Method for determining optimal parameters for parallel transmit MRI
The Bayesian optimization of radiofrequency waveform parameters in MRI addresses signal dropout and heating issues in ultra-high field MRI, improving image accuracy and safety by minimizing energy deposition and adhering to regulatory standards.
Patent Information
- Application Number
- PCT/GB2024/051235
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-11
- Filing Date
- 2024-05-10
- Publication Date
- 2024-11-14
AI Technical Summary
Ultra-high field MRI applications are limited by signal dropout areas due to transmit field inhomogeneity and regulatory constraints on radiofrequency heating, which hinder clear and accurate imaging.
A Bayesian optimization process is used to determine optimized parameters for a radiofrequency waveform in MRI, minimizing signal dropout and reducing RF energy deposition, thereby improving image accuracy and safety while complying with regulatory requirements.
The method effectively reduces signal dropout and minimizes RF energy exposure, enhancing the clarity and safety of MRI images while adhering to regulatory limits.
Smart Images

Figure GB2024051235_14112024_PF_FP_ABST
Abstract
Description
[0001] METHOD FOR DETERMINING OPTIMAL PARAMETERS FOR PARALLEL TRANSMIT MRI
[0002] Field
[0003] The present techniques generally relate to determining magnetic resonance imaging, MRI, process parameters. In particular, the present techniques provide a method for identifying parameters for a radiofrequency, RF, waveform in a magnetic resonance imaging, MRI, process using a Bayesian optimisation process.
[0004] Background
[0005] Ultra-high field (i.e. field strengths of 7T and above) MRI provides exceptional signal- to-noise and spatial resolution. However, in practice, applications of ultra-high field MRI are limited by signal dropout areas due to transmit field Bf inhomogeneity, as well as the need to comply with regulatory limits on radiofrequency, RF, heating. This means that in some regions of a human or animal body, there may be signal dropouts that prevent the region from being imaged clearly or accurately.
[0006] The applicant has therefore identified the need for an improved technique for determining magnetic resonance imaging process parameters.
[0007] Summary
[0008] In a first approach of the present techniques, there is provided a computer- implemented method for identifying optimised parameters for a radiofrequency, RF, waveform in a magnetic resonance imaging, MRI, process, the method comprising: receiving at least one calibration dataset comprising magnetic field information for imaging a target object using the MRI process; processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of an RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and optimised the parameters of the RF waveform for imaging the target object.
[0009] Advantageously, the present techniques provide a way to design an RF waveform to image a specific target object using MRI, which is designed in a way that reduces or minimises signal dropout in the target object. This means that the designed RF waveform can produce better MRI images in terms of accuracy and clarity. Additionally or alternatively, the present techniques can reduce the RF energy deposition which allows for shorter repetition times. Hence, exposure to RF energy in, for example, a patient may be reduced making high- resolution imaging techniques safer and ensuring these techniques comply with regulatory and legal requirements. Thus, it is clear that the present techniques may be used to find a suitable balance between a minimum necessary amount of RF energy deposited and minimising signal dropout.
[0010] The at least one calibration dataset may take a variety of forms. Preferably, the calibration dataset includes at least Boand B + maps of the target object. That is, the calibration dataset comprises magnetic field information, such as the ABo offset and per-channel Bi+fields. In general, the ABo and per-channel Bi+field data may be in the form of a 3D volume, from which a slice is interpolated to aid the design of the RF waveform. Additionally or alternatively, the calibration dataset may be or may comprise at least one image obtained via the MRI process, for example, by using one or more specialist MRI imaging sequences. The calibration dataset may be acquired via the MRI process, or be obtained by any other suitable method. For example, the calibration dataset may be estimated using machine learning techniques.
[0011] The RF waveform which is to be optimised / designed by the present techniques may correspond to a spokes pulse comprising a sequence of subpulses interspersed with gradient blips. Spokes pulses are often used for single-slice imaging, also known as slice-by-slice (SBS) imaging, but can also be used for simultaneous multi-slice (SMS) imaging. The gradient blips define a trajectory in excitation kT-space. Each subpulse corresponds to a different point in the (x,y) plane in excitation kT-space and to a slice-select gradient that moves in the z direction in kT-space. That is, each subpulse excites the same slice, but the gradient blips between the subpulses create phase variation in the (x,y) direction across a slice which can create a beneficial interference pattern and help avoid signal dropouts in crucial regions of a target object being imaged. Thus, the step of determining the parameters may comprise using the Bayesian optimisation process to optimise a location, in kT-space, of each subpulse. That is, the present technique for identifying optimised parameters of an RF pulse may involve identifying optimised parameters for subpulses when the RF waveform corresponds to a spokes pulse.
