Self-supervised learning-based magnetic resonance pararecombinamagnetic source separation imaging method, electronic device, and computer-readable storage medium

Abstract

This invention provides a magnetic resonance pararecombinamagnetic source separation imaging method based on self-supervised learning. [Solution] The present invention belongs to the field of medical imaging. This method includes the steps of: acquiring total susceptibility images and transverse relaxation rate images using multi-gradient echo data; constructing a deep learning model that integrates physical constraints; training the model in a self-supervised manner using the two types of images as input to learn spatially changing intermediate parameters; and finally calculating high-precision paramagnetic material distribution maps and diamagnetic material distribution maps based on these parameters and physical relationships. The present invention eliminates the need for time-consuming additional spin echo sequence acquisition, significantly reduces scan time, and enables more accurate and efficient quantitative separation and analysis of iron deposition and myelin sheathing changes in the brain.

Description

[Technical Field]

[0001] This invention belongs to the field of medical imaging, and more particularly to a magnetic resonance pararecombinamagnetic source separation imaging method based on self-supervised learning. [Background technology]

[0002] In the field of magnetic resonance imaging, quantitative susceptibility imaging (QSM) techniques can reconstruct tissue susceptibility distribution maps, providing an important tool for the quantitative evaluation of neurological diseases such as intracranial iron deposition and myelin changes. However, the total susceptibility of tissue is a superposition of the positive susceptibility of paramagnetic materials (such as iron) and the negative susceptibility of diamagnetic materials (such as myelin). Existing QSM reconstruction results fail to effectively separate the contributions of these two types of materials, which directly leads to bias in the quantitative analysis of specific pathological changes (e.g., iron concentration), affecting the accuracy of disease diagnosis and monitoring.

[0003] To separate paramagnetic and diamagnetic signal sources, existing techniques typically require reliance on additional spin echo sequences to acquire additional transverse relaxation rate R'2 information, resulting from local magnetosusceptibility effects, as an auxiliary constraint for solving physical models. This method has clear limitations in clinical practice. First, acquiring additional sequences significantly extends scan time, increasing patient burden and discomfort, and also increases the risk of motion-induced artifacts. Second, reliance on specific sequences limits the versatility of the method and its direct application to standard clinical scan protocols. [Overview of the project] [Means for solving the problem]

[0004] To solve the above technical problems, the present invention provides a magnetic resonance normal retamagnetic source separation imaging method based on self-supervised learning. The steps include acquiring a total susceptibility image and a transverse relaxation image based on the collected multi-gradient echo magnetic resonance data, constructing a deep learning model, wherein an input of the model is the total susceptibility image and the transverse relaxation rate image, and an output of the model is a plurality of spatially varying intermediate parameter maps; training the deep learning model by self-supervised learning based on the total susceptibility image and the transverse relaxation rate image, wherein a loss function used for training incorporates a constraint term based on a physical relationship; obtaining the intermediate parameter maps based on a total susceptibility image and a transverse relaxation rate image corresponding to input magnetic resonance data to be processed by using the trained deep learning model, and performing calculation based on the intermediate parameter maps and a predetermined physical relationship formula to obtain a paramagnetic substance distribution map and a diamagnetic substance distribution map.

[0005] Optionally, the step of obtaining the total susceptibility image and the transverse relaxation rate image comprises: generating the total susceptibility image through unwrapping, background magnetic field removal and susceptibility inversion processing based on phase information of the multi-gradient echo magnetic resonance data; generating the transverse relaxation rate image by fitting a signal attenuation curve based on amplitude information of the multi-gradient echo magnetic resonance data.

[0006] Optionally, the plurality of spatially varying intermediate parameter maps include a paramagnetic relaxation contribution parameter, a diamagnetic relaxation contribution parameter, and a proportional parameter and an offset parameter for describing a linear relationship between the transverse relaxation rate and a relaxation rate induced by magnetic susceptibility.

