Automatic detection of water-fat exchange in Dixon magnetic resonance imaging

Abstract

A medical system (100, 300, 500) is disclosed herein, comprising a memory (110) storing machine executable instructions (120) and a convolutional neural network (122). The convolutional neural network is configured to receive an initial Dixon magnetic resonance image (124, 126) as input. The convolutional neural network is configured to identify one or more water-fat exchange regions (128) in the initial Dixon magnetic resonance image. The medical system also includes a processor (104) for controlling the medical system. Execution of the machine executable instructions causes the processor to: receive (200) the initial Dixon magnetic resonance image; and receive (204) the one or more water-fat exchange regions from the convolutional neural network in response to inputting the initial Dixon magnetic resonance image into the convolutional neural network.

Description

Technical Field

[0001] The present invention relates to magnetic resonance imaging, and in particular to the Dixon method of magnetic resonance imaging. Background Art

[0002] As part of the process for producing images within a patient's body, a large static magnetic field is used by a magnetic resonance imaging (MRI) scanner to align the nuclear spins of atoms. This large static magnetic field is called the B0 field.

[0003] During an MRI scan, radio frequency (RF) pulses generated by a transmitter coil cause disturbances to the local magnetic field, and the RF signals emitted by the nuclear spins are detected by a receiver coil. These RF signals are used to construct MRI images. These coils may also be referred to as antennas. In addition, the transmitter and receiver coils may also be integrated into a single transceiver coil that performs both functions. It should be understood that the use of the term transceiver coil also refers to systems in which separate transmitter and receiver coils are used. The transmitted RF field is referred to as the B1 field.

[0004] MRI scanners can construct images of either slices or volumes. A slice is a thin volume that is only one voxel thick. A voxel is a small volume over which the MRI signal is averaged and represents the resolution of the MRI image. A voxel may also be referred to herein as a pixel.

[0005] The Dixon method of magnetic resonance imaging comprises a family of techniques for producing separated water and lipid (fat) images. Various Dixon techniques, such as but not limited to the two-point Dixon method, the three-point Dixon method, the four-point Dixon method, and the six-point Dixon method, are collectively referred to herein as Dixon techniques or methods. There are often multiple solutions when reconstructing water and fat images. It is possible to inadvertently switch voxels between water and fat images; this is referred to herein as water-fat exchange.

[0006] US patent application publication US20170082713A1 discloses a method and apparatus for automatically assigning at least one combined image of an object under examination to a spin species represented in the combined image, wherein a relationship determined from an existing database and involving the assignment of spin species to the combined image is loaded into a computer. At least two MR data sets at one of at least two echo times are provided to the computer in each case after excitation by means of a multi-contrast measurement. At least one combined image is determined in the computer from the at least two MR data sets. The spin species represented in the at least one combined image are assigned in the computer based on the loaded relationship. By using the relationship determined from the existing database, an automatic unambiguous global assignment of the correct spin species is enabled.

[0007] The article "Correction of fat-water swaps in Dixon MRI" by B. Glocker et al. (in MICCAI 2016) addresses the problem of fat-water swaps in Dixon MRI. This corresponds to the goal of the present invention. However, this paper proposes a solution (Dixon-Fix) based on regression methods and learning the conditional distribution of image appearance. Summary of the invention

[0008] The present invention provides a medical system, a computer program product and a method.In the following description, embodiments are given.

[0009] A difficulty in reconstructing fat images and water images during a Dixon magnetic resonance imaging protocol is water-fat exchange. The water-fat exchange of a water-fat exchange area is an area or voxel group in which a water area and a fat area have been misidentified within a Dixon magnetic resonance image (or Dixon image). An embodiment provides a method for automatically identifying one or more water-fat exchanges using a convolutional neural network. One or more water-fat exchanges are identifications of voxels in a magnetic resonance image. One or more water-fat exchanges can be given, for example, in the form of voxel positions. More conveniently, a mask can be used to identify one or more water-fat exchanges. A mask is an image of the location of one or more water-fat exchanges that can be identified. If there is no water-fat exchange, the mask can be empty and / or indicate that there is no water-fat exchange. Reference to a mask in this article refers to an image that can be used to identify the location of a water-fat exchange in the presence of a water-fat exchange in a Dixon magnetic resonance image.

[0010] The convolutional neural network can identify the location of the one or more water-fat exchange regions to aid human interpretation or correction of the Dixon image. The convolutional neural network can also be incorporated into an automatic algorithm that iteratively removes the water-fat exchange regions from the Dixon image.

[0011] The present invention provides a medical system including a memory storing machine executable instructions and a convolutional neural network. The convolutional neural network is configured or trained to receive an initial Dixon magnetic resonance image as input. The convolutional neural network is configured to identify one or more water-fat exchange regions in the initial Dixon magnetic resonance image.

[0012] The initial Dixon magnetic resonance image as used herein includes a fat and / or water image and / or a fat fraction image and / or an R2* image reconstructed according to the Dixon magnetic resonance imaging protocol. Water-fat exchange will be visible in any of these images reconstructed using the Dixon method. The convolutional neural network is configured to take different types of inputs. The input of the convolutional neural network can be, for example, any of the following: a water Dixon image, a fat Dixon image, or both a water Dixon image and a fat Dixon image, a fat fraction image, an R2* image, and a combination thereof. Similarly, it is possible to identify water-fat exchange areas by looking only at the water Dixon image, similarly as it is possible to identify it by looking only at the fat Dixon image. If the convolutional neural network takes both paired water and fat Dixon magnetic resonance images as input for training, it may be easier to train and have higher accuracy. However, this is not necessary, because the convolutional neural network can be trained to also look only at water or only fat images.

[0013] A medical system as used herein includes a system configured for performing image processing and / or numerical techniques on medical data or medical imaging data.

[0014] The medical system also includes a processor configured to control the medical system. The execution of the machine executable instructions causes the processor to receive an initial Dixon magnetic resonance image. The operation of the machine executable instructions also causes the processor to receive the one or more water-fat exchange regions from the convolutional neural network in response to inputting the initial Dixon magnetic resonance image into the convolutional neural network. This embodiment may be beneficial because it may provide an automatic method for identifying water-fat exchange regions in a Dixon magnetic resonance image. For example, this may be used for quality control of Dixon magnetic resonance images and incorporated into an automatic algorithm for reconstructing Dixon magnetic resonance images.

