A method and system for processing brain magnetic resonance images

Abstract

The application belongs to the technical field of image processing, and discloses a brain magnetic resonance image processing method and system. The method iteratively processes the image with artifacts and the original k-space data in the mixed k-space and image space domain. The core is: through a differentiable magnetic resonance physical simulation module, a simulation k-space is generated based on the predicted artifact field, and an image processing network is optimized by minimizing the physical simulation error of the simulation k-space and the original k-space data; at the same time, a graph neural network is used to model the brain anatomy to impose structural constraints; finally, features are extracted from the processed residual artifacts to update a digital twin model associated with a specific scanner to realize active equipment management. The application realizes physically interpretable and anatomically reasonable artifact correction, and is extended to device-level active quality control.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method and system for processing brain magnetic resonance images. Background Technology

[0002] Brain MRI is an indispensable imaging technique for modern neuroscience research and clinical diagnosis. However, its imaging process is lengthy and highly sensitive to the physical environment, easily affected by various factors such as patient physiological movements (e.g., involuntary head movements, breathing, heartbeat) and imperfections in equipment hardware (e.g., gradient field nonlinearity, main magnetic field drift), resulting in complex artifacts on the images. These artifacts severely degrade image quality, affecting not only doctors' interpretation and diagnosis but also causing devastating errors in subsequent high-precision quantitative analyses that rely on the images (e.g., functional connectivity analysis, tumor volume measurement).

[0003] While existing deep learning-based artifact correction methods have made some progress, they generally suffer from the following deep-seated and insurmountable bottlenecks:

[0004] 1. Most methods treat artifact correction as a purely image-to-image transformation task, operating within the image pixel domain. This approach ignores the fundamental physical cause of artifacts: phase and amplitude errors during k-space data acquisition. Therefore, the correction process is like a black box, making it difficult to guarantee the physical fidelity and reliability of the results, and sometimes even producing new, anatomically illogical, illusory structures.

[0005] 2. Traditional convolutional neural networks (CNNs) or Transformer models treat brain images as a uniform grid of pixels, lacking an understanding of the fact that the brain itself is an organ composed of brain regions with different functions and physical characteristics and complex topological connections. This results in the model being blind in predicting deformation fields, failing to guarantee that the correction results fully conform to the inherent constraints of biological anatomy.

[0006] 3. Existing artifact correction tools are all reactive, meaning they passively repair the image after a poor scan has occurred. They cannot feed the information extracted from the artifacts back into the scanning device itself to form a closed-loop management system. Therefore, they cannot prevent or proactively compensate for the generation of artifacts at the source, nor can they monitor the health status of the scanner in the long term. Summary of the Invention

[0007] The present invention aims to at least partially solve the aforementioned technical problems. Therefore, the objective of the present invention is to provide a method and system for processing brain magnetic resonance imaging (MRI) images.

[0008] To achieve the above objectives, this invention proposes a method for processing brain MRI images. Specifically, it involves iteratively processing a target image with artifacts and its corresponding original k-space data in a hybrid k-space and image space domain. In each iteration, an image processing network predicts an artifact field (e.g., a motion field). Then, an innovative differentiable MRI physics simulation module applies this predicted artifact field to an ideal reference k-space to generate a simulated k-space. By minimizing the physics simulation error between this simulated k-space and the patient's original k-space data, the image processing network is back-optimized, forcing it to learn an artifact process consistent with physical principles. During artifact field prediction, a graph neural network is used to model the brain's anatomical structure, with nodes corresponding to anatomical regions in a brain atlas. Anatomical consistency constraints are imposed through message passing between nodes. After iterative processing, the method extracts residual artifact features that cannot be explained by motion models and uses them to update a scanner digital twin model associated with a specific scanner. Finally, the system outputs a high-quality, post-corrected image.

[0009] In a preferred embodiment, the iterative processing is an alternating optimization process, for example, predicting a fine motion artifact field in the image space, then returning to the k-space, performing data consistency correction and updates based on the motion field, and repeating this process until convergence, thereby finding the optimal solution between the two domains.

