A method for quantifying supraspinatus fat infiltration using magnetic resonance techniques

Abstract

This invention provides a method for quantifying supraspinatus muscle fat infiltration using magnetic resonance imaging (MRI). Belonging to the field of MRI technology, this invention involves performing IDEAL-IQ sequence scanning on the shoulder joint of a subject to acquire multi-echo complex raw three-dimensional matrix data. A three-dimensional tensor unsupervised low-rank sparse decomposition blind-source water-fat separation model is used to decouple the real tissue signal from the field distortion phase term in the signal algebraic structure. After manually delineating the region of interest (ROI) of the supraspinatus muscle at a standard research level, a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction is employed to perform sub-voxel-level partial volume effect correction on the boundary voxels. Finally, the results are combined with artificial intelligence and MRI morphological typing to calculate and output a quantitative assessment report of the supraspinatus muscle fat fraction. This invention solves the technical problem of insufficient quantitative accuracy of the supraspinatus muscle fat fraction caused by phase contamination and systematic interference from partial volume effects in the water-fat separation results under shoulder joint edge scanning conditions with highly non-uniform main magnetic fields.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance technology, and more specifically, relates to a method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance technology. Background Technology

[0002] Supraspinatus fat infiltration is an important quantitative indicator of the pathological progression of rotator cuff injuries, and magnetic resonance imaging (MRI) is currently the main clinical method for assessing supraspinatus fat infiltration. In existing technologies, proton density fat fraction measurement based on IDEAL-IQ sequences has been widely used in uniform magnetic field environments such as the abdomen and liver. Its water-fat separation typically relies on iterative least squares methods or image-domain phase-constrained algorithms, and a single relaxation model is used to fit the transverse relaxation time. Regarding partial volume effect correction, traditional methods often employ simple linear interpolation or mixed signals that ignore boundary voxels; the delineation of the region of interest directly determines the accuracy of the fat fraction calculation. In existing technologies, due to the complex anatomy of the shoulder joint and its location at the edge of the magnet aperture, the main magnetic field... The degree of non-uniformity is significantly higher in the central part than in the central part. The phase constraint algorithm will produce a water-fat phase reversal error when the local field distortion exceeds the model assumption range. The single relaxation model cannot independently decouple the lateral relaxation time of muscle and fat. Furthermore, the anatomical boundaries of the supraspinatus muscle become blurred during its degeneration and complete tear. The mixed signals from the boundary voxels, under the influence of microscopic magnetic susceptibility anisotropy, produce systematic phase deviations. These factors collectively lead to a significant systematic error in the fat fraction calculation results. In other words, existing technologies suffer from insufficient accuracy in quantitatively determining the supraspinatus muscle fat fraction due to phase contamination and partial volume effects during shoulder joint edge scanning under conditions of highly non-uniform main magnetic field. Summary of the Invention

[0003] In view of this, the present invention provides a method for quantifying the fat infiltration of the supraspinatus muscle using magnetic resonance technology, which can solve the technical problem in the prior art where, under the condition of shoulder joint edge scanning with highly uneven main magnetic field, the water-fat separation results are systematically interfered with by phase contamination and part of the volume effect, resulting in insufficient quantitative accuracy of the supraspinatus muscle fat fraction.

[0004] This invention is achieved as follows: This invention provides a method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging (MRI), comprising the following steps:

[0005] The shoulder joint of the subject was scanned using the IDEAL-IQ sequence to acquire multi-echo complex raw three-dimensional matrix data. The standard research plane with the scapula in the oblique sagittal position was selected for the scanning.

[0006] Based on the original three-dimensional matrix data of multi-echo complex numbers, a three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model is constructed. The main magnetic field distortion phase term and artifacts in the original signal are separated by unsupervised blind source separation, and the pure water phase three-dimensional matrix and lipid phase three-dimensional matrix are extracted.

[0007] A spatial phase decoupling cyclic skipping generative adversarial network model, driven by inputting the three-dimensional matrices of the aqueous and lipid phases into physical information, is used to globally correct water-lipid inversion artifacts, dynamically compensate for phase and amplitude errors, and adjust the transverse relaxation times of the water and lipid components. Independent two-component decoupling fitting was performed, and the water-lipid separation refinement map and the two-component lateral relaxation time were output simultaneously. Mapping diagram;

[0008] On the standard research-level water image, manually delineate the region of interest along the anatomical boundary of the supraspinatus muscle, and simultaneously copy the region of interest to the fat ratio image.

[0009] Based on the fat fraction of each voxel within the region of interest, a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction is introduced to perform sub-voxel-level partial volume effect correction on the boundary voxels of the region of interest and calculate the fat fraction of the supraspinatus muscle.

[0010] By combining the fat fraction of the supraspinatus muscle with the results of magnetic resonance morphology typing, the subjects were grouped into normal group, degenerative group, partial tear group and complete tear group, and a quantitative assessment report was generated.

[0011] The optimization objective function of the three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model is specifically composed of the low-rank term constrained by the tensor nuclear norm and... The sparse terms of the norm penalty are weighted and superimposed to satisfy the constraints. ,in For high-dimensional tensors, For low-rank component tensors, It is a sparse component tensor.

[0012] The value of the regularization tradeoff coefficient in the optimization objective function is obtained by conducting multiple rounds of cross-validation experiments on a training and validation dataset covering four types of pathological subtypes, with the water-lipid separation purity index as the evaluation criterion for selecting the optimal value.

[0013] Specifically, the solution of the three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model is to use the alternating direction multiplier method to iteratively optimize in the multidimensional tensor space. The iteration termination condition is that the relative change of the Frobenius norm of the low-rank component tensor between two adjacent iterations is less than the convergence threshold.

[0014] The generator of the physical information-driven spatial phase decoupling cyclic jump generative adversarial network model is specifically composed of dual-channel three-dimensional residual dense blocks. The two channels correspond to the real part channel and the imaginary part channel of the multi-echo complex data, respectively. The two channels are independently encoded and then feature fusion is performed in the intermediate layer.

[0015] The generator introduces multi-scale cyclic skip connections in the middle layer of the network. Specifically, it modulates the shallow three-dimensional spatial detail feature map along the time echo dimension using gated cyclic units and then fuses it with the deep physical feature map element by element.

[0016] The generator embeds a physical iteration operation unit between every two three-dimensional residual dense blocks. After receiving the preliminary water-fat separation parameter estimate, the physical iteration operation unit inputs it into the Bloch equation simulator to calculate the physical residual between the theoretical signal and the actual acquired multi-echo signal. The physical residual is superimposed with the backpropagation gradient of the main network using weighted coefficients.

[0017] Specifically, the discriminator of the spatial phase decoupling cyclic jump generative adversarial network model driven by physical information adopts a Markov texture discriminator architecture, with a receptive field covering local pixel blocks, to specifically evaluate the local authenticity of spatial pathological textures in the decoupled aqueous and lipid three-dimensional images.

[0018] The training loss function of the physical information-driven spatial phase decoupling cyclic jump generative adversarial network model is specifically composed of three weighted parts: pixel-level reconstruction loss, physical constraint loss generated by the physical iteration operation unit, and Markov texture discriminator adversarial loss. The weight coefficients are determined by conducting grid search experiments on the validation set.

[0019] The weight coefficient of the physical constraint loss is adjusted by a dynamic weight adjustment function, which is calculated based on three data items: the mean physical residual, the mean adversarial loss, and the mean pixel-level reconstruction loss on the current round of the validation set. The weight coefficient range is adaptively switched according to the calculation results.

[0020] Specifically, the independent two-component decoupled fitting is implemented by extending the multi-echo complex signal into a complex-domain dual-relaxation spectrum fitting dynamic equation, and using the Levenberg-Marquardt iterative algorithm to measure the proton density of the water component, the proton density of the lipid component, and the transverse relaxation time of the water component. Transverse relaxation time of fatty acid components The six parameters, namely the initial phase of the water component and the initial phase of the fat component, are estimated nonlinearly using least squares estimation.

