Magnetic resonance imaging method and system

Abstract

The embodiment of the invention provides a magnetic resonance imaging method. The method may include acquiring magnetic resonance (MR) images of a subject, at least two of the MR images being acquired by an MRI scanner according to different imaging parameters. The method may further include processing the MR image by the trained machine learning model, obtaining one or more target MR mappings corresponding to at least a portion of the MR image. The second number of target MR maps is less than the first number of MR images. The trained machine learning model includes at least two sub-models, and each sub-model processes at least one MR image.

Description

Cross-referencing

[0001] This application claims priority to U.S. Application No. 18 / 829,251, filed on September 9, 2024, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This specification relates to the field of magnetic resonance imaging, and in particular to magnetic resonance imaging methods and systems based on machine learning techniques. Background Technology

[0003] In magnetic resonance imaging (MRI), T1 mapping provides information about longitudinal relaxation time of tissues, which is crucial for differentiating between healthy and diseased tissues. T2 mapping measures transverse relaxation time, providing insights into tissue hydration and edema. T1rho (T1ρ) mapping, on the other hand, focuses on the interactions between water molecules and macromolecules, providing additional biochemical information. However, existing T1, T2, and T1rho mapping techniques suffer from long scan times and motion artifacts due to these extended times. Furthermore, these mappings require separate scans, further increasing the overall scan time and patient discomfort. This is particularly problematic in cardiac scans, where breath-holding is required, further complicating the procedure.

[0004] Therefore, there is an urgent need for a method that can shorten scanning time or quickly obtain mapped images. Summary of the Invention

[0005] One or more embodiments of this disclosure provide a method for magnetic resonance imaging, the method comprising: acquiring magnetic resonance (MR) images of an object, each MR image being acquired by an MRI scanner according to target imaging parameters, at least two MR images corresponding to different target imaging parameters; processing the MR images by a trained machine learning model to acquire one or more target MR maps corresponding to at least a portion of the MR images, a second number of target MR maps being less than a first number of MR images; wherein the trained machine learning model comprises at least two sub-models, and each sub-model processes at least one MR image.

[0006] In some embodiments, processing MR images with a trained machine learning model to obtain one or more target MR maps corresponding to at least a portion of the MR images may include: obtaining relaxation times corresponding to at least one MR image; and processing MR images and relaxation times corresponding to at least one MR image with a trained machine learning model to obtain one or more target MR maps.

[0007] In some embodiments, the first quantity may be determined based on the second quantity.

[0008] In some embodiments, the method may further include determining a second quantity based on one or more target quantitative parameters of the object, wherein the one or more target MR maps include at least one of a T1 map, a T2 map, or a T1rho map.

[0009] In some embodiments, the MR image includes at least one MR image, and the type of the target imaging parameter corresponding to the at least one MRI image is the same as the type of one of one or more target quantitative parameters.

[0010] In some embodiments, one of the sub-models includes at least one of a fully connected (FC) network, a convolutional neural network (CNN), a recurrent neural network (RNN), or a Transformer.

[0011] In some embodiments, a trained machine learning model can be obtained by: acquiring multiple training samples, wherein each training sample includes a sample MR image and a reference map; performing multiple iterations on an initial machine learning model based on the multiple training samples to obtain a trained machine learning model; the initial machine learning model includes at least two sub-models; wherein at least one of the multiple iterations includes: obtaining a prediction map by inputting the sample MR image into the initial machine learning model; determining the value of a target loss function based on the prediction map and the reference map; and updating the network parameters of at least two sub-models based on the value of the target loss function.

[0012] In some embodiments, the target loss function may include at least two loss terms, each of which corresponds to a sub-model.

[0013] In some embodiments, at least one of the at least two loss terms may include a weighting factor, which is updated when the network parameters of at least two sub-models are updated.

[0014] In some embodiments, the method may further include updating the network parameters and weight factors of at least two sub-models based on the value of the target loss function.

[0015] In some embodiments, processing an MR image using a trained machine learning model to obtain one or more target MR maps corresponding to at least a portion of the MR image may include: processing a first portion of the MR image using a first sub-model to obtain a first MR map corresponding to a target quantitative parameter; processing a second portion of the MR image using a second sub-model to obtain a second MR map corresponding to the target quantitative parameter; wherein a third and a fourth quantity are less than or equal to the first quantity; and obtaining a target MR map corresponding to the target quantitative parameter by weighting the first MR map and the second MR map based on the weight parameters of the first and second sub-models.

[0016] In some embodiments, the first number may be equal to 2, and MR may include a first MR image corresponding to a first target imaging parameter and a second MR image corresponding to a second target imaging parameter that is different from the first target imaging parameter. One or more target MR maps correspond to target quantitative parameters, and the type of the target quantitative parameters is the same as the type of the first target imaging parameter or the second target imaging parameter.

[0017] One or more embodiments of this disclosure provide a magnetic resonance imaging system, comprising: an acquisition module configured to acquire magnetic resonance (MR) images of an object, each MR image being acquired by an MRI scanner according to target imaging parameters, at least two MR images corresponding to different target imaging parameters; and a processing module configured to process the MR images using a trained machine learning model to acquire one or more target MR maps corresponding to at least a portion of the MR images, a second number of target MR maps being less than a first number of MR images; wherein the trained machine learning model includes at least two sub-models, and each sub-model processes at least one MR image.

[0018] One or more embodiments of this disclosure provide a magnetic resonance imaging apparatus, the apparatus including a processor, wherein the processor is configured to perform any of the magnetic resonance imaging methods described above.

[0019] One or more embodiments of this disclosure provide a computer-readable storage medium for storing computer instructions, wherein when the computer instructions in the storage medium are read, the computer executes any of the above-described magnetic resonance imaging methods. Attached Figure Description

[0020] This specification will be further explained by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, wherein the same reference numerals denote the same structures, wherein:

[0021] Figure 1 This is a schematic diagram illustrating exemplary application scenarios of a magnetic resonance (MR) imaging system according to some embodiments of the present disclosure;

[0022] Figure 2 This is an exemplary flowchart of an MR imaging process according to some embodiments of the present disclosure;

[0023] Figure 3 This is an exemplary schematic diagram illustrating the acquisition of MR mappings according to other embodiments of this disclosure;

[0024] Figure 4 These are exemplary schematic diagrams illustrating the acquisition of a specific type of MR mapping according to some embodiments of this disclosure;

[0025] Figure 5This is an exemplary schematic diagram illustrating the acquisition of multiple types of MR mappings according to some embodiments of this disclosure;

[0026] Figure 6 This is an exemplary flowchart illustrating the process of training a trained machine learning model according to some embodiments of this disclosure; and

[0027] Figure 7 This is an exemplary block diagram of an MR imaging system according to some embodiments of the present disclosure. Specific Implementation

[0028] The technical solutions of the embodiments of this disclosure will be described more clearly below, and the accompanying drawings that need to be configured in the embodiment description will be briefly described below. Obviously, the accompanying drawings in the following description are only some examples or embodiments of this disclosure, and can be applied to other similar scenarios based on these drawings without creative effort. Unless obviously obtained from the context or otherwise indicated by the context, the same numbers in the drawings represent the same structures or operations.