[0010] In a preferred embodiment, the differentiable magnetic resonance physics simulation module implements the physical equations of MRI imaging (such as simplified forms of the Bloch equations) in the form of neural network layers or differentiable operators, capable of accurately simulating how at least one or more physical processes selected from the following introduce phase and amplitude errors during k-space data acquisition: patient physiological motion, main magnetic field (B0) inhomogeneity, and radio frequency field (B1) inhomogeneity.

[0011] In a preferred embodiment, the graph neural network introduces strong biological priors. By learning the physical connections and motion conduction constraints between different brain regions (nodes) (e.g., intracranial brain tissue undergoes rigid motion as a whole, while non-rigid deformations may exist between different lobes), it provides high-level structural guidance for artifact field prediction, ensuring that the corrected image does not exhibit distortions or deformations that violate anatomical common sense.

[0012] In a preferred embodiment, the residual artifact features are precisely defined as systematic and repetitive artifact patterns that cannot be explained by modeled physical processes (such as motion). These features are considered a direct reflection of the scanner hardware state (such as nonlinearity of gradient coils, drift of the magnetic field over time, etc.).

[0013] In a preferred embodiment, the scanner digital twin model is a dedicated longitudinal model built for each physical scanner. It continuously receives and learns the residual artifact characteristics generated by the device, thereby accurately tracking and modeling the unique hardware personality and state of the device over time.

[0014] Based on the digital twin model, this invention achieves a groundbreaking proactive quality control function: before processing new scanning data for this specific scanner, the system can use its digital twin model to generate a feedforward pre-correction field, which proactively compensates for the known, inherent systematic artifacts of the device during the image reconstruction stage, thereby improving the imaging quality from the source.

[0015] Furthermore, this invention also implements predictive maintenance functionality: the system compares the residual artifact features extracted from the current scan with a health baseline stored long-term in the digital twin model, representing the normal operating status of the equipment. Once the current deviation exceeds a preset threshold, the system generates an alarm indicating an abnormal equipment status, prompting the need for hardware calibration or maintenance.

[0016] The present invention also provides a corresponding processing system, which is implemented by one or more processors, and whose internal functional modules closely correspond to the methods. The core includes: a hybrid domain iterative processing module, a graph neural network guidance module, and a scanner digital twin module.

[0017] In the system of this invention, these three core modules do not work independently, but rather in a highly collaborative manner. The hybrid domain iterative module provides interpretability at the physical level, the graph neural network guidance module ensures rationality at the biological anatomical level, and the digital twin module enables proactive quality control at the device level. Together, they constitute a complete closed-loop intelligent solution from the physical to the biological to the device level.

[0018] The beneficial effects of this invention are as follows:

[0019] This invention introduces differentiable physical simulation and hybrid domain iteration, transforming the correction process from a black box into a reverse engineering of the real physical process, resulting in extremely high fidelity and reliability. By introducing graph neural networks to model brain structure, the correction process is endowed with anatomical knowledge, effectively avoiding artifacts and ensuring the biological rationality of the output image. By constructing a digital twin of the scanner, this invention, for the first time, transforms artifact information into longitudinal tracking, feedforward pre-correction, and predictive maintenance of device status, expanding the value boundary of artifact correction technology.

[0020] This invention provides the most solid and reliable data foundation for various downstream clinical applications, such as building more accurate stroke prediction models and conducting detailed brain surgery planning, by outputting unprecedented high-quality, high-fidelity, and high-reliability MRI images. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the hybrid domain iterative loop of a brain magnetic resonance image processing method provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of a multi-module collaborative architecture of a processing system provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] It should be understood that, and also noted, in the embodiments, the functions / actions may appear in a different order than those shown in the figures. For example, depending on the functions / actions involved, they may actually be performed substantially concurrently, or sometimes the two figures shown consecutively may be performed in reverse order.

[0025] This embodiment describes the complete workflow of the method and system of the present invention when processing a brain MRI image with motion artifacts.

