sequence generation module takes the T1 map mask output by the quantitative parameter map segmentation module as the first input variable and the sequence of reverse motion fields output by the reversible motion estimation module as the second input variable, and generates a sequence of image segmentation masks as the output through a deformation operator.

[0114] Among them, in step S1, the definitions of the 4 loss functions are:

[0115] Model loss: The model loss takes the sequence of deformation maps output by the motion correction module as the first input variable and the sequence of simulated images output by the simulated image sequence generation module as the second input variable. By minimizing the per-pixel mean square error of the two input variables, it makes the two input variables as close as possible.

[0116] Segmentation loss: The segmentation loss takes the sequence of image masks output by the image mask sequence generation module as the first input variable and the sequence of manually annotated label masks as the second input variable. By minimizing the Dice coefficient of the two input variables, it makes the two input variables as close as possible.

[0117] Smoothing loss: The smoothing loss takes the sequence of velocity fields output by the reversible motion estimation module as the input variable. By minimizing the second-order thin plate bending energy of the velocity field sequence, it makes the changes of the velocity field sequence and the motion field sequence generated by it as smooth as possible in the spatial dimension.

[0118] Low-rank loss: The low-rank loss takes the sequence of deformation maps output by the motion correction module as the input variable. By maximizing the low-rank property between the images after motion correction, it improves the robustness of motion correction. Specifically, each frame image in the deformation map sequence is flattened into a column vector and arranged in chronological order to form a Casorati matrix, and the nuclear norm of this Casorati matrix is calculated as the low-rank loss.

[0119] Step S2: Construct a dataset of motionless simulation signals for the STONE sequence, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module based on this dataset.

[0120] Among them, in step S2, the dataset of motionless simulation signals includes paired T1 values, signal acquisition timestamps, and corresponding time-domain simulation signals. Among them, the T1 values are generated by random sampling within a specified range. Specifically, the range of T1 is specified as 50 ms to 3000 ms, and then 1000 T1 values are randomly and uniformly sampled. The signal acquisition timestamps are generated by randomly sampling the RR intervals of the heartbeat within a specified range and adding random noise of a specified intensity. Specifically, for each T1 value, the range of the RR interval is specified as 500 ms to 3000 ms, and then 10,000 RR intervals are randomly and uniformly sampled, and Gaussian random noise with an intensity of 0% - 50% of the RR interval is added. In this way, a total of 10 million combinations of T1 values and signal acquisition timestamps are obtained. The time-domain simulation signals corresponding to each combination of T1 value and signal acquisition timestamp are generated by the signal relaxation model of the STONE sequence.

[0121] Among them, in step S2, the single-pixel parameter estimation network takes the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the T1 value through a neural network. The loss function of the single-pixel parameter estimation network takes the estimated T1 value as the first input variable and the gold-standard T1 value as the second input variable, and promotes the two input variables to be as close as possible by minimizing the absolute value error between the two input variables.

[0122] Step S3: Construct a real-image dataset of the STONE sequence, fix the network parameters pre-trained in S2, and train the multi-task model in S1 to determine the values of the remaining parameters in the model. Among them, the real-image dataset of the STONE sequence is obtained by collecting the image data of the subject.

[0123] Step S4: For the newly acquired STONE image sequence, perform motion correction, quantitative parameter map reconstruction, and segmentation of the region of interest through the trained multi-task model, and output the reconstructed T1 map and the segmentation mask of the T1 map.

