Method and system for processing MRI data

Abstract

The present technology relates generally to a method of processing MRI data using a trained machine learning, ML, model. In particular, the present application relates to a method of quantifying tissue properties from MRI data using a trained ML model.

Description

Technical Field

[0001] This technology generally relates to a method for processing MRI data using a trained machine learning (ML) model. Specifically, this application relates to a method for quantifying tissue properties from MRI data using a trained ML model. Background Technology

[0002] Chemical shift encoded MRI (CPE) PDFF is a reliable and rapid method for quantifying proton density fat fraction (PDFF) in various organs and disease states. PDFF measurement is now established for measuring hepatic steatosis and is increasingly being used in other organs, including the pancreas, muscle, and bone marrow. Multi-echo gradient echo acquisition is most commonly used, which means that it is possible to extract from the same acquisition. Relaxation rate measurement provides additional quantitative information about iron or calcium. Therefore, gradient echo acquisition is used. It is a flexible way to explore multiple pathological processes. The speed of these acquisitions means they are also suitable for whole-body MRI. Whole-body MRI is becoming an important tool for cancer staging and is also used in other applications such as the assessment of inflammatory arthritis.

[0003] From a signal processing perspective, The core challenge is to disentangle or separate the signals generated by different components within a tissue, such as water and fat. Water-dominated and fat-dominated tissues produce extremely similar signals, leading to ambiguity. Existing techniques utilize the phase of complex signals to separate water and fat, provided that the dominant magnetic field can be estimated and distinguished. The contribution to phase. An alternative approach is to use signal amplitude for parameter estimation. Recent findings suggest that subtle differences in amplitude signal oscillations under multi-echo conditions can resolve the ambiguity between water-dominant and fat-dominant tissues. For both complex signal-based and amplitude-based methods, the computational cost of water-fat separation can be very high, hindering its use in research settings and clinical practice. Furthermore, both complex signal-based and amplitude-based methods may be subject to limitations due to… The estimation error or choice of the likelihood maximum can lead to water-lipid exchange and other artifacts.

[0004] Research has focused on using deep learning to reduce the cost of fat-water separation and minimize artifacts. The advantage of using deep learning for model fitting lies in upfront computational cost: model training can be performed before image acquisition, and the computational resources required to process these images are typically far lower than traditional fitting methods. Existing techniques use convolutional neural networks (CNNs) to separate fat and water, leveraging spatial relationships within the image and information contained in individual voxels along the echo dimension. However, the drawbacks of CNN-based approaches are that their results may not generalize beyond the specific anatomical structures used to train the network and are susceptible to variations in acquisition parameters (such as the image matrix). Furthermore, training networks using this approach is typically extremely slow, taking hours or even days, and requires large training datasets that are difficult to obtain and costly to construct.

[0005] The applicant has identified a need for an improved method for performing MRI parameter estimation using deep learning. Summary of the Invention

[0006] In a first aspect of this technology, a computer-implemented method is provided for quantifying tissue characteristics from magnetic resonance imaging (MRI) data using a trained ML model comprising at least two trained machine learning (ML) sub-models. The method includes: acquiring MRI data of a human or animal subject, wherein the MRI data comprises a plurality of voxels corresponding to regions of the subject; processing the plurality of voxels of the MRI data using each of the at least two trained ML sub-models to generate at least two quantified predictions of tissue characteristics, wherein the at least two trained ML sub-models are trained using different regions of a parameter space corresponding to the tissue characteristics; determining, for the plurality of voxels, which of the at least two quantified predictions is most accurate; and based on the determination, outputting the most accurate quantified prediction of the tissue characteristic for the plurality of voxels.

[0007] The specific human / animal tissue characteristics to be quantified may be, for example, proton density fat fraction and / or MRI relaxation rate. However, this technique can be applied to quantify any human / animal tissue characteristics that may be meaningful for assessing the clinical risk of certain diseases (e.g., heart disease or diabetes). Furthermore, the MRI data can be any type of MRI data, including but not limited to the aforementioned chemical shift encoded MRI (…). ) and whole-body MRI ( ).

[0008] Advantageously, this technique provides a rapid and flexible method for quantifying human / animal tissue characteristics across a variety of physiological structures. That is, this technique can be applied to MRI data associated with any part of the human or animal body. This is advantageous because existing techniques (such as those using convolutional neural networks (CNNs)) are trained on MRI data associated with specific parts of the human body and cannot generalize. Furthermore, existing techniques are generally slow and inefficient. For example, traditional fitting methods without employing ML models can be used to quantify tissue characteristics. However, such fitting methods are time-consuming, and training models using existing ML schemes with CNNs can take hours or even days. This technique addresses this problem by providing a rapid and flexible method for quantifying tissue characteristics through at least two ML sub-models trained to quickly provide quantitative predictions, regardless of the anatomical structure in the MRI data. By using at least two ML sub-models, the method achieves flexibility and accuracy across a variety of anatomical structures. For example, the at least two ML sub-models can be applied to organs such as the liver, bone, muscle, or adipose tissue without requiring training / retraining with data specific to such organs or tissues. To construct non-anatomy-specific ML models, this technique uses voxel-based or patch-based MRI data, rather than complete MRI image data, as described below. However, regardless of the anatomical structure, such data is not always applicable to different parts of the same anatomical structure. To address this issue, this technique employs at least two ML sub-models.

[0009] MRI data can be monovinyl or multivinyl data. In either case, MRI data cannot infer overall anatomical structure. This segmental prediction method based on a relatively small amount of MRI data allows for the construction of attribute maps on the target anatomical structure. As mentioned above, using at least two ML sub-models addresses the problem of not being able to construct a single general model that is not anatomy-specific. Using at least two sub-models enables the use of small image regions.

[0010] Therefore, in some cases, multiple voxels comprise multiple monomers. In this case, the generation step may include processing the multiple monomers individually using each of the at least two trained ML sub-models.

[0011] Preferably, in this case, the determination step may include, for each of the plurality of voxels, determining which of the at least two quantization predictions is most accurate. Therefore, the accuracy determination is performed on a per-voxel basis.

[0012] In an alternative scenario, the plurality of voxels comprises a plurality of multi-voxel groups. In this case, the generation step may include processing the plurality of multi-voxel groups individually using each of the at least two trained ML sub-models.

[0013] Preferably, in these cases, the determination step may include, for each of the plurality of multi-voxel sets, determining which of the at least two quantization predictions is most accurate. Therefore, the accuracy determination is performed on a per-voxel basis.

