Reduce off-resonance effects in magnetic resonance imaging

Abstract

This document discloses a medical system including a memory (110) storing machine-executable instructions (120) and a trained neural network (122). The trained neural network is configured to output rectified magnetic resonance image data (130) in response to receiving a set of magnetic resonance images (126) as input, each of the magnetic resonance images in the set having a different spatially constant frequency deviation resonance factor. The medical system also includes a computing system (106) configured to control the medical system, wherein the execution of the machine-executable instructions causes the computing system to: receive (200) k-space data (124) acquired according to a magnetic resonance imaging protocol; reconstruct (202) a set of magnetic resonance images (126) according to the magnetic resonance imaging protocol, wherein each of the magnetic resonance images in the set of magnetic resonance images is reconstructed assuming different spatially constant frequency deviation resonance factors selected from a list of frequency deviation resonance factors (128); and receive (204) rectified magnetic resonance image data in response to inputting the set of magnetic resonance images into the trained neural network.

Description

Technical Field

[0001] This invention relates to magnetic resonance imaging, and more particularly to reducing the effects of deviation from resonance (e.g., blurring). Background Technology

[0002] As part of the process used to generate images of a patient's body, magnetic resonance imaging (MRI) scanners use a large static magnetic field to align the nuclear spins of atoms. This large static magnetic field is called the B0 field or main magnetic field. The strength of the B0 field and any applied gradient magnetic field determine the frequency of spin precession (typically protons in hydrogen nuclei). Inhomogeneities in the B0 field can cause protons to precess at a frequency different from the desired frequency. Protons or other spins then deviate from frequency resonance. A B0 field inhomogeneity map, or an equivalent frequency deviation resonance map, can be measured and used for correction during the reconstruction of MRI images. Several difficulties may arise. In some cases, the B0 inhomogeneity map may be unavailable or invalid, for example, when the object changes position or is moved.

[0003] US Patent Application Publication US 2016 / 0202335 A1 discloses a method for reducing deviated resonance blur in acquired magnetic resonance imaging data. The method includes acquiring a first set of inter-spiral leaf data for each of one or more spiral entry / exit interleavings by performing a first sampling along a first redundant spiral entry / exit trajectory to each of one or more locations in k-space, and acquiring a second set of inter-spiral leaf data for each of one or more spiral entry / exit interleavings by performing a second sampling along a second redundant spiral entry / exit trajectory to each of one or more locations in k-space, wherein the second redundant spiral entry / exit trajectory corresponds to a time-reversed trajectory of the first redundant spiral entry / exit trajectory. The method may further include combining the first and second sets of inter-spiral leaf data with an averaging operation to reduce artifacts. Summary of the Invention

[0004] This invention provides a magnetic resonance imaging system, a computer program product, and a method.

[0005] As described above, inhomogeneities in the B0 magnetic field can lead to deviations from resonance, such as blurring in reconstructed magnetic resonance images. Embodiments can provide a means of reducing or eliminating deviations from resonance, such as blurring, by using a trained neural network. Instead of acquiring a map of the B0 magnetic field inhomogeneities, acquired k-space data is used to reconstruct a set of magnetic resonance images. For each image in the set of magnetic resonance images, a constant frequency deviation from resonance is selected. This is done because it produces a set of magnetic resonance images that are blurred except in the regions where the deviation from resonance is correct. The trained neural network is trained to take the set of magnetic resonance images as input and then output corrected magnetic resonance image data.

[0006] The trained neural network can be incorporated into the computing system of the magnetic resonance imaging system. Alternatively, the computing system can be provided with access to the trained neural network located remotely from the physical hardware of the magnetic resonance imaging system, such as on a local server at a medical institution, or the trained neural network can be in the cloud.

[0007] The magnetic resonance imaging system is configured to reconstruct the set of magnetic resonance images when reconstruction software is installed on a computing system or when the computing system has access to a remote reconstruction facility. The reconstruction software can be installed on a remote server, for example, in a medical institution with access to a data network, as the reconstruction software can be used in the "cloud." In these remote configurations, the computing system is equipped with the capability to reconstruct the set of magnetic resonance images at a remotely located reconstruction facility.

[0008] The neural network can be trained or configured in several different ways. In some examples, the corrected magnetic resonance image data is a reconstructed magnetic resonance image. The neural network outputs a fully reconstructed image. In other examples, the neural network can output a voxel map. The voxel map indicates the selection or choice of each voxel in the voxel map of one of the magnetic resonance images in the set of magnetic resonance images. This provides several things. Voxel maps can, for example, be used to assemble composite magnetic resonance images by using voxel maps to indicate which images should be used to supply the values ​​of a particular voxel in the composite magnetic resonance image. Since each of the magnetic resonance images in the set has a constant frequency deviation resonance factor, voxel maps can be used to assemble frequency deviation resonance maps (or equivalently, B0 magnetic field inhomogeneity maps).

[0009] Another potential benefit of trained neural networks is that they operate in image space and may be independent of the k-space sampling pattern. When the k-space sampling pattern is spiral, the implementation may be particularly beneficial in correcting blur artifacts. However, trained neural networks can be used with a wide variety of sampling patterns.

[0010] This invention provides a medical system, a method, and a computer program, and provides embodiments.

[0011] In one aspect, the present invention provides a medical system comprising a memory storing machine-executable instructions and a trained neural network. The trained neural network may, for example, be a convolutional neural network. The trained neural network is configured to output rectified magnetic resonance image data in response to receiving a set of magnetic resonance images as input, each magnetic resonance image having a different spatially constant frequency deviation resonance factor. In magnetic resonance imaging, the so-called B0, or main magnetic field, may have inhomogeneities. Differences in the B0 magnetic field cause protons to resonate with a slight deviation from the resonance factor. Therefore, the frequency deviation resonance factor corresponds to a deviation from the ideal value of the B0 field.