[0021] Specifically, the multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction decomposes the mixed signal within the boundary voxel of the region of interest into three sub-voxel-level components: muscle matrix component, endothelial fat component, and interstitial water component, and iteratively solves the problem using non-negative matrix factorization technique with spatial prior distribution constraints.

[0022] The nonnegative matrix decomposition uses the spatial gradient of fat fraction between adjacent voxels as a spatial prior, and performs iterative optimization of the volume ratio of each sub-voxel component under nonnegative constraints. The iterative convergence threshold is determined by experimental analysis of the trade-off between partial volume error and the number of iterations on phantom experimental data.

[0023] The method for calculating the fat fraction of the supraspinatus muscle is as follows: ,in The intensity of the lipid phase signal. The signal intensity is for the aqueous phase, and the fat fraction ranges from 0 to 1. When reported as a percentage, multiply by 100%.

[0024] The reference range for the regularization tradeoff coefficient is 0.01 to 10.00; the reference range for the convergence threshold is... ~ The reference range for the weighting coefficient of the physical constraint loss is 0.1 to 1.0; the transverse relaxation time of the water components... The initial estimate is in the range of 20 to 40. transverse relaxation time of fatty components The initial estimate is in the range of 5 to 15. The reference range for the convergence threshold of the nonnegative matrix factorization iteration is: ~ .

[0025] This invention employs a three-dimensional tensor unsupervised low-rank sparse decomposition blind-source water-fat separation model. It completely decouples the real tissue signal from the field distortion phase term within the algebraic structure of the multi-echo complex original signal, providing high-purity prior input for subsequent water-fat inversion artifact correction. This fundamentally suppresses the systematic interference of phase contamination caused by inhomogeneous main magnetic fields on fat fraction. Based on this, a physically-driven spatial phase decoupling cyclic jump generative adversarial network embeds the Bloch equation as a physical iterative computation unit within the network, achieving global correction of water-fat inversion artifacts and two-component lateral relaxation time. The independent decoupled fitting eliminates the systematic bias of the single-relaxation model. Furthermore, a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction performs sub-voxel-level partial volume effect correction on the voxels at the region of interest boundary, eliminating the systematic bias of microscopic interface magnetic susceptibility differences in fat fraction calculation. In summary, this invention solves the technical problem mentioned in the background art, where the water-fat separation results are systematically interfered with by phase contamination and partial volume effect under the condition of shoulder joint edge scanning with highly non-uniform main magnetic field, leading to insufficient quantitative accuracy of supraspinatus muscle fat fraction. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 The image shows the supraspinatus muscle hydrograph in the standard "Y"-shaped oblique sagittal plane of the scapula in this embodiment.

[0028] Figure 3 This is a standard research plane supraspinatus muscle fat ratio image after processing by a physical information-driven spatial phase decoupling cyclic jump generative adversarial network in the embodiment.

[0029] Figure 4 The box plot shows the distribution of fat fraction in the supraspinatus muscle for each pathological group (normal group, degeneration group, partial tear group, and complete tear group) in the examples.

[0030] Figure 5 The transverse relaxation time of water components in each pathological group in the examples is shown. Lateral relaxation time of fatty components Estimated value distribution chart.

[0031] Figure 6 This is a distribution map of the difference in fat fraction before and after correction for the volume effect of the voxel portion of the region of interest boundary for each pathological group in the examples.

[0032] Figure 7 The dynamic weight adjustment function in the example Output value distribution and corresponding physical constraint loss weighting coefficients Value distribution chart. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0034] like Figure 1 The diagram shown is a flowchart of a method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging (MRI) provided by this invention. This method includes the following steps:

[0035] S01. Perform IDEAL-IQ sequence scanning on the shoulder joint of the subject to acquire multi-echo complex raw three-dimensional matrix data. The data includes real and imaginary images. The standard research plane with the scapula oblique sagittal plane in a "Y" shape is selected for the scanning plane.

[0036] S02. Based on the acquired multi-echo complex original three-dimensional matrix data, a three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model is constructed. Unsupervised blind source separation is performed on the main magnetic field distortion phase term and artifacts in the original signal to extract the pure water phase three-dimensional matrix and the lipid phase three-dimensional matrix.

[0037] S03. A spatial phase decoupling cyclic skipping generative adversarial network model is driven by inputting the three-dimensional matrix of the water phase and the three-dimensional matrix of the fat phase into the physical information. The model performs global correction on the water-fat phase inversion artifacts caused by the inhomogeneity of the main magnetic field, dynamically compensates for the phase and amplitude errors caused by small movements during the multi-echo acquisition process, and adjusts the lateral relaxation time of muscle and fat. Independent two-component decoupling fitting was performed, and the water-lipid separation refinement map and the two-component lateral relaxation time were output simultaneously. Mapping diagram;

[0038] S04. On the standard research-level water image, manually delineate the region of interest along the anatomical boundary of the supraspinatus muscle, and simultaneously copy the region of interest to the fat ratio image.

[0039] S05. Based on the fat fraction of each voxel in the region of interest, a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction is introduced to perform sub-voxel-level partial volume effect correction on the boundary voxels of the region of interest, and finally calculate the fat fraction of the supraspinatus muscle.

[0040] S06. Combine the supraspinatus fat fraction with the magnetic resonance morphological classification results to classify the subjects into normal group, degenerative group, partial tear group and complete tear group, and output a quantitative assessment report.

[0041] Among them, the IDEAL-IQ sequence is a magnetic resonance multi-echo acquisition sequence based on the principle of iterative least squares water-lipid separation. By acquiring complex signals at multiple echo time points, it simultaneously calculates proton density fat fraction and transverse relaxation time. Mapping and main magnetic field Distortion mapping enables non-invasive and quantitative determination of tissue fat content.

[0042] Among them, multi-echo complex original three-dimensional matrix data refers to a complete three-dimensional spatial encoded dataset containing real and imaginary signal matrices, which is collected at multiple echo time points. The real and imaginary parts together carry the chemical shift phase evolution information between water molecules and fat molecules in the tissue.

[0043] The construction and solution process of the three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model (3D-TLRSD) is as follows: the multi-echo complex original three-dimensional matrix data is regarded as a high-dimensional tensor. Based on the high spatial correlation between the spatially continuous muscle tissue and fat distribution, which manifests as low-rank characteristics and a main magnetic field, The phase distortion, noise, and motion artifacts caused by inhomogeneity exhibit spatial sparsity, a priori physical property, transforming the water-fat separation problem into a convex optimization problem using tensor robust principal component analysis. The optimization objective function consists of low-rank terms constrained by the tensor kernel norm and... The norm penalty is constructed by a weighted superposition of sparse terms, i.e. Satisfying constraints ,in For low-rank component tensors, For sparse component tensors, This is the regularization tradeoff coefficient. The value range was obtained through multiple rounds of cross-validation experiments on a training and validation dataset covering four pathological subtypes: normal group, degenerative group, partial tear group, and complete tear group. Specifically, in Candidate values ​​are set at logarithmic intervals within the interval, and the water-lipid separation purity index (a comprehensive score of signal-to-noise ratio and fat fraction error) is used as the evaluation criterion. The value that results in the optimal comprehensive score is selected. The solution process employs the alternating direction multiplier method, iteratively optimizing within a multidimensional tensor space. Each iteration utilizes the tensor singular value soft thresholding operator. The update step extracts the low-rank component representing the true tissue structure, namely the pure water lipid evolution signal, and applies it to the sparse soft threshold operator. The update steps remove field distortion phase and artifacts; the iteration termination condition is that the relative change in the Frobenius norm of the low-rank component tensor between two adjacent iterations is less than the convergence threshold. , The value of was obtained through analysis experiments on the iterative convergence curve on a typical case dataset. The threshold corresponding to the slope of the curve entering the plateau region was selected, and the reference range is as follows. ~ This model does not rely on any labeled training data and achieves blind source separation of high-purity aqueous phase three-dimensional matrix and lipid phase three-dimensional matrix from the underlying physical signal.

