Determination of B0 inhomogeneity in magnetic resonance images

Abstract

Disclosed herein is a medical system 100, 300 having a memory 110 storing machine-executable instructions 120 and a convolutional neural network 122 configured to output a number of deblurred magnetic resonance images 126, which are slices of a deblurred magnetic resonance imaging dataset, in response to receiving a group of partially deblurred magnetic resonance images for each of the slices. Execution of the machine-executable instructions causes the computing system 104 to receive (200) a group of partially deblurred magnetic resonance images, receive (202) a predetermined number of deblurred magnetic resonance images in response to inputting the group of partially deblurred magnetic resonance images for each of the slices into a convolutional neural network, calculate (204) a group of difference images 128 for each of the slices by calculating the difference between the deblurred magnetic resonance image and each of the group of partially deblurred magnetic resonance images, and calculate (206) a determined B0 inhomogeneity map 130 for each of the slices by fitting a smooth manifold to B0 values determined from the group of difference images, the demodulation frequency map, and the assigned demodulation frequency for each of the group of difference images.

Description

[Technical Field]

[0001] The present invention relates to magnetic resonance imaging, and in particular to the determination of B0 inhomogeneity. [Background technology]

[0002] In magnetic resonance imaging (MRI) scanners, a large static magnetic field is used to align the nuclear spins of atoms as part of the procedure to generate images of the patient's body. This large static magnetic field is called the B0 field or main magnetic field. The strength of the B0 field and any applied gradient fields determine the frequency at which the spins (usually protons in hydrogen nuclei) precess. As a result of inhomogeneity in the B0 field, protons may precess at a frequency different from the desired one. In that case, the protons or other spins are off-resonance in frequency. A B0 field inhomogeneity map, or equivalently, a frequency off-resonance map, can be measured and used to make corrections during magnetic resonance image reconstruction. However, several difficulties can exist. In some cases, the B0 inhomogeneity map may not be available or may be invalid (e.g., if the subject changes position or moves).

[0003] International Patent Application Publication No. WO 2021 / 197955 discloses a medical system including a memory storing machine-executable instructions and a trained neural network. The trained neural network is configured to output corrected magnetic resonance image data in response to receiving as input a group of magnetic resonance images, each having a different spatially constant frequency resonance off-factor. The medical system further includes a computing system configured to control the medical system, wherein execution of the machine-executable instructions causes the computing system to: receive k-space data acquired according to a magnetic resonance imaging protocol; reconstruct a group of magnetic resonance images according to the magnetic resonance imaging protocol, each of the group of magnetic resonance images being reconstructed assuming a different spatially constant frequency resonance off-factor selected from a list of frequency resonance off-factors; and receive the corrected magnetic resonance image data in response to inputting the group of magnetic resonance images to the trained neural network. Summary of the Invention

[0004] The present invention provides a medical system, a computer program and a method as set forth in the independent claims. Embodiments are set forth in the dependent claims.

[0005] It is known to use neural networks to determine B0 inhomogeneity maps or to correct for B0 inhomogeneity. A difficulty is that neural networks are susceptible to out-of-distribution (OOD) errors or can sometimes result in erroneous data, such as so-called "neural hallucinations." This can be problematic for medical imaging because errors caused by neural networks can result in misleading or inaccurate medical images. Embodiments may provide an improved means of estimating B0 inhomogeneity maps that may be less likely to result in errors. This will be described in the context of a single two-dimensional magnetic resonance image or slice. The following description can be extended to three-dimensional data sets containing multiple slices.

[0006] To accurately estimate the B0 inhomogeneity map (for a single slice or image), a group of partially deblurred magnetic resonance images is created by applying different demodulation frequencies to a single magnetic resonance image. These varying demodulation frequencies have the effect of deblurring the single magnetic resonance image if the demodulation frequency is correct. A convolutional neural network constructs a deblurred magnetic resonance image from the group of partially deblurred magnetic resonance images. Difference images are then calculated by subtracting the deblurred magnetic resonance image from each of the partially deblurred magnetic resonance images, or vice versa. These difference images can then be used to algorithmically determine which or which of the partially deblurred magnetic resonance images provides the correct demodulation frequency. Instead of doing this directly, the B0 inhomogeneity value is determined by fitting a smooth manifold (or smooth surface) to the data derived from the group of difference images, the demodulation frequency map, and the assigned demodulation frequency for each difference image. The smooth surface or manifold fitting has the effect of reducing or eliminating errors caused by the neural network not providing a properly deblurred magnetic resonance image.

[0007] The concept of a partially deblurred image in the framework of the present invention relates to a version of a magnetic resonance image of a particular slice associated with demodulation frequency values specified by the demodulation frequency map of that slice. The demodulation frequency is associated with the demodulation of the acquired magnetic resonance signals (k-space data) that have the Larmor frequency as their radio frequency carrier frequency. This demodulation can be performed before the reconstruction of the magnetic resonance image from the k-space data. Each partially deblurred image has one or more correctly deblurred portions or patches. That is, in these portions or patches, the demodulation frequency values correspond to the actual spatial main magnetic field strength (related to the local (Larmor) radio frequency carrier frequency), which may be spatially inhomogeneous. For each slice, the demodulation frequency map represents the spatially dependent values of the demodulation frequency over the region of the image. The slice-specific demodulation frequency map can be set to a constant, i.e., a flat, uniform map. The demodulation frequency map can vary from slice to slice, in that for each slice, the demodulation frequency map is offset by a slice-specific offset value.

[0008] In one aspect, the present invention provides a medical system having a memory storing machine-executable instructions and a convolutional neural network configured to output a predetermined number of deblurred magnetic resonance images, the number being slices of a deblurred magnetic resonance imaging dataset, in response to receiving a group of partially deblurred magnetic resonance images for each of the slices.

[0009] In magnetic resonance imaging, so-called off-resonance blurring can occur. The deblurred magnetic resonance imaging dataset can be either a single slice or a stack of slices forming a three-dimensional magnetic resonance imaging dataset. Off-resonance blurring in magnetic resonance imaging results from a lack of knowledge of the actual B0 field during the imaging procedure. A baseline B0 field can be measured for a magnetic resonance imaging system and used to correct or deblur the magnetic resonance images. A difficulty with this is that for a particular magnetic resonance imaging protocol, the gradient magnetic fields can induce eddy currents at various locations in the magnetic resonance imaging scanner or magnet. Because these eddy currents can vary with a particular magnetic resonance imaging protocol, correcting for these B0 field inhomogeneities can be very difficult. The approach taken in this example is to provide a set of partially deblurred magnetic resonance images for each slice of the measured magnetic resonance imaging dataset. Each of these images can be prepared by assuming a particular B0 inhomogeneity. The result of performing this process is that the original magnetic resonance imaging slice can have certain regions blurred or deblurred depending on whether the assumptions about the B0 field are correct or incorrect. A convolutional neural network can receive a group of partially deblurred magnetic resonance images and construct a deblurred image from them. Essentially, this is a synthesis of the group of partially deblurred magnetic resonance images. There can be several variations on this.