[0012] The medical system also includes a computing system configured to control the medical system. The execution of machine-executable instructions causes the computing system to receive k-space data acquired according to a magnetic resonance imaging (MRI) protocol. The execution of the machine-executable instructions also causes the computing system to reconstruct a set of MRI images according to the MRI protocol. Assuming different spatially constant frequency deviation resonance factors are selected from a list of frequency deviation resonance factors, each of the set of MRI images is reconstructed.

[0013] If the B0 map or frequency deviation resonance factor map is known, this can be used to directly reconstruct magnetic resonance images from k-space data. However, this may be unknown, or the data may be corrupted. Instead of providing the B0 map or frequency deviation resonance factor map or mapping, k-space data is reconstructed by assuming different frequency deviation resonance factors for each reconstructed image. Image portions with actual frequency deviation resonance factors close to a constant frequency deviation resonance factor will be substantially deblurred or sharpened in that region. A trained neural network can therefore be trained to look at a set of input magnetic resonance images and select regions with the correct frequency deviation resonance factors. This can be useful in directly reconstructing magnetic resonance images or using it to reconstruct B0 maps or frequency deviation resonance factor maps.

[0014] The execution of the machine-executable instructions also causes the computing system to receive corrected magnetic resonance image data in response to inputting the set of magnetic resonance images into a trained neural network. This embodiment can be advantageous because the neural network is capable of correcting for deviations from the resonance factor or B0 inhomogeneity without a pre-measured B0 map or frequency deviation map. For example, this can be useful in avoiding the acquisition of the B0 map or even correcting it if it is corrupted. For instance, the B0 map can be measured, and then the object can be moved when k-space data is acquired. In this case, the B0 map will no longer be valid. The embodiment provides a method for correcting B0 inhomogeneity without measuring the B0 map.

[0015] In another embodiment, the rectified magnetic resonance imaging data includes a magnetic resonance image corrected for inhomogeneity. In this embodiment, the rectified magnetic resonance imaging data is directly reconstructed into a rectified image.

[0016] In another embodiment, the output corrected magnetic resonance image data is complex-valued. This could mean that the output of the neural network corresponds to voxels, and for each voxel there are two components: either real and imaginary, or amplitude and phase.

[0017] In another embodiment, the rectified magnetic resonance images include a pixel map. Each set of magnetic resonance images has a voxel matrix of the same size. The voxel mapping involves selecting one image from a set of magnetic resonance images for each voxel in the same-sized voxel matrix. In this embodiment, the rectified magnetic resonance imaging data includes a graph that identifies which pixels from which the image is substantially deblurred or sharpened. This can have several different uses. For example, an image can be reconstructed by acquiring appropriate pixels using the pixel map. In other examples, the pixel map can be converted into a B0 inhomogeneity map. This, for example, allows it to first use k-space data and reconstruct a B0 map, and then use a conventional magnetic resonance imaging algorithm to reconstruct an image using that B0 inhomogeneity map.

[0018] In another embodiment, the voxel matrix of each set of magnetic resonance images has the same size. Essentially, these images are spatially consistent. The pixel mapping can have a mapping for each voxel in the voxel matrix.

[0019] In another embodiment, the execution of the machine-executable instructions further enables the computing system to assemble a composite magnetic resonance image by selecting voxels from the set of magnetic resonance images according to a pixel map. In this embodiment, the composite magnetic resonance image is assembled by selecting voxels identified by the pixel map. For example, this allows the computing system to assemble the composite magnetic resonance image using the best available voxels with the least blur.

[0020] In another embodiment, the execution of the machine-executable instructions further enables the computing system to assemble B0 inhomogeneity maps and / or frequency deviation resonance maps by assigning constant frequency deviation resonance factors from the magnetic resonance image set based on voxels. When each set of magnetic resonance images is reconstructed, a specific frequency deviation resonance factor is assigned and used for reconstruction. Knowledge of the voxel mapping and these frequency deviation resonance factors can be used to reconstruct the B0 inhomogeneity map.

[0021] In other embodiments, spatial filters or smoothing algorithms are applied to the B0 non-uniformity map.

[0022] In another embodiment, the execution of the machine-executable instructions also enables the computing system to reconstruct a B0 inhomogeneity-corrected magnetic resonance image from k-space data and a B0 inhomogeneity map or frequency deviation resonance map. In this embodiment, instead of stitching images together or having the neural network directly output the resulting image, a B0 inhomogeneity map determined by the neural network and using voxel mapping is applied within a conventional magnetic resonance imaging algorithm. This, for example, allows for great system flexibility. The system can be substantially used to reconstruct images regardless of the type of magnetic resonance imaging protocol chosen. The neural network is simply trained to reconstruct the B0 inhomogeneity map or its equivalent, and then the reconstruction is applied using a normal algorithm.

[0023] In another embodiment, the machine-executable instructions further instruct the computing system to assemble rectified magnetic resonance image data by applying a trained neural network to portions of the magnetic resonance image using a spatial sliding window algorithm. A particular challenge in using neural networks is assuming specific input and output sizes for the data. In this embodiment, the neural network can be trained to take an image smaller than the actual size of the magnetic resonance image as input. Then, to reconstruct the complete image, the spatial sliding window algorithm divides the complete image into several parts and computes the resulting image for each part. Taking these multiple images and combining them into a larger image is then relatively straightforward.

