Device and method for stabilizing magnetic resonance-based image acquisition using artificial intelligence

Abstract

The invention relates to a magnetic resonance imaging (MRI) scanner with artificial intelligence for stabilizing image acquisition, a method for operating the MRI scanner, and a method for training the artificial intelligence. The MRI scanner has a control unit that sends a control signal to an output unit to initiate image acquisition and generate an output signal for the image acquisition. It also includes a pre-distortion device with artificial intelligence that takes into account different environmental parameters to minimize deviations in the output signal.

Description

The invention relates to a magnetic resonance tomography system having an artificial intelligence for stabilizing the image acquisition, a method for operating the magnetic resonance tomography system and a method for training the artificial intelligence. The magnetic resonance tomography apparatus comprises a control unit which outputs a control signal for performing image acquisition to an output unit to generate an output signal for image acquisition, and a pre-distortion device having an artificial intelligence to minimize deviations of the output signal.Magnetic resonance tomography apparatuses are imaging devices which, for imaging an object under examination, align nuclear spins of the object under examination with a strong external magnetic field and excite them by means of an alternating magnetic field for precession about this alignment. The precession or return of the spins from this excited to a state of lower energy in turn generates an alternating magnetic field in response which is received via antennas.With the aid of magnetic gradient fields, spatial coding is imposed on the signals, which then enables an assignment of the received signal to a volume element. The received signal is then evaluated and a three-dimensional imaging representation of the examination object is provided. For receiving the signal, local receiving antennas, so-called local coils, are preferably used, which are arranged directly on the object under examination to achieve a better signal-to-noise ratio. The receiving antennas can also be installed in a patient couch.For imaging, the highest demands are made on homogeneity and constancy of the magnetic and electromagnetic fields. At the same time, however, the high powers and currents used lead to changes in the signals. Firstly, short-term changes result from thermal effects such as expansion and the associated changes and also changes in the resistance of the materials. Furthermore, long-term changes result from aging processes in semiconductors or contact points. All these changes interact in a complex manner, so that their effects can hardly be predicted and compensated by analytical methods.It is therefore an object of the present invention to provide a magnetic resonance tomography system with more stable operating behavior.The object is achieved by a magnetic resonance tomography system according to the invention as claimed in claim 1, a method for operating the magnetic resonance tomography system as claimed in claim 8 and a method for training an artificial intelligence as claimed in claim 9.The magnetic resonance tomography system according to the invention has a control unit which is designed to output a control signal for carrying out an image acquisition using the magnetic resonance tomography system to an output unit.The output unit can be, for example, a radio-frequency power amplifier for amplifying the excitation pulses or else gradient amplifiers for generating the currents for the gradient coil. A current source for shim coils for compensating magnetic field inhomogeneities can also be the output unit. Likewise, an oscillator for generating the high-frequency signal for the high-frequency power amplifier is a possible output unit, or a combination of oscillator and high-frequency power amplifier.The control signal can be a digital or an analog signal. In particular, the control signal can be a digital signal which is supplied to a D / A converter or else its analog output signal. However, an analog signal is also to be understood as a control signal that is not generated directly by a D / A converter, but is generated in an analog manner, for example by analog mixing of an oscillator signal with another analog signal.The output unit outputs an output signal for image acquisition in response to the control signal. The output signal can be, for example, as described above, a radio-frequency excitation pulse for the nuclear spins or a current for generating gradient fields or shim fields.The magnetic resonance tomography system has a pre-distortion device. A pre-distortion device is understood to mean a device which changes the control signal such that the output unit with the changed control signal outputs an output signal which differs from an output signal predetermined by the control signal to a lesser extent than if the unchanged control signal were supplied to the output unit. For example, many output units are idealized through linear systems, i.e., the output signal is to be proportional to the control or input signal by a constant factor. Deviations can occur as a result of a change in the factor or as a result of non-linearities which are associated with the control signal itself, its previous history or, as explained below with respect to a possible embodiment, one or more state parameters. The predistortion device changes the supplied control signal in such a way that it supplies the output unit in the changed form, and generates an output signal that deviates less from the ideal output signal, in this example linearly or in proportion to the original control signal. The deviation can be defined, for example, by a square sum of the deviation between the ideal output signal and the real output signal integrated over a predetermined period of time.In one possible embodiment, the adaptation of the control signal or the generation of the output signal takes place as a function of a state parameter of the magnetic resonance tomography system. The state parameter, as described below, indicates a variable property of the magnetic resonance tomography apparatus that influences the output signal.The adaptation of the control signal can, however, also take place as a function of the control signal itself, including its immediate previous history, e.g. the last previous output signal or parts of the control signal in the immediate past of this output unit, its temporal interval and / or course.The predistortion device of the magnetic resonance tomography system according to the invention has an artificial intelligence, preferably realized by a neural network, which is