[0012] The RF waveform which is to be optimised may take any suitable form. There are many possible forms that the subpulses could take. For example, the subpulses may have sine [sin(t) / t] shapes and a constant slice-select z-gradient, but it will be understood that this is a non-limiting example.
[0013] Using a Bayesian optimisation process to optimise the subpulses may involve selecting a cost function to be optimised. For example, the root-mean-squared error in excitation flipangle may be used, as well as a nominal prescribed flip angle and a Tikhonov regularisation factor that favours low-energy deposition. However, other things could be used, such as: magnitude of a desired echo in the extended phase graph formalism; predicted image fidelity (such as predictions of the uncertainty (Cramer Rao bounds) for derived parameters like the apparent diffusion coefficient or diffusion tensor, or T1 or T2 or T2* or tissue susceptibility); robustness to subject motion for the pulse; minimised acoustic noise from the pulse; cost functions that minimise excitation of unwanted signals such as fat by considering off-resonance performance at the appropriate frequency; and robustness to flow. It will be understood that these are non-limiting and non-exhaustive examples of cost functions that may be used during the optimisation process.
[0014] Determining the optimised parameters may further comprise an iterative process. The iterative process may comprise: in a first step, picking the location in kT-space of one subpulse (in the sequence of RF subpulses); in a second step, using an RF optimisation algorithm to obtain the RF parameters of the RF waveform given the fixed location (in kT-space) of one subpulse; and repeating the first and second steps for all subpulses. Thus, in each iteration, a different subpulse is used for the first step, and the process continues until convergence.
[0015] Determining the parameters may comprise starting with the last subpulse in the sequence of subpulses. Starting with the last subpulse is advantageous because it is desirable for the subpulses with major contributions to be at the end of the sequence of subpulses. The subpulses at the beginning of the sequence are ‘small corrections’. This is because as soon as a subpulse is applied, the signal from this subpulse will start to exhibit a T2* decay (loss of signal). So, if the subpulses giving more energy appears later, the overall decay is smaller, and the whole spokes pulse (i.e. sequence of subpulses) is more efficient.
[0016] In the second step of the iterative process, using the RF optimisation algorithm may comprise using one of: a magnitude least squares algorithm; and a large tip angle RF optimisation algorithm. It will be understood that these are exemplary, non-limiting optimisation algorithms and other suitable algorithms may be used.
[0017] In the first step of the iterative process, there are many ways the initial location in k - space may be chosen for a subpulse. For example, it may be advantageous to pick a location around a circle in k-space for each subsequent preceding subpulse. This is because minima in the optimization landscape tend to be located around the origin, where the location of the final subpulse is fixed. That is, the location of the last subpulse in the sequence of subpulses is fixed at the centre of k-space, and it is advantageous to pick a location of any of the preceding subpulses to be in a circle that has the centre of k-space as its centre.
[0018] As mentioned above, the at least one calibration dataset may take a variety of forms. Preferably, the calibration dataset includes at least a Bo map of the target object. In some cases, receiving the at least one calibration dataset may comprise receiving at least one calibration dataset corresponding to any of the following images acquired from an MRI process: multiple images; multiple images for each slice; simultaneous multi slice images; and a 3D volumetric RF image. It will be understood these are non-limiting and non-exhaustive techniques that may be used to obtain the calibration dataset (and in particular, the Boand B^ maps of the target object). Thus, the calibration dataset may be obtained using the data and images captured during the calibration MRI of the target object.
[0019] Alternatively, receiving the at least one calibration dataset may comprise receiving at least one generated calibration dataset that is generated using a trained machine learning, ML, model. That is, instead of using data from a full MRI scan of the target object, a calibration dataset may be generated using other data, and then this generated calibration dataset may be used to optimise the parameters of the RF waveform needed to image the target object using MRI.