[0014] In all cases, the step of outputting a quantization prediction may include combining the most accurate quantization predictions from each single voxel or group of voxels to generate a combined quantization prediction for the acquired MRI data. In other words, quantization predictions from a single voxel or a group of voxels can be combined to output a holistic quantization prediction for the entire acquired MRI data. This allows clinicians to make diagnoses about areas in human or animal subjects using MRI imaging.

[0015] The determination process may include: calculating the likelihood value of each of the at least two quantized predictions; and selecting the quantized prediction with the highest likelihood value. Therefore, it is possible to accurately determine which prediction is the most accurate.

[0016] Alternatively, the determining step may include: calculating the error value of each of the at least two quantization predictions; and selecting the quantization prediction with the lowest error value.

[0017] Alternatively, the determination step may include processing the at least two quantization predictions using another trained ML model and determining the most accurate quantization prediction. Therefore, the method may include using a second trained ML model to identify the most accurate prediction, rather than using a calculated likelihood value. Using a second trained ML model can improve the overall accuracy of the quantization predictions. Compared to using the quantization prediction with the highest likelihood value or the lowest error value, using a second trained ML model can incorporate a selection strategy based on human input, which cannot be easily encoded through an analytical function.

[0018] As mentioned above, MRI data can be either monomorphic or multimorphic. In either case, MRI data cannot infer overall anatomical structures. This segmental prediction method based on a relatively small amount of MRI data allows for the construction of attribute maps on the target anatomical structure. As stated above, using at least two ML sub-models addresses the problem of not being able to construct a single, general model that is not anatomy-specific. Using at least two sub-models enables the use of small image regions.

[0019] At least two quantitative predictions of tissue properties can be generated, including: generating a quantified value for fat content using a first trained ML sub-model, and generating a quantified value for water content using a second trained ML sub-model. That is, the first trained ML sub-model can be trained to provide high-accuracy predictions for human tissues that are primarily composed of water. A large portion of the human body is composed of water-dominant tissues, therefore using a trained ML sub-model to predict the properties of such tissues can provide accurate predictions across multiple anatomical structures. Similarly, the second trained ML sub-model can provide significantly higher prediction accuracy for human tissues whose composition is predominantly fat. The human body also includes a large portion of tissues whose composition is predominantly fat.

[0020] MRI data can be acquired by receiving MRI data from the MRI scanning device. In this case, the MRI data is acquired directly from the MRI scanning device "online" to improve speed.

[0021] Alternatively, MRI data can be obtained by reading MRI data from the storage medium of the user's device. In this case, clinicians can analyze the MRI data, for example, in a hospital or laboratory, for research or other purposes.

[0022] The at least two trained ML sub-models can be any combination of the following: multilayer perceptron, decision tree, gradient boosting decision tree, random forest, convolutional neural network, recurrent neural network, generative adversarial neural network, and diffusion model. The method can use any combination of such ML sub-models as long as it can perform regression or an equivalent prediction task.

[0023] In a second aspect of this technology, a system is provided for quantifying tissue properties from magnetic resonance imaging (MRI) data using a trained ML model comprising at least two trained machine learning (ML) sub-models. The system includes: an MRI (scanning) apparatus; and at least one processor coupled to a memory, configured to: acquire MRI data of a human or animal subject from the MRI apparatus, wherein the MRI data includes a plurality of voxels corresponding to regions of the subject; process the plurality of voxels of the MRI data using each of the at least two trained ML sub-models, generating at least two quantified predictions of tissue properties; determining, for each of the plurality of voxels, which of the at least two quantified predictions is most accurate; and based on the determination, outputting the most accurate quantified prediction of the tissue property for the plurality of voxels.

[0024] The MRI scanning device and the at least one processor can be integrated together. In this case, by integrating the at least one processor that executes the trained ML model with the scanning device, the implementation of the above-described method in a clinical setting can be simplified.

[0025] Alternatively, the MRI scanning apparatus and the at least one processor can be independent of each other, i.e., housed in different hardware devices (e.g., separate computing devices). In this case, the at least one processor can operate independently of the MRI scanning apparatus. This can be advantageous, for example, when interfacing with other software already present on the computing device.

[0026] In a third aspect of this technology, a computer-implemented training method is provided for training a machine learning (ML) model to quantify tissue characteristics from magnetic resonance imaging (MRI) data. The ML model includes at least two ML sub-models. The method includes: generating at least two training datasets, wherein each training dataset includes MRI data corresponding to each ML sub-model to be trained; and training each of the at least two ML sub-models using the corresponding training datasets to generate quantitative predictions of tissue characteristics.

[0027] Advantageously, each of the at least two ML sub-models is provided with a training dataset. Therefore, instead of training each sub-model using the same training dataset, each sub-model is provided with its own training dataset to enhance the predictive power of the overall ML model.

[0028] The step of training each of the at least two ML sub-models may include: training each ML sub-model separately using the corresponding training dataset. In other words, joint training of the sub-models is not performed.

[0029] The step of generating at least two training datasets may include: generating a multidimensional parameter space having parameters defined by a mathematical model of MRI signals in voxels; dividing the parameter space into at least two regions, where the number of regions corresponds to the number of training datasets to be generated; and generating multiple simulated MRI signals for each of the at least two regions, wherein the simulated MRI signals for each region form a training dataset. That is, a statistical / mathematical model can be used to generate training datasets including simulated signals. For example, the statistical / mathematical model can be a model known in the art capable of accurately simulating MRI signals in voxels. Furthermore, this avoids the need to generate training datasets from real data, which is time-consuming and difficult to obtain.

[0030] The at least two regions can be non-overlapping. That is, each ML sub-model can be trained on a dataset independent of the dataset used by other sub-models in the training model. In this way, the domain in which each sub-model has high accuracy can be clearly distinguished.

[0031] Dividing the parameter space into at least two regions may include: calculating the likelihood values ​​of multiple points in the parameter space using a likelihood function; identifying at least one inflection point in the likelihood values ​​of the parameter space; and, when at least one inflection point is identified, partitioning the parameter space based on the position / location of each identified stationary point in the parameter space. Advantageously, potentially ambiguous regions can be identified based on the statistical / mathematical model used to generate the training dataset. Stationary points or inflection points in the likelihood function may cause inaccurate predictions from the ML submodel due to degeneracy of the underlying solutions. Partitioning the training dataset based on stationary points of the likelihood function mitigates the degeneracy problem.