[0044] The 3D-TLRSD model, by combining the inherent low-rank geometric topology of high-dimensional tensors with spatially sparse priors, completely decouples the real tissue water-fat evolution signal from the intrinsic algebraic structure of the original multi-echo complex signal with the field inhomogeneity phase distortion term and motion artifact term, without requiring manually labeled data. The low-rank constraint is mathematically equivalent to a physical prior modeling of the spatially continuous distribution of muscle and fat, allowing the real biological tissue signal to naturally cluster in a few principal components within the singular value decomposition domain of the tensor; the sparsity constraint incorporates the discontinuous jump characteristics of local magnetic field distortion and random noise into the optimization framework. This mechanism fundamentally solves the phase contamination problem caused by the inhomogeneity of the principal magnetic field, laying the foundation for subsequent water-fat phase inversion artifact correction and two-component lateral relaxation time. Decoupling provides a priori input with high signal-to-noise ratio and high purity, effectively improving the physical accuracy and spatial consistency of quantifying supraspinatus fat fraction under extremely non-uniform magnetic field conditions.

[0045] Tensor nuclear norm is a generalization of matrix nuclear norm in higher-order tensor spaces. It is defined as the weighted sum of the singular values ​​of a tensor on all modal expansion matrices and is used to measure and constrain the low-rank nature of a tensor.

[0046] Among them, the tensor singular value soft thresholding operator is a surrogate operator that applies a soft thresholding shrinkage operation to each singular value after performing singular value decomposition on the tensor and then reconstructs the tensor. It is used as a proximal operator to approximate the tensor nuclear norm in the optimization iteration.

[0047] Among them, the alternating direction multiplier method is a distributed optimization algorithm that decomposes the overall convex optimization problem into multiple subproblems and solves them alternately. By introducing Lagrange multipliers to update the low-rank components, sparse components and dual variables in each iteration, it ensures convergence to the global optimum under constraints.

[0048] The specific structure of the Physical Information-Driven Spatial Phase Decoupling Cyclic Skip Generative Adversarial Network (PI-PDRGAN) model is as follows: The generator consists of a dual-channel 3D residual dense block, where the dual channels correspond to the real and imaginary channels of the multi-echo complex data, respectively. The two channels are independently encoded and then fused in an intermediate layer. The 3D residual dense block is composed of densely connected 3D convolutional layers superimposed with residual skip connections, and the convolutional kernel size is [missing information]. The number of feature map channels is ~ The number of convolutional kernels increases progressively between layers. The number of kernels was determined experimentally by balancing training convergence speed and memory usage, with candidate ranges of 32, 64, and 128. The final kernel was selected to achieve the fastest convergence of the validation set loss while minimizing memory usage per GPU. The network configuration includes: introducing multi-scale cyclic skip connections in the middle layer; these connections modulate the shallow 3D spatial detail feature map along the time echo dimension using gated cyclic units and then fuse it with the deep physical feature map element-wise, enabling adaptive recombination of deep abstract physical features and shallow anatomical spatial details; embedding physical iteration units between every two 3D residual dense blocks within the network; these units receive preliminary water-fat separation parameter estimates from the current layer and input them into an integrated Bloch equation simulator to dynamically simulate the theoretical signals at each echo time point, calculating the physical residuals between the theoretical signals and the actual acquired multi-echo signals; and then processing these physical residuals. The convolutional mapping is a physical loss gradient map with the same dimension as the main network feature map, weighted by coefficients. Superimposed with the gradient of the backpropagation of the main network, The value range was determined through ablation experiments on four types of pathological classification datasets, with a reference interval of [value missing]. The discriminator employs a Markov texture discriminator architecture, consisting of multiple stride convolutional layers, with a receptive field coverage of [missing information]. Pixel blocks are specifically used to evaluate the local realism of spatial pathological textures in decoupled aqueous and lipid 3D images; the generator has approximately [number of parameters]. ~ The specific values ​​were determined experimentally by minimizing inference latency while ensuring separation accuracy. The CUDA stream allocation strategy is to distribute the dual-channel encoding computation, physical iteration computation unit simulation computation, and discriminator forward propagation to three independent CUDA streams for parallel execution, with the real and imaginary parts of the dual-channel encoding each allocated to GPU memory. ~ GB, physical iteration operation unit allocates video memory ~ GB, Discriminator allocates video memory ~ GB, total video memory usage controlled at 24 GB per card Within; feature map transfer between different network layers adopts a memory buffer pool mechanism based on dynamic allocation of layer output size to avoid frequent memory allocation and deallocation; batch data prefetching in the data loop adopts asynchronous CUDA data stream, and the prefetch buffer size is [missing information]. ~ In batches; the weight gradients between neurons are allocated bandwidth resources using a circular full reduction communication strategy during multi-GPU training.

[0049] The steps for establishing the training dataset for the physically-informed spatial phase decoupling cyclic skipping generative adversarial network model specifically include: collecting IDEAL-IQ multi-echo complex raw three-dimensional matrix data from subjects in the normal group, degeneration group, partial tear group, and complete tear group. The total sample size is determined through statistical power analysis, and the reference interval is for each group. ~ Example: Performing main magnetic field analysis on the raw data. Distortion physics map annotation and lateral relaxation time Mapping and annotation were performed independently by two senior radiologists, with consistent results. Randomized affine transformations, multi-echo phase noise injection, and magnetic field distortion simulations were used to augment the dataset to its original size. ~ The training set, validation set, and test set are divided into three parts according to a ratio of 7:2:1.

[0050] The specific steps for training the physically information-driven spatial phase decoupling cyclic skip generative adversarial network model include: using the Adam optimizer, with an initial learning rate set to... ~ The specific value is determined through a learning rate warm-up experiment at the point where the slope of the loss curve is maximum; the reference range for the total number of training rounds is [missing information]. ~ Wheel, to verify the continuity of physical residual loss of the set. Early stopping is defined as the wheel not decreasing in descent; the loss function is a weighted sum of three parts: pixel-level reconstruction loss, physical constraint loss generated by the physical iteration unit, and adversarial loss from the Markov texture discriminator, with each weight coefficient determined through grid search experiments on the validation set; a dynamic batch normalization strategy is used during training, with the batch size set to... ~ To adapt to the memory limitations of 3D volumetric data; after training, the final performance was verified on the test set using the mean fat fraction error and the main magnetic field correction residual as evaluation indicators.

[0051] A physically-driven spatial phase-decoupled cyclic-skip generative adversarial network (GAN) embeds the Bloch equations of magnetic resonance physics into the network's internal iterative process as physical iterative computational units, ensuring that the feature transformations of each layer are constrained by the evolution of real physical signals. Multi-scale cyclic-skip connections ensure the synergistic preservation of deep physical decoupling capabilities and superficial anatomical spatial resolution, effectively addressing the coexistence of anatomical boundary blurring and signal heterogeneity during the process of supraspinatus muscle degeneration leading to complete tearing. A Markov texture discriminator provides adversarial constraints from the perspective of the authenticity of local pathological textures, further suppressing the generation of pseudo-structures in the water-lipid separation results. This mechanism enables the network to possess both physical interpretability and data-driven generalization capabilities under limited sample conditions, achieving robust elimination of water-lipid inversion artifacts and two-component lateral relaxation time at the shoulder joint margins where the main magnetic field is highly inhomogeneous. Decoupling ensures the physical accuracy of the quantitative results of the supraspinatus muscle fat fraction.

[0052] Among them, a dynamic weight adjustment function is designed. This is used to adaptively adjust the physical constraint loss weight coefficients in physically-driven spatial phase decoupling cyclic skipping generative adversarial networks during training. , Based on the mean of the physical residuals on the current round of validation set (Unit: Signal strength in any unit), Mean of adversarial loss (Dimensionless normalized value) and mean pixel-level reconstruction loss The calculation of the three data items (dimensionless normalized value) is as follows: ,in and The dimensionless normalization coefficient, The unit is the reciprocal of any unit of signal strength. These are dimensionless coefficients; their specific values ​​were determined through multiple rounds of ablation experiments on the validation set. The output value and the actual optimal Minimizing the error between them is the objective determination, and the reference ranges are respectively , ;when At that time, the physical constraint loss weighting coefficient Take the smaller value. Prioritizing pixel-level reconstruction accuracy; when hour, This achieves a balanced adjustment between physical constraints and reconstruction accuracy; when hour, Strengthen physical constraints to suppress water-lipid inversion artifacts in scenes with severe field inhomogeneity.