designed to perform the predistortion by application to the control signal, so that the control signal is changed, as described above, in such a way that the output signal deviates less from a predetermined output signal compared to an output signal of the output unit for an unchanged control signal.Advantageously, an artificial intelligence or a neural network can be made by training to compensate for influences which can hardly be controlled with conventional control techniques on account of the relationships which are complex and analytically difficult to acquire.The method according to the invention for operating the magnetic resonance tomography system shares the advantages of the magnetic resonance tomography system according to the invention.The method according to the invention for training the artificial intelligence of the magnetic resonance tomography system is provided in order to provide the artificial intelligence which optimizes the properties of the control signal in the pre-distortion device according to the invention.The method comprises a step in which a predetermined control signal is provided. Predetermined control signals of this type can be obtained, for example, by logging from the operation of magnetic resonance tomography systems. These are preferably magnetic resonance tomography units which use output units comparable in terms of construction and performance. The predetermined control signals are preferably from comparable operating situations in which the magnetic resonance tomography apparatus according to the invention is to be used. For example, sequences that are usually used may differ depending on the location or intended use of the magnetic resonance tomography system. In orthopedic practice, for example, sequences other than in neurology or internal medicine are used.In a further step of the method according to the invention, the control signal is output via the output unit. For this purpose, the control signal is supplied to the output unit via the predistortion device with the artificial intelligence. It is conceivable as an initial state that the pre-distortion device first forwards the control signal transparently to the output unit or as an initial state. However, it is also conceivable for the artificial intelligence to be preconfigured with a parameter set which was obtained, for example, by training with a magnetic resonance tomography system with output units different in hardware or in other application scenarios or which originates from a simulation of the hardware.In one possible embodiment, while the predetermined control signal is output, a state parameter of the output unit or of the magnetic resonance tomography unit is acquired which has an influence on the output signal of the output unit. State parameters can be, for example, temperature or radio-frequency load through the body tissue of a patient, given by its weight, but also operating duration or operating parameters of other output units, which interact, for example, via the power supply or via induced currents in the construction of the magnetic resonance tomography system. Also conceivable as state parameters are parameters which are set by the control unit for image acquisition, such as an amplitude of the high-frequency signal or of the gradient currents. Conceivable state parameters can also be vibrations or stray fields detected by sensors. Properties of the acquired magnetic resonance signals, k-space data or reconstructed images can also be provided by an evaluation unit as state parameters.The control signal differs from one of the state parameters in that the predistortion device merely modifies and outputs the control signal. The modified control signal is similar to the original control signal in temporal variation and spectral distribution. For example, it can arise from the control signal by multiplication by a variable factor, the factor depending on a state parameter, but also on its previous history or the previous history of the control signal itself. A shift in the frequency range or a temporal shift / phase shift depending on a state parameter or the previous history of the control signal is also conceivable.The control signal can also itself take on the role of a state parameter for other control signals, as the control signal another control signal is modified as an input or state parameter depending on the control signal.In one possible embodiment, the state parameter of the artificial intelligence is provided as an input parameter, so that the artificial intelligence can carry out the pre-distortion of the control signal as a function of the state parameter, i.e. the state parameter with suitable setting of the parameters of the artificial intelligence, the output signal of which is dependent on the state parameter.In a further step, the output signal of the output unit is compared with a predetermined output signal in response to the predetermined control signal. The predetermined output signal can be considered, for example, as an ideal output signal which arises in the case of a purely linear response of the output unit as a result of a scaling of the control signal. It is also conceivable that the predetermined output signal already takes into account known and preferably deterministically determinable effects, for example a temperature dependence of the resistance of the gradient coils.The comparison of predetermined output signal and output signal output by the output unit can be carried out, for example, by forming the difference between measured values of a sensor for the measured values of the sensor to be expected at the output of the output unit with measured values of the sensor to be expected for a predetermined output signal. For the high-frequency power amplifier, for example, a sensor for monitoring SAR limit values is already present.In a further step, one or more parameters of the artificial intelligence are adapted in such a way that a deviation of the output signal from the predetermined output signal is reduced. In other words, the parameters of the artificial intelligence are set so that the difference of the predetermined output from the predetermined control signal output unit decreases. For neural networks, this can be achieved, for example, by the known method of back propagation.In one embodiment of the method according to the invention, the decrease in the deviation and the adaptation of the parameters take place taking into account at least one detected state parameter supplied to the artificial intelligence.In this case, the artificial intelligence parameter or parameters are adapted in such a way that the distortion device changes the control signal in such a way that the deviation of the output signal is reduced below different values for the state