[0029] It should be understood that the terms “system,” “device,” “unit,” and / or “module” used in this document are methods of distinguishing different components, elements, parts, sections, or components at different levels. However, these terms may be substituted if other terms can serve the same purpose.

[0030] As stated in this disclosure and claims, unless the context clearly indicates otherwise, the words “a,” “an,” “one,” and / or “the” do not specifically refer to the singular and may also include the plural. Generally, the terms “comprising” and “including” simply mean that expressly identified steps and elements are included and do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0031] This disclosure uses flowcharts to illustrate the operations performed by a system according to embodiments of this disclosure. It should be understood that preceding or subsequent operations are not necessarily performed in a precise order. Rather, these steps may be processed in reverse order or simultaneously. Furthermore, additional steps may be added to these flowcharts, or one or more steps may be deleted from them.

[0032] Currently, fully connected networks (FC) are commonly used in related technologies for parametric mapping of weighted images acquired from magnetic resonance scans. However, reconstructing weighted images and achieving high-precision mapping imaging using fully connected networks remains challenging, particularly in terms of robustness to noise and motion interference. Furthermore, in related technologies, at least three images are required during the inversion recovery process when acquiring mapped images (also known as MR mapping, such as T1 mapping). Insufficient image numbers may result in the inability to acquire the appropriate mapped images. However, as the number of images increases, it inevitably leads to an increase in scan time, thereby reducing imaging efficiency.

[0033] Figure 1 This is a schematic diagram illustrating exemplary application scenarios of a system for magnetic resonance imaging (MRI) according to some embodiments of the present disclosure.

[0034] like Figure 1 As shown, an MRI system (also referred to as an MRI system) 100 may include a magnetic resonance imaging (MRI) scanner 110, a processing device 120, a storage device 130, one or more terminals 140, and a network 150. In some embodiments, the MRI scanner 110, processing device 120, storage device 130, and / or one or more terminals 140 may be connected to and / or communicate with each other via wireless connections, wired connections, or a combination of both. The connections between components in the MRI system 100 may be variable. For example, the MRI scanner 110 may be connected to the processing device 120 via the network 150. As another example, the MRI scanner 110 may be directly connected to the processing device 120.

[0035] In some embodiments, the MRI system 100 may include a single-modality imaging system and / or a multimodality imaging system. The single-modality imaging system may include, for example, a system for MR imaging. The multimodality imaging system may include, for example, a system for X-ray imaging magnetic resonance imaging (X-ray-MRI), a system for single-photon emission computed tomography MRI (SPECT-MRI), a system for digital subtraction angiography MRI (DSA-MRI), a system for MRI computed tomography (MRI-CT), a system for positron emission tomography MRI (PET-MRI), and the like.

[0036] MRI scanner 110 can be configured to scan an object (or a portion of an object) to obtain image data of the object, such as MR-weighted images. In some embodiments, MRI scanner 110 may include, for example, a main magnet, gradient coils (also known as spatial coding coils), radio frequency (RF) coils, etc. The object scanned by MRI scanner 110 can be biological or non-biological. For example, the object may include a patient, an artificial object, etc. As another example, the object may include a specific site, organ, tissue, and / or body part of a patient. By way of example only, the object may include the head, brain, neck, body, shoulder, arm, chest, heart, stomach, blood vessels, soft tissue, knee, foot, etc., or combinations thereof.

[0037] Processing device 120 can be configured to process data and / or information acquired from MRI scanner 110, storage device 130, and / or one or more terminals 140. For example, processing device 120 can acquire MR images of an object, each image being acquired by the MRI scanner according to target imaging parameters. At least two MR images can correspond to different target imaging parameters. Processing device 120 can process the MR images using a trained machine learning model to acquire one or more target MR maps corresponding to at least a portion of the MR images. The second number of target maps can be less than the first number of MR images. In some embodiments, processing device 120 can be a single server or a group of servers. The server group can be centralized or distributed. In some embodiments, processing device 120 can access information and / or data from MRI scanner 110, storage device 130, and / or one or more terminals 140 via network 150. As another example, processing device 120 can be directly connected to MRI scanner 110, one or more terminals 140, and / or storage device 130 to access information and / or data. In some embodiments, processing device 120 can be implemented on a cloud platform. For example, cloud platforms can include private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, inter-cloud, multi-cloud, or combinations thereof.

[0038] Storage device 130 may store data, instructions, and / or any other information. In some embodiments, storage device 130 may store data acquired from MRI scanner 110, processing device 120, and / or one or more terminals 140. In some embodiments, storage device 130 may store data and / or instructions that processing device 120 may execute or use to perform the exemplary methods described in this disclosure. In some embodiments, storage device 130 may include a mass storage device, a removable storage device, a volatile read-write memory, a read-only memory (ROM), or a combination thereof. In some embodiments, storage device 130 may be implemented on a cloud platform, as described elsewhere in this disclosure.

[0039] In some embodiments, storage device 130 may be connected to network 150 to communicate with one or more other components of MRI system 100 (e.g., MRI scanner 110, processing device 120, and / or one or more terminals 140). One or more components of MRI system 100 may access data or instructions stored in storage device 130 via network 150. In some embodiments, storage device 130 may be part of processing device 120 or one or more terminals 140.

[0040] At least one of the one or more terminals 140 may be configured to allow interaction between the user and the MRI system 100. For example, at least one of the one or more terminals 140 may receive instructions from the user to scan a target using the MRI scanner 110. As another example, at least one of the one or more terminals 140 may receive processing results (e.g., a mapped image of the object) from the processing device 120 and display the processing results to the user. In some embodiments, at least one of the one or more terminals 140 may be connected to and / or communicate with the MRI scanner 110, the processing device 120, and / or the storage device 130. In some embodiments, at least one of the one or more terminals 140 may include a mobile device 140-1, a tablet computer 140-2, a laptop computer 140-3, or a combination thereof. In some embodiments, at least one of the one or more terminals 140 may be part of the processing device 120 or the MRI scanner 110.

[0041] Network 150 may include any suitable network that facilitates the exchange of information and / or data between the MRI system 100 and the MRI system 100. In some embodiments, one or more components of the MRI system 100 (e.g., MRI scanner 110, processing device 120, storage device 130, terminal 140, etc.) may transmit information and / or data to one or more other components of the MRI system 100 via the network. For example, processing device 120 may acquire image data (e.g., MR images) from MRI scanner 110 via network 150. As another example, processing device 120 may acquire user instructions from at least one of one or more terminals via network 150. In some embodiments, network 150 may include one or more network access points. For example, network 150 may include wired and / or wireless network access points, such as base stations and / or internet switching points, through which one or more components of the MRI system 100 may connect to network 150 to exchange data and / or information.