[0053] The specific implementation of the multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction is as follows: Utilizing the multi-echo complex phase evolution characteristics of the aqueous phase 3D matrix and the lipid phase 3D matrix, the mixed signal within a single voxel is decomposed into three sub-voxel-level components at the boundary voxel of the region of interest: muscle matrix component, endothelial fat component, and interstitial water component. A non-negative matrix factorization technique with spatial prior distribution constraints is employed, using the spatial gradient of fat fraction between adjacent voxels as a spatial prior, to iteratively optimize the volume ratio of each sub-voxel component under non-negative constraints. The iterative convergence threshold of the non-negative matrix factorization is determined experimentally by analyzing the trade-off between partial volume error and the number of iterations on phantom experimental data, with a reference range of [reference range missing]. ~ Finally, the fat fraction of the boundary voxels was corrected based on the volume ratio of each sub-voxel component to eliminate the systematic bias of the fat fraction calculation results caused by the difference in magnetic susceptibility of the microscopic muscle-fat interface and the partial volume effect.

[0054] The microscopic magnetic susceptibility anisotropy refers to the fact that at the microscopic interface between muscle fibers and fat cells, due to the different magnetic susceptibility of the two tissues and the orientation dependence of the interface, the local magnetic field produces an orientation-related distortion effect at the subvoxel scale, thereby affecting the phase evolution characteristics of complex signals.

[0055] The endothelial fat component refers to the collective term for the interstitial fat located between myofibril cells (intramuscular fat) and myofibrils (intermuscular fat). In cases of severe fat infiltration in the supraspinatus muscle, the volume ratio of the endothelial fat component within a single voxel is significantly increased, which is the main source of distortion in the partial volume effect.

[0056] Here, fat fraction refers to the ratio of the fat proton density signal to the total fat proton density signal within the region of interest, i.e., the proton density fat fraction, which is calculated as follows: ,in The intensity of the lipid phase signal. The range of values ​​for the fat fraction is: (The value represents the signal intensity in the aqueous phase). When reported as a percentage, multiply by 100%.

[0057] Among them, the two-component transverse relaxation time The specific implementation of decoupled fitting is as follows: the traditional single relaxation model is extended to a complex domain double relaxation spectrum fitting dynamic equation, that is, the multi-echo complex signal in the first... echo time The model at the location is ,in and These are the proton densities of the water and fat components, respectively. and These represent the independent transverse relaxation times of the water component and the fatty component, respectively. and These represent the initial phases of the water and fat components, respectively. The Levenberg-Marquardt iterative algorithm is used to perform nonlinear least-squares estimation of the six parameters in the above equations. Simultaneously, a spatial smoothing constraint term is introduced into the objective function. The values ​​of the constraint coefficients are determined experimentally on a phantom dataset with the goal of minimizing the relaxation time estimation error. The reference range is [reference range missing]. The initial estimated reference range is: ~ , The initial estimated reference range is: ~ The initial values ​​were determined using experimental data on the transverse relaxation time of normal shoulder joint muscles and adipose tissue reported in the literature, and were verified and corrected by prefitting data from normal subjects during the experimental preprocessing stage of this protocol.

[0058] Among them, the Bloch equation is a set of physical equations describing the evolution of the magnetization vector under the action of the main magnetic field and radio frequency pulse in the magnetic resonance phenomenon. It is the basic mathematical expression of the magnetic resonance signal generation mechanism. The physical iterative operation unit embeds it into the network in the form of differentiable numerical integration, so that the water-lipid separation parameters output by the network are explicitly constrained by the physical evolution law of the signal in each forward propagation.

[0059] Among them, multi-scale cyclic skip connection refers to the connection structure established between different depth levels of the network, in which the shallow feature map is temporally modulated in the echo time dimension by gated cyclic units and then fused with the deep feature map, so that the network can simultaneously use spatial anatomical details and deep physical abstract features at different scales to perform water-lipid separation inference.

[0060] Among them, the Markov texture discriminator is a discriminator architecture based on local receptive fields. By only distinguishing between real and fake local blocks in the input image rather than making global judgments on the entire image, it effectively captures and evaluates the authenticity of the spatial distribution of local pathological textures in aqueous and lipid images, thereby imposing adversarial constraints on the generator at the level of local details.

[0061] Optionally, the present invention also provides a computer-based method for forming a system for quantifying supraspinatus muscle fat infiltration using magnetic resonance imaging (MRI), wherein the computer is provided with a readable storage medium storing program instructions, which, when executed in the computer, are used to perform the above-described method.

[0062] The specific implementation of step S01 is as follows: During shoulder joint MRI scans of the subject, the IDEAL-IQ sequence is used to acquire complete three-dimensional spatial encoded data, including both real and imaginary signal matrices, at multiple echo time points. The real and imaginary parts together carry information about the phase evolution of chemical shifts between water and fat molecules. The standard research plane is selected based on the "Y" shape of the scapula in the oblique sagittal plane, which maximizes the visualization of the supraspinatus muscle's anatomical morphology and facilitates accurate delineation of the region of interest. The acquired data provides the original physical signal basis for all subsequent processing steps, and its signal quality directly determines the upper limit accuracy of water-lipid separation and fat fraction quantification.

[0063] The specific implementation of step S02 is as follows: the multi-echo complex original three-dimensional matrix data is regarded as a high-dimensional tensor. The high spatial correlation between the spatially continuous muscle tissue and fat distribution, exhibiting low-rank characteristics and a dominant magnetic field, is utilized. The phase distortion, noise, and motion artifacts caused by inhomogeneity exhibit spatial sparsity as a priori physical property, transforming the water-fat separation problem into a convex optimization problem using tensor robust principal component analysis. The optimization objective function consists of low-rank terms constrained by the tensor kernel norm and... The norm penalty is constructed by a weighted superposition of sparse terms, i.e. Satisfying constraints ,in For low-rank component tensors, For sparse component tensors, regularization tradeoff coefficients The optimal value was selected through multiple rounds of logarithmic-interval cross-validation on four types of pathological classification training datasets, with a reference range of 0.01–10.00. The solution process employed the alternating direction multiplier method. In each iteration, the low-rank components were updated using a tensor singular value soft thresholding operator, and the field distortion phase term was removed using a sparse soft thresholding operator. The iteration terminated when the relative change in the Frobenius norm of the low-rank component tensor between two adjacent iterations was less than the convergence threshold, with a reference range of [missing information]. ~ This model does not rely on any labeled data and separates high-purity aqueous phase three-dimensional matrices and lipid phase three-dimensional matrices from blind sources in the original signal algebraic structure.