parameters.In one possible embodiment, this takes place as a function of the detected state parameter, so that the deviation of the output signal is reduced below different values for the state parameters.Finally, the preceding steps are repeated for a plurality of different predetermined control signals with associated predetermined output signals and / or a plurality of different state parameters.Finally, the trained artificial intelligence is provided for the magnetic resonance tomography system according to the invention or for a plurality of corresponding magnetic resonance tomography systems. This can be effected, for example, by the determined parameters being provided on a storage medium. The parameters are then loaded into a processor or its memory or also into dedicated hardware of the artificial intelligence, for example a neural network. The parameters can also be an integral part of a stored program which, during execution, executes the artificial intelligence and the method according to the invention for operation in the control unit.The method according to the invention advantageously makes it possible to control the output signal more accurately for different scenarios and operating states even without analytical knowledge of the relationships of state parameters and output signal and thus to achieve better imaging.Further advantageous embodiments are specified in the dependent claims.In one conceivable embodiment of the magnetic resonance tomography apparatus according to the invention, the output unit is a unit from a group of radio-frequency line amplifiers, gradient amplifiers, magnetic shim control or signal synthesizers.The pre-distortion device advantageously makes it possible to compensate for wear phenomena on units which are subject to particularly high loads and / or which are particularly influenceable for the image acquisition.In a possible embodiment of the magnetic resonance tomography apparatus according to the invention, the state parameter or the state variable is a temperature of the output unit or of a part of the output unit or of the environment. The state parameter can also be a magnetic field strength, for example that of the B0 field, a gradient field, a B1 alternating field or else of an external magnetic field in the surroundings of the magnetic resonance tomography system. A further conceivable state parameter is a power consumption, for example of the output unit, of another output unit, of the cooling unit, a sum of a plurality of these units or of the entire magnetic resonance tomography system. Another conceivable state parameter can also be an operating duration of one or more output units, for example since startup or since the last startup.These state parameters have an influence on the signal processing of the various output units, so that detection and consideration enables a better quality of the output signals.In one conceivable embodiment of the magnetic resonance tomography system according to the invention, the neural network can be of the recurrent neural network (RNN) type, which, by feeding signals in one layer or in a preceding layer back, is capable of taking account of the temporal behavior of the control signal and / or the state parameters, that is to say carries a memory for these signals.In particular, an RNN of the long short-term memory (LSTM) network is suitable in its simplified form, in which the temporal parameter is no longer trained, but rather is set to a predetermined value on the basis of the knowledge about the signals to be processed and their temporal profiles or characteristics. It is also conceivable that temporal parameters in such a network are fixed to a predetermined time window and are trained only within this time window. The LSTM in this case has in particular two separate time windows, a long-term window and a short-term window, wherein the long-term window relates to time periods which relate to substantially greater time periods relative to the short-term window, for example are longer by more than the factor 5, 10, 50 or 100. The time windows are in particular limited to a maximum time duration taken into account, but also to a minimum time duration.An RNN advantageously allows the time curves or previous history of the signals to be taken into account. In particular, it is thus conceivable to apply the distortion device also solely to a control signal without further state parameters.The LSTM network, in turn, can be designed more simply and thus faster during the passing through or the playing out of the signals to be processed, on account of the prior knowledge about the time slots of the prior history to be taken into account. In particular, an LSTM has both a long-term memory and a short-term memory, the time response of which, in particular respective time windows, are also adapted during training. As a result, compared to a neural network with a continuous memory over all time ranges, the circuit outlay and the cycle time are reduced.It is thus possible to also change fast signals, such as radio-frequency signals of an excitation pulse, directly by the neural network and at the same time to keep the outlay for the hardware of the neural network and its energy requirement within limits.In a possible embodiment of a magnetic resonance tomography system according to the invention, the pre-distortion device is designed to change the control signal as a function of a plurality of state parameters. In other words, a plurality or all of the available state data are supplied to the predistortion device and the predistortion of the control signal takes place as a function of a plurality or all of the state parameters. Accordingly, in an implementation of the predistortion device according to the invention with a neural network, all state parameters and the control signal are supplied to the neural network.In one conceivable embodiment of the magnetic resonance tomography system according to the invention, the pre-distortion device has a filter. A filter is referred to as a circuit or digital function that changes the control signal beyond a linear amplification or scaling. These can be, in particular, non-linear changes in the amplitude, changes in the spectrum of the control signal in the frequency space and / or in the time domain. The filter can be, for example, amplitude compressors, low-pass filters, high-pass filters or band-pass filters. More complex filters such as leaky bucket filters and combinations of the aforementioned functions are also conceivable.The filter is configured to generate a weighted average for a state parameter. The mean value contains, in particular, information about the past of the state parameter. The predistortion device is designed to change the control signal as a function of the weighted mean