[0010] In one example, the convolutional neural network operates on only one group at a time. This is equivalent to deblurring only a single slice or two-dimensional data set of the measured magnetic resonance imaging data set. In this case, the number of deblurred magnetic resonance images is simply one. In the case of a three-dimensional magnetic resonance imaging data set consisting of multiple slices, there is a deblurred magnetic resonance image provided for each of these slices. This means that there is a group of partially deblurred magnetic resonance images provided for each slice, and the neural network can generate a deblurred image for each of these groups. This has several advantages over processing a single slice at a time. For example, the convolutional neural network can be trained to receive all of this data simultaneously, and then data from adjacent slices is essentially used to assist in the deblurring of the individual slice. For example, if brain anatomy or other anatomical structures vary from slice to slice but have similarities within various slices, a properly trained convolutional neural network can use this data and perform better deblurring than processing a single slice at a time.

[0011] Using a convolutional neural network that accepts a single slice may have the advantage that it may be easier or more straightforward to train. For example, if only a single slice is processed, it may not even be necessary to use medical imaging data to train it. For example, if a regular optical image is captured from a camera, various parts of this image will be blurred. This can be used as training data, and it may be much simpler to provide this version of the convolutional neural network.

[0012] The medical system further includes a computing system. As used herein, a computing system can take several different forms. In some cases, the computing system can be a remote or virtual computing system provided, for example, as a cloud service. In other examples, the computing system can be a workstation or computer located in a radiology department or other medical facility. In yet other examples, the computing system can be part of a computer or control system for a magnetic resonance imaging system. Execution of the machine-executable instructions causes the computing system to receive a group of partially deblurred magnetic resonance images for each of the slices. Each of the group of partially deblurred magnetic resonance images has an assigned demodulation frequency that specifies an offset in a slice-specific demodulation frequency map. Thus, in this feature, the offset is a demodulation frequency offset, and the slice-specific demodulation frequency map is an assumption regarding which demodulation frequency will be correct. For example, previously measured B0 inhomogeneity measurements can provide an appropriate estimate or starting point for determining B0 inhomogeneity for a particular pulse sequence or a particular activation sequence of a gradient coil system of a magnetic resonance imaging system. Since B0 inhomogeneity is of course spatially varying, a slice-specific demodulation frequency map is thus created by slicing this existing or pre-measured B0 inhomogeneity map, where the offset can be used to move this slice-specific demodulation frequency map in intervals of demodulation frequencies.

[0013] In some cases, there may be no prior knowledge of B0 inhomogeneity, or it may be better to make no assumptions about it. In this case, the slice-specific demodulation frequency map may be set to a constant or near a null value for all slices. In this case, the assigned demodulation frequency for each partially deblurred magnetic resonance image is simply specified by an offset value. In this example, each of a group of partially deblurred magnetic resonance images has an assigned demodulation frequency that specifies a single offset demodulation frequency or value.

[0014] Execution of the machine-executable instructions further causes the computing system to receive a predetermined number of deblurred magnetic resonance images in response to inputting the group of partially deblurred magnetic resonance images for each slice into a convolutional neural network, where the group of partially deblurred magnetic resonance images for each slice are simultaneously input into the neural network and, in response, receive as output the predetermined number of deblurred magnetic resonance images.

[0015] Execution of the machine-executable instructions further causes the computing system to calculate a set of difference images for each slice by calculating the difference between the deblurred magnetic resonance image and each of the set of partially deblurred magnetic resonance images. The difference may be calculated by performing a pixel-by-pixel subtraction. This process is repeated for each slice. Execution of the machine-executable instructions further causes the computing system to calculate the determined B0 inhomogeneity map for each slice by fitting a smooth manifold to values determined from the set of difference images, the demodulation frequency map, and the assigned demodulation frequency for each of the set of difference images.

[0016] Essentially, the set of difference images provides information about which of the partially deblurred magnetic resonance images correctly deblurred the deblurred magnetic resonance imaging data for a particular slice. Knowing this, and then knowing the slice-specific demodulation frequency maps and assigned demodulation frequencies, the specific values of the B0 inhomogeneity map for those regions can be known.

[0017] The smooth manifold fitting provides a very effective means of accurately determining a B0 inhomogeneity map. A difficulty with using convolutional neural networks in medical imaging is that they can provide spurious or erroneous data. For example, certain voxels may provide incorrect values, or regions may be partially incorrect. If a convolutional neural network had to be used, it could be trained to directly output a correctly determined B0 inhomogeneity map for each slice. However, there would be no adequate method for detecting or correcting errors. The technique detailed above is very robust, and the subtraction of a set of difference images and smooth manifold fitting automatically removes small imperfections or inaccuracies in the determined B0 inhomogeneity map for each slice. If only a single slice is present, the smooth manifold can be a smooth surface. If a stack of slices is present, the manifold can be a smoothly varying function in three-dimensional space.

[0018] The smooth manifold can be fitted in various ways. In one example, the smooth manifold is fitted directly to the difference values in the set of difference images, spanning the space where the difference images are smallest. In this case, knowing the demodulation frequency map and the assigned demodulation frequency for each of the set of difference images allows for calculation of the determined B0 inhomogeneity map from the manifold. As another example, on a pixel-by-pixel basis, the set of difference images can be used to select or interpolate the demodulation frequency for each pixel. The smooth manifold can be fitted to these demodulation frequency values or to the B0 inhomogeneity map calculated from the demodulation frequency values. In all of these examples, the smooth manifold can be used to smooth out discontinuities and / or smooth out potential errors.

[0019] In one example, fitting a smooth manifold is solved as an optimization problem in which the values of the set of difference images are minimized along the manifold, subject to constraints such as smoothness and maximum slope that are known from physical knowledge about the properties of the B0 map. For example, a template representing a typical or previous B0 map can be used.

[0020] The convolutional neural network can be implemented, for example, as a U-net. To configure the U-net to receive a set of partially deblurred magnetic resonance images per slice, the number of inputs or encoding branches can be increased to have an input branch or encoding input for each image x number of slices in the set of partially deblurred magnetic resonance images. Similarly, the output can be increased so that there is an output layer or branch for each individual slice. Skip connections can be present between the various outputs or encoding branches of the U-net to share data between them, essentially extending the conventional U-net architecture to provide a convolutional neural network. Instead of using a U-net, a RESNET can be used, which extends the number of inputs and outputs.

[0021] Training a convolutional neural network can be performed in several different ways. When there are multiple slices, perhaps the best approach is to take a magnetic resonance imaging dataset without visible blur and artificially provide a set of partially deblurred magnetic resonance images for each slice. This can be done, for example, by applying blurring kernels with various varying spatial patterns to portions of the image. It can also be achieved by acquiring the original k-space data and artificially resampling it using a simulated B0 inhomogeneity map. In either case, you have a set of blurred images, as well as the original image, either a two-dimensional slice or a full three-dimensional dataset formed from a stack of two-dimensional slices, which can be used as ground truth data. The artificially blurred samples can be input into the convolutional neural network, compared to the original, unblurred image, and a deep learning algorithm can be used to train the neural network. For a convolutional neural network that operates on only a single slice, the training data can be significantly more extensive; for example, a variety of photographic images can be used to train the convolutional neural network. In this example, a photograph is taken and a set of partially deblurred magnetic resonance images is created by locally blurring different regions of the original photograph. A major advantage of this approach is the availability of vast amounts of training data for a variety of tasks.