[0042] The above description is intended to be illustrative and does not limit the scope of this disclosure. Many alternatives, modifications, and variations will be apparent to those skilled in the art. The features, structures, methods, and characteristics of the exemplary embodiments described herein can be combined in various ways to obtain additional and / or optional exemplary embodiments. In some embodiments, the MRI system 100 may include one or more additional components, and one or more of the aforementioned components may be omitted. Alternatively or additionally, two or more components of the MRI system 100 may be integrated into a single component. For example, processing device 120 may be integrated into MRI scanner 110. As another example, components of the MRI system 100 may be replaced by other components capable of performing the functions of the components. In some embodiments, storage device 130 may be a data storage device including a cloud computing platform, such as a public cloud, private cloud, community cloud, hybrid cloud, etc. However, these variations and modifications do not depart from the scope of this disclosure.

[0043] Figure 2 This is an exemplary flowchart of a magnetic resonance imaging process according to some embodiments of the present disclosure. In some embodiments, process 200 may be performed by an MR imaging system (e.g., MRI system 100) or a processing device (e.g., processing device 120). Figure 2 As shown, process 200 includes the following steps.

[0044] Step 202 allows for the acquisition of magnetic resonance (MR) images.

[0045] According to MR imaging technology, an MRI scanner can acquire MR images by scanning a subject. MR imaging technology can be a multi-parameter imaging technique. In other words, each MR image can be acquired by an MRI scanner based on multiple imaging parameters. For example, multiple imaging parameters may include echo time, repetition time, inversion time, inversion angle, etc.

[0046] In some embodiments, MR images can be generated in a single MRI scan and can be used to simultaneously generate one or more types of target MR maps (e.g., T1, T2, and T1rho maps). As used herein, a single MRI scan refers to a complete MRI scan performed by the MRI scanner between the on and off states of the MRI scanner.

[0047] In some embodiments, MR data (e.g., MR images) generated in a single MRI scan can be used to simultaneously generate T1, T2, and T1rho maps.

[0048] In some embodiments, MR imaging techniques may include MR weighted imaging techniques, inversion recovery sequence imaging, fast imaging, etc. For example, MR weighted imaging techniques may include T1-weighted imaging techniques, T2-weighted imaging techniques, T2*-weighted imaging techniques, R*-weighted imaging methods, T1rho-weighted imaging techniques, etc. In some embodiments, different MR imaging techniques may be used to acquire every two MR images. In some embodiments, the same MR imaging technique may be used to acquire at least two MR images. For example, every two MR images may be a T1-weighted image and a T2-weighted image, or both may be T1-weighted images.

[0049] In some embodiments, each of a plurality of MR images may correspond to a target imaging parameter associated with an object. MR images can represent the relationship between target imaging parameter values ​​between different parts of an object through pixel values ​​within the MR image. In other words, an MR image corresponding to a target imaging parameter can provide qualitative information about the target imaging parameter for different parts of an object. Target imaging parameters can be used to represent the characteristics of different parts of an object. The type of target imaging parameter may include T1, T2, T2*, T1rho, R*, etc. In some embodiments, each of a plurality of MR images may correspond to a different type of target imaging parameter. In some embodiments, a plurality of MR images may correspond to the same type of target imaging parameter. In some embodiments, at least two of a plurality of MR images may correspond to different types of target imaging parameters. The target imaging parameter may be determined by the user according to actual clinical requirements, such as based on the type of object, lesions in the object, etc. The MR imaging technique and / or imaging parameters used to acquire the MR images may be determined based on the target imaging parameters corresponding to the MR images.

[0050] MR images acquired using different weighted imaging techniques can include different image types. In some embodiments, the image type of an MR image acquired using MR imaging techniques can be defined by the target imaging parameters (also known as the imaging target) that the MR image primarily presents through pixel values ​​in the image. In some embodiments, the image type of an image acquired using MR imaging techniques can be defined by the type of MR imaging technique.

[0051] In some embodiments, two MR images can be acquired based on different imaging parameters. For example, MR images can be acquired using the same MR imaging technique (e.g., T1-weighted imaging, T2-weighted imaging, T2*-weighted imaging, and R*-weighted imaging) based on different imaging parameters. As another example, MR images can be acquired using the same MR imaging technique based on different MR pulse sequences. As another example, every two MR images can be acquired using different MR imaging techniques. As another example, at least two MR images can be acquired using the same MR imaging technique. As yet another example, at least two MR images can be acquired using different MR imaging techniques (e.g., T1-weighted imaging and T2-weighted imaging).

[0052] For example, an MR image may include a T1-weighted image acquired using T1-weighted imaging techniques, where the target imaging parameter may include T1, and the pixel values ​​of the T1-weighted image may primarily represent the T1 distribution or differences in different parts of the object. As another example, an MR image may include a T2-weighted image acquired using T1-weighted imaging techniques, a T2*-weighted image acquired using T2-weighted imaging techniques, an R*-weighted image acquired using R*-weighted imaging techniques, and a T1rho-weighted image acquired using T1rho-weighted imaging techniques, or combinations thereof.

[0053] As another example, for a T1-weighted image acquired using T1-weighted imaging techniques, the image grayscale can be primarily determined by T1, and the MR image can be called a T1-weighted image. If the image grayscale is primarily determined by T2, then the MR image can be called a T2-weighted image. Therefore, "weighting" in weighted image can refer to specific target imaging parameters of the image, such as T1, T2, or T1rho. For different MR images, at least some of the values ​​of multiple imaging parameters may differ. The imaging parameters used to acquire MR images can be arranged in a time series to obtain an MR pulse sequence. An MRI scanner can acquire MR images by applying an MR pulse sequence. For example, MR data acquired through a single scan. By setting different imaging parameters, different imaging targets can be acquired through different MR pulse sequences. For example, pulse sequences used to acquire T1-weighted images can include spin echo (SE) sequences, fast spin echo (FSE), gradient echo (GRE) sequences, inversion recovery (IR) sequences, etc. Pulse sequences used for T2-weighted imaging may include spin echo (SE) sequences, fast spin echo (FSE) or fast spin echo (RSE) sequences, gradient echo (GRE) sequences, dual echo sequences, inversion recovery (IR) sequences, etc.

[0054] In some embodiments, the first quantity may be at least 2.

[0055] The specific value of the first quantity can be directly specified by the user, for example, based on the imaging target, or it can be determined by other methods. For example, the first quantity can be determined based on the second quantity or based on image features of one of the MR images. For example, the first quantity can be determined based on the acquired T1-weighted image, T2-weighted image, T1rho-weighted image, etc. Image features of MR images can include image sharpness, contrast, signal-to-noise ratio, etc. Further description of the second quantity can be found elsewhere in this disclosure.

[0056] For example, the first quantity equals the second quantity plus n, where n is a positive integer greater than 1, and the second quantity is a positive integer greater than 1. Therefore, the first quantity can be 2, 3, 4, 6, 9, 10, etc. In some embodiments, the first quantity can be equal to the second quantity * n. For example, C1 = C2 + n, where C1 is the first quantity and C2 is the second quantity.