[0024] In another embodiment, the machine-executable instructions are configured to cause a computing system to assemble rectified magnetic resonance image data by applying a trained neural network to subgroups of a set of magnetic resonance images. In some cases, the number of layers in a magnetic resonance image may vary. Therefore, having a trained neural network that can accept any number of layers may be impractical. One approach to this problem is to have a trained neural network and then divide the data into subgroups and process these subgroups separately using the trained neural network. For example, if the trained neural network can accept both voxels with three layers and magnetic resonance images with nine layers, the magnetic resonance images can first be divided into three groups of three layers each, yielding one result. These three results can then be fed back into the trained neural network to obtain the final result.

[0025] In another embodiment, the magnetic resonance imaging protocol is a parallel imaging magnetic resonance imaging protocol. The trained neural network operates in image space, so the underlying scheme used for image acquisition can technically work in many different situations.

[0026] In another embodiment, each set of magnetic resonance images is complex-valued. Each voxel in the set of magnetic resonance images has a complex value. This can be represented by two components: either real and imaginary parts, or amplitude and phase.

[0027] In another embodiment, the trained neural network is a U-net convolutional neural network. U-net convolutional neural networks are highly effective in processing medical images. In U-net, there are contraction and expansion paths. This is why it has a U-shaped structure and is named as such. U-net is effective for medical image processing because it allows for comparison of correlations on both small and large scales.

[0028] In another embodiment, the medical system further includes a magnetic resonance imaging (MRI) system configured to acquire k-space data from an imaging region. The memory also contains pulse sequence commands configured to control the MRI system to acquire k-space data. The execution of the machine-executable instructions further enables the computing system to control the MRI system to acquire k-space data.

[0029] In another embodiment, the k-space data has a non-Cartesian sampling mode.

[0030] In another embodiment, the k-space data has a helical sampling pattern. This embodiment may be advantageous because helical imaging patterns in k-space are particularly prone to deviating from resonance effects. This can provide a means of correcting magnetic resonance images with helical imaging patterns in k-space.

[0031] In another aspect, the present invention provides a method for training a neural network. The method includes configuring the topology of the neural network to receive a set of magnetic resonance images as input, each magnetic resonance image having a different spatially constant frequency deviation resonance factor. This may, for example, include configuring the neural network to receive a set of magnetic resonance images of the same size. The method also includes configuring the topology of the neural network to output corrected magnetic resonance image data. The method further includes receiving training data. The training data includes datasets comprising deblurred magnetic resonance images reconstructed using spatially varied frequency deviation resonance maps and multiple magnetic resonance images, each image reconstructed using a frequency deviation resonance map with different spatial constraints. In each of these datasets, there are magnetic resonance images reconstructed using the correct spatially varied frequency deviation resonance maps.

[0032] Multiple training magnetic resonance images are then reconstructed from the same k-space data, but instead of using spatially varied frequency deviations from the resonance map, spatially constant values ​​are applied to the deviation frequencies for each. The training data can then be used, for example, to train a neural network using a deep learning scheme. The method also includes generating a trained neural network by repeatedly training the network with each dataset.

[0033] In another aspect, the present invention provides a computer program comprising machine-executable instructions for execution by a computing system controlling a medical system. The execution of the machine-executable instructions causes the computing system to receive k-space data acquired according to a magnetic resonance imaging (MRI) protocol. The execution of the machine-executable instructions also causes the computing system to reconstruct a set of magnetic resonance images according to the MRI protocol. Each of the set of MRI images is reconstructed, assuming different spatially constant frequency deviation resonance factors selected from a list of frequency deviation resonance factors. The execution of the machine-executable instructions further causes the computing system to receive calibrated MRI image data in response to inputting the set of MRI images into a trained neural network. The trained neural network is configured to output calibrated MRI image data in response to receiving a set of MRI images as input, each MRI image having a different spatially constant frequency deviation resonance factor.

[0034] It should be understood that one or more of the foregoing embodiments of the present invention may be combined, as long as the combined embodiments are not mutually exclusive.

[0035] As those skilled in the art will recognize, several aspects of the invention can be implemented as apparatus, method, or computer program product. Therefore, aspects of the invention can take the form of entirely hardware embodiments, entirely software embodiments (including firmware, resident software, microcode, etc.), or embodiments combining software and hardware aspects, which can be collectively referred to herein as “circuit,” “module,” or “system.” Furthermore, aspects of the invention can take the form of computer program products implemented in one or more computer-readable media having computer-executable code implemented thereon.

[0036] Any combination of one or more computer-readable media can be used. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. As used herein, "computer-readable storage medium" includes any tangible storage medium that can store instructions executable by a processor or computing system of a computing device. The computer-readable storage medium may be referred to as a "computer-readable non-transient storage medium." The computer-readable storage medium may also be referred to as a tangible computer-readable medium. In some embodiments, the computer-readable storage medium may also be able to store data accessible by the computing system of the computing device. Examples of computer-readable storage media include, but are not limited to: floppy disks, magnetic hard disk drives, solid-state drives, flash memory, USB thumb drives, random access memory (RAM), read-only memory (ROM), optical disks, magneto-optical disks, and register files of computing systems. Examples of optical disks include compact optical disks (CDs) and digital multi-purpose optical disks (DVDs), such as CD-ROMs, CD-RWs, CD-Rs, DVD-ROMs, DVD-RWs, or DVD-R discs. The term computer-readable storage medium also refers to various types of recording media accessible by the computer device via a network or communication link. For example, data can be retrieved via a modem, via the Internet, or via a local area network. Computer-executable code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, or any suitable combination of the foregoing.