[0015] In another embodiment, execution of the machine executable instructions further causes the processor to reconstruct a corrected Dixon magnetic resonance image using the received one or more water-fat exchange regions. This embodiment may be beneficial because it may provide a method of improving the quality of a Dixon magnetic resonance image.

[0016] In another embodiment, the Dixon magnetic resonance image is corrected by exchanging voxels of one or more water-fat exchange regions between the fat image and the water image. The Dixon magnetic resonance image comprises the fat image and the water image. In this example, which is particularly suitable for a two-point Dixon method, the voxels identified as the water-fat exchange regions are simply exchanged with each other. This can provide a simple method of providing a corrected Dixon magnetic resonance image.

[0017] In another embodiment, the Dixon magnetic resonance image is corrected by reconstructing the Dixon magnetic resonance image from Dixon magnetic resonance imaging data according to a Dixon magnetic resonance imaging reconstruction algorithm. The Dixon magnetic resonance imaging reconstruction algorithm is configured to use the one or more water-fat exchange regions as constraints and / or for correcting the B0 inhomogeneity estimate. In the presence of constraints, the algorithm may be constrained to correctly identify the regions as water or fat when determining a solution. Likewise, using the one or more water-fat exchange regions to correct the B0 inhomogeneity estimate may also provide a corrected Dixon magnetic resonance image.

[0018] The Dixon MRI data is MRI data acquired according to a Dixon MRI protocol and can be reconstructed into a water-fat image according to the Dixon MRI protocol.

[0019] In another embodiment, the execution of the machine executable instructions further causes the processor to receive training data. Each of the training data includes a training Dixon magnetic resonance image and a training water-fat exchange mask. The training water-fat exchange mask can identify the ground truth of the one or more water-fat exchange areas. Some of the training Dixon magnetic resonance images will not have water-fat exchange, and their masks can indicate this. The execution of the machine executable instructions also causes the processor to train the convolutional neural network using the training data. Similarly, in this embodiment, the training Dixon magnetic resonance image can simply be a water image, a fat image, or a combination of water and fat images. In this way, the convolutional neural network can be trained to use one image as input or two images. If there is a single image as input, it may be only a water or fat image. If both, then a pair of matching water and fat images are input.

[0020] In another embodiment, the operation of the machine executable instructions also causes the processor to receive non-swap Dixon magnetic resonance images. The non-swap Dixon magnetic resonance images are correct Dixon magnetic resonance images, wherein they do not have regions that cause their water and fat regions to switch. In addition, the non-swap Dixon magnetic resonance images can be only non-swap water Dixon magnetic resonance images, non-swap fat Dixon magnetic resonance images, or a matching pair of water and fat images. The operation of the machine executable instructions also causes the processor to generate the training data by generating synthetic water-fat exchange regions in each of the non-swap Dixon magnetic resonance images in the non-swap Dixon magnetic resonance images and constructing the training water-fat exchange mask. This embodiment may be beneficial because it starts with an image that is guaranteed not to have any water-fat exchange regions. In order to make the training data, the algorithm then manually adds these water-fat exchange regions. This can be performed in various ways, for example, the regions can be simply exchanged or it can be done according to any of the following algorithms described in detail below.

[0021] In another embodiment, the training data is generated at least in part by first determining one or more random variables, then determining at least some of the one or more synthetic water-fat exchange regions by inputting the one or more random variables into a spatially variable function, and then thresholding the spatially variable function to determine one or more water-fat exchange regions for training Dixon magnetic resonance images. In this embodiment, the locations of the water-fat exchange regions are randomly determined.

[0022] In another embodiment, the training data is generated at least in part by identifying high water-fat exchange areas in a Dixon magnetic resonance image having both fat content and water content above a predetermined water-fat threshold. The training data is generated at least in part by selecting at least part of the high water-fat area as the synthetic water-fat exchange area. The high water-fat area is an area or voxel containing both high water and fat content. If the voxel contains equal parts of water and fat, there will be 50% water and 50% fat. In this case, the Dixon magnetic resonance image is more likely to have errors when it is reconstructed by an automatic algorithm. When using an automatic algorithm, these voxels with both high water and fat content may then result in water-fat exchange. Identifying these high water-fat areas and then selecting at least part of them is a way to generate realistic training water-fat exchange areas.

[0023] For example, a voxel with a non-negligible content of both water and fat indicating a content of both components above 30% may qualify. This can occur, for example, at tissue boundaries or in conditions such as fatty liver disease.

[0024] In another embodiment, the training data is generated at least in part by receiving a B0 inhomogeneity map for the non-exchange Dixon magnetic resonance image. For example, when acquiring magnetic resonance data for the non-exchange Dixon magnetic resonance image, the B0 inhomogeneity map may also be stored and used at this time. The training data is also generated at least in part by thresholding the B0 inhomogeneity map using a predetermined B0 inhomogeneity threshold to identify high B0 inhomogeneity regions in the non-exchange Dixon magnetic resonance image. The training data is also generated at least in part by selecting at least part of the high B0 inhomogeneity regions as the training water-fat exchange regions. This embodiment may be beneficial because in regions with higher B0 inhomogeneity, an automatic algorithm is more likely to misidentify regions as water or fat.

[0025] This can also be thought of as the probability of exchange that a risk factor for development or a particular region might be exchanged in a real Dixon MRI. These artificial or synthetic exchange regions can then be generated based on this map or the B0 inhomogeneity map. This can result in a more realistic training data set than using, for example, random regions.

[0026] In another embodiment, the training data is generated at least in part by identifying tissue boundary regions in the non-exchange Dixon magnetic resonance images using an image segmentation algorithm. For example, standard image segmentation techniques can be used to identify different tissue regions. The training data is generated at least in part by selecting at least part of the tissue boundary regions or regions surrounding these as the training water-fat exchange regions. This embodiment can also be further beneficial because in regions adjacent to transitions between tissue types, it may also be possible or likely that the automatic Dixon magnetic resonance imaging algorithm incorrectly identifies regions as water or fat. This is because there may be strong local signal gradients that may make it more difficult to correctly solve the equations and distinguish between fat and water regions.

