System of and method for magnetic resonance image guided radiotherapy

Abstract

An image taken, e.g. in a combined magnetic resonance imaging and radiotherapy apparatus, is corrected for motion of the patient support during imaging using a trained neural network. The invention is

Description

This disclosure relates to magnetic resonance imaging (MRI), and in particular to the generation of an updated magnetic resonance imaging (MRI) data set using a machine learning model such as a neural network trained to reduce one or more distortions in MRI data sets. Background Radiotherapy can be described as the use of ionising radiation to damage or destroy unhealthy cells in both humans and animals. The ionising radiation may be directed to tumours on the surface of the skin or deep inside the body. Common forms of ionising radiation include X-rays and charged particles. Volumetric modulated arc therapy (VMAT) is a radiotherapy technique which involves the delivery of radiation to a target region, such as a tumour, while the source of radiation and / or beam generation system rotates. A particular implementation of VMAT, known as helical volumetric modulated arc therapy (helical VMAT), involves the delivery of radiation to a tumour via a rotatable source of radiation whilst a patient support is translated along the axis of rotation, allowing the radiation beam to be applied to the tumour in a helical fashion. In other words, the movement of the patient support and the rotation of the source of radiation produce a radiation beam in a helical trajectory. Such a technique provides an effective treatment for a tumour with multiple isocenters, for example a tumour located along the craniospinal axis. It is desirable to have a means of imaging a patient during a radiotherapy treatment. Imaging a patient during a radiotherapy treatment enables real-time adaptive treatment techniques such as gating and MLC tracking, which can optimize treatment by enabling a prescribed dose to be provided to a target region while minimising or limiting the radiation delivered to other parts of the patient (i.e. to healthy tissue). It can also be desirable to have such a means of imaging the patient to assist in treatment planning. Magnetic Resonance Imaging, MRI, is one such imaging technique. Figure 1 illustrates a partially cut-away view of an example system including a combined radiotherapy device and an MR imaging device, known as an MR linac, where the treatment beam is emitted from an accelerator 9 while the gantry 1 is rotated around a patient support 19, and the patient support is simultaneously translated in the longitudinal direction. The resulting treatment beam 13 is a volumetric modulated helical arc beam. The acquisition of useful MR images or data signals during helical VMAT has been thought to be an extremely difficult, if not impossible, problem. MR images require an acquisition time, and during helical VMAT there is motion of the patient during this acquisition time, between the MR imaging device and the patient. The MR signal phase in each proton spin inside the patient’s body is distorted due to the relative motion of the patient support and components of the MR imaging device, for example the fixed RF coils used for MR signal detection installed inside the magnet 17. In other words, the relative motion between the patient support and components of the MR imaging device, such as the one or more RF coils installed inside the magnet, causes distortions in the resulting MR image. Therefore, conventional MR image reconstruction algorithms do not function properly when used for MR images taken during helical VMAT treatment. Accordingly, despite the advantages of MR imaging for standard VMAT, these advantages cannot currently be realized for helical VMAT. The present invention seeks to address these and other disadvantages encountered in the prior art. Summary According to an aspect, there is provided a computer-implemented method for generating an updated magnetic resonance imaging, MRI, data set using a neural network trained to reduce one or more distortions in MRI data sets, the distortions being caused, at least in part, by relative motion between a patient support and a component of an MR imaging device. The method comprises receiving a first MRI data set, wherein the first MRI data set is based on MRI data acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device occurred. The received first MRI data set is then applied as an input to a trained neural network, and converted, using the trained neural network, to an updated MRI data set in which distortions in the first MRI data set are reduced. The relative motion between the patient support and the component of the MR imaging device may be caused, at least in part, by translation of the patient support with respect to the component. The translation may have occurred during delivery of helical volumetric modulated arc therapy treatment. The received first MRI data set may comprise a first magnetic resonance image, or first magnetic resonance signals. The updated MRI data set may comprise one of an updated magnetic resonance image, or updated magnetic resonance signals. The neural network may have been trained using a first and a second plurality of MRI data sets, each of the first plurality of MRI data sets comprising MRI data acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device occurred, and each of the second plurality of MRI data sets comprising MRI data acquired at an acquisition time in which no relative motion between the patient support and the component of the MR imaging device occurred, and where each one of the second plurality of MRI data sets corresponds to a respective one MRI data set of the first plurality of MRI data sets. The method can further comprise receiving a first plurality of MRI data sets, each of the first plurality of MRI data sets comprising MRI data acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device occurred, receiving a second plurality of MRI data sets, each of the second plurality of MRI data sets comprising MRI data acquired at an acquisition time in which no relative motion between the patient support and the component of the MR imaging device occurred, each one of the second plurality of MRI data sets corresponding to a respective one MRI data set of the first plurality of MRI data sets, and training a neural network using the corresponding first plurality of MRI data sets and second plurality of MRI data sets to reduce one or more distortions in the first plurality of MRI data sets. Training the neural network can comprise, for each of the first plurality of MRI data sets, applying the neural network to provide an estimated distortion-reduced MRI data set, and updating parameters of the neural network based on a comparison of the estimated distortion-reduced MRI data set with the one of the second plurality of MRI data sets which corresponds to the respective one of the first plurality of MRI data sets. The MRI data sets of the first plurality of MRI data sets may be one of magnetic resonance images, or magnetic resonance signals. The MRI data sets of the second plurality of MRI data sets may be one of magnetic resonance images, or magnetic resonance signals. The first plurality of MRI data sets may not comprise the received first MRI data set. Converting the received MRI data set to an updated MRI data set using the trained neural network may further comprise outputting the updated MRI data set. The method may further comprise using the updated MRI data set in a treatment plan. The method may further comprise using the updated MRI data set in real time cine imaging to monitor a target position. The updated MRI data set may be a corrected MRI data set. The corrected data set may be one in which the distortions caused, at least in part, by the relative motion between the patient support and the component of the MR imaging device have been corrected. The neural network may be a deep learning neural network, and may be a convolutional neural network. The neural network may be a generative adversarial network (GAN), and may be a cycleGAN. The component of the MR imaging device may comprise one or more RF coils. The one or more RF coils may comprise an RF transmitter coil. The one or more RF coils may be fixed, and may be fixed inside the bore of the MR imaging device. A second aspect provided herein is a system for generating an updated magnetic resonance imaging, MRI, data set using a neural network trained to reduce one or more distortions in MRI data sets, the distortions being caused, at least in part, by relative motion between a patient support and a component of an MR imaging device. The system comprises a memory storing machine executable instructions and the trained neural network, a processor for controlling the system, wherein execution of the machine executable instructions causes the processor to control the system to perform any method described herein. A third aspect provided herein is a computer-readable medium containing instructions which, when executed by the processor, cause the processor to perform any method described herein. Figures Specific embodiments are now described, by way of example only, with reference to the drawings, in which: Figure 1 illustrates a partially cut-away view of an example system including a combined radiotherapy device and an MR imaging (MRI) device. Figure 2 depicts a radiotherapy device suitable for delivering, and configured to deliver, a beam of radiation to a patient during radiotherapy treatment. Figure 3 depicts a method 300 according to the present disclosure. Figure 4 depicts a method 400 for training a neural network according to the present disclosure. Figures 5a-5b depict a method of training a neural network according to the present disclosure, along with the corresponding trained neural network. Figures 6a-6b depict a method of training a neural network according to the present disclosure, along with the corresponding trained neural network. Figures 7a-7b depict a method of training a neural network according to the present disclosure, along with the corresponding trained neural network. Figures 8a-8b depict a method of training a neural network according to the present disclosure, along with the corresponding trained neural network. Figure 9 illustrates a block diagram of one implementation of a radiotherapy system according to the present disclosure. Detailed Description In overview, and without limitation, the application discloses generating an updated magnetic resonance imaging (MRI) data set using a machine learning model such as a neural network. The model is trained to reduce one or more distortions in MRI data sets, the distortions being at least partly caused by relative motion between a patient support and a component of an MR imaging device such as might occur during helical VMAT treatment. The method comprises receiving a first MRI data set, acquired during an acquisition time in which relative motion between the patient support and a component of the MR imaging device occurred, applying that MRI data set to the neural network, and converting the first MRI data set to an updated MRI data set in which distortions in the first MRI data set are reduced, using the trained neural network. The model is trained using a first plurality of MRI data sets acquired during an acquisition time in which relative motion between the patient support and the component occurred, and a second plurality of MRI data sets acquired during an acquisition time in which no relative motion occurred. Advantageously, the presently disclosed method allows for the reduction of distortions in MR images caused by relative motion of a patient support, and an MR imaging device. The reduction of these distortions allows for more accurate MRI data sets to be produced during acquisition times where there is relative motion between the patient support and an MR imaging device. These MRI data sets can then be utilised for improved treatment planning, real-time imaging and the like. In particular, where an MR imaging device is part of an MR-linac, or is otherwise used in conjunction with a radiotherapy device to provide radiotherapy treatment, the present method allows for the accurate MR imaging of a treatment subject during periods where there is relative motion of the MR imaging apparatus and the patient support. For example, the presently disclosed method allows for more accurate MR imaging of a treatment subject during a helical radiotherapy treatment, or a helical VMAT treatment. Thus, using the presently disclosed method, a region of interest (e.g. a tumour) can be treated using a helical treatment, with the provision of MR images that cover the total treatment volume, thereby minimising the complexity of quality assurance and treatment workflow, and minimising session time whilst also maximising treatment safety. Figure 2 depicts a radiotherapy device suitable for delivering, and configured to deliver, a beam of