Magnetic resonance imaging methods, systems, and computer program products

Abstract

The present invention provides a technique for removing artifacts such as RF interference and / or noise from magnetic resonance data. The technique includes: acquiring (302) input magnetic resonance data using at least one radio frequency coil (526) of a magnetic resonance imaging system (500); and generating a magnetic resonance image based on the input magnetic resonance data, at least in part, by using a neural network model (130) for suppressing (304, 308) at least one artifact in the input magnetic resonance data.

Description

[0001] Cross-references to related applications

[0002] Pursuant to 35 § 119(e) of the United States Code, this application claims priority to U.S. Provisional Application Serial No. 62 / 764,742, filed August 15, 2018, entitled “MAGNETIC RESONANCE IMAGE DENOISING USING K-SPACE DEEP LEARNING MODEL”, and U.S. Provisional Application Serial No. 62 / 820,119, filed March 18, 2019, entitled “END-TO-END LEARNABLE MR IMAGE RECONSTRUCTION”, each of which is incorporated herein by reference in its entirety. Background Technology

[0003] Magnetic resonance imaging (MRI) provides an important imaging modality for many applications and is widely used in clinical and research settings to produce images of the interior of the human body. MRI is based on the detection of magnetic resonance (MR) signals, which are electromagnetic waves emitted by atoms in response to state changes induced by an applied electromagnetic field. For example, nuclear magnetic resonance (NMR) techniques involve detecting MR signals emitted from the nuclei of excited atoms when their nuclear spins realign or relax in the object being imaged (e.g., atoms in human tissue). The detected MR signals can be processed to produce images, enabling the investigation of internal structures and / or biological processes within the body for diagnostic, therapeutic, and / or research purposes in the context of medical applications.

[0004] MRI offers an attractive imaging modality for bioimaging due to its ability to produce non-invasive images with relatively high resolution and contrast without the safety issues of other modalities (e.g., no exposure of the subject to ionizing radiation (such as X-rays) or the introduction of radioactive materials into the body). Furthermore, MRI is particularly well-suited for providing soft tissue contrast, which can be used to image subjects that cannot be satisfactorily imaged using other modalities. In addition, MRI techniques can capture information related to structure and / or biological processes that cannot be obtained through other modalities. However, conventional MRI techniques have several drawbacks, which for a given imaging application may include the relatively high cost of equipment, limited availability (e.g., the difficulty and expense of obtaining access to clinical MRI scanners), and the length of the image acquisition process.

[0005] To improve image quality, the trend in clinical and research MRI has been to increase the field strength of MRI scanners to improve scan time, image resolution, and image contrast, thereby driving up the cost of MRI imaging. The vast majority of installed MRI scanners operate at at least 1.5 or 3 Tesla (T), where 1.5 or 3 Tesla refers to the field strength of the scanner's main magnetic field B0. A rough cost estimate for a clinical MRI scanner is approximately one million US dollars per Tesla, which does not even include the substantial operational, servicing, and maintenance costs involved in operating such a scanner. Furthermore, traditional high-field MRI systems typically require large superconducting magnets and associated electronics to generate a strong, homogeneous static magnetic field (B0) for imaging the subject (e.g., a patient). Superconducting magnets further require cryogenic equipment to maintain the conductor in a superconducting state. Such systems are quite large, with typical MRI setups comprising multiple rooms (including specially shielded rooms to isolate the magnetic components of the MRI system) for magnetic components, electronics, thermal management systems, and control console areas. The size and cost of MRI systems generally limit their use to facilities such as hospitals and academic research centers with sufficient space and resources to purchase and maintain them. The high cost and significant space requirements of high-field MRI systems limit the availability of MRI scanners. Consequently, clinical situations often arise where MRI scanning would be beneficial, but is impractical or impossible due to the aforementioned limitations. Summary of the Invention

[0006] Some embodiments relate to a method comprising: obtaining input magnetic resonance data, i.e., input MR data, using at least one radio frequency coil, i.e., at least one RF coil, of a magnetic resonance imaging system, i.e., an MRI system; and generating an MR image based on the input MR data, at least in part, using a neural network model for suppressing at least one artifact in the input MR data.

[0007] Some embodiments relate to a system comprising: at least one computer hardware processor; and at least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by the at least one computer hardware processor, cause the at least one computer hardware processor to: acquire input magnetic resonance data, i.e., input MR data, using at least one radio frequency coil, i.e., at least one RF coil, of a magnetic resonance imaging system, i.e., an MRI system; and generate an MR image based on the input MR data, at least in part, using a neural network model for suppressing at least one artifact in the input MR data.

[0008] Some embodiments relate to at least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by at least one computer hardware processor, cause the at least one computer hardware processor to: acquire input magnetic resonance data, i.e., input MR data, using at least one radio frequency coil, i.e., at least one RF coil, of a magnetic resonance imaging system, i.e., an MRI system; and generate an MR image based on the input MR data, at least in part, using a neural network model for suppressing at least one artifact in the input MR data.

[0009] Some embodiments relate to a magnetic resonance imaging system, i.e., an MRI system, comprising: a magnetic system including: a B0 magnet configured to provide a B0 field for the MRI system; a gradient coil configured to provide a gradient field for the MRI system; and at least one RF coil configured to detect magnetic resonance signals, i.e., MR signals; a controller configured to: control the magnetic system to acquire input MR data using the at least one RF coil; and generate an MR image based on the input MR data, at least in part, using a neural network model for suppressing at least one artifact in the input MR data.

[0010] The above is a non-limiting summary of the invention as defined by the appended claims. Attached Figure Description

[0011] Various aspects and embodiments of the disclosed technology will be described with reference to the following accompanying drawings. It should be understood that the drawings are not necessarily drawn to scale.

[0012] Figure 1A An exemplary processing pipeline is shown, according to some embodiments of the techniques described herein, for generating MR images from input MR data using a neural network model designed to suppress one or more artifacts in the input MR data.

[0013] Figure 1B The architecture of an exemplary convolutional neural network block with a "U"-shaped structure and an average pooling layer is shown according to some embodiments of the techniques described herein. This block may be part of a neural network model used to suppress artifacts in input MR data.

[0014] Figure 1C Some embodiments of the technology described herein are shown. Figure 1B A specific example of the architecture of the exemplary convolutional neural network block is shown.

[0015] Figure 1DThe architecture of an exemplary convolutional neural network block with a "U"-shaped structure and a spectral depooling layer is shown according to some embodiments of the techniques described herein. This block may be part of a neural network model used to suppress artifacts in input MR data.

[0016] Figure 1E An exemplary architecture of a spectrum depooling layer is shown, representing some embodiments of the techniques described herein.

[0017] Figure 2A The architecture of an exemplary neural network with a spectral depooling layer for suppressing RF interference in input MR data is shown according to some embodiments of the techniques described herein.

[0018] Figure 2B Some embodiments of the technology described herein are shown. Figure 2A The exemplary neural network shown is used to suppress RF interference in input MR data.

[0019] Figure 3 This is a flowchart of an exemplary process 300 for using a neural network model to suppress one or more artifacts in input MR data, based on some embodiments of the techniques described herein.

[0020] Figure 4A The present invention illustrates some embodiments of the techniques described herein for generating training data for training a neural network model, wherein the neural network model is used to suppress one or more artifacts in MR data.

[0021] Figure 4B The following are exemplary examples of generating training data for training a neural network model according to some embodiments of the techniques described herein, wherein the neural network model is used to suppress one or more artifacts in MR data.

[0022] Figure 5 This is a schematic diagram of a low-field MRI system according to some embodiments of the techniques described herein.

[0023] Figure 6 and Figure 7 A biplane permanent magnet structure of B0 magnet is shown, according to some embodiments of the technology described herein.

[0024] Figure 8A and Figure 8B A view of a portable MRI system according to some embodiments of the technology described herein is shown.

[0025] Figure 9A A portable MRI system for performing head scans is shown, according to some embodiments of the technology described herein.

[0026] Figure 9B A portable MRI system for performing knee scans is shown, according to some embodiments of the technology described herein.

[0027] Figure 10 A diagram is shown of an exemplary computer system that can implement the embodiments described herein. Detailed Implementation

[0028] As mentioned above, conventional clinical MRI systems require a specially shielded room to operate correctly, which is one of the many reasons for the cost, lack of availability, and inaccessibility of currently available clinical MRI systems. In addition to protecting personnel and equipment from the magnetic fields generated by the MRI system, the shielded room also prevents artifacts such as RF interference from various external electronic devices (e.g., other medical devices) from affecting the operation of the MRI system and the quality of the resulting images. The inventors have realized that, in order to operate outside a specially shielded room, and more particularly to allow MRI systems that are generally portable, vehicle-mounted, or otherwise transportable, MRI systems must be able to operate in relatively uncontrolled electromagnetic environments (e.g., in unshielded or partially shielded rooms) and must be able to account for and / or compensate for the presence of interference, noise, and / or other artifacts typically present in such environments.

[0029] The inventors have developed a deep learning technique for reducing or eliminating the impact of environmental artifacts, such as RF interference and noise, on the operation of MRI systems and the resulting image quality. The deep learning technique developed by the inventors allows MRI systems to be operated outside of specially shielded rooms, thus facilitating both portable / transportable MRI and fixed MRI installations in rooms that do not require special shielding. Furthermore, while the technique developed by the inventors and described herein allows for the operation of MRI systems outside of specially shielded rooms, since aspects of the technique described herein are not limited thereto, it can also be used to mitigate the effects of interference, noise, and / or other artifacts on the operation of MRI systems in shielded environments (e.g., less expensive, loosely or temporarily shielded environments), and therefore can be used in conjunction with areas already equipped with limited shielding.

[0030] For example, deep learning techniques developed by the inventors and described herein can be used to facilitate the deployment of MRI systems (e.g., generally mobile, portable, or vehicle-mounted systems) in various settings such as emergency rooms, operating rooms, intensive care units, offices, and / or clinics. These settings are particularly susceptible to the presence of artifacts such as RF interference and noise, which many conventional MRI systems are largely unaffected by because they are installed in dedicated rooms with extensive shielding. However, conventional MRI systems are simply unavailable in these settings due to cost, lack of portability, size, and shielding requirements, despite the clear need for MR imaging. The techniques developed by the inventors are particularly valuable for facilitating the deployment of MRI systems in these settings.

[0031] The deep learning techniques developed by the inventors can be used to suppress (e.g., reduce and / or eliminate) artifacts in MR data acquired by any suitable type of MR scanner. For example, the techniques developed by the inventors can be used to reduce and / or eliminate artifacts in MR data collected by “low-field” MR systems, which operate at a lower field strength than the “high-field” MR systems that dominate the MRI system market (particularly for medical or clinical MRI applications). The lower magnetic field strength of low-field MRI systems makes them particularly susceptible to the presence of RF interference, noise, and / or other artifacts, which can adversely affect the performance of such systems. For example, the deep learning techniques developed by the inventors can be used to reduce and / or eliminate artifacts in MR data acquired by any suitable type of MR scanner described herein and / or in U.S. Patent No. 10,222,434 entitled “Portable Magnetic Resonance Imaging Methods and Apparatus” (a system that has matured since U.S. Patent Application Serial No. 15 / 879,254, filed January 24, 2018), which is incorporated herein by reference in its entirety. It should be understood that the techniques described herein are not limited to low-field MRI systems or any particular type of MRI system, and can be used in high-field and / or any other suitable type of MRI system. It should also be understood that, in addition to deep learning techniques, other machine learning techniques may be employed in some embodiments, as the aspects of the techniques described herein are not limited thereto.

[0032] Many MRI systems (including some described herein) utilize electromagnetic shielding to reduce the impact of artifacts on the operation of the MRI system and the quality of the resulting images. The installation and maintenance costs of such electromagnetic shielding can be high, and any errors or defects in the electromagnetic shielding can degrade the quality of MR images produced by the MRI system. The techniques developed by the inventors and described herein can reduce the amount of electromagnetic shielding required for MRI systems, thereby reducing their cost, and can compensate for any errors or defects in the electromagnetic shielding and / or its installation.

[0033] As used herein, “high field” generally refers to MRI systems currently used in clinical settings, and more specifically, to MRI systems operating with a main magnetic field of 1.5 T or higher (i.e., B0 field), although clinical systems operating between 0.5 T and 1.5 T are often also characterized as “high field.” Field strengths between approximately 0.2 T and 0.5 T are characterized as “mid field,” and as the field strength increases further within the high field region, the range between 0.5 T and 1 T is also characterized as mid field. In contrast, “low field” generally refers to MRI systems operating with a B0 field less than or equal to approximately 0.2 T, but systems with a B0 field between 0.2 T and approximately 0.3 T are sometimes also characterized as low field due to the increased field strength at the higher end of the high field region. Within the low field region, low-field MRI systems operating with a B0 field less than 0.1 T are referred to herein as “very low field,” and low-field MRI systems operating with a B0 field less than 10 mT are referred to herein as “ultra-low field.”

[0034] In some embodiments, the deep learning techniques developed by the inventors involve using a neural network model to process input MR spatial frequency data to suppress one or more artifacts in the input MR data (e.g., reduce or eliminate their presence and / or influence).