[0026] Reference Figure 1 The figure visually illustrates the core hybrid domain iterative correction loop of this invention. Once a target image with artifacts is acquired, along with its original k-space data collected during scanning, the hybrid domain iterative processing module initiates this closed-loop optimization process.

[0027] Artifact Field Prediction (Image Space): As shown at the top of the loop, the processing flow begins in the image space. The target image with artifacts is input into an image processing network. The core task of this network (e.g., a UNet) is to predict the primary physical field causing the artifacts, in this embodiment, the motion artifact field. This field describes the three-dimensional displacement and rotation that occurs at each pixel of the head during the scan. In this step, the graph neural network guidance module plays a crucial role. The system first automatically segments the input image into multiple anatomical regions using a standard brain atlas and abstracts these regions as nodes in a graph. The graph neural network imposes anatomical constraints by message passing at these nodes, guiding the image processing network to generate an anatomically more plausible motion field. This ensures that the predicted deformations do not violate the physical structural characteristics of the brain.

[0028] Physical Simulation and Error Calculation (k-space): Next, as shown in the flow chart on the right side of the loop, the predicted motion field is fed into the differentiable MRI physical simulation module. The processing flow then enters the k-space domain. This module is a software component that implements the MRI signal acquisition equations (e.g., Fourier transforms involving patient motion modulation, main magnetic field bias, or radio frequency field amplitude modulation). It calculates what the resulting k-space data would look like if an ideal object underwent one or more of these physical processes during acquisition. This calculation result is the simulated k-space. Subsequently, the system calculates the difference between this simulated k-space and the patient's actual, artifact-laden raw k-space data. This difference, called the physical simulation error, quantifies the degree of deviation between the predicted artifact field and the actual physical process.

[0029] Network Optimization and Iteration: As shown in the feedback path on the left side of the loop, the physical simulation error is used as the loss function to update the weights of the image processing network through gradient backpropagation. The essence of this process is: if the predicted artifact field is incorrect, then the simulated k-space will not match the real k-space, and the large error will drive the network to predict a more accurate artifact field in the next iteration. This process of predicting in image space, calculating the error in k-space, and feeding back the optimization will alternate and iterate several times until the physical simulation error converges to a minimum. At this point, the system believes it has found the true physical process that best explains the artifacts in the original k-space data.

[0030] Final Image Generation: After iterative convergence, the system uses the final, most accurate artifact field to perform a final inverse correction on the original image with artifacts or k-space, generating a clear, artifact-free, and highly faithful post-corrected image. This image is of extremely high quality, providing reliable data input for subsequent clinical applications, such as establishing high-precision stroke prediction models.

[0031] The above process perfectly corrects artifacts caused by specific physical processes (such as motion), but there are usually some small, systematic residual artifacts between the final corrected image and the ideal image. These residuals are sent to the scanner's digital twin module for further processing.

[0032] Reference Figure 2 The figure illustrates the overall collaborative architecture of the processing system of the present invention, clearly depicting how the three core modules work together and how information flows.

[0033] The system input consists of a target image with artifacts and raw k-space data. This data first enters the core correction engine (corresponding to...) which is a collaborative process of a hybrid domain iterative processing module and a graph neural network guidance module. Figure 1 (The loop). After completing the iterative correction, the engine produces two intermediate outputs: one is the corrected high-quality image, and the other is the residual artifact features that cannot be explained by the modeled physical process.

[0034] The scanner's digital twin module then receives these residual artifact features. This module is a dedicated longitudinal model built for each physical scanner. It continuously receives and learns the residual artifact features generated by the device, thereby accurately tracking and modeling the unique hardware characteristics and state of the device over time.

[0035] Ultimately, as Figure 2 As shown, the system produces two types of key outputs:

[0036] Output 1: High-quality post-processed image after correction. This is the final product after physical interpretation and reasonable anatomical correction, which can be directly used for clinical diagnosis or scientific research analysis.