[0057] In some embodiments, the value of n can be any specified value or determined based on quality parameters. Quality parameters can include image resolution, signal-to-noise ratio, sharpness, etc. For example, regarding resolution, when one of the MR images has a high resolution, the value of n may be larger. It can be understood that higher resolution means more information in the MR image, and using a larger value of n (including more input information) can yield a higher quality image.

[0058] By associating the first quantity with the second quantity, the required first quantity can be precisely controlled, achieving the goal of not increasing additional scanning time or reducing prediction effectiveness.

[0059] The correspondence between imaging parameters and n-values ​​can be pre-established and stored. The processing device can retrieve the n-value corresponding to the imaging parameters by looking up the relationship table. Here, imaging parameters refer to parameters corresponding to all previously acquired MR images, where different MR images may have different corresponding n-values ​​for their imaging parameters.

[0060] In some embodiments, the processing device can construct a vector database (also called a first vector database) based on imaging parameters and corresponding n values. The first vector database may include reference vectors and corresponding reference n values. Reference vectors can be obtained through vector transformation of imaging parameters. Since the amount of imaging parameter data is large, converting imaging parameters into reference vectors can effectively improve the efficiency of querying the n values ​​corresponding to the imaging parameters. During the query process, the processing device can construct a target feature vector based on the current imaging parameters (e.g., through vector transformation or feature extraction methods); based on the target feature vector, match at least one reference vector in the vector database that meets a preset condition, where the preset condition may be that the vector distance is less than a distance threshold, and the vector distance may be Euclidean distance, cosine distance, etc.; and determine the target n value based on the n values ​​corresponding to the reference vectors that meet the preset condition. The target n value may be any one of the n values ​​queried, or it may be the average of the n values ​​queried; this embodiment does not limit this.

[0061] In some embodiments, the value of n can be determined by a trained machine learning model, also known as a first trained machine learning model, such as a convolutional neural network, a recurrent neural network, or a long short-term memory network.

[0062] The input data for training the first machine learning model can be imaging parameters, and the output data can be the predicted value of n.

[0063] The first machine learning model can be trained based on a training dataset. The training dataset may include sample image data collected by an MRI scanner under different sample imaging parameters. In some embodiments, the sample imaging parameters of the sample image data can be determined from the sample image data used as training samples. Labels (true values ​​or reference n-values) can be assigned to each sample image data based on actual imaging results or experience.

[0064] For example, five different n-value evaluation network configurations (Config1, Config2, ..., Config5) can be defined, each with a different n-value. These five configurations can then be configured to process the same imaging parameter set A, and an image quality evaluation metric (e.g., sharpness scores of 80, 85, 90, 75, 82) can be calculated for each configuration. By comparing the evaluation results, if Config3 (sharpness score of 90) performs best across all metrics, then the n-value of Config3 can be used as the label for imaging parameter set A.

[0065] In some embodiments, the training of the first training machine learning model can be performed in various ways. For example, after obtaining the training dataset, the first trained machine learning model to be trained can be iterated multiple times using the training dataset. At least one iteration may include selecting one or more training samples from the training dataset, inputting the one or more samples into the first initial machine learning model to be trained, and obtaining the predicted output of the first initial machine learning model corresponding to the one or more samples; determining the value of a loss function based on the predicted output of the first initial machine learning model corresponding to the one or more training samples and the labels corresponding to the one or more training samples; and updating the model parameters in the first initial machine learning model to be trained based on the value of the loss function. For example, the model parameters may be updated based on a gradient descent algorithm. When a termination condition is met (e.g., the loss function value converges, the number of iterations reaches a preset number, etc.), the iteration is terminated, and the first initial machine learning model is obtained.

[0066] In some embodiments, the value of n can be determined based on image features of the MR image. For example, a correspondence between image features and the value of n can be established using a vector database (also known as a second vector database) or a trained machine learning model (also known as a "second trained machine learning model").

[0067] For example, the second vector database may include a second reference vector (obtained through vector transformation of imaging parameters) and its corresponding reference n value. The second trained machine learning model may be a deep learning model, such as a convolutional neural network (CNN) combined with a recurrent neural network (RNN) or a long short-term memory network (LSTM). The input to the second trained machine learning model includes imaging parameters, and the output is the n value. The second trained machine learning model can be trained based on second historical data, which includes MR images under different imaging parameters and their corresponding n values ​​(labels). These labels can be obtained through manual annotation or other methods.

[0068] For more information on vector database and model determination, please refer to the above explanation regarding the determination of imaging parameters and the value of n. In certain implementations, the imaging parameters are replaced with image features. Further details will not be repeated here.

[0069] In some embodiments, the processing device can acquire MR images by controlling an MRI scanner to scan objects or by retrieving them from a storage device or database.

[0070] Step 204: The MR image can be processed by a trained machine learning model to obtain one or more target MR maps corresponding to at least a portion of the MR image.

[0071] A target MR map can represent the values ​​of target quantitative parameters for different parts of an object. Target quantitative parameters may include T1, T2, T1rho, R*, etc. A target MR map corresponding to an MR image means that the type of target quantitative parameters is the same as the type of target imaging parameters in the MR image. In other words, a target MR map corresponding to an MR image can provide quantitative information on the target imaging parameters of different parts of the object represented in the MR image. In some embodiments, the MR map can be presented as an image, and the MR map can also be called an MR map image.

[0072] One or more MR maps may include T1 maps, T2 maps, T1rho maps, etc., or combinations thereof. Different MR maps can provide corresponding information on different quantitative parameters. For example, a T1 map can provide information on the longitudinal relaxation time of different parts of an object (e.g., tissue), which is crucial for distinguishing healthy tissue from diseased tissue; a T2 map can provide information on the transverse relaxation time of different parts of an object (e.g., tissue) and can provide insights into tissue hydration and edema; a T1rho map can provide information on the spin lattice relaxation time of different parts of an object (e.g., tissue) and focuses on the interactions between water molecules and macromolecules, providing additional biochemical environmental information.

[0073] A trained machine learning model refers to a composite network consisting of two or more sub-models, each capable of independently performing its function. Sub-models within a trained machine learning model can independently predict MR mappings based on input data, without relying on other networks.

[0074] In some embodiments, each sub-model can be a machine learning model, such as a fully connected (FC) network, a convolutional neural network (CNN), a recurrent neural network (RNN), a Transformer, or any combination thereof. That is, each sub-model can be a single network or a composite network. For example, assuming a trained machine learning model includes a first sub-model and a second sub-model, the first sub-model can be a fully connected network, and the second sub-model can be a CNN. Alternatively, the first sub-model can be a combination of an FC network and a CNN, while the second sub-model can be an RNN.

[0075] In some embodiments, each sub-model of the trained machine learning model can process at least one MR image. For example, the processing device can input a first number of MR images into the trained machine learning model, and each sub-model in the trained machine learning model can process the first number of MR images. As another example, the first number of MR images can be divided into a first part and a second part, and at least one of at least two sub-models can process the MR images in the first part, while the remaining sub-models can process these MR images in the second part.