[0035] In some embodiments, input MR data can be processed in multiple stages, where one or more stages may involve suppressing artifacts in the input MR data. For example, in some embodiments, different processing stages may be used to suppress different types of artifacts (e.g., RF interference from one or more devices outside the MRI system may be suppressed in one stage, and noise generated by the MR receiver chain may be suppressed in another stage). As another example, in some embodiments, multiple processing stages may be used to suppress the same type of artifact (e.g., multiple stages may be used to suppress RF interference generated by one or more devices outside the MRI system).

[0036] Circuitry involving the processing of signals recorded by at least one RF coil may be referred to as an "MR receiver chain". An MR receiver chain can include various types of circuitry, such as analog circuitry (e.g., one or more amplifiers, decoupling circuitry, RF transmit / receive switching circuitry, etc.), digital circuitry (e.g., a processor), and / or any suitable combination thereof. Examples of MR receiver chain circuitry are described in U.S. Patent Application Serial No. 16 / 418,414, entitled "Radio-Frequency Coil Signal Chain For a Low-Field MRI System," filed May 21, 2019, and is incorporated herein by reference in its entirety.

[0037] In some embodiments, a neural network model for suppressing one or more artifacts in input MR data may include multiple parts, and each of these parts may be applied during a corresponding processing stage. For example, in some embodiments, the neural network model may include two parts—a first part and a second part—where the first part is configured to suppress RF interference generated by devices located outside the MRI system (also referred to herein as “external RF interference”) and / or RF interference generated by one or more components within the MRI system located outside its imaging area (also referred to herein as “internal RF interference”), and the second part is configured to suppress noise generated by circuitry in the MR receiver chain and / or noise generated by the subject (or object) being imaged. In this example, the input MR data may be processed in multiple stages, where one stage involves applying the first part of the neural network to suppress (external and / or internal) RF interference, and another stage involves applying the second part of the neural network to suppress noise generated by the subject / object being imaged. Reference is made below. Figure 1A The processing pipeline shown is used to illustrate another example, which involves a neural network with three parts applied in a (non-continuous) sequence across three processing stages.

[0038] It should be understood that while input MR data can be processed in multiple levels, not every level involves artifact suppression, because one or more processing levels can be used to perform functions other than artifact suppression. For example, one of these levels (e.g., Figure 1A The stage 108 shown can involve a reconstruction step that generates an image from the input MR data using any suitable reconstruction technique.

[0039] In some embodiments, the input MR data can be processed using one or more levels not in the image domain (e.g., before image reconstruction) and one or more levels in the image domain (e.g., after image reconstruction). For example, in some embodiments, a portion of a neural network model can be applied in the sensor domain or the spatial frequency domain to (e.g., in...) Figure 1A During stage 106 (as shown), RF interference is suppressed, and different parts of the neural network model can be applied in the image domain to (e.g., in...) Figure 1A During stage 112 (as shown), RF interference and / or noise generated by the MR receiver chain or the subject (or object) being imaged are suppressed. However, it is not necessary to apply artifact suppression processing both before and after image reconstruction (e.g., in the sensor domain or spatial frequency domain and in the image domain). For example, in some embodiments, artifact suppression may be performed only before image reconstruction or only in the image domain.

[0040] Furthermore, in some embodiments, artifact suppression can be performed in one or more domains other than the sensor domain, spatial frequency domain, and image domain. In such embodiments, data can be transformed to another domain via a suitable reversible transform (e.g., 1D or 2D or 3D wavelet transform, 1D or 2D or 3D Fourier transform, 1D or 2D or 3D short-time Fourier transform, and / or any other suitable time-frequency and / or time-scale transform), wherein artifact suppression can be performed before applying a suitable inverse transform to the processed data.

[0041] Data in the "sensor domain" can include raw sensor measurements obtained by the MRI system. Sensor domain data can include measurements acquired row-by-row for a set of coordinates specified for a sampling pattern. Measurement rows can be referred to as "readout" rows. Each measurement can be a spatial frequency. Therefore, sensor domain data can include multiple readout rows. For example, if p readout rows are measured and each readout row includes m samples, the sensor domain data can be organized as an m×p matrix. Knowing the k-space coordinates associated with each sample in the m×p samples, the sensor domain data can be reorganized into corresponding k-space data, which can then be considered spatial frequency domain data. Image domain data can be obtained by applying an inverse Fourier transform to the k-space data (e.g., an inverse fast Fourier transform if the samples fall on a grid).

[0042] Therefore, some embodiments provide a deep learning artifact suppression technique involving: (1) accessing MR data obtained using at least one radio frequency (RF) coil of an MRI system; and (2) generating an MR image based on the input MR data, at least in part, using a neural network model (e.g., a model comprising one or more convolutional layers) for suppressing at least one artifact in the input MR data. In some embodiments, the first action of the deep learning artifact suppression technique may involve using at least one RF coil to obtain the input MR data (instead of simply accessing data previously obtained using at least one RF coil).

[0043] In some embodiments, at least one artifact includes RF interference, and generating an MR image includes using a neural network model to suppress the RF interference. In some embodiments, the RF interference may include external RF interference generated by a device located outside the MRI system. This device may be located in the same room as the MRI system and / or sufficiently close to the MRI system (e.g., within a threshold distance of the MRI system) such that the electromagnetic waves generated by the device can be detected by the MRI system. This device may be a medical device, such as a cardiac monitor, pulse oximeter, infusion pump, or other electrical equipment (e.g., transformer, motor) located in the same room and / or sufficiently close to the MRI system.

[0044] In some embodiments, RF interference may include internal RF interference generated by one or more components of the MRI system located outside the imaging region of the MRI system. For example, internal RF interference may be generated by one or more magnetic components (e.g., gradient coils, magnets, etc.) and / or one or more electrical components (e.g., one or more gradient power amplifiers, one or more power distribution units, one or more power supplies, one or more switches, one or more thermal management components, etc.) of the MRI system. However, it should be understood that any other component of the MRI system located outside its imaging region besides the components listed above may also generate internal RF interference, as the aspects of the technology described herein are not limited thereto.

[0045] In some embodiments, at least one artifact may include noise generated by the MR receiver chain and / or noise generated by the subject or object being imaged. In some embodiments, the MRI system may include at least one RF coil configured to detect MR signals in the imaging region of the MRI system.

[0046] The inventors have realized that certain types of artifacts can be suppressed more effectively in domains other than the image domain (e.g., the sensor domain or the spatial frequency domain (sometimes referred to as "k-space")). In particular, the inventors have realized that external RF interference can be effectively suppressed in the sensor domain or the spatial frequency domain because, in these domains, external RF interference sometimes manifests as a set of complex exponential components superimposed on the detected MR signal. The inventors have realized that suppression of this type of external RF interference can be performed more effectively in the sensor domain or the spatial frequency domain compared to the image domain.

[0047] Therefore, in some embodiments, the neural network model for artifact suppression includes a first neural network portion configured to process data in the sensor domain or spatial frequency domain, and wherein using the neural network model to suppress at least one artifact in the input MR spatial frequency domain data includes utilizing the first neural network portion to process the sensor domain data or spatial frequency domain data obtained from the input MR data. An example of the first neural network portion is... Figure 1D The diagram shows the neural network section 150, which will be described in more detail here.

[0048] In some embodiments, the first neural network portion includes a "U"-shaped structure, wherein convolutional layers are applied along a "downsampling path" to successive lower-resolution versions of the data, and then along an "upsampling path" to successive higher-resolution versions of the data. In some embodiments, the resolution of the data may be reduced using one or more pooling layers (e.g., along the downsampling path) and increased using one or more corresponding depooling layers (e.g., along the upsampling path).

[0049] As described above, the first neural network portion can be configured to process data in the sensor domain or the spatial frequency domain. In some embodiments, the first neural network portion may include and be configured to process data in the sensor domain or the spatial frequency domain using a spectral depooling layer developed by the inventors. In some embodiments, applying the spectral depooling layer includes applying a pointwise multiplication layer for combining a first feature having a first resolution provided via a skip connection with a second feature having a second resolution lower than the first resolution. In some embodiments, applying the spectral depooling layer includes zero-padding the second feature before combining the first feature with the second feature using the pointwise multiplication layer. Figure 1E An exemplary example of a spectral pooling layer is shown. In some embodiments, where the first neural network portion includes a spectral depooling layer, the first neural network portion also includes a corresponding spectral pooling layer. Additionally, the first neural network portion may include multiple convolutional layers and at least one skip connection.

[0050] As described above, a neural network model may include multiple parts that use artifact suppression at different levels of processing MR data. In some embodiments, the neural network model includes: (1) a first neural network part configured to suppress RF interference (e.g., external and / or internal RF interference); and (2) a second neural network part configured to suppress noise (e.g., noise generated by the MR receiver chain and / or the subject (or object) being imaged). Each of these parts may include one or more convolutional layers, one or more pooling layers, and / or one or more skip connections, as aspects of the technique described herein are not limited thereto. For example, in some embodiments, the neural network may include a first part and a second part, wherein the first part is configured as... Figure 1A The exemplary example shows that the process pipeline 100 is part of stage 106 to suppress RF interference, and the second part is configured to suppress noise as part of stage 108 or stage 112 of the same process pipeline.

[0051] In some embodiments, the neural network model further includes a third neural network portion configured to suppress noise from image domain data obtained using input MR spatial frequency data. For example, the neural network may include, as... Figure 1A The exemplary example shows a third part of a stage 112 of the processing pipeline 100.

[0052] The inventors have also developed a technique for training a neural network model for artifact suppression in MR data. The technique involves generating training data by: (1) synthesizing and / or measuring RF artifact measurements; (2) synthesizing and / or measuring MR measurements; and (3) combining the obtained RF artifact measurements and MR measurements to obtain artifact-correputed MR data. Furthermore, the artifact-correputed MR data (along with corresponding individual artifact components and MR data components) can be used to train one or more neural network models for suppressing artifacts in MR data.

[0053] Therefore, in some embodiments, the technique for training a neural network for suppressing artifacts in MR data includes: (e.g., in the absence of MR signals in the imaging region of the MRI system) obtaining RF artifact measurements using at least one RF coil of the MRI system during a first time period, wherein the RF artifact measurements include measurements of RF interference and / or noise; obtaining MR measurements of a subject in the imaging region of the MRI system during a second time period different from the first time period; generating artifact-damaged MR data by combining the RF artifact measurements with the MR measurements of the subject; and training a neural network model using the artifact-damaged MR data.

[0054] In some embodiments, a technique for training a neural network for suppressing artifacts in MR data includes: synthesizing RF artifact measurements, wherein the RF artifact measurements include synthesized measurements of RF interference and / or noise; obtaining MR measurements of a subject in an imaging region of an MRI system; generating artifact-damaged MR data by combining the synthesized RF artifact measurements with the MR measurements of the subject; and training a neural network model using the artifact-damaged MR data.

[0055] In some embodiments, a technique for training a neural network for suppressing artifacts in MR data includes: (e.g., in the absence of MR signals in the imaging region of an MRI system) using at least one RF coil of an MRI system to obtain RF artifact measurements, wherein the RF artifact measurements include measurements of RF interference and / or noise; synthesizing MR measurements of the MRI system of the subject; generating artifact-damaged MR data by combining the obtained RF artifact measurements with the synthesized MR measurements of the subject; and training a neural network model using the artifact-damaged MR data.

[0056] In some embodiments, the measured RF artifact measurements and / or measured MR measurements may be taken using an MRI system to train a neural network for suppressing artifacts in subsequent MR data obtained by the same MRI system. Furthermore, the RF artifact measurements and / or MR measurements may be obtained using the MRI system when it is placed in the environment in which it will subsequently be used for imaging. In this way, the training data will accurately reflect the types of interference that may be present during subsequent operation of the MRI system.

[0057] For example, in some embodiments, an MRI system can be calibrated for subsequent artifact suppression by: (1) placing the MRI system in the environment in which it will be used for imaging (e.g., an emergency room, office, operating room, patient room, intensive care unit, etc.); (2) obtaining one or more measurements of RF artifacts (e.g., external RF interference generated by medical devices in a medical facility where the MRI system is placed) and / or MR data in that environment; (3) using these measurements to generate training data for training a neural network for artifact suppression; (4) using this training data (e.g., learning at least some parameters of the neural network by using only the training data obtained in the environment or by updating / adapting the neural network parameters to the training data obtained in the environment); and (5) using the trained neural network to suppress artifacts in MR data subsequently collected by the MRI system in the environment. In this way, the neural network can learn to suppress and / or can be adapted to precisely suppress the types of interference present in the environment during imaging.

[0058] The following describes in more detail various concepts and embodiments thereof related to methods and apparatus for suppressing artifacts in MR data using neural networks. It should be understood that the aspects described herein can be implemented in any of a variety of ways. Examples of specific implementations are provided herein for illustrative purposes only. Furthermore, the aspects described in the following embodiments can be used alone or in any combination, and are not limited to the combinations explicitly described herein.