[0037] Output 2: Device status analysis report for a specific scanner. This report, generated by the digital twin module, includes two proactive management functions:

[0038] Feedforward pre-calibration: Before processing new scan data, a pre-calibration field is generated using a digital twin model to actively compensate for known systematic artifacts in the equipment.

[0039] Predictive maintenance alerts: By comparing the current residual with the historical health baseline, an alert is generated when the deviation exceeds a threshold, indicating that the equipment is in an abnormal state and that hardware calibration or maintenance is required.

[0040] In summary, through Figure 1 The iterative correction process shown and Figure 2The system collaborative architecture shown in this invention constructs a complete intelligent processing system. It can not only correct artifacts in a physically interpretable and anatomically reasonable way, but also transform each correction task into a check-up of the equipment's health status, realizing proactive and intelligent management of the entire MRI imaging chain.

[0041] This invention is not limited to the above-described optional embodiments. Anyone can derive other various forms of products under the guidance of this invention. However, regardless of any changes made in their shape or structure, any technical solution that falls within the scope of the claims of this invention shall be protected by this invention.

Claims

1. A method for processing brain magnetic resonance imaging (MRI) images, characterized in that, include: In a hybrid k-space and image space domain, an iterative process is performed on a target image with artifacts and its corresponding original k-space data. In each iteration, an image processing network is used to predict an artifact field, and a differentiable magnetic resonance physics simulation module is used to apply the artifact field to a reference k-space to generate a simulated k-space. The image processing network is then optimized by minimizing the physics simulation error between the simulated k-space and the original k-space data. When predicting the artifact field, a graph neural network is used to model the anatomical structure of the brain. The nodes of the graph neural network correspond to preset brain anatomical regions, and anatomical structure consistency constraints are applied by message passing between the nodes. Extract a residual artifact feature from the final result of the iterative processing, and use the feature to update a scanner digital twin model associated with a specific magnetic resonance scanner; Finally, a corrected processed image is generated.

2. The method according to claim 1, characterized in that, The iterative processing includes: predicting motion artifact fields in the image space, performing data consistency correction in the k-space, and alternating between the two until convergence.

3. The method according to claim 1, characterized in that, The differentiable magnetic resonance physics simulation module can simulate the effects of at least one or more physical processes selected from the following on k-space data encoding: patient physiological motion, main magnetic field inhomogeneity, and radio frequency field inhomogeneity.

4. The method according to claim 1, characterized in that, The graph neural network learns the physical connections and motion constraints between anatomical regions of the brain, providing structural priors for the artifact field predicted by the image processing network, thereby ensuring the anatomical rationality of the corrected image.

5. The method according to claim 1, characterized in that, The residual artifact features include systematic artifact features that cannot be explained by the motion model and are related to the scanner hardware state.

6. The method according to claim 1 or 5, characterized in that, The scanner digital twin model is a longitudinal model used to track and learn the inherent artifact patterns of a specific scanner over time.

7. The method according to claim 6, characterized in that, The method further includes generating a feedforward pre-calibration field using the scanner's digital twin model before processing new scan data, in order to actively compensate for the scanner's inherent systematic artifacts.

8. The method according to claim 6, characterized in that, The method further includes: comparing the currently extracted residual artifact features with the health baseline stored in the scanner's digital twin model, and generating an alarm for abnormal device status when the deviation exceeds a preset threshold.

9. A system for processing brain magnetic resonance imaging (MRI) images, characterized in that, The system includes a processor configured to perform the method of any one of claims 1-8, the system comprising: The hybrid domain iterative processing module is used to alternately optimize the artifact correction process between k-space and image space; The graph neural network guidance module is used to introduce topological constraints of brain anatomy during processing; The scanner digital twin module is used to longitudinally track and model the hardware status characteristics of a specific scanner, and based on this, perform feedforward pre-calibration or generate device status alarms.

10. The system according to claim 9, characterized in that, The hybrid domain iterative processing module, graph neural network guidance module, and scanner digital twin module work together to achieve physical interpretability correction of artifacts, ensure the anatomical rationality of the output image, and proactive quality control of the scanning equipment.

Images