value, i.e. to take the state parameter into account when changing the control signal.It is also conceivable that the filter is an integral component of the artificial intelligence or of the neural network.The filter advantageously provides a memory with the weighted mean value of the predistortion device, so that the previous history of the magnetic resonance tomography system can be taken into account in the correction of the control signal and thus of the output signal.In a possible embodiment of the method according to the invention, the steps of providing a predetermined control signal, outputting the predetermined control signal, comparing the output signal of the output unit and adapting the parameters of the artificial intelligence with a plurality of different sequences of the different control signals are repeated. In other words, in the training, not only different control signals are fed to the predistortion device, but the predetermined control signals are fed in different orders. Preferably, the predetermined control signals are executed in pseudo random sequences for training in order to prevent the artificial intelligence from being optimized to a sequence occurring only during training.Training with different sequences advantageously leads to the predistortion device functioning reliably even in different application scenarios and under changing conditions.In one conceivable embodiment of the method according to the invention, the magnetic resonance tomography system has a plurality of output units. For example, in a magnetic resonance tomography system, as already described, this is the radio-frequency power amplifier, the gradient amplifiers or the shim control. In the embodiment of the method according to the invention, a plurality of predetermined control signals is provided and output via the at least one predistortion device by means of the plurality of output units while detecting a state parameter. It is conceivable that one pre-distortion device is provided for each output unit. However, it is likewise possible for a predistortion device to have a plurality of inputs for control signals and a plurality of signal outputs for a signal connection to the plurality of output units, so that a predistortion device supplies or controls a plurality of output units.The plurality of output signals of the plurality of output units is compared in one step with a plurality of predetermined output signals for the output units. This takes place as already described above for the comparison of an individual output signal.In a further step, the parameters of the at least one artificial intelligence are adapted in such a way that a deviation of the output signals from the predetermined output signals is reduced. For predistortion devices which each receive only one control signal and thus control an output unit, this is done as already described. For predistortion devices having a plurality of signal inputs and signal outputs which actuate a plurality of output units, there are corresponding multidimensional adaptation methods, for example a multidimensional variant of back propagation for a multidimensional neural network.The method according to the invention advantageously enables optimization for a plurality of output units. In particular, if a predistortion device jointly supplies a plurality of output units, this enables a consideration of the interactions and thus a further optimization.In one possible embodiment of the method according to the invention, a phantom is arranged in a recording area of the magnetic resonance tomography system. The deviation of the output signal of the output unit from the predetermined output signal is then determined on the basis of an image acquisition of the phantom. For example, it is possible to evaluate the flip angle, which is output by an excitation pulse via the high-frequency power amplifier, by means of the captured image of the phantom.A phantom advantageously enables an analysis of the entire signal chain and thus an optimal correction of the imaging.In another conceivable embodiment of the magnetic resonance tomography system according to the invention, the magnetic resonance tomography system has a comparison device. The comparison device is designed to determine a deviation of the at least one output signal of the at least one output unit from the predetermined at least one output signal. This can be effected, for example, by forming the difference between the signals. Preferably, the data quantity in its output signal is reduced compared to the input signals by the comparison device.The comparison device provides the result of the artificial intelligence as adaptation parameters via a signal connection and thus simplifies the evaluation by the artificial intelligence and allows simpler, more economical and faster architectures.The above-described characteristics, features and advantages of this invention and the manner in which these are achieved become clearer and more clearly comprehensible in conjunction with the following description of the exemplary embodiments, which are explained in more detail in conjunction with the drawings.The following are shown: FIG. 1 shows a schematic representation of an exemplary magnetic resonance tomography system according to the invention; FIG. 2 shows a schematic illustration of the signal flow in an embodiment of the magnetic resonance tomography system according to the invention; FIG. 3 shows a schematic flow diagram of an embodiment of a method according to the invention for training artificial intelligence of the magnetic resonance tomography system according to the invention; FIG. 4 shows a schematic flow chart for an embodiment of the method according to the invention for operating the magnetic resonance tomography system according to the invention.FIG. 1 shows a schematic representation of an embodiment of a magnetic resonance tomography system 1 according to the invention.The magnet unit 10 of the magnetic resonance tomography system 1 has a field magnet 11 which generates a static magnetic field B0for aligning nuclear spins of samples or of the patient 100 in a recording region. The receiving region is characterized by an extremely homogeneous static magnetic field B0, wherein the homogeneity relates in particular to the magnetic field strength or the amount. The receiving region is virtually spherical and is arranged in a patient tunnel 16 which extends through the magnet unit 10 in a longitudinal direction 2. A patient bed 30 is movable in the patient tunnel 16 by the displacement unit 36. Typically, the field magnet 11 is a superconducting magnet that can provide magnetic fields having a magnetic flux density of up to 3T, even higher in recent devices. However, permanent magnets or electromagnets with