[0032] In a specific example, generating at least two training datasets may include: calculating the likelihood values, which includes calculating the likelihood value of the proton density fat fraction (PDFF) for each of the plurality of points; identifying at least one stationary point, which includes identifying a stationary point where the likelihood value switches from high PDFF likelihood to low PDFF likelihood; and partitioning the parameter space, which includes dividing the parameter space into two regions based on the stationary point, wherein the first region corresponds to the low PDFF likelihood value and the second region corresponds to the high PDFF likelihood value.

[0033] The method may further include training another ML model to process the quantization predictions generated by the at least two sub-models and determine the most accurate quantization prediction.

[0034] In the related solutions of this technology, a computer program is provided, the computer program including instructions, which, when executed by a computer, cause the computer to perform any of the methods described herein.

[0035] Those skilled in the art will understand that this technology can be embodied as a system, method, or computer program product. Therefore, this technology can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0036] Furthermore, this technology can take the form of a computer program product embodied on a computer-readable medium, on which computer-readable program code is embodied. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable medium can be, for example, but not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination of the foregoing.

[0037] The computer program code used to perform the operations of this technology can be written in any combination of one or more programming languages, including object-oriented programming languages ​​and traditional procedural programming languages. Code components can be represented as procedures, methods, etc., and may include sub-components, which can take the form of instructions or instruction sequences at any level of abstraction, from direct machine instructions of the native instruction set to those constructed in high-level compiled or interpreted languages.

[0038] Embodiments of this technology also provide a non-transitory data carrier carrying code that, when implemented on a processor, causes the processor to perform any of the methods described herein.

[0039] This technology also provides processor control code for implementing the methods described above, for example, on a general-purpose computer system or a digital signal processor (DSP). This technology also provides a carrier that carries the processor control code to implement any of the methods described above at runtime, particularly on a non-transitory data carrier. The code can be provided on media such as disks, microprocessors, CDs, etc. or DVD On a carrier of ROM, programmable memory (e.g., non-volatile memory such as flash memory), or read-only memory (firmware), or provided on a data carrier such as an optical or electrical signal carrier. Code (and / or data) used to implement embodiments of the technology described herein may include source code, object code, or executable code in a conventional programming language (interpreted or compiled) (e.g., C) or assembly code, code for setting up or controlling an ASIC (Application-Specific Integrated Circuit) or FPGA (Field-Programmable Gate Array), or code for a hardware description language (e.g., Verilog (RTM) or VHDL (Very High Speed ​​Integrated Circuit Hardware Description Language)). Those skilled in the art will understand that such code and / or data may be distributed among multiple coupled components communicating with each other. The technology may include a controller comprising a microprocessor, working memory, and program memory coupled to one or more components of the system.

[0040] Those skilled in the art will also appreciate that all or part of the logical methods according to embodiments of the present technology can be suitably embodied in a logic device including logic elements for performing the steps of the methods described above, and such logic elements may include components (such as logic gates) in, for example, programmable logic arrays or application-specific integrated circuits. Such logical arrangements may also be embodied in enabling elements for temporarily or permanently establishing logical structures in such arrays or circuits using, for example, a virtual hardware description language, which can be stored and transmitted using a fixed or transportable carrier medium.

[0041] In one embodiment, this technology may be implemented using multiple processors or control circuits. This technology may be adapted to run on or be integrated into the device's operating system.

[0042] In one embodiment, the technology can be implemented as a data carrier having functional data thereon, the functional data including a functional computer data structure, which, when loaded into and operated on a computer system or network, causes the computer system to perform all the steps of the above-described method. Attached Figure Description

[0043] The embodiments of this technology will now be described by way of example only, with reference to the accompanying drawings, wherein:

[0044] Figure 1A shows the Gaussian likelihood of PDFF as a function of water-dominant elements;

[0045] Figure 1B shows the Gaussian likelihood of PDFF as a function of fat-dominant voxels;

[0046] Figure 2A illustrates the training process for a single network scheme;

[0047] Figures 2B and 2C illustrate the training process of the dual-network scheme according to this technology;

[0048] Figure 3 is a flowchart of the steps for training an ML model to quantify organizational characteristics according to this technique;

[0049] Figures 4A and 4B show the inference process for single-network and dual-network schemes, respectively;

[0050] Figure 5 is a flowchart of the steps for quantifying tissue characteristics using a trained ML model according to this technique;

[0051] Figure 6 shows a block diagram of a system 100 for predicting human tissue attributes using a trained ML model;

[0052] Figures 7A and 7B show the performance of a single network trained using parameter values ​​that are uniformly distributed across the entire parameter space.

[0053] Figures 8A and 8B illustrate the combined performance of the two networks according to this technology;

[0054] Figure 9 shows the analysis results of the multi-site phantom dataset;

[0055] Figure 10 shows a PDFF image of the volunteer's pelvis and picture. Detailed Implementation

[0056] In a broader sense, this technology generally relates to a method for processing MRI data using a trained machine learning (ML) model. Specifically, this application relates to a method for quantifying tissue characteristics from MRI data using a trained ML model. Advantageously, this technology enables the quantitative prediction of tissue characteristics in a computationally efficient manner that distinguishes signals generated by different tissue components.

[0057] This technique includes a fast, anatomically independent deep learning-based method that addresses model degeneracy. Method (also referred to herein as "RAIDER"). RAIDER allows parameter estimation to be performed separately in each voxel using deep learning. The applicant utilizes pairs of separately trained deep neural networks, each with a restricted training distribution to avoid degeneracy, and then selects the optimal parameter estimate from the two networks based on likelihood differences.

[0058] Organization and noise model Utilizing gradient echo-based The noiseless complex signal S acquired at echo time t can be modeled as:

[0059] (1)

[0060] in and The amplitudes of water and fat components are respectively. The frequency of each spectral fatty acid component, The relative amplitude of each spectral fatty acid component, This represents the total number of fatty acid components in the spectrum. for Frequency shift caused by inhomogeneity Let be an unknown relaxation constant. The conventional assumption is that the relative amplitude and frequency shifts of each fat component are known a priori; therefore, the unknown parameter is... , , and .Sure and Subsequently, proton density fat fraction was analyzed using PDFF. calculate.

[0061] In the presence of Gaussian noise (present in both real and imaginary channels), the log-likelihood of the measured signal set is given by the following formula:

[0062] (2)

[0063] in, For the measurement signal set, For the corresponding signal set based on parameter estimation, The number of measurements (twice the number of echo times for complex data, and the number of echo times for amplitude data).

