Cardiac magnetic resonance quantitative imaging frame and automatic post-processing system

Abstract

The invention relates to the technical field of medical imaging, and discloses an integrated cardiac magnetic resonance quantitative imaging framework, which realizes rapid and accurate quantitative image reconstruction and automatic analysis through a deep learning technology, simplifies an imaging process and performs standardized processing. The core of the system comprises S1, a graphical sequence construction module, a user can graphically combine a magnetic resonance quantitative sequence on line, programming is not needed, pulse events are flexibly configured, and T1, T2, T1rho and simultaneous T1 / T2 quantitative imaging are supported; s2, an end-to-end quantitative image reconstruction module which accurately reconstructs quantitative images of different sequences and parameters based on a recurrent neural network and a full-connection neural network; and S3, an automatic analysis module which adopts a two-dimensional convolutional neural network to realize automatic analysis of the quantitative image.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a cardiac magnetic resonance quantitative imaging framework and analysis system. Background Art

[0002] Magnetic resonance imaging (MRI) technology uses the interaction between hydrogen protons in the human body and an external magnetic field to obtain two relaxation times, and uses the adjustment of multiple scanning parameters to generate images with different contrasts. This enables MRI to perform multi-dimensional assessments of tissue structure, function, composition, blood flow, and metabolites, making it an important medical imaging examination method. Compared with imaging technologies such as computed tomography (CT), ultrasound, and positron emission tomography (PET), MRI has the advantages of high sensitivity, high resolution, wide spatial coverage, multi-angle imaging, no radiation, and good soft tissue contrast. Therefore, it has been widely used in the diagnosis of various diseases.

[0003] Cardiovascular Magnetic Resonance Imaging (CMR) parameter quantitative imaging technology (English: mapping) provides a unique non-invasive method to quantify the changes in myocardial tissue composition and measure the absolute relaxation time of the myocardium (such as T1, T2, T2 * and T1rho), changes in these parameters can indicate changes in myocardial tissue composition, such as fibrosis, which has made significant progress in precision medicine. CMR technology has gradually become an important tool for precision medicine because of its ability to detect myocardial fibrosis, edema and hemorrhage. In particular, T1rho quantitative imaging, which reflects the slow exchange rate between water and macromolecules, can quantify myocardial fibrosis without the need for contrast agents and has gradually attracted attention in clinical research. However, despite the unique capabilities of CMR, typical CMR parameter quantitative imaging examinations involve executing dedicated sequences, collecting different contrast images based on specific parameters, fitting relaxation times pixel by pixel to create magnetic resonance parameter quantitative images, and manually segmenting images to view global and local abnormalities. Existing technologies are complex, labor-intensive and time-consuming, and face challenges in accessibility, robustness and automation, which affect their widespread application.

[0004] Currently, modified Look-Locker Inversion Recovery (MOLLI), single-shot balanced steady-state free precession with T2 preparation pulse (T2-prep bSSFP), and gradient and spin echo (GraSE) sequences are widely used by major vendors for T1 and T2 quantitative imaging. However, many of these sequences are not widely available and can be challenging to implement, which limits their adoption and hinders clinical research and practice. The application of cardiac magnetic resonance imaging has the following problems: 1) Sequences for quantitative cardiac magnetic resonance imaging are usually pre-designed and programmed. Many sequences do not allow radiologists or researchers to adjust them according to the needs of specific subjects; 2) There is no universal framework in the field of cardiac magnetic resonance parameter quantitative imaging to standardize the implementation of sequences, and therefore it is impossible to enhance the reproducibility of cardiac magnetic resonance parameter quantitative imaging between different imaging centers and suppliers; 3) The most commonly used curve fitting method for constructing quantitative imaging is affected by multiple factors, including the algorithm used, the signal-to-noise ratio of the image, the initial conditions and the number of samples; 4) Data post-processing is complex and time-consuming, and existing CMR parameter quantitative imaging technology requires a lot of manual operations during the post-imaging process. This process usually involves manual segmentation, which is not only time-consuming, but also relies on experienced readers, and the results are affected by observer variations; 5) Most deep learning models are specifically trained and optimized for single parameter quantitative imaging, such as T1 or T2. For routine myocardial tissue examinations involving multiple cardiac magnetic resonance parameter quantitative imaging, multiple models are required, which increases the complexity of the analysis.