normally conducting coils can also be used for lower magnetic field strengths.Furthermore, the magnet unit 10 has gradient coils 12 which are designed to superimpose temporally and spatially variable magnetic fields in three spatial directions on the magnetic field B0for the spatial differentiation of the captured imaging regions in the examination volume. The gradient coils 12 are usually coils made of normally conducting wires which can generate fields in the examination volume that are orthogonal to one another.The magnet unit 10 also has a body coil 14 which is designed to emit a radio-frequency signal, which is supplied via a signal line, into the examination volume and to receive resonance signals emitted by the patient 100 and to emit them via a signal line.A control unit 20 supplies the magnet unit 10 with the various signals for the gradient coils 12 and the body coil 14 and evaluates the received signals.Thus, the control unit 20 has the gradient controller 21, which is designed to supply the gradient coils 12 via supply lines with variable currents, which provide the desired gradient fields in the examination volume in a temporally coordinated manner.Furthermore, the control unit 20 has a radio-frequency unit 22, which is designed to generate a radio-frequency pulse with a predefined temporal profile, amplitude and spectral power distribution for exciting a magnetic resonance of the nuclear spins in the patient 100. In this case, pulse powers in the range of kilowatts can be achieved. The excitation signals can be emitted into the patient 100 via the body coil 14 or also via a local transmitting antenna.A central controller 23 of the control unit 20 communicates via a signal bus 25 with the gradient controller 21 and the radio frequency unit 22 in order to coordinate the individual units for image acquisition.For receiving the magnetic resonance signal, a local coil 50 according to the invention is arranged on the patient 100 in the patient tunnel 16 in order to acquire magnetic resonance signals from an examination region in the immediate vicinity with the largest possible signal-to-noise ratio. The local coil 50 is in signal connection via a connecting line 33 with a receiver in the high-frequency unit 22. However, the signal connection can also be wireless.The magnetic resonance tomography system 1 has different output units. This can be, for example, a high-frequency line amplifier 24 of the high-frequency unit 22. This is used in the generation of an excitation pulse in order to amplify a small signal of a signal synthesizer of the radio-frequency unit 22. The properties of the output signal change, for example, depending on the temperature of the high-frequency power amplifier 24 or on a variable load of the high-frequency power amplifier 24 by the patient 100 as a state parameter.The signal synthesizer can likewise be an output unit within the meaning of the invention. For example, a preferably digital control command of the control unit 20 can be used to preset the frequency, phase and / or amplitude to be output. However, the ambient temperature can be used as a state parameter to change, for example, a reference frequency and thus also the frequency of the output signal.Another conceivable output unit is the gradient amplifier. Here too, the gradient controller 21 specifies gradient currents and temporal characteristics by means of a control signal which is converted by gradient amplifiers as output unit. The output signal changes with the resistance of the gradient coil, which changes with the temperature as the state parameter. The time profile with time as a state parameter can also be changed by eddy currents.A further conceivable output unit is a shim amplifier which supplies shim coils with currents that can be adjusted by the control unit 20 for compensating B0 inhomogeneities. A possible state parameter can be the B0field strength, which is present, for example, in the form of a B0map.The predistortion device 70 according to the invention can be provided separately for an output unit, i.e. the predistortion device 70 receives a control signal which changes it depending on one or more state parameters and passes it on to the output unit. This reduces the requirements of this predistortion device and simplifies the training of an artificial intelligence by means of which the predistortion device is realized.However, if there are interactions or dependencies between the output units or the output signals to be generated by them, a common pre-distortion device for these control signals may be advantageous. A control signal can also be considered a state parameter for another control signal. In this case, an artificial intelligence or a neural network is advantageously able to capture and compensate non-analytical relationships by training.FIG. 2 schematically illustrates an exemplary signal flow in a magnetic resonance tomography system 1. Only parts of the magnetic resonance tomography system 1 that participate directly therein are shown.FIG. 2 shows the signal flow for the excitation pulse. The high-frequency unit 22 generates a control signal for driving a high-frequency power amplifier 24 as part of a sequence, on instruction from the controller 23, said control signal describing a high-frequency signal which specifies the frequency, the phase and the amplitude of the excitation pulse as a function of time.It is conceivable that the control signal is digital, for example a continuous sequence of amplitude values. Then, the pre-distortion can advantageously take place in the pre-distortion device 70 by means of digital signal processing. Conversion into an analog signal can then take place, for example, in a D / A converter at the output of the predistortion device 70 or at a signal input of the high-frequency power amplifier.However, it is also conceivable for the control signal to already be an analog low-power signal. It is then conceivable for the predistortion device to have one or more analog phase and / or amplitude actuators 75 which, applied to the control signal, provide the control signal modified by the predistortion device 70.The predistortion device 70 receives state parameters of the magnetic resonance tomography system 1 from one or more sensors such as a temperature sensor 71 for detecting the temperature of the radio-frequency power amplifier 24, a directional coupler 73 for detecting a reflected power and a pick-up coil 72 for detecting a B1 field in the patient tunnel 16.Sensors for vibrations or, as a sensor, an evaluation unit for properties of the received magnetic resonance signals, k-space data or of the images reconstructed therefrom are also conceivable.The control signal is also supplied to the artificial intelligence 74 as an input signal and can be evaluated