[0124] Figure 3 It is a schematic diagram of the results of this embodiment. Among them, the three rows correspond to three subjects respectively; the first column is the T1 map before motion correction; the second to seventh columns are the T1 maps corrected by different contrast methods; the eighth column is the T1 map corrected by the method of the present invention; the ninth column is the segmented T1 map corrected by the method of the present invention. In the T1 map before motion correction, affected by respiratory motion, severe blurring, artifacts, and even missing appear in the cardiac region. After correction by the contrast method, the quality of the T1 map is improved, but obvious blurring and artifacts, as well as partial myocardial missing, can still be observed. In contrast, after correction by the method of the present invention, the myocardium in the T1 map becomes complete, the boundary is clearer, and no obvious motion artifacts are included, and the overall quality is significantly improved. On this basis, the method of the present invention also accurately segments the myocardial region from the T1 map after motion correction, providing a basis for subsequent quantitative analysis of the myocardial T1 value.

[0125] Example 4:

[0126] The following will introduce in detail an embodiment of the present invention applied to free-breathing cardiac magnetic resonance T1 and T2 quantitative imaging based on the Multimapping sequence with reference to the accompanying drawings. Figure 1 It is the implementation flowchart of the present invention, Figure 2 It is the algorithm schematic diagram of the present invention. Specifically, this embodiment includes the following steps:

[0127] Step S1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction, and segmentation. The input is the acquired Multimapping image sequence, and the outputs are the reconstructed T1 and T2 maps and the segmentation masks of the T1 and T2 maps. This model includes six modules and four loss functions. The six modules are the reversible motion estimation module, the motion correction module, the quantitative parameter map reconstruction module, the simulated image sequence generation module, the quantitative parameter map segmentation module, and the image mask sequence generation module; the four loss functions are the model loss, the segmentation loss, the smoothing loss, and the low-rank loss.

[0128] Among them, in step S1, the definitions of the six modules are as follows:

[0129] Reversible motion estimation module: The reversible motion estimation module takes the acquired Multimapping image sequence as the input variable, generates a predicted velocity field sequence through the registration network, then performs scaling and squaring operations on the velocity field sequence, and finally generates a forward motion field sequence and a backward motion field sequence. The velocity field sequence, the forward motion field sequence, and the backward motion field sequence are the outputs of this module. Among them, the architecture of the registration network is a three-dimensional U-Net.

[0130] Motion correction module: The motion correction module takes the acquired image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, and generates a sequence of deformed images after motion correction as the output through the deformation operator.

[0131] Quantitative parameter map reconstruction module: The quantitative parameter map reconstruction module takes the sequence of deformed images output by the motion correction module as the first input variable and the acquisition timestamps (heartbeat RR intervals) of each image in the Multimapping sequence as the second input variable, and generates the corresponding T1 and T2 maps of the sequence of deformed images through the single-pixel parameter estimation network. Among them, the architecture of the single-pixel parameter estimation network is a multi-layer perceptron mechanism.

[0132] Simulated image sequence generation module: The simulated image sequence generation module takes the T1 and T2 maps output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in the Multimapping sequence as the second input variable, and the sequence of deformed images output by the motion correction module as the third input variable. First, it generates the simulated signal at each pixel through the single-pixel signal simulation network, and then corrects the amplitude of the simulated signal with the actual signal in the sequence of deformed images as a reference to generate the simulated image sequence as the output.

[0133] Quantitative parameter map segmentation module: The quantitative parameter map segmentation module takes the T1 and T2 maps output by the quantitative parameter map reconstruction module as input variables, and segments the regions of interest in the T1 and T2 maps through a segmentation network, generating segmentation masks for the T1 and T2 maps as outputs. Among them, the architecture of the segmentation network is a two-dimensional U-Net, and the three channels of the segmentation mask correspond to the left ventricular myocardium, the left ventricular blood pool, and the background respectively.

[0134] Image mask sequence generation module: The image mask sequence generation module takes the T1 and T2 map masks output by the quantitative parameter map segmentation module as the first input variable, and the reverse motion field sequence output by the reversible motion estimation module as the second input variable, and generates an image segmentation mask sequence as the output through a deformation operator.

[0135] Among them, in step S1, the definitions of the four loss functions are as follows:

[0136] Model loss: The model loss takes the deformation map sequence output by the motion correction module as the first input variable, and the simulation map sequence output by the simulation map sequence generation module as the second input variable, and promotes the two input variables to be as close as possible by minimizing the per-pixel mean square error of the two input variables.