[0059] Figure 1A An exemplary processing pipeline 100, illustrating some embodiments of the techniques described herein, is used to generate MR images from input MR data using a neural network model designed to suppress one or more artifacts in the input MR data.

[0060] like Figure 1A As shown, the data processing pipeline 100 includes multiple stages for processing input MR data 102, including: a preprocessing stage 104, an RF interference removal stage 106, a noise removal stage 108, a reconstruction stage 110, and a noise removal stage 112. These processing stages are applied to the input MR spatial frequency data 102 to produce an output MR image 114.

[0061] exist Figure 1A In the example, levels three (i.e., levels 106, 108, and 112) are shaded, thus indicating that these levels are subject to artifact suppression. Figure 1AIn the example, stages 106 and 108 are processed in the spatial frequency domain, while stage 112 performs artifact suppression processing in the image domain. As described above, in some embodiments, any one or more of these stages can be performed in any other suitable domain. For example, in some embodiments, one or both of stages 106 and 108 may perform artifact suppression in the sensor domain instead of the spatial frequency domain. In such an embodiment, a preprocessing stage 104 that transforms data from the sensor domain to the spatial frequency domain may be placed between stages 108 and 110, instead of as... Figure 1A As shown, it is placed before level 106.

[0062] exist Figure 1A In the example, each stage in stages 106, 108, and 112 uses a corresponding neural network portion to suppress artifacts in the data provided as input to that stage. In this example, the entire neural network model comprises three parts: a first neural network portion, a second neural network portion, and a third neural network portion. The first neural network portion is configured to suppress RF interference in the MR data as part of the processing performed during stage 106, the second neural network portion is configured to suppress noise in the MR data as part of the processing performed during stage 108, and the third neural network portion is configured to suppress noise from the MR data as part of the processing performed during stage 112. In some embodiments, the three parts of the neural network model can be trained jointly (e.g., the output of one neural network portion can influence the input of another neural network portion).

[0063] While this example involves a data processing pipeline with three artifact suppression stages, this is not a limitation of the techniques described herein. In some embodiments, the data processing pipeline may be used with any one or two of stages 106, 108, and 112, rather than all three. Furthermore, it is possible to... Figure 1A Any one or two or all of the stages shown in the exemplary data processing pipeline 100 and / or replacing Figure 1A The exemplary data processing pipeline 100 may use any one or two or all of the stages shown in the example data processing pipeline 100 to use one or more artifact suppression stages.

[0064] The data processing pipeline 100 can be applied to any suitable type of input sensor data 102. Data 102 can be collected by one or more RF coils of an MRI system. Data 102 can be collected using Cartesian sampling trajectories or any suitable type of non-Cartesian sampling trajectory (e.g., radial, spiral, rosette, variable density, Lissajou, etc.). Data 102 can be fully sampled data (data collected by sampling the spatial frequency space such that the corresponding Nyquist criterion is not violated). Data 102 can be undersampled data (data containing fewer points than required by the spatial Nyquist criterion). In some embodiments, data 102 may exhibit artifacts due to the presence of external RF interference, internal RF interference, and / or noise generated by the MR receiver chain and / or the subject (or object) being imaged.

[0065] Initially, as part of preprocessing stage 104, one or more preprocessing steps may be applied to the input MR data 102. For example, in some embodiments, the input MR data 102 may be sensor domain data, and the preprocessing stage may transform the sensor domain data (e.g., by performing a 1D Fourier transform along the readout line). As another example, in some embodiments, preprocessing stage 104 may involve removing some data from the input data 102. For example, some data in 102 may be removed when it is determined that the data is corrupt (e.g., due to sensor readings indicating unreliable data).

[0066] Next, as part of the data processing pipeline 100 in stage 106, the first part of the neural network model is applied to suppress (external and / or internal) RF interference in the data provided as input to stage 106.

[0067] In some embodiments, the neural network applied during level 106 may have a “U”-shaped structure, wherein convolutional layers are first applied (along the downsampling path) to a sequence of consecutive lower-resolution versions of the data, and then (along the upsampling path) to a sequence of consecutive higher-resolution versions of the data.

[0068] For example, the first part of a neural network model can have Figure 1B The architecture shown is 130. (As shown) Figure 1B As shown, in the downsampling path, convolutional layers 132a and 132b are applied to input 131. Then, an average pooling layer 133 is applied to the output of convolutional layer 132b, and convolutional layers 134a and 134b are applied to the low-resolution data produced by the average pooling layer 133. Next, another average pooling layer 135 is applied to the output of convolutional layer 134b, and convolutional layers 136a, 136b, and 136c are applied to the output of average pooling layer 135.

[0069] Next, in the upsampling path, the output of convolutional layer 136c is processed by average depooling layer 137. The output of average depooling layer 137 is processed by convolutional layers 138a and 138b. The output of convolutional layer 138b is processed by average depooling layer 139, and the output of average depooling layer 139 is processed by convolutional layers 140a-c to generate output 145.

[0070] Architecture 130 also includes skip connections 141 and 142, which indicate that the input to the average depooling layer comprises the output of the immediately preceding convolutional layer and the higher-resolution output of another (not immediately preceding) convolutional layer. For example, the input to the average depooling layer 137 is the output of convolutional layers 134b (indicated by skip connection 142) and 136c. The output of convolutional layer 134b has a higher resolution than the output of layer 136c. As another example, the input to the average depooling layer 139 is the output of convolutional layers 132b (indicated by skip connection 142) and 138b. The output of convolutional layer 132b has a higher resolution than the output of layer 138b. In this way, high-frequency information lost by applying pooling layers along the downsampling path is reintroduced (without loss) as input to the depooling layer along the upsampling path.

[0071] Despite Figure 1B Not explicitly shown, however, nonlinear layers (e.g., rectified linear units or ReLU, sigmoid, etc.) can be applied after one or more layers shown in architecture 130. For example, nonlinear layers can be applied after... Figure 1B It is applied after one or more of the convolutional layers shown. Additionally, batch normalization can be applied at one or more points along architecture 130 (e.g., at the input layer).

[0072] Figure 1C Some embodiments of the technology described herein are shown. Figure 1B A specific example of the architecture of the exemplary convolutional neural network block is shown. Figure 1CAs shown, all convolutional layers apply 3×3 kernels. In the downsampling path, the input at each level is processed by repeatedly applying two (or three at the bottom level) convolutions with 3×3 kernels, followed by a non-linear, average 2×2 pooling operation with a stride of 2 for downsampling. In each downsampling step, the number of feature channels increases from 64 to 128 to 256. At the bottom level, the number of feature channels also increases from 256 to 512. In the upsampling path, the feature maps are repeatedly upsampled (using 3×3 kernels) by an average depooling step that halves the number of feature channels (e.g., from 256 to 128 to 64), and the data is processed by concatenating the corresponding feature maps from the downsampling path and one or more convolutional layers, with a non-linearity applied after each convolutional layer. The final convolutional layer 140c reduces the number of feature maps to 2.

[0073] As described above, the inventors have developed a novel type of depooling layer (referred to herein as a "spectral depooling layer") for use in neural network models applicable in the sensor domain or spatial frequency domain to, for example, suppress artifacts in input MR data. Figure 1D Architecture 150 is shown, featuring a convolutional neural network block with a "U"-shaped structure and a spectral depooling layer. Architecture 150 is compared to... Figure 1B The architecture shown is the same as 130, but the average depooling layer is replaced by the spectral depooling layer.

[0074] like Figure 1D As shown, in the downsampling path, convolutional layers 152a and 152b are applied to input 151. Then, a spectral pooling layer 153 is applied to the output of convolutional layer 152b, and convolutional layers 154a and 154b are applied to the low-resolution data generated by the spectral pooling layer 153. Another spectral pooling step 155 is applied to the output of convolutional layer 154b, and convolutional layers 136a, 136b, and 136c are applied to the output of spectral pooling layer 155. In the upsampling path, the output of convolutional layer 156c is processed by a spectral depooling layer 157, and the output of spectral depooling layer 157 is further processed by convolutional layers 158a and 158b. The output of convolutional layer 158b is processed by a spectral depooling layer 159, and the output of spectral depooling layer 159 is processed by convolutional layers 160a-c to produce output 165.

[0075] In some embodiments, spectral pooling layers can be implemented by pruning the data. This is similar to discarding only the higher spatial frequency content from the data, and is implemented very efficiently because the data is already in the spatial frequency domain, eliminating the need to apply a discrete Fourier transform to implement the spectral pooling layer. Various aspects of spectral pooling are described in “Spectral representations for convolutional neural networks” by Rippel, O., Snoek, J., and Adams, RP (Advances in Neural Information Processing Systems, pp. 2449–2457, 2015), which is incorporated herein by reference in its entirety.

[0076] like Figure 1D As shown, architecture 150 also includes skip connections 161 and 162. Therefore, the input to the spectral depooling layer 157 is the output of convolutional layers 154b and 156c (where the output of layer 154b includes a higher frequency content compared to the output of layer 156c). The input to the spectral depooling layer 159 is the output of convolutional layers 152b and 158b (where the output of layer 152b includes a higher frequency content compared to the output of layer 158b).

[0077] In some embodiments, architecture 150 can be configured with, for example, Figure 1C The architecture 130 shown is implemented in a similar manner. For example, a 3×3 core can be used, and the number of feature channels can increase from 64 to 128 to 256 to 512 along the downsampling path, and decrease from 512 to 256 to 128 to 64 and down to 2 along the upsampling path. However, it should be understood that any other suitable implementation (e.g., the number of feature channels, core size, etc.) can be used, as the aspects of the techniques described herein are not limited thereto.

[0078] Figure 1E An exemplary architecture of a spectrum depooling layer according to some embodiments of the techniques described herein is shown. In particular, Figure 1E Show as Figure 1D The architecture shown is a portion of the spectrum depooling layer 157 of the architecture 150. (See also:) Figure 1E As shown, the output 180 of the spectral depooling layer 157 is generated by two inputs: (1) via skip connection 162 (from such...) Figure 1D (2) The high-resolution features 170 provided by the output of the convolutional layer 152b shown; and (3) from the output of the convolutional layer 152b shown. Figure 1D The convolutional layer 158b shown provides low-resolution features 174 as output. High-resolution features 170 are called this because they include a higher (spatial) frequency content compared to low-resolution features 174.

[0079] In the illustrated embodiment, the spectral depooling layer 157 combines high-resolution features 170 and low-resolution features 174 by: (1) zero-padding the low-resolution features 174 with zero-padding blocks 176; and (2) calculating a weighted combination of the zero-padding low-resolution features (weighted with weight 178) and the high-resolution features (weighted with weight 172). In some embodiments, weights 172 and 178 are learned from data rather than preset. However, in other embodiments, at least some weights may be manually set rather than learned from data.

[0080] As a specific example of using a spectral pooling layer, low-resolution feature 174 may include one or more (e.g., 128) feature channels (each comprising 64×64 complex values), and high-resolution feature 174 may include one or more (e.g., 64) feature channels (each comprising 128×128 complex values). The high-resolution 128×128 feature channels and the corresponding low-resolution 64×64 feature channels may be combined by: (1) zero-padding the 64×64 feature channels to obtain a 128×128 zero-padding value set; and (2) adding the high-resolution 128×128 feature channels (weighted by weight 172) to the 128×128 zero-padding value set (weighted by weight 178).

[0081] As described above, some neural network architectures used for artifact suppression can use average pooling (and depooling) layers or spectral pooling (and depooling) layers. In other embodiments, max pooling (and depooling) layers can be used. Still in other embodiments, pooling layers can be omitted entirely, and longer kernel strides can be used to efficiently downsample the data, while transposed convolutional layers can be used to upsample the data.

[0082] Return to Figure 1A The data processing pipeline 100 shown includes a noise removal stage 108 following the RF interference removal stage 106. As part of stage 108, a second part of a neural network model is applied to suppress noise in the data provided as input to stage 108. For example, the second part of the neural network model can be used to suppress noise generated by the MR receiver chain during the collection of input MR data. As another example, the second part of the neural network model can be used to suppress noise generated by the subject (or object) being imaged.

[0083] In some embodiments, the second part of the neural network model may have the same or similar architecture as the first part (used as part of stage 106). For example, the second part may have an architecture similar to that of the referenced part. Figures 1B-1EThe “U”-shaped architectures 130 and 150 are described. However, it should be understood that one or more other architectures can be used, such as the ResNet architecture including convolutional blocks with residual connections, as described by He K, Zhang X, Ren S, and Sun J in “Deep residual learning for image recognition” (IEEE Conference Proceedings on Computer Vision and Pattern Recognition, pp. 770–778, 2016).

[0084] like Figure 1A As shown in the example, noise removal stage 108 is applied in the spatial frequency domain. However, in other embodiments, the noise removal stage can be applied in other domains after appropriate transformation (e.g., sensor domain, logarithmic spectrum domain, time domain, spectrum domain, etc.), because the aspects of the technique described herein are not limited thereto.