[0007] Optionally, the step of performing calculation based on a predetermined physical relationship between the intermediate parameter maps and the total susceptibility image as well as the transverse relaxation rate image to obtain a paramagnetic substance distribution map and a diamagnetic substance distribution map specifically comprises: obtaining an intermediate relaxation amount by subtracting the offset parameter from the transverse relaxation rate image and then dividing by the proportional parameter based on the paramagnetic relaxation contribution parameter, the diamagnetic relaxation contribution parameter, the proportional parameter, the offset parameter, the total susceptibility image and the transverse relaxation rate image; performing calculation by combining the total susceptibility image, the paramagnetic relaxation contribution parameter and the diamagnetic relaxation contribution parameter, and separating the paramagnetic substance distribution map and the diamagnetic substance distribution map.

[0008] Optionally, the deep learning model comprises a physics-aware fusion module, which is used for fusing the features of the total susceptibility image and the features of the transverse relaxation rate image, the physics-aware fusion module calculates the interaction relationship between two types of features by using a low-rank bilinear attention mechanism, and introduces physics guidance embedding information composed of the two types of features and the calculation result thereof to adjust feature interaction.

[0009] Optionally, the physics-aware fusion module outputs a final fused feature by an adaptive residual connection method, wherein the adaptive residual connection method dynamically performs weighted addition on the feature processed by the physics-aware fusion module and the direct addition result of the two types of original input features through a learnable spatial weight map.

[0010] Optionally, the self-supervised loss function comprises a pre-training loss term and a physics constraint loss term, the pre-training loss term is used for constraining the numerical values of the intermediate parameter map to approach pre-initial values, the physics constraint loss term at least comprises a first constraint term for constraining that the sum of the paramagnetic substance distribution map and the diamagnetic substance distribution map is equal to the total susceptibility image, a second constraint term for constraining that the paramagnetic substance distribution map is non-negative and the diamagnetic substance distribution map is non-positive, and a regularization constraint term for promoting spatial smoothness of a separation result.

[0011] Optionally, the self-supervised loss function further includes a regular constraint term used to constrain the orthogonality of the low-rank projection matrix used by the attention mechanism in the physical perception fusion module.

[0012] On the other hand, the present invention also provides an electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the method is realized when the processor executes the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium that stores a computer program and realizes the method when the computer program is executed by a processor. [Effects of the Invention]

[0014] Compared to existing technologies, the present invention has the following advantages and technical effects.

[0015] This invention proposes a magnetic resonance paradiamagnetic source separation imaging method based on self-supervised learning. This method utilizes only multi-gradient echo data and identifies the susceptibility of paramagnetic materials (χ²). para ) and diamagnetic material magnetic susceptibility (χ dia Based on the physical constraint relationship between ) and the total magnetic susceptibility χ, a series of loss functions integrating prior physical information are constructed and used for self-supervised training of deep learning models. This enables effective separation of paramagnetic and diamagnetic material signals and high-precision magnetic resonance imaging reconstruction without the need for manually labeled data and additional R'2 measurement information, improving the accuracy of quantitative analysis of intracranial iron deposition and myelin changes in neurological diseases, and promoting early diagnosis, progress evaluation, and widespread clinical application of related diseases. [Brief explanation of the drawing]

[0016] The drawings, which constitute part of this Application, are used to provide a further understanding of this Application, and the exemplary embodiments and descriptions thereof are used to illustrate this Application and do not constitute an unreasonable limitation of this Application. In the drawings: [Figure 1] This is a comparison diagram of the results and residuals of separation of normal retamagnetic materials using different methods according to the embodiments of the present invention. [Figure 2] This is a feature visualization diagram showing the effectiveness of the physical perception fusion module for verifying βθ in an embodiment of the present invention. [Figure 3] This is a visualization of the four spatially changing feature parameters learned by the model of the embodiment of the present invention. [Figure 4] This is a diagram of a self-supervised paradiamagnetic material separation model of an embodiment of the present invention, showing (A) a model architecture diagram, (B) a physical perception fusion module, and (C) a low-rank bilinear attention system. [Figure 5] This is a technical flowchart of an embodiment of the present invention. [Modes for carrying out the invention]

[0017] Furthermore, if there are no conflicts, the embodiments and features described herein can be combined with each other. The embodiments will now be described in detail with reference to the drawings.