[0027] In another embodiment, the training data is generated at least in part by determining a spatially correlated signal-to-noise ratio map for the Dixon magnetic resonance image. The training data is also generated at least in part by selecting at least a portion of a training water-fat exchange region to include in the training data set using a signal-to-noise ratio weighting factor determined using the spatially correlated signal-to-noise ratio map. The signal-to-noise ratio weighting factor increases as the spatially correlated signal-to-noise ratio map decreases. For example, the lower the signal-to-noise ratio, the more likely water-fat exchange is occurring in an automated Dixon magnetic resonance imaging algorithm.

[0028] In another embodiment, the training data is generated at least in part by receiving an isocenter position for the non-exchange Dixon magnetic resonance image. The training data is also generated at least in part by selecting at least a portion of a synthetic water-fat exchange region for inclusion in the training data using a distance weighting factor determined by the distance from the isocenter position. The distance weighting factor increases with increasing distance from the isocenter position. Generally, as the distance from the isocenter increases, water-fat exchange is more likely to increase. This is often associated with an increase in B0 inhomogeneity. However, using the distance from the isocenter as a weighting factor can enable better selection of regions to generate synthetic water-fat exchange regions.

[0029] In another embodiment, the training data is generated at least in part by increasing the size of the training data by applying an image transform to the training Dixon magnetic resonance images and a training water-fat exchange mask for each of the training Dixon magnetic resonance images to generate additional training magnetic resonance images. The image transform includes any of: an image translation, an image rotation, an image deformation, an image flip transform, a mirror transform, and combinations thereof. This may be a way to increase the training database or data used to train the convolutional neural network. This may enable training with a smaller amount of training data.

[0030] The medical imaging system also includes a magnetic resonance imaging system. The memory also includes pulse sequence commands configured to control the magnetic resonance imaging system to acquire magnetic resonance imaging data according to the Dixon magnetic resonance imaging protocol. The execution of the machine executable instructions also causes the processor to control the magnetic resonance imaging system to acquire the magnetic resonance image by using the pulse sequence commands.

[0031] Execution of the machine executable instructions further causes the processor to reconstruct the initial Dixon magnetic resonance image using the magnetic resonance imaging data according to the Dixon magnetic resonance imaging protocol. This embodiment may be beneficial because when the magnetic resonance imaging data is acquired by the magnetic resonance imaging system, the convolutional neural network is used to automatically identify the one or more water-fat exchange regions and can, for example, be used to automatically improve the quality of the Dixon magnetic resonance image.

[0032] In the above embodiments, the medical system is a magnetic resonance imaging system. Embodiments may also be implemented in other types of systems. For example, in one embodiment, the medical system is an online service that provides correction of fat and / or water images as part of image reconstruction software. In another embodiment, the medical system is a retrospective water-fat exchange correction tool that can be executed, for example, on a diagnostic workstation. In yet another example, the medical system may be an exchange correction method that may be independent of a specific supplier or manufacturer of the medical system. For example, this may be implemented or sold as a cloud-based service that may be provided.

[0033] In another aspect, the present invention provides a computer program product comprising machine executable instructions for being executed by a processor controlling the medical system and a convolutional neural network. The convolutional neural network is configured to receive an initial Dixon magnetic resonance image as input. In addition, the initial Dixon magnetic resonance image can be a fat Dixon magnetic resonance image, a water Dixon magnetic resonance image, or a matched pair of fat and water Dixon magnetic resonance images. The convolutional neural network is configured to identify one or more water-fat exchange regions in the initial Dixon magnetic resonance image. The execution of the machine executable instructions also causes the processor to receive the initial Dixon magnetic resonance image. The execution of the machine executable instructions also causes the processor to receive the one or more water-fat exchange regions from the convolutional neural network in response to inputting the initial Dixon magnetic resonance image into the convolutional neural network. The advantages of the system have been previously discussed.

[0034] In another aspect, the present invention provides a method for training a convolutional neural network. The convolutional neural network is configured to receive an initial Dixon magnetic resonance image as input. The convolutional neural network is configured to identify or is trained to identify one or more water-fat exchange regions in the initial Dixon magnetic resonance image. The method includes receiving training data. Each of the training data includes a training Dixon magnetic resonance image and a training water-fat exchange mask. The training Dixon magnetic resonance image includes one or more water-fat exchange regions. The training water-fat exchange mask identifies standard data for the one or more water-fat exchange regions.

[0035] The method further comprises training the convolutional neural network using the training data. As previously mentioned above, the training Dixon MRI images may be, for example, water Dixon MRI images, fat Dixon MRI images, or may be a pair of matched images, one of which is a water Dixon MRI image and the other is a fat Dixon MRI image.

[0036] In another embodiment, the method further comprises receiving a non-switch Dixon magnetic resonance image. The method further comprises generating the training data by generating synthetic water-fat exchange regions in the non-switch Dixon magnetic resonance image. The synthetic water-fat exchange regions are then used as the training water-fat exchange mask.

[0037] It should be understood that one or more of the aforementioned embodiments of the present invention may be combined as long as the combined embodiments are not mutually exclusive.

[0038] As will be appreciated by those skilled in the art, various aspects of the present invention may be implemented as an apparatus, method, or computer program product. Accordingly, various aspects of the present invention may take the form of a fully hardware embodiment, a fully software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects (all of which may be generally referred to herein as "circuits," "modules," or "systems"). In addition, various aspects of the present invention may take the form of a computer program product implemented in one or more computer-readable media having computer executable code implemented thereon.

[0039] Any combination of one or more computer-readable media can be utilized. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. As used herein, "computer-readable storage medium" encompasses any tangible storage medium that can store instructions that can be executed by a processor of a computing device. A computer-readable storage medium can be referred to as a computer-readable non-transitory storage medium. A computer-readable storage medium can also be referred to as a tangible computer-readable medium. In some embodiments, a computer-readable storage medium can also be capable of storing data that can be accessed by a processor 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 memories, USB thumb drives, random access memories (RAM), read-only memories (ROM), optical disks, magneto-optical disks, and register files of processors. Examples of optical disks include compact disks (CDs) and digital versatile disks (DVDs), such as CD-ROMs, CD-RWs, CD-Rs, DVD-ROMs, DVD-RWs, or DVD-R disks. The term computer-readable storage medium also refers to various types of recording media that can be accessed by a computer device via a network or communication link. For example, data can be retrieved on a modem, the Internet, or a local area network. Computer executable code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0040] A computer readable signal medium may include a propagated data signal having computer executable code implemented 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 may be any computer readable medium that is not a computer readable storage medium and is capable of conveying, propagating, or transmitting a program used by or in conjunction with an instruction execution system, apparatus, or device.