[0085] In some embodiments, a second part of the neural network model can be jointly trained with a first part of the neural network model. For example, training data can be generated such that the input to the second part of the neural network model can be the output of the first part of the neural network model. As a specific example, training data corrupted by both RF interference and MR receiver chain noise can be provided as input to the first part of the neural network model, and the output (with at least some RF interference suppressed by the first part) can be provided as input to the second part of the neural network model. In other embodiments, the first and second parts can be trained independently of each other. As a specific example, the second part of the neural network model can be trained using training data corrupted by noise (e.g., MR receiver chain noise) but not by RF interference. The aspects of training the first and second neural network model parts are further described below.

[0086] like Figure 1A As shown, image reconstruction stage 110 follows noise suppression stage 108. During image reconstruction stage 110, the spatial domain frequency data output from stage 108 is transformed to the image domain to generate image domain data. Image reconstruction can be performed in any suitable manner. For example, when MR data is sampled along a Cartesian grid, the data can be transformed to the image domain using a 2D (or 3D) inverse Fourier transform (e.g., using a 2D or 3D inverse fast Fourier transform). As another example, when MR data is undersampled, the data can be transformed using a gridding operation and subsequent inverse Fourier transform, a non-uniform inverse Fourier transform, a neural network model for reconstructing image data from non-Cartesian k-space data, compressed sensing, and / or any other suitable method, as the aspects of the techniques described herein are not limited thereto.

[0087] As a concrete example, in some embodiments, when using non-Cartesian sampling trajectories, MR data can be mapped to a regular grid in the spatial frequency domain (sometimes referred to as “grinding” the data), and the gridded data can be transformed to the image domain using a 2D inverse fast Fourier transform to obtain the corresponding image. A regular grid in k-space refers to a grid of regularly spaced points in k-space such that there is a fixed distance Δ between each indexable k-space coordinate. In some embodiments, gridding can be performed by applying an interpolation matrix transformation to the data. In some embodiments, the entries of the interpolation weight matrix can be computed using optimization methods (such as those described by Fessler, JA, and Sutton BP in “Non-uniform fast Fourier transforms using min-max interpolation” (IEEE Transactions on Signal Processing, Vol. 51(2), pp. 560–574, 2013, which are incorporated herein by reference in their entirety). Aspects of image reconstruction in non-Cartesian settings are described in U.S. Patent Application Serial No. 16 / 524,598, filed July 30, 2019, entitled “Deep Learning Techniques for Magnetic Resonance Image Reconstruction,” which is incorporated herein by reference in its entirety.

[0088] like Figure 1A As shown, noise removal stage 112 follows reconstruction stage 110. As part of noise removal stage 112, a third part of the neural network model is applied to suppress noise in the image domain MR data. However, in this example, unlike the first and second neural network model parts applied in the spatial frequency domain, the third part of the neural network model is applied in the image domain. The third neural network part can have the same or similar architecture as the first or second part, and for example, can have an architecture similar to... Figures 1B-1E The architecture described (where appropriate Fourier transforms are used to perform these layers when employing spectral pooling and spectral depooling layers) is described. The third part of the neural network model can be trained jointly with or independently of the first and second neural network model parts, as the aspects of the technique described herein are not limited thereto. The aspects of training the third part are further described below.

[0089] like Figure 1A As shown, the output of the noise removal stage is MR image 114. It should be understood that... Figure 1AThe data processing pipeline 100 shown is exemplary and variations exist. As described, in some embodiments, one or more artifact suppression stages (106, 108, and 112) may be omitted. As another example, one or more additional processing stages may be added to the pipeline for artifact suppression or any other function. As yet another example described above, stages 106 and 108 may be applied in the sensor domain rather than the spatial frequency domain.

[0090] Next, we will discuss neural network models used for artifact suppression (such as...). Figures 1A-1E Other aspects and details of the neural network model shown (e.g., the neural network model). It should be noted that although we described the three neural network parts of a single network model above in conjunction with the neural network processing as part of levels 106, 108, and 112, we may refer to the neural network part simply as the neural network in the following text.

[0091] First, let's introduce some notation. An MRI system can have one or more RF coils configured to detect MR signals in an imaging region of the MR system. Let the number of such RF coils be N. C Indicated. For each RF coil c configured to detect MR signals in the imaging region, let s c This indicates the detected signal. The detected signal contains the following three different components: (1) the target MR signal data x of coil c. c (2) Noise n that damages the signal c (e.g., noise generated by the MR receiver chain of coil c, noise generated by the subject (or object) being imaged); and (3) external and / or internal RF interference i c Therefore, s c =x c +n c +i c Furthermore, by using N P With one receiver coil positioned outside the system, we can acquire the noise observed outside the system (which is related to S). c Related), referred to as Therefore, the observed signal can be written as:

[0092]

[0093] As described above, in some embodiments, neural network models can be used to suppress RF interference. c For example, the first part of a neural network model used as part of level 106 can be used to suppress RF interference. c To generate a specific value for coil c The neural network model used to suppress RF interference can be trained jointly or independently with any other neural network part of the data processing pipeline. In some embodiments, in order to train a model to suppress i c The model is created by creating models that include s c Training data for all components of the data makes ground truth available. This can be done in any suitable manner described here. For example, x c n c and i c Each can be synthesized using computer-based simulations and / or observed using an MRI system. For example, to generate i c Structured noise lines can be synthesized and added to s c Or, to obtain s when there is no object within the system. c As another example, an MRI system may have one or more RF coils outside the imaging area, which can be used to observe artifacts outside the imaging area (without needing to detect the MR signal), and the coils can be used to measure RF interference.

[0094] In some embodiments, the architecture of the neural network used to remove RF interference can be as described in the reference. Figures 1B-1E The described architectures are U-shaped architectures 130 and 150. Alternatively, a ResNet-type architecture with residual connections in the convolutional blocks can be used. The network input can be: (1) the signals s of each coil. c This allows the neural network to suppress the RF interference of each coil individually; (2) the signals s of all coils as individual channels c This allows the neural network to simultaneously suppress RF interference from all coils; or (3) the signals s of each coil as separate channels. c And signals that serve as additional information in other channels. (Not suppressed, but used to suppress RF interference in the signal). The output generated by the neural network corresponding to the input can be: (1) the output of each coil. Or (2) as the entirety of a separate channel (When the input is one of the latter two cases). Additionally, in some embodiments, the input to this block can be any N. avg average total s c This incorporates more information. In this case, the output will be the total denoised coil data, calculated using all averages. This can be helpful when multiple observations are made for each coil.

[0095] A neural network for suppressing RF interference can be trained using any of a variety of loss functions, and various examples of loss functions are provided here. As an example, the following loss function can be used to train a neural network for suppressing RF interference in data acquired using a single coil:

[0096]

[0097] Where W is the weighting matrix and F is the 1D Fourier transform. It represents the image gradient, and θ represents the convolutional neural network f. CNN The parameters.

[0098] In a multi-channel setup, the following loss function can be used:

[0099]

[0100] Where N coil It is the number of coils, and f CNN (s) c It is noise reduction sensor data for the coil.

[0101] As described above, in some embodiments, neural networks can be used to suppress RF interference. c (For example, as part of stage 106 of pipeline 100), and another neural network can be used to suppress noise n. c (For example, as part of stage 108 of pipeline 100). As described here, it is used to suppress n c The architecture of the neural network can be used to suppress RF interference. c The architectures are the same or similar (e.g., "U"-shaped structured networks, ResNet structured networks, and / or reference networks). Figures 1B-1E (Any architecture described).

[0102] In some embodiments, the input to the noise removal neural network may be: (1) separate s for suppressing noise from each coil c. c (2) All s as individual channels c (3) All s as individual channels, to be used to suppress noise in all coils simultaneously; c And data detected by coils outside the imaging area as additional information for noise reduction. In some embodiments, the output of the trained neural network may be: (1)x c Or (2) all x of multiple coils c .

[0103] A neural network for noise suppression can be trained using any of a variety of loss functions, and various examples of loss functions are provided here. As an example, the following loss function can be used to train a neural network for suppressing noise in data acquired using a single coil:

[0104]

[0105] In some embodiments, when training a neural network to suppress noise in data acquired using multiple coils, the following loss function may be employed:

[0106]

[0107] As described above, in some embodiments, neural networks can be used to suppress artifacts in the image domain (e.g., as part of stage 112 of pipeline 100). The architecture of this neural network, as described herein, can be the same as or similar to the architectures of other neural networks described herein (e.g., "U"-shaped structured networks, ResNet structured networks, and / or reference networks). Figures 1B-1E (Any architecture described).

[0108] Suppressing artifacts in the image domain facilitates the reduction or removal of noise generated by the acquisition system (e.g., MR receiver chains). The effects of this noise are more pronounced in low-field MRI systems, resulting in a lower signal-to-noise ratio. Conventional techniques for suppressing noise in MR images involve parametric filtering techniques such as anisotropic diffusion or nonlocal mean filtering. These parametric filtering techniques aim to remove noise in homogeneous image regions while preserving the sharpness of edges around anatomical structures. When noise levels are high (as in low-field systems), applying parametric filters often results in a smooth-looking image with loss of detail in low-contrast image regions. In contrast, using deep learning to leverage the techniques developed by the inventors to suppress artifacts (e.g., noise) in the image domain yields a sharp-looking image while preserving structure even in low-contrast regions.

[0109] The neural network architecture used to suppress artifacts in the image domain can be any architecture described herein, and can be, for example, a convolutional neural network with convolutional blocks having residual connections (as in the ResNet architecture), as referenced. Figures 1B-1E The described "U" shaped structure, or any other suitable structure.

[0110] In some embodiments, training data can be created to reflect the effect of noise on MR images. Noise can be measured or synthesized (e.g., using an MRI system). For example, a synthetic noise signal e c You can add it to image x as follows c : The noise can be derived from Gaussian e c ~N(0,σ c Take from either the Ricean distribution (assuming, for simplicity, there is no correlation between the coils).

[0111] In some embodiments, in a given dataset In this case, a neural network for suppressing artifacts in the image domain can be trained using content loss (structural similarity index (SSIM) loss or mean squared error loss) and adversarial loss given by the following formula:

[0112]

[0113] In the above expression for the loss, the generator G is a filtering network, and the discriminator D is trained to optimally distinguish between images filtered by network G and the original noise-free image (ground truth). In some embodiments, this can be achieved by adding a filter to the generator (θ). G Neural networks and discriminators (θ) D A minimax game is established between the neural networks to optimize the parameters of the generator and discriminator neural networks. The generator network can be trained to produce filtered images that are as close as possible to the ground truth, thus fooling the discriminator neural network. On the other hand, the discriminator network can be trained to classify the input image as either filtered or ground truth. As described above, using adversarial loss helps to achieve a sharp-looking filtered image while preserving structure even in low-contrast regions.

[0114] In some embodiments, the neural network for suppressing artifacts in the image domain can be trained jointly or independently with any other artifact suppression network. In the former case, the input to the neural network can be (e.g., generated by reconstruction stage 110) the final reconstructed image, and the network can be trained using the target image x. In some embodiments, the network can be trained using the target x0 before resizing. In this way, the filter will learn to upsample the image in an optimal manner.

[0115] Next, we discuss exemplary examples of neural networks used to suppress RF interference in the spatial frequency domain, based on some embodiments of the techniques described herein.

[0116] The inventors have realized that, in some instances, RF interference itself can manifest as one or more bright zipper-like scratches along the phase-encoding direction in an image, since the frequency components of the interference captured in the spatial frequency domain are generally consistent throughout the scan. Image recovery of regions damaged by zipper-like artifacts is challenging because their appearance has a complex structure, which typically significantly degrades the underlying image in the image domain. However, in the k-space domain, although the noise is not localized as much as in the image domain and therefore more areas are affected, the damage is not destructive because the noise frequency components with small amplitudes are superimposed on the signal data with much larger amplitudes.

[0117] Figure 2A The architecture of an exemplary neural network with a spectral depooling layer for suppressing RF interference in MR data is shown, according to some embodiments of the techniques described herein. Figure 2A Neural network implementation reference Figure 1D and 1E The version of the architecture described.

[0118] In particular, Figure 2A The neural network comprises multiple convolutional residual blocks (e.g., n = 8) used as interpolation filters after learning. In the downsampling path, spectral pooling is applied after every two convolutional blocks to project the data onto a low-dimensional frequency basis. In the upsampling path, spectral depooling layers are used to upsample the lower-level k-space features and combine them with higher-level features (from the skip connections). In this example, the spectral depooling layer applies convolutions, batch normalization, ReLU, and pointwise multiplication layers (with learning coefficients) to both the lower-level and higher-level features (skips). The lower-level features are then zero-padded and added to the processed skips.

[0119] exist Figure 2A In this specific implementation, the input image is resized to 128×128 blocks. In the downsampling path, spectral pooling reduces the height and width of each activation by 2; the convolutional layers utilize 3×3 kernels, with output dimensions of 16 (before the first spectral pooling), 32 (between the first and second spectral pooling), 48 (between the second and final spectral pooling), and 64 (after the final spectral pooling). In the upsampling path, the convolutional kernels are 1×1, and the output dimension is the same as the input dimension.