[0005] Therefore, there is an urgent need to develop an integrated cardiac magnetic resonance quantitative imaging framework that can construct the required imaging sequences online, and realize fast and accurate quantitative image reconstruction and automated analysis online, simplify the quantitative imaging process and quantitative imaging standardization processing, and simplify the application of cardiac magnetic resonance parameter quantitative imaging in myocardial tissue characterization, quickly realize the online construction of graphical sequence combinations, and realize a fast cardiac magnetic resonance quantitative image reconstruction and automated post-processing method to solve the problems existing in the above-mentioned cardiac magnetic resonance technology and further improve its capacity and promotion. Summary of the Invention

[0006] In response to the above-mentioned shortcomings of the existing technology, the present invention proposes a cardiac magnetic resonance quantitative imaging framework and automated post-processing system, which mainly solves the problems of difficult adjustment of cardiac magnetic resonance parameter quantitative imaging sequences, lack of a universal framework to standardize sequences, complex and time-consuming data post-processing, and the complexity of processing multiple cardiac magnetic resonance parameter quantitative imaging through different deep learning models.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: a cardiac magnetic resonance quantitative imaging framework and an automated post-processing system, comprising:

[0008] S1, a graphical online sequence construction module, provides a graphical interface on the MRI scanner for users to configure pulse events for each cardiac cycle through graphical combination, enabling online construction of target cardiac MRI quantitative imaging sequences such as T1, T2, and T1rho, without the need for programming;

[0009] S2, a quantitative image reconstruction module based on recurrent neural networks and fully connected neural networks, can automatically complete the reconstruction of T1, T2, and T1rho quantitative images of different sequences without manual user intervention;

[0010] S3 is an automated analysis module for cardiac magnetic resonance quantitative images based on a two-dimensional convolutional neural network. This module can automatically segment and analyze T1, T2, and T1rho quantitative images of different sequences.

[0011] Furthermore, the above-mentioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, the graphical online sequence construction module includes an online sequence construction framework, which consists of two parts: a graphical user interface for interactively defining magnetic resonance pulse events in each cardiac cycle, and a lookup algorithm module for executing sequences and magnetic resonance events in each cardiac cycle.

[0012] Furthermore, the aforementioned cardiac MRI quantitative imaging framework and automated post-processing system includes a graphical online sequence construction module that provides a graphical interface allowing users to interactively define MRI pulse events for each cardiac cycle, creating cardiac quantitative MRI sequences in real time and streamlining the sequence execution process. The constructed sequences are stored in a table, with each row representing a cardiac cycle and each column corresponding to a specific MRI event.

[0013] Furthermore, in the graphical online sequence construction module of the aforementioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, various mapping sequences can be constructed by using graphical sequence combinations on a user graphical interface. Each cardiac cycle is regarded as an independent unit block, allowing the user to arrange magnetic resonance radiofrequency pulses for each cardiac cycle, including saturation pulses, flip pulses, T2 preparation pulses, delay time, respiratory navigation, image acquisition, etc., and set specific parameters for them, defining the order of pulse events for each cardiac cycle to adapt to different scanning requirements. This graphical sequence combination method greatly simplifies the sequence construction process.

[0014] Furthermore, within the aforementioned quantitative cardiac MRI framework and automated post-processing system, the graphical online sequence construction module allows activation of the preparation pulse, saturation pulse, flip pulse, and image acquisition for each cardiac cycle by entering "yes" or "no." Delay times can also be manually set. T2 and T1rho preparation pulses can be indicated by entering an echo time greater than 0, and can be easily inserted before image acquisition. The remaining MRI pulse settings remain consistent throughout the cardiac cycle, eliminating the need for individual cycle-specific settings.