radiation to a patient during radiotherapy treatment. The device and its constituent components will be described generally for the purpose of providing useful accompanying information for the present invention. The device depicted in Figure 2 is in accordance with the present disclosure and is suitable for use with the disclosed systems and apparatuses. The device 200 depicted in Figure 2 is an MR-linac. The device 200 comprises both MR imaging apparatus 212 and radiotherapy (RT) apparatus which may comprise a linac device. The MR imaging apparatus 212 is shown in cross-section in the diagram. In operation, the MR scanner produces MR images of the patient, and the linac device produces and shapes a beam of radiation and directs it toward a target region within a patient’s body in accordance with a radiotherapy treatment plan. The depicted device does not have the usual ‘housing’ which would cover the MR imaging apparatus 212 and RT apparatus in a commercial setting such as a hospital. The MR-linac device depicted in Figure 2 comprises a source of radiofrequency waves 202, a waveguide 204, a source of electrons 206, a source of radiation 206, a collimator 208 such as a multi-leaf collimator configured to collimate and shape the beam, MR imaging apparatus 212, and a patient support surface 214. In use, the device would also comprise a housing (not shown) which, together with the ring-shaped gantry, defines a bore. The moveable support surface 214 can be used to move a patient, or other subject, into the bore when an MR scan and / or when radiotherapy is to commence. The MR imaging apparatus 212, RT apparatus, and a subject support surface actuator are communicatively coupled to a controller or processor. The controller is also communicatively coupled to a memory device comprising computer-executable instructions which may be executed by the controller. The RT apparatus comprises a source of radiation and a radiation detector (not shown). Typically, the radiation detector is positioned diametrically opposed to the radiation source. The radiation detector is suitable for, and configured to, produce radiation intensity data. In particular, the radiation detector is positioned and configured to detect the intensity of radiation which has passed through the subject. The radiation detector may also be described as radiation detecting means, and may form part of a portal imaging system. The radiation source may comprise a beam generation system. Fora linac, the beam generation system may comprise a source of RF energy 202, an electron gun 206, and a waveguide 204. The radiation source is attached to the rotatable gantry 216 so as to rotate with the gantry 216. In this way, the radiation source is rotatable around the patient so that the treatment beam 210 can be applied from different angles around the gantry 216. In a preferred implementation, the gantry is continuously rotatable. In other words, the gantry can be rotated by 360 degrees around the patient, and in fact can continue to be rotated past 360 degrees. The gantry may be ring-shaped. In other words, the gantry may be a ringgantry. The source 202 of radiofrequency waves, such as a magnetron, is configured to produce radiofrequency waves. The source 202 of radiofrequency waves is coupled to the waveguide 204 via circulator 218, and is configured to pulse radiofrequency waves into the waveguide 204. Radiofrequency waves may pass from the source 202 of radiofrequency waves through an RF input window and into an RF input connecting pipe or tube. A source of electrons 206, such as an electron gun, is also coupled to the waveguide 204 and is configured to inject electrons into the waveguide 204. In the electron gun 206, electrons are thermionically emitted from a cathode filament as the filament is heated. The temperature of the filament controls the number of electrons injected. The injection of electrons into the waveguide 204 is synchronised with the pumping of the radiofrequency waves into the waveguide 204. The design and operation of the radiofrequency wave source 202, electron source and the waveguide 204 is such that the radiofrequency waves accelerate the electrons to very high energies as the electrons propagate through the waveguide 104. The design of the waveguide 204 depends on whether the linac accelerates the electrons using a standing wave or travelling wave, though the waveguide typically comprises a series of cells or cavities, each cavity connected by a hole or ‘iris’ through which the electron beam may pass. The cavities are coupled in order that a suitable electric field pattern is produced which accelerates electrons propagating through the waveguide 204. As the electrons are accelerated in the waveguide 104, the electron beam path is controlled by a suitable arrangement of steering magnets, or steering coils, which surround the waveguide 204. The arrangement of steering magnets may comprise, for example, two sets of quadrupole magnets. Once the electrons have been accelerated, they may pass into a flight tube. The flight tube may be connected to the waveguide by a connecting tube. This connecting tube or connecting structure may be called a drift tube. The electrons travel toward a heavy metal target which may comprise, for example, tungsten. Whilst the electrons travel through the flight tube, an arrangement of focusing magnets act to direct and focus the beam on the target. To ensure that propagation of the electrons is not impeded as the electron beam travels toward the target, the waveguide 204 is evacuated using a vacuum system comprising a vacuum pump or an arrangement of vacuum pumps. The pump system is capable of producing ultra-high vacuum (UHV) conditions in the waveguide 204 and in the flight tube. The vacuum system also ensures UHV conditions in the electron gun. Electrons can be accelerated to speeds approaching the speed of light in the evacuated waveguide 204. The source of radiation is configured to direct a beam 210 of therapeutic radiation toward a patient positioned on the patient support surface 214. The source of radiation may comprise a heavy metal target toward which the high energy electrons exiting the waveguide are directed. When the electrons strike the target, X-rays are produced in a variety of directions. A primary collimator may block X-rays travelling in certain directions and pass only forward travelling X-rays to produce a treatment beam 210. The X-rays may be filtered and may pass through one or more ion chambers for dose measuring. The beam can be shaped in various ways by beam-shaping apparatus, for example by using a multi-leaf collimator 208, before it passes into the patient as part of radiotherapy treatment. In some implementations, the source of radiation is configured to emit either an X-ray beam or an electron particle beam. Such implementations allow the device to provide electron beam therapy, i.e. a type of external beam therapy where electrons, rather than X-rays, are directed toward the target region. It is possible to ‘swap’ between a first mode in which X-rays are emitted and a second mode in which electrons are emitted by adjusting the components of the linac. In essence, it is possible to swap between the first and second mode by moving the heavy metal target in or out of the electron beam path and replacing it with a so-called ‘electron window’. The electron window is substantially transparent to electrons and allows electrons to exit the flight tube. The patient support surface, also known as a subject support surface, subject support, or a patient support, 214 is configured to move between a first position substantially outside the bore, and a second position substantially inside the bore. In the first position, a patient or subject can mount the patient support surface. The support surface 214, and patient, can then be moved inside the bore, to the second position, in order for the patient to be imaged by the MR imaging apparatus 212 and / or imaged or treated using the RT apparatus. The movement of the patient support surface is effected and controlled by a subject support surface actuator, which may be described as an actuation mechanism. The actuation mechanism is configured to move the subject support surface in a direction parallel to, and defined by, the central axis of the bore. The terms subject and patient are used interchangeably herein such that the subject support surface can also be described as a patient support surface. The subject support surface may also be referred to as a moveable or adjustable couch or table. The radiotherapy apparatus / device depicted in Figure 2 also comprises MR imaging apparatus 212. The MR imaging apparatus 212 is configured to obtain images of a subject positioned, i.e. located, on the patient support surface 214. The MR imaging apparatus 212 may also be referred to as the MR imager. The MR imaging apparatus 212 may be a conventional MR imaging apparatus operating in a known manner to obtain MR data, for example MR images. The skilled person will appreciate that such a MR imaging apparatus 212 may comprise a primary magnet, one or more gradient coils, one or more receive coils, and an RF pulse applicator. The operation of the MR imaging apparatus is controlled by the controller. The controller is a computer, processor, or other processing apparatus. The controller may be formed by several discrete processors; for example, the controller may comprise an MR imaging apparatus processor (also referred to as an MR controller), which controls the MR imaging apparatus 212; an RT apparatus processor, which controls the operation of the RT apparatus; and a patient support surface processor which controls the operation and actuation of the patient support surface 214. The controller is communicatively coupled to a memory, e.g. a computer readable medium. The linac device also comprises several other components and systems as will be understood by the skilled person. For example, in order to ensure the linac does not leak radiation, appropriate shielding is also provided. The radiotherapy device may be configured to perform any of the method steps presently disclosed and may comprise computer executable instructions which, when executed by a processor, cause a processor to perform any of the method steps presently disclosed. Any of the steps that the radiotherapy device is configured to perform may be considered as method steps of the present disclosure and may be embodied in computer executable instructions for execution by a processor. The radiotherapy apparatus 200 can be configured to perform a helical treatment, based on a corresponding treatment plan. In a helical treatment, the radiotherapy apparatus is configured to rotate the gantry 216 supporting the radiation head and to apply continuous irradiation from the radiation head as it rotates about a subject (e.g. a patient or region of interest), whilst the subject is moved via translation of the patient support 214. By moving the subject through the radiation plane as the radiation source is rotated, the radiation beam sweeps out a helical shape from the patient’s perspective. The pitch of the helix can be adjusted by varying the speed of movement of the patient support 214. The helical treatment can start in a first anatomical location, for example at a particular vertebrae of a patient’s spine, and end at a second anatomical location, for example a different vertebrae, applying radiation to the segment of the spine located in between the two locations as an elongate target region. An example of such a helical treatment is a helical volumetric modulated arc therapy treatment (helical VMAT). MR imaging involves placing a subject in a strong magnetic field which aligns the magnetic moments of protons in the subject to produce a net magnetization. Irradiating the subject with radiofrequency (RF) pulses of a particular resonant frequency tips the net magnetization of these protons by a flip-angle a° into a plane perpendicular to the strong magnetic field. Immediately after the RF pulse is completed, the tipped net magnetization of these protons realigns with the strong magnetic field. The changing magnetic flux generated during realignment induces a voltage in a coil. This is measured and analysed to provide information on the distribution of different tissues within the subject. The skilled person will be familiar with MR imaging techniques. The MR imaging apparatus 212 may be comprised in the device 200. The MR imaging apparatus 212 comprises a primary magnetic field coil, a magnetic gradient coil and a radiofrequency (RF) coil 206. Each of these may have cylindrical symmetry around the bore and each may be positioned concentrically around the bore. The primary magnetic field coil is configured to produce a constant magnetic field along the bore of the MR imaging apparatus. This constant magnetic field may be oriented in the z-direction in Figure 2. This constant magnetic field coil is of high strength, typically of the order of one or more