by the latter as a state parameter for the changes of the control signal itself. This also makes it possible for the artificial intelligence 74 or the neural network self to assume the function of the actuator 75 and to change the control signal for the output unit, as stated below.The artificial intelligence 74 distorts the control signal depending on the state parameters. It is thus conceivable, for example, for the amplification factor of the high-frequency power amplifier 24 to decrease as the temperature of the high-frequency power amplifier increases. The predistortion device 70 or the artificial intelligence 74 must then increase the signal level at its output, which can be effected, for example, by an adjustable amplifier as an actuator 75. In the same way, an increasing reflected power at directional coupler 73 may indicate a phase shift, which may be compensated by artificial intelligence 74 by a phase shifter as actuator 75.However, it is also conceivable in principle for the predistortion device 70 to change the control signal only on the basis of its properties without further state parameters.If the control signal is still digital, then the artificial intelligence 74, in particular in the form of a neural network, can advantageously simultaneously take over the role of the actuator 75. The weighting factors can realize the different amplification and the weighted path of the control signal via one or more neurons can realize a variable phase shift.In the previous example, the control signal was already used to indicate the high-frequency signal. However, it is also conceivable that the output unit 24 in the high-frequency path comprises not only the amplification of the high-frequency signal, but also the provision of the high-frequency signal from parameterized details such as frequency, phase and amplitude as a function of time. For example, a high-frequency pulse can be provided by means of a numerically controlled oscillator (NCO) for providing a carrier frequency, a modulator having a D / A converter for modulating an envelope curve and an adjustable phase shifter. The artificial intelligence 74 or the neural network can then be realized in an advantageous manner, even in the case of digital signal processing of a digital control signal, owing to the lower data rate, with slower, but possibly also more complex, networks and thus carry out the predistortion in a better form taking account of a multiplicity of state parameters.Conversely, it is also conceivable to use special forms of a neural network.Recurrent neural network (RNN) type neural networks have a feedback of signals in one layer or in a preceding layer. These enable the neural network to provide a memory for the time profile of a signal, whether it be a control signal or state parameter. Thus, in an embodiment with this type, it becomes possible, by suitable training, for example in the case of a gradient controller 21 or a high-frequency power amplifier 24, to predict heating on the basis of the previous load and to counteract changes in the output signals.From the family of RNNs, the type of longshort-term memory (LSTM) networks is again used in one embodiment. In these, the temporal parameter is no longer freely trained. Instead, the time parameters or time windows for these values are set to a predetermined value or limited to a predetermined time window on the basis of a priori knowledge about the signals to be processed and their time profiles or characteristics. On the one hand, this limits the flexibility of the neural network with regard to the time-dependent parameters, but allows this to be realized much more easily and with a smaller depth. In this way, the cycle time of the control signals, the effort and the energy consumption can be reduced. It is thus also conceivable to change directly fast signals such as baseband signals or radio frequency signals without actuator 75 by means of the neural network in the predistortion device 70.In the same way, for example, a pre-distortion device 70 in the gradient controller 21 can correct the gradient current as an output signal. State parameters can be, for example, the temperature of a gradient amplifier and / or the temperature of the gradient coil 12.The artificial intelligence 74 has a particularly advantageous effect if many state parameters and control signals interact. An example of this is gradient currents and excitation pulse. The gradient currents or the changes in the gradient fields caused thereby cause eddy currents in the metal parts of the patient tunnel, which in turn generate a magnetic field that temporarily changes the static magnetic field B 0. This can in turn be compensated by a predistortion device 70 by changing the frequency of the excitation pulse over the NCO. For this purpose, the pre-distortion device 70 for the excitation pulse must be supplied with the gradient signals as state parameters. A combined pre-distortion device 70 would also be conceivable, which receives control signals for the excitation pulse and the gradient currents and outputs correspondingly pre-distorted control signals as input signals for the high-frequency power amplifier and the gradient amplifier, wherein the pre-distorted control signal for the excitation pulse is provided as a function of the control signal for the gradient currents.In the same way, other or further interactions between different control and output signals can also be taken into account. For example, it would also be conceivable to correct eddy currents by shim currents in corresponding shim coils or to combine these with the pre-distortion of the excitation pulse.FIG. 3 shows an embodiment of a sequence of a method according to the invention for training a corresponding artificial intelligence.In a step S 10, a predetermined control signal is provided. Such a control signal can be, for example, an analog or digital signal that describes an output signal of the output unit. In the simplest case, the control signal can be linearly dependent on the output signal, that is to say can be a small signal which requires only one power amplifier as output unit, such as a radio-frequency power amplifier for the excitation pulses, or a current amplifier for gradient pulses or shim currents. However, the control signal can also be more abstract, as for a signal synthesizer, in which a fundamental frequency and phase and an envelope curve preset the output signal in their temporal profile. The signal synthesizer may also be combined with a high frequency power amplifier in an output unit.In a further step S 20, the predetermined control signal is output via the predistortion device 70 by means of the output unit. In this case, the predistortion device processes the control signal preferably as a function of one or