[0015] Furthermore, in the above-mentioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, in the graphical online sequence construction module, the magnetic resonance events and timing of each cardiac cycle are designed as follows: each cardiac cycle is regarded as an independent block, and the duration of each cardiac cycle is defined as T RR ; Image acquisition (ACQ) is set to be performed during diastole, and its duration is represented by T ACQ ; The idle time after acquisition is defined as T Dummy ;T Trigger represents the interval between the R wave and the start of acquisition; the fixed delay time after the R wave is denoted as T PostR , used to adapt to arrhythmic events.

[0016] Furthermore, in the above-mentioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, in the online sequence construction module, saturation and inversion pulses are used to prepare T1 weighting before image acquisition, and the time period between the two pulses and after the inversion pulse is designated as T Sat and T Inv ; Insert a delay period T before the saturation pulse pre , to adjust T Sat and T Inv , thereby modifying the T1 weighting. pre 、T Sat 、T Inv 、T Trigger With T PostR The relationship between them is formula (1):

[0017] T pre +T Sat +T Inv =T Trigger -T PostR

[0018] Furthermore, in the above-mentioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, a T2 or T1rho preparation pulse is inserted in the graphical online sequence construction module before image acquisition, and its duration is TE Prep, realize T2 or T1rho weighting; before image acquisition or T2 / T1rho preparation pulse, insert diaphragm breathing navigation, its duration is T Nav , construct a free-breathing imaging sequence.

[0019] Furthermore, the aforementioned cardiac MRI quantitative imaging framework and automated post-processing system includes a table within the online sequence construction framework that records the MRI events for each cardiac cycle in each user-created sequence. A table lookup algorithm module is then used to execute each sequence, each cardiac cycle, and each MRI event using two loops. Each cardiac cycle loop is triggered by the ECG, and for each heartbeat, the table lookup algorithm is used to retrieve and execute each MRI event.

[0020] Furthermore, in the above-mentioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, in S2, after the sequence scanning is completed, the deep learning quantitative image reconstruction model can encode the sampling signal and its corresponding preparation time online to generate the corresponding quantitative image on the scanner.

[0021] Furthermore, the aforementioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system utilizes a deep learning-based quantitative image reconstruction module that employs a one-dimensional recurrent neural network (RNN) containing eight bidirectional hidden layers. Each RNN input node contains three channels, representing the sampling signal and its corresponding T1 / T2 preparation time. Following the RNN, a one-dimensional fully connected neural network uses the least squares method to predict the A, B, T1, and T2 parameters, thereby achieving quantitative image reconstruction. The relationship between the four parameters is expressed as Equation (2):

[0022]

[0023] Where A is the image signal corresponding to the initial longitudinal magnetization vector, B is the image signal corresponding to the complete recovery of the longitudinal magnetization vector, td and te correspond to the delay time of T1 preparation and T2 / T1rho-prep, respectively. For T1 mapping, td represents the T in the T1 preparation module of Inversion-Recovery (IR). Inv , or T in the T1 preparation module of Saturation-Recovery (SR) Sat , and te is set to 0; for T2 or T1rho mapping, te corresponds to TE PREP , and td is set to infinity; for simultaneous T1 / T2 or T1 / T1rho mapping, td and te correspond to the delay times of T1 preparation and T2 / T1rho preparation, respectively.

[0024] Furthermore, in the quantitative image reconstruction module of the aforementioned cardiac magnetic resonance imaging framework and automated post-processing system, based on recurrent neural networks and fully connected neural networks, the last two layers of the fully connected neural network are repeated three times to improve accuracy. Each repetition receives the output of the sixth layer of the fully connected neural network and the mean absolute error between the measured signal and the predicted signal from the previous iteration. Furthermore, each hidden layer of the recurrent neural network and the fully connected network uses a leaky rectified linear unit (ReLU) as the activation function.

[0025] Furthermore, in the above-mentioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, S3, the cardiac magnetic resonance quantitative image automated analysis module includes an automated segmentation neural network that can automatically segment the left ventricular myocardium, ventricular septum, left ventricular blood, and identify the left and right ventricular sub-insertion points, and perform comprehensive global and local myocardial analysis of T1, T2, and T1rho quantitative images.