[0020] When a trained ML model is used, the method may comprise generating the at least one generated calibration dataset by using the trained machine learning model to process a set of localiser images. Localiser images (also known as scout images) are often used to ensure accurate positioning of the target object within an MRI machine. The localiser images may not be of diagnostic quality, but they may provide a useful starting point to be processed by a trained ML model to generate a calibration dataset that can be used for optimising the parameters of the RF pulse used in an MRI process. The output of the processing may be or may comprise magnetic field information, such as a Boand a B±+ map.
[0021] Alternatively, the method may comprise selecting, using the trained machine learning model, from a database of pre-computed pulses for an MRI process, at least one pre-computed pulse most suitable for imaging the target object. That is, the RF pulse and gradient are precomputed. The database of possible pre-computed pulses may be a database of pulses that have been designed using the present techniques, but which need to be further optimised for the specific target object being imaged using MRI.
[0022] Generally speaking, the calibration data may be 3D data, and may be sampled for each of the slices that needs designing. For slice-by-slice imaging, one slice at a time may sampled, while for simultaneous-multi-slice (SMS) imaging, more than one slice may be sampled at a time. Thus, in some cases, receiving at least one calibration dataset may comprise receiving at least two calibration images (or calibration datasets) corresponding to different slices, and determining the parameters may comprise determining the parameters for each of the at least two calibration images (datasets) separately. Then, the separately determined parameters may be combined in some way. The at least two calibration datasets may correspond to the Boand B + map of the target object for different slices (i.e. as obtained using slice-by-slice MRI). The parameters may be determined for each target slice. In this way, the optimisation for each slice happens separately.
[0023] Alternatively, receiving at least one calibration dataset may comprise receiving at least two calibration images (or calibration datasets) corresponding to different slices, and determining the parameters may comprise determining the parameters for each of the at least two calibration images (datasets) simultaneously. The at least two calibration datasets may correspond to the Boand map of the target object for different slices (i.e. as obtained using simultaneous multi-slice MRI). The parameters may be determined for the target slices together. In this way, the optimisation for each slice happens jointly / simultaneously.
[0024] The present techniques may also be extended to non-selective volumetric excitation (“kT-points”), such that the present techniques may be used to pick the gradient blip locations. The kT-points are like spokes pulses except that there are no slice-select z gradients inbetween the subpulses. So, kT-points excite a 3D volume rather than a 2D slice. Thus, determining the parameters may comprise using the Bayesian optimisation process to optimise a location, in kyspace, of gradient blips in a / cT-points pulse, wherein each kT-points pulse comprises a sequence of RF subpulses interspersed with gradient blips.
[0025] In a second approach of the present techniques there is described an apparatus for identifying optimised parameters for a radiofrequency, RF, waveform fora magnetic resonance imaging, MRI, process, the apparatus comprising: at least one processor coupled to memory, and arranged for: receiving at least one calibration dataset comprising magnetic field information for imaging a target object using the RI process; processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of an RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and the optimised parameters of the RF waveform for imaging the target object.
[0026] The features described above with respect to the first approach apply equally to the second approach and are therefore not repeated.
[0027] The apparatus may further comprise a communication module for transmitting the determined parameters to an external apparatus for use in an MRI process to image the target object. Thus, once the parameters have been determined, they may be sent to an external apparatus, such as an MRI machine, for use. The external apparatus may preferably be the same apparatus used during the calibration process.
[0028] Alternatively, the apparatus may itself be the MRI machine which is being used to image the target object or may be used to control the MRI machine. In this case, the determined parameters may be output for use by the apparatus itself (or an operator of the apparatus). Thus, the processor may be further arranged for generating instructions for an MRI machine to emit an RF waveform using the determined parameters of the optimised RF waveform for the target object.
[0029] In a third approach of the present techniques there is provided a system for identifying optimised parameters for a radiofrequency, RF, waveform of a magnetic resonance imaging, MRI, process, the system comprising: an MRI machine; and an apparatus comprising: at least one processor coupled to memory, and arranged for: receiving at least one calibration dataset comprising magnetic field information for imaging a target object using the MRI process; processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of an RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and the optimised parameters of the RF waveform for imaging the target object.
[0030] The features described above with respect to the first approach apply equally to the third approach and are therefore not repeated.