[0037] Computer-readable signal media may include propagated data signals having computer-executable code implemented therein, for example, in baseband or as part of a carrier wave. Such propagated signals may take any of a variety of forms, including, but not limited to, electromagnetic, optical, or any suitable combination thereof. A computer-readable signal medium may be any computer-readable medium that is not a computer-readable storage medium and is capable of transmitting, propagating, or conveying a program for use by or in connection with an instruction execution system, apparatus, or device.

[0038] "Computer memory" or "memory" is an example of a computer-readable storage medium. Computer memory is any memory that can be directly accessed by a computing system. "Computer storage device" or "storage device" is another example of a computer-readable storage medium. A computer storage device is any non-volatile computer-readable storage medium. In some embodiments, a computer storage device may also be computer memory, or vice versa.

[0039] As used herein, "computing system" encompasses electronic components capable of executing programs or machine-executable instructions or computer-executable code. References to computing systems that include examples of "computing systems" should be interpreted as potentially including more than one computing system or processing core. A computing system can, for example, be a multi-core processor. A computing system can also refer to a collection of computing systems within a single computer system or distributed across multiple computer systems. The term computing system should also be interpreted as potentially referring to a collection or network of computing devices, each including a processor or multiple computing systems. Machine-executable code or instructions can be executed by multiple computing systems or processors, which may be within the same computing device or even distributed across multiple computing devices.

[0040] Machine-executable instructions or computer-executable code may include instructions or programs that cause a processor or other computing system to perform one aspect of the invention. Computer-executable code for performing operations targeting the aspects of the invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, Smalltalk, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages, and compiled into machine-executable instructions. In some cases, the computer-executable code may be used in the form of a high-level language or in a pre-compiled form in conjunction with an interpreter that generates machine-executable instructions in flight. In other cases, the machine-executable instructions or computer-executable code may be in the form of programming against a programmable gate array.

[0041] The computer-executable code can run as a standalone software package entirely on the user's computer, partially on the user's computer, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer via any type of network (including a local area network (LAN) or a wide area network (WAN)) or via a connection to an external computer (e.g., via the Internet using an Internet service provider).

[0042] Various aspects of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block or portion of a block in a flowchart, illustration, and / or block diagram can be implemented, where applicable, by computer program instructions in the form of computer-executable code. It should also be understood that combinations of blocks in different flowcharts, illustrations, and / or block diagrams can be combined when not mutually exclusive. These computer program instructions can be provided to the memory of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine, such that instructions executable via the computer's memory or other programmable data processing apparatus create units for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0043] These machine-executable instructions or computer program instructions may also be stored in a computer-readable medium that is capable of directing a computer, other programmable data processing apparatus or other device to operate in a particular manner, such that the instructions stored in the computer-readable medium produce an article of writing comprising instructions that implement the functions / actions specified in flowcharts and / or one or more block diagrams.

[0044] The machine-executable instructions or computer program instructions may also be loaded onto a computer, other programmable data processing apparatus or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer-implemented process, such that the instructions running on the computer or other programmable apparatus provide for implementing the functions / actions specified in the flowchart and / or one or more block diagram boxes.

[0045] As used herein, a "user interface" is an interface that allows a user or operator to interact with a computer or computer system. A "user interface" can also be referred to as a "human-machine interface device." A user interface can provide information or data to an operator and / or receive information or data from an operator. A user interface enables input from an operator to be received by the computer and can provide output from the computer to the user. In other words, a user interface allows an operator to control or manipulate a computer, and the interface allows the computer to indicate the effects of the operator's control or manipulation. The display of data or information on a monitor or graphical user interface is an example of providing information to an operator. Receiving data via a keyboard, mouse, trackball, touchpad, pointing stick, graphics tablet, joystick, gamepad, webcam, helmet, pedal, wired gloves, remote control, and accelerometer are all examples of user interface components that implement the receiving of information or data from an operator.

[0046] As used herein, "hardware interface" encompasses any interface that enables a computer system to interact with or control external computing devices and / or apparatuses. A hardware interface allows a computing system to send control signals or instructions to external computing devices and / or apparatuses. A hardware interface also enables a computing system to exchange data with external computing devices and / or apparatuses. Examples of hardware interfaces include, but are not limited to: Universal Serial Bus (USB), IEEE 1394 port, parallel port, IEEE 1284 port, serial port, RS-232 port, IEEE-488 port, Bluetooth connectivity, wireless LAN connectivity, TCP / IP connectivity, Ethernet connectivity, control voltage interface, MIDI interface, analog input interface, and digital input interface.

[0047] As used herein, the terms "display" or "display device" encompass output devices or user interfaces suitable for displaying images or data. Displays can output visual, audio, and tactile data. Examples of displays include, but are not limited to: computer monitors, television screens, touchscreens, haptic electronic displays, Braille screens, cathode ray tubes (CRTs), memory tubes, bistable displays, electronic paper, vector displays, flat panel displays, vacuum fluorescent displays (VFs), light-emitting diode (LED) displays, electroluminescent displays (ELDs), plasma display panels (PDPs), liquid crystal displays (LCDs), organic light-emitting diode (OLED) displays, projectors, and head-mounted displays.

[0048] k-space data are defined in this paper as recorded measurements of radio frequency signals emitted by atomic spins via the antenna of a magnetic resonance imaging (MRI) device during a magnetic resonance imaging (MRI) scan. Magnetic resonance data is an example of tomographic medical image data.