[0041] "Computer memory" or "memory" is an example of a computer-readable storage medium. Computer memory is any memory directly accessible by a processor. "Computer storage" 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.

[0042] As used herein, "processor" encompasses an electronic component capable of executing a program or machine executable instructions or computer executable code. References to computing devices including "processors" should be interpreted as being capable of containing more than one processor or processing core. The processor may be, for example, a multi-core processor. A processor may also refer to a collection of processors within a single computer system or distributed between multiple computer systems. The term computing device should also be interpreted as being capable of referring to a collection or network of computing devices each including one or more processors. Computer executable code may be executed by multiple processors that may be within the same computing device or may even be distributed between multiple computing devices.

[0043] The computer executable code may include machine executable instructions or programs that cause a processor to perform aspects of the present invention. The computer executable code for performing operations for aspects of the present invention may be written in any combination of one or more programming languages ​​and compiled into machine executable instructions, including object-oriented programming languages ​​such as Java, Smalltalk, C++, etc. and conventional procedural programming languages ​​such as "C" programming language or similar programming languages. In some instances, the computer executable code may take the form of a high-level language or take a precompiled form and be used in conjunction with a decoder that generates machine executable instructions at work.

[0044] The computer executable code may be executed 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 case, the remote computer may be connected to the user's computer through any type of network including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0045] Aspects of the present invention are described with reference to the flow chart, diagram and / or block diagram of the method, device (system) and computer program product according to an embodiment of the present invention. It should be understood that, when applicable, each square frame or part of the square frame of the flow chart, diagram and / or block diagram can be implemented by computer program instructions in the form of computer executable code. It should also be understood that, when mutually non-exclusive, the combination of the square frames in different flow charts, diagrams and / or block diagrams can be combined. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device that produces the machine, so that the instructions executed by the processor of the computer or other programmable data processing device create a unit for implementing the function / action specified in the flow chart and / or one or more block diagram frames.

[0046] These computer program instructions may also be stored in a computer-readable medium, which can direct a computer, other programmable data processing apparatus, or other device to work in a specific manner so that the instructions stored in the computer-readable medium produce a product including instructions for implementing the functions / actions specified in the flowchart and / or one or more block diagrams.

[0047] The computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operating steps to be performed on the computer, other programmable apparatus, or other device, thereby producing a computer-implemented process, so that the instructions executed on the computer or other programmable apparatus provide a process for the functions / actions specified in the flowchart and / or one or more block diagram boxes.

[0048] As used herein, "user interface" is an interface that allows a user or operator to interact with a computer or computer system. "User interface" can also be referred to as "human-machine interface device". The user interface can provide information or data to the operator and / or receive information or data from the operator. The user interface can enable the input from the operator to be received by the computer and can provide output to the user from the computer. In other words, the user interface can allow the operator to control or manipulate the computer, and the interface can allow the computer to indicate the effect of the control or manipulation of the operator. The display of data or information on a display or graphical user interface is an example of providing information to the operator. The reception of data by keyboard, mouse, tracking ball, touch pad, pointing stick, graphic input board, joystick, game controller, webcam, earphone, pedal, wired gloves, remote controller and accelerometer is all examples of user interface components that realize the reception of information or data from the operator.

[0049] As used herein, "hardware interface" encompasses an interface that enables a processor of a computer system to interact with and / or control an external computing device and / or apparatus. A hardware interface may allow a processor to send control signals or instructions to an external computing device and / or apparatus. A hardware interface may also enable a processor to exchange data with an external computing device and / or apparatus. 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 LAN connection, a TCP / IP connection, an Ethernet connection, a control voltage interface, a MIDI interface, an analog input interface, and a digital input interface.

[0050] As used herein, a "display" or "display device" encompasses an output device or user interface suitable for displaying images or data. A display may output visual, audio, and / 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), memory 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.

[0051] Magnetic resonance (MR) data is defined herein as the measurement of radio frequency signals emitted by atomic spins recorded using the antenna of a magnetic resonance device during a magnetic resonance imaging scan. Magnetic resonance data is an example of medical image data. 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 within the magnetic resonance imaging data. This visualization can be performed using a computer. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Preferred embodiments of the invention will be described hereinafter by way of example only and with reference to the accompanying drawings, in which:

[0053] Figure 1 The illustration shows an example of a medical system;

[0054] Figure 2 The diagram shows the operation Figure 1 A flowchart of a method of a medical system;

[0055] Figure 3 Another example of a medical system is illustrated;

[0056] Figure 4 The diagram shows the operation Figure 1 A flowchart of a method of a medical system;

[0057] Figure 5 Another example of a medical system is illustrated;

[0058] Figure 6 The diagram shows the operation Figure 1 A flowchart of a method of a medical system;

[0059] Figure 7 A method for generating a synthetic water-fat exchange is illustrated;

[0060] Figure 8 A medical system having a convolutional neural network is illustrated; and

[0061] Fig. 9 Explained Figure 8 The use of convolutional neural networks.

[0062] Reference numerals list

[0063] 100 Medical System

[0064] 102 Computer

[0065] 104 Processor

[0066] 106 Hardware Interface

[0067] 108 Optional User Interface

[0068] 110 Memory

[0069] 120 machine executable instructions

[0070] 122 Convolutional Neural Networks

[0071] 124 Initial Dixon Fat MRI

[0072] 126 Initial Dixon Water MRI

[0073] 128 One or more water-fat exchange areas

[0074] 124 Dixon-corrected fat MRI

[0075] 132 Corrected Dixon water MRI

[0076] 200 Receive initial Dixon MRI

[0077] 204 In response to inputting the initial Dixon magnetic resonance image into the convolutional neural network, receiving one or more water-fat exchange regions from the convolutional neural network

[0078] 206 Reconstructing the corrected Dixon magnetic resonance image 300 using the received one or more water-fat exchange regions of the medical system