[0064] The specific implementation of step S03 is as follows: The three-dimensional water phase matrix and the three-dimensional lipid phase matrix output from step S02 are used as inputs and fed into a physically-driven spatial phase decoupling cyclic skip generative adversarial network model. The generator consists of dual-channel three-dimensional residual dense blocks. The real and imaginary channels are independently encoded and then fused in the intermediate layer. The three-dimensional residual dense blocks are composed of densely connected three-dimensional convolutional layers superimposed with residual skip connections. The convolutional kernel size is [missing information]. The network introduces multi-scale cyclic skip connections in the middle layer. The shallow 3D spatial detail feature map is modulated along the time echo dimension by a gated cyclic unit and then fused element-wise with the deep physical feature map. This allows for the synergistic preservation of the deep physical decoupling capability and the shallow anatomical spatial resolution. A physical iteration computation unit is embedded between every two 3D residual dense blocks. This unit inputs the preliminary water-fat separation parameter estimate from the current layer into the internal Bloch equation simulator to dynamically simulate the theoretical signal at each echo time point. It calculates the physical residual between the theoretical signal and the actual multi-echo signal. The physical residual is then processed... After convolution mapping, the weighted coefficients are used Superimposed with the gradient of the backpropagation of the main network, The reference range is 0.1 to 1.0, and is adjusted by a dynamic weighting function. The network adaptively adjusts the mean of physical residuals, the mean of adversarial loss, and the mean of reconstruction loss based on the current validation set. The discriminator employs a Markov texture discriminator architecture, with a receptive field covering local pixel blocks to evaluate the local realism of spatial pathological textures in 3D images of aqueous and lipid phases. Simultaneously, the lateral relaxation time of the aqueous and lipid components is adjusted within the network. Independent two-group decoupled fitting was performed, and the dynamic equations were fitted using complex-domain double relaxation spectra. The Levenberg-Marquardt iterative algorithm was then used to estimate the six parameters using nonlinear least-squares estimation. The initial estimate is in the range of 20 to 40. , The initial estimate is in the range of 5 to 15. The objective function incorporates a spatial smoothing constraint term with a reference coefficient ranging from 0.01 to 0.10. The Adam optimizer is used during training, and the loss function consists of a weighted average of pixel-level reconstruction loss, physical constraint loss, and adversarial loss. The network synchronously outputs a refined water-lipid separation map and the lateral relaxation time of the two components. Mapping diagram.

[0065] The specific implementation of step S04 is as follows: On the standard research-level aqueous image output in step S03, an experienced radiologist manually delineates the region of interest (ROI) along the anatomical boundaries of the supraspinatus muscle, using the muscle's anatomical boundaries as the reference, avoiding including adjacent adipose tissue within the ROI. After delineation, the ROI is simultaneously copied to the fat comparison image to ensure strict consistency in the spatial position of the ROI between the aqueous and fat comparison images, eliminating registration errors caused by manual copying. It is recommended that the ROI data for the supraspinatus muscle of all subjects be measured multiple times and averaged to reduce random errors from manual delineation.

[0066] The specific implementation of step S05 is as follows: At the boundary voxel of the region of interest, utilizing the multi-echo complex phase evolution characteristics of the aqueous phase 3D matrix and the lipid phase 3D matrix, the mixed signal within a single voxel is decomposed into three sub-voxel-level components: muscle matrix component, endothelial fat component, and interstitial water component. A non-negative matrix factorization technique with spatial prior distribution constraints is employed, using the spatial gradient of the fat fraction between adjacent voxels as the spatial prior. The volume ratio of each sub-voxel component is iteratively optimized under non-negative constraints, with the iterative convergence threshold reference range being [reference range missing]. ~ This threshold was determined by weighing the volumetric error against the number of iterations in the phantom experimental data. Finally, the fat fraction of the boundary voxels was corrected based on the volume ratio of each sub-voxel component, and the fat fraction of the supraspinatus muscle was adjusted accordingly. The calculation eliminates the systematic bias in the calculation of fat fraction caused by microscopic magnetic susceptibility anisotropy and partial volume effect.

[0067] The specific implementation of step S06 is as follows: The fat fraction of the supraspinatus muscle calculated in step S05 is analyzed together with the magnetic resonance morphological typing results. The morphological typing is based on the continuity of the supraspinatus tendon, the uniformity of signal intensity, and the degree of muscle atrophy, classifying the subjects into normal group, degenerative group, partial tear group, and complete tear group. The quantitative results of fat fraction and the morphological typing results together constitute the core content of the quantitative assessment report. The output report covers the fat fraction distribution characteristics of each group, providing objective quantitative basis for clinical assessment of the degree of fat infiltration in the supraspinatus muscle.

[0068] It should be noted that the key technologies of this invention include: three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation, the core of which is to utilize the low-rank geometric topology of muscle and fat signals and the spatial sparsity of field distortion to form a physical dual, so as to decouple the real tissue signal and phase contamination term from the signal algebraic structure without labeled data, and cut off the interference propagation path of the main magnetic field inhomogeneity to the subsequent quantitative steps from the root; physical information-driven spatial phase decoupling cyclic jump generative adversarial network, the core of which is to embed the Bloch equation in a differentiable form into the network iteration process, so that the feature transformation of each layer is constrained by the evolution law of the real physical signal, and at the same time, through multi-scale cyclic jump connection, the deep physical decoupling ability and shallow anatomical resolution are retained in synergy, and physical interpretability and generalization ability are combined under limited sample conditions; multi-component super-resolution demixing based on microscopic magnetic susceptibility anisotropy correction, the core of which is to decompose the boundary voxel mixed signal to the sub-voxel scale, and eliminate the systematic bias of the magnetic susceptibility difference of the microscopic interface through non-negative matrix decomposition with spatial prior constraints. The three technologies are interconnected. The high-purity output of the previous step provides the necessary conditions for the physical constraints of the next step. Together, they form a physical self-consistent closed loop from signal acquisition to boundary correction, which enables the quantitative analysis of the supraspinatus muscle fat fraction to maintain physical accuracy and spatial consistency at the edge of the shoulder joint where the main magnetic field is highly uneven.

[0069] It should be noted that under conditions of complete tearing or severe degeneration of the supraspinatus muscle, the tendon ends retract significantly, muscle atrophy is pronounced, and the degree of fat infiltration is extremely high. At this point, the fat fraction in some voxels within the region of interest is close to 1, the aqueous phase signal is extremely weak, and the signal-to-noise ratio drops sharply. Under these conditions, the transverse relaxation time of the aqueous and fatty components... The modulation effect of differences on complex signals is difficult to distinguish from noise under low signal-to-noise ratio (SNR) conditions. Traditional single-relaxation models directly fit the mixed signal to a single relaxation time, systematically overestimating the fat fraction and failing to differentiate the contribution ratios of intramuscular and interstitial fat, leading to a significant decrease in the discriminative power of quantitative results across different pathological subtypes. The reasons for these technical problems are: phase distortion caused by inhomogeneity of the main magnetic field, phase amplitude errors caused by minute motions, and low SNR are coupled within the single-relaxation model framework, making it impossible to independently estimate the physical parameters of water and fat components. Traditional methods, under this strongly coupled multi-parameter condition, are prone to getting trapped in local optima during nonlinear optimization, and the estimation results are highly sensitive to initial values. Furthermore, the mixed signals of boundary voxels exhibit direction-dependent phase deviations due to microscopic magnetic susceptibility anisotropy, further exacerbating the uncertainty in parameter estimation. Common solutions to these problems include increasing the number of echoes to improve the fitting degrees of freedom or introducing regularization constraints to stabilize the nonlinear optimization. However, increasing the number of echoes prolongs the scan time and introduces more motion artifacts. Simple regularization constraints cannot distinguish between the real spatial variations of the physical signal and the pseudo-spatial variations caused by noise. Even in voxels with heavy fat infiltration, the independent decoupling accuracy of water-fat parameters cannot be guaranteed, and the systematic bias caused by the anisotropy of the microscopic magnetic susceptibility of boundary voxels cannot be eliminated. This invention effectively solves this technical problem. The physical iteration computation unit feeds back the physical residual between the theoretical signal of the Bloch equation and the actual acquired signal to the network weight update in real time, reducing the transverse relaxation time of the two components. The independent decoupled fitting is subject to explicit physical constraints in each forward propagation, anchoring parameter estimates within the physically feasible region even under low signal-to-noise ratio conditions, thus avoiding the sensitive dependence of traditional nonlinear optimization on initial values. Multi-scale cyclic skip connections use gated cyclic units to temporally modulate shallow spatial detail features in the echo time dimension, enabling the network to leverage the phase evolution correlation of adjacent voxels in the time dimension to improve the robustness of parameter estimation for low signal-to-noise ratio voxels. Furthermore, a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction unmixes boundary voxels at the sub-voxel scale. Utilizing nonnegative matrix factorization techniques with spatial prior distribution constraints, the systematic bias of microscopic interface magnetic susceptibility differences is eliminated from the fat fraction calculation results, effectively ensuring both physical accuracy and spatial consistency in the quantitative results of supraspinatus muscle fat fraction under severe fat infiltration conditions.