[0022] In another embodiment, execution of the machine-executable instructions further causes the computing system to determine a demodulation frequency for each voxel of the deblurred magnetic resonance image for each slice using the set of partially deblurred magnetic resonance images, the assigned demodulation frequency for each of the set of difference images, and the demodulation frequency map. A B0 inhomogeneity value is determined from the demodulation frequency for each voxel. In this embodiment, the demodulation frequency for each voxel is determined individually. The deblurred magnetic resonance image can be used to select a value, or a curve can be fitted to all of the values for a particular voxel using all available images, and the best value can be interpolated. However, it should be noted that the specific value of the demodulation frequency is not used directly, but rather a smooth manifold is still used. As previously mentioned, this helps to reduce the impact of errors and inaccuracies caused by using convolutional neural networks. This is highly beneficial for generating medical images because it reduces the likelihood of out-of-distribution errors and artifacts caused by convolutional neural networks.

[0023] In another embodiment, voxels of the deblurred magnetic resonance image having magnitudes below a predetermined magnitude or magnitudes below a predetermined tolerance, at least within a contiguous predetermined volume, are ignored or suppressed from being enhanced during smooth manifold fitting. When viewing a magnetic resonance image, one notices image regions where voxels have very low values. For example, proton-weighted magnetic resonance images show volumes of hydrogen protons or water in a spatially varying manner.

[0024] Outside the subject's body, or where bones are located, this signal will be zero or very small. Since the signal is always small, it is not useful to fit a manifold to this location. Similarly, a region of an image or magnetic resonance image may have a constant value or a value that varies with a certain noise level. If this is the case, the inhomogeneity of the B0 field will not appear. Therefore, in this case, if a voxel has a certain size range (meaning that the voxel varies by a certain amount within a continuous, predetermined volume, which means that at least a certain spatial volume or region exists), this region will also be ignored or suppressed from being enhanced during fitting. This also helps to provide a more accurate estimate of the determined B0 inhomogeneity map.

[0025] In another embodiment, execution of the machine-executable instructions further causes the computing system to receive a single magnetic resonance image for each of the slices. Execution of the machine-executable instructions further causes the computing system to calculate a group of partially deblurred magnetic resonance images for each of the slices by applying off-resonance demodulation determined by the demodulation frequency map and the assigned demodulation frequencies. For each of these calculated images, the demodulation frequency to be used on a voxel-by-voxel basis is determined using the demodulation frequency map and a particular value of the assigned demodulation frequency. The assigned demodulation frequency can be selected from a set of demodulation frequencies, each member of which corresponds to one of the group of partially deblurred magnetic resonance images.

[0026] The off-resonance demodulation can be applied in either image space or k-space, for example, using a demodulation kernel. The demodulation kernel can be the Fourier transform of the phase modulation map in k-space. Demodulation in k-space or image space is mathematically identical and can be done in either image or k-space, but is applied slightly differently. In image space, it is a convolution with a "blurring kernel" (or "deblurring kernel"). In k-space, it is a point-wise multiplication with a phasor (frequency * time at which each point was acquired).

[0027] Execution of the machine-executable instructions further causes the computing system to receive measured k-space data, the measured k-space data having a helical sampling pattern or a non-Cartesian sampling pattern. Execution of the machine-executable instructions further causes the computing system to reconstruct a single magnetic resonance image for each slice in the measured k-space data. This embodiment is particularly useful when using a helical sampling pattern or a non-Cartesian sampling pattern, so the above-described techniques for calculating a determined B0 inhomogeneity map may be beneficial.

[0028] In another embodiment, the medical system further comprises a magnetic resonance imaging system, wherein the memory further comprises pulse sequence commands configured to control the magnetic resonance imaging system to acquire the measured k-space data according to a magnetic resonance imaging protocol. For example, the magnetic resonance imaging protocol may, in some instances, use a helical sampling pattern or a non-Cartesian sampling pattern. Execution of the machine-executable instructions further causes the computing system to acquire the measured k-space data by controlling the magnetic resonance imaging system with the pulse sequence commands.

[0029] In another embodiment, the single magnetic resonance image for each slice is further reconstructed using a previous B0 inhomogeneity map. Execution of the machine-executable instructions further causes the computing system to calculate a corrected B0 inhomogeneity map by correcting the previous B0 inhomogeneity map with the determined B0 inhomogeneity map. In this example, the single magnetic resonance image for each slice is initially corrected using the previous B0 inhomogeneity map. Thus, in this case, some off-resonance blur should be at least partially corrected. At the end of the procedure, a corrected B0 inhomogeneity map is calculated by correcting the previous B0 inhomogeneity map with the determined B0 inhomogeneity map. This process can be beneficial, for example, because for a particular set of pulse sequence commands, eddy currents from acquisition to acquisition are likely to be similar. The corrected B0 inhomogeneity map can be used, for example, for acquisitions using these particular pulse sequence commands.

[0030] In another embodiment, execution of the machine-executable instructions further causes the computing system to calculate a corrected magnetic resonance image using the measured k-space data and the corrected B0 inhomogeneity map. In this example, the corrected B0 inhomogeneity map is used to re-calculate a corrected magnetic resonance image using the corrected B0 inhomogeneity map. There are various ways to examine and calculate a corrected magnetic resonance image. This method returns to the original k-space data and uses the corrected B0 inhomogeneity map. This embodiment may be advantageous because it may provide a more accurate and potentially less blurred magnetic resonance image. The corrected magnetic resonance image in this example may be, for example, a single slice, or may be a stack of slices if the dataset is a three-dimensional dataset.

[0031] In another embodiment, execution of the machine-executable instructions further causes the computing system to acquire additional k-space data by controlling the magnetic resonance imaging system with pulse sequence commands. Execution of the machine-executable instructions further causes the computing system to reconstruct additional magnetic resonance images for each slice using the additional k-space data. The reconstruction of the additional magnetic resonance images is corrected using the corrected B0 inhomogeneity map. In this example, the pulse sequence commands are the same as the pulse sequence commands used during the process of determining the corrected B0 inhomogeneity map. Any eddy currents generated by the gradient magnetic fields are very likely to be similar in different acquisitions. This is most likely to be true for the same acquisition on the same individual at the same location, but even across individuals, the eddy currents will still be very close. This provides a means of determining the correct B0 inhomogeneity map for a particular magnetic resonance imaging pulse sequence command.

[0032] In another embodiment, execution of the machine-executable instructions further causes the computing system to determine a spatially varying demodulation frequency using the determined B0 inhomogeneity map. Execution of the machine-executable instructions further causes the computing system to calculate a corrected magnetic resonance image for each slice by demodulating a single magnetic resonance image with off-resonance frequency demodulation using the spatially varying demodulation frequency. This process can be performed, for example, in either k-space or image space. As previously mentioned, off-resonance demodulation can be applied in either image space or k-space, for example, using a demodulation kernel.

[0033] In another embodiment, the corrected magnetic resonance image for each slice is a motion corrected magnetic resonance image.

[0034] In another embodiment, the corrected magnetic resonance images are cyclic cardiac magnetic resonance images.

[0035] In another embodiment, the corrected magnetic resonance images are respiratory phase-resolved magnetic resonance images.