more state parameters such as temperature, load, geometric influences, magnetic field measurement values, which are supplied to the predistortion device as digital or analog signals. Also, control signals may simultaneously serve as state parameters with respect to other control signals.However, it is also conceivable for the predistortion device 70 to take into account only the control signal itself or its temporal profile.In a further step S 30, the output signal of the output unit is compared with a predetermined output signal. The predetermined output signal is defined by the control signal. The output signal is detected by a sensor, for example a voltage or current sensor, in the case of high-frequency signals also by means of directional couplers or induction loops, for gradients also by means of Hall sensors. In this case, a comparison can be made in a simple case by analog difference formation between an analog control signal and an analog sensor signal of the output signal. More complex comparisons can also be determined digitally by digital signal processing.In a step S 40, parameters of the artificial intelligence are adapted in such a way that a deviation of the output signal output by the output unit from the predetermined output signal, which is predefined by the control signal upstream of the pre-distortion device, is reduced. Reduced can be understood, for example, to mean that a square of the difference in the difference formation is reduced during the comparison, as explained for the preceding step. If a predistortion device for a plurality of output signals is used, the reduction of the deviation can be defined by reducing the sum of the squares of the differences. It is also conceivable that the individual summands are weighted differently than are multiplied by a weighting factor.In addition to the control signal, one or more state parameters are preferably also taken into account in the context of the predistortion, so that the parameters of the artificial intelligence are adapted in S 40 taking into account the at least one state parameter.There are different methods from the prior art for the suitable adaptation of the parameters. For neural networks, the method is referred to as "backpropagation", for example, in which the connection parameters of the individual network nodes or artificial neurons are adapted.Steps S 10 to S 40 are repeated with a plurality of different control signals and / or state parameters by the method jumping to step S 10. The method can also be ended in step S 50 if all provided datasets or combinations of control signals and state parameters have been used in the training. It is also conceivable that the training is ended when a test criterion is fulfilled, for example for predetermined test scenarios based on predetermined control signals and state parameters the deviations between the predetermined output signal and the output signal output by the output unit with the predistorted control signal do not exceed a predetermined threshold value. For the determination of the deviation, reference is made to the comparison described above.Finally, in a step S 60, the trained artificial intelligence is provided for use in a magnetic resonance tomography system according to the invention. This can be done, for example, by the parameter set that was determined in the training. This can be distributed over a data medium or over a network. It is also conceivable that the generated parameter set is distributed as part of a program for implementing the artificial intelligence. It is also conceivable, in particular for high-frequency generation or even gradient control with high real-time requirements, to implement the generated parameters in hardware, for example a programmable logic such as an FPGA or dedicated hardware for a neural network.FIG. 4 shows a schematic flow chart for an embodiment of the method according to the invention for operating the magnetic resonance tomography system according to the invention.In a step S 120, the control unit 20 outputs a control signal for performing image acquisition using the magnetic resonance tomography apparatus 1. The control signal can be generated centrally by the controller 20 or can be called up from a memory. However, it is also conceivable for the control signal to be stored or at least temporarily stored, for example, in the high-frequency unit 22 or the gradient controller 21. The control signal can be analog, for example a voltage or a current which is proportional to the predetermined output signal, or else digital information which indicates the predetermined output signal. The control signal is fed to a predistortion device. This can be implemented separately for a control signal. However, it is also conceivable, in particular if dependencies exist between the control signals or the associated output signals, for a predistortion device to be supplied with a plurality of control signals and to provide a plurality of predistorted output signals accordingly. In this case, other control signals which do not indicate the corresponding output signal can be considered as additional state parameters for an output signal.In a step S 121, the pre-distorter 70 distorts the control signal so that an output signal of the output unit deviates from a predetermined output signal less compared to an output signal of the output unit for an unchanged control signal. The warping takes place by means of the trained artificial intelligence, so that, as a result of the pre-warping of the control signal, the output signal generated from the latter in the output unit deviates less from the predetermined or intended output signal than an output signal generated by the undistorted control signal in the output unit. This is advantageously achieved in that the artificial intelligence can predict and compensate for the influence of the control signal and / or the state parameters and their development over time by the training.The predistortion device 70 preferably takes into account at least one state parameter.In a step S122, the output signal thus generated is output from the image acquisition output unit. This can be the excitation pulse for exciting the nuclear spins via the radio-frequency unit 22, the gradient currents for generating the gradient fields from the gradient controller 21 or other additional signals required for the image acquisition, such as shim currents for compensating B0field deviations by eddy currents or the like.Although the invention has been illustrated and described in more detail by the preferred exemplary embodiment, the invention is not restricted by the disclosed examples and other variations can be derived therefrom by the person skilled in the art without departing from the scope of protection of the invention.