[0018] The steps shown in the flowchart can be executed by a computer system, such as a series of computer-executable instructions, and although the flowchart shows a logical order, in some cases the steps may be executed in a different order than those shown or described herein.

[0019] Example 1

[0020] As shown in Figure 5, this embodiment provides a magnetic resonance paradiamagnetic source separation imaging method based on self-supervised learning. 1. To collect amplitude and phase data based on a magnetic resonance multi-gradient echo sequence, 2. For amplitude data, R*2 is calculated using the ARLO algorithm (autoregression of linear operations), and phase data is obtained through unwrapping, background removal and inversion to obtain the total magnetic susceptibility χ, 3. R*2, χ and the paramagnetic-diamagnetic map χ para (>0) and χ dia (<0), designing a calculation method for χ para and χ dia that does not require R'2, 4. Designing a physical perception fusion module to efficiently fuse features of R*2 and χ, 5. Constructing a self-supervised deep learning network model, wherein the model only learns four spatially varying intermediate parameter maps, 6. Designing a self-supervised loss function based on prior information, 7. Training the deep learning model in a self-supervised manner using R*2 and χ obtained in post-processing, and verifying the performance, 8. Using the trained model to complete separation of paramagnetic and diamagnetic sources of a substance in forward data, and constructing a paramagnetic substance distribution map χ para and a diamagnetic substance distribution map χ dia .

[0021] Constructing the self-supervised loss function based on prior information comprises: 1. A pre-training loss function L based on prior information pre , 2. A model adjustment loss function L based on physical constraints phy , 3. A regularization loss L for preventing collapse of low-rank spatial attention otrh .

[0022] The physical perception fusion module comprises: 1. A bilinear low-rank spatial attention system, 2. A physical guidance module based on the cross feature space of R*2 and χ, 3. An adaptive residual connection method based on dynamic maps.

[0023] The four spatially varying intermediate parameter maps to be learned are: 1. R'2 and χ para and χ dia Linear relationship between

number

number

[0024] The specific process includes the following:

[0025] Step 1: Collect amplitude and phase data based on a multi-gradient echo sequence: A subject is scanned using a multi-gradient echo magnetic resonance imaging sequence, collecting amplitude and phase data corresponding to multiple echo time points during the same scanning process. The collected amplitude data represents signal intensity information from different tissues, and the phase data reflects local magnetic field changes caused by differences in the magnetic susceptibility of the tissues. The multi-echo acquisition method provides amplitude and phase information with a time dimension, providing fundamental data for subsequent susceptibility χ calculations, R*2 estimation, and normal-reactive source separation.

[0026] Step 2: Preprocess amplitude and phase data: R*2 is obtained by fitting the amplitude image of the multi-gradient echo data using the ARLO method. The original phase data is sequentially preprocessed using a multi-step process including Laplacian phase unwrapping, brain tissue extraction based on FSL BET, background magnetic field removal using the V-SHARP method, and susceptibility inversion using msQSM to generate high-quality QSM images (χ) for magnetic source separation. R2 is the model

number

[0027] Step 3: χ para and χ dia Construct a calculation method for: The total magnetic susceptibility χ is the paramagnetic susceptibility (χ para >0) and diamagnetic susceptibility (χ²) dia Being affected by both <0) simultaneously:

number

[0028] In a static magnetic field, a material undergoes transverse relaxation R2 under specific radio frequency pulse excitation. Under static dephase conditions, magnetic field inhomogeneity caused by an external susceptibility source induces additional transverse relaxation R'2, satisfying the following relationship:

number

number

number

number

[0029] Step 4: Design the physical perception fusion module: From equation (5), using only the total susceptibility χ calculated from the phase image and R*2 calculated from the amplitude image, χ para and χ dia We can calculate and complete the paradiamagnetic material separation task. In this embodiment, we design a deep learning module based on physical perception to efficiently fuse χ and R*2 feature information. This module mainly consists of three parts: a low-rank bilinear attention system, physical guidance, and adaptive residual connections.