[0036] In another embodiment, the corrected magnetic resonance image is a diffusion-weighted magnetic resonance image.

[0037] In another embodiment, the corrected magnetic resonance image is a diffusion tensor weighted magnetic resonance image.

[0038] In another embodiment, the corrected magnetic resonance image is an arterial spin-labeling magnetic resonance image.

[0039] The image types mentioned above can benefit from corrected B0 inhomogeneity using a determined B0 inhomogeneity map, as these techniques are particularly sensitive to errors in B0 inhomogeneity.

[0040] In another embodiment, the predetermined number of deblurred magnetic resonance images is 1. In this case, there is only a single group of partially deblurred magnetic resonance images and there is only one slice.

[0041] In other embodiments, the demodulation frequency map has a constant value. This may be a constant number, which may be, for example, zero. In this case, a group of partially deblurred magnetic resonance images is constructed such that only a single demodulation frequency is applied to the entire image. This may be useful, for example, when a preliminary measurement of B0 inhomogeneity is not available or when one does not want to make assumptions about that inhomogeneity.

[0042] In other embodiments, the single magnetic resonance image, the deblurred magnetic resonance image, and the group of partially deblurred magnetic resonance images are three-dimensional or two-dimensional. If they are three-dimensional, this means that there is a stack of slices, as previously described. If the group of partially deblurred magnetic resonance images is only two-dimensional, this means that there is only one group, and only a single magnetic resonance image that is two-dimensional and deblurred.

[0043] In another aspect, the present invention provides a computer program or computer program product including machine-executable instructions for execution by a computing system controlling a medical system. The computer program product may be stored, for example, on a non-transitory storage medium. Execution of the machine-executable instructions causes the computing system to receive a group of partially deblurred magnetic resonance images for each of the slices. Each of the group of partially deblurred magnetic resonance images has an assigned demodulation frequency that specifies an offset and a slice-specific demodulation frequency map. The specified or assigned demodulation frequency may vary spatially within each of the partially deblurred magnetic resonance images.

[0044] Execution of the machine-executable instructions further causes the computing system to receive a predetermined number of deblurred magnetic resonance images in response to inputting the group of partially deblurred magnetic resonance images for each of the slices into a convolutional neural network, the convolutional neural network being configured to output the predetermined number of deblurred magnetic resonance images, the predetermined number being slices of the deblurred magnetic resonance image dataset, in response to receiving the group of partially deblurred magnetic resonance images for each of the slices. Execution of the machine-executable instructions further causes the computing system to calculate a group of difference images for each of the slices by subtracting the deblurred magnetic resonance image from each of the group of partially deblurred magnetic resonance images, or vice versa. Execution of the machine-executable instructions further causes the computing system to calculate the determined B0 inhomogeneity map for each of the slices by fitting a smooth manifold to values determined from the group of difference images, the demodulation frequency map, and the assigned demodulation frequency for each of the group of difference images.

[0045] In another aspect, the present invention provides a method of medical imaging, the method comprising receiving a group of partially deblurred magnetic resonance images for each of a plurality of slices, wherein the plurality of slices are referred to as "slice." Each of the group of partially deblurred magnetic resonance images has an assigned demodulation frequency that specifies an offset of a slice-specific demodulation frequency map. The method further comprises receiving a predetermined number of deblurred magnetic resonance images in response to inputting the group of partially deblurred magnetic resonance images for each of the slices into a convolutional neural network. The convolutional neural network is configured to output the predetermined number of deblurred magnetic resonance images, which are slices of a deblurred magnetic resonance imaging dataset, in response to receiving the group of partially deblurred magnetic resonance images for each of the slices.

[0046] The method further comprises calculating a set of difference images for each of the slices by calculating the difference between the deblurred magnetic resonance image and each of the set of partially deblurred magnetic resonance images. The method further comprises calculating a determined B0 inhomogeneity map for each of the slices by fitting a smooth manifold to values determined from the set of difference images, the demodulation frequency map, and the assigned demodulation frequency for each of the set of difference images.

[0047] It is understood that one or more of the above-described embodiments of the present invention can be combined, unless the combined embodiments are mutually exclusive.

[0048] As will be appreciated by one skilled in the art, aspects of the present invention may be embodied as an apparatus, a method, or a computer program product. Accordingly, aspects of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects, collectively referred to herein as a "circuit," "module," or "system." Furthermore, aspects of the present invention may take the form of a computer program product embodied in one or more computer-readable medium(s) having computer-executable code embodied therein.

[0049] Any combination of one or more computer-readable media may be used. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. As used herein, "computer-readable storage medium" includes any tangible storage medium capable of storing instructions executable by a processor or computing system of a computing device. The computer-readable storage medium may also be referred to as a computer-readable non-transitory 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 store data accessible by the computing system of a computing device. Examples of computer-readable storage media include, but are not limited to, floppy disks, magnetic hard disk drives, solid-state hard disks, flash memory, USB thumb drives, random access memory (RAM), read-only memory (ROM), optical disks, magneto-optical disks, and computing system register files. Examples of optical disks include compact discs (CDs) and digital versatile discs (DVDs), such as CD-ROMs, CD-RWs, CD-Rs, DVD-ROMs, DVD-RWs, and DVD-R discs. The term computer-readable storage medium also refers to various types of storage media that can be accessed by a computer device over a network or communications link. For example, data can be obtained via a modem, over the Internet, or over a local area network. Computer-executable code embodied on a computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, fiber optic cable, RF, etc., or any suitable combination thereof.

[0050] A computer-readable signal medium may include a propagated data signal with computer-executable code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated signal 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 is not a computer-readable storage medium, but may be any computer-readable medium that can communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0051] "Computer memory" or "memory" is one example of a computer-readable storage medium. Computer memory is any memory that is directly accessible by a computing system. "Computer storage" or "storage" is another example of a computer-readable storage medium. Computer storage is any non-volatile computer-readable storage medium. In some embodiments, computer storage is computer memory, or vice versa.

[0052] As used herein, a "computing system" includes an electronic element capable of executing a program, machine-executable instructions, or computer-executable code. References to a computing system, including examples of a "computing system," should be interpreted as possibly including two or more computing systems or processing cores. A computing system may be, for example, a multi-core processor. A computing system may 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 referring to a collection or network of computing devices, each of which includes a processor or computing system. Machine-executable code or instructions may be executed by multiple computing systems or processors within the same computing device or distributed across multiple computing devices.

[0053] Machine-executable instructions or computer-executable code may include instructions or programs that cause a processor or other computing system to perform an aspect of the present invention. Computer-executable code for carrying out processes for aspects of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, Smalltalk®, or C++, and conventional procedural programming languages such as the "C" programming language or similar programming languages, and may be compiled into machine-executable instructions. In some cases, the computer-executable code may be in the form of a high-level language or in a compiled form, and may be used in combination with an interpreter that generates machine-executable instructions on the fly. In other cases, the machine-executable instructions or computer-executable code may be in the form of programming for a programmable logic gate array.

[0054] The computer-executable code may run entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may 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 the connection may be to an external computer (e.g., via the Internet using an Internet Service Provider).