[0076] In some embodiments, the number of sub-models in a trained machine learning model can be determined based on imaging parameters or image features of MR images. Furthermore, the number of sub-models can be determined based on a vector database (also known as a third vector database) or a machine learning model (also known as a trained third machine learning model). Specific methods for determining this number can be found in the description of the relationship between imaging parameters and the value of n above, and will not be repeated here.

[0077] For example, a third vector database may include third reference vectors (obtained through vector transformations of imaging parameters and / or image features) and the number of sub-models corresponding to these third reference vectors. The third trained machine learning model can be a deep learning model, such as a convolutional neural network (CNN) combined with a recurrent neural network (RNN) or a long short-term memory network (LSTM). The input to the third trained machine learning model includes the imaging parameters and / or image features of the MR images, and the output is the number of sub-models. It can be trained based on third historical data, which includes the actual number of sub-models (labels) under different imaging parameters and / or image features. These labels can be obtained through manual annotation or other methods.

[0078] In some embodiments, the second number of target MR maps may be less than the first number. For example, if the first number is at least 2, then the second number is at least 1, and both the first and second numbers are positive integers.

[0079] In some embodiments, the second quantity may be determined based on one or more target quantitative parameters of the object, which are expected to be obtained based on MR scans or user requirements, reflecting the user's imaging needs for magnetic resonance scans.

[0080] In some embodiments, one or more target quantitative parameters may include at least one of T1 mapping, T2 mapping, T1rho mapping, proton density, diffusion coefficient, diffusion tensor, etc.

[0081] In some embodiments, MR images may include at least one MR image whose target imaging parameters are the same as the target quantitative parameters. For example, if one or more target quantitative parameters include T1 and T2, then a first number of MR images may include at least one T1-weighted image and at least one T2-weighted image. As another example, if one or more target quantitative parameters include T1, T2, and T1rho, then a first number of MR images may include at least one T1-weighted image, at least one T2-weighted image, and at least one T1rho-weighted image. In some embodiments, assuming a second number is less than a first number, MR images may include MR images whose target imaging parameters differ from one or more target quantitative parameters. Preferably, MR images may include more T1-weighted images to provide more information. In some embodiments, to provide more information to obtain accurate images when acquiring multiple types of MR maps simultaneously, MR images may include one or more T1-weighted images.

[0082] In some embodiments, different types of MR images can be input simultaneously to obtain a target mapping corresponding to one of them. For example, suppose the input quantitative parameters corresponding to the MR images include T1 and T2. In this case, the trained machine learning model can specify which MR mapping to output in several ways. For example, the model can first output the MR mapping corresponding to the target quantitative parameters of the MR image input. If the type of the target quantitative parameter corresponding to the first input MR image is T1, the model outputs the T1 mapping. If the type of the target quantitative parameter corresponding to the first input MR image is T2, the model outputs the T2 mapping. Furthermore, the input MR images can be labeled, specific features can be input, or model parameters can be set so that the model outputs the MR mapping corresponding to the specified type of target quantitative parameter. This embodiment does not limit the specific method. In some embodiments, the processing device can input MR images into a trained machine learning model, which can process the input data and output a second number of target MR mappings.

[0083] In some embodiments, the MR images may comprise only one type of MR image (e.g., T1, T2, or T1rho). In this case, after the MR images are input into a trained machine learning model, the trained machine learning model can process the input data and output MR mappings corresponding to the type. For example, if the input is a first number of T1-weighted images, the trained machine learning model outputs a second number of T1 mappings. The first number is at least 2, and the second number is less than the first number; for example, the second number could be 1. Similarly, when the input is a first number of T2-weighted images, the trained machine learning model outputs a second number of T2 mappings.

[0084] In some embodiments, the MR image may include two or more weighted images of different types (e.g., any combination of two or more of T1, T2, and T1rho). In this case, after the MR image is input into a trained machine learning model, the trained machine learning model processes the input data and outputs multiple MR maps corresponding to the input types. For example, if the input is a first number of T1-weighted images and T2-weighted images, the trained machine learning model outputs a second number of T1 maps and T2 maps. The first number is at least 2, and the second number is less than the first number; for example, the second number could be 1.

[0085] In some embodiments, each sub-model within the trained machine learning model can process MR images, or each sub-model can process a portion of an MR image independently. Preferably, each sub-model processes an MR image.

[0086] For example, a trained machine learning model can include three sub-models (labeled 1, 2, and 3), an MR image can include two T1-weighted images and one T2-weighted image, and the final output of the trained machine learning model can include one T1 map and one T2 map. The trained machine learning model can process MR images as follows.

[0087] Two T1-weighted images and one T2-weighted image can each be input into one of the three sub-models. For example, two T1-weighted images and one T2-weighted image can be input into the input layer of a trained machine learning model, and the input layer can then input the two T1-weighted images and one T2-weighted image into the three sub-models respectively.

[0088] 1) Input two T1-weighted images and one T2-weighted image into sub-model 1. Sub-model 1 outputs T1 mapping (represented as T1_1) and T2 mapping (represented as T2_1).

[0089] 2) Input two T1-weighted images and one T2-weighted image into sub-model 2. Sub-model 2 outputs T1 mapping (represented as T1_2) and T2 mapping (represented as T2_2).

[0090] 3) Input two T1-weighted images and one T2-weighted image into sub-model 3. Sub-model 3 outputs T1 mapping (represented as T1_3) and T2 mapping (represented as T2_3).

[0091] 4) Obtain the final T1 mapping based on T1_1, T1_2, and T1_3. The final T1 mapping can be obtained by averaging or weighting T1_1, T1_2, and T1_3. The weights can be based on the weights of each sub-model in the trained machine learning model, or they can be based on pre-set weights. This embodiment does not impose any limitations. The final T1 mapping can be expressed as λ1T1_1+λ2T1_2+λ3T1_3. λ1, λ2, and λ3 represent the weights corresponding to the three sub-models, which can be the system's default settings or user-defined settings.

[0092] 5) Obtain the final T2 mapping based on T2_1, T2_2, and T2_3. The specific process is similar to the weighting in step 4).

[0093] 6) Obtain the second number of Tx mappings, namely one T1 mapping and one T2 mapping.

[0094] The above example is for illustrative purposes only. In practical applications, the input data for each sub-model can be adjusted as needed. For example, in the example above, sub-model 1 can process only two T1-weighted images and output a T1_1 mapping, sub-model 2 can process one T1-weighted image and one T2-weighted image and output a T2_2 mapping, and sub-model 3 can process two T1-weighted images and one T2-weighted image simultaneously and output a T1_3 mapping and a T2_3 mapping. Finally, based on the outputs of the three sub-models, a second number of MR mappings are obtained.