[0079] 302 No Exchange Dixon MRI

[0080] 304 Synthetic Water-Fat Exchange Area

[0081] 306 Training Dixon MRI

[0082] 308 Training Water-Fat Exchange Mask

[0083] 400 Receive No Exchange Dixon MRI

[0084] 402 Generate training data by generating synthetic water-fat exchange regions in each of the non-exchange Dixon MRI images and constructing a training water-fat exchange mask 404 Receive training data

[0085] 406 Use training data to train convolutional neural network

[0086] 500 Medical Systems

[0087] 502 Magnetic Resonance Imaging System

[0088] 504 Magnet

[0089] 506 Magnet bore

[0090] 508 Imaging Area

[0091] 509 Area of ​​Interest

[0092] 510 Magnetic Field Gradient Coil

[0093] 512 Magnetic Field Gradient Coil Power Supply

[0094] 514 RF Coil

[0095] 516 Transceiver

[0096] 518 objects

[0097] 520 Object Support

[0098] 530 Pulse sequence command

[0099] 532 MRI data

[0100] 600 Acquiring magnetic resonance imaging data by controlling the magnetic resonance imaging system using pulse sequence commands 602 Reconstructing an initial Dixon magnetic resonance image using the magnetic resonance imaging data according to the Dixon magnetic resonance imaging protocol

[0101] 700 Artificial Water-Fat Exchange Generator

[0102] 702 training data

[0103] 900 Dixon water image with swapping artifacts

[0104] 902 Water-Fat Exchange Area

[0105] 904 Corrected Dixon Water Image DETAILED DESCRIPTION

[0106] In these figures, like numbered elements are equivalent elements or perform the same function. Elements that have been discussed previously will not necessarily be discussed in later figures if the functionality is equivalent.

[0107] Figure 1 An example of a medical system 100 is illustrated. In this example, the medical system includes a computer 102 having a processor 104. The processor is shown connected to an optional hardware interface 106. The hardware interface 106 may, for example, enable the processor 104 to communicate with other components of the medical system 100 or even exchange data with other computer systems or information systems. The processor 104 is also shown connected to an optional user interface 108 and a memory 110. The memory 110 may, for example, represent different types of memory that the processor 104 can access.

[0108] Memory 110 may be any combination of memory accessible to processor 104. This may include things such as main memory, cache memory, and non-volatile memory such as flash RAM, a hard drive, or other storage devices. In some examples, memory 110 may be considered a non-transitory computer-readable medium.

[0109] The memory 110 is shown as containing machine executable instructions 120. The machine executable instructions 120 enable the processor 104 to perform various control functions and / or data processing and image processing techniques. The memory 110 is also shown as containing an embodiment of a convolutional neural network 122. The convolutional neural network is configured to identify one or more water-fat exchange regions in response to having an initial Dixon magnetic resonance image input thereto. The memory 110 is shown as containing an initial Dixon fat magnetic resonance image 124 and an initial Dixon water magnetic resonance image 126.

[0110] The convolutional neural network 122 can be configured to work with the initial fat MRI 124 or the initial Dixon MRI 126. In some other cases, the convolutional neural network 122 takes both the initial Dixon fat MRI 124 and the initial Dixon water MRI 126 as input. In these various examples, the convolutional neural network then outputs one or more water-fat exchange regions 128 in response. The memory 110 is shown as containing the one or more water-fat exchange regions 128. The one or more water-fat exchange regions 128 are identifications of one or more voxels within the initial Dixon fat MRI 124 or the initial Dixon water MRI 126 where water regions have exchanged with fat regions. The identification of the one or more water-fat exchange regions 128 is useful for evaluating the Dixon MRI images 124, 126.

[0111] In some cases, the machine executable instructions 120 will also be programmed to employ the one or more water-fat exchange regions 128 and calculate a corrected Dixon fat MRI image 130 and a corrected Dixon water MRI image 132. In some cases, this may simply involve switching voxels between the two images 130 and 132. In other cases, the magnetic resonance data used to construct the initial Dixon MRI images 124, 126 is input into an algorithm that performs correction using the one or more water-fat exchange regions 128. For example, the one or more water-fat exchange regions 128 may be used as constraints or may be used to correct the B0 inhomogeneity estimate.

[0112] Figure 2 The diagram shows the operation Figure 1Flowchart of a method of a medical system 100 of the present invention. First, in step 200, an initial Dixon magnetic resonance image 124 and / or 126 is received. Next, in step 202, the initial Dixon magnetic resonance image 124 and / or 126 is input into the convolutional neural network 122, and as an output, one or more water-fat exchange regions 128 are received. In some examples, step 204 is performed. In step 204, a corrected Dixon magnetic resonance image 130 and / or 132 is calculated using the one or more water-fat exchange regions 128.

[0113] Figure 3 Another example of a medical system 300 is shown. It should be noted that the features of the medical system 300 may be similar to those of Figure 1 The features of the illustrated medical system 100 are freely combined. The memory 110 in this example is shown as again containing the machine executable instructions 120 and the convolutional neural network 122. In addition, the memory 110 is shown as containing a non-exchange Dixon magnetic resonance image 302. A non-exchange Dixon magnetic resonance image is a Dixon magnetic resonance image that does not contain any water-fat exchange. The non-exchange Dixon magnetic resonance image 302 can be a water image, a fat image, or both water and fat images as a pair. The memory 110 is also shown as containing a synthetic water-fat exchange region 304. Various algorithms can be applied to the non-exchange Dixon magnetic resonance images 302 to generate a water-fat exchange region 304 for each of them.

[0114] The memory 110 is also shown as containing training data combining training Dixon MRI images 306 and training water-fat exchange masks 308. Again, the training Dixon MRI images 306 can be fat, water, or both types of images that match the same type as the no-exchange Dixon MRI images 302. The training water-fat exchange masks 308 identify the locations of the synthetic water-fat exchange regions 304. The training data 306, 308 can then be used to train the convolutional neural network 122, for example using deep learning.

[0115] Figure 4 The diagram shows the use of Figure 3 Flow chart of a method of a medical system 300 . Figure 4 The method can be used with Figure 2 For example, you can first execute Figure 4 method, and then follow with Figure 2 The steps of the method are illustrated.