[0026] Furthermore, the above-mentioned cardiac magnetic resonance quantitative imaging framework and automated post-processing system, based on the cardiac magnetic resonance quantitative image automated analysis module of the two-dimensional convolutional neural network, uses a U-shaped two-dimensional convolutional neural network, consisting of a downsampling (encoder) block and four upsampling blocks (decoders). The four upsampling blocks share the encoded features extracted from different levels of the downsampling module to predict the masks of the left ventricular myocardium, the ventricular septum, the left ventricular blood region of interest, and the right ventricular inferior insertion point. The composite convolution layer includes a two-dimensional convolution layer, batch normalization and activated ReLU, followed by another two-dimensional convolution layer, batch normalization and ReLU. And the Sigmoid activation function is applied to the last layer of each of the four upsampling blocks to ensure that the predicted mask value is between 0 and 1. The predicted mask of the left ventricular myocardium is integrated into the last layer of the other three upsampling blocks as a spatial attention mechanism.

[0027] The beneficial effects of the present invention are as follows: the present invention provides a cardiac magnetic resonance quantitative imaging framework and an automated post-processing system, which can construct the required imaging sequence online, significantly simplify the cardiac quantitative imaging process, and realize fast and accurate quantitative image reconstruction and automated analysis through deep learning technology. This system simplifies the quantitative imaging process and quantitative imaging standardization processing, improves the operational convenience of cardiac magnetic resonance imaging, reduces dependence on technical personnel, and effectively reduces the complexity and time cost of manual operation. Through automated data post-processing and feature extraction, the system can perform comprehensive global and local myocardial analysis of cardiac magnetic resonance parameter quantitative images, thereby improving the efficiency of cardiology research and clinical diagnosis. In addition, the system supports quantitative imaging of multiple cardiac magnetic resonance parameters, enhances the detection ability of myocardial lesions, thereby providing strong technical support for precision medicine and promoting the widespread promotion of cardiac magnetic resonance imaging in clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the present invention;

[0029] Figure 2 A deployment of the present invention for implementing graphical online sequence construction through a user interface;

[0030] Figure 3 To realize the construction of simultaneous T1 / T2 quantitative imaging sequence in the graphical user interface in the present invention;

[0031] Figure 4 Schematic diagram of online construction of MOLLI sequence using graphical sequence combination in the present invention;

[0032] Figure 5 A table lookup algorithm module for executing sequences and pulse events in each cardiac cycle in the online sequence building framework of the present invention;

[0033] Figure 6 The network model for quantitative image reconstruction of T1, T2 and T1rho based on recurrent neural network and fully connected neural network in the present invention;

[0034] Figure 7 A network model for automatically analyzing and segmenting the left ventricular myocardium, septum, and left ventricular blood mask based on a two-dimensional convolutional neural network in cardiac magnetic resonance quantitative images, and identifying the right ventricular inferior insertion point;

[0035] Figure 8 The results of simultaneous T1 and T2 mapping and automatic segmentation performed by the present invention, as well as in vivo mapping and automatic segmentation of a commonly used cardiac T1 mapping sequence, are presented. DETAILED DESCRIPTION

[0036] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present application will be further described in detail with reference to the accompanying drawings. The described embodiments should not be regarded as limiting the present application. For ordinary technicians in this technical field, as long as various changes are within the scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

[0037] The present invention relates to the field of medical imaging technology and discloses an integrated cardiac magnetic resonance quantitative imaging framework that can construct the required imaging sequences online and achieve rapid and accurate quantitative image reconstruction and automated analysis through deep learning technology, aiming to simplify the quantitative imaging process and standardize quantitative imaging processing. The present invention includes three core components: S1, a graphical sequence construction module. The graphical sequence construction module of the present invention allows users to construct magnetic resonance cardiac quantitative sequences online through graphical combination without any programming requirements. Through the graphical sequence construction module of the present invention, users can flexibly configure the magnetic resonance pulse events of each cardiac cycle, including radiofrequency preparation pulses, signal recovery time, image acquisition modules, etc., to construct complex cardiac T1, T2, T1rho quantitative imaging sequences, including advanced simultaneous T1 / T2 quantitative imaging sequences. S2, an end-to-end quantitative image reconstruction module based on recurrent neural networks and fully connected neural networks. The present invention establishes a one-dimensional deep learning prediction model that can complete cardiac magnetic resonance quantitative image reconstruction for different sequences, different quantitative parameters (T1, T2, T1rho), and different signal lengths. S3, automated analysis module: The present invention establishes a two-dimensional convolutional neural network to perform automated analysis of quantitative images of different cardiac magnetic resonance parameters.