[0095] In some embodiments of this disclosure, the trained machine learning model method can simultaneously generate T1, T2, and T1rho mapping images in a single scan, solving the problem of prolonged scan time caused by individual scans. Furthermore, by integrating multiple sub-models into the trained machine learning model, T1, T2, and / or T1rho mapping images can be obtained using only two or more weighted images, reducing patient breath-holding time and overcoming the difficulty of obtaining accurate mapping images for many patients due to breath-holding time limitations, thereby improving the accuracy and robustness of MRI imaging. For example, if the number of MR images is equal to two, each sub-model in the trained machine learning model can independently process these two MR images to generate two results, and these two results can be correlated (e.g., weighted summation) to generate the final result.

[0096] On the other hand, the method disclosed in this embodiment can simultaneously provide T1, T2, and T1rho maps while collecting only three or more (e.g., five, six, etc.) shared weighted images, effectively reducing scan time and improving patient comfort. The trained machine learning model utilizes a combination of fully connected (FC) networks and three-dimensional deep convolutional neural networks (3DCNNs) to leverage the nonlinear relationships between MR images and between them and MR maps. Compared to training models individually to acquire MR map images, this method effectively utilizes the autocorrelation and cross-correlation between MR images, improving the quality and accuracy of MR map imaging results.

[0097] In other words, a trained machine learning model can output one target mapping at a time or multiple target mappings at a time, depending on the input data. For example, when the input consists of two or more MR-weighted images of the same type, the trained machine learning model can output the corresponding MR mapping based on the type of MR-weighted image input; when the input consists of two or more MR-weighted images, the trained machine learning model can output multiple types of MR mappings, and the type of the output MR mapping can correspond to the two or more MR-weighted images in the input.

[0098] Related techniques for generating MR maps can use a single network model to obtain a mapped image based on input from at least three weighted images, and cannot simultaneously obtain multiple different types of MR maps. According to some embodiments of this disclosure, a trained machine learning model can include at least two sub-models, and weights can be assigned to the sub-models, thereby reducing the number of input images for the trained machine learning model (at least two weighted images) and enabling the simultaneous acquisition of multiple different types of MR maps.

[0099] Figure 3 This is an exemplary schematic diagram illustrating the acquisition of MR maps according to other embodiments of this disclosure. In some embodiments, the operations shown in FIG300 may be performed by an MR imaging system (e.g., MRI system 100) or a processing device (e.g., processing device 120).

[0100] In some embodiments, the processing device is capable of acquiring the relaxation time corresponding to at least one MR image; and processing the MR image and the relaxation time corresponding to the at least one MR image through a trained machine learning model to obtain a second number of MR maps.

[0101] In MRI, relaxation time can be a time constant that describes the decay of the MRI signal over time, including T1 relaxation time and T2 relaxation time.

[0102] The T1 relaxation time (also known as the longitudinal relaxation time or spin lattice relaxation time) refers to the time constant during which the longitudinal magnetization vector of nuclear magnetic resonance in an external magnetic field recovers to its equilibrium state. Specifically, the T1 relaxation time can be defined as the time required for the longitudinal magnetization vector to recover to 63% of its equilibrium value.

[0103] The T2 relaxation time (also known as the transverse relaxation time or spin-spin relaxation time) can be defined as the time constant required for the transverse magnetization vector of nuclear magnetic resonance in an external magnetic field to decay to 37% of its initial value.

[0104] T1 relaxation time and T2 relaxation time are very important parameters in MRI, which can affect image contrast and signal intensity. Due to the difference between T1 and T2 relaxation times, different tissues exhibit different signal intensities in MRI images, which can be configured to differentiate and diagnose various pathological and tissue types.

[0105] Furthermore, relaxation time can also include T1rho relaxation time (also known as longitudinal relaxation time in a rotating frame), which refers to the relaxation time of the longitudinal magnetization vector in a rotating reference frame. It differs from the conventional T1 relaxation time by involving the maintenance of the magnetization vector's processing state through the continuous application of a radio frequency field (e.g., a low-frequency B1 field) within the rotating frame. T1rho imaging can be used to assess microscopic motion and interactions in tissues; for example, in studies of cartilage, brain, heart, and liver, T1rho imaging plays a crucial role in detecting early tissue degeneration and other pathological changes.

[0106] like Figure 3 As shown, the processing device can input a first number of MR images 310 and a relaxation time 320 corresponding to the Tx-weighted images (MR images) into a trained machine learning model 330. After processing the input data, the trained machine learning model 330 can output a second number of MR maps 340.

[0107] For more information on how the trained machine learning model processes the input MR image and relaxation time, please see below. Figure 4 and Figure 5 The description.

[0108] Figure 4 This is an exemplary schematic diagram illustrating the acquisition of a specific type of MR mapping according to some embodiments of this disclosure.

[0109] In some embodiments, the obtained MR mapping can be any one of T1 mapping, T2 mapping, and T1rho mapping.

[0110] The processing device can input a first number of MR images 410 into a trained machine learning model 420. In the trained machine learning model, each sub-model, such as sub-model 1, sub-model 2, ..., processes the MR images and generates its own output. For example, sub-model 1 can output prediction result 1, and sub-model 2 can output prediction result 2. The prediction result refers to the MR mapping predicted by the sub-model.

[0111] The processing device can weight the outputs of multiple sub-models to obtain one or more target MR maps 430. For example, the processing device can weight prediction result 1 and prediction result 2 to obtain a target MR map. As another example, prediction result 1 may be a first T1 map and prediction result 2 may be a second T1 map, and the processing device can determine the target T1 map by weighting the first T1 map and the second T1 map.

[0112] It should be noted that, although Figure 4 An example of each sub-model processing MR images is shown, but in practice, each sub-model can also process a different number (less than the first number) of MR images. Please refer to specific examples. Figure 5 .

[0113] Figure 5 This is an exemplary schematic diagram illustrating the acquisition of multiple types of MR mappings according to some embodiments of this disclosure.

[0114] In some embodiments, the multiple MR maps acquired simultaneously can be any combination of two or more of T1 maps, T2 maps, and T1rho maps. For example, they can be T1 and T2 maps, T1 and T1rho maps, or T1, T2, and T1rho maps, etc. The final type of MR map acquired depends on the user's image acquisition target and the type of the input MR image. Figure 2 The relevant description can be found in the instructions.

[0115] The processing device can input MR image 510 into a trained machine learning model 520. The type of input MR image needs to correspond to the MR mapping to be acquired. For example, if the target quantitative parameters include T1 mapping and T2 mapping, the MR image may include at least one T1-weighted image and one T2-weighted image; if the target quantitative parameters include T1 mapping, T2 mapping, and T1rho mapping, the MR image may include at least one T1-weighted image, one T2-weighted image, and one T1rho-weighted image.

[0116] Within the trained machine learning model 520, each sub-model can process a different number of MR images (less than or equal to a first number). For example, sub-model 1 can process a first portion of an MR image, sub-model 2 can process a second portion of an MR image, and so on. The number of first portions of an MR image and the number of second portions of the MR image (not shown in the figure) can be the same or different, and both are less than or equal to the number of first portions.