[0070] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the aforementioned core technical problems lies in the fact that its technical solution forms a physically self-consistent closed-loop processing link from four levels: signal acquisition, water-lipid separation, physical constraint correction, and boundary correction. Firstly, the essence of phase distortion caused by the inhomogeneity of the main magnetic field at the shoulder joint edge is that the original multi-echo complex signal is superimposed with a spatially varying distortion phase term. This term algebraically manifests as an increase in the rank of the signal matrix and spatial sparse perturbation, forming a duality with the low-rank characteristics of the real tissue signal. The three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model utilizes this physical duality to transform the high-dimensional tensor decomposition problem into a convex optimization problem of tensor robust principal component analysis. It extracts the low-rank real tissue water-lipid evolution signal through tensor nuclear norm constraints, and then... Norm penalty strips away sparse field distortion phase terms, achieving blind source separation without any labeled data, thus cutting off the path of phase contamination propagating to subsequent steps from the signal source. Secondly, the high-purity water-lipid matrix obtained after blind source separation still retains phase amplitude errors caused by minute movements during multi-echo acquisition, and water-lipid phase inversion artifacts caused by extreme local inhomogeneities in the main magnetic field. These two types of errors are difficult to eliminate simultaneously within a single physical model framework. The physically driven spatial phase decoupling cyclic skip generative adversarial network embeds physical iteration units between every two 3D residual dense blocks within the network, feeding back the physical residuals of the theoretical signal from the Bloch equation and the actual acquired signal in real-time as gradients to the network weight update process. This ensures that the feature transformation of each layer is constrained by the physical evolution of the real signal, maintaining physical interpretability under limited training sample conditions. Multi-scale cyclic skip connections use gated cyclic units to temporally modulate shallow spatial detail features in the echo time dimension, ensuring the synergistic preservation of deep physical decoupling capability and shallow anatomical resolution, effectively addressing scenarios where anatomical boundary blurring and signal heterogeneity coexist in severe supraspinatus muscle tears. A Markov texture discriminator applies adversarial constraints from the perspective of local pathological texture authenticity, further suppressing the generation of pseudostructures in the water-lipid separation results. Finally, the mixed signals of the boundary voxels of the supraspinatus muscle region of interest exhibit direction-dependent phase biases under the influence of microscopic magnetic susceptibility anisotropy. Traditional linear interpolation cannot distinguish the volume ratios of the three components—muscle matrix, endothelial fat, and interstitial water—at the sub-voxel scale. A multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction utilizes multi-echo complex phase evolution characteristics, combined with non-negative matrix factorization techniques with spatial prior distribution constraints, to perform sub-voxel-level component unmixing on the boundary voxels. The fat fraction is corrected according to the volume ratios of each sub-voxel component, mathematically eliminating the systematic biases caused by differences in microscopic interface magnetic susceptibility and partial volume effects, ensuring the physical accuracy and spatial consistency of the final quantitative result of the supraspinatus muscle fat fraction. These three-level processing mechanisms progressively constitute a complete technical logic for solving the core technical problem.

[0071] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0072] The specific implementation of step S01 is as follows: The subject's shoulder joint undergoes an IDEAL-IQ sequence scan. The scanning plane is selected as the standard research plane of the scapula in an oblique sagittal position, forming a "Y" shape. Multi-echo complex raw three-dimensional matrix data is acquired. The data is obtained in the [missing information - likely a specific data point or section]. echo time The complex signal matrix at point is denoted as ,in , , These are the matrix dimensions for the three spatial encoding directions, Represents the field of complex numbers, all The data from each echo are stacked along the echo dimension to form a high-dimensional tensor. , This represents the total number of echoes.

[0073] The specific implementation of step S02 is as follows: convert the high-dimensional tensor... Inputting a three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model, the water-lipid separation problem is transformed into the following convex optimization problem:

[0074] ;

[0075] The constraints are ,in The low-rank component tensor represents the signal of real tissue water and lipid evolution. The sparse component tensor represents the principal magnetic field. Phase distortion and motion artifacts caused by inhomogeneity For regularization tradeoffs, in Candidate values ​​were set with logarithmic intervals within the interval, and multiple rounds of cross-validation were performed on training and validation datasets for four types of pathological subtyping. The optimal comprehensive score for water-lipid separation purity was used as the criterion for determination. Tensor nuclear norm. Defined as the weighted sum of the singular values ​​of the tensor across all mode expansion matrices, the formula is as follows:

[0076] ;

[0077] In the formula, For tensor Along the first The expansion matrix of each pattern, The nuclear norm of a matrix is ​​the sum of all its singular values. For the first The weight coefficients for each pattern are empirically calculated as follows: , The total number of patterns in the tensor, here ,correspond , , , Four dimensions For sparse component tensors The sum of the moduli of all elements. Solved using the alternating direction multiplier method, the [number]th [element / unit]... The update steps for each iteration are as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] In the formula, This is the index of the current iteration round. For the first The dual variable tensor of the wheel, Initialize to All-zero tensors of the same dimension To augment the Lagrange penalty coefficient, the empirical value is... ~ The convergence speed on the validation set can be determined through preliminary experiments before training. A soft thresholding operator for singular values ​​of tensors, applying a threshold to the singular values ​​of the expansion matrices of each mode of the tensor. The reconstructed tensor after soft contraction works as follows:

[0082] ;

[0083] In the formula, and They are respectively The left and right singular vector matrices obtained from singular value decomposition. For the corresponding singular value vector, This represents constructing a diagonal matrix using vectors. for The conjugate transpose, here , This is an element-wise non-negation operation. Sparse soft thresholding operator. The process of action is as follows:

[0084] ;

[0085] In the formula, For sparse component tensors In spatial location , No. The element at each echo location, For symbolic functions, For modulo operation, here The iteration termination condition is:

[0086] ;

[0087] In the formula, For the Frobenius norm, The convergence threshold was determined by analyzing the slope plateau region of the iterative convergence curve on a typical case dataset, with a reference range of [missing information]. ~ After iterative convergence, from the low-rank component tensor Extracting the three-dimensional matrix of pure water phase With lipid phase three-dimensional matrix , Represents the real number field.

[0088] The specific implementation method of step S03 is as follows: ... and A spatial phase-decoupled cyclic skip generative adversarial network model driven by input physical information. The generator consists of dual-channel 3D residual dense blocks. The real and imaginary channels are independently encoded and then fused in an intermediate layer. The kernel size within the 3D residual dense block is [size missing]. The number of feature map channels increases progressively between 32 and 128, and the total number of generator parameters is approximately ~ Within the network, a physical iteration computation unit is embedded between every two 3D residual dense blocks. This unit receives the preliminary water-lipid separation parameter estimates output from the current layer. ,in , The number of 3D residual dense blocks in the generator, i.e., the total number of embeddings of physical iteration computation units, is determined by the network depth design. Input a differentiable distributed Loch equation simulator, at the... Calculate the theoretical signal at each echo time:

[0089] ;

[0090] In the formula, and The first The estimated proton density values ​​of the water and lipid components output by the layer are in units of arbitrary signal intensity. and The first Estimates of the transverse relaxation times of the water and fat components output from the layer, in units of time and... Consistent and The first The initial phase estimates of the water and fat components output by the layer, in radians. The imaginary unit, The dimension is arbitrary, representing signal strength. The physical residual is:

[0091] ;

[0092] In the formula, For the actual collection of the first A complex echo signal, with dimensions of arbitrary units of signal strength. The dimension is also any unit of signal strength. through The convolutional mapping is a physical loss gradient map with the same dimension as the main network feature map, weighted by coefficients. Superimposed with the gradient of the backpropagation of the main network, The reference range is The dynamic weight adjustment function was determined through ablation experiments. The calculation formula is:

[0093] ;