[0038] like Figures 1 to 7 As shown, the present invention provides a cardiac magnetic resonance quantitative imaging framework and an automated post-processing system, comprising:

[0039] S1, a graphical online sequence construction module, provides a graphical interface on the MRI scanner for users to configure pulse events for each cardiac cycle through graphical combination, enabling online construction of target cardiac MRI quantitative imaging sequences such as T1, T2, and T1rho, without the need for programming;

[0040] S2, a quantitative image reconstruction module based on recurrent neural networks and fully connected neural networks, can automatically complete the reconstruction of T1, T2, and T1rho quantitative images of different sequences without manual user intervention;

[0041] S3 is an automated analysis module for cardiac magnetic resonance quantitative images based on a two-dimensional convolutional neural network. This module can automatically segment and analyze T1, T2, and T1rho quantitative images of different sequences.

[0042] The present invention develops an application that uses graphical online sequence construction to simplify quantitative imaging of cardiac parameters. In terms of sequence construction, the present invention eliminates the need for programming during sequence design, allowing users to easily construct multiple complex mapping sequences and customize flexible sequences for specific patients. In terms of quantitative image reconstruction, the present invention simplifies the complexity of quantitative image construction. In terms of data post-processing analysis, the present invention facilitates comprehensive examination of myocardial tissue under different parameter mappings.

[0043] The present invention comprises a framework constructed using a graphical online sequence, which consists of two parts: a graphical user interface for interactively defining magnetic resonance pulse events in each cardiac cycle, and a lookup algorithm module for executing the sequence and the magnetic resonance events in each cardiac cycle.

[0044] To facilitate the configuration of magnetic resonance events within each cardiac cycle and the construction of target sequences, the present invention provides a graphical interface for online sequence construction. This allows users to interactively define magnetic resonance pulse events for each cardiac cycle, creating cardiac quantitative imaging magnetic resonance sequences in real time and simplifying the sequence execution process. The constructed sequence is stored in a table, with each row representing a cardiac cycle and each column corresponding to a specific magnetic resonance event.

[0045] In the online sequence construction module of the present invention, various mapping sequences can be constructed using graphical sequence combinations on a graphical user interface. Each cardiac cycle is treated as an independent unit block, allowing the user to schedule magnetic resonance radiofrequency pulses for each cardiac cycle, including saturation pulses, inversion pulses, T2 preparation pulses, delay time, respiratory navigation, image acquisition, and other parameters, and define the sequence of pulse events for each cardiac cycle to accommodate different scanning requirements. This graphical sequence combination greatly simplifies the sequence construction process.

[0046] In the online sequence construction module of the present invention, the preparation pulse, saturation pulse, flip pulse, and image acquisition for each cardiac cycle can be activated by entering "yes" or "no"; the delay time can also be manually set; the T2 and T1rho preparation pulses can be indicated by entering an echo time greater than 0 to indicate whether to execute, and can be easily inserted before image acquisition. The settings of the remaining magnetic resonance pulses remain consistent throughout the cardiac cycle and do not need to be specified separately for each cardiac cycle. Therefore, the present invention can use different combinations of T1 preparation, T2 preparation, T1rho preparation pulses and image acquisition, etc., to construct various parameter quantitative imaging sequences through graphical combinations.

[0047] In the magnetic resonance event and timing design of each cardiac cycle of the present invention, each cardiac cycle is regarded as an independent block, and the duration of each cardiac cycle is defined as T RR ; Image acquisition (ACQ) is set to be performed during diastole, and its duration is represented by T ACQ ; The idle time after acquisition is defined as T Dummy ;T Trigger represents the interval between the R wave and the start of acquisition; the fixed delay time after the R wave is denoted as T PostR , used to adapt to arrhythmic events.