[0120] In this example, training is performed using a loss function consisting of k-space loss (k-space MSE and conjugate symmetric loss) and image domain loss (structural similarity index) (in some instances, the Adam optimizer with a learning rate of 0.001 is used). Figure 2AThe neural network. Specifically, we define the loss as... Where y is the target image. It is a denoised image, F is the Fourier transform, L conj It is a conjugate symmetry loss, and SSIM is defined in the same way as Wang, Z., Bovik, AC, Sheikh, HR, and Simoncelli, EP in “Image quality assessment: from errorvisibility to structural similarity” (IEEE Transactions on Image Processing, Vol. 13(4), pp. 600–612, 2004). Using this type of k-space loss helps suppress zipper artifacts, while the structural similarity index ensures that the image is sharp.

[0121] exist Figure 2A and 2B In the example, a database of disturbed and destructive images is used as input paired with undisturbed data as the ground truth of the target ground to train the system. Figure 2A The neural network. The interference-free image is corrupted by RF interference synthesized using a generative statistical model. In this example, the inventors have recognized that the RF interference has a specific structure, which can be expressed by equations... To model, where α m It is the intensity of the interference, β m The location of the interfering noise in the image space is determined, where ∈ is the added Gaussian noise. The neural network is then trained using the corrupted image and the original, interference-free image.

[0122] Figure 2B Some embodiments of the technology described herein are shown. Figure 2A The exemplary neural network shown is used to suppress RF interference in MR images. It utilizes λ in the loss function. c =1e-5, λ s =1e-2 and the degradation function M~Uniform(0,60), α m ~Uniform(0,1) and β m ~Uniform(0,1) for training Figure 2A The neural network was trained for 1000 epochs, with 8192 training volumes per epoch and a batch size of 64.

[0123] Figure 2B For each of the two different MR images, the following three images are shown: (1) ground truth (original image without interference); (2) interference-damaged MR image; and (3) image obtained by using... Figure 2AThe model was used to suppress RF interference to obtain cleaned images. Results showed that the model suppressed the zipper lines in the images while preserving other brain features. For Figure 2B For the two slices shown, the overall SSIM improved from 0.862 to 0.919 and from 0.884 to 0.918, respectively; for the areas affected by interference noise (the 20 pixels closest to the interference line in the vertical direction), the SSIM improved significantly from 0.377 to 0.744 and from 0.472 to 0.743.

[0124] Figure 3 This is a flowchart of an exemplary process 300 for using a neural network model to suppress one or more artifacts present in input MR data, according to some embodiments of the techniques described herein. Process 300 can be performed using any suitable computing device. For example, in some embodiments, process 300 may be performed by a computing device located in the same location (e.g., in the same room) as the MRI system that acquires MR data by imaging a subject (or object). As another example, in some embodiments, process 300 may be performed by one or more processors located at a location remote from the MRI system that acquires the input MR data (e.g., as part of a cloud computing environment).

[0125] Process 300 begins at action 302, where input MR data is acquired. In some embodiments, the input MR data has previously been acquired and stored by the MRI system for subsequent analysis, making it accessible at action 302. In other embodiments, as part of process 300, the input MR data may be acquired by an MRI system (including any MRI system described herein). In some embodiments, the data may have been acquired using a Cartesian sampling trajectory. In other embodiments, the data may have been acquired using a non-Cartesian sampling trajectory, examples of which are provided herein.

[0126] After one or more preprocessing steps (which may be optional and may involve transforming the input MR data from the sensor domain to the spatial frequency domain), as described above, processing 300 proceeds to action 304, where at least one artifact is suppressed in the spatial frequency domain using a neural network model.

[0127] In some embodiments, spatial frequency domain processing is performed in two stages. For example, during the first stage, a first part of a neural network model can be used to suppress RF interference in the spatial frequency domain, as described here (including stage 106 of reference processing pipeline 100, and...). Figures 1A-1E As described in 2A-2B). Then, during another stage (in some embodiments immediately following), noise in the MR data can be suppressed in the spatial frequency domain, as described here (including stage 108 of the reference processing pipeline 100, and...). Figures 1A-1EAs described herein. However, in other embodiments, different numbers of artifact suppression stages (e.g., one, three, four, five, etc.) may be used, as the aspects of the technique described herein are not limited thereto. For example, in some embodiments, a single artifact suppression stage may be used at action 304 to simultaneously suppress RF interference and noise in the spatial frequency domain.

[0128] Next, process 300 proceeds to action 306, where image reconstruction is performed to transform the spatial domain MR data to image domain data. Reconstruction can be performed in any suitable manner. For example, when the spatial frequency domain data is spaced on a Cartesian grid, a 2D inverse Fourier transform (e.g., using a 2D inverse fast Fourier transform) can be used to transform the data. As another example, when the spatial frequency domain data is undersampled, a non-uniform inverse Fourier transform, a neural network model for reconstructing image data from non-Cartesian k-space data, compressed sensing, and / or any other suitable method can be used to transform the data, as the aspects of the techniques described herein are not limited thereto.

[0129] Next, processing 300 proceeds to action 308, where a neural network model is applied to suppress artifacts present in the image obtained at action 306. The neural network model can be applied in the image domain and can have any suitable architecture (including any architecture described herein). In some embodiments, the processing at action 308 can be performed as described herein (including stage 112 of reference processing pipeline 100, and...). Figures 1A-1E As described above. After action 308 is completed, processing 300 proceeds to action 310, in which the resulting MR image is output (e.g., saved for later access, sent to a recipient via a network, etc.).

[0130] In some embodiments, as part of processing 300, a neural network model with three parts may be employed. The first part may be configured to suppress RF interference in the spatial frequency domain. The second part may be configured to suppress noise in the spatial frequency domain. The third part may be configured to suppress noise in the image domain. These parts may be trained jointly or independently of each other. In some embodiments, all three parts are used, but in other embodiments, one or two of these parts may be omitted.

[0131] exist Figure 3 In the exemplary example, action 304 relates to suppressing artifacts in MR data in the spatial frequency domain. However, in other embodiments, artifacts can be suppressed in the sensor domain or any other suitable domain, as the aspects of the technique described herein are not limited thereto.

[0132] Next, we will discuss additional aspects of training the aforementioned neural network model.

[0133] In some embodiments, a residual training strategy can be employed to train a neural network to suppress artifacts in corrupted training data. As part of this residual training strategy, the input data can include a superposition of the signal of interest (e.g., an MR signal) and unwanted artifacts that corrupt the signal of interest (e.g., RF interference, noise, etc.). The target data can be the unwanted artifact data (e.g., RF interference signals added to the MR signal of interest to generate the input data). The output data is then the unwanted artifacts (e.g., the RF interference signals)—therefore, the goal is to estimate the unwanted artifact data rather than the clean MR signal. This approach is sometimes referred to as "residual training." Neural network models trained in this way focus on learning the difference between data with and without artifacts, rather than on learning to generate artifact-free data, which helps with convergence during training.

[0134] In some embodiments, one or a linear combination of multiple loss functions may be used to train the neural network model described herein:

[0135] • L2 loss between output and target data.

[0136] • L1 loss between output and target data.

[0137] • L2-weighted loss between output and target data. Weights can be calculated based on k-space coordinates. The higher the spatial frequency (the farther away from the center of k-space), the higher the weight. Using such weights ensures that the resulting model maintains a high spatial frequency with more noise compared to lower frequencies.

[0138] • L1-weighted regularization of the output. A sparse prior can be enforced on the neural network output by optionally applying an L1 norm after weighting. Weights can be computed based on k-space coordinates. Higher spatial frequencies (further from the center of k-space) correspond to smaller weights. This promotes sparsity.

[0139] • Generative adversarial network loss.

[0140] • Structural similarity index loss.

[0141] • Here (including combinations) Figures 1A-1E And any other loss function described in 2A to 2B).

[0142] It should be understood that regardless of the domain in which the neural network operates, the loss function described herein can be computed in any suitable domain (e.g., the sensor domain, the k-space domain, the image domain) or (e.g., any weighted representation). For example, a neural network designed to operate on spatial frequency data can take spatial frequency data as input and produce spatial frequency data as output, but during training, its loss function can be computed in the image domain (e.g., after the data has been appropriately transformed using a Fourier transform). This is helpful because some loss functions may not be directly or possibly computed in one domain, but may be easier to compute in another (e.g., it is natural to compute the SSIM loss in the image domain). Therefore, the input data can be transformed using any suitable transformation (e.g., by meshing to k-space and / or performing a Fourier transform) before applying the neural network.

[0143] Any suitable optimization technique can be used to estimate neural network parameters from the data. For example, one or more of the following optimization techniques can be used: stochastic gradient descent (SGD), mini-batch gradient descent, momentum SGD, Nesterov accelerated gradient, Adagrad, Adadelta, RMSprop, adaptive moment estimation (Adam), AdaMax, Nesterov accelerated adaptive moment estimation (Nadam), and AMSGrad.

[0144] In some embodiments, training data for training the neural network model described herein can be obtained by: (1) synthesizing and / or measuring RF artifact measurements; (2) synthesizing and / or measuring MR measurements; and (3) combining the obtained RF artifact measurements and MR measurements to obtain artifact-damaged MR data. The artifact-damaged MR data (and the corresponding individual artifact and MR data components) can then be used to train one or more neural network models for suppressing artifacts in MR data.

[0145] In some embodiments, synthetic and / or measured RF artifact measurements can represent various sources of interference and / or noise. For example, synthetic and / or measured RF artifact measurements can represent external RF interference generated by one or more electronic devices, including but not limited to computers, monitors, mobile phones, Bluetooth devices, medical devices (e.g., EEG, ECG, pulse oximeter, cardiac monitor, blood pressure cuff, etc.), transformers, motors, pumps, fans, and ventilators. As another example, synthetic and / or measured RF artifact measurements can represent internal RF interference generated by the electrical and / or magnetic components of the MRI system (e.g., gradient coils, power amplifiers, etc.). This interference itself can manifest in a predictable manner as a function of the pulse sequence used for imaging. This interference can be effectively measured using one or more pulse sequences during operation of the MRI system. As another example, synthetic and / or measured RF artifact measurements can represent noise generated by the MR receiver chain or the subject (or object) being imaged.

[0146] In some embodiments, RF artifact measurements can be obtained using any sensor configured to directly or indirectly capture any RF artifacts present in the MRI system environment. Such a sensor may include the MRI system sensor itself (e.g., one or more RF coils) and / or any auxiliary sensors that may be located near the MRI system or even in other locations (e.g., other rooms in the hospital).

[0147] In some embodiments, RF artifact measurements can be obtained by one or more sensors when the MRI system performs one or more pulse sequences during MR acquisition, whether the MRI system has a subject (or object) in the imaging area of ​​the MRI system or not. For example, in some embodiments, RF artifacts can be measured by one or more sensors without emitting any RF excitation pulses (which avoids the generation of MR signals).

[0148] In some embodiments, RF artifact measurement results can be obtained based on one or more characteristics of the pulse sequence (e.g., sampling rate, readout duration, repetition time, etc.). For example, in some embodiments, RF artifact measurement results can be acquired by one or more sensors at a sampling rate and readout duration consistent with the pulse sequence of interest. In some embodiments, the repetition time between successive artifact measurements can be consistent with the pulse sequence and can be matched with its repetition time (TR).

[0149] Figure 4A The present invention illustrates some embodiments of the techniques described herein for generating training data for training a neural network model, wherein the neural network model is used to suppress artifacts in MR data.

[0150] like Figure 4A As shown, generating training data 420 may include: generating RF artifact measurement results, generating MR measurement results, and combining them using combination block 416 to obtain training data 420.

[0151] In some embodiments, generating RF artifact measurement results may include (e.g., using one or more generative statistical models, one or more physics-based models, etc.) synthesizing RF artifact measurement results 412. In some embodiments, generating RF artifact measurements may include using one or more sensors 402 to obtain RF artifact measurement results 408.

[0152] In some embodiments, generating MR measurement results may include (e.g., using one or more generative statistical models, one or more physics-based models, etc.) synthesizing MR measurement results 410. In some embodiments, generating MR measurement results may include using one or more sensors 402 to obtain clean RF MR measurement results 406 and / or artifact-damped measurement results 404. The measurement results may be measurements of a real subject (e.g., a patient) and / or an object (e.g., a phantom).

[0153] In some embodiments, any measurements obtained using sensor 402 may be preprocessed. For example, the measurement results may be resampled, compressed, pre-denoised, pre-whitened, filtered, amplified, and / or preprocessed in any other suitable manner.

[0154] In some embodiments, MR and RF artifact measurements can be transformed to any suitable domain via domain transform 418. For example, the measurements can be transformed to any other domain using analytical or learned transforms (e.g., Fourier transform, wavelet transform, etc.).

[0155] In some embodiments, after collecting training data, artifact-damaged data (input) is paired with its clean version (target) and used to estimate the parameters of the neural network model using any of the optimization algorithms described above.

[0156] Figure 4B The following are exemplary examples of generating training data for training a neural network model, according to some embodiments of the techniques described herein, wherein the neural network model is used to suppress artifacts in MR data.