[0030] Low-rank bilinear attention systems aim to efficiently model the quadratic interaction relationship between bimodals. Conventional self-attention calculations focus on the projected features.

number

number

number

[0031] The total computational complexity is 4rdN + 2rN 2 Since r ≪ d ≪ N, the computational complexity is significantly reduced. Next, we calculate self-attention using the scaling dot product method:

number

[0032] Physical guidance modulates the interactions between multimodal features according to their inherent physical relationships. During the feature projection process, projected features K and V, derived from χ and R*2, may lose their intrinsic physical attributes. Therefore, physical guidance embedding is introduced to inject prior physical knowledge with strong susceptibility or relaxation sensitivity into the feature space:

number

number

number

number

number

number

[0033] Adaptive residual connections are used to balance information within the residual pathway. If we denote the physical perceptual fusion module as PAF, the features after adaptive connection are expressed as follows:

number

number

[0034] Step 5: Build a self-supervised deep learning network model: Figure 4 shows the architecture diagram of the deep learning network model, where (A) is the overall flowchart, (B) is the physical perception fusion module in the above step, and (C) is the multi-head low-rank bilinear attention system. The model takes χ and R*2 as inputs and first goes through a grouping convolution encoding layer to process the respective feature representations F χ and

number

[0035] Unlike conventional deep learning models that directly learn the final result, this embodiment proposes an innovative approach in which four intermediate parameter maps are learned, and then the separation result is calculated using a physical model equation. para and χ dia This guarantees that equation (1), which is a fundamental physical constraint, is completely satisfied, and the accuracy of the separation results is greatly improved.

[0036] Step 6: Design a self-supervised loss function based on prior information: In this embodiment, a self-supervised loss function is designed based on prior information. First, using an existing method, α is calculated using several fixed brain regions. p =α d The fitting was done with =137, k=1.1781 and b=12.9629, but in this embodiment α p , α d k and b are designed as spatially variable parameter maps to adapt to the differences in the physical properties of each tissue in the whole brain. In this embodiment, a pre-trained loss function is designed using whole-brain fixed parameters fitted by an existing method as prior information:

number

number

number

number

[0037] Finally, obtain the physical loss:

number

[0038] Furthermore, to prevent attention collapse from occurring in multi-head low-rank upper-linear attention systems, multi-head regularization loss is used:

number

number

[0039] Step 7: Model Training: Based on the input features obtained in the above step, a self-supervised deep learning model is constructed and trained. During the training process, quantitative susceptibility imaging (χ) data and its derived data (R*2) are used as model input, and intermediate parameter maps necessary for paramagnetic and diamagnetic separation are obtained through forward propagation. A self-supervised loss function is constructed by combining these with pre-designed physical constraints, and the model parameters are backpropagated and iteratively updated. The training process does not require manually labeled data or additional R'2 measurement information. Model weights are optimized stepwise through multiple rounds of iterations, ensuring the model satisfies physical consistency constraints while simultaneously improving the stability and accuracy of the paradiamagnetic material separation results. This process continues until the model converges and reaches a predetermined performance index.

[0040] Step 8: Using the Model: After model training is complete and predetermined performance metrics are reached, the trained model is applied to the processing of the target magnetic resonance data. Specifically, data is acquired using the method in "Step 1," preprocessing is completed using the method in "Step 2," the preprocessed data is input into the trained model, and intermediate parameter maps necessary for paramagnetic and diamagnetic separation are automatically generated by forward propagation inference. The corresponding paramagnetic susceptibility distribution χ is then determined based on the physical model equation (5). para and diamagnetic susceptibility distribution χ dia This process calculates the following without requiring the collection of additional T2 mapping sequences or artificial intervention, enabling efficient and stable completion of paradiamagnetic source separation imaging and providing a reliable basis for subsequent quantitative analysis of intracranial iron deposition and myelin changes.

[0041] Example 2

[0042] This embodiment provides a magnetic resonance paradiamagnetic source separation imaging method based on self-supervised learning, which includes the following:

[0043] 1. Effect of separation and reconstruction of normal-reactive magnetic sources: The self-supervised learning-based magnetic resonance pararemagnetic source separation imaging method provided in this application uses only gradient echo data and combines it with a self-supervised model based on physical perception and prior information to separate the total susceptibility χ from the paramagnetic portion χ. para and diamagnetic portion χ dia It separates into. In clinical neurological disease applications, χ para For quantitative analysis of iron distribution, χ dia This can be used to quantify myelin distribution and offers a new perspective to disease research.