[0137] Segmentation loss: The segmentation loss takes the image mask sequence output by the image mask sequence generation module as the first input variable, and the manually labeled label mask sequence as the second input variable, and promotes the two input variables to be as close as possible by minimizing the Dice coefficient of the two input variables.

[0138] Smoothing loss: The smoothing loss takes the velocity field sequence output by the reversible motion estimation module as the input variable, and promotes the changes in the velocity field sequence and the motion field sequence generated by it to be as smooth as possible in the spatial dimension by minimizing the second-order thin plate bending energy of the velocity field sequence.

[0139] Low-rank loss: The low-rank loss takes the deformation map sequence output by the motion correction module as the input variable, and improves the robustness of motion correction by maximizing the low-rank property between the images after motion correction. Specifically, each frame image in the deformation map sequence is flattened into a column vector and arranged in a Casorati matrix in chronological order, and the nuclear norm of this Casorati matrix is calculated as the low-rank loss.

[0140] Step S2: Construct a dataset of motionless simulation signals for the Multimapping sequence, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module and the single-pixel signal simulation network in the simulation map sequence generation module based on this dataset.

[0141] Among them, in step S2, the motion-free simulation signal dataset includes paired T1 values, T2 values, signal acquisition timestamps, and corresponding time-domain simulation signals. Among them, the combination of T1 and T2 values is generated by random sampling within a specified range. Specifically, the range of T1 is specified as 50 ms to 3000 ms, and the range of T2 is specified as 5 ms to 250 ms, and then 2500 combinations of T1 and T2 values are randomly and evenly sampled. The signal acquisition timestamp is generated by randomly sampling the RR interval of the heartbeat within a specified range and adding random noise with a specified intensity. Specifically, for each combination of T1 and T2 values, the range of the RR interval is specified as 500 ms to 3000 ms, and then 5000 RR intervals are randomly and evenly sampled, and Gaussian random noise with an intensity of 0% to 50% of the RR interval is added. In this way, a total of 12.5 million combinations of T1 values, T2 values, and signal acquisition timestamps are obtained. The time-domain simulation signals corresponding to each combination of T1 value, T2 value, and signal acquisition timestamp are generated by the signal relaxation model of Multimapping.

[0142] Among them, in step S2, the single-pixel parameter estimation network takes the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the T1 and T2 values through a neural network. The loss function of the single-pixel parameter estimation network takes the estimated T1 and T2 values as the first input variable and the gold-standard T1 and T2 values as the second input variable, and promotes the two input variables to be as close as possible by minimizing the absolute value error between the two input variables.

[0143] The single-pixel signal simulation network takes the combination of T1 and T2 values as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the simulation signal through a neural network. The loss function of the single-pixel signal simulation network takes the estimated simulation signal as the first input variable and the gold-standard simulation signal as the second input variable, and promotes the two input variables to be as close as possible by minimizing the mean square error between the two input variables.

[0144] Step S3: Construct a real image dataset of the Multimapping sequence, fix the network parameters pre-trained in S2, and train the multi-task model in S1 to determine the values of the remaining parameters in the model. Among them, the real image dataset of the Multimapping sequence is obtained by collecting the image data of the subject.

[0145] Step S4: For the newly collected Multimapping image sequence, perform motion correction, quantitative parameter map reconstruction, and segmentation of the region of interest through the trained multi-task model, and output the reconstructed T1 and T2 maps and the segmentation masks of the T1 and T2 maps.