[0055] Aspects of the present invention will be described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block or portion of the blocks in the flowcharts, diagrams, and / or block diagrams, where applicable, can be implemented by computer program instructions in the form of computer-executable code. It will also be understood that combinations of blocks in different flowcharts, diagrams, and / or block diagrams can be combined, where not mutually exclusive. These computer program instructions can be supplied to a general-purpose computer, special-purpose computer, or other programmable data processing device computing system to form a machine, such that the instructions, executed via the computer or other programmable data processing device computing system, generate means for performing the function(s) / acts indicated in the block or blocks of the flowcharts and / or block diagrams.

[0056] These machine-executable instructions or computer program instructions may also be stored on a computer-readable medium that can directly instruct a computer, other programmable data processing apparatus, or other device to function in a specific manner, such that the instructions stored on the computer-readable medium form an article of manufacture including instructions that implement the function / acts specified in a block or blocks of the flowcharts and / or block diagrams.

[0057] The machine-executable instructions or computer program instructions can be loaded into a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be executed on the computer, other programmable apparatus, or other device, thereby generating a computer-implemented process, such that the instructions, when executed on the computer or other programmable apparatus, provide a process for performing the function(s) / act(s) specified in the flowchart and / or block diagram block or blocks.

[0058] 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" is also referred to as a "human interface device." A user interface can provide information or data to an operator or receive information or data from an operator. A user interface can allow input from an operator to be received by a computer and can provide output from the computer to a user. In other words, a user interface allows an operator to control or manipulate a computer, and the interface allows the computer to show the effects of the operator's control or manipulation. Displaying data or information on a display or graphic user interface is an example of providing information to an operator. Receiving data via a keyboard, mouse, trackball, touchpad, pointing stick, graphic tablet, joystick, gamepad, webcam, headset, pedals, wired gloves, remote control, and accelerometer are all examples of user interface elements that allow information or data to be received from an operator.

[0059] As used herein, a "hardware interface" includes an interface that allows a computer system to interact with and / or control external computing devices and / or devices. A hardware interface may allow a computing system to send control signals or commands to external computing devices and / or devices. A hardware interface may also allow a computing system to exchange data with external computing devices and / or devices. Examples of hardware interfaces include, but are not limited to, a universal serial bus, an IEEE 1394 port, a parallel port, an IEEE 1284 port, a serial port, an RS-232 port, an IEEE-488 port, a Bluetooth® connection, a wireless local area network connection, a TCP / IP connection, an Ethernet connection, a control voltage interface, a MIDI interface, an analog input interface, and a digital input interface.

[0060] As used herein, a "display" or "display device" includes an output device or user interface adapted to display images or data. A display can output visual, audio, or tactile data. Examples of displays include, but are not limited to, computer monitors, television screens, touch screens, tactile electronic displays, Braille screens, cathode ray tubes (CRTs), storage tubes, bi-stable 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 displays (OLEDs), projectors, and head-mounted displays.

[0061] K-space data is defined herein as the recorded measurements of radio frequency signals emitted by atomic spins using the antenna of a magnetic resonance machine during a magnetic resonance imaging scan. Magnetic resonance data is an example of tomographic medical image data.

[0062] A magnetic resonance imaging (MRI) image or MR image is defined herein as a reconstructed two-dimensional or three-dimensional visualization of anatomical data contained in magnetic resonance imaging data, which visualization can be performed using a computer. [Brief explanation of the drawings]

[0063] [Figure 1] FIG. 1 shows an example of a medical device. [Figure 2] FIG. 2 is a flow chart illustrating a method of using the medical device of FIG. [Figure 3] FIG. 3 shows an example of a medical device. [Figure 4] FIG. 4 is a flow chart illustrating a method of using the medical device of FIG. [Figure 5] FIG. 5 shows the construction of the determined B0 heterogeneity map. [Figure 6] FIG. 6 shows the construction of a group of difference images of a single slice. [Figure 7] FIG. 7 illustrates one method for determining an appropriate resonance outlier for a particular voxel. [Figure 8] FIG. 8 shows a cross section of a row or column of the difference image. DETAILED DESCRIPTION OF THE INVENTION

[0064] Preferred embodiments of the present invention will now be described, by way of example only, with reference to the drawings in which:

[0065] Like numbered elements in the figures are equivalent elements or perform the same function. In cases of functional equivalence, an element described earlier is not necessarily described in a later figure.

[0066] FIG. 1 illustrates an example of a medical system 100. The medical system 100 in this example is shown as having a computer 102 with a computing system 104. The computer 102 is intended to represent one or more computing devices, which may be co-located or distributed. The computing system 104 is intended to represent one or more computing systems, such as one or more processors or processing cores. The various computing systems 104 may be located in the same location or in different locations. The medical system 100 in this example is intended to represent various configurations; for example, the medical system 100 may be a remote computer in a cloud computing or other network-accessible computing system.

[0067] In another example, the medical system 100 may be located in, for example, a hospital or radiology department. In another example, the medical system 100 is incorporated into a magnetic resonance imaging system. The computer 102 is shown as including a hardware interface 106. The hardware interface may, for example, allow the computing system 104 to communicate with and control other components of the medical system 100, if the medical system also includes a magnetic resonance imaging system. The hardware interface 106 may also represent a network connection that allows the computing system 104 to function at a remote location. The computer 102 is further shown as including an optional user interface 108, which may, for example, allow a user to operate and / or control the medical system 100. The computer 102 is further shown as including memory 110 in communication with the computing system 104. The memory 110 is intended to represent various types of memory accessible by the computing system 104, such as volatile or non-volatile memory and non-transitory storage media.

[0068] The memory 110 is shown as including machine-executable instructions 120. The machine-executable instructions 120 enable the computing system 104 to perform various data processing and control tasks. For example, the computing system 104 can reconstruct magnetic resonance images from k-space data and perform various other data manipulation and image processing tasks. The memory 110 is further shown as including a convolutional neural network 122. The convolutional neural network 122 is configured to output a predetermined number of deblurred magnetic resonance images, each of which is a slice of the deblurred magnetic resonance imaging dataset, in response to receiving a group of partially deblurred magnetic resonance images for each slice. The deblurred magnetic resonance imaging dataset can be, for example, a single slice, in which case the deblurred magnetic resonance imaging dataset is a two-dimensional magnetic resonance image. In other cases, the deblurred magnetic resonance imaging dataset can be a stack of slices forming a three-dimensional magnetic resonance imaging dataset or image.

[0069] The memory 110 is further shown as containing a set of partially deblurred magnetic resonance images for each slice. The memory is further shown as containing a number of deblurred magnetic resonance images 126 received from the convolutional neural network 122 by inputting the set of partially deblurred magnetic resonance images for each slice 124 into the convolutional neural network. The memory 110 is further shown as containing a set of difference images 128 for each slice. The memory 110 is further shown as containing a B0 inhomogeneity map 130 determined for each slice.