[0048] The present invention uses saturation (Sat) and inversion (Inv) pulses to prepare T1 weighting before image acquisition, and the time period between the two pulses and after the inversion pulse is designated as T Sat and T Inv ; Insert a delay period T before the saturation pulse pre , to adjust T Sat and T Inv , thereby modifying the T1 weighting. pre 、T Sat 、T Inv 、T Trigger With T PostR The relationship between them is:

[0049] T pre +T Sat +T Inv =T Trigger -T PostR

[0050] In order to achieve T2 or T1rho weighting, the present invention inserts a T2 or T1rho preparation pulse before image acquisition, and its duration is TE Prep ; and in order to construct a free breathing imaging sequence, a diaphragm breathing navigation is inserted before image acquisition or before the T2 / T1rho preparation pulse, and its duration is T Nav .

[0051] The present invention utilizes a table for MRI data acquisition, which records the MRI events for each user-defined sequence and each cardiac cycle. A table lookup algorithm module then executes each sequence, each cardiac cycle, and each MRI event using two loops. Each cardiac cycle loop is triggered by the ECG, and for each heartbeat, the table lookup algorithm retrieves and executes each MRI event.

[0052] Specifically, in the graphical online sequence construction module of the present invention, some common parameter quantitative image sequences are graphically constructed as follows: the SASHA sequence first acquires the first image in the Mz equilibrium state without using a saturation pulse. After the first image, SASHA collects 9 saturation preparation images in 9 cardiac cycles; the MOLLI sequence uses 3 flip pulses to obtain 3, 3 and 5 T1-weighted images respectively, and there is an idle period of 3 heartbeats between the 2 flip pulses to ensure the complete recovery of Mz; the MOLLI5(3)3 sequence uses 2 flip pulses to collect 5 and 3 images respectively, and there is also an idle period of 3 heartbeats between the 2 flip pulses to ensure the complete recovery of Mz; the T2-prep The bSSFP sequence acquires four 2DbSSFP images in a single shot. A T2 preparation pulse is added before the acquisition of the last three images to change the T2 weighting, with echo times of 25ms, 35ms, and 45ms, respectively. The 3D free-breathing T2 quantitative imaging sequence uses a multi-acquisition GRE sequence to read and acquire three left ventricular layers. The corresponding T2 preparation pulses have echo times of 0ms, 25ms, and 45ms. A saturation pulse is also applied before each T2 preparation pulse, followed by a fixed delay time, and diaphragmatic breathing navigation is used.

[0053] In addition, in the user interface of the graphical online sequence construction module proposed in the present invention, saturation pulses and T2-prep pulses are combined to construct a simultaneous T1 / T2 quantitative imaging sequence: the first image is obtained without using T1 or T2-prep pulses, and then 12 hybrid T1 / T2-weighted images are collected. Among the 12 hybrid T1 / T2 weighted images, the echo time of T2-prep of every 4 hybrid T1 / T2 weighted images is 0 milliseconds, 25 milliseconds, 35 milliseconds, and 45 milliseconds, respectively. The saturation recovery time is T Sat The minimum value increases linearly from the maximum available value. Furthermore, the improved simultaneous T1 / T2 quantitative imaging sequence collects 8 images in 12 cardiac cycles, of which the first 4 images take one heartbeat each and the last 4 images take two heartbeats each. The first image is sampled when Mz is fully recovered, and a saturation pulse is applied before each subsequent image to adjust the T1 weighting, and a T2 preparation pulse is added before the last 3 images. TE Prep The time is 25ms, 35ms, and 45ms.

[0054] The present invention constructs a deep learning model based on a one-dimensional recurrent neural network and a one-dimensional fully connected network to reconstruct quantitative images. After the sequence scan is completed, the deep learning quantitative image reconstruction model can encode the sampling signal and its corresponding preparation time online, generating the corresponding quantitative image on the scanner.