[0146] Figure 4Schematic diagram of the results of this embodiment. Among them, the first to third rows are the T1 maps of three subjects respectively; the fourth to sixth rows are the T2 maps of these three subjects respectively; the first column is the T1 and T2 maps before motion correction; the second to seventh columns are the T1 and T2 maps after correction using different contrast methods; the eighth column is the T1 and T2 maps after correction using the method of the present invention; the ninth column is the T1 and T2 maps after correction and segmentation using the method of the present invention. In the T1 and T2 maps before correction, affected by respiratory motion, there are serious blurs, artifacts, and even missing parts in the cardiac region. The contrast method suppresses the errors caused by respiratory motion to a certain extent, but there are still some blurs, artifacts, and myocardial defects remaining. In contrast, the method of the present invention not only enables the myocardium to be completely presented, but also eliminates obvious blurs and artifacts, significantly improving the quality of the T1 and T2 maps. On this basis, the method of the present invention also accurately segments the myocardial region from the corrected T1 and T2 maps, providing a basis for subsequent quantitative analysis of myocardial T1 and T2 values.

[0147] Those skilled in the art can understand this embodiment as a more specific illustration of Embodiment 1 and Embodiment 2.

[0148] Those skilled in the art know that in addition to implementing the systems, devices, modules, and units provided by the present invention in the form of pure computer-readable program codes, the method steps can be logically programmed to enable the systems, devices, modules, and units provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers to achieve the same functions. Therefore, the systems, devices, modules, and units provided by the present invention can be regarded as a kind of hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as the structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as software modules for implementing the method and the structures within the hardware component.

[0149] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined arbitrarily.

Claims

1. A method for cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation, characterized in that The method includes the following steps: Step S1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction, and segmentation. The input is the image sequence of the selected quantitative imaging task collected, and the outputs are the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map. Step S2: Construct a motion-free simulation signal dataset for the selected quantitative imaging task, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module and the single-pixel signal simulation network in the simulation map sequence generation module based on this dataset. Step S3: Construct a real image dataset for the selected quantitative imaging task, fix the network parameters obtained from pre-training in Step S2, and train the multi-task model in Step S1 to determine the values of the remaining parameters in this model. Step S4: For the newly collected image sequence of the selected quantitative imaging task, perform motion correction, quantitative parameter map reconstruction, and segmentation of the region of interest through the trained multi-task model, and output the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map.

2. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation method according to claim 1, wherein The multi-task model in Step S1 includes 6 modules and 2 loss functions; the 6 modules are respectively the reversible motion estimation module, the motion correction module, the quantitative parameter map reconstruction module, the simulation map sequence generation module, the quantitative parameter map segmentation module, and the image mask sequence generation module; the 2 loss functions are the model loss and the segmentation loss.

3. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation method according to claim 2, wherein The definitions of the 6 modules are as follows: Reversible motion estimation module: Taking the collected image sequence as the input variable, generating a predicted velocity field sequence through the registration network, then performing scaling and squaring operations on the velocity field sequence, and finally generating a forward motion field sequence and a backward motion field sequence; the velocity field sequence, the forward motion field sequence, and the backward motion field sequence are the outputs of this module. Motion correction module: Taking the collected image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, generating a sequence of deformed maps after motion correction as the output through the deformation operator. Quantitative parameter map reconstruction module: Taking the sequence of deformed maps output by the motion correction module as the first input variable and the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, generating T1- and T2-related quantitative parameter maps corresponding to the sequence of deformed maps through the single-pixel parameter estimation network as the output. Simulation map sequence generation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, and the sequence of deformed maps output by the motion correction module as the third input variable, first generating the simulation signal at each pixel through the single-pixel signal simulation network or the analytical physical model, and then correcting the amplitude of the simulation signal with the actual signal in the sequence of deformed maps to generate a sequence of simulation maps as the output. Quantitative parameter map segmentation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the input variable, segmenting the region of interest in the parameter map through the segmentation network to generate the segmentation mask of the parameter map as the output. Image mask sequence generation module: Using the parameter map mask output by the quantitative parameter map segmentation module as the first input variable, and the reverse motion field sequence output by the reversible motion estimation module as the second input variable, it generates an image segmentation mask sequence as the output through a deformation operator; The definitions of the two loss functions are as follows: Model loss: Using the deformation map sequence output by the motion correction module as the first input variable, and the simulated image sequence output by the simulated image sequence generation module as the second input variable; Segmentation loss: Using the image mask sequence output by the image mask sequence generation module as the first input variable, and the manually annotated label mask sequence as the second input variable; The multi-task model also includes two additional loss functions to improve the accuracy of motion correction, including smoothness loss and low-rank loss: Smoothness loss: Using the velocity field sequence output by the reversible motion estimation module as the input variable, by minimizing the spatial regularization term of the velocity field sequence, it promotes the smooth change of the velocity field sequence and the motion field sequence generated by it in the spatial dimension. The spatial regularization terms adopted are the first-order diffusion regularization term and the second-order thin plate bending energy; Low-rank loss: Using the deformation map sequence output by the motion correction module as the input variable, by maximizing the low-rank property between the images after motion correction to improve the robustness of motion correction; Flatten each frame image in the deformation map sequence into a column vector and arrange them in chronological order to form a Casorati matrix, and calculate the nuclear norm of this Casorati matrix as the low-rank loss.

4. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation method according to claim 1, characterized in that In step S2, the motionless simulation signal dataset includes paired parameter value combinations, signal acquisition timestamps, and corresponding time-domain simulation signals; The parameter value combinations are generated by randomly sampling the parameter values; The signal acquisition timestamps are generated by randomly sampling the RR intervals of the heartbeat and adding random noise; The time-domain simulation signals corresponding to the combinations of each parameter and the signal acquisition timestamp are generated by a physical model describing the evolution of magnetic resonance signals; The single-pixel parameter estimation network uses the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the parameter value combination through a neural network; The loss function of the single-pixel parameter estimation network uses the estimated parameter value combination as the first input variable and the standard parameter value combination as the second input variable; The single-pixel signal simulation network uses the parameter value combination as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the simulation signal through a neural network; The loss function of the single-pixel signal simulation network uses the estimated simulation signal as the first input variable and the standard simulation signal as the second input variable.

5. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation method according to claim 1, wherein In step S3, the real image dataset of the selected quantitative imaging task is obtained by collecting the image data of the subject.

6. A cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation system, characterized in that, The system includes the following modules: Module M1: Construct a multi-task model for joint motion correction, quantitative parameter map reconstruction, and segmentation. The input is the image sequence of the selected quantitative imaging task collected, and the output is the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map; Module M2: Construct a motion-free simulation signal dataset for the selected quantitative imaging task, and pre-train the single-pixel parameter estimation network in the quantitative parameter map reconstruction module and the single-pixel signal simulation network in the simulation map sequence generation module based on this dataset; Module M3: Construct a real image dataset for the selected quantitative imaging task, fix the network parameters pre-trained in Module M2, and train the multi-task model in Module M1 to determine the values of the remaining parameters in this model; Module M4: For the newly acquired image sequence of the selected quantitative imaging task, perform motion correction, quantitative parameter map reconstruction, and segmentation of the region of interest through the trained multi-task model, and output the reconstructed quantitative parameter map and the segmentation mask of the quantitative parameter map.

7. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation system according to claim 6, wherein The multi-task model in Module M1 includes 6 modules and 2 loss functions; the 6 modules are respectively the reversible motion estimation module, the motion correction module, the quantitative parameter map reconstruction module, the simulation map sequence generation module, the quantitative parameter map segmentation module, and the image mask sequence generation module; the 2 loss functions are the model loss and the segmentation loss.

8. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation system according to claim 7, characterized in that, The definitions of the 6 modules are as follows: Reversible motion estimation module: Taking the acquired image sequence as the input variable, generating a predicted velocity field sequence through the registration network, then performing scaling and squaring operations on the velocity field sequence, and finally generating a forward motion field sequence and a backward motion field sequence; the velocity field sequence, the forward motion field sequence, and the backward motion field sequence are the outputs of this module; Motion correction module: Taking the acquired image sequence as the first input variable and the forward motion field output by the reversible motion estimation module as the second input variable, generating a sequence of deformed maps after motion correction as the output through the deformation operator; Quantitative parameter map reconstruction module: Taking the sequence of deformed maps output by the motion correction module as the first input variable and the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, generating T1 and T2 related quantitative parameter maps corresponding to the sequence of deformed maps through the single-pixel parameter estimation network as the output; Simulation map sequence generation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the first input variable, the acquisition timestamps of each image in magnetic resonance quantitative imaging as the second input variable, and the sequence of deformed maps output by the motion correction module as the third input variable, first generating the simulation signal at each pixel through the single-pixel signal simulation network or the analytical physical model, and then correcting the amplitude of the simulation signal with the actual signal in the sequence of deformed maps as the reference, generating a sequence of simulation maps as the output; Quantitative parameter map segmentation module: Taking the quantitative parameter map output by the quantitative parameter map reconstruction module as the input variable, segmenting the region of interest in the parameter map through the segmentation network, and generating the segmentation mask of the parameter map as the output; Image mask sequence generation module: Taking the parameter map mask output by the quantitative parameter map segmentation module as the first input variable and the backward motion field sequence output by the reversible motion estimation module as the second input variable, generating a sequence of image segmentation masks as the output through the deformation operator; The definitions of the 2 loss functions are as follows: Model loss: The first input variable is the sequence of deformation maps output by the motion correction module, and the second input variable is the sequence of simulated images output by the simulated image sequence generation module; Segmentation loss: The first input variable is the sequence of image masks output by the image mask sequence generation module, and the second input variable is the sequence of manually annotated label masks; The multi-task model also includes 2 additional loss functions to improve the accuracy of motion correction, including smoothness loss and low-rank loss: Smoothness loss: The input variable is the sequence of velocity fields output by the reversible motion estimation module. By minimizing the spatial regularization term of the velocity field sequence, it promotes the smooth change of the velocity field sequence and the motion field sequence generated by it in the spatial dimension. The spatial regularization term adopted is the first-order diffusion regularization term and the second-order thin plate bending energy; Low-rank loss: The input variable is the sequence of deformation maps output by the motion correction module. By maximizing the low-rank property between the images after motion correction, the robustness of motion correction is improved; Each frame image in the deformation map sequence is flattened into a column vector and arranged in chronological order to form a Casorati matrix, and the nuclear norm of this Casorati matrix is calculated as the low-rank loss.

9. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation system according to claim 6, characterized in that, In the module M2, the motionless simulation signal data set includes paired parameter value combinations, signal acquisition timestamps, and corresponding time-domain simulation signals; The parameter value combinations are generated by randomly sampling the parameter values; The signal acquisition timestamps are generated by randomly sampling the RR intervals of the heartbeat and adding random noise; The time-domain simulation signals corresponding to each combination of parameters and signal acquisition timestamps are generated by a physical model describing the evolution of magnetic resonance signals; The single-pixel parameter estimation network takes the simulation signal as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the parameter value combination through a neural network; The loss function of the single-pixel parameter estimation network takes the estimated parameter value combination as the first input variable and the standard parameter value combination as the second input variable; The single-pixel signal simulation network takes the parameter value combination as the first input variable and the signal acquisition timestamp as the second input variable, and estimates the simulation signal through a neural network; The loss function of the single-pixel signal simulation network takes the estimated simulation signal as the first input variable and the standard simulation signal as the second input variable.

10. The cardiac MRI quantitative imaging motion correction, parameter map reconstruction and segmentation system according to claim 6, wherein In the module M3, the real image data set of the selected quantitative imaging task is obtained by collecting the image data of the subjects.

Images