Tesla. The magnetic gradient coil is configured to generate a spatially varying magnetic field in the volume of the bore, i.e. within a subject positioned on the patient support 214. The spatially varying magnetic field is in the same direction as the constant magnetic field produced by the primary magnetic field coil, i.e. in the z-direction. However, this magnetic field in the z-direction may vary (in strength) in dependence on location along x, y or z axes. The magnetic gradient coil 204 may include an x-direction magnetic gradient coil and / or a y-direction magnetic gradient coil and / or a z-direction magnetic gradient coil. Each of these gradient coils may be configured to generate a spatially varying field in the z-direction which varies along a particular axis. For example, the x-direction magnetic gradient coil may generate a magnetic field in the z-direction which varies along the x-axis. Similarly, the y-direction magnetic gradient coil may generate a magnetic field in the z-direction which varies along the y-axis. The z-direction magnetic gradient coil may generate a magnetic field in the z-direction which varies along the z-axis. The RF coil is configured to generate RF pulses and to detect MR signals produced in the subject in response to these RF pulses. When a current is passed through the RF coil, an oscillating / rotating magnetic field is produced in the bore. This magnetic field is perpendicular to the constant magnetic field produced by the primary magnetic field coil. Conversely, changing magnetic fields of the protons of the subject induce a voltage in the RF coil which can be analysed to derive information on the distribution of tissue within the subject. The RF coil may comprise an RF transmitter coil configured to generate the RF pulses and an RF receiver coil configured to detect the MR signals. Alternatively, the same coil may be used for both transmitting and receiving. In some implementations, the RF transmitter coil is fixed inside the bore of the MR linac, whereas the RF receiver coil is configured to move with the patient support 214. In other words, the RF transmitter coil remains fixed within the bore (i.e. inside the MR magnet), and the RF receiver coil moves relative to the RF transmitter coil, together with the patient support (and, for example, a patient on the support). As such, relative motion between the RF receiver coil and the RF transmitter coil may occur during a translation or other movement of the patient support. For example, the RF receiver coil may move relative to the RF transmitter coil as the patient support undergoes a translation during a helical treatment. The MR imaging apparatus 212 may additionally comprise one or more of thermal insulation, magnetic shielding and RF shielding. These may be disposed between and / or around the outside of one or more of the coils described above. The MR imaging apparatus 212 may comprise an MR controller which controls the MR imaging apparatus 212. The MR controller may be a computer, processor, or other processing apparatus. The MR controller may be communicatively coupled to a memory, e.g. a computer readable medium. The MR controller may be communicatively coupled to one or more other components of the device 200 and / or to one or more controllers thereof. As noted further above, during a helical treatment the radiotherapy apparatus is configured to rotate the gantry 216 supporting the radiation head and to apply continuous irradiation from the radiation head as it rotates about the subject, whilst the subject (e.g. a patient) is moved via translation of the patient support 214. The translation of the patient support 214 also causes the subject to move relative to the MR imaging apparatus. In particular, the translation of the patient support causes the subject to move relative to the components of the MR imaging apparatus. The subject support can move relative to one or more components of the MR imaging apparatus (or MR imaging device) during MR imaging. In some implementations, the subject support may move relative to the MR magnet. In such implementations, relative motion between the subject support and the MR magnet during MR imaging causes distortions in the resulting MR data set, for the same or similar reasons as those described below in relation to relative motion between the subject support and the one or more RF coils. In other implementations, the subject support may move relative to one or more RF coils of the MR imaging device, as described herein. In such implementations, relative motion between the subject support and the one or more RF coils during MR imaging causes distortions in the resulting MR data set. The motion of a subject relative to the one or more RF coils of the MR imaging apparatus during MR imaging of the subject creates distortions in the resulting MR data set, which can render a resulting MR image inaccurate. The one or more RF coils may comprise only the RF transmitter coil, or both the RF transmitter coil and the RF receiver coil. In some implementations the one or more RF coils are fixed in the bore of the MR apparatus (or MR-linac). In a preferred implementation, the RF transmitter coil is fixed in the bore of the MR apparatus. This movement of the subject can be caused by the movement (e.g. a translation) of the patient support, for example during a helical treatment. These distortions are due, at least in part, to the movement of the patient support (and thus the subject of the MR imaging) causing inaccuracies in the MR reconstruction process. During a typical MR reconstruction process, an assumption is made that the subject does not move relative to the RF transmitter coil and the RF receiver coil. In other words, a typical MR reconstruction process assumes that a subject is stationary relative to the one or more RF coils of the MR apparatus. Such an assumption does not hold true when a patient support moves relative to one or more of the RF coils of the MR imaging apparatus during a period in which MR data is being acquired. For example, for an MR-linac where the RF transmitter coil is fixed in the bore, when a two-dimensional MR imaging sequence is executed on a coronal plane or a sagittal plane, a received time-domain MR signal has a phase information of: z G(z) + B0(z) where G(z) is the gradient field in the z-direction, B0(z) is the variable magnetic field at z, and where z is an MRI-space coordinate of a particular body position in a patient in a headfoot direction, defined in reference to the MR magnet and gradient coils. To reconstruct a spatial domain MR image from time-domain MR signals received during an MR acquisition time, a Fourier transform is used. An assumption that the particular body position (z) of the patient does not move during the MR data acquisition period is made during a typical MR reconstruction process. In other words, an assumption that the position of the patient remains stationary relative to the one or more coils of the MR apparatus is made during a typical MR reconstruction process. If the patient support moves, for example via translation in the z-direction, this assumption made during the reconstruction process is no longer accurate. A movement of the patient support, and thus the patient, while the imaging sequence is executed (i.e. during the MR acquisition time) leads to phase distortion in the received time-domain MR signals, resulting in image distortions in a resultant MR image. The distortions caused by the relative motion between a patient support and a component of the MR imaging device may be spatial distortions in an MR image and / or phase distortions in MR signals. In other words, one or more features of the MR data set may be spatially distorted or phase-distorted when compared to the corresponding one or more features on an equivalent or corresponding MR data set acquired when there was no relative motion between the patient support and the MR imaging device. The one or more features on the MRI data set acquired during a period of relative motion may be, for example, translated, blurred, noisy, deformed or curved when compared to the corresponding one or more features on an equivalent or corresponding MRI data set acquired of the same subject when there was no relative motion between the patient support and a component of the MR imaging device. The one or more distortions may appear, for example, as blurring, streaking, noise or shading on a resultant MR image. The one or more distortions can affect the entirety of the MR data set, as the entire subject may be displaced relative to the component of the MR imaging device at the acquisition time of the MR data set. This differs from individual, localised, motion artefacts caused by motion of part of the subject (e.g. a tumour or region of interest), which occur when a subject is on a stationary patient support, where there is no relative motion between the patient support and the MR imaging apparatus. Unlike motion artefacts which can appear on an MR image due to the motion of part of a subject on a stationary support surface, the distortions caused by relative motion of the patient support and one or more RF coils are not governed by random, or stochastic, processes. Instead, distortions caused by relative motion of the patient support and one or more RF coils are governed by deterministic processes. In other words, the relative motion is planned, for example as part of a treatment plan. Distortions in the MR data may also be caused, at least in part, by relative motion between a subject on a patient support and a component of an MR imaging device. For example, distortions may be caused by relative motion between a subject and a fixed RF transmitter coil, the relative motion occurring during an acquisition time in which MR data is being acquired. In some implementations, distortions in the MR data may be caused by both the relative motion between a patient support and a component of the MR imaging device, and relative motion between a subject on a patient support and a component of an MR imaging device. The presence of distortions on an MRI dataset can render the entire data set inaccurate and such data sets are not suitable for use in, for example, treatment planning or real-time imaging. Provided herein is a computer-implemented method for generating an updated magnetic resonance imaging (MRI) data set using a neural network trained to reduce one or more distortions in MRI data sets, the distortions being caused by relative motion between a patient support and a component of an MR imaging device. Figure 3 depicts a method 300 according to the present disclosure. The method is suitable for being performed, for example, by system 900 described further below. At block 310, a first magnetic resonance imaging (MRI) data set is received, wherein the first MRI data set is based on MRI data acquired during an acquisition time in which relative motion between a patient support (e.g. 214) and a component, such as one or more RF coils, of an MR imaging device (e.g. 212) occurred. The relative motion between the patient support and the component may be due to, for example, a translation of the patient support in any of the X-, Y-, orZ-planes relative to the component. In some implementations, relative motion between the patient support and the component may be due to a rotation or a tilting of the patient support relative to the RF coils. For example, the first MRI data set may be based on MRI data acquired during a helical treatment, wherein the translation of the patient support occurred as part of the helical treatment. That is, the first MRI data may be based on MRI data acquired during a helical treatment wherein at the time of acquisition, the radiotherapy apparatus 200 rotated the gantry 216 supporting the radiation head to apply continuous irradiation from the radiation head as it rotated about the subject, whilst the subject (e.g. a patient) was moved via translation of the subject support surface 214. The MRI data upon which the first received MRI data set is based may be received by the system 900. In some implementations, the MRI data is received by the computing system 910. In further implementations, the first received MRI data set is received by the MR Image processing circuitry. The MRI data may be received directly from an MRI acquisition device. The MRI data upon which the first received MRI data set is based can comprise a first magnetic resonance image. That is, the first received MRI data set may be based on a magnetic resonance image acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device 212 occurred. This may be, for example, an MR image of a patient, a target, or a region of interest acquired during a radiotherapy treatment. In some implementations, the MR image may have been acquired during a helical treatment, or a helical VMAT treatment. In other implementations, the MR data upon which the first received MRI data set is based can comprise a first magnetic resonance signals. That is, the first