[0117] Finally, the processing device weights the outputs of multiple sub-models to obtain a second number of MR maps 530. When multiple types of MR maps are acquired simultaneously, the second number is the sum of the number of MR maps of each type. For example, if the final output is one T1 map, one T2 map, and one T1rho map, then the second number is 3.

[0118] It should be noted that, although Figure 4 An example of obtaining a certain type of MR mapping is shown. Figure 5 An example of acquiring multiple types of MR images simultaneously is shown, but it should be understood that... Figure 4 The scenario shown can also be used to acquire multiple MR maps simultaneously. Figure 5 The scenario shown can also be used to obtain a certain type of MR mapping.

[0119] Figure 6 This is an exemplary flowchart illustrating the training process of a machine learning model according to some embodiments of this disclosure. In some embodiments, process 600 may be performed by an MR imaging system (e.g., MRI system 100) or a processing device (e.g., processing device 120). Figure 6 As shown, process 600 includes the following steps.

[0120] Step 602 allows you to obtain multiple training samples.

[0121] In some embodiments, each training sample may include a sample MR image and a reference map. The sample MR image refers to the MR image used for model training, and the reference map can serve as the gold standard during model training.

[0122] Sample MR images can be obtained based on historical data, and reference mappings can be determined manually.

[0123] Step 604 involves performing multiple iterations on the initial machine learning model based on multiple training samples to obtain a trained machine learning model.

[0124] In some embodiments, the initial machine learning model may include at least two sub-models. The processing device can batch-feed multiple training samples into the initial machine learning model, obtain its output, and update its parameters based on the output. After multiple iterations, when the iteration stopping condition is met, such as reaching a certain number of iterations or the loss function converging, the machine learning model is obtained.

[0125] In some embodiments, at least one of the multiple iterations includes the following operations.

[0126] Step 6042: The predicted mapping can be obtained by inputting the sample MR image into the initial machine learning model.

[0127] Predictive mapping can be the output of an initial machine learning model after processing input sample MR images.

[0128] Step 6044 allows determining the value of the target loss function based on the prediction mapping and the reference mapping.

[0129] In some embodiments, the target loss function may include at least two loss terms, each corresponding to a sub-model. The target loss function may be constructed based on a prediction map and a reference map. For example, if the initial machine learning model includes multiple sub-models, then the target loss function = first loss term + second loss term + ... + nth loss term.

[0130] In some embodiments, the target loss function may also include other terms, such as regularization terms.

[0131] Step 6046 allows updating the network parameters of at least two sub-models based on the value of the target loss function.

[0132] Synchronous updates refer to updating the parameters of at least two sub-models in each iteration. For example, the initial machine learning model might include an input layer, a fully connected layer, a convolutional neural network layer, and an output layer, where the fully connected layer and the convolutional neural network can be considered as sub-models respectively.

[0133] The input layer can be configured to feed x (input data) into a trained machine learning model.

[0134] A fully connected layer configured to take input x passes through an FC layer to obtain the output f(x) and the output f(x) = W. fc x+b fc Among them, W fc b represents the weight of the FC layer. fc This indicates the bias in the FC layer.

[0135] A convolutional neural network layer configured as input x obtains output g(x) through CNN, where g(x) = CNN(x), and g(x) can be obtained through a series of operations, such as convolution, pooling, activation functions, etc.

[0136] The output layer f(x) configured as the output of the FC layer is added together with the weighted output of the CNN to obtain the final output y, where y = f(x) + λ·g(x), and λ is the weight parameter of the CNN, also known as the weight factor.

[0137] In some embodiments, the weight parameter λ can be a fixed value or it can be updated along with the parameters of the initial machine learning model. The fixed value can be obtained empirically or specified directly by the user.

[0138] To update the weight parameter λ, the backpropagation algorithm can be used. For example, assuming the target loss function is represented by L, the gradient of λ can be represented by the following formula (1). in, so The value of λ can be updated based on this gradient; for example, the processing device can update the value of λ using a gradient descent algorithm. As another example, updating the value of λ can include obtaining the gradient by taking the derivative of the loss function L. The value of λ is then updated according to the update rule. The update rule can be expressed as the following formula (2): in, Let λ represent the gradient. new This represents the updated value of λ. old Let λ represent the value before the update, and η represent the learning rate, which controls the step size of the update. If the gradient... If the gradient is positive, it means that the current λ is large, and the value of λ should be reduced; if the gradient is positive... If the value is negative, it means that the current λ is small and λ should be increased.

[0139] In some embodiments, based on the value of the target loss function, the network parameters and weight parameters λ of at least two sub-models can be updated synchronously.

[0140] In some embodiments, the processing device can also synchronously update the network parameters of at least two sub-models based on the value of the target loss function, and update the weight parameters λ according to a preset frequency. Updating according to a preset frequency means updating at certain iteration intervals. For example, if the preset frequency is 1, it means that the weight parameters are updated every time the network parameters of a sub-model are updated; if the preset frequency is 2, it means that the weight parameters are updated once every two updates to the network parameters of a sub-model.

[0141] In some embodiments, a preset frequency can be determined based on imaging parameters and image features of MR images. If the image resolution in the imaging parameters is high, the motion artifacts in the image features are large, and the signal-to-noise ratio is low, the preset frequency value may be small (minimum 1). Image features can refer to features of multiple images, such as the average image features of multiple historical MR images.

[0142] In some embodiments, the preset frequency can also be determined based on a vector database. For example, a vector database is constructed, which includes multiple reference vectors (imaging parameters, image features of MR images) and corresponding reference preset frequencies (constructed based on historical actual preset frequencies). During application, the processing device can construct a target feature vector based on the current imaging parameters and image features of the MR image, match at least one reference vector that meets preset conditions in the vector database based on the target feature vector, and calculate the average value based on the preset frequency corresponding to the reference vector that meets the preset conditions, and use it as the preset frequency.

[0143] In some embodiments, the processing device can also determine a preset frequency using a frequency determination model. Given the complexity of imaging parameters and image features, the frequency determination model can employ a deep learning model such as a CNN combined with an RNN or LSTM network. The input to the frequency determination model includes imaging parameters and image features of the MR image, and the output is the preset frequency.

[0144] For more information on training vector databases and frequency determination models, please refer to [link to relevant documentation]. Figure 2 Similar content to that in [the previous section]. The difference lies in configuring the data used to build the vector database, the model's inputs and outputs, while other aspects (such as model training) can reference each other.

[0145] It should be noted that the above descriptions of various processes are for illustrative purposes only and do not limit the scope of this disclosure. Those skilled in the art can make various modifications and changes to the processes under the guidance of this disclosure. However, these modifications and changes remain within the scope of this disclosure. For example, adding a storage step to various processes.

[0146] Figure 7 These are exemplary block diagrams of an MR imaging system shown according to some embodiments of the present disclosure. Figure 7 As shown, system 700 may include acquisition module 710 and processing module 720.