[0049] Magnetic resonance imaging (MRI) images, or MR images, are defined herein as reconstructed two-dimensional or three-dimensional visualizations of anatomical data contained within magnetic resonance imaging data. Such visualizations can be performed using a computer. Attached Figure Description

[0050] In the following description, preferred embodiments of the invention will be illustrated by way of example only and with reference to the accompanying drawings, in which:

[0051] Figure 1 An example of a medical system is illustrated;

[0052] Figure 2 A flowchart illustrating an example of a method for operating a medical system is shown;

[0053] Figure 3 This illustration shows another example of a medical system;

[0054] Figure 4The illustration shows a method for reconstructing a magnetic resonance image using a set of magnetic resonance images;

[0055] Figure 5 The diagram illustrates the training of a neural network;

[0056] Figure 6 This is an example of a magnetic resonance image reconstructed using the B0 magnetic field inhomogeneity map;

[0057] Figure 7 A magnetic resonance image with blur artifacts is shown;

[0058] Figure 8 The image shown is a magnetic resonance image reconstructed using a trained neural network, which can correct... Figure 7 Blurred artifacts in the image;

[0059] Figure 9 The diagram showing the B0 magnetic field inhomogeneity predicted by the neural network is presented; and

[0060] Figure 10 The diagram illustrates the reconstruction. Figure 6 The measured B0 magnetic field inhomogeneity diagram.

[0061] List of reference numerals

[0062] 100 Medical Systems

[0063] 102 Computer

[0064] 104 Hardware Interfaces

[0065] 106 Computing System

[0066] 108 User Interface

[0067] 110 Memory

[0068] 120 Machine-executable instructions

[0069] 122 Trained Neural Network

[0070] 124k spatial data

[0071] A collection of 126 magnetic resonance images

[0072] List of 128 frequency deviations from resonance factors

[0073] 130 Corrected magnetic resonance imaging data

[0074] 200 Receives k-space data acquired according to the magnetic resonance imaging protocol

[0075] 202 Reconstructing a set of magnetic resonance images according to magnetic resonance imaging protocols

[0076] 204 Receiving calibrated magnetic resonance image data in response to inputting the set of magnetic resonance images into a trained neural network.

[0077] 300 Medical System

[0078] 302 Magnetic Resonance Imaging System

[0079] 304 magnet

[0080] 306 Magnet Chamber

[0081] 308 Imaging Area

[0082] 309 Areas of Interest

[0083] 310 Magnetic Gradient Coil

[0084] 312 Magnetic Gradient Coil Power Supply

[0085] 314 RF coil

[0086] 316 transceiver

[0087] 318 Objects

[0088] 320 Object Support

[0089] 330 Pulse Sequence Command

[0090] 400 First Image

[0091] 402 Second Image

[0092] 404 Composite Image

[0093] Dataset with 500 training data points

[0094] 502 Deblurred MRI Images

[0095] 504+ training MRI images

[0096] 506 Clear image areas

[0097] 600 original magnetic resonance images

[0098] Original image after 700-degree spiral blur

[0099] 702 Blurred Area

[0100] 800 deblurred images output by a trained neural network

[0101] 900 Predicted B0 Inhomogeneity Plot

[0102] B0 non-uniformity map measured at 1000 Detailed Implementation

[0103] In these figures, similarly numbered elements are equivalent elements or perform the same function. If the functions are equivalent, elements that have been discussed previously will not necessarily be discussed in later figures.

[0104] Figure 1 An example of a medical system 100 is illustrated. The medical imaging system 100 is shown as including a computer 102. The computer 102 is shown as including a computing system 106 connected to a hardware interface 104, a user interface 108, and a memory 110. The computing system 106 is intended to represent one or more processing units, which may also be distributed among multiple computers 102. For example, if the medical system 100 includes a magnetic resonance imaging system, the hardware interface 104 may be used by the computing system 106, for example, to control external or additional components of the medical system 100. For example, the memory may be any type of memory accessible to the computing system.

[0105] Computer 102 may also be a virtual machine or other machine located in the cloud or in a remote location and used for computationally intensive tasks.

[0106] The memory is shown to contain machine-executable instructions 120. These machine-executable instructions can, for example, be executed by a computing system 106. The machine-executable instructions, for example, enable the computing system to control components of the medical system 100 and perform various data and image analysis functions.

[0107] Memory 110 is also shown to include a trained neural network 122. The trained neural network 122 has been trained or configured to output rectified magnetic resonance image data in response to receiving a set of magnetic resonance images as input, each magnetic resonance image having a different spatially constant frequency deviation resonance factor. Memory 110 is also shown to contain k-space data 124. Memory 110 is also shown to contain a set of magnetic resonance images 126, which have been reconstructed from the k-space data 124 by assuming different spatially constant frequency deviation resonance factors selected from a list of frequency deviation resonance factors 128. Memory 110 is also shown to include rectified magnetic resonance image data 130 received from the trained neural network 122 in response to a set of input magnetic resonance images 126.

[0108] Figure 2 The illustrated operation is shown. Figure 1The flowchart describes a method for a medical system 100. First, in step 200, k-space data 124 is received. Next, in step 202, a set of magnetic resonance images 126 is reconstructed according to a magnetic resonance imaging protocol. It is assumed that different spatially constant frequency deviation resonance factors are selected from a list of frequency deviation resonance factors 128 to reconstruct each image in the set of magnetic resonance images. Finally, in step 204, calibrated magnetic resonance image data 130 is received in response to inputting the set of magnetic resonance images 126 into a trained neural network 122.

[0109] Figure 3 Another example of a medical system 300 is illustrated. Figure 3 The medical system depicted in the text is similar to 300 Figure 1 The medical system 100 includes, in addition to the medical system, a magnetic resonance imaging system 302.