[0055] The deep learning-based quantitative image reconstruction module in this invention uses a one-dimensional recurrent neural network (RNN) with eight hidden layers and a bidirectional design to explore the relationship between the sampled signal and the corresponding time. Each input node of the RNN contains three channels, namely the sampled signal and its corresponding T1 / T2 preparation time. After the RNN, a one-dimensional fully connected neural network uses the least squares method to predict the A, B, T1, and T2 parameters to achieve quantitative image reconstruction. The relationship between the four parameters is:

[0056]

[0057] In the present invention, A is the image signal corresponding to the initial longitudinal magnetization vector, B is the image signal corresponding to the longitudinal magnetization vector after complete recovery, td and te correspond to the delay time of T1 preparation and T2 / T1rho-prep respectively. For T1 mapping, td represents the T in the T1 preparation module of Inversion-Recovery (IR). Inv , or T in the T1 preparation module of Saturation-Recovery (SR) Sat , and te is set to 0; for T2 or T1rho mapping, te corresponds to TE PREP , and td is set to infinity; for simultaneous T1 / T2 or T1 / T1rho mapping, td and te correspond to the delay times of T1 preparation and T2 / T1rho preparation, respectively.

[0058] In the quantitative image reconstruction module based on recurrent neural networks and fully connected neural networks of the present invention, the number of neurons in each layer of the fully connected network is 400, 200, 100, 50, 30, and 3, respectively. The last two layers are repeated three times to improve accuracy. For each repetition, the output of the sixth layer of the fully connected neural network and the mean absolute error between the measured signal and the predicted signal in the previous iteration are input. Each hidden layer of the recurrent neural network and the fully connected network uses a linear rectified linear unit (ReLU) as the activation function. Thus, the present invention simplifies the processing after the data is collected through a deep learning model.

[0059] The present invention constructs a U-shaped two-dimensional convolutional neural network installed on a data post-processing server as an automated analysis module for quantitative cardiac magnetic resonance images. This network includes an automated segmentation neural network to automatically analyze T1, T2, and T1rho mapping images. This neural network can automatically segment the left ventricular myocardium, ventricular septum, and left ventricular blood, as well as identify the left and right ventricular inferior insertion points, and conduct comprehensive global and regional myocardial analysis of T1, T2, and T1rho quantitative images.

[0060] The automated analysis module for quantitative cardiac magnetic resonance images based on a two-dimensional convolutional neural network in the present invention uses a U-shaped two-dimensional convolutional neural network, which consists of a downsampling (encoder) block and four upsampling blocks (decoders). The four upsampling blocks share the encoded features extracted from different levels of the downsampling module to predict the masks of the left ventricular myocardium, the ventricular septum, the left ventricular blood region of interest, and the right ventricular inferior insertion point. The composite convolution layer includes a two-dimensional convolution layer, batch normalization, and an activated ReLU, followed by another two-dimensional convolution layer, batch normalization, and ReLU. And the Sigmoid activation function is applied to the last layer of each of the four upsampling blocks to ensure that the predicted mask value is between 0 and 1. The predicted mask of the left ventricular myocardium is integrated into the last layer of the other three upsampling blocks as a spatial attention mechanism. The present invention realizes rapid automatic analysis after data acquisition and image reconstruction.

[0061] The above are merely embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A cardiac magnetic resonance quantitative imaging framework and automated post-processing system, characterized in that: include: A graphical online sequence construction module provides a graphical interface on the MRI scanner, allowing users to configure pulse events for each cardiac cycle through graphical combinations, enabling online construction of target cardiac MRI quantitative imaging sequences such as T1, T2, and T1rho, without the need for programming; A quantitative image reconstruction module based on recurrent neural networks and fully connected neural networks, which can automatically complete the reconstruction of T1, T2, and T1rho quantitative images of different sequences without manual user intervention; An automated analysis module for cardiac magnetic resonance quantitative images based on a two-dimensional convolutional neural network. This module can automatically segment and analyze T1, T2, and T1rho quantitative images of different sequences.

2. The cardiac magnetic resonance quantitative imaging framework and automated post-processing system according to claim 1, characterized in that: The online sequence construction framework provides a graphical interface, allowing users to interactively define magnetic resonance pulse events for each cardiac cycle, create cardiac quantitative imaging magnetic resonance sequences in real time, and simplify the sequence implementation process.