[0147] The acquisition module 710 can be configured to acquire MR images of an object, each MR image being acquired by an MRI scanner based on target imaging parameters, wherein at least two MR images correspond to different target imaging parameters.

[0148] The processing module 720 can be configured to process MR images using a trained machine learning network to obtain MR maps corresponding to at least a portion of the MR images, wherein the second number of MR maps is less than the first number of MR images. The trained machine learning network may include at least two sub-models, each of which processes at least one MR image.

[0149] The basic concepts described above have obviously been detailed as stated above and do not constitute a limitation of the present invention. Although not explicitly stated herein, various modifications, improvements, and alterations can be made to this disclosure by those skilled in the art. Such modifications, improvements, and corrections are recommended in this disclosure and therefore remain within the spirit and scope of the exemplary embodiments of this disclosure.

[0150] Furthermore, this disclosure uses specific terms to describe embodiments of the invention. "An embodiment," "a specific embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic of at least one implementation of the invention. Therefore, it should be emphasized and understood that multiple references to "an embodiment," "a specific embodiment," or "an alternative embodiment" throughout the various parts of this disclosure do not necessarily refer to the same embodiment. Additionally, certain features, structures, or characteristics of one or more embodiments of this disclosure may be combined.

[0151] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names in this disclosure are not intended to limit the order of the procedures and methods of this disclosure. Although the foregoing disclosure has discussed various useful embodiments currently considered to be part of this disclosure by way of various examples, it should be understood that these details are for that purpose only, and the appended claims are not limited to the disclosed embodiments, but are instead intended to cover modifications and equivalent arrangements within the spirit and scope of the disclosed embodiments. For example, while the implementation of the various components described above can be implemented in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0152] Similarly, it should be understood that in the above description of embodiments of this disclosure, various features are sometimes combined in one embodiment, figure, or description thereof in order to simplify the disclosure and facilitate understanding of one or more embodiments. However, this disclosure does not imply that the purpose of this disclosure requires more features than those mentioned in the claims. Rather, the claimed subject matter may lie in fewer than all features of a single foregoing disclosed embodiment.

[0153] In some embodiments, the numbers used to describe and claim certain embodiments of this application representing quantities, properties, etc., should be understood to be modified in some cases by the terms "approximately," "approximately," or "substantially." Unless otherwise stated, "approximately," "approximately," or "substantially" may represent a variation of ±20% of their described values. Therefore, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may vary depending on the characteristics required for each embodiment. In some embodiments, numerical parameters should take into account specified significant figures and employ common methods of number preservation. Although in some embodiments, the numerical fields and parameters used to determine their range width are approximate values, in particular embodiments these values ​​are set as precisely as feasible.

[0154] For each patent, patent application, patent application publication, and other material, such as articles, books, manuals, publications, documents, etc., cited in this disclosure, the entire contents are incorporated herein by reference. Historical application documents that are inconsistent with or conflict with this disclosure are excluded, as are documents (currently or subsequently appended to this disclosure) that limit the broadest scope of the claims of this disclosure. It should be noted that in the event of any inconsistency or conflict between the descriptions, definitions, and / or terminology used in the materials appended to this disclosure and the descriptions herein, the descriptions, definitions, and / or terminology used in this disclosure shall prevail.

[0155] Finally, it should be understood that the embodiments described in this disclosure are merely illustrative of the principles of embodiments of this disclosure. Other modifications may be employed within the scope of this disclosure. Therefore, alternative configurations of the embodiments of this disclosure can be utilized in accordance with the teachings herein, rather than as limitations. Consequently, the embodiments of this disclosure are not limited to the precise embodiments shown and described.

Claims

1. A method of magnetic resonance imaging, the method comprising: acquiring MR images of a subject; processing the MR images by a trained machine learning model to obtain one or more target MR maps corresponding to at least a portion of the MR images, a second number of target MR maps being less than a first number of MR images; wherein the trained machine learning model comprises at least two sub-models, and each sub-model processes at least one of the MR images.

2. The method of claim 1, wherein, the processing the MR images by the trained machine learning model to obtain one or more target MR maps corresponding to at least a portion of the MR images comprises: acquiring relaxation times corresponding to at least one of the MR images; and processing the MR images and the relaxation times corresponding to at least one of the MR images by the trained machine learning model to obtain the one or more target MR maps.

3. The method of claim 1, further comprising: determining the second number based on one or more target quantitative parameters of the subject, each of the one or more target MR maps corresponding to one of the one or more target quantitative parameters.

4. The method of claim 3, wherein, the MR images comprise at least one MR image, a type of target imaging parameter corresponding to at least one of the MR images is the same as a type of one of the one or more target quantitative parameters.

5. The method of claim 1, wherein, one of the sub-models comprises at least one of a fully connected network, a convolutional neural network, a recurrent neural network, or a Transformer.

6. The method of claim 1, wherein, the trained machine learning model is obtained by: obtaining a plurality of training samples, wherein each training sample of the plurality of training samples comprises a sample MR image and a reference map; performing a plurality of iterations on an initial machine learning model based on the plurality of training samples to obtain the trained machine learning model; the initial machine learning model comprises at least two sub-models.

7. The method of claim 1, wherein, the processing the MR images by the trained machine learning model to obtain one or more target MR maps corresponding to at least a portion of the MR images comprises: processing a first portion of the MR images by a first sub-model to obtain a first MR map corresponding to a target quantitative parameter; processing a second portion of the MR images by a second sub-model to obtain a second MR map corresponding to the target quantitative parameter; based on weight parameters of the first sub-model and the second sub-model, obtaining a target MR map corresponding to the target quantitative parameter by weighting the first MR map and the second MR map; or processing the MR images by a first sub-model to obtain a first MR map corresponding to a target quantitative parameter; processing the MR images by a second sub-model to obtain a second MR map corresponding to the target quantitative parameter; based on weight parameters of the first sub-model and the second sub-model, obtaining a target MR map corresponding to the target quantitative parameter by weighting the first MR map and the second MR map.

8. The method of claim 1, wherein, The first number is equal to 2, and the MR images include a first MR image corresponding to a first target imaging parameter and a second MR image corresponding to a second target imaging parameter, and the one or more target MR maps correspond to a target quantitative parameter, and the type of the target quantitative parameter is the same as the type of the first target imaging parameter or the second target imaging parameter.

9. The method of claim 1, wherein, The MR images are acquired in one scan, and the one or more target MR maps include at least one of a T1 map, a T2 map, or a T1rho map.

10. A system of magnetic resonance imaging, comprising: at least one processor and at least one memory, wherein at least one of the memories is configured to store computer instructions; and at least one of the processors is configured to execute at least a portion of the computer instructions: acquire MR images of a subject, at least two of the MR images being acquired by an MRI scanner according to different imaging parameters; process the MR images by a trained machine learning model to acquire one or more target MR maps corresponding to at least a portion of the MR images, the second number of target MR maps being less than the first number of MR images; wherein the trained machine learning model includes at least two sub-models, and each sub-model processes at least one of the MR images.

Images