[0116] Figure 4The method starts at step 400. In step 400, a non-switch Dixon magnetic resonance image 302 is received. Next, in step 402, training data 306, 308 is generated by creating a synthetic water-fat exchange region 304 for the non-switch Dixon magnetic resonance image 302. Next, in step 404, the training data 306, 308 is received. Then finally in step 406, the convolutional neural network 122 is trained using the training data 306, 308.

[0117] Figure 5 Another example of a medical system 500 is shown. The medical system 500 is similar to Figure 1 The medical system is shown except that it additionally includes a magnetic resonance imaging system 502. The medical imaging system 500 is also shown as optionally including Figure 3 Features of the medical imaging system 300.

[0118] The magnetic resonance imaging system 502 comprises a magnet 504. The magnet 504 is a superconducting cylindrical magnet having a bore 506 passing through it. It is also possible to use different types of magnets; for example it is also possible to use both a split cylindrical magnet and a so-called open magnet. 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 isoplane of the magnet, such a magnet can be used, for example, in conjunction with charged particle beam therapy. An open magnet has two magnet parts, one above the other, with a space in between that is large enough to accommodate the object: the arrangement of the two partial areas is similar to that of a Helmholtz coil. Open magnets are popular because the object is less confined. Inside the cryostat of the cylindrical magnet there is a collection of superconducting coils.

[0119] Within the bore 506 of the cylindrical magnet 504, there is an imaging zone 508 in which the magnetic field is strong and uniform enough to perform magnetic resonance imaging. A region of interest 509 within the imaging zone 508 is shown. Magnetic resonance data is typically acquired for the region of interest. An object 518 is shown supported by an object support 520 so that at least a portion of the object 518 is within the imaging zone 508 and the region of interest 509.

[0120] Also within the bore 506 of the magnet is a set of magnetic field gradient coils 510 for collecting preliminary magnetic resonance data to spatially encode magnetic spins within the imaging zone 508 of the magnet 504. The magnetic field gradient coils 510 are connected to a magnetic field gradient coil power supply 512. The magnetic field gradient coils 510 are intended to be representative. Typically the magnetic field gradient coils 510 include three independent coil groups for spatially encoding in three orthogonal spatial directions. The magnetic field gradient power supply supplies current to the magnetic field gradient coils. The current supplied to the magnetic field gradient coils 510 is controlled according to time and can be ramped or pulsed.

[0121] Adjacent to the imaging zone 508 is a radio frequency coil 514, which is used to manipulate the orientation of the magnetic spins within the imaging zone 508 and to receive radio frequency transmissions from the spins also within the imaging zone 508. The radio frequency antenna may contain multiple coil elements. The radio frequency antenna may also be referred to as a channel or antenna. The radio frequency coil 514 is connected to a radio frequency transceiver 516. The radio frequency coil 514 and the radio frequency transceiver 516 may be replaced by independent transmitting coils and receiving coils and independent transmitters and receivers. It should be understood that the radio frequency coil 514 and the radio frequency transceiver 516 are representative. The radio frequency coil 514 is intended to also represent a dedicated transmitting antenna and a dedicated receiving antenna. Similarly, the transceiver 116 may also represent a separate transmitter and receiver. The radio frequency coil 514 may also have multiple receiving / transmitting elements, and the radio frequency transceiver 516 may have multiple receiving / transmitting channels. For example, if a parallel imaging technique such as SENSE is performed, the radio frequency coil 514 will have multiple coil elements.

[0122] The transceiver 516 and the gradient controller 512 are shown as being connected to the hardware interface 106 of the computer system 102. The memory 110 is also shown as containing pulse sequence commands. The pulse sequence commands 530 are commands that control the magnetic resonance imaging system 502 to acquire magnetic resonance imaging data according to the Dixon magnetic resonance imaging protocol or data that can be converted into such commands.

[0123] The memory 110 is further shown as containing magnetic resonance imaging data 532 that has been acquired by controlling the magnetic resonance imaging system 502 with the pulse sequence commands 530. The magnetic resonance imaging data 532 may be reconstructed into an initial Dixon fat magnetic resonance image 124 and an initial Dixon water magnetic resonance image 126 using the Dixon magnetic resonance imaging protocol.

[0124] Figure 6 Graphical operation Figure 5 First, in step 600, the magnetic resonance imaging system 502 is controlled using a pulse sequence command 530 to acquire magnetic resonance imaging data 532. Next, in step 602, the initial Dixon magnetic resonance images 124 and 126 are reconstructed based on the magnetic resonance imaging data 532. The method then proceeds to Figure 2 Steps 200, 202 and 204 are illustrated.

[0125] Examples may provide methods for automatically detecting and correcting water-fat exchange in Dixon images. Examples may use a neural network (convolutional neural network 122) that is trained to identify water-fat exchange using a data set (training data 702) and a standard data exchange mask (training water-fat exchange mask 308). During use, the output of the network may be used for direct correction of the exchange, or as input to a second Dixon reconstruction that produces a Dixon image without water-fat exchange (corrected Dixon image 124).

[0126] Fat suppression is an essential element to improve the diagnostic value of magnetic resonance (MR) imaging, since fat often hides details of other structures due to its high signal intensity. Suppression of fat is also important because it can help delineate structures adjacent to water-retaining tissue. Two-thirds of MR studies utilize some form of fat suppression. Dixon imaging has become a widely used technique to achieve more robust fat suppression in acquisition than conventional methods based on spectral or T1 selection. Some embodiments provide improved fat suppression in large fields of view in oncology and musculoskeletal applications, but also enable new applications such as non-subtracted MR angiography based on water-only mDIXON images acquired with contrast agents. Water-fat separation based on the two-point Dixon method faces a fundamental ambiguity: two solutions corresponding to water and fat signals with different phase errors coincide with the acquired signal. This can lead to the so-called water-fat exchange, i.e., water signals are misassigned to image areas of fat signals and vice versa.

[0127] Many variations of the original Dixon method have been developed to achieve robust separation of the two species, often relying on the assumption of spatially smooth main magnetic field (B0) inhomogeneities. Nevertheless, water-fat exchanges still occur. They occur primarily in body parts located far from the isocenter, in areas where large B0 inhomogeneities are present, or in areas around weakly magnetic materials, such as in implants.

[0128] In MRI, water-fat exchange is problematic for several reasons:

[0129] If the Dixon method is used for fat suppression, a fat signal with a high signal intensity appears in areas with exchanged water storage tissue.