[0070] 2 shows a flowchart illustrating a method for operating the medical system 100 of FIG. 1. First, a group of partially deblurred magnetic resonance images 124 for each slice is received. Each image in the group of partially deblurred magnetic resonance images 124 has an assigned demodulation frequency that specifies an offset in a slice-specific demodulation frequency map. Next, in step 202, a predetermined number of deblurred magnetic resonance images 126 are received in response to inputting the group of partially deblurred magnetic resonance images 124 for each slice into a convolutional neural network 122. Then, in step 204, a group of difference images 128 for each slice is calculated by pixel-by-pixel subtraction of the deblurred magnetic resonance image from each image in the group of partially deblurred magnetic resonance images, or vice versa. Finally, in step 206, a determined B0 inhomogeneity map 130 for each slice is calculated by fitting a smooth manifold to the values determined from the set of difference images, the demodulation frequency map, and the assigned demodulation frequency for each of the set of difference images.

[0071] Figure 3 illustrates another example medical system 300. The medical system 300 illustrated in Figure 3 is similar to the medical system 100 of Figure 1, except that it further includes a magnetic resonance imaging system 302 controlled by the computing system 104.

[0072] The magnetic resonance imaging system 302 includes a magnet 304. The magnet 304 is a superconducting cylindrical magnet with a bore 306 extending therethrough. Different types of magnets can be used, including both split cylindrical magnets and so-called open magnets. Split cylindrical magnets are similar to standard cylindrical magnets except that they are split into two sections to allow a cryostat to access the magnet's equal surface. Such magnets can be used, for example, in conjunction with charged particle beam therapy. Open magnets have two magnet sections, one above the other, with enough space between them to accommodate a subject; the arrangement of these two sections is similar to that of a Helmholtz coil. Open magnets are popular because they provide less restraint on the subject. Within the cryostat of the cylindrical magnet is a collection of superconducting coils.

[0073] Within the bore 306 of the cylindrical magnet 304 is an imaging zone 308 where the magnetic field is strong and uniform enough to perform magnetic resonance imaging. A field of view 309 is shown within the imaging zone 308. Acquired k-space data is typically acquired relative to the field of view 309. A region of interest may be identical to the field of view 309 or may be a subvolume of the field of view 309. A subject 318 is shown supported by a subject support 320 such that at least a portion of the subject 318 is located within the imaging zone 308 and the field of view 309.

[0074] Also present within the magnet bore 306 is a set of gradient magnetic field coils 310, which are used to acquire preliminary k-space data for spatially encoding magnetic spins within the imaging zone 308 of the magnet 304. The gradient magnetic field coils 310 are connected to a gradient coil power supply 312. The gradient magnetic field coils 310 are intended to be representative. Typically, the gradient magnetic field coils 310 include three separate sets of coils for spatial encoding in three orthogonal spatial directions. The gradient magnetic field coil power supply supplies current to the gradient magnetic field coils. The current supplied to the gradient magnetic field coils 310 is controlled as a function of time and can be ramped or pulsed.

[0075] Adjacent to the imaging zone 308 is a radio frequency coil 314 for manipulating the orientation of magnetic spins within the imaging zone 308 and for receiving radio frequency transmissions from the spins within the imaging zone 308. The radio frequency coil (radio frequency antenna) can include multiple coil elements. A radio frequency antenna is also referred to as a channel or antenna. The radio frequency coil 314 is connected to a radio frequency transmitter / receiver (transceiver) 316. The radio frequency coil 314 and the radio frequency transceiver 316 can be replaced by separate transmit and receive coils and separate transmitters and receivers. The radio frequency coil 314 and the radio frequency transceiver 316 are understood to be representative. The radio frequency coil 314 is also intended to represent a dedicated transmit antenna and a dedicated receive antenna. Similarly, the transceiver 316 can also represent separate transmitters and receivers. The radio frequency coil 314 can have multiple receive / transmit elements, and the radio frequency transceiver 316 can have multiple receive / transmit channels.

[0076] The transceiver 316 and the gradient controller 312 are shown as being connected to the hardware interface 106 of the computer system 102 .

[0077] The memory 110 is further shown as containing pulse sequence commands 330. Pulse sequence commands are commands, or data that can be converted into commands, that enable the computing system 104 to control the magnetic resonance imaging system 302 to acquire k-space data. The memory 110 is further shown as containing measured k-space data 332, which is acquired by the magnetic resonance imaging system 302 by controlling the imaging system with the pulse sequence commands 330. The memory 110 is further shown as containing a single magnetic resonance image 334 for each slice reconstructed from the measured k-space data 332. The memory 110 is further shown as containing a measured prior B0 inhomogeneity map 336. The memory 110 is further shown as containing a corrected B0 inhomogeneity map 338, which is calculated from the prior B0 inhomogeneity map 336 and the determined B0 inhomogeneity map for each slice 130. It will be understood that references to the B0 inhomogeneity map and / or the corrected B0 inhomogeneity map 338 may refer to individual slices or the inhomogeneity map as a whole. The memory 110 is further shown as including additional k-space data 340, which was also acquired by controlling the magnetic resonance imaging system 302 with pulse sequence commands 330. The memory 110 is further shown as including additional magnetic resonance images for each of the slices 342, which were reconstructed from the additional k-space data 340 and the corrected B0 inhomogeneity map 338, which may be, for example, slice-specific corrected B0 inhomogeneity maps 338.

[0078] FIG. 4 is a flowchart illustrating a method of operating the magnetic resonance imaging system 300 of FIG. 3. The method illustrated in FIG. 4 is similar to the method illustrated in FIG. 2, but includes additional steps. The method begins at step 400, in which the computing system 104 controls the magnetic resonance imaging system with pulse sequence commands 330 to acquire measured k-space data 332. Next, at step 402, the measured k-space data 332 is received. This data may be retrieved, for example, from memory 110. The measured k-space data may have a helical sampling pattern or a non-Cartesian sampling pattern, in some examples. Next, at step 404, a single magnetic resonance image 334 is reconstructed for each of the slices from the measured k-space data 332. Next, at step 406, the single magnetic resonance image 334 for each slice is received. This may involve retrieving from memory 110 or receiving over a network, for example. Next, in step 408, a set of partially deblurred magnetic resonance images 124 for each slice is calculated using off-resonance demodulation at the demodulation frequency set by the demodulation frequency map and the assigned demodulation frequency. The method then proceeds to steps 200, 202, 204, and 206 as shown in FIG. 2. After step 206 is performed, the method proceeds to step 410. In step 410, a corrected B0 inhomogeneity map 338 is calculated for each slice from the previous B0 inhomogeneity map 336 and the determined B0 inhomogeneity map 130. Then, in step 412, the calculation system 104 controls the magnetic resonance imaging system 302 with pulse sequence commands 330 to acquire additional k-space data 340. Finally, in step 414, additional magnetic resonance images for each slice are reconstructed from the additional k-space data 340. During this reconstruction, the corrected B0 inhomogeneity map 338 is used to correct for B0 field inhomogeneity.

[0079] As an example, we could improve existing prototypes that perform "deblurring" of spiral imaging using trained "artificial intelligence" (AI) neural networks. In spiral or other non-Cartesian imaging, off-resonance causes blurring of the reconstructed image. We have shown that, given a set of input images sharpened with uniform frequency, AI can be used to directly generate deblurred images. However, this method risks introducing clinically relevant artifacts. An alternative approach is to use AI to predict off-resonance field maps (i.e., "B0 maps") and then pass these to a "traditional" deblurring algorithm to correct for blurring. This approach is more robust, but requires the AI network to learn a transformation from the image domain to the B0 domain, which increases the risk of obtaining unreliable or inaccurate predictions.