received MRI data set may be based on magnetic resonance signals acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device 212 occurred. This may be, for example, MR signals of a patient, a target, or a region of interest acquired during a radiotherapy treatment. In some implementations, the MR signals may have been acquired during a helical treatment, or a helical VMAT treatment. In some implementations, the MRI data upon which the first received MRI data set is based has undergone one or more pre-processing steps. For example, the MRI data may have undergone one or more steps of segmentation (e.g. organ segmentation), target identification (e.g. tumour identification), tumour contouring, noise removal, and / or smoothing. Any pre-processing of the MRI data may take place prior to step 320. The preprocessing steps may be carried out by the computing system 910. In some implementations, the pre-processing steps may be carried out by the MR image processing circuitry 919. In such implementations, the MRI data is received by the MR image processing circuitry, and undergoes one or more pre-processing steps prior to being applied as an input to the trained neural network in block 320. At block 320, the received first MRI data set is applied as an input to a trained neural network. Herein, applying a data set as an input may comprise inputting the data set into the trained neural network. The trained neural network is a neural network which has been trained to reduce one or more distortions in MRI data sets, the distortions being caused by relative motion between a patient support and a component of the MR imaging device. In other words, the neural network is one which has been trained to mitigate, compensate or ameliorate one or more distortions in MRI data sets, the distortions being caused by relative motion between a patient support and a component of the MR imaging device. The trained neural network can be, for example, a deep learning network, a convolutional neural network, a generative adversarial network, or a CycleGAN. The features of the trained neural network, and the training of the neural network, are described further below in relation to figure 4. At block 330, the received MRI data set is converted to an updated MRI data set in which the distortions in the first MRI data set are reduced, using the trained neural network, the distortions being those caused by the relative motion of the patient support and a component of the MR imaging device. In other words, the first received MRI data set is converted by the trained neural network to an updated MRI data set, in which one or more distortions in the data set have been reduced, mitigated, compensated or ameliorated. The updated MRI data set can comprise an updated MR image. That is, the updated MRI data set may comprise an MR image in which one or more distortions caused by the relative motion between a patient support and a component of the MR imaging device are reduced, mitigated, compensated or ameliorated. In other implementations, the updated MRI data set can comprise updated MR signals. That is, the updated MRI data set may comprise MR signals in which one or more distortions caused by the relative motion between a patient support and a component of the MR imaging device are reduced, mitigated, compensated or ameliorated. In such an implementation, standard MR reconstruction algorithms can be used to generate an updated, or ‘corrected’, MR image in which the distortions have been reduced or removed. Accordingly the method may comprise an additional step of generating an MR image based on the updated MR data signals. The updated MRI data set may be a corrected MRI data set, wherein the corrected MRI data set is one in which the distortions caused by the relative motion between the patient support and the component of the MR imaging device have been corrected. In other words, the updated MRI data set may present a corrected MRI data set which corresponds to or is equivalent to that of an MR image acquired by the MR apparatus at an acquisition time at which there is no relative motion between the patient support and the component of the MR imaging device. By providing an updated MRI data set in which one or more distortions caused by relative motion between the patient support and the component are reduced, the present method allows for the provision of a more accurate MR image during, for example, radiotherapy treatments in which a patient support is moved relative to the RF coils of an MR imaging device. Such a radiotherapy treatment may be a helical treatment, or a helical volumetric modulated arc therapy treatment. In particular, by providing an updated MRI data set in which one or more distortions caused by relative motion between the patient support and the component of the MR imaging device are corrected, the present method allows for the provision of an MR image which corresponds to or is equivalent to that of an MR image acquired by the MR apparatus at an acquisition time at which there is no relative motion between the patient support and the component of the MR imaging device. As such, it is possible to provide a more accurate MR image during, for example, radiotherapy treatments in which a patient support is moved relative to a component, for example the RF coils, of an MR imaging device. As noted above, such a radiotherapy treatment may be a helical treatment, or a helical volumetric modulated arc therapy treatment. The updated MRI data set may be used in the generation of a treatment plan. That is, the updated MRI data set may be used to plan a radiotherapy treatment for a patient. The updated MRI data set may provide information on target area location during a radiotherapy treatment, in particular during a treatment in which there is relative motion between a patient support and a component of an MR imaging device, which may otherwise be obscured or rendered inaccurate due to distortions caused by the aforementioned relative motion. This is particularly useful in relation to helical treatment, and helical volumetric modulated arc therapy, as it allows for the use of MRI data to aid treatment planning for such treatments, without being affected by the distortions caused by the translation of the patient support. For MR-linacs, such as system 200 described above, it is advantageous to use an updated MRI data set to generate a treatment plan, and / or to generate revisions to a treatment plan in real-time. An MR-linac such as that described above can acquire geometric information regarding the shape and / or location of a target (e.g. a tumour, or one or more organs-at-risk) in real-time, or in breaks / pauses during delivery of a radiotherapy treatment, using MR imaging techniques. The treatment plan can thus be revised if or when the target changes position or shape during the treatment. However, in order to revise the treatment accurately and safely, it is desirable to avoid or omit any distortions caused by the relative motion of the patient support (and thus the target being imaged and treated) and the MR imaging device. As such, the provision of an updated MRI data set in which distortions are reduced allows for more accurate revision of a treatment plan in real-time, or in breaks / pauses during delivery of radiotherapy treatment. For MR-linacs, such as system 200 as described above, it is advantageous to facilitate realtime imaging of a subject, for example during a radiotherapy treatment. Such real-time imaging may take place during delivery of the radiotherapy treatment, or during breaks / pauses in the delivery of the radiotherapy treatment Real-time MR imaging of a radiotherapy treatment in which there may be relative motion between the patient support, and thus the target being treated, and the MR imaging device, is hindered by distortions to the MR data caused by the relative motion. The present method provides an updated MR data set which reduces the distortions caused by such relative motion, and as such can facilitate the real-time imaging or cine imaging of a subject during, for example, a radiotherapy treatment. In regard to helical treatments, and helical volumetric arc therapy treatments, the present method would facilitate the real-time imaging of the subject during treatment. Facilitating real-time imaging of a subject allows for adjustments to be made to treatment in real-time. For example, it can be verified whether a target of interest is within the region being treated by the radiotherapy beam, and where the target has moved, at least in part, outside the region being treated by the beam, the treatment can be stopped to prevent radiation being directed to, for example, an organ-at-risk. Similarly, real-time imaging can be used to inform a clinician or a control system when a treatment plan may need to be revised in real-time. Optionally, the updated MRI data set is output In other words, converting the received MRI data set to an updated MRI data set using the trained neural network can further comprise outputting the updated MRI data set. The outputted updated MRI data set may comprise an updated MR image, or may comprise MR signals. The outputted updated MRI data set may be provided on a display device, such as a screen or output device 930 as described further below. The outputted updated MRI data set can be used to inform a clinician or a control system in regard to the progression of a treatment plan, and / or whether a treatment plan may need to be revised in real-time. Figure 4 depicts a method 400 for training a neural network according to the present disclosure. The method is suitable for being performed, for example, by system 200 and / or system 900, to train a neural network such as that utilized in method 300. At block 410, a first plurality of MRI data sets are received, wherein each of the first plurality of MRI data sets comprise MRI data acquired during an acquisition time in which relative motion between a patient support and a component of an MR imaging device occurred. The relative motion between a patient support and a component may be due to, for example, a translation of the patient support in any of the X-, Y-, or Z-planes relative to the RF coils. In some implementations, relative motion between the patient support and the component may be due to a rotation or a tilting of the patient support relative to the RF coils. The first plurality of MRI data sets may be received from image acquisition device 940, or from input device 920. In some implementations, for example those where the first plurality of MRI data sets have undergone one or more pre-processing steps, the first plurality of MRI data sets may be received from MR image processing circuitry 919. In some implementations, the first plurality of MRI data sets may be received by the computing system 910, and / or may be received by training circuitry 911. As noted above in relation to method 300, the MRI data upon which one or more of the first plurality of MRI data sets is based may have undergone one or more pre-processing steps. For example, some or all of the MRI data may have undergone one or more steps of segmentation, target identification (e.g. tumour identification), noise removal, and / or smoothing. The pre-processing steps may be carried out by the computing system 910. In some implementations, the pre-processing steps may be carried out by the MR image processing circuitry 919. The MRI data upon which each MRI data set of the first plurality of MRI data sets is based can comprise a magnetic resonance (MR) image. That is, each MRI data set of the plurality of data sets may be based on a magnetic resonance image acquired during a respective acquisition time in which relative motion between a patient support and a component of an MR imaging device occurred. For example, the first plurality of MRI data sets may comprise a plurality of MR images of one or more patients, one or more targets, or one or more regions of interest acquired during respective imaging sessions and / or radiotherapy treatments wherein relative motion between a patient support and a component of an MR imaging device occurred. In some implementations, the first plurality of MRI data sets comprises a plurality of MR images of patients, targets, and / or regions of interest acquired during respective radiotherapy treatments wherein relative motion between a patient support and a component of an MR imaging device occurred. In some implementations, one or more of the first plurality of MR images may have been acquired during helical treatments, or helical VMAT treatments. In other implementations, the MRI data upon which each MRI data set of the first plurality of MRI data sets is based can comprise first magnetic resonance (MR) signals. That is, each MRI data set of the plurality of data sets may be based on magnetic resonance signals acquired during a respective acquisition time in which relative motion between a patient support and a component of an MR imaging device occurred. For example, the