[0157] Figure 5 This is a block diagram of typical components of an MRI system 500. Figure 5In an exemplary example, MRI system 500 includes workstation 504, controller 506, pulse sequence storage unit 508, power management system 510, and magnetic assembly 520. It should be understood that system 500 is exemplary, and in addition to... Figure 5 Other than or in place of the components shown Figure 5 As shown in the diagram, the MRI system may have one or more other components of any suitable type.

[0158] like Figure 5 As shown, the magnetic assembly 520 includes a B0 magnet 522, a shimming coil 524, an RF transmit and receive coil 526, and a gradient coil 528. The B0 magnet 522 can be used to at least partially generate the main magnetic field B0. The B0 magnet 522 can be any suitable type of magnet capable of generating the main magnetic field (e.g., a low field strength of about 0.2 T or lower) and can include one or more B0 coils, correction coils, etc. The shimming coil 524 can be used to contribute a magnetic field to improve the uniformity of the B0 field generated by the magnet 522. The gradient coil 528 can be arranged to provide a gradient field and, for example, can be arranged to generate gradients in the magnetic field along three substantially orthogonal directions (X, Y, Z) to locate the position where the MR signal is induced.

[0159] RF transmit and receive coil 526 may include one or more transmit coils that can be used to generate RF pulses to sense magnetic field B1. The transmit / receive coils may be configured to generate any suitable type of RF pulse, which is configured to excite the MR response of the subject and detect the emitted MR signal. RF transmit and receive coil 526 may include one or more transmit coils and one or more receive coils. The structure of the transmit / receive coil varies depending on the implementation and may include a single coil for both transmitting and receiving, separate coils for transmitting and receiving, multiple coils for transmitting and / or receiving, or any combination for implementing a single-channel or parallel MRI system. Therefore, the transmit / receive magnetic assembly is often referred to as a Tx / Rx or Tx / Rx coil to generally refer to various structures of the transmit and receive assemblies of an MRI system.

[0160] Each magnetic component 520 can be of any suitable type and can be constructed in any suitable manner. For example, in some embodiments, the BO magnet 522 can be an electromagnet or a permanent magnet (e.g., as referenced below). Figure 6 , 7 (As described in 8A-8B). As another example, in some embodiments, lamination techniques may be used to fabricate one or more magnetic components 520 (e.g., shimming coils 524 and / or gradient coils 528).

[0161] The power management system 510 includes electronic devices for providing operating power to one or more components of the low-field MRI system 500. For example, the power management system 510 may include one or more power supplies, gradient power amplifiers, transmit coil amplifiers, and / or any other suitable power electronics required to provide appropriate operating power to power and operate the components of the low-field MRI system 500.

[0162] like Figure 5 As shown, the power management system 510 includes a power supply 512, an amplifier 514, a transmit / receive switch 516, and a thermal management component 518. The power supply 512 includes electronics for providing operating power to the magnetic components 520 of the low-field MRI system 500. For example, in some embodiments, the power supply 512 may include electronics for providing operating power to one or more Bo coils (e.g., Bo magnet 522), one or more shimming coils 524, and / or one or more gradient coils 528 used to generate the main magnetic field of the low-field MRI system. In some embodiments, the power supply 512 may be a unipolar continuous wave power supply; however, any suitable power supply may be used. The transmit / receive switch 516 can be used to select whether the RF transmit coil or the RF receive coil is being operated.

[0163] In some embodiments, amplifier 514 may include one or more RF receive (Rx) preamplifiers for amplifying MR signals detected by one or more RF receive coils (e.g., coil 524), one or more RF transmit (Tx) amplifiers configured to power one or more RF transmit coils (e.g., coil 526), ​​one or more gradient power amplifiers configured to power one or more gradient coils (e.g., gradient coil 528), and / or one or more shimming amplifiers configured to power one or more shimming coils (e.g., shimming coil 524).

[0164] In some embodiments, thermal management component 518 provides cooling for components of the low-field MRI system 500 and may be configured to provide cooling by facilitating the transfer of heat generated by one or more components of the low-field MRI system 500 away from those components. Thermal management component 518 may include, but is not limited to, components for performing water-based or air-based cooling, which may be integrated with or located near the heat-generating MRI components, including, but not limited to, BO coils, gradient coils, shimming coils, and / or transmit / receive coils. Thermal management component 518 may include any suitable heat transfer medium for facilitating heat transfer away from the components of the low-field MRI system 500, including but not limited to air and water.

[0165] like Figure 5As shown, the low-field MRI system 500 includes a controller 506 (also referred to as a console) having control electronics for sending instructions to and receiving information from a power management system 510. The controller 506 may be configured to implement one or more pulse sequences for determining instructions sent to the power management system 510 to operate the magnetic component 520 using a desired sequence. For example, the controller 506 may be configured to control the power management system 510 to operate the magnetic component 520 according to equilibrium steady-state free precession (bSSFP) pulse sequences, low-field gradient echo pulse sequences, low-field spin echo pulse sequences, low-field inversion recovery pulse sequences, arterial spin labeling, diffusion-weighted imaging (DWI), and / or any other suitable pulse sequences. The controller 506 may be implemented in hardware, software, or any suitable combination of hardware and software, as the aspects of the invention provided herein are not limited thereto.

[0166] In some embodiments, the controller 506 may be configured to implement the pulse sequence by obtaining information related to the pulse sequence from a pulse sequence repository 508, which stores information for one or more pulse sequences. The information stored in the pulse sequence repository 508 for a particular pulse sequence may be any suitable information that allows the controller 506 to implement that particular pulse sequence. For example, the information stored in the pulse sequence repository 508 for a pulse sequence may include one or more parameters for operating the magnetic component 520 according to the pulse sequence (e.g., parameters for operating the RF transmitting and receiving coil 526, parameters for operating the gradient coil 528, etc.), one or more parameters for operating the power management system 510 according to the pulse sequence, one or more programs containing instructions that, when executed by the controller 506, cause the controller 506 to control the system 500 to operate according to the pulse sequence, and / or any other suitable information. The information stored in the pulse sequence repository 508 may be stored on one or more non-transitory storage media.

[0167] like Figure 5 As shown, in some embodiments, controller 506 can interact with computing device 504 programmed to process received MR data (which in some embodiments may be sensor-domain MR data or spatial-frequency domain MR data). For example, computing device 504 can process the received MR data to generate one or more MR images using any suitable image reconstruction processing (including any techniques described herein that utilize neural network models to generate MR images from input MR data). For example, computing device 504 can perform [further details omitted]. Figure 3Any of the aforementioned processing. Controller 506 may provide computing device 504 with information relating to one or more pulse sequences for data processing by the computing device. For example, controller 506 may provide computing device 504 with information relating to one or more pulse sequences, and the computing device may perform image reconstruction processing based at least in part on the provided information.

[0168] In some embodiments, computing device 504 may be any electronic device configured to process acquired MR data and generate one or more images of the subject being imaged. In some embodiments, computing device 504 may include a fixed electronic device, such as a desktop computer, server, rack-mount computer, or any other suitable fixed electronic device configured to process MR data and generate one or more images of the subject being imaged. Optionally, computing device 504 may be a portable device, such as a smartphone, personal digital assistant, laptop computer, tablet computer, or any other portable device configured to process MR data and generate one or more images of the subject being imaged. In some embodiments, computing device 504 may include multiple computing devices of any suitable type, as the aspects of the technology described herein are not limited thereto.

[0169] In some embodiments, user 502 can interact with computing device 504 to control aspects of the low-field MR system 500 (e.g., programming the system 500 to operate according to a specific pulse sequence, adjusting one or more parameters of the system 500, etc.) and / or view images acquired by the low-field MR system 500. According to some embodiments, computing device 504 and controller 506 form a single controller, while in other embodiments, computing device 504 and controller 506 each include one or more controllers. It should be understood that the functions performed by computing device 504 and controller 506 can be distributed in any way across any combination of one or more controllers, because the aspects of the technology described herein are not limited to use with any particular implementation or architecture.

[0170] Figure 6 and 7 A biplane permanent magnet structure of a B0 magnet according to some embodiments of the technology described herein is shown. Figure 6 A permanent BO magnet 600 according to some embodiments is shown. In the illustrated embodiment, the BO magnet 600 is formed by permanent magnets 610a and 610b arranged in a biplane geometry and a yoke 620 for capturing the electromagnetic flux generated by the permanent magnets and transferring that flux to opposing permanent magnets to increase the flux density between the permanent magnets 610a and 610b. The permanent magnets 610a and 610b are each formed from a plurality of concentric permanent magnet rings. In particular, as shown... Figure 6As can be seen, the permanent magnet 610b includes an outer permanent magnet ring 614a, a middle permanent magnet ring 614b, an inner permanent magnet ring 614c, and a central permanent magnet disk 614d. Although shown as having four concentric permanent magnet rings, the permanent magnet 610b (and permanent magnet 610a) can have any suitable number of permanent magnet rings, as the aspects of the technology described herein are not limited thereto. The permanent magnet 610a can be formed substantially the same as the permanent magnet 610b, and, for example, includes the same set of permanent magnet rings as the permanent magnet 610b.

[0171] The permanent magnet material used can be selected based on the system's design requirements. For example, according to some embodiments, the permanent magnet (or a portion thereof) can be made of NdFeB, which, once magnetized, generates a magnetic field in a manner that results in a relatively high magnetic field per unit volume of material. In some embodiments, SmCo material is used to form the permanent magnet or a portion thereof. While NdFeB produces a higher field strength (and is generally cheaper than SmCo), SmCo exhibits less thermal drift, thus providing a more stable magnetic field in the face of temperature fluctuations. Other types of permanent magnet materials can also be used, as the aspects of the technique described herein are not limited to these. Generally, the type of permanent magnet material utilized will depend at least in part on the field strength, temperature stability, weight, cost, and / or ease-of-use requirements to be achieved for a given B0 magnet.

[0172] In some embodiments, the size and arrangement of the permanent magnet rings are designed to generate a uniform field of desired intensity in the imaging region (field of view) between permanent magnets 610a and 610b. Figure 6 In the exemplary embodiments shown, each permanent magnet ring includes multiple segments, each segment formed using multiple permanent magnet blocks stacked radially and positioned adjacent to each other around its outer periphery to form a corresponding ring. The inventors have realized that by varying the width of each permanent magnet (in the direction tangential to the ring), less wasted useful space can be achieved while using less material. For example, by varying the width of the blocks (e.g., as a function of the radial position of the blocks), the space between stacks that do not generate a useful magnetic field can be reduced, allowing for a tighter fit to reduce wasted space and maximize the amount of magnetic field that can be generated in a given space. The size of the blocks can also be varied in any desired manner to facilitate the generation of a magnetic field with desired strength and uniformity. For example, in some embodiments, the heights of the blocks in different rings may differ from each other, and / or the heights of one or more blocks within a particular ring may differ from each other to achieve a magnetic field with desired strength and uniformity.

[0173] like Figure 6As shown, the B0 magnet 600 also includes a yoke 620 configured and arranged to capture the magnetic flux generated by permanent magnets 610a and 610b and direct it to opposite sides of the B0 magnet to increase the flux density between permanent magnets 610a and 610b, thereby increasing the field strength within the field of view of the B0 magnet. By capturing the magnetic flux and directing it to the region between permanent magnets 610a and 610b, the desired field strength can be achieved using less permanent magnet material, thereby reducing the size, weight, and cost of the B0 magnet 600. Alternatively, for a given permanent magnet, the field strength can be increased, thereby improving the system's SNR without using an increased amount of permanent magnet material. For a typical B0 magnet 600, the yoke 620 includes a frame 622 and plates 624a and 624b. Plates 624a and 624b can capture the magnetic flux generated by permanent magnets 610a and 610b and guide it to frame 622 for circulation via the magnetic loop of the yoke, thereby increasing the flux density in the field of view of the B0 magnet. The yoke 620 can be constructed of any desired ferromagnetic material (e.g., low-carbon steel, CoFe, and / or silicon steel, etc.) to provide the desired magnetic properties to the yoke. In some embodiments, plates 624a and 624b (and / or frame 622 or portions thereof) can be constructed of silicon steel, etc., in regions where eddy currents are most commonly induced by gradient coils.

[0174] A typical frame 622 includes arms 623a and 623b attached to plates 624a and 624b respectively, and supports 625a and 625b providing a magnetic loop for the flux generated by the permanent magnet. The arms are generally designed to reduce the amount of material required to support the permanent magnet while providing sufficient loop cross-section for the magnetic flux generated by the permanent magnet. Frame 622 has two supports within the magnetic loop for the B0 field generated by the B0 magnet. A gap 627 is formed between supports 625a and 625b, thereby providing sufficient cross-section for the magnetic flux generated by the permanent magnet while providing a measure of frame stability and / or structural lightness. For example, the required cross-section for the magnetic flux loop can be divided between the two support structures, thereby providing sufficient loop while increasing the structural integrity of the frame.