Claims

Magnetic resonance tomography apparatus having a control unit (20) which is designed to output a control signal for carrying out image acquisition using the magnetic resonance tomography apparatus (1) to an output unit, such that the output unit outputs an output signal for the image acquisition, wherein the magnetic resonance tomography apparatus (1) has a pre-distortion device (70) having an artificial intelligence (74), preferably a neural network, which is designed to change the control signal in such a way that the output signal deviates less from a predetermined output signal compared to an output signal of the output unit for an unchanged control signal.Magnetic resonance tomography apparatus according to claim 1, wherein the pre-distortion device (70) is configured to change the output signal depending on at least one state parameter.The magnetic resonance tomography apparatus according to claim 1 or 2, wherein the output unit is one of a group of radio frequency line amplifiers (24), gradient amplifiers, magnetic shim control or signal synthesizers.The magnetic resonance tomography apparatus according to claim 2 or 3, wherein the state parameter is one of a group of temperature, magnetic field strength, power consumption, vibrations, imaging properties, stray fields or operating timeMagnetic resonance tomography apparatus according to one of the preceding claims, wherein the neural network is of the recurrent neural network type, in particular a recurrent neural network of the long short-term memory network type.Magnetic resonance tomography apparatus according to one of the preceding claims, wherein the pre-distortion device is designed to change the control signal as a function of a plurality of state parameters.Magnetic resonance tomography apparatus according to one of the preceding claims, wherein the pre-distortion device (70) has a filter which is designed to generate a weighted average value for a state parameter and the pre-distortion device (70) is designed to change the control signal as a function of the weighted average value.Method for operating a magnetic resonance tomography apparatus (1) according to one of the preceding claims, wherein the method has the steps: (S120) outputting a control signal for carrying out image acquisition using the magnetic resonance tomography apparatus (1) by the control unit (20) of the magnetic resonance tomography apparatus (1), (S121) pre-distortion of the control signal by the pre-distortion device (70) as a function of at least one state parameter, such that an output signal of the output unit deviates from a predetermined output signal less compared to an output signal of the output unit for an unchanged control signal; (S122) outputting the output signal for image acquisition by the output unit.Method for training the artificial intelligence (74) of the magnetic resonance tomography apparatus (1) of one of claims 1 to 7, wherein the method comprises the steps of: (S10), by the control unit (20), providing a predetermined control signal; (S20), by means of the output unit, outputting the predetermined control signal via the predistortion device (70); (S30), comparing the output signal of the output unit with a predetermined output signal; (S40), adjusting the parameters of the artificial intelligence (74) such that a deviation of the output signal from the predetermined output signal is reduced; (S50), repeating the steps (S10) to (S40) with a plurality of different control signals and / or state parameters; (S60), providing the trained artificial intelligence (74).Method according to claim 9, wherein in step (S20) outputting the output signal is output via the pre-distortion device (70) depending on a state parameter, and in step (S40) adjusting the parameters the parameters are adjusted such that the deviation below the state parameter is reduced.The method of claim 9 or 10, wherein steps (S10) through (S40) are repeated with a plurality of different sequences of the different control signals.Method according to one of Claims 9 to 11, wherein the magnetic resonance tomography system (1) has a plurality of output units, wherein in step (S10) a plurality of control signals is provided; in step (S20) the plurality of predetermined control signals is output to the plurality of output units via at least one predistortion device (70); in step (S30) the plurality of output signals of the plurality of output units are compared with a plurality of predetermined output signals and in step (S40) the parameters of the artificial intelligence (74) are adapted in such a way that a deviation of the output signals from the predetermined output signals is reduced.Method according to one of Claims 9 to 12, wherein a phantom is arranged in a recording region of the magnetic resonance tomography system (1) and a deviation of the output signal of the output unit from the predetermined output signal is determined on the basis of an image acquisition of the phantom.Method according to one of Claims 9 to 13, wherein the magnetic resonance tomography system (1) has a comparison device which is designed to determine a deviation of the at least one output signal of the at least one output unit from the predetermined at least one output signal and to provide it to the artificial intelligence (74) as an adaptation parameter.A computer program or computer readable storage medium comprising instructions which, when executed on the control unit (20) of the magnetic resonance tomography apparatus (1) according to any one of claims 1 to 7, cause the control unit (20) to carry out the method according to claim 8.

Images