first plurality of MRI data sets may comprise a plurality of MR signals of one or more patients, one or more targets, or one or more regions of interest acquired during respective imaging sessions and / or radiotherapy treatments wherein relative motion between a patient support and a component of an MR imaging device occurred. In a preferred implementation, the first plurality of MRI data sets comprises a plurality of MR signals of patients, targets, and / or regions of interest acquired during respective radiotherapy treatments wherein relative motion between a patient support and a component of an MR imaging device occurred. In some implementations, one or more of the first plurality of MRI signals may have been acquired during helical treatments, or helical VMAT treatments. The first plurality of MRI data sets forms part of the training data for the neural network. The first received MRI data set referred to in method 300 does not form part of the first plurality of MRI data sets. At block 420, a second plurality of MRI data sets are received, wherein each of the second plurality of MRI data sets comprise MRI data acquired at an acquisition time in which no relative motion between a patient support and a component of an MR imaging device occurred. MRI data acquired at an acquisition time in which no relative motion between a patient support and a component of an MR imaging device occurred may be MRI data acquired at a time where a subject (e.g. a patient, target, and / or region of interest) is static or substantially static, and / or wherein there is no or substantially no relative movement between the subject and the component of the MR imaging device. The second plurality of MRI data sets may be received from image acquisition device 940, or from input device 920. In some implementations, for example those where the second plurality of MRI data sets have undergone one or more pre-processing steps, the second plurality of MRI data sets may be received from MR image processing circuitry 919. In some implementations, the second plurality of MRI data sets may be received by the computing system 910, and / or may be received by training circuitry 911. The MRI data upon which each MRI data set of the second plurality of MRI data sets is based can comprise a magnetic resonance (MR) image. That is, each MRI data set of the plurality of data sets may be based on a magnetic resonance image acquired during a respective acquisition time in which no relative motion between a patient support and a component of an MR imaging device occurred. In other implementations, the MRI data upon which each MRI data set of the second plurality of MRI data sets is based can comprise magnetic resonance (MR) signals. That is, each MRI data set of the plurality of data sets may be based on magnetic resonance signals acquired during a respective acquisition time in which no relative motion between a patient support and a component of an MR imaging device occurred. The second plurality of MRI data sets forms part of the training data for the neural network. Each one of the second plurality of MRI data sets corresponds to a respective one of the first plurality of MRI data sets. In other words, each MRI data set in the first plurality of MRI data sets corresponds to at least one MRI data set in the second plurality of data sets, and vice versa. In some implementations, each MRI data set in the first plurality of data sets corresponds to an MRI data set in the second plurality of data sets. In other implementations, a plurality of MRI data sets from the first plurality of MRI data sets correspond to a plurality of MRI data sets in the second plurality of MRI data sets. As such, the training data for the neural network comprises groups of corresponding MRI data sets, wherein each group comprises a first MRI data set acquired at an acquisition time in which relative motion between a patient support and a component of an MR imaging device occurred, and a second MRI data set acquired at an acquisition time in which no relative motion between the patient support and the component of the MR imaging device occurred. For each group of MRI data sets in the training data, the circumstances in which the MRI data sets were acquired differ only in that the first MRI data set was acquired at an acquisition time in which relative motion between a patient support and a component of an MR imaging device occurred, and the second MRI data set was acquired at an acquisition time in which no such relative motion occurred. That is to say, the MRI data sets in a group relate to the same subject (i.e. patient, target or region of interest), include corresponding views of that subject, and were acquired using the same apparatus. Blocks 410 and 420 provide a description of how suitable training data may be collected. In a simple example, a patient may be positioned on the patient support of an MR-linac. The patient, positioned on the patient support, is translated into the bore of the machine. As this movement occurs, an MR image is acquired. This first MR image, and / or the MR signals generated during the image acquisition process, form a data set according to block 410. This data set is likely to comprise distortions due to the movement of the patient support during the image acquisition process. Another MR image may then be acquired, this time while the patient is positioned on a stationary patient support. In other words, a second MR image is acquired while the patient support is not translated. This second MR image, and / or the MR signals generated during the image acquisition process, form a data set according to block 420. In an example, this process can be repeated multiple times, preferably with many different patients, with the aim of building a body of training data comprising images and / or signals containing distortions due to the relative movement, and images and / or signals which are representative of what the MR image should or would have looked like if no relative movement had occurred. At block 430, a neural network is trained to reduce one or more distortions in the first plurality of MRI data sets, using the first plurality of MRI data sets and the corresponding second plurality of MRI data sets. In other words, the neural network is trained to reduce one or more distortions in MRI data sets caused by relative motion between the patient support and a component of an MR imaging device. Training data for the neural network comprises the first plurality of MRI data sets and the second plurality of MRI data sets, as shown in figure 5a. Figure 5a depicts a method of training a neural network according to the present disclosure. In particular, figure 5a illustrates the training of a neural network using a first plurality of MRI data sets acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and a second plurality of MRI data sets acquired during an acquisition time in which there is no such relative motion. Figure 5b depicts the trained neural network which takes as an input an MRI data set acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and outputs an updated MRI data set in which one or more distortions caused by the relative motion are reduced. Figure 6a depicts an exemplary method of training a neural network according to the present disclosure. In particular, figure 6a illustrates the training of a neural network using a first plurality of MR images acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and a second plurality of MR images acquired during an acquisition time in which there is no such relative motion. Figure 6b depicts the trained neural network which takes as an input an MR image acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and outputs an updated MR image in which one or more distortions caused by the relative motion are reduced. Figure 7a depicts another exemplary method of training a neural network according to the present disclosure. In particular, figure 7a illustrates the training of a neural network using a first plurality of MR signals acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and a second plurality of MR signals acquired during an acquisition time in which there is no such relative motion. Figure 7b depicts the trained neural network which takes as an input a first set of MR signals acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and outputs updated MR signals in which one or more distortions caused by the relative motion are reduced. Figure 8a depicts a further exemplary method of training a neural network according to the present disclosure. In particular, figure 8a illustrates the training of a neural network using a first plurality of MR signals acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and a second plurality of MR images acquired during an acquisition time in which there is no such relative motion. Figure 8b depicts the trained neural network which takes as an input a first set of MR signals acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and outputs an updated MR image in which one or more distortions caused by the relative motion are reduced. The training may be carried out by training circuitry 918 of computing system 910. A neural network is a machine learning process involving a predictive model which can be trained to learn a relationship between an input and output. A neural network comprises an input layer, an output layer and one or more interconnected hidden layers, with each hidden layer comprising a number of nodes (i.e. neurons) arranged in often complex geometries. There may be two stages to building a neural network. The first stage involves training a predictive model to learn the relationship between two data sets, using training data comprising an origin data set and a destination data set. Where it is desired to learn the relationship between two sets of images, the training data can comprise a set of origin images, and a set of destination images. Training the neural network involves using the training data to learn the parameters or weights of a predictive model representing the relationship between the origin data set and the destination data set. Each node in a neural network has an associated weight, where the weights are adjusted during the learning process to give more importance to the features which contribute more towards learning the relationship between the origin data set and the destination data set. The weights of the nodes are adjusted throughout the learning process, according to the importance of each node to the predictive process. In other words, nodes that contribute more to an accurate prediction of destination data sets associated with each origin data set are given greater weighting in the network, and nodes that contribute less to the accurate prediction of a destination data set are given lesser weighting in the network. In this way, the predictive model learns the relationship between the origin data sets and the destination data sets. In the present case, a neural network can be trained to learn the relationship between a first MRI data set acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and a second MRI data set acquired during an acquisition time in which there is no such relative motion. In other words, a neural network can be trained to learn the relationship between the first plurality of MRI data sets and the second plurality of MRI data sets as described above. In the present case, the training involves using the first and second pluralities of MRI data sets to learn the parameters or weights of a predictive model representing the relationship between the first and second pluralities of MRI data sets. Training the neural network may comprise, for each of the first plurality of MRI data sets, applying the neural network to provide an estimated distortion-reduced MRI data set. In other words, each MRI data set in the first plurality of MRI data sets is used as an input to the neural network, and an estimated distortion-reduced MRI data set is provided as an output of the neural network. The parameters or weights of the neural network are then updated or adjusted based on a comparison of the estimated distortion-reduced MRI data set with the one of the second plurality of MRI data sets which corresponds to the respective one of the first plurality of MRI data sets used as an input for the estimated distortion-reduced image. Other methods and variations of training the neural network will be familiar to those skilled in the art, and can be applied here. The trained neural network can then be used to generate an updated MRI data set wherein distortions due to the relative motion of the patient support and the component of the MR imaging device are reduced, as discussed in relation to method 300 above. In a preferred implementation, the neural network is a deep learning neural network, and may be a convolutional neural network. A deep learning neural network is a neural network which comprises many hidden layers, typically on the order