[0094] In the formula, The mean of the physical residuals of the validation set in the current round, with units of arbitrary signal strength. To counteract the loss mean, a dimensionless normalized value is used. is the mean of pixel-level reconstruction loss, and is the dimensionless normalized value. This is the dimensionless normalization coefficient, where the dimension is the reciprocal of the signal strength in any unit, and the reference range is... , This is a dimensionless coefficient, with a reference range of [missing information]. Both were validated through multiple rounds of ablation experiments. Output value and actual optimal Minimizing the error between them is the objective. It is a dimensionless value. When hour, ;when hour, ;when hour, Two-component transverse relaxation time The decoupling fitting model is:

[0095] ;

[0096] In the formula, and These represent the proton densities of the water and fat components, respectively, with dimensions of arbitrary units of signal intensity. and These are the transverse relaxation times of the water and fatty components, respectively, with the dimension being time units. The initial estimate is within the range of 20–40 ms. The initial estimated values ​​were within the range of 5–15 ms. These initial values ​​were determined using experimental data from the measurement of transverse relaxation time of normal shoulder joint muscles and adipose tissue reported in the literature. Furthermore, the values ​​were validated and corrected through pre-fitting of data from the normal group of subjects during the experimental preprocessing stage. and These represent the initial phases of the water and fat components, respectively, with dimensions in radians. The Levenberg-Marquardt iterative algorithm is used to evaluate the six parameters. Perform nonlinear least squares estimation, with the objective function as follows:

[0097] ;

[0098] In the formula, The theoretical signal is obtained by substituting the current parameter estimates into the two-component relaxation model. Here are the spatial smoothing constraint coefficients, which are dimensionless coefficients with a reference range of [missing information]. The method was determined through experiments on a phantom dataset with the objective of minimizing the relaxation time estimation error. The spatial smoothing constraint term penalizes the difference in relaxation time parameters between adjacent voxels, with dimensions equal to... Consistent, where the square of the signal strength is any unit. The dimensions are consistent with those of the summation term. The training loss function consists of a weighted sum of three parts, as described below:

[0099] ;

[0100] In the formula, For pixel-level reconstruction loss, The physical constraint loss generated for the physical iteration computation unit. For the Markov texture discriminator adversarial loss, all three terms are dimensionless normalized values. , , The weighting coefficients for each part satisfy the following conditions: This was determined through grid search experiments on the validation set. This is a dimensionless normalized value.

[0101] The specific implementation of step S04 is as follows: On the standard research-level water map, the physician manually delineates the region of interest along the anatomical boundary of the supraspinatus muscle. The region of interest is simultaneously copied to the fat ratio map to obtain the initial value of the fat fraction of each voxel within the region of interest.

[0102] The specific implementation of step S05 is as follows: a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction is introduced into the voxels at the boundary of the region of interest to decompose the mixed signal within a single voxel into muscle matrix components. endothelial fat content Interstitial moisture content The volume proportions of the three sub-voxel components are solved using nonnegative matrix factorization with spatial prior distribution constraints. The objective function is:

[0103] ;

[0104] In the formula, The multi-echo complex signal vector of the boundary voxel to be corrected. This is a dictionary matrix, with its three columns representing muscle matrix, endothelial fat, and interstitial water, respectively. The reference signal vector at each echo time point is obtained by the pure voxel within the region of interest through the two-component transverse relaxation time output in step S03. The mapping graph is obtained by averaging. Let be the volume ratio vector of the three sub-voxel components, satisfying and , For spatial prior constraint coefficients, the dimensions are... Consistent This represents the spatial gradient of fat fraction between the current voxel and its neighboring voxels, and is a dimensionless value. For dimensionless values, the reference range for the iterative convergence threshold is: ~ After convergence Subvoll-level corrections were performed on the boundary voxel fat fraction. The final formula for calculating the supraspinatus muscle fat fraction is as follows:

[0105] ;

[0106] In the formula, The intensity of the lipid phase signal within the region of interest. The two values ​​represent the water phase signal intensity within the region of interest; both have the same dimensions. It is a dimensionless ratio, and its range is [value range missing]. When reported as a percentage, multiply by 100%.

[0107] The specific implementation of step S06 is as follows: the supraspinatus muscle fat fraction is analyzed in conjunction with the magnetic resonance morphological classification results, and the subjects are grouped and classified into normal group, degenerative group, partial tear group and complete tear group, and a quantitative assessment report is output.

[0108] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0109] This embodiment selected 120 subjects and volunteers who underwent shoulder joint MRI examinations at a tertiary-level Class A hospital. All subjects were over 23 years old and presented with related clinical symptoms such as shoulder pain or weakness. Those with a history of acute shoulder trauma, surgery, arthritis, or other systemic musculoskeletal diseases were excluded. All subjects underwent IDEAL-IQ sequence scanning using a 3.0T MRI scanner, acquiring multi-echo complex raw three-dimensional matrix data.

[0110] After the scan was completed, the standard research plane in which the scapula appears as a "Y" shape in the oblique sagittal plane was selected as the analysis plane, such as... Figure 2 As shown, the standard research-level water image clearly presents the anatomical outline of the supraspinatus muscle and its relative positional relationship with surrounding structures. The multi-echo complex raw three-dimensional matrix data is input into a three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model. Using a convex optimization framework based on tensor robust principal component analysis, the model is iteratively solved using the alternating direction multiplier method, with regularized tradeoff coefficients. Set the threshold to 0.05, and the iterative convergence threshold to... High-purity aqueous phase three-dimensional matrix and lipid phase three-dimensional matrix were separated from the original signal algebraic structure, effectively removing the main magnetic field. Phase distortion term caused by inhomogeneity.

[0111] The aforementioned three-dimensional water phase matrix and three-dimensional lipid phase matrix are input into a spatial phase decoupled cyclic skip generative adversarial network model driven by physical information. The model generator uses dual-channel encoding of the real and imaginary parts of the signals. A Bloch equation simulator is embedded in the physical iteration computation unit to dynamically calculate the physical residuals between the theoretical and acquired signals at each echo time point, and the physical constraint loss weighting coefficients. The comprehensive score in this embodiment is adaptively determined by a dynamic weight adjustment function. The mean is 0.52, corresponding to Take 0.50. Transverse relaxation time of water components. The initial estimate is 28. transverse relaxation time of fatty components The initial estimate is 10. After convergence, the Levenberg-Marquardt iterative algorithm outputs six-parameter estimation results, with a spatial smoothing constraint coefficient of 0.05. The network synchronously outputs a refined water-lipid separation map and the transverse relaxation time of the two components. Mapping diagram, such as Figure 3 As shown, Figure 3 The images show fat ratios at the same standard research level, clearly demonstrating the spatial distribution of fat fraction within the supraspinatus muscle region. The differences in the degree of fat infiltration between different pathological groups are visually represented in the images.

[0112] On standard research-level water images, two experienced radiologists manually delineated the region of interest (ROI) along the anatomical boundaries of the supraspinatus muscle, and simultaneously copied the ROI to the fat ratio image. Supraspinatus muscle fat fraction data for all subjects were measured three times and averaged to reduce random error. Subsequently, a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction was introduced to perform sub-voxel-level partial volume effect correction on the voxels at the ROI boundary. The non-negative matrix factorization iterative convergence threshold was set to... Using the spatial gradient of fat fraction between adjacent voxels as a spatial prior, the boundary voxel mixing signal is decomposed into three sub-voxel components: muscle matrix, endothelial fat, and interstitial water. The boundary voxel fat fraction is then corrected according to the volume ratio, and finally... The supraspinatus muscle fat fraction of each subject was calculated and reported as a percentage.

[0113] Based on the morphological findings of magnetic resonance imaging, 120 subjects were divided into four groups: normal group (35 cases), degeneration group (31 cases), partial tear group (28 cases), and complete tear group (26 cases). The mean fat fraction of the supraspinatus muscle in each group is shown in Table 1.

[0114] Table 1. Summary of mean fat fraction in supraspinatus muscle for each pathological group

[0115]

[0116] As shown in Table 1, the mean fat fraction in the supraspinatus muscle increased stepwise as the pathological classification progressed from the normal group to the complete tear group. Significant differences in fat fraction distribution existed among the groups, and the quantitative results were highly consistent with the morphological classification. The fat fraction distribution characteristics of each pathological group are as follows: Figure 4 As shown.