[0064] For the signal amplitude, the noise-free signal in equation (1) becomes:

[0065] (3)

[0066] It has only three unknown parameters , and Equation (3) above can be used to generate multiple analog signals with a range of parameter values, where the parameters can be, for example, PDFF and / or The range of parameter values ​​can be predetermined. That is, given a range of parameter values, a simulated signal can be generated by providing parameter values ​​sampled from that range as input to equation (3). The parameter values ​​can be sampled uniformly from that range. Alternatively, the parameter values ​​can be sampled non-uniformly from that range. Multiple simulated signals can form a training dataset for training the ML model to perform tissue attribute / characteristic prediction and quantization.

[0067] Using amplitude implies that the noise now follows a Ricean distribution. The log-likelihood of the measured signal set becomes:

[0068] (4)

[0069] in This is a set of measured amplitude signals at different echo times. For the corresponding predicted amplitude signal set, It is a zeroth-order modified Bessel function of the first kind.

[0070] The problem of double optimal value and Sources of degeneracy Figures 1A and 1B are graphs illustrating the degeneracy of the likelihood maximum measured using amplitude-based fitting. With amplitude-based fitting, the likelihood function defined by equations (2) and (4) has two optimal values: a “true” solution very close to the true value, and a PDFF value at the other end of the range (i.e., a PDFF value close to 1). The error "swap" solution in PDFF) is resolved.

[0071] Figure 1A shows the Gaussian likelihood as a function of PDFF for water-dominant voxels (PDFF=20%), and Figure 1B shows the Gaussian likelihood as a function of PDFF for fat-dominant voxels (PDFF=80%). The true value estimates and maximum / optimal likelihoods are shown as dashed lines. For both water-dominant and fat-dominant voxels, there are two optimal values ​​at low and high PDFF. This creates a source of degeneracy—for a given set of signals, there are two candidate solutions with very similar likelihoods. Existing techniques (e.g., MAGO [1] and MAGORINO [2] methods) use two-point search to ensure that two optimal values ​​are explored, thus ensuring that the global optimum corresponds to the true solution. However, as described below, this degeneracy poses a problem for deep learning. It can also be seen from Figures 1A and 1B that there is a switching point near 60% PDFF, below which any initial guess may converge to a low PDFF solution, and above which the initial guess may converge to a high PDFF solution.

[0072] Parameter estimation using deep learning The use of deep learning for curve fitting is increasingly prevalent in qMRI parameter estimation. Similar to traditional fitting, the model is fitted to the acquired measurements. However, instead of finding the point of minimum error / maximum likelihood on the objective function, deep learning trains a deep neural network to directly map the monomer signal to the corresponding qMRI parameters. Therefore, the unknowns in this model are the network weights, not the parameters in the signal model. Once the network is trained, parameter estimation can be achieved by feeding unseen data into the network; then, the parameter estimates are computed all at once. This effectively front-loads computational costs, meaning that the cost of applying the network is significantly reduced compared to traditional fitting.

[0073] Degeneracy in Deep Learning Neural networks can represent one-to-one and many-to-one mappings, but typically cannot represent one-to-many (“multivalued”) mappings. When degenerate samples exist during training, the function learned by the neural network during training maps the degenerate signal to the empirical mean of the organizational properties on a degenerate subset. In other words, in this case, the network will not learn any correct solution. This is... This presents a specific problem where two candidate solutions (the "true" solution and the "exchange" solution) have very similar likelihoods. One potential technical solution to the degeneracy problem is to exclude a region of the output space. However, this approach assumes that a region of the output space can be discarded based on its improbability in experiments. This is not a satisfactory technical solution because low PDFF and high PDFF values ​​may coexist in real organizations.

[0074] The applicant addresses the degeneracy problem during training by partitioning the parameter space. As mentioned above, equation (3) can be used to generate multiple simulated signals for a range of parameter values. The range of parameter values ​​can be predetermined, and multiple simulated signals can be used to obtain the training dataset. For example, for PDFF and Parameters, the complete range of values ​​for a parameter exists in a two-dimensional parameter space, in each dimension. and The parameter space can be divided into regions corresponding to the range of parameter values, allowing for the training of multiple ML sub-models. That is, the ML model of this technique includes at least two ML sub-models. For example, the parameter space can be divided into two regions, such that the first sub-model can be trained on the first region of the parameter space, and the second sub-model can be trained on the second region of the parameter space, without degeneracy. Therefore, the first region can correspond to a first training dataset distribution, and the second region can correspond to a second training dataset distribution.

[0075] Training machine learning (ML) models to quantify organizational characteristics

[0076] This technique provides a method for training an ML model to quantify organizational characteristics, wherein the ML model comprises at least two sub-models, each trained on a different region of the parameter space. To highlight the advantages of this technique, the training of both single-network and dual-network models is explained below.

[0077] Training dataset distribution — single sub-model . Figure 2 A illustrates the process of training a single network. In a simple single-network scheme, a network is trained using a uniform distribution across the entire space of reasonable parameter values ​​(i.e., using a uniformly distributed training dataset across a single region representing the complete parameter space). Specifically, in the traditional scheme, training involves using an unsplit training dataset consisting of 20,000 analog signals across the entire range of parameter values: , PDFF / Generated in a uniformly distributed two-dimensional parameter space. PDFF and The values ​​are sampled from the full range and provided as input to equation (3), using a typical value with ten echo times. Acquisition protocol, where TE1=1.1ms, ΔTE=1.1ms.

[0078] Training dataset distribution—multiple sub-modelsFigures 2B and 2C illustrate the process of training a dual-network system according to the present technique. In this technique, two separate networks are trained on separate portions of a reasonably defined parameter space to avoid degeneracy. Figures 2B and 2C also show an example of how a training dataset is obtained to train two networks according to the present technique. In this example, the parameter space is divided into two regions: a low PDFF region (PDFF values ​​below the switching point, which may also be interchangeably referred to as the stationary point and the inflection point) and a high PDFF region (PDFF values ​​above the switching / stationary point). The low PDFF region corresponds to the first training dataset distribution, and the high PDFF region corresponds to the second training dataset distribution. In this way, two sub-models can be trained, each accurate within its respective parameter space region. For example, the first sub-model can be trained to have particularly high accuracy on water-dominant tissues, and the second sub-model can be trained to have particularly high accuracy on fat-dominant tissues.

[0079] Figures 2B and 2C show two training datasets, each consisting of 20,000 signals, constructed according to the following parameter distribution:

[0080] The training dataset for the first sub-model (“Water Network”) is shown in Figure 2B, corresponding to the low PDFF region of the full parameter space, i.e., the range below the switching point: , The parameter values ​​being sampled.