[0080] The example uses a hybrid approach that takes advantage of the advantages of both techniques while avoiding their drawbacks. First, as described above, an AI module (convolutional neural network 122) is used to generate a sharpened image (a predetermined number of deblurred magnetic resonance images 126). This process is a simple task for the AI because it does not require transformation to a new domain. The AI output and the single-frequency deblurred input are then fed into a second algorithm that calculates an "implicit" B0 map. This process avoids the domain transformation required to generate a B0 map directly with AI, while also reducing the risk of network-generated artifacts. The resulting hybrid approach is thus more stable and accurate than either previous method.

[0081] Spiral images often contain blurring due to off-resonance. This is typically corrected using an off-resonance (or "B0") map acquired during a prescan. However, some blurring remains due to eddy currents, inaccuracies in the B0 map, and / or field drift. As part of the development of spiral imaging, we are exploring ways to use AI to sharpen images beyond what can be achieved based on measured B0 maps. Such features are particularly desirable for diffusion imaging, where strong eddy current effects and long readout times result in significant residual blurring.

[0082] Two example configurations are described below. In the first example, the AI network produces a sharpened image as output. In the second example, the network produces a prediction of the off-resonance field, which can then be passed to a "traditional" deblurring algorithm. The advantages and disadvantages of each approach are summarized in the following table.

[0083] [Table 1]

[0084] The present invention utilizes the advantages of image sharpening methods while avoiding the increased risks. It should be noted that the risks mentioned above prevent the solution from being implemented in production due to its predicted reliability.

[0085] The invention may include one or more of the following elements: An AI network that is trained to take a set of one or more blurry input images and produce a sharpened output image.

[0086] A method for predicting an off-resonance map using both the input(s) and output of the network. In a first embodiment, this element does not use AI, but a second AI network can serve this purpose.

[0087] Figure 5 illustrates the construction of a determined B0 inhomogeneity map. Figure 5 shows the processing for a single slice. First, there is a set of partially deblurred magnetic resonance images 124 for this slice. These are then input into a convolutional neural network 122 (in this case a U-net). The U-net outputs a deblurred magnetic resonance image 126. Steps 204 and 206 are then performed to output a determined B0 inhomogeneity map 130.

[0088] The input to the U-net is the initial (blurred) data as a series of frequencies f0 to f N-1 The demodulation is performed based on the known acquisition time of each point in k-space:

number

[0089] where m refers to the initial data and m ~ i refers to the ith demodulated data set. This approach is based on the conjugate phase reconstruction (CPR) algorithm. Each of the N input images has a local resonance offset of f i In this case, U-Net can be trained to generate a sharpened image from the set of partially sharpened inputs.

[0090] The second part of the invention is the generation of an off-resonance map from the inputs and outputs of the network. This process begins by subtracting the output from each input and taking the magnitude of the result, as shown in Figure 6 below.

[0091] 6 shows the construction of a set of difference images 128 for a single slice. A deblurred magnetic resonance image 126 is subtracted from a partially deblurred magnetic resonance image 124 for this slice. Since absolute values are taken in this example, the order of subtraction could in principle be reversed. In this case, the result of the subtraction is a set of difference images 128 for this particular slice. The darkest areas in the image are the areas where the particular deblurred magnetic resonance image best corrects for B0 inhomogeneity.

[0092] These difference images are then fed into an algorithm that generates a smoothly varying off-resonance map. The algorithm calculates the difference image for the i-th image with an off-resonance frequency of f i This is illustrated in the figures below: the top row shows the amplitude difference images, while the bottom row shows where the off-resonance map is closer to fi than any other frequency (in white).

[0093] Figure 7 illustrates one method for determining an appropriate resonance outlier for a particular voxel. In this figure, the top row 128 shows a collection of different difference images. These difference images 128 are used to determine f i A mask point of the nearest bin equal to .

[0094] Another way to look at this is to look at the rows or columns of magnitude difference data "on the edge," i.e., f i The solution is to look at the σ dimension as a plane axis. By doing this, we can see a band of values near zero. The curve shows the true off-resonance frequency for this group of pixels.

[0095] 8 shows a cross section of a row or column of the difference image 128. The plot shows image pixel vs. f i By looking for the pixel closest to zero, the resonance outliers for these voxels can be determined.

[0096] To generate the final B0 map, use any algorithm that meets the following requirements: A certain degree of smoothness and maximum slope is imposed on the output frequency map f(x,y) based on the prediction of the off-resonance map.

[0097] i min (x,y) is the smallest |In(f i ,x,y)-Out(x,y)|, then the output frequency map is |f(x,y)-f_(i min (x,y))| is minimized.

[0098] While the first example is based on the magnitude of the difference between the AI's output and input, any algorithm that uses prior knowledge of both the input and output, as well as the expected properties of the B0 map, can serve this purpose.

[0099] The advantage of this method is that AI is used for a simple task (converting a set of partially clear images into a fully clear image), while the risk of clinically relevant artifacts from the AI is limited by the use of a second step to infer the B0 map.

[0100] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description is to be considered illustrative or exemplary and not restrictive, i.e., the invention is not limited to the disclosed embodiments.

[0101] Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprises" does not exclude other elements or steps, and the singular does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used to advantage. A computer program can be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, as well as distributed in other forms, such as via the Internet or other wired or wireless communication systems. Reference signs in the claims should not be construed as limiting the scope. [Explanation of symbols]

[0102] 100 Healthcare Systems 102 Computer 104 Computing Systems 106 Hardware Interface 108 User Interface 110 memory 120 machine-executable instructions 122 Convolutional Neural Networks 124 A set of partially deblurred magnetic resonance images for each slice 126 A predetermined number of deblurred magnetic resonance images A set of difference images for each of the 128 slices 130 Determined B0 inhomogeneity map for each slice receiving a set of partially deblurred magnetic resonance images for each of the 200 slices; receiving a predetermined number of deblurred magnetic resonance images in response to inputting a set of partially deblurred magnetic resonance images for each of the 202 slices into a convolutional neural network; 204. Compute a set of difference images for each of the slices by calculating the difference between the deblurred magnetic resonance image and each of the set of partially deblurred magnetic resonance images. 206. Calculate the determined B0 inhomogeneity map for each of the slices by fitting a smooth manifold to the values determined from the set of difference images, the demodulation frequency map, and the assigned demodulation frequency for each of the set of difference images. 300 Healthcare Systems 302 Magnetic Resonance Imaging System 304 Magnet 306 Magnet Bore 308 Imaging Zone 309 Field of view 310 Gradient Magnetic Field Coil 312 Gradient magnetic field coil power supply 314 Radio Frequency Coil 316 Transceiver 318 subjects 320 Subject Support 330 Pulse Sequence Commands 332 measured k-space data A single magnetic resonance image for each of the 334 slices 336 Previous B0 heterogeneity maps 338 Corrected B0 inhomogeneity map 340 additional k-space data Additional magnetic resonance images for each of the 342 slices 400 The magnetic resonance imaging system is controlled by pulse sequence commands to obtain measured k-space data. 402 Receive measured k-space data A single magnetic resonance image for each of the 404 slices is reconstructed from the measured k-space data. Receive a single magnetic resonance image for each of the 406 slices. 408. Compute a set of partially deblurred magnetic resonance images by applying off-resonance demodulation at demodulation frequencies set by the demodulation frequency map and the assigned demodulation frequencies. 410. Calculate a corrected B0 inhomogeneity map by correcting the previous B0 inhomogeneity map with the determined B0 inhomogeneity map. 412 Additional k-space data is acquired by controlling the magnetic resonance imaging system with pulse sequence commands. 414 Reconstructing additional magnetic resonance images for each slice using the additional k-space data, wherein the reconstruction of the additional magnetic resonance images is corrected using the corrected B0 inhomogeneity map. 700 f i The mask point of the nearest bin equal to