[0044] Figure 1 shows the results of separating diamagnetic materials using different methods and a comparison of residuals. Here, χ-sepnet is a supervised separation method that requires training labels, while APART-QSM and χ-separation are conventional separation methods that use R'2 to assist the calculation. In the figure, the top two rows are χ para and its residuals, the bottom two lines are χ diaThe residuals are shown and magnified three times to demonstrate the advanced nature of this embodiment's method. From the residual plot, it can be seen that this embodiment's method still has the closest results to the benchmark and the smallest residuals, even without R'2 and training labels.

[0045] Table 1 shows the quantitative comparison results. Here, SSIM represents structural similarity, PSNR represents the peak signal-to-noise ratio, NRMSE represents the normalized root mean square error, and HFEN represents the high-frequency error. From the table, it can be seen that the method of this embodiment shows superior reconstruction results compared to general QSM reconstruction methods in all metrics.

[0046] [Table 1]

[0047] 2. Effects of different loss functions: Table 2 shows a quantitative comparison of the degradation of model performance that occurs when different loss function components are removed in a self-supervised learning framework. The results in the table show that removing any of the loss function components degrades model performance to varying degrees. This indicates that the self-supervised loss function designed in this embodiment plays a crucial role and is effective in the model training process.

[0048] [Table 2]

[0049] 3. Effects of the Physical Perception Fusion Module: Figure 2 shows β θ The visualization results for ∈[0,1] demonstrate the effectiveness of the physical-perceptual fusion module. If we denote the physical-perceptual fusion module as PAF, the feature fusion output of the model is expressed as follows:

number

[0050] 4. The effect of intermediate parameters learned by the model: Unlike conventional deep learning models that directly output the target result, this embodiment outputs the model output using four spatially varying feature parameters α p , α d Design as k and b. Existing methods typically use constant values ​​(α) for the above parameters. p =α d The parameters are set as (=137, k=1.1781, and b=12.9629), ignoring the differences in different tissues and spatial locations. In this example, we innovatively propose dynamically learning the spatial variation characteristics of the above parameters through a deep learning model to improve the accuracy of separating paramagnetic and diamagnetic materials.

[0051] Figure 3 shows the visualization results of the four characteristic parameters in two examples. In each example, the first row shows the spatial distribution of the parameters, and the second row shows the statistical histogram of the corresponding parameter and its mean value. From the figure, all four parameters show clear spatial non-uniform distribution characteristics, which is consistent with the physical expectation that they should vary depending on the tissue type and magnetic material distribution, rather than being fixed constants. Specifically, the parameter values ​​are relatively small in the lateral ventricular region due to the low magnetic susceptibility of cerebrospinal fluid, while the parameter values ​​increase accordingly in the high-magnetic-susceptibility cortical and deep gray matter regions, indicating that the parameter distribution can adaptively reflect the magnetic properties of materials with different characteristics.

[0052] Furthermore, the parameter mean values ​​clearly show that the parameters obtained through dynamic learning are distinctly different from the fixed parameters fitted based on a small number of brain regions using conventional methods. This indicates that there were limitations to the conventional modeling method using global constant parameters. In this example, by learning the spatial variation characteristics of the parameters, the above shortcomings are effectively overcome, providing a more detailed and reliable modeling basis for paradiamagnetic source separation.

[0053] On the other hand, this embodiment also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, and realizes the method when the processor executes the computing program.

[0054] On the other hand, this embodiment also provides a computer-readable storage medium that stores a computer program, and realizes the method when the computer program is executed by a processor.

[0055] While the most favorable specific embodiments of this application have been described above, the scope of protection is not limited thereto. Any modification or substitution that a person skilled in the art could easily conceive within the technical scope disclosed herein should be included within the scope of protection. Therefore, the scope of protection should be determined by the scope of protection of the claims.