of tens or up to hundreds of layers of nodes. Using a deep learning neural network allows for more accuracy and speed when solving complex problems, such as those found with image classification. In some implementations, a convolutional neural network, may be trained to learn the relationship between the first plurality of MRI data sets and the second plurality of MRI data sets as described above. A convolutional neural network (CNN) is a deep learning neural network which comprises one or more convolution layers, one or more pooling layers and a fully connected layer. Convolutional neural networks can also comprise deconvolution layers and unpooling layers. The first layer in a CNN is always a convolution layer. A convolution layer is one which applies a convolution operation to the input to the convolution layer. The convolution layer comprises a 2D or 3D filter with associated weights. The filter determines the size of the receptive field to which the convolution operation is applied. The convolution operation involves computing the dot product of the weights of the filter, and the values of the input in the receptive field. The filter is then moved across the input according to a stride length, and the convolution operation is repeated on the new receptive field. The output of each respective convolution operation forms a feature map. The weights of the filter for each convolutional layer, along with any biases, can be learned during the training of the model. For example, if the filter is an NxN filter, and the input data is an RGB image of size MxMx3, the filter is applied to an NxN area of the RGB image, wherein the pixel values in that NxN area undergo the convolution operation. The filter is then moved across the RGB image according to a stride length, and the convolution operation is applied to a new NxN area of the image. This process continues until the convolution operation has been applied to the entire RGB image. The resulting feature map contains the output of each convolution operation. An activation function, such as a ReLu function, can be applied to the feature map after each convolution operation. The feature map can indicate the importance of features in the input. In addition, a convolution layer may be followed by another convolution layer. A pooling layer, also known as a downsampling layer, can be used to reduce the size of a feature map output from a convolution layer. The pooling operation is also carried out using a filter which is moved across the feature map according to its size and stride length, but the pooling filter does not include weights. Instead, the pooling filter takes the values in the relevant receptive field and uses these values to compute a single value. The single values of each receptive field form the output of the pooling layer, reducing the size of the feature map accordingly. Several types of pooling layers may be employed in a CNN, e.g. max pooling layers, and average pooling layers. The last layer in a convolutional neural network is a fully connected layer. Each node in a fully connected layer is connected to a node of the directly preceding layer in the CNN. The fully connected layer receives the outputs of the nodes of the preceding layer, and performs a classification operation, and outputs the classification predicted for the given input. A CNN may also comprise other types of layers, for example unpooling layers, or deconvolution layers. An unpooling layer is also known as an upsampling layer, and increases the size of a feature map output from a convolution layer. A deconvolution layer performs a deconvolution operation on the input to the deconvolution layer. Unpooling layers, and deconvolution layers can be utilised to convert an input to a larger size following a convolution and / or pooling operation. In the present case, the CNN may be trained to learn the relationship between a first MRI data set acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device, and a second MRI data set acquired during an acquisition time in which there is no such relative motion. In other words, a CNN can be trained to learn the relationship between the first plurality of MRI data sets and the second plurality of MRI data sets as described above. The trained CNN can then be used to generate an updated MRI data set wherein distortions due to the relative motion of the patient support and the component are reduced. Such a CNN may comprise a first convolutional layer, one or more stacks of convolution layers grouped together, one or more pooling layers and a fully connected layer. In some implementations, the CNN may additionally comprise one or more unpooling layers, and / or one or more deconvolution layers. A first MRI data set may be used as an input to the trained CNN, and can be passed to the first convolution layer. A convolution operation is applied to the MRI data set, and a feature map is generated as an output, where the feature map indicates the importance of the features of the input, in this case the MRI data set. During the training of the CNN, the filter weights used in the convolution operation to indicate the importance of features in the input MRI data set were learned from the training data. These features may be ones which allow for the identification of distortions caused by relative motion between a patient support and a component of an MR imaging device in the first MRI data set. An activation function may be applied to the feature map, before it is passed to the next layer of the CNN. The next layer may be a pooling layer, to reduce the size of the feature map, or it may be a further convolution layer. The CNN may include many further layers comprising convolution layers, pooling layers, deconvolution layers and / or unpooling layers. The final output layer of the CNN uses a feature map or feature maps received from the one or more convolutional layer to generate an output MRI data set. In some implementations, a feature map or feature maps output from one or more convolutional layer(s) is used in combination with the input MRI data set to generate an output MRI data set. The output MRI data set is an updated MRI data set wherein one or more distortions caused by relative motion between a patient support and a component of an MR imaging device have been reduced. The convolutional neural network can be trained using the first plurality of MRI data sets, and the second plurality of MRI data sets, as defined further above. Training the CNN may comprise, for each of the first plurality of MRI data sets, applying the CNN to provide an estimated distortion-reduced MRI data set. In other words, each MRI data set in the first plurality of MRI data sets is used as an input to the CNN, and an estimated distortion-reduced MRI data set is provided as an output of the CNN. The parameters of the CNN are then updated or adjusted based on a comparison of the estimated distortion-reduced MRI data set with the one of the second plurality of MRI data sets which corresponds to the respective one of the first plurality of MRI data sets used as an input. The parameters updated comprise the weights of the filters, along with any biases, used in each of the convolutional layers of the CNN. The CNN may have one of the following architectures: LeNet, AlexNet, VGG, ResNet. Alternatively, the neural network may be a generative adversarial network, or preferably a cycleGAN. A generative adversarial network (GAN) comprises two models, a generator model and a discriminator model. Each model may be a respective convolutional neural network, with features as described above. The generator model is trained to generate new data sets based on training data. The discriminator model then determines whether a new data set generated by the generator model is the same as or different to the ground truth data in the training data. The aim of the GAN is for the generator network to learn the predictive model to an accuracy where the discriminator model can only correctly identify whether a data set is generated by the generator model or is from the ground truth training data 50% of the time. Once the generative model has been trained, it can be used to generate new data sets according to the predictive model. In the present case, the generator model generates new MRI data sets based on the first plurality of MRI data sets in the training data. The discriminator model then determines whether the new MRI data sets generated by the generator model are the same as or different to the updated MRI data sets in the second plurality of MRI data sets in the training data. Only the discriminator model has access to the second plurality of MRI data sets. In other words, the generator model will take an MRI data set from the first plurality of MRI data sets, and use it to generate a new data set. The discriminator model is then used to determine whether that data set is the same as or different to the corresponding updated MRI data set found in the second plurality of MRI data sets in the training data. The trained generator model is one that has learned the predictive model to an accuracy where the discriminator model can only correctly identify whether a data set is generated by the generator model or is from the training data 50% of the time. Once the generative model has been trained, it can be used to generate updated MRI data sets wherein distortions caused by relative motion of a patient support and a component of an MR imaging device have been reduced. A cycleGAN comprises two GANs, and thus comprises a first generator model, a second generator model, a first discriminator model and a second discriminator model. An advantage of a cycleGAN is that the training data does not need to be paired. In a cycleGAN, the training data comprises two categories of unpaired data, which will be referred to as a first data set and a second data set for the purpose of this discussion. The first generator model generates new data sets using the first data set as an input, and the first discriminator model determines whether a new data set generated by the first generator model is the same as or different to the ground truth data in the second data set. Conversely, the second generator model generates new data sets using the second data set as an input, and the second discriminator model determines whether a new data set generated by the second generator model is the same as or different to the ground truth data in the first data set. Again, the aim of each GAN is for the generator network to learn the predictive model to an accuracy where the discriminator model can only correctly identify whether a data set is generated by the generator model or is from the ground truth training data 50% of the time. In addition, each output generated by the first generator model can be used as an input to the second generator model. That is, the second generator model receives an output from the first generator model as an input, and uses it to generate a new data set. The output from the second generator model can then be compared to the initial data set used as an input for the first generator model, and the weights of the first and second GAN can be adjusted accordingly. The new data set is referred to as a ‘cycle’ data set, and can be used to compute cycle losses, and thus to adjust the weights of each of the first and second GAN. Each output generated by the second generator model can be used as an input to the first generator model in a similar fashion. Once each of the first and second generator model have reached the required accuracy, the generator models can be used to generate new data sets corresponding to each of the second data set and first data set respectively. In some implementations of the present invention, the neural network is a cycleGAN, and thus comprises two GANs. Each GAN may be a trained convolutional network. In the present case, the training data comprises the first plurality of MRI data sets, and the second plurality of MRI data sets, as defined further above. The first generator model generates updated MRI data sets using data sets from the first plurality of MRI data sets as an input, and the first discriminator model determines whether an updated MRI data set generated by the first generator model is the same as or different to the ground truth data in the second plurality of MRI data sets. In other words, the first generator model generates an updated data set using a data set acquired during an acquisition time in which there is relative motion between a patient support and a component of an MR imaging device as an input. The first discriminator model then determines whether the updated MRI data set is the same as or different to an MRI data set acquired during a time when there was no such relative motion. The second generator model generates new data sets using data sets from the second plurality of data sets as an input, and the second discriminator model determines whether a new data set generated by the second generator