[0081] The training dataset for the second sub-model (“fat network”) is shown in Figure 2C, corresponding to the high PDFF region of the full parameter space, i.e., the range above the switching point: , The parameter values ​​being sampled.

[0082] It will be understood that the above will be a two-dimensional PDFF / The parameter space is divided into two regions only as an example; the parameter space can be divided into two or more regions. Furthermore, it will be understood that the above two-dimensional PDFF / Provided as an example only, the parameter space may include two or more dimensions, where each dimension corresponds to any relevant parameter predicting the organizational attributes of a person.

[0083] Figure 3 is a flowchart of the steps for training an ML model to quantify tissue characteristics according to the present technology. The training process is a computer-implemented method. The ML model includes at least two ML sub-models. The method includes: generating at least two training datasets for each of the at least two sub-models, wherein each training dataset includes MRI data (step S100); and training each of the at least two ML sub-models using the corresponding training datasets to generate a quantitative prediction of at least one tissue characteristic (step S102).

[0084] Step S100 of generating at least two training datasets may include: generating a multidimensional parameter space having parameters defined by a mathematical model of MRI signals in voxels; dividing the parameter space into at least two regions, wherein the number of regions corresponds to the number of training datasets to be generated; and generating multiple simulated MRI signals for each of the at least two regions, wherein the simulated MRI signals of each region form a training dataset.

[0085] That is, each of the at least two sub-models has a corresponding training dataset. Therefore, multiple analog signals can represent a labeled training dataset for supervised or semi-supervised learning. In the case of supervised or semi-supervised learning, the multiple analog signals form features of the labeled training dataset, and the corresponding parameter values ​​used to generate the analog signals form the labels of the labeled training dataset. (Unsupervised learning can also be used, but in this case, the training dataset may not include any labeled data. In this case, the training dataset may only include features obtained from the analog signals.) Each of the multiple input parameters can correspond to the human tissue attribute to be predicted and can be any of the following: proton density fat fraction, relaxation rate, signal at time 0, fat spectral model, and chemical shift.

[0086] As mentioned above, a given likelihood function may have at least two optimal values, where the likelihood function values ​​at each optimal value are similar. For example, as shown in Figures 1A and 1B, the likelihood function may have two local maxima, where the likelihood value at the first maxima is substantially similar to the likelihood value at the second maxima. In other words, the likelihood function may have at least one stationary point or inflection point. This degeneracy stems from the fact that the likelihood similarity at the optimal values ​​is not due to the stationary / inflection point itself. Stationary / inflection points provide useful notations for partitioning the parameter space to overcome the degeneracy problem.

[0087] Dividing the parameter space into at least two regions may include: calculating the likelihood values ​​of multiple points in the parameter space using a likelihood function; identifying at least one stationary point among the likelihood values ​​across the parameter space; and, when at least one stationary point is identified, partitioning the parameter space based on the localization / position of each identified stationary point. Advantageously, regions can be identified based on potential ambiguities generated by the statistical / mathematical model used to generate the training dataset. Stationary points or inflection points in the likelihood function may cause inaccurate predictions from the ML submodel due to the degeneracy of the underlying solutions. The degeneracy problem is mitigated by partitioning the training dataset based on stationary points of the likelihood function.

[0088] In a specific example, generating at least two training datasets may include: calculating the likelihood values, which includes calculating the likelihood value of the proton density fat fraction (PDFF) for each of the plurality of points; identifying at least one stationary point, which includes identifying a stationary point where the likelihood value switches from high PDFF likelihood to low PDFF likelihood; and partitioning the parameter space, which includes dividing the parameter space into two regions based on the stationary point, wherein the first region corresponds to the low PDFF likelihood value and the second region corresponds to the high PDFF likelihood value.

[0089] At step S102, the step of training each of the at least two ML sub-models may include: training each ML sub-model separately using the corresponding training dataset. In other words, joint training of the sub-models is not performed.

[0090] The method may further include (not shown) training another ML model to process the quantization predictions generated by the at least two sub-models and determining the most accurate quantization prediction.

[0091] Network architecture As described above, the machine learning (ML) model of this technology includes at least two ML sub-models. The at least two ML sub-models may consist of a first neural network and a second neural network. Each of the first and second neural networks may be a multilayer perceptron (MLP) network. Each of the first and second neural networks may include five fully connected hidden layers, wherein each fully connected hidden layer may include the same number of nodes as the number of signal measurements (i.e., echo time). The neural network may also include an output layer, wherein the output layer includes nodes for each individual organizational attribute to be predicted (i.e., one output node for PDFF, one for...). It has the Exponential Linear Unit (ELU) activation function.

[0092] Network trainingEach of the first and second neural networks can be trained using supervised learning as described above, where multiple analog signals form features of the labeled training dataset, and the corresponding parameter values ​​of the analog signals form the labels of the labeled training dataset. The networks can be implemented using the Exponential Linear Unit (ELU) activation function. It will be understood that the ELU activation function is provided only as an example, and any suitable activation function can be used.

[0093] To enable the network to predict changes in signal strength (i.e. The change is robust, and the multi-echo signal passes through the input network before being processed. The rough approximation is normalized; this step is applied both during training and inference (see [reference]). Figure 2 B. Figure 2 C and Figure 4 Note that this step is not expected to produce accurate parameter estimates, and the network is expected to learn to compensate for the inaccuracies of this step.

[0094] This rough Approximate value is denoted as It obtains the measured signals of the two echo times closest to the first and second "in-phase" echo times (assuming a single-peak model), calculates the attenuation between these echo times (assuming a single exponential attenuation), and then extrapolates back. To obtain it. To ensure this does not lead to... If the signal strength is lower than the maximum measured signal strength, then if Then Set to equal to .

[0095] For simulation experiments, this normalization step is not absolutely necessary because It is specified, and therefore known a priori. However, this step is included in all these experiments (and for real data) to ensure that the results represent the expected situation in vivo as fairly as possible. unknown.