[0175] Figure 7 A B0 magnet 700 is shown according to some embodiments. The B0 magnet 700 can be used with... Figure 6The B0 magnet 600 shown shares a common design component. Specifically, the B0 magnet 700 is formed by permanent magnets 710a and 710b arranged in a biplane geometry and a yoke 720 coupled to the permanent magnets 710a and 710b to capture the electromagnetic flux generated by the permanent magnets and transfer the flux to the opposing permanent magnets to increase the flux density between the permanent magnets 710a and 710b. The permanent magnets 710a and 710b are each formed from multiple concentric permanent magnets, as shown in permanent magnet 710b, which includes an outer permanent magnet ring 714a, a middle permanent magnet ring 714b, an inner permanent magnet ring 714c, and a central permanent magnet disk 714d. Permanent magnet 710a may include the same set of permanent magnet elements as permanent magnet 710b. The permanent magnet material used can be selected according to the design requirements of the system (e.g., NdFeB, SmCo, etc., depending on the desired characteristics).

[0176] The size and arrangement of the permanent magnet rings are designed to generate a uniform field of desired intensity in the central region (field of view) between permanent magnets 710a and 710b. Figure 7 In an exemplary embodiment, each permanent magnet ring includes multiple arc segments whose size and position are designed to generate a desired B0 magnetic field. Figure 6 Similar to the illustrated yoke 620, the yoke 720 is configured and arranged to capture the magnetic flux generated by permanent magnets 710a and 710b and direct it to opposite sides of the B0 magnet to increase the flux density between the permanent magnets 710a and 710b. The yoke 720 thereby increases the field strength in the field of view of the B0 magnet, which has less permanent magnet material, reducing the size, weight, and cost of the B0 magnet. The yoke 720 also includes a frame 722 and plates 724a and 724b, which capture the magnetic flux generated by permanent magnet 710a in a manner similar to that described above in conjunction with the yoke 720 and circulate it via the magnetic loop of the yoke to increase the flux density in the field of view of the B0 magnet. The structure of the yoke 720 can be similar to the structure described above to provide sufficient material to accommodate the magnetic flux generated by the permanent magnets and provide sufficient stability, while minimizing the amount of material used to reduce, for example, the cost and weight of the B0 magnet.

[0177] Because a permanent B0 magnet generates its own persistent magnetic field once magnetized, no electricity is required to operate the permanent B0 magnet to generate its magnetic field. As a result, the significant (often dominant) contribution to the overall power consumption of the MRI system is eliminated by using a permanent magnet (as opposed to, for example, an electromagnet that requires electricity), thus facilitating the development of MRI systems that can be powered by mains electricity (e.g., via a standard wall socket or a common large household appliance outlet). As described above, the inventors have developed a low-power, portable low-field MRI system that can be deployed in virtually any environment and can be taken to the patient undergoing the imaging procedure. In this way, patients in emergency rooms, intensive care units, operating rooms, and many other locations can benefit from MRI when it is typically unavailable.

[0178] Figure 8A and 8B A view of a portable MRI system 800 according to some embodiments of the technology described herein is shown. The portable MRI system 800 includes a B0 magnet 810 partially formed by an upper magnet 810A and a lower magnet 810b, the B0 magnet 810 having a yoke 820 coupled to the B0 magnet 810 to increase flux density within the imaging region. The B0 magnet 810 may be housed together with a gradient coil 815 (e.g., any gradient coil described in U.S. Application No. 14 / 845,652, entitled "Low Field Magnetic Resonance Imaging Methods and Apparatus," filed September 4, 2015, which is incorporated herein by reference in its entirety) within a magnet housing 812. In some embodiments, the B0 magnet 810 includes an electromagnet. In some embodiments, the B0 magnet 810 includes a permanent magnet (e.g., any permanent magnet described in U.S. Application No. 15 / 640,369, filed June 30, 2017, entitled “LOW-FIELD MAGNETIC RESONANCE IMAGING METHODS AND APPARATUS,” which is incorporated herein by reference in its entirety). For example, in some embodiments, the B0 magnet 810 may be a reference magnet. Figure 6 The permanent magnet 600 or reference Figure 7 The permanent magnet 700 is described above.

[0179] The exemplary portable MRI system 800 also includes a base 850 for housing electronics that operate the MRI system. For example, the base 850 may house electronics including, but not limited to, one or more gradient power amplifiers, an on-system computer, a power distribution unit, one or more power supplies, and / or any other power components configured to operate the MRI system using mains power (e.g., via a connection to a standard wall outlet and / or a large electrical outlet). For example, the base 870 may house low-power components such as those described herein, at least in part enabling power supply to the portable MRI system from an readily available wall outlet. Thus, the portable MRI system 800 can be taken to a patient and plugged into a nearby wall outlet.

[0180] The portable MRI system 800 also includes a movable slider 860, which can be opened, closed, and positioned in various configurations. The slider 860 includes an electromagnetic shield 865, which can be made of any suitable conductive or magnetic material to form the movable shield, thereby attenuating electromagnetic noise in the operating environment of the portable MRI system to shield the imaging area from at least some electromagnetic noise. As used herein, the term electromagnetic shielding refers to a conductive or magnetic material configured to attenuate electromagnetic fields in the spectrum of interest and positioned or arranged to shield the space, object, and / or component of interest. In the context of an MRI system, electromagnetic shielding can be used to shield electronic components of the MRI system (e.g., power components, cables, etc.), shield the imaging area of ​​the MRI system (e.g., the field of view), or both.

[0181] The degree of attenuation achieved by electromagnetic shielding depends on many factors, including the type of material used, the material thickness, the spectrum for which electromagnetic shielding is desired or required, the size and shape of the apertures in the electromagnetic shield (e.g., the size of the space in the conductive mesh, the size of the unshielded portion or gap in the shield, etc.), and / or the orientation of the apertures relative to the incident electromagnetic field. Therefore, electromagnetic shielding generally refers to any conductive or magnetic barrier used to attenuate at least some electromagnetic radiation and positioned to at least partially shield a given space, object, or component by attenuating at least some electromagnetic radiation.

[0182] It should be understood that the spectrum for which shielding (attenuation of the electromagnetic field) is desired can vary depending on what is being shielded. For example, the electromagnetic shielding of certain electronic components may be configured to attenuate different frequencies compared to the electromagnetic shielding of the imaging region of an MRI system. Regarding the imaging region, the spectrum of interest includes frequencies that alter, influence, and / or reduce the ability of the MRI system to excite and detect MR responses. Generally, the spectrum of interest for the imaging region of an MRI system corresponds to frequencies associated with the rated operating frequency (e.g., Larmor frequency) at a given B0 magnetic field strength that the receiving system is configured to or able to detect. This spectrum is referred to herein as the operating spectrum of the MRI system. Therefore, electromagnetic shielding that provides shielding for the operating spectrum refers to conductive or magnetic materials arranged or positioned to attenuate frequencies within the operating spectrum that are located at least in at least a portion of the imaging region of the MRI system.

[0183] exist Figure 8A and 8B In the portable MRI system 800 shown, movable shielding elements 18A and 18B can therefore be configured to provide shielding in different arrangements that can be adjusted as needed to accommodate the patient, provide access to the patient, and / or conform to a given imaging protocol. For example, for imaging procedures such as brain scans, once the patient has been positioned, the slider 860 can be closed, for example, using the handle 862, to provide electromagnetic shielding 865 around the imaging area in addition to an opening to accommodate the patient's upper torso. As another example, for imaging procedures such as knee scans, the slider 860 can be arranged to have openings on both sides to accommodate the patient's legs. Thus, the movable shielding elements allow the shielding to be configured in an arrangement suitable for the imaging procedure and facilitate proper patient positioning within the imaging area.

[0184] In some embodiments, a noise reduction system comprising one or more noise reduction and / or compensation techniques may be implemented to suppress at least some electromagnetic noise that is not blocked or sufficiently attenuated by the shield 865. In particular, the inventors have developed noise reduction systems configured to suppress, avoid, and / or reject electromagnetic noise in the operating environment of an MRI system. According to some embodiments, these noise suppression techniques work in conjunction with a movable shield to facilitate operation in various shielding structures with arrangeable sliders. For example, when slider 960 is open, an increased level of electromagnetic noise may enter the imaging area via the opening. As a result, the noise suppression component will detect the increased electromagnetic noise level and accordingly adapt the noise suppression and / or avoidance response. Due to the dynamic nature of the noise suppression and / or avoidance techniques described herein, the noise reduction system is configured to respond to changing noise conditions, including those resulting from different arrangements of the movable shield. Therefore, a noise reduction system according to some embodiments may be configured to work in conjunction with a movable shield to suppress electromagnetic noise in the operating environment of an MRI system in any available shielding structure, including structures that are substantially unshielded (e.g., structures without movable shields).

[0185] To ensure that the movable shield provides shielding regardless of the arrangement of the sliding element, electrical pads can be arranged to provide continuous shielding along the outer periphery of the movable shield. For example, as Figure 8B As shown, electrical pads 867a and 867b may be disposed at the interface between the slider 860 and the magnet housing to maintain continuous shielding along the interface. According to some embodiments, the electrical pads are beryllium pawls or beryllium copper pawls (e.g., aluminum pads) used to maintain an electrical connection between the shield 865 and ground during and after the slider 860 is moved to a desired position around the imaging area.

[0186] To facilitate transport, a mobility component 880 is provided to allow the portable MRI system to be moved from one location to another, for example, using controls such as joysticks or other control mechanisms located on or off the MRI system. In this way, the portable MRI system 800 can be transported to the patient and moved to the bedside for imaging.

[0187] The portable MRI system described herein can be operated from portable electronic devices such as laptops, tablets, smartphones, etc. For example, a tablet computer 875 can be used to operate the portable MRI system to run desired imaging protocols and view the resulting images. The tablet computer 875 can be connected to a secure cloud to transmit images for data sharing on datasets, telemedicine, and / or deep learning. Any technology utilizing network connectivity described in U.S. Application No. 14 / 846158, filed September 4, 2015, entitled “Automatic Configuration of a Low Field Magnetic Resonance Imaging System,” can be incorporated herein by reference in its entirety from the portable MRI system described herein.

[0188] As described above, Figure 9A A portable MRI system 900 has been transported to the patient's bedside for a brain scan. Figure 9B A portable MRI system 900 is shown, which has been transported to the patient's bedside for a knee scan. Figure 9B As shown, the shielding component 960 has an electrical gasket 867c.

[0189] It should be understood that Figures 8A-8B The electromagnetic shielding shown in 9A-9B is typical, and providing shielding for an MRI system is not limited to the exemplary electromagnetic shielding described herein. Electromagnetic shielding can be implemented using any suitable material and in any suitable manner. For example, electromagnetic shielding can be formed using conductive mesh, fabric, etc., capable of providing a movable “curtain” to shield the imaging area. Electromagnetic shielding to shield the imaging area from electromagnetic interference can be formed using one or more conductive strips (e.g., one or more strips of conductive material) coupled to the MRI system as fixed, movable, or configurable components, some examples of which will be described in further detail below. Electromagnetic shielding can be provided by embedding material into any movable or fixed part of a door, slide, or housing. Electromagnetic shielding can be deployed as a fixed or movable component, as aspects are not limited thereto.

[0190] Figure 10 This is a diagram of an exemplary computer system that can implement the embodiments described herein. Figure 10 The diagram illustrates an exemplary implementation of a computer system 1000 that can be used in conjunction with any embodiment of the invention provided herein. For example, see reference to... Figure 3 and 4AThe aforementioned processing can be implemented on and / or using computer system 1000. As another example, computer system 1000 can be used to train and / or use any neural network statistical model described herein. Computer system 1000 may include one or more processors 1002 and one or more articles of manufacture containing non-transitory computer-readable storage media (e.g., memory 1004 and one or more non-volatile storage devices 1006). Processor 1002 can control data writing and data reading relative to memory 1004 and non-volatile storage devices 1006 in any suitable manner, as the aspects of the invention provided herein are not limited thereto. To perform any of the functions described herein, processor 1002 can execute one or more processor-executable instructions stored in one or more non-transitory computer-readable storage media (e.g., memory 1004), which can be used as a non-transitory computer-readable storage medium storing processor-executable instructions executable by processor 1002.

[0191] Therefore, having described several aspects and embodiments of the technology set forth herein, it should be understood that various changes, modifications, and improvements will be readily apparent to those skilled in the art. Such changes, modifications, and improvements are intended to be within the spirit and scope of the technology described herein. For example, those skilled in the art will readily conceive of various other components and / or structures for performing functions and / or obtaining results and / or one or more advantages described herein, and these changes and / or modifications are each considered within the scope of the embodiments described herein. Those skilled in the art will recognize or be able to identify many equivalent embodiments using only conventional experimentation, i.e., the specific embodiments described herein. Therefore, it should be understood that the above embodiments are presented by way of example only, and that embodiments of the invention may be practiced in other ways besides the specific descriptions, within the scope of the appended claims and their equivalents. Furthermore, any combination of two or more such features, systems, articles, materials, kits, and / or methods is included within the scope of the invention if they are not inconsistent with each other.