model is the same as or different to the ground truth data in the first plurality of MRI data sets. Again, the aim of each GAN is for the generator network to learn the predictive model to an accuracy where the discriminator model can only correctly identify whether a data set is generated by the generator model or is from the ground truth training data 50% of the time. In other words, the second generator model generates a new data set using a data set acquired during an acquisition time in which was no relative motion between a patient support and a component of an MR imaging device as an input. The second discriminator model then determines whether the new MRI data set is the same as or different to an MRI data set acquired during an acquisition time when there was relative motion between a patient support and a component of an MR imaging device. In addition, each output generated by the first generator model can be used as an input to the second generator model. That is, the second generator model receives an output from the first generator model as an input, and uses it to generate a new data set. The output from the second generator model can then be compared to the initial MRI data set used as an input for the first generator model, and the weights of the first and second GAN can be adjusted accordingly. Each output generated by the second generator model can be used as an input to the first generator model in a similar fashion. Once each of the first and second generator model have reached the required accuracy, the generator models can be used to generate new data sets corresponding to each of the second data set and first data set respectively. In particular, the first generator model is trained to generate an updated MRI data set using data sets from the first plurality of MRI data sets as an input, i.e. the first generator model is trained to generate an updated MRI data set wherein distortions due to relative motion between a patient support and a component of an MR imaging device are reduced. Figure 9 illustrates a block diagram of one implementation of a radiotherapy system 200 / 900. The radiotherapy system 900 comprises a computing system 910 within which a set of instructions, for causing the computing system 910 to perform any one or more of the methods discussed herein, may be executed. The computing system 910 shall be taken to include any number or collection of machines, e.g. computing device(s), that individually or jointly execute a set (or multiple sets) of instructions to perform any one or more of the methods discussed herein. That is, hardware and / or software may be provided in a single computing device, or distributed across a plurality of computing devices in the computing system. In some implementations, one or more elements of the computing system may be connected (e.g., networked) to other machines, for example in a Local Area Network (LAN), an intranet, an extranet, or the Internet. One or more elements of the computing system may operate in the capacity of a server or a client machine in a client-server network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. One or more elements of the computing system may be a personal computer (PC), a tablet computer, a set-top box (STB), a Personal Digital Assistant (PDA), a cellular telephone, a web appliance, a server, a network router, switch or bridge, or any machine capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by that machine. The computing system 910 includes controller circuitry 911 and a memory 913 (e.g., readonly memory (ROM), flash memory, dynamic random access memory (DRAM) such as synchronous DRAM (SDRAM) or Rambus DRAM (RDRAM), etc.). The memory 913 may comprise a static memory (e.g., flash memory, static random access memory (SRAM), etc.), and / or a secondary memory (e.g., a data storage device), which communicate with each other via a bus (not shown). Controller circuitry 911 represents one or more general-purpose processors such as a microprocessor, central processing unit, accelerated processing units, or the like. More particularly, the controller circuitry 911 may comprise a complex instruction set computing (CISC) microprocessor, reduced instruction set computing (RISC) microprocessor, very long instruction word (VLIW) microprocessor, processor implementing other instruction sets, or processors implementing a combination of instruction sets. Controller circuitry 911 may also include one or more special-purpose processing devices such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), network processor, or the like. One or more processors of the controller circuitry may have a multicore design. Controller circuitry 911 is configured to execute the processing logic for performing the operations and steps discussed herein. The computing system 910 may further include a network interface circuitry 915. The computing system 910 may be communicatively coupled to an input device 920 and / or an output device 930, via input / output circuitry 917. In some implementations, the input device 920 and / or the output device 930 may be elements of the computing system 910. The input device 920 may include an alphanumeric input device (e.g., a keyboard or touchscreen), a cursor control device (e.g., a mouse or touchscreen), an audio device such as a microphone, and / or a haptic input device. The output device 930 may include an audio device such as a speaker, a video display unit (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)), and / or a haptic output device. In some implementations, the input device 920 and the output device 930 may be provided as a single device, or as separate devices. In some implementations, computing system 910 includes training circuitry 918. The training circuitry 918 is configured to train a method of generating an updated magnetic resonance imaging (MRI) data set. In particular, training circuitry 918 may train a model to perform a method of generating an MRI data set using a neural network trained to reduce one or more distortions in MRI data sets, the distortions being caused by relative motion between a patient support and a component of an MR imaging device. The model may comprise a deep neural network (DNN), such as a convolutional neural network (CNN) and / or recurrent neural network (RNN). Training circuitry 918 may be configured to execute instructions to train a model that can be used to generate an updated magnetic resonance imaging (MRI) data set, wherein distortions caused by relative motion between a patient support and a component of an MR imaging device, as described with reference to figure 4. Training circuitry 918 may be configured to access training data and / or testing data from memory 913 or from a remote data source, for example via network interface circuitry 915. In some examples, training data and / or testing data may be obtained from an external component, such as image acquisition device 940 and / or treatment device 950. In some implementations, training circuitry 918 may be used to update, verify and / or maintain the model for performing a method of generating an MRI data set using a neural network trained to reduce one or more distortions in MRI data sets, the distortions being caused by relative motion between a patient support and a component of an MR imaging device. In some implementations, the computing system 910 may comprise MR image processing circuitry 919. MR image processing circuitry 919 may be configured to process MR image data 980 (e.g. images, MR signals or imaging data), such as medical images obtained from one or more MR imaging data sources, a treatment device 950 and / or an MR image acquisition device 940. MR image processing circuitry 919 may be configured to process, or pre-process, MR image data. For example, MR image processing circuitry 919 may convert received image data into a particular format, size, resolution or the like. In some implementations, MR image processing circuitry 919 may be combined with controller circuitry 911. Image acquisition device 940 may be configured to output image data 980, which may be accessed by computing system 910. Treatment device 950 may be configured to output treatment data 960, which may be accessed by computing system 910. Computing system 910 may be configured to access or obtain treatment data 960, planning data 970 and / or image data 980. Treatment data 960 may be obtained from an internal data source (e.g. from memory 913) or from an external data source, such as treatment device 950 or an external database. Planning data 970 may be obtained from memory 913 and / or from an external source, such as a planning database. Planning data 970 may comprise information obtained from one or more of the image acquisition device 940 and the treatment device 950. The various methods described above may be implemented by a computer program. The computer program may include computer code (e.g. instructions) arranged to instruct a computer to perform the functions of one or more of the various methods described above. The steps of the methods described above may be performed in any suitable order. For example, step 310 of method 300 may be performed before, after, simultaneously or substantially simultaneously with step 320. The computer program and / or the code for performing such methods may be provided to an apparatus, such as a computer, on one or more computer readable media or, more generally, a computer program product)). The computer readable media may be transitory or non-transitory. The one or more computer readable media could be, for example, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, or a propagation medium for data transmission, for example for downloading the code over the Internet. Alternatively, the one or more computer readable media could take the form of one or more physical computer readable media such as semiconductor or solid state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disc, and an optical disk, such as a CD-ROM, CD-R / W or DVD. The instructions may also reside, completely or at least partially, within the memory 913 and / or within the controller circuitry 911 during execution thereof by the computing system 910, the memory 913 and the controller circuitry 911 also constituting computer-readable storage media. In an implementation, the modules, components and other features described herein can be implemented as discrete components or integrated in the functionality of hardware components such as ASICS, FPGAs, DSPs or similar devices. A “hardware component” is a tangible (e.g., non-transitory) physical component (e.g., a set of one or more processors) capable of performing certain operations and may be configured or arranged in a certain physical manner. A hardware component may include dedicated circuitry or logic that is permanently configured to perform certain operations. A hardware component may comprise a special-purpose processor, such as an FPGA or an ASIC. A hardware component may also include programmable logic or circuitry that is temporarily configured by software to perform certain operations. In addition, the modules and components can be implemented as firmware or functional circuitry within hardware devices. Further, the modules and components can be implemented in any combination of hardware devices and software components, or only in software (e.g., code stored or otherwise embodied in a machine-readable medium or in a transmission medium). Unless specifically stated otherwise, as apparent from the following discussion, it is appreciated that throughout the description, discussions utilizing terms such as "receiving”, “determining”, “comparing ”, “enabling”, “maintaining,” “identifying,” or the like, refer to the actions and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantities within the computer system memories or registers or other such information storage, transmission or display devices. It will be understood that the above description of specific embodiments is by way of example only and is not intended to limit the scope of the present disclosure. Many modifications of the described embodiments, some of which are now described, are envisaged and intended to be within the scope of the present disclosure. It is to be understood that the above description is intended to be illustrative, and not restrictive. Many other implementations will be apparent to those of skill in the art upon reading and understanding the above description. Although the present disclosure has been described with reference to specific example implementations, it will be recognized that the disclosure is not limited to the implementations described, but can be practiced with modification and alteration within the spirit and scope of the appended claims. Accordingly, the specification and drawings are to be regarded in an illustrative sense rather than a restrictive sense. The scope of the disclosure should, therefore, be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.