[0110] The magnetic resonance imaging system 302 includes a magnet 304. Magnet 304 is a superconducting cylindrical magnet with a bore 306 passing through it. Different types of magnets are also possible; for example, split cylindrical magnets and so-called open magnets can also be used. A split cylindrical magnet is similar to a standard cylindrical magnet, except that the cryostat has been split into two parts to allow access to the isoplanar surface of the magnet, thus allowing the magnet to be used, for example, in conjunction with charged particle beam therapy. An open magnet has two magnet sections, one on top of the other, with a space in between large enough to accommodate the object: the arrangement of the two sections is similar to that of a Helmholtz coil. Open magnets are popular because the object is less restricted. An assembly of superconducting coils is located inside the cryostat of the cylindrical magnet.

[0111] Within the bore 306 of the cylindrical magnet 304, an imaging region 308 exists, in which the magnetic field is sufficiently strong and uniform to perform magnetic resonance imaging. A region of interest 309 within the imaging region 308 is shown. The acquired magnetic resonance data is typically acquired for the region of interest. An object 318 is shown supported by an object support 320, such that at least a portion of the object 318 is within both the imaging region 308 and the region of interest 309.

[0112] The magnet chamber 306 also contains an assembly of magnetic field gradient coils 310, which are used to acquire primary magnetic resonance data for spatial encoding of magnetic spins within the imaging region 308 of the magnet 304. The magnetic field gradient coils 310 are connected to a magnetic field gradient coil power supply 312. The magnetic field gradient coils 310 are intended to be representative. Typically, the magnetic field gradient coil 310 comprises an assembly of three discrete coils for spatial encoding in three orthogonal spatial directions. The magnetic field gradient power supply provides current to the magnetic field gradient coils. The current supplied to the magnetic field gradient coils 310 is time-controlled and can be either slanted or pulsed.

[0113] Adjacent to the imaging region 308 is an RF coil 314, which is used to manipulate the orientation of the magnetic spins within the imaging region 308 and to receive RF transmissions from spins also located within the imaging region 308. The RF antenna may comprise multiple coil elements. The RF antenna may also be referred to as a channel or antenna. The RF coil 314 is connected to an RF transceiver 316. The RF coil 314 and the RF transceiver 316 may be replaced by separate transmit and receive coils, and separate transmitters and receivers. It is to be understood that the RF coil 314 and the RF transceiver 316 are representative. The RF coil 314 is intended to also represent a dedicated transmit antenna and a dedicated receive antenna. Similarly, the transceiver 316 may also represent separate transmitters and receivers. The RF coil 314 may also have multiple receive / transmit elements, and the RF transceiver 316 may have multiple receive / transmit channels. For example, if performing a parallel imaging technique such as SENSE, the RF coil 314 may have multiple coil elements.

[0114] Transceiver 316 and gradient controller 312 are shown as hardware interface 106 connected to computer system 102.

[0115] The memory 110 is also shown to include pulse sequence commands 330. Pulse sequence commands are commands or data that can be converted into commands or data that enable the computing system 106 to control the magnetic resonance imaging system 302 to acquire k-space data 124.

[0116] As mentioned above, non-Cartesian MR imaging techniques (such as helical imaging) are prone to deviating resonance effects, resulting in significant blur artifacts. If the spatial deviating resonance is quantitatively known (e.g., by acquiring a B0 field map), this effect can be corrected during reconstruction. However, this requires acquiring additional MR preparation scans, and scan parameters and eddy currents can affect its accuracy.

[0117] Conjugate phase reconstruction (CPR) for non-Cartesian MRI

[0118] The MR signal s(t) obtained under the condition of deviating from resonance Δω(r) can be written as:

[0119] [1]

[0120] Where m(r) represents the proton density and It is a k-space trajectory. In conventionally reconstructed images, deviations from resonance can lead to artifacts, such as spiral blurring, derived from:

[0121] [2]

[0122] in, This represents the weighting factor used to compensate for sampling density and k-space traversal speed. Blurring can be eliminated by conjugate phase reconstruction (CPR), where the acquired signal is multiplied by the conjugate phase factor (1):

[0123] [3]

[0124] The premise is that the deviation from resonance can be known from the additional B0 field diagram.

[0125] However, CPR is computationally expensive because the integral must be solved for each pixel based on its own deviation resonance Δω®.

[0126] A faster alternative to the precise CPR expressed by the above approximation is proposed, such as frequency-segmented CPR. For frequency-segmented CPR, the integral is approximated by the superposition of integrals with a fixed demodulation frequency:

[0127] [4]

[0128] For a selected demodulation frequency set, the integral term can be efficiently solved using gridding and FFT. Interpolator Basically, the demodulation frequency that is closest to the offset resonance of the selected pixel is chosen. Another method of frequency-segmented CPR is time-segmented CPR, in which the integral in equation [3] is decomposed into small time intervals, using a constant offset resonance.

[0129] U-NET (3) is a convolutional neural network (CNN) topology that has been proposed for biomedical image segmentation tasks.

[0130] The network consists of contraction paths and expansion paths, forming a U-shaped architecture.

[0131] The shrinking path is a typical convolutional network, composed of repeatedly applied convolutions, each followed by a rectified linear unit (ReLU) and max pooling operation. During shrinking, spatial information decreases while feature information increases. The expanding path combines the features and spatial information with the high-resolution features from the shrinking path through a series of up-convolutions and connections.

[0132] As mentioned above, CPR requires a B0 field map. Acquiring the field map requires additional time and must be repeated to compensate for resonant frequency drift. Furthermore, the field map may degrade due to spatial phase variations caused by eddies and is susceptible to motion and other confounding factors.

[0133] Examples can be made by applying a properly configured multidimensional neural network (e.g., U-net) (trained neural network 122) to make decisions among many possible potential local deviation resonances, which is the best way to make the image look sharp, or in other words, to deblur the image without knowing the field map.