[0117] To further verify the quantitative robustness of this method in the shoulder joint marginal region where the main magnetic field is highly uneven, the two-component transverse relaxation time of all subjects was extracted. The estimated results show that the transverse relaxation time of water components Lateral relaxation time of fatty components The distribution in each pathological group is as follows: Figure 5 As shown, each group and All can be independently and stably estimated, complete tearing group It was shorter than the normal group. The changes were relatively stable among the groups, which is consistent with the physical law of accelerated relaxation of extracellular water molecules after muscle fiber injury, indicating that the results of the two-group decoupling fitting are physically reasonable.

[0118] In this embodiment, the partial volume effect correction of the region of interest boundary voxels is evaluated by comparing the difference in fat fraction before and after correction. The distribution of the difference in fat fraction before and after correction for each pathological group is as follows: Figure 6 As shown, the boundary voxel correction amplitude was the largest in the completely torn group, reflecting that the systematic overestimation effect of microscopic magnetic susceptibility anisotropy on the fat fraction of the boundary voxel was most significant under the extremely heavy fat infiltration conditions. The corrected fat fraction is closer to the true fat content of the tissue.

[0119] Dynamic weight adjustment function Output distribution and corresponding results during the training reasoning process of all 120 subjects Value distribution as follows Figure 7 As shown, Figure 7 As shown, The values ​​showed a concentrated distribution in the middle range in the degeneration group and the partial tear group, while in the complete tear group... The value is too high. Adaptively taking a larger value strengthens physical constraints and suppresses water-oil phase inversion artifacts in severely non-uniform field scenarios, which is consistent with the physical logic of the scheme design.

[0120] The advancements of this invention compared to traditional methods are reflected in the following aspects: Traditional water-fat separation methods rely on image-domain phase constraint algorithms. However, in the shoulder joint area where the main magnetic field is highly uneven, the assumptions of the phase constraint model are violated, and water-fat inversion artifacts cannot be eliminated. In contrast, this invention uses unsupervised low-rank sparse decomposition of three-dimensional tensors to a priori decouple the field distortion phase term from the signal algebraic structure. Then, through a physical iterative computation unit, the Bloch equation is explicitly embedded into the network constraint, making water-fat separation physically self-consistent and fundamentally solving the problem of phase contamination propagation paths. Traditional single-relaxation models cannot distinguish the lateral relaxation times of muscle and fat. Significant systematic biases exist at voxels with high fat infiltration. This invention addresses this by using a complex-domain dual-relaxation spectrum fitting kinetic equation and the Levenberg-Marquardt iterative algorithm to independently estimate the relaxation parameters of the two tissue types, eliminating the systematic errors of single-relaxation models. Traditional calculations of fat fraction in regions of interest neglect boundary voxel mixing signals, leading to systematic overestimation under the influence of microscopic magnetic susceptibility anisotropy. This invention, through sub-voxel-level multi-component super-resolution unmixing, accurately separates boundary voxel components and corrects the fat fraction according to volume ratio, ensuring the physical accuracy of quantitative results at tissue interfaces.

[0121] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0122] Table 2. Variable Explanation Table (Part 1)

[0123]

[0124] Table 3. Variable Explanation Table (Part Two)

[0125]

[0126] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging, characterized in that, Includes the following steps: The shoulder joint of the subject was scanned using the IDEAL-IQ sequence to acquire multi-echo complex raw three-dimensional matrix data. The standard research plane with the scapula in the oblique sagittal position was selected for the scanning. Based on the original three-dimensional matrix data of multi-echo complex numbers, a three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model is constructed. The main magnetic field distortion phase term and artifacts in the original signal are separated by unsupervised blind source separation, and the pure water phase three-dimensional matrix and lipid phase three-dimensional matrix are extracted. A spatial phase decoupling cyclic skipping generative adversarial network model, driven by inputting the three-dimensional matrices of the aqueous and lipid phases into physical information, is used to globally correct water-lipid inversion artifacts, dynamically compensate for phase and amplitude errors, and adjust the transverse relaxation times of the water and lipid components. Independent two-component decoupling fitting was performed, and the water-lipid separation refinement map and the two-component lateral relaxation time were output simultaneously. Mapping diagram; On the standard research-level water image, manually delineate the region of interest along the anatomical boundary of the supraspinatus muscle, and simultaneously copy the region of interest to the fat ratio image. Based on the fat fraction of each voxel within the region of interest, a multi-component super-resolution unmixing model based on microscopic magnetic susceptibility anisotropy correction is introduced to perform sub-voxel-level partial volume effect correction on the boundary voxels of the region of interest and calculate the fat fraction of the supraspinatus muscle. By combining the fat fraction of the supraspinatus muscle with the results of magnetic resonance morphology typing, the subjects were grouped into normal group, degenerative group, partial tear group and complete tear group, and a quantitative assessment report was generated.

2. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 1, characterized in that, The optimization objective function of the three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model is specifically composed of the low-rank term constrained by the tensor nuclear norm and... The sparse terms of the norm penalty are weighted and superimposed to satisfy the constraints. ,in For high-dimensional tensors, For low-rank component tensors, It is a sparse component tensor.

3. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 2, characterized in that, The value of the regularization tradeoff coefficient in the optimization objective function is obtained by conducting multiple rounds of cross-validation experiments on a training and validation dataset covering four types of pathological subtypes, with the water-lipid separation purity index as the evaluation criterion for selecting the optimal value.

4. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 3, characterized in that, The solution to the three-dimensional tensor unsupervised low-rank sparse decomposition blind source water-lipid separation model is specifically achieved by using the alternating direction multiplier method to iteratively optimize in the multidimensional tensor space. The iteration termination condition is that the relative change in the Frobenius norm of the low-rank component tensor between two adjacent iterations is less than the convergence threshold.

5. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 4, characterized in that, The generator of the physical information-driven spatial phase decoupling cyclic jump generative adversarial network model is specifically composed of dual-channel three-dimensional residual dense blocks. The two channels correspond to the real part channel and the imaginary part channel of the multi-echo complex data, respectively. The two channels are independently encoded and then feature fusion is performed in the intermediate layer.

6. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 5, characterized in that, The generator introduces multi-scale cyclic skip connections in the middle layer of the network. Specifically, it modulates the shallow three-dimensional spatial detail feature map along the time echo dimension using gated cyclic units and then fuses it with the deep physical feature map element by element.

7. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 6, characterized in that, The generator embeds a physical iteration operation unit between every two three-dimensional residual dense blocks. After receiving the preliminary water-fat separation parameter estimate, the physical iteration operation unit inputs it into the Bloch equation simulator to calculate the physical residual between the theoretical signal and the actual acquired multi-echo signal. The physical residual is superimposed with the backpropagation gradient of the main network using weighted coefficients.

8. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 7, characterized in that, The discriminator of the physical information-driven spatial phase decoupling cyclic jump generative adversarial network model specifically adopts a Markov texture discriminator architecture, with the receptive field covering local pixel blocks, to evaluate the local authenticity of spatial pathological textures in the decoupled aqueous and lipid 3D images.

9. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 8, characterized in that, The training loss function of the physical information-driven spatial phase decoupling cyclic jump generative adversarial network model is specifically composed of three weighted parts: pixel-level reconstruction loss, physical constraint loss generated by the physical iteration operation unit, and Markov texture discriminator adversarial loss. The weight coefficients are determined by conducting grid search experiments on the validation set.

10. The method for quantifying fat infiltration in the supraspinatus muscle using magnetic resonance imaging according to claim 9, characterized in that, The weight coefficients of the physical constraint loss are adjusted by a dynamic weight adjustment function, which is calculated based on three data points: the mean physical residual, the mean adversarial loss, and the mean pixel-level reconstruction loss on the current round of the validation set. The weight coefficient range is adaptively switched according to the calculation results.

Images