Claims

1. A computer-implemented method for generating an updated magnetic resonance imaging, MRI, data set using a neural network trained to reduce one or more distortions in MRI data sets, the distortions being caused, at least in part, by relative motion between a patient support and a component of an MR imaging device, the method comprising:receiving a first MRI data set, wherein the first MRI data set is based on MRI data acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device occurred;applying the received first MRI data set as an input to the trained neural network; andconverting, using the trained neural network, the received first MRI data set to an updated MRI data set in which distortions in the first MRI data set are reduced.

2. The method of claim 1, wherein the relative motion between the patient support and the component of the MR imaging device is caused, at least in part, by translation of the patient support with respect to the component.

3. The method of claim 2 wherein the translation occurred during delivery of helical volumetric modulated arc therapy treatment.

4. The method of any preceding claim, wherein the received first MRI data set comprises:a first magnetic resonance image; or first magnetic resonance signals.

5. The method of any preceding claim, wherein the updated MRI data set comprises one of:an updated magnetic resonance image; orupdated magnetic resonance signals.

6. The method of any preceding claim, wherein the neural network has been trained using a first and a second plurality of MRI data sets, each of the first plurality of MRI data sets comprising MRI data acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device occurred, and each of the second plurality of MRI data sets comprising MRI data acquired at an acquisition time in which no relative motion between the patient support and the component of the MR imaging device occurred; andwherein each one of the second plurality of MRI data sets corresponds to a respective one MRI data set of the first plurality of MRI data sets.

7. The method of any preceding claim, further comprising:receiving a first plurality of MRI data sets, each of the first plurality of MRI data sets comprising MRI data acquired during an acquisition time in which relative motion between the patient support and the component of the MR imaging device occurred;receiving a second plurality of MRI data sets, each of the second plurality of MRI data sets comprising MRI data acquired at an acquisition time in which no relative motion between the patient support and the component of the MR imaging device occurred;each one of the second plurality of MRI data sets corresponding to a respective one MRI data set of the first plurality of MRI data sets;training a neural network using the corresponding first plurality of MRI data sets and second plurality of MRI data sets to reduce one or more distortions in the first plurality of MRI data sets.

8. The method of any preceding claim, wherein training the neural network comprises, for each of the first plurality of MRI data sets:applying the neural network to provide an estimated distortion-reduced MRI data set; andupdating parameters of the neural network based on a comparison of the estimated distortion-reduced MRI data set with the one of the second plurality of MRI data sets which corresponds to the respective one of the first plurality of MRI data sets.

9. The method of any of claims 6 to 8, wherein the MRI data sets of the first plurality of MRI data sets are one of:magnetic resonance images; or magnetic resonance signals.

10. The method of any of claims 6 to 9, wherein the MRI data sets of the second plurality of MRI data sets are one of:magnetic resonance images; or magnetic resonance signals.

11. The method of any of claims 6 to 10, wherein the first plurality of MRI data sets does not comprise the received first MRI data set.

12. The method of any preceding claim, wherein converting the received MRI data set to an updated MRI data set using the trained neural network further comprises outputting the updated MRI data set.

13. The method of any preceding claim, further comprising using the updated MRI data set in a treatment plan.

14. The method of any preceding claim, further comprising using the updated MRI data set in real time cine imaging to monitor a target position.

15. The method of any preceding claim, wherein the updated MRI data set is a corrected MR I data set.

16. The method of claim 15, wherein the corrected MRI data set is one in which the distortions caused, at least in part, by the relative motion between the patient support and the component of the MR imaging device have been corrected.

17. The method of any preceding claim, wherein the neural network is a deep learning neural network.

18. The method of claim 17, wherein the deep learning neural network is a convolutional neural network.

19. The method of any of claims 1 to 16, wherein the neural network is a generative adversarial network (GAN).

20. The method of claim 19, wherein the neural network is a cycleGAN.

21. The method of any preceding claim, wherein the component of the MR imaging device comprises one or more RF coils.

22. The method of claim 21, wherein the one or more RF coils comprise an RF transmitter coil.

23. The method of claim 21 or 22, wherein the one or more RF coils are fixed.

24. The method of any of claims 21 to 23, wherein the one or more RF coils are fixed inside the bore of the MR imaging device.

25. A system for generating an updated magnetic resonance imaging, MRI, data set using a neural network trained to reduce one or more distortions in MRI data sets, the distortions being caused, at least in part, by relative motion between a patient support and a component of an MR imaging device, the system comprising:a memory storing machine executable instructions and the trained neural network;a processor for controlling the system, wherein execution of the machine executable instructions causes the processor to control the system to perform the method of any of claims 1 to 24.

26. A computer-readable medium containing instructions which, when executed by the processor, cause the processor to perform the method of any of one of claims 1 to 24.

Images