Claims

1. A magnetic resonance normal retamagnetic source separation imaging method based on self-supervised learning, The steps include acquiring a total susceptibility image and a transverse relaxation image based on the collected multi-gradient echo magnetic resonance data, Step 1: Construct a deep learning model, the input to which the model is the total susceptibility image and the transverse relaxation rate image, and the output to which the model is a plurality of spatially varying intermediate parameter maps. Based on the total susceptibility image and transverse relaxation image, the deep learning model is trained by self-supervised learning, where the self-supervised loss function used for training incorporates constraint terms based on physical relationships. The process includes the steps of: using a trained deep learning model, obtaining an intermediate parameter map based on a total susceptibility image and a transverse relaxation image corresponding to the input magnetic resonance data to be processed; performing calculations based on the intermediate parameter map and predetermined physical relationships to obtain a paramagnetic material distribution map and a diamagnetic material distribution map; The deep learning model includes a physical perception fusion module, which is used to fuse the features of the total susceptibility image with the features of the transverse relaxation image. The aforementioned physical perception fusion module is a self-supervised learning-based magnetic resonance pararecombinamagnetic source separation imaging method characterized by calculating the interaction relationship between two types of features using a low-rank bilinear attention mechanism, and adjusting the interaction relationship between the two types of features by introducing physical guidance embedding information composed of the calculation results of the two types of features and the interaction relationship between the two types of features.

2. The steps of acquiring the total magnetic susceptibility image and the transverse relaxation rate image are as follows: The steps include generating the total susceptibility image based on the phase information of the multi-gradient echo magnetic resonance data, through unwrapping, background magnetic field removal, and susceptibility inversion processing, The method according to claim 1, comprising the step of generating the transverse relaxation rate image by fitting a signal attenuation curve based on the amplitude information of the multi-gradient echo magnetic resonance data.

3. The method according to claim 1, characterized in that the spatially varying intermediate parameter maps include a paramagnetic relaxation contribution parameter, a diamagnetic relaxation contribution parameter, and proportional and offset parameters for describing a linear relationship between the transverse relaxation rate and the relaxation rate caused by the magnetic susceptibility.

4. The step of obtaining the paramagnetic material distribution map and the diamagnetic material distribution map involves separating the paramagnetic material distribution map and the diamagnetic material distribution map based on the paramagnetic relaxation contribution parameter, the diamagnetic relaxation contribution parameter, the proportional parameter, the offset parameter, the total magnetic susceptibility image, and the transverse relaxation rate image, and specifically, The steps include: obtaining the intermediate relaxation amount by subtracting the offset parameter from the transverse relaxation rate image and then dividing by the proportional parameter; The method according to claim 3, characterized by comprising the step of performing a calculation by combining the intermediate relaxation amount, the total susceptibility image, the paramagnetic relaxation contribution parameter, and the diamagnetic relaxation contribution parameter, and separating the paramagnetic material distribution map and the diamagnetic material distribution map.

5. The method according to claim 1, characterized in that the physical perception fusion module outputs a final fused feature by an adaptive residual connection method, and the adaptive residual connection method dynamically weights and adds the result of direct addition of the feature processed by the physical perception fusion module and two types of original input features through a learnable spatial weight map.

6. The self-supervised loss function includes a pre-training loss term and a physical constraint loss term, The aforementioned pre-training loss term is used to constrain the values ​​of the intermediate parameter map to approach the pre-initial values. The method according to claim 1, characterized in that the physical constraint loss term includes at least a first constraint term that constrains the sum of the paramagnetic material distribution map and the diamagnetic material distribution map to be equal to the total susceptibility image, a second constraint term that constrains the paramagnetic material distribution map to be non-negative and the diamagnetic material distribution map to be non-positive, and a regularization constraint term that promotes spatial smoothness of the separation result.

7. The method according to claim 6, characterized in that the self-supervised loss function further includes a regular constraint term, the regular constraint term is used to constrain the orthogonality of the low-rank projection matrix used by the low-rank bilinear attention mechanism in the physical perception fusion module and to prevent the collapse of multi-head attention.

8. An electronic device comprising a memory, a processor, and a calculation program stored in the memory and executable on the processor, wherein when the processor executes the calculation program, the method according to any one of claims 1 to 7 is realized.

9. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it realizes the method according to any one of claims 1 to 7.

Images