[0130] - In some cases, such as when fat suppression is used to provide contrast between water-reservoir tissue embedded in adipose tissue, image interpretation is complicated and corresponding lesions may be missed.

[0131] - Quantitative analysis of Dixon scans, for example for total water-fat composition analysis, yields erroneous results in the presence of water-fat exchange.

[0132] - Using Dixon images to calculate radiation attenuation maps in radiotherapy planning may lead to incorrect dose calculations in the presence of water-fat exchange, since water and fat have different attenuation properties. The same applies to attenuation correction in PET-MR.

[0133] To address one or more of these issues, in an example, a neural network 122 is trained on an artificially created dataset (synthetic water-fat exchange regions 304 and training Dixon magnetic resonance images 306) to automatically detect and localize water-fat exchange in Dixon images. After training, the network can be applied to unseen data to identify voxels with exchanged water-fat signals (902). This information can be used to directly correct the image (904), or as input to a second Dixon reconstruction that produces a Dixon image without exchange.

[0134] Convolutional neural networks typically consist of convolutional layers and max pooling layers, followed by one or more fully connected layers. For image segmentation purposes, fully convolutional networks are often used. Unlike classification networks, they present a symmetrical structure that allows efficient generation of predictions at the pixel level (with deconvolution and unpooling layers replacing fully connected layers).

[0135] For training of such a network, a dataset with suitable images and standard data segmentation is used.A large dataset with artificial water-fat exchange can be easily generated based on a set of Dixon scans (non-exchanged Dixon magnetic resonance images 302) of the anatomical structure of interest.

[0136] The artificial exchange region can be defined, for example, by applying a suitable threshold on a randomly generated slowly varying two-dimensional function (polynomial, radial, Gaussian, etc.). The water and fat signals in this region described by the exchange mask (training water-fat exchange mask 308) are then exchanged between the Dixon water and fat images, and the exchange mask represents the standard data segmentation. Both images as well as the mask are then saved to disk for use during training. By applying data augmentation schemes such as translation, rotation, deformation or flipping to the Dixon images before exchange, a large training dataset with a high degree of anatomical variability can be generated.

[0137] Once this training data is available, a suitable objective function can be chosen that represents the accuracy of the predicted segmentation. A common choice is the pixel-wise cross entropy loss. The weights of the network can then be optimized using optimization techniques such as stochastic gradient descent based on backpropagation. Figure 7 and Figure 8 A schematic overview of the proposed method is shown.

[0138] Figure 7A method of generating training data 702 is illustrated. In this example, there is an artificial water-fat exchange generator 700, which is an algorithm. The artificial water-fat exchange generator 700 takes as input a non-exchange Dixon magnetic resonance image 302. It then outputs a training Dixon magnetic resonance image 306 and a training water-fat exchange mask 308 as output. This is an example of training data 702 generated by making synthetic water-fat exchange regions. Figure 7 In , the non-exchange Dixon data (water and fat image pairs) are artificially modified to include water-fat exchange. A standard data mask is generated for each image pair.

[0139] Figure 8 The functionality of the convolutional neural network 122 is illustrated. An initial Dixon fat magnetic resonance image 124 and / or an initial Dixon water magnetic resonance image 126 may be input into the convolutional neural network 122, and as an output, one or more water-fat exchange regions 128 are output. The Dixon image 124 or 126 is referred to as clinical Dixon data with exchange in the image, and the one or more water-fat exchange regions 128 are identified as estimated exchange masks. When using the training data 702, the neural network 122 may be used or trained using deep learning, for example. During use of the neural network 122, the network 122 is able to estimate the exchange regions given the Dixon data with exchange as input.

[0140] During use, the exchange mask (one or more water-fat exchange regions 128) predicted by the network can be used in two ways. First, for Dixon images where a true exchange of water and fat signals occurs (two echoes, a unimodal spectral model of fat, etc.), the exchange can be easily corrected by simply exchanging the two signals in the exchange region. Second, in all other cases, the intrinsic estimate of B0 inhomogeneity can be corrected in the exchange region and the water-fat separation can be repeated.

[0141] Fig. 9 The performance of the convolutional neural network 122 is illustrated. The image in column 900 is an example of a Dixon water image with swapping artifacts 900. The swapping artifacts in example 900 are artificial. Column 902 shows the output of the convolutional neural network after the images 900 have been each input to the convolutional neural network. Also, 902 is an example of a water-fat exchange region 902 identified by the neural network. The image in column 904 illustrates a corrected Dixon water image 904, which is corrected by swapping the water-fat signal within the mask and it produces a nearly artifact-free result.

[0142] In an additional example, the exchange mask of the training dataset is at least partially created using manual segmentation. Although this requires additional manual work, it can be assumed that the resulting dataset is more realistic than a manually created dataset, potentially leading to improved network performance.

[0143] In another additional example, since water-fat exchange in clinical practice occurs mainly in areas with large B0 deviations, B0 maps obtained using standard B0 mapping techniques or B0 maps based on representations such as Legendre polynomials can be used interchangeably. Areas where B0 inhomogeneity or its spatial gradient or a combination of both exceeds a defined threshold can be used to define an exchange mask.

[0144] In another additional example, one or more of several discontinuous regions in the combined water-fat image may be used to define a swap mask.

[0145] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed embodiments.

[0146] Those skilled in the art can understand and implement other variations of the disclosed embodiments when practicing the claimed invention by studying the drawings, the specification and the claims. In the claims, the word "comprising" does not exclude other elements or steps, and the words "one" or "an" do not exclude multiple. A single processor or other unit can satisfy the functions of several items recorded in the claims. Although specific elements are recorded in mutually different dependent claims, this does not indicate that the combination of these elements cannot be used to advantage. The computer program can be stored and / or distributed on appropriate media, such as optical storage media or solid-state media provided with or as part of other hardware, but the computer program can also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. Any figure mark in the claims shall not be interpreted as limiting the scope.

Claims

1. A medical system (100, 300, 500), comprising: - a memory (110) storing machine executable instructions (120) and a convolutional neural network (122), wherein the convolutional neural network is configured to receive an initial Dixon magnetic resonance image (124, 126) as input, wherein the convolutional neural network is configured to identify one or more water-fat exchange regions (128) in the initial Dixon magnetic resonance image; - a processor (104) for controlling the medical system, wherein execution of the machine executable instructions causes the processor to: - receiving (200) said initial Dixon magnetic resonance image; and - receiving (204) the one or more water-fat exchange regions from the convolutional neural network in response to inputting the initial Dixon magnetic resonance image into the convolutional neural network.