[0134] Examples can be found in many applications of non-Cartesian MR imaging, such as spiral MR imaging.

[0135] To perform frequency-segmented CPR, the blurred spiral image must first be demodulated using a set of demodulation frequencies that cover the actual field map range with granularity, resulting in a multi-frequency dataset (a set of MRI images 126). In the second step, each pixel is acquired from the demodulated image deviating from the resonance at a demodulation frequency equal to or close to the selected pixel. If no field map is available, this may also involve dividing the image into small regions (patches) and selecting a modulation frequency for each patch to minimize blurring within that patch (see below). Figure 4 This is simple, but very complicated, and can be performed by a trained neural network (see...). Figure 5 Different demodulation frequencies correspond to the input channels of the network. The output is a deblurred image. In an improved implementation, the field map can be stored as a complex phase of the output image (mapping the frequency range ±fMax to ±PI).

[0136] Figure 4 The diagram illustrates how an image is deblurred. Two images represent a set of magnetic resonance images 126. Image 404 below is a composite image made from the first image 400 and the second image 402. Both images 400 have unblurred, sharp image regions 506. The composite image 404 is a synthesis made from two image regions 506 from each of images 400 and 402. The deblurred image 404 can be obtained by manually selecting sharp regions (ellipses 506) from a multi-frequency dataset to combine with the unblurred image 404. The data for this diagram was synthesized using a Shepp-Logan phantom and assumes random quadratic deviations from resonance in the x and y directions.

[0137] Figure 5The diagram illustrates a method for training a trained neural network 122. In this example, neural network 122 is a U-net. A dataset of training data 500 exists, comprising deblurred magnetic resonance images 502 and multiple training magnetic resonance images 504. Neural network 122 can be trained using a repeated dataset of training data 500 and a deep learning training scheme for deep networks. In use, the frequency-modulated spiral image is used as the input channel, and the complex deblurred images (including field maps in the complex phase) represent the two output channels (real and imaginary parts). A Shepp-Logan phantom is used to synthesize the data, and random quadratic deviation resonance terms in the x and y directions are assumed.

[0138] The Shepp-Logan head (software) phantom is used to synthesize helical images blurred due to deviated resonance. Field maps with constant, linear, and quadratic deviated resonance terms are used to generate the blur.

[0139] To improve the robustness of learning, data augmentation is performed by rotating and scaling the model and changing the field plot mode.

[0140] Synthetic spiral Shepp-Logan images (256x256 image size) are generated and used to train U-Net for deblurring due to off-resonance. U-Net consists of four downsampling steps (each preceded by two convolutional steps) and corresponding upsampling steps followed by convolutions. Stochastic gradient descent using least squares is used as the optimizer. Data augmentation is performed as described above to improve the robustness of the learning process.

[0141] Assuming a spiral trajectory (50 ms duration) and a random field map (maximum ±20 Hz), the in vivo brain image is blurred. The blurred brain image is used as input to train the network. Figure 3 The image shown includes the original image, as well as blurred and predicted (deblurred) images. Furthermore, the predicted field map is compared with the underlying field map used for data synthesis. Figure 1 The predicted anatomical image and field map closely match the unblurred image and the base field map. Note that the base field map is only used to synthesize the blurred image and is therefore not explicitly fed into the U-NET.

[0142] Figure 6-10 Used to illustrate the effectiveness of the method. Figure 6 The original magnetic resonance image 600 is shown. Figure 6 The image in the image is reconstructed using the measured B0 inhomogeneity map, as follows: Figure 10 As shown.

[0143] Figure 7 The image 700 depicts a blurred magnetic resonance image. It can be seen that the area circled in 702 is somewhat blurred. This is due to the use of a spiral k-space sampling mode.

[0144] Figure 8 The image shown is a deblurred magnetic resonance image 800 deblurred using a trained neural network. It can be seen that... Figure 7 The blurred area 702 is now in Figure 8 The middle part was correctly reconstructed.

[0145] Neural networks can also be used to compute the predicted B0 non-uniformity map 900. The actual measured B0 non-uniformity map is shown below. Figure 10 As shown, it is marked as 1000. It can be seen that the two non-uniformity maps are similar.

[0146] Although the invention has been illustrated and described in detail in the accompanying drawings and the foregoing description, such illustrations and descriptions should be considered illustrative or exemplary, and not restrictive. The invention is not limited to the disclosed embodiments.

[0147] Those skilled in the art, through studying the accompanying drawings, disclosure, and claims, will understand and implement other variations of the disclosed embodiments when practicing the claimed invention. In the claims, the word "comprising" does not exclude other elements or steps, and the words "a" or "an" do not exclude a plurality. A single computing system processor or other unit can implement the functions of several items recited in the claims. Although specific measures are recited in dissimilar dependent claims, this does not imply that combinations of these measures cannot be advantageously used. Computer programs can be stored / distributed on suitable media such as optical storage media or solid-state media provided with or as part of other hardware, but can also be distributed in other forms such as via the Internet or other wired or wireless telecommunications systems. Any reference numerals in the claims should not be construed as limiting the scope.