Claims

1. A memory that stores machine-executable instructions and a convolutional neural network that outputs a predetermined number of deblurred magnetic resonance images, which are slices of a deblurred magnetic resonance imaging dataset, in response to receiving a group of partially deblurred magnetic resonance images for each of the slices, and Computation system A medical system having the execution of the machine-executable instructions in the computing system The group of partially deblurred magnetic resonance images for each of the slices is received, where each of the group of partially deblurred magnetic resonance images has an assigned demodulation frequency that specifies an offset of the slice-specific demodulation frequency map. In response to inputting the group of partially deblurred magnetic resonance images for each of the slices into the convolutional neural network, a predetermined number of deblurred magnetic resonance images are received. A group of difference images for each of the slices is calculated by calculating the difference between the deblurred magnetic resonance image and each of the group of partially deblurred magnetic resonance images, and The determined B0 heterogeneity map for each of the slices is calculated by fitting a smooth manifold to the values ​​determined from the group of difference images, the demodulation frequency map, and the assigned demodulation frequencies for each of the group of difference images. Healthcare system.

2. The medical system according to claim 1, wherein the execution of the machine-executable instructions causes the computing system to determine the demodulation frequency of each voxel in the deblurred magnetic resonance image for each of the slices, using the group of partially deblurred magnetic resonance images, the assigned demodulation frequencies for each of the group of difference images, and the demodulation frequency map, where the B0 heterogeneity value is determined from the demodulation frequency of each voxel.

3. The medical system according to claim 1 or 2, wherein voxels of the deblurred magnetic resonance image having a size less than a predetermined size or a size less than a predetermined tolerance within at least a continuous predetermined volume are ignored or their emphasis is suppressed during the fitting of the smooth manifold.

4. The execution of the aforementioned machine-executable instructions is performed by the computing system, A single magnetic resonance image is received for each of the slices, and The group of partially de-blurred magnetic resonance images is calculated by applying de-resonance demodulation at the demodulation frequency set by the demodulation frequency map and the assigned demodulation frequency. The medical system according to claim 1 or 2.

5. The execution of the aforementioned machine-executable instructions is performed by the computing system, The measured k-space data is received, and the measured k-space data has a helical sampling pattern or a non-Cartesian sampling pattern, and A single magnetic resonance image for each of the slices is reconstructed from the measured k-space data. The medical system according to claim 4.

6. The medical system according to claim 5, further comprising a magnetic resonance imaging system, wherein the memory further includes pulse sequence commands for controlling the magnetic resonance imaging system to acquire the measured k-space data in accordance with a magnetic resonance imaging protocol, and the execution of the machine-executable instructions causes the computing system to acquire the measured k-space data by further controlling the magnetic resonance imaging system with the pulse sequence commands.

7. The medical system according to claim 5, wherein the single magnetic resonance image for each slice is further reconstructed using a previous B0 heterogeneity map, and the execution of the machine-executable instruction causes the computing system to further calculate a corrected B0 heterogeneity map by modifying the previous B0 heterogeneity map with the determined B0 heterogeneity map.

8. The medical system according to claim 7, wherein the execution of the machine-executable instructions causes the computing system to further calculate a corrected magnetic resonance image using the measured k-space data and the corrected B0 heterogeneity map.

9. The execution of the machine-executable instructions further involves the computing system, By controlling the magnetic resonance imaging system with the aforementioned pulse sequence command, additional k-space data is acquired, and An additional magnetic resonance image for each slice is reconstructed using the additional k-space data, wherein the reconstruction of the additional magnetic resonance image is corrected using the corrected B0 heterogeneity map. The medical system according to claim 8.

10. The execution of the machine-executable instructions further involves the computing system, Using the B0 heterogeneity map determined above, the spatially varying demodulation frequency is determined, and The corrected magnetic resonance image for each slice is calculated by demodulating the single magnetic resonance image using out-of-resonance frequency demodulation with the spatially varying demodulation frequency. The medical system according to claim 1 or 2.

11. The medical system according to claim 9, wherein the corrected magnetic resonance image for each slice is one of motion-corrected magnetic resonance images, periodic cardiac magnetic resonance images, respiratory phase-resolved magnetic resonance images, diffusion-weighted magnetic resonance images, diffusion tensor-weighted magnetic resonance images, and arterial spin-labeled magnetic resonance images.

12. The medical system according to claim 1 or 2, wherein the number of predetermined deblurred magnetic resonance images is 1.

13. The medical system according to claim 1 or 2, wherein the demodulation frequency map has a constant value.

14. A computer program comprising machine-executable instructions for execution by a computing system that controls a medical system, wherein the execution of the machine-executable instructions is performed by the computing system, A group of partially deblurred magnetic resonance images is received for each of multiple slices, where each of the group of partially deblurred magnetic resonance images has an assigned demodulation frequency that specifies the offset of the slice-specific demodulation frequency map. In response to inputting the group of partially deblurred magnetic resonance images for each of the slices into a convolutional neural network, a predetermined number of deblurred magnetic resonance images are received, and the convolutional neural network, in response to receiving the group of partially deblurred magnetic resonance images for each of the slices, outputs the predetermined number of deblurred magnetic resonance images, which are slices of the deblurred magnetic resonance imaging dataset. A set of difference images for each of the slices is calculated by calculating the difference between the deblurred magnetic resonance image and each of the set of partially deblurred magnetic resonance images; and The determined B0 heterogeneity map for each of the slices is calculated by fitting a smooth manifold to the values ​​determined from the group of difference images, the demodulation frequency map, and the assigned demodulation frequencies for each of the group of difference images. Computer program.

15. Receiving a group of partially deblurred magnetic resonance images relating to each of one or more slices, wherein each of the group of partially deblurred magnetic resonance images has an assigned demodulation frequency that specifies an offset of the slice-specific demodulation frequency map. A step of receiving a predetermined number of deblurred magnetic resonance images in response to inputting a group of partially deblurred magnetic resonance images for each of the slices into a convolutional neural network, wherein the convolutional neural network outputs the predetermined number of deblurred magnetic resonance images, which are slices of a deblurred magnetic resonance imaging dataset, in response to receiving the group of partially deblurred magnetic resonance images for each of the slices. A step of calculating a group of difference images for each of the slices by calculating the difference between the deblurred magnetic resonance image and each of the group of partially deblurred magnetic resonance images, and The step of calculating the determined B0 heterogeneity map for each of the slices by fitting a smooth manifold to a value determined from the group of difference images, the demodulation frequency map, and the assigned demodulation frequencies for each of the group of difference images. A medical imaging method having [a specific feature].

Images