[0096] In some cases, a self-supervised learning scheme (a form of unsupervised learning) can be used to train both the first and second neural networks. When applied to quantitative MRI parameter estimation where degeneracy exists, the dual-network scheme may be advantageous for self-supervised learning. This is because simulation experiments have shown that training a single network across the entire parameter space leads to unsatisfactory parameter estimates. The performance of the self-supervised dual-network scheme was compared to the single-network scheme during simulation experiments (simulation evaluation at SNR=60). The single network was trained without any restrictions on the distribution of the training dataset. Simulation experiments showed that the dual-network scheme improved the estimation of fat-dominated and water-dominated tissues compared to the single network. The simulation experiments also highlighted how the presence of degeneracy affects self-supervised learning differently compared to supervised learning. Unlike supervised learning, where degeneracy causes the single-network scheme to produce the mean of multiple degenerate solutions, self-supervised learning of a single network produces one of the degenerate solutions. This was observed through the nearly identical performance of the single network and the water network in the simulation experiments. The self-supervised dual-network method has also been shown to be equally important in vivo. Due to the presence of degeneracy, the single network learns either water-dominated or fat-dominated solutions. However, the combined predictions of the water network and the fat network are similar to conventionally fitted parameter maps.

[0097] After the ML model has been trained and predictions have been obtained on the test dataset, another set of “composite” or final predictions can be obtained by selecting the output with the highest likelihood from the trained ML sub-model based on equation (4).

[0098] Figure 4 A and Figure 4 B illustrates the reasoning process for single-network and dual-network schemes, respectively. Figure 4 A illustrates how a single network receives normalized measurement signals and generates organizational predictions. Figure 4 B illustrates an example of how the final prediction is derived according to this technique. As described above, the ML model may include a first neural network and a second neural network. The first neural network has high accuracy for water-dominant tissue (“water network”), and the second neural network has high accuracy for fat-dominant tissue (“fat network”). The most accurate prediction between the two networks can be selected based on the difference in likelihood between the two predictions (as defined in Equation (4)). Additionally, if one network generates physically unreasonable parameter estimates, the other network is assumed to be the correct choice. That is, the outputs of multiple sub-models can be used to provide the final prediction at inference time. Specifically, at inference time, for a two-network scheme, the outputs from both networks (fat network and water network) are considered as “candidate” solutions. That is, the likelihood of each candidate is calculated based on the measured signal, and the one with the higher likelihood is selected. References will now be made to Figure 5 Describe the reasoning method in more detail.

[0099] Figure 5This is a flowchart illustrating the steps of predicting human tissue attributes using a trained ML model according to the present technology. The inference process is a computer-implemented method. The ML model includes at least two sub-models and is trained according to the above-described technology. The method includes: acquiring MRI data using each of the at least two trained ML sub-models (step S200), wherein the MRI data includes multiple voxels corresponding to regions of the subject; processing the multiple voxels of the MRI data using each of the at least two trained ML sub-models to generate at least two quantitative predictions of tissue attributes / characteristics (step S202); determining, for the multiple voxels, which of the at least two quantitative predictions is the most accurate (step S204); and based on the determination, outputting the quantitative prediction of at least one tissue attribute / characteristic that is the most accurate for the multiple voxels (step S206).

[0100] At step S200, the MRI data can be monovinyl data. Alternatively, the MRI data can be multivinyl data. MRI data can be acquired by receiving MRI data from an MRI scanning device. That is, the MRI data can be acquired directly from the corresponding scanning device. Alternatively, the MRI data can be read from the storage medium of a user device (e.g., a personal computer or laptop). In a clinical setting, this device can be a computer in a laboratory or hospital.

[0101] At step S202, the at least two trained ML sub-models are used to generate corresponding predictions for at least one human tissue attribute. Each of the at least two trained ML sub-models processes MRI data, and each sub-model generates corresponding predictions for at least one human tissue attribute. That is, each of the at least two trained ML sub-models processes all MRI data, rather than a portion of the MRI data. The prediction generated by each of the at least two trained ML sub-models can be a prediction of proton density fat fraction, relaxation rate, or any combination thereof.

[0102] The at least two sub-models may include a first trained ML sub-model and a second trained ML sub-model. That is, the number of trained ML sub-models can be two. In this case, for human tissues whose composition is water-dominant (e.g., above the switching / stationary point), the prediction accuracy of the first ML sub-model may be significantly higher than that of the second trained ML sub-model. Additionally or alternatively, for human tissues whose composition is fat-dominant (e.g., below the switching / stationary point), the prediction accuracy of the second trained ML sub-model may be significantly higher than that of the first trained ML sub-model.

[0103] As described above, each of the at least two trained ML sub-models may be a multilayer perceptron (MLP). Alternatively, the at least two trained ML sub-models may be any combination of the following: MLP, decision tree, gradient boosting decision tree, random forest, convolutional neural network, recurrent neural network, linear regression model, or generative adversarial neural network.

[0104] At step S204, it is determined which of the generated predictions is the most accurate. As described above, the predictions from the at least two trained ML sub-models can be used to derive a composite or final prediction. This determination may include: calculating the likelihood value of each prediction generated by the at least two trained ML sub-models; and identifying the most accurate prediction as the one with the highest likelihood value. For example, the likelihood function may be given by a function such as equation (4) above. That is, the likelihood function may be derived from a statistical model that simulates MRI signals in voxels.

[0105] Alternatively, the determination may include: generating a final prediction using a second trained ML model, wherein the second trained ML model processes each prediction generated by the at least two trained ML sub-models, and the final prediction corresponds to the most accurate prediction.

[0106] Figure 6 shows a block diagram of a system 100 for predicting human tissue attributes using trained ML models. System 100 includes: an MRI scanning device 102; and a computing device 104. The computing device 104 includes: a memory 106; a trained ML model 108 comprising at least two trained ML sub-models; and a processor 110 coupled to the memory 106, configured to: acquire MRI data using each of the at least two trained ML sub-models; generate corresponding quantitative predictions of at least one tissue attribute / characteristic using each of the at least two trained ML sub-models, wherein each of the at least two trained ML sub-models processes the MRI data; determine which of the generated predictions is most accurate; and output the determined prediction for mapping human tissue attributes. The MRI scanning device 102 and the computing device 104 may be integrated together. Alternatively, the MRI scanning device 102 and the computing device 104 may be independent of each other.

[0107] experiment

[0108] The present invention will now be evaluated using experiments in an exemplary manner. Simulation, phantom, and in vivo experiments were performed. For comparison with known techniques, simulation experiments were also performed using exemplary schemes of prior art, which use a single neural network trained on a single training dataset.

[0109] Train an ML model comprising two MLP networks using the Adam optimizer with the following hyperparameter set: learning rate 0.001, mini-batch size 32, validation patience value 50, maximum number of epochs 1000, and no L2 regularization. It will be understood that this hyperparameter set and optimizer selection are provided as examples only in the following experiments, and any suitable combination of hyperparameters and / or optimizers can be used. (Using data from PDFF / ...) The parameter values ​​(labels) are randomly sampled from the parameter space (using a uniform distribution within the defined parameter range to minimize bias) and the corresponding simulated signal strength (features), and the ML model is trained in a supervised manner. Noise is added to the training data, and the SNR values ​​are randomly sampled from a uniform distribution between 20 (very low SNR) and 120 (very high SNR, designed to exceed the upper limit of the expected SNR range in vivo). The network training time for the dual-network scheme is approximately 90 minutes.