Claims

1. A medical system comprising: A memory (110) is provided for storing machine-executable instructions (120) and a trained neural network (122), wherein the trained neural network is configured to output corrected magnetic resonance image data (130) in response to receiving a set of magnetic resonance images (126) as input, each of the set of magnetic resonance images being reconstructed from k-space data assuming different spatially constant frequency deviations from the resonance factor. A computing system (106) configured to control the medical system, wherein the execution of machine-executable instructions enables the computing system to: Receive the k-space data (124) acquired according to the magnetic resonance imaging protocol. According to the magnetic resonance imaging protocol, a set of reconstructed (202) magnetic resonance images (126) are arranged, wherein each magnetic resonance image in the set of magnetic resonance images is reconstructed by assuming that different spatially constant frequency deviation resonance factors are selected from a list of frequency deviation resonance factors (128); and In response to inputting the set of magnetic resonance images into the trained neural network, rectified magnetic resonance image data is received. The trained neural network is trained using a method including the following operations: The topology of the neural network is configured to receive a set of magnetic resonance images as input, each magnetic resonance image having a different spatially constant frequency deviation resonance factor; The topology of the neural network is configured to output calibrated magnetic resonance image data; Receive training data, wherein the training data includes a dataset comprising a deblurred magnetic resonance image (502) reconstructed from a k-space dataset using spatially varied frequency deviation resonograms and multiple training images (504), each training image being reconstructed from the same k-space dataset using a different spatially constant frequency deviation resonogram; and A trained neural network is generated by repeatedly training the neural network using each dataset in the dataset.

2. The medical system according to claim 1, wherein, The corrected magnetic resonance imaging data includes magnetic resonance images corrected for inhomogeneity.

3. The medical system according to claim 1 or 2, wherein, The corrected magnetic resonance image data is complex-valued.

4. The medical system according to claim 1 or 2, wherein, The corrected magnetic resonance imaging data includes voxel mapping, wherein each magnetic resonance image in the set of magnetic resonance images has a voxel matrix of the same size, and wherein the voxel mapping includes selecting one magnetic resonance image from the set of magnetic resonance images for each voxel of the same size voxel matrix.

5. The medical system according to claim 4, wherein, The execution of the machine-executable instructions also enables the computing system to assemble and synthesize magnetic resonance images by selecting voxels from the set of magnetic resonance images according to the pixel mapping (404).

6. The medical system according to claim 4, wherein, The execution of the machine-executable instructions also enables the computing system to assemble B0 inhomogeneity maps (900) and / or frequency deviation resonance maps by assigning constant frequency deviation resonance factors to the set of magnetic resonance images according to the voxel mapping.

7. The medical system according to claim 6, wherein, The execution of the machine-executable instructions also enables the computing system to reconstruct B0 inhomogeneity-corrected magnetic resonance images from k-space data and B0 inhomogeneity maps or frequency deviation resonance maps.

8. The medical system according to claim 1 or 2, wherein, The machine-executable instructions also enable the computing system to assemble the corrected magnetic resonance image data by applying the trained neural network to portions of the set of magnetic resonance images using a spatial sliding window algorithm.

9. The medical system according to claim 1 or 2, wherein, The machine-executable instructions are configured to cause the computing system to assemble the corrected magnetic resonance image data by applying the trained neural network to a subgroup of the set of magnetic resonance images.

10. The medical system according to claim 1 or 2, wherein, Each magnetic resonance image in the set of magnetic resonance images is complex-valued.

11. The medical system according to claim 1 or 2, wherein, The medical system further includes a magnetic resonance imaging system (302) configured to acquire k-space data from an imaging region (308), wherein the memory further contains pulse sequence commands configured to control the magnetic resonance imaging system to acquire the k-space data, wherein the execution of the machine-executable instructions further enables the computing system to control the magnetic resonance imaging system to acquire the k-space data.

12. The medical system according to claim 1 or 2, wherein, The k-space data has a non-Cartesian sampling mode.

13. The medical system according to claim 12, wherein, The k-space data has a spiral sampling mode.

14. A method for training a neural network, wherein, The method includes: The topology of the neural network is configured to receive a set of magnetic resonance images as input, each magnetic resonance image having a different spatially constant frequency deviation resonance factor; The topology of the neural network is configured to output calibrated magnetic resonance image data; Receive training data, wherein the training data includes a dataset comprising a deblurred magnetic resonance image (502) reconstructed from a k-space dataset using spatially varied frequency deviation resonograms and multiple training images (504), each training image being reconstructed from the same k-space dataset using a different spatially constant frequency deviation resonogram; and A trained neural network is generated by repeatedly training the neural network using each dataset in the dataset.

15. A computer program comprising machine-executable instructions (120) for execution by a computing system (106) controlling a medical system (100, 300), wherein, The execution of the machine-executable instructions enables the computing system to: Receive k-space data (124) acquired according to the magnetic resonance imaging protocol. A set of magnetic resonance images (126) is reconstructed according to a magnetic resonance imaging protocol (202), wherein each magnetic resonance image in the set of magnetic resonance images is reconstructed by assuming that different spatially constant frequency deviation resonance factors are selected from a list of frequency deviation resonance factors (128); and In response to inputting the set of magnetic resonance images into a trained neural network, rectified magnetic resonance image data is received (130), wherein the trained neural network is configured to output the rectified magnetic resonance image data in response to receiving a set of magnetic resonance images as input, each of the set of magnetic resonance images having been reconstructed based on k-space data assuming different spatially constant frequency deviations from the resonance factor. The trained neural network is trained using a method including the following operations: The topology of the neural network is configured to receive a set of magnetic resonance images as input, each magnetic resonance image having a different spatially constant frequency deviation resonance factor; The topology of the neural network is configured to output calibrated magnetic resonance image data; Receive training data, wherein the training data includes a dataset comprising a deblurred magnetic resonance image (502) reconstructed from a k-space dataset using spatially varied frequency deviation resonograms and multiple training images (504), each training image being reconstructed from the same k-space dataset using a different spatially constant frequency deviation resonogram; and A trained neural network is generated by repeatedly training the neural network using each dataset in the dataset.

Images