3. The cardiac magnetic resonance quantitative imaging framework and automated post-processing system according to claim 1, characterized in that: In the online sequence construction framework, each cardiac cycle is regarded as an independent unit block on the user graphical interface, allowing the user to arrange magnetic resonance radiofrequency pulses for each cardiac cycle, including saturation pulses, inversion pulses, T2 preparation pulses, delay time, respiratory navigation, image acquisition, etc.

4. In the online sequence construction framework according to claim 3, it is characterized in that, The saturation pulse, flip pulse, and image acquisition for each cardiac cycle can be activated by entering "yes" or "no"; the delay time can also be set manually; the preparation pulse for T2 and T1rho can be indicated by entering an echo time greater than 0.

5. The cardiac magnetic resonance quantitative imaging framework according to claim 3, characterized in that: In the online sequence construction framework, the magnetic resonance events and timing of each cardiac cycle are designed as follows: The duration of each cardiac cycle is defined as T RR ; Image acquisition (ACQ) is set to be performed during diastole, and its duration is represented by T ACQ ; The idle time after acquisition is defined as T Dummy ;T Trigger represents the interval between the R wave and the start of acquisition; the fixed delay time after the R wave is denoted as T PostR , used to adapt to arrhythmic events; before image acquisition, saturation and inversion pulses are used to prepare T1 weighting, and the time period between the two pulses and after the inversion pulse is designated as T Sat and T Inv ; Insert a delay period T before the saturation pulse pre , to adjust T Sat and T Inv , thereby modifying T1 weighting.

6. The cardiac magnetic resonance quantitative imaging framework according to claim 3, characterized in that: Before image acquisition, T2 or T1rho preparation pulses are inserted to achieve T2 or T1rho weighting; before image acquisition or before T2 / T1rho preparation pulses, diaphragm breathing navigation is inserted to construct a free breathing imaging sequence.

7. The cardiac magnetic resonance quantitative imaging framework according to claim 3, characterized in that: The online sequence construction framework includes a table for recording the magnetic resonance events of each cardiac cycle of each sequence constructed by the user; the online sequence construction framework includes a table lookup algorithm module for sequentially executing each sequence, each cardiac cycle, and each magnetic resonance event.

8. The cardiac magnetic resonance quantitative imaging framework and automated post-processing system according to claim 1, characterized in that: In S2, after the serial scanning is completed, the deep learning quantitative image reconstruction model can encode the sampling signal and its corresponding preparation time online to generate the corresponding quantitative image on the scanner.

9. The cardiac magnetic resonance quantitative imaging framework and analysis system according to claim 8, characterized in that: Quantitative image reconstruction based on deep learning uses a one-dimensional recurrent neural network. Each input node of the recurrent neural network contains three channels, namely the sampling signal and its corresponding T1 / T2 preparation time. The recurrent neural network contains eight hidden layers. After the recurrent neural network, a one-dimensional fully connected neural network is used to predict the A, B, T1, and T2 parameters. The relationship between the four parameters is expressed as Equation (1): Where A is the image signal corresponding to the initial longitudinal magnetization vector, B is the image signal corresponding to the complete recovery of the longitudinal magnetization vector, td and te correspond to the delay time of T1 preparation and T2 / T1rho-prep, respectively. For T1 mapping, td represents the T in the T1 preparation module of Inversion-Recovery (IR). Inv , or T in the T1 preparation module of Saturation-Recovery (SR) Sat , and te is set to 0; for T2 or T1rho mapping, te corresponds to TE PREP , and td is set to infinity; for simultaneous T1 / T2 or T1 / T1rho mapping, td and te correspond to the delay times of T1 preparation and T2 / T1rho preparation, respectively.

10. The cardiac magnetic resonance quantitative imaging framework and automated post-processing system according to claim 1, characterized in that: Automated post-processing includes an automated segmentation neural network that can automatically segment the left ventricular myocardium, ventricular septum, left ventricular blood, and identify the left and right ventricular inferior insertion points, and perform comprehensive global and regional myocardial analysis on T1, T2, and T1rho quantitative images.

Images