[0110] In the experiment, deep learning-based parameter prediction was performed as described above. After generating the parameter map, fat fractions were obtained from individual vials of the phantom by taking the median from the ROI placed on a single test tube, and then compared with known reference values ​​using linear regression. The network performance was then compared with traditional fitting.

[0111] To determine which of the two networks' predictions should be used as the final output, the likelihood of the predicted signals from each network is compared given the measured signal. For this purpose, equation (3) is restated as follows:

[0112] (5)

[0113] in Given a real-valued scalar, it is estimated from the RAIDER parameters using the following equation:

[0114] (6)

[0115] Once noise has been added, the maximum Gaussian likelihood can be obtained using the following matrix formulation. New estimate (Note that this is different from the previous rough approximation) different):

[0116] (7)

[0117] Where M is an N×1 vector representing the amplitude of the measured signal, and A is a vector containing the amplitude of each echo time. The values ​​are an N×1 vector, where N is the total number of echo times. It can be estimated using the following equation:

[0118] (8)

[0119] Calculated Then, for each parameter estimate of the two networks, the log-likelihood of the predicted signal based on the measured signal is calculated using equation (4), and the network prediction with the higher likelihood is output.

[0120] To facilitate the handling of occasional situations where a network produces highly likely but physically unreasonable parameter estimates (especially negative ones). Values ​​that may appear when the network is exposed to signals outside its training distribution, such as the negative value, are used to select the correct network output. The value is considered incorrect, and the output of another network is selected.

[0121] Results: Using a single network with a uniform training distribution (noise-free) . Figure 7 A and Figure 7 B shows the performance of a single network trained using parameter values ​​uniformly distributed across the entire parameter space. The training dataset consists of a set of 100,000 amplitude signals from PDFF / Parameter space ( , The parameter combinations are randomly generated in a uniform distribution. Figure 7 A shows PDFF / Prediction of PDFF values ​​in parameter space, and associated error plots ( Figure 7 B). Figure 7 Discrete positive error regions exist in the upper left, lower left, and upper right regions of B.

[0122] Results: Dual network using modified training distribution (noise-free) . Figure 8 A and Figure 8 B illustrates the "combined" performance (i.e., the performance of the outputs of the two networks based on likelihood selection) of the two networks according to this technique. Figure 7 A and Figure 7 Compared to B, performance is significantly improved, and the three main error regions (top left, top right, and bottom left) are reduced. Some small error regions still exist in the lower PDFF and the upper right region of the graph.

[0123] Results: Dual network (with noisy data) using modified training distribution. To investigate the performance of this technique in the presence of noise, the test dataset was expanded to include more data for each PDFF / Combine (a total of 101 × 21 × 100 = 212100 instances across all parameter values) to generate 100 noisy signals (SNR = 60, which is...) Typical values). The performance of the composite prediction (i.e., the selected outputs of the two networks) from the two trained networks is evaluated based on the accuracy and precision on these 100 noisy signals and compared with conventional fitting. Conventional fitting is implemented using the previously described MAGORINO algorithm and the internally implemented MAGO algorithm.

[0124] Results: Phantom Experiment To evaluate the feasibility of DL-based fitting across multiple scanners, this technique was evaluated on publicly available multi-site, multi-vendor, multi-field phantom datasets. These datasets consist of fat... The composition of the water mixture, with a known and variable fat fraction, was scanned at 1.5T and 3T at six centers (two each at GE, Philips, and Siemens). Data acquisition was performed using a multi-echo 3D scrambled gradient echo CSE sequence version for each site, with reproducibility across different acquisition parameters tested using two different protocols (both performed at 1.5T and 3T). Protocols 1 and 2 were performed at 1.5T, and protocols 2 and 4 at 3T. Protocols 1 and 3 generated approximately in-phase and out-of-phase echoes, while protocols 2 and 4 used the shortest achievable echoes.

[0125] Figure 9 shows the analysis results of the multi-site phantom dataset. The PDFF median for all 11 phantom vials across all sites, acquisition protocols, and field strengths is plotted relative to the reference fat score. An agreement plot is shown for each of the four protocols, with a separate point for each of the six sites. The solid diagonal line indicates perfect agreement with the reference PDFF value. “Site 7” refers to a repeat scan of Site 1. Except for Protocol 2, the predictions show good agreement with the reference values, exhibiting high accuracy, high linearity, and small bias across suppliers and field strengths (reflected in R-squared coefficients close to 1, slopes close to 1, and intercepts close to 0). Consistent with the simulation results, all protocols exhibit small negative biases at high PDFF. Overall performance is slightly inferior to conventional fitting.

[0126] Results: In vivo imaging To assess the feasibility of this technique in vivo, the pelvis and lower legs of two healthy volunteers were imaged. These scans were performed with the approval of the institutional review board (Queen Square Research Ethics Committee, London, REC 15 / LO / 147), and both subjects provided written informed consent. Data were acquired on a 3T Philips Ingenia system using a multi-echo 3D scrambled gradient echo sequence with retrace gradient and single-pole readout, TE1 = 1.2 ms, ΔTE = 1.6 ms, flip angle = 5°, TR = 25 ms, matrix size = 320 × 320, and pixel pitch = 1.8 × 1.8 mm. Coil combination was performed using SENSE (factor 1). Parameter estimation based on deep learning was performed again as described above.

[0127] Figure 10 A PDFF diagram (top row) showing the volunteer's pelvis and Figure (bottom row). Composite prediction across images yields satisfactory water-lipid separation and high quality. The results are highly consistent with those obtained through conventional fitting, although this technique exhibits a small negative bias in adipose tissue. The outputs from the two networks, based on likelihood selection, produce reasonable predictions across multiple tissue types, including adipose tissue (almost pure fat), muscle (almost pure water), and bone marrow (with roughly equal amounts of fat and water). Figure 9 It was also shown that RAIDER exhibits a small bias in subcutaneous fat compared to conventional fitting.