2. The medical system according to claim 1, wherein: Execution of the machine executable instructions further causes the processor to reconstruct (206) a corrected Dixon magnetic resonance image (130, 132) using the received one or more water-fat exchange regions, wherein the Dixon magnetic resonance image is corrected using any of: - wherein the Dixon magnetic resonance image further comprises a fat image (124) and a water image (126), wherein the corrected Dixon magnetic resonance image is reconstructed by exchanging voxels of the one or more water-fat exchange regions between the fat image and the water image; and - wherein the corrected Dixon MRI is reconstructed from Dixon MRI data according to a Dixon MRI reconstruction algorithm, wherein the Dixon MRI reconstruction algorithm is configured to use the one or more water-fat exchange regions as constraints and / or for correcting B0 inhomogeneity estimation.

3. The medical system according to claim 1 or 2, wherein: Execution of the machine executable instructions further causes the processor to: - receiving (404) training data (702), wherein each of the training data comprises a training Dixon magnetic resonance image (306) and a training water-fat exchange mask (308), wherein the training water-fat exchange mask identifies standard data locations of the one or more water-fat exchange regions; and - Training (406) the convolutional neural network using the training data.

4. The medical system according to claim 3, wherein: Execution of the machine executable instructions further causes the processor to: - receiving (400) a non-switchable Dixon magnetic resonance image (302); and - generating (402) the training data by generating synthetic water-fat exchange regions in each of the non-exchange Dixon MRI images and constructing the training water-fat exchange masks.

5. The medical system according to claim 4, wherein: The training data is generated at least in part by determining at least some of the one or more synthetic water-fat exchange regions by: - determine one or more random variables; - inputting the one or more random variables into a spatial variable function; - thresholding the spatially variable function to determine the one or more water-fat exchange regions for the training Dixon magnetic resonance images.

6. The medical system according to claim 4, wherein: The training data is generated at least in part by: - identifying in said Dixon magnetic resonance image a high water-fat area having both a fat content and a water content above a predetermined water-fat threshold; - selecting at least part of said high water-fat zone as said synthetic water-fat exchange area.

7. The medical system according to claim 4, wherein: The training data is generated at least in part by: - receiving a B0 inhomogeneity map for said exchange-free Dixon magnetic resonance image; - identifying areas of high B0 inhomogeneity in the Dixon magnetic resonance image by thresholding the B0 inhomogeneity map using a B0 inhomogeneity threshold; and - selecting at least part of said high B0 inhomogeneity zone as said synthetic water-fat exchange area.

8. The medical system according to claim 5, wherein: The training data is generated at least in part by: - using an image segmentation algorithm to identify tissue boundary regions in said non-exchange Dixon magnetic resonance image; and - selecting at least a portion of said tissue border zone as said synthetic water-fat exchange area.

9. The medical system according to claim 5, wherein: The training data is generated at least in part by: - determining a spatially correlated signal-to-noise ratio map for said Dixon magnetic resonance image; and - selecting at least part of the synthetic water-fat exchange area for inclusion in the training data using a signal-to-noise weighting factor determined using the spatially correlated signal-to-noise map, wherein the signal-to-noise weighting factor increases as the spatially correlated signal-to-noise map decreases.

10. The medical system according to claim 4, wherein: The training data is generated at least in part by: - receiving an isocenter position for said non-switchable Dixon magnetic resonance image; and - selecting at least part of a resultant water-fat exchange region for inclusion in said training data using a distance weighting factor determined by the distance from said isocenter position, wherein said distance weighting factor increases as said distance from said isocenter position increases.

11. The medical system according to claim 3, wherein: The training data is generated at least in part by increasing the size of the training data by the following operation: generating additional training standard data magnetic resonance images by applying an image transformation to both the training Dixon magnetic resonance images and the training water-fat exchange masks, wherein the image transformation includes any of the following: image translation, image rotation, image deformation, image flip transformation, mirror transformation, and combinations thereof.

12. The medical system according to claim 1 or 2, wherein: The medical system further comprises a magnetic resonance imaging system (502), wherein the memory further comprises pulse sequence commands (530) configured to control the magnetic resonance imaging system to acquire magnetic resonance imaging data (532) according to a Dixon magnetic resonance imaging protocol, wherein execution of the machine executable instructions further causes the processor to: - acquiring (600) the magnetic resonance imaging data by controlling the magnetic resonance imaging system using the pulse sequence commands; and - reconstructing (602) the initial Dixon magnetic resonance image using the magnetic resonance imaging data according to the Dixon magnetic resonance imaging protocol.

13. A computer program product comprising machine executable instructions (120) and a convolutional neural network (122) for execution by a processor (104) controlling a medical system (100, 300, 500), wherein: The convolutional neural network is configured to receive an initial Dixon magnetic resonance image (124, 126) as input, wherein the convolutional neural network is configured to identify one or more water-fat exchange regions (128) in the initial Dixon magnetic resonance image, wherein execution of the machine executable instructions causes the processor to: - receiving said initial Dixon magnetic resonance image; and - receiving the one or more water-fat exchange regions from the convolutional neural network in response to inputting the initial Dixon magnetic resonance image into the convolutional neural network.

14. A method for training a convolutional neural network (122), wherein: The convolutional neural network is configured to receive an initial Dixon magnetic resonance image (124, 126) as input, wherein the convolutional neural network is configured to identify one or more water-fat exchange regions (128) in the initial Dixon magnetic resonance image, wherein the method comprises: - receiving training data (404), wherein each of the training data comprises a training Dixon magnetic resonance image (306) and a training water-fat exchange mask (308), wherein the training Dixon magnetic resonance image comprises the one or more water-fat exchange regions, wherein the training water-fat exchange mask identifies standard data locations of the one or more water-fat exchange regions; and -Training the convolutional neural network using the training data.

15. The method according to claim 14, wherein: The method further comprises: - receiving a non-switchable Dixon magnetic resonance image (302); - generating said training data by generating synthetic water-fat exchange regions in said non-exchange Dixon magnetic resonance images.

Images