[0192] The above embodiments can be implemented in any number of ways. One or more aspects and embodiments of the invention relating to the performance of processing or methods can utilize program instructions executable by means (e.g., a computer, processor, or other means) to perform the processing or method or control the performance of the processing or method. In this regard, various inventive concepts can be embodied in a computer-readable storage medium (or multiple computer-readable storage media) (e.g., a computer memory, one or more floppy disks, compact disks, optical disks, magnetic tapes, flash memory, circuit structures in field-programmable gate arrays or other semiconductor devices, or other tangible computer storage media) encoded with one or more programs, which, when executed on one or more computers or other processors, perform the methods for implementing one or more of the various embodiments described above. The computer-readable medium can be transportable, such that a program stored thereon can be loaded onto one or more different computers or other processors to implement various aspects of the above aspects. In some embodiments, the computer-readable medium can be a non-transitory medium.

[0193] As used herein, the terms "program" or "software" generally refer to any type of computer code or set of computer-executable instructions that can be used to program a computer or other processor to implement the aspects described above. Furthermore, it should be understood that, according to one aspect, one or more computer programs that carry out the methods of the invention during execution do not need to reside on a single computer or processor, but can be distributed in a modular manner among multiple different computers or processors to implement the aspects of the invention.

[0194] Computer-executable instructions can take many forms, such as program modules executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, etc., for performing specific tasks or implementing specific abstract data types. Typically, the functionality of program modules can be combined or distributed in various embodiments as needed.

[0195] Furthermore, data structures can be stored in any suitable form on a computer-readable medium. For simplicity, a data structure might be shown as having fields related by position within the data structure. Similarly, such relationships can be implemented by allocating storage in a computer-readable medium with locations for conveying relationships between fields. However, any suitable mechanism can be used to establish relationships between information in the fields of a data structure, including by using pointers, labels, or other mechanisms that establish relationships between data elements.

[0196] When implemented in software, the software code can be executed on any suitable processor or set of processors, whether provided in a single computer or distributed among multiple computers.

[0197] Furthermore, it should be understood that, as a non-limiting example, a computer can be embodied in any of a variety of forms, such as a rack-mount computer, a desktop computer, a laptop computer, or a tablet computer. Additionally, a computer can be embedded in a device that is not generally considered a computer but has suitable processing capabilities, including a personal digital assistant (PDA), a smartphone, or any other suitable portable or stationary electronic device.

[0198] Additionally, a computer may have one or more input and output devices. These devices can be used to present a user interface, etc. Examples of output devices that can be used to provide a user interface include a printer or display screen for visual presentation and a speaker or other sound-generating device for audible presentation. Examples of input devices that can be used for a user interface include keyboards and pointing devices such as mice, touchpads, and digital tablets. As another example, a computer may receive input information through voice recognition or in other audible formats.

[0199] Such computers can be interconnected through one or more networks of any suitable form, including local area networks (LANs) or wide area networks (WANs), such as enterprise networks and intelligent networks (INs) or the Internet. These networks can be based on any suitable technology and can operate according to any suitable protocol, and can include wireless networks, wired networks, or fiber optic networks.

[0200] Furthermore, as described, some aspects can be embodied as one or more methods. Actions performed as part of a method can be ordered in any suitable manner. Therefore, embodiments can be constructed in which actions are performed in a different order than those illustrated, which may include simultaneous performance of some actions, even if shown as sequential actions in the exemplary embodiments.

[0201] All definitions defined and used herein should be understood as control dictionary definitions, definitions in referenced and incorporated literature, and / or the usual meaning of the defined terms.

[0202] In the specification and claims, unless there is an explicit indication to the contrary, the indefinite articles “a” and “an” used herein shall be understood to mean “at least one”.

[0203] In the specification and claims, the phrase “and / or” as used herein should be understood to mean “any one or both” of the elements thus combined (i.e., elements presented together in some cases and separately in others). Multiple elements listed using “and / or” should be interpreted in the same way, i.e., “one or more” of the elements thus combined. Other elements may optionally be present in addition to those specifically identified by the “and / or” clause, whether related to or unrelated to those specifically identified elements. Thus, as a non-limiting example, in one embodiment, a reference to “A and / or B” when used in conjunction with open-ended language such as “comprising” may refer only to A (optionally including elements other than B); in another embodiment, it may refer only to B (optionally including elements other than A); in yet another embodiment, it may refer to both A and B (optionally including other elements); and so on.

[0204] In the specification and claims, the phrase "at least one" as used herein, when referring to a list of one or more elements, should be understood to mean at least one element selected from any one or more elements in the list, but does not necessarily include at least one element from every element specifically listed in the list, and does not exclude any combination of elements in the list. This definition also allows for the optional presence of elements other than those specifically identified within the list of elements referenced by the phrase "at least one," whether related to or unrelated to the specifically identified elements. Thus, as a non-limiting example, "at least one of A and B" (or equivalently "at least one of A or B," or equivalently "at least one of A and / or B") in one embodiment may refer to optionally including at least one more A without the presence of B (and optionally including elements other than B); in another embodiment may refer to optionally including at least one more B without the presence of A (and optionally including elements other than A); in yet another embodiment may refer to optionally including at least one more A and optionally including at least one more B (and optionally including other elements); and so on.

[0205] In the claims, and in the description above, all transitional phrases such as “comprising,” “including,” “carrying,” “having,” “containing,” “involving,” “holding,” and “covering” should be understood as open-ended, meaning including but not limited to. Only the transitional phrases “consisting of…” and “consisting substantially of…” should be closed or semi-closed transitional phrases, respectively.

[0206] The terms “about” and “approximately” can be used to mean within ±20% of the target value in some embodiments, within ±10% of the target value in some embodiments, within ±5% of the target value in some embodiments, and within ±2% of the target value in some embodiments. The terms “about” and “approximately” can include the target value.

Claims

1. A magnetic resonance imaging method, comprising: At least one radio frequency coil, i.e., at least one RF coil, of a magnetic resonance imaging system, i.e., an MRI system, is used to obtain input magnetic resonance data, i.e., input MR data. Noise that corrupts the input MR data is obtained using multiple receiver coils outside the MRI system; as well as An MR image is generated, at least in part, from the input MR data using a neural network model designed to suppress at least one noise artifact that corrupts the input MR data, wherein the generation includes: The reconstruction step is performed by generating an image in the image domain based on the input MR data; and Using the neural network model, prior to the reconstruction step, the at least one noise artifact is suppressed by processing the input MR data in a domain other than the image domain.

2. The method according to claim 1, wherein, The at least one noise artifact from the plurality of receiver coils includes RF interference, and the generation includes using the neural network model to suppress the RF interference.

3. The method according to claim 2, wherein, The RF interference includes external RF interference generated by devices located outside the MRI system.

4. The method according to claim 3, wherein, Devices located outside the MRI system include medical devices located in the same room as the MRI system.

5. The method according to claim 2, wherein, The MRI system includes an imaging region, and the RF interference includes internal RF interference generated by at least one component of the MRI system located outside the imaging region.

6. The method according to claim 5, wherein, The at least one component of the MRI system includes one or more magnetic components of the MRI system.

7. The method according to claim 6, wherein, The one or more magnetic components of the MRI system include the gradient coils of the MRI system.

8. The method according to claim 1, wherein, The at least one noise artifact includes noise generated by the circuitry in the MR receiver chain and / or noise generated by the subject or object being imaged.

9. The method according to claim 1, wherein, The neural network model includes a first neural network portion configured to process data in the spatial frequency domain; and Using the neural network model to suppress at least one noise artifact that corrupts the input MR data includes utilizing the first neural network portion to process spatial frequency domain data obtained from the input MR data.

10. The method according to claim 1, wherein, The input MR data is in the sensor domain; The neural network model includes a first neural network portion configured to process data in the sensor domain; as well as Using the neural network model to suppress at least one noise artifact that corrupts the input MR data includes processing the input MR data using the first neural network portion.

11. The method according to claim 1, wherein, The neural network model includes a first neural network portion configured to process input MR data in domains other than the image domain.

12. The method according to claim 11, wherein, The first neural network portion includes a spectral depooling layer, and Processing the input MR data using the first neural network portion includes applying the spectral depooling layer.

13. The method according to claim 11, wherein, The first neural network portion also includes a spectral pooling layer, multiple convolutional layers, and skip connections.

14. The method according to claim 12, wherein, The application of the spectrum depooling layer includes applying a pointwise multiplication layer for combining a first feature having a first resolution provided via a skip connection with a second feature having a second resolution lower than the first resolution.

15. The method according to claim 14, wherein, Applying the spectral depooling layer includes zero-padding the second feature before combining the first feature with the second feature using the pointwise multiplication layer.

16. The method according to claim 1, wherein, The neural network model includes at least one convolutional layer.

17. The method according to claim 1, further comprising: During a first time period, RF artifact measurements are obtained using at least one RF coil of the MRI system, wherein the RF artifact measurements include measurements of RF interference and / or noise; MR measurements of the subject in the imaging region of the MRI system are obtained during a second time period, which is different from the first time period. Artifact-damped MR data is generated by combining the RF artifact measurement results with the MR measurement results of the subject; and The neural network model is trained using the artifact-damaged MR data.

18. The method according to claim 1, further comprising: Synthesized RF artifact measurement results, wherein the RF artifact measurement results include synthesized measurement results of RF interference and / or noise; Obtain MR measurement results of the subject in the imaging region of the MRI system; Artifact-damped MR data is generated by combining the synthetic RF artifact measurements with the MR measurements of the subject; and The neural network model is trained using the artifact-damaged MR data.

19. The method according to claim 1, further comprising: RF artifact measurements are obtained using at least one RF coil of the MRI system, wherein the RF artifact measurements include measurements of RF interference and / or noise; MR measurement results of the synthetic subject MRI system; Artifact-damped MR data are generated by combining the obtained RF artifact measurements with the synthetic MR measurements of the subject; and The neural network model is trained using the artifact-damaged MR data.

20. A magnetic resonance imaging method, comprising: At least one radio frequency (RF) coil of a magnetic resonance imaging (MRI) system is used to obtain input magnetic resonance (MR) data. as well as An MR image is generated from the input MR data, at least in part, by using a neural network model for suppressing RF interference in the input MR data. The generation includes: The reconstruction step is performed by generating an image in the image domain based on the input MR data; and Using the neural network model, prior to the reconstruction step, the RF interference in the input MR data is suppressed by processing the input MR data in domains other than the image domain. The neural network model includes: A first neural network portion, configured to suppress RF interference in the input MR data in a domain other than the image domain prior to the reconstruction step, includes one or more convolutional layers; and A second neural network portion, configured to suppress noise in the input MR data in a domain other than the image domain prior to the reconstruction step, includes one or more convolutional layers.

21. The method according to claim 20, wherein, The neural network model also includes: The third neural network component is configured to suppress noise from image domain data obtained using the input MR data.

22. A magnetic resonance imaging system, comprising: At least one computer hardware processor; as well as At least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by the at least one computer hardware processor, cause the at least one computer hardware processor to perform the following operations: At least one radio frequency coil, i.e., at least one RF coil, of the magnetic resonance imaging system, i.e., the MRI system, is used to obtain input magnetic resonance data, i.e., input MR data. Noise that corrupts the input MR data is obtained using multiple receiver coils outside the MRI system; as well as An MR image is generated, at least in part, from the input MR data using a neural network model designed to suppress at least one noise artifact that corrupts the input MR data, wherein the generation includes: The reconstruction step is performed by generating an image in the image domain based on the input MR data; and Using the neural network model, prior to the reconstruction step, the at least one noise artifact is suppressed by processing the input MR data in a domain other than the image domain.

23. At least one non-transitory computer-readable storage medium storing processor-executable instructions that, when executed by at least one computer hardware processor, cause the at least one computer hardware processor to perform the following operations: At least one radio frequency coil, i.e., at least one RF coil, of a magnetic resonance imaging system, i.e., an MRI system, is used to obtain input magnetic resonance data, i.e., input MR data. Noise that corrupts the input MR data is obtained using multiple receiver coils outside the MRI system; as well as An MR image is generated, at least in part, from the input MR data using a neural network model designed to suppress at least one noise artifact that corrupts the input MR data. The generation includes: The reconstruction step is performed by generating an image in the image domain based on the input MR data; as well as Using the neural network model, prior to the reconstruction step, the at least one noise artifact is suppressed by processing the input MR data in a domain other than the image domain.

24. A magnetic resonance imaging system, i.e., an MRI system, comprising: Magnetic system, including: B0 magnets are configured to provide a B0 field for the MRI system; Gradient coils, which are configured to provide a gradient field for the MRI system; At least one RF coil is configured to detect a magnetic resonance signal, i.e., an MR signal; and At least one receiver coil, configured to detect noise or interference; and The controller is configured as follows: Control the magnetic system to acquire input MR data using the at least one RF coil; and An MR image is generated, at least in part, from the input MR data using a neural network model for suppressing at least one noise artifact in the input MR data, wherein generating the MR image comprises: The reconstruction step is performed by generating an image in the image domain based on the input MR data; and Using the neural network model, prior to the reconstruction step, the at least one noise artifact is suppressed by processing the input MR data in a domain other than the image domain.

25. A computer program product comprising instructions that, when executed by a processor, implement the method of any one of claims 1-21.

Images