[0128] Fitting speed Compared to the 10ms of traditional fitting using MAGORINO, the deep learning-based fitting is approximately 700 times faster, taking 14μs per voxel. All fittings were performed on an Apple iMac running MATLAB 2021b (MathWorks, Natick, MA) with a 3.8GHz 8-core Intel i7 processor. Traditional fitting was performed using the fmincon minimizer with an interior-point algorithm; deep learning-based fitting was implemented using the network on the central processing unit (CPU). Both traditional and deep learning-based fittings utilized parallel computation across 8 cores. No graphics processing unit (GPU) was used.

[0129] References:

[0130] Triay Bagur A, Hutton C, Irving B, Gyngell ML, Robson MD, Brady M.Magnitude-intrinsic water-fat ambiguity can be resolved with multipeak fatmodeling and a multipoint search method. Magn Reson Med. 2018;82(1):460–475

[0131] Bray TJP, Bainbridge A, Hall-Craggs MA, Zhang H. MAGORINO: Magnitude-only fat fraction and estimation with Rician noise modeling.Magn Reson Med. 2022;89:1173-1192

[0132] Those skilled in the art will understand that, although the foregoing description has described what is considered the best mode for carrying out this technology, as well as other modes where appropriate, this technology should not be limited to the specific configurations and methods disclosed in the description of the preferred embodiments. Those skilled in the art will recognize that this technology has a wide range of applications, and various modifications can be made to the embodiments without departing from the inventive concept defined by the appended claims.

Claims

1. A computer-implemented method for quantifying tissue properties from magnetic resonance imaging (MRI) data using a trained ML model comprising at least two trained machine learning (ML) sub-models, the method comprising: Acquire MRI data of a human or animal subject, wherein the MRI data includes multiple voxels corresponding to regions of the subject; At least two quantitative predictions of tissue properties are generated by processing the plurality of voxels of the MRI data using each of the at least two trained ML sub-models, wherein the at least two trained ML sub-models are trained using different regions of the parameter space corresponding to the tissue properties. For the plurality of voxels, determine which of the at least two quantization predictions is the most accurate; as well as Based on the determination, the most accurate quantitative prediction of tissue characteristics for the plurality of voxels is output.

2. The method according to claim 1, wherein, The plurality of voxels includes a plurality of monomers, and the generation step includes processing the plurality of monomers individually using each of the at least two trained ML sub-models.

3. The method according to claim 2, wherein, The determination step includes, for each of the plurality of monomeric voxels, determining which of the at least two quantization predictions is the most accurate.

4. The method according to claim 1, wherein, The plurality of voxels includes a plurality of multi-voxel groups, and the generation step includes processing the plurality of multi-voxel groups individually using each of the at least two trained ML sub-models.

5. The method according to claim 4, wherein, The determination step includes, for each of the plurality of multi-voxels, determining which of the at least two quantization predictions is the most accurate.

6. The method according to claim 3 or 5, wherein, The step of outputting the quantization prediction includes combining the most accurate quantization prediction for each monovinyl or multivinyl group to generate a combined quantization prediction for the acquired MRI data.

7. The method according to any one of the preceding claims, wherein, The determination includes: Calculate the likelihood value for each of the at least two quantized predictions; as well as Select the quantization prediction with the highest likelihood value.

8. The method according to any one of the preceding claims, wherein, The determination includes: Calculate the error value for each of the at least two quantized predictions; as well as Select the quantization prediction with the lowest error value.

9. The method according to any one of the preceding claims, wherein, The determination includes: The at least two quantization predictions are processed using another trained ML model, and the most accurate quantization prediction is determined.

10. The method according to any one of the preceding claims, wherein, The generation of at least two quantitative predictions of tissue properties includes: generating a quantification of fat content using a first trained ML sub-model, and generating a quantification of water content using a second trained ML sub-model.

11. The method according to any one of the preceding claims, wherein, The at least two trained ML sub-models are any of the following models: multilayer perceptron, decision tree, gradient boosting decision tree, random forest, convolutional neural network, recurrent neural network, generative adversarial neural network, and diffusion model.

12. A system for quantifying tissue properties from magnetic resonance imaging (MRI) data using a trained ML model comprising at least two trained machine learning (ML) sub-models, the system comprising: MRI device; as well as At least one processor, coupled to memory, is used for: MRI data of a human or animal subject are acquired from the MRI device, wherein the MRI data includes multiple voxels corresponding to regions of the subject; At least two quantitative predictions of tissue properties are generated by processing the plurality of voxels of the MRI data using each of the at least two trained ML sub-models, wherein the at least two trained ML sub-models are trained using different regions of the parameter space corresponding to the tissue properties. For the plurality of voxels, determine which of the at least two quantization predictions is the most accurate; as well as Based on the determination, the most accurate quantitative prediction of tissue characteristics for the plurality of voxels is output.

13. A computer-implemented training method for training a machine learning (ML) model to quantify tissue properties from magnetic resonance imaging (MRI) data, said ML model comprising at least two ML sub-models, the method comprising: Generate at least two training datasets, each of which includes MRI data corresponding to each ML sub-model to be trained; as well as Each of the at least two ML sub-models is trained using the corresponding training dataset to generate quantitative predictions of organizational characteristics.

14. The method according to claim 13, wherein, Training each of the at least two ML sub-models involves training each ML sub-model separately using the corresponding training dataset.

15. The method according to claim 13 or 14, wherein, Generating at least two training datasets includes: A multidimensional parameter space is generated, the parameter space having parameters defined by a mathematical model of MRI signals in voxels; The parameter space is divided into at least two regions, where the number of regions corresponds to the number of training datasets to be generated; and For each of the at least two regions, multiple simulated MRI signals are generated, wherein the simulated MRI signals of each region form one of the training datasets.

16. The method according to claim 15, wherein, Dividing the parameter space into at least two regions includes: The likelihood function is used to calculate the likelihood values ​​for multiple points in the parameter space; Identify at least one stationary point among the likelihood values ​​across the parameter space; and The parameter space is divided based on the position of each identified stationary point in the parameter space.

17. The method according to claim 16, wherein: Calculating the likelihood includes calculating the likelihood of the proton density fat fraction PDFF for each of the plurality of points; Identifying at least one stationary point includes identifying the stationary point where the likelihood value transitions from high PDFF likelihood to low PDFF likelihood; and Dividing the parameter space includes dividing the parameter space into two regions based on the stationary point, wherein the first region corresponds to the low PDFF likelihood value and the second region corresponds to the high PDFF likelihood value.

18. The method of any one of claims 13 to 17, further comprising training another ML model to process the quantization predictions generated by each of the at least two sub-models and determining the most accurate quantization prediction.

19. A computer-readable storage medium comprising instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 11 and 13 to 18.

Images