[0031] In the system, the apparatus may not be able to control the MRI machine itself. Thus, the apparatus may further comprise a communication module for transmitting the determined parameters to the MRI machine for use in an MRI process to image the target object. Alternatively, the processor of the apparatus may be used to control the MRI machine itself. Thus, the processor of the apparatus may further be arranged for generating instructions for the MRI machine to emit an RF waveform using the determined parameters of the optimised RF waveform for the target object; and the apparatus may further comprise a communication module for transmitting the generated instructions to the MRI machine for use in an MRI process to image the target object.
[0032] In a further approach to the present techniques there is described a computer-readable storage medium comprising instructions which, when executed by a processor, causes the processor to carry out any of the methods, processes and techniques described herein.
[0033] As will be appreciated by one skilled in the art, the present techniques may be embodied as a system, method or computer program product. Accordingly, present techniques may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.
[0034] Furthermore, the present techniques may take the form of a computer program product embodied in a computer readable medium having computer readable program code embodied thereon. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. A computer readable medium may be, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing.
[0035] Computer program code for carrying out operations of the present techniques may be written in any combination of one or more programming languages, including object oriented programming languages and conventional procedural programming languages. Code components may be embodied as procedures, methods or the like, and may comprise subcomponents which may take the form of instructions or sequences of instructions at any of the levels of abstraction, from the direct machine instructions of a native instruction set to high- level compiled or interpreted language constructs. Embodiments of the present techniques also provide a non-transitory data carrier carrying code which, when implemented on a processor, causes the processor to carry out any of the methods described herein.
[0036] The techniques further provide processor control code to implement the abovedescribed methods, for example on a general purpose computer system or on a digital signal processor (DSP). The techniques also provide a carrier carrying processor control code to, when running, implement any of the above methods, in particular on a non-transitory data carrier. The code may be provided on a carrier such as a disk, a microprocessor, CD- or DVD- ROM, programmed memory such as non-volatile memory (e.g. Flash) or read-only memory (firmware), or on a data carrier such as an optical or electrical signal carrier. Code (and / or data) to implement embodiments of the techniques described herein may comprise source, object or executable code in a conventional programming language (interpreted or compiled) such as C, or assembly code, code for setting up or controlling an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array), or code for a hardware description language such as Verilog (RTM) or VHDL (Very high speed integrated circuit Hardware Description Language). As the skilled person will appreciate, such code and / or data may be distributed between a plurality of coupled components in communication with one another. The techniques may comprise a controller which includes a microprocessor, working memory and program memory coupled to one or more of the components of the system.
[0037] It will also be clear to one of skill in the art that all or part of a logical method according to embodiments of the present techniques may suitably be embodied in a logic apparatus comprising logic elements to perform the steps of the above-described methods, and that such logic elements may comprise components such as logic gates in, for example a programmable logic array or application-specific integrated circuit. Such a logic arrangement may further be embodied in enabling elements for temporarily or permanently establishing logic structures in such an array or circuit using, for example, a virtual hardware descriptor language, which may be stored and transmitted using fixed or transmittable carrier media.
[0038] In an embodiment, the present techniques may be implemented using multiple processors or control circuits. The present techniques may be adapted to run on, or integrated into, the operating system of an apparatus.
[0039] In an embodiment, the present techniques may be realised in the form of a data carrier having functional data thereon, said functional data comprising functional computer data structures to, when loaded into a computer system or network and operated upon thereby, enable said computer system to perform all the steps of the above-described method.
[0040] Brief description of the drawings Implementations of the present techniques will now be described, by way of example only, with reference to the accompanying drawings, in which:
[0041] Figure 1a shows a radiofrequency, RF, and gradient form of a 2-spoke sine pulse;
[0042] Figure 1 b shows a kyspace trajectory of the 2-spoke pulse;
[0043] Figure 2 is a flowchart of example steps an iterative Fourier transform approach for identifying parameters for a radiofrequency, RF, waveform in a magnetic resonance imaging, MRI, process, as known in the prior art;
[0044] Figure 3 is a flowchart of example steps for using the present techniques for identifying parameters for a radiofrequency, RF, waveform in a magnetic resonance imaging, MRI, process;
[0045] Figure 4 shows an example of calibration data that may be contained in the calibration dataset;
[0046] Figure 5 is a flowchart of further steps to identify the optimised RF waveform parameters;
[0047] Figure 6 shows the cost function landscape plotted as a function of spokes position in fcT-space;
[0048] Figure 7 shows three iterations of the Bayesian optimization Gaussian process mean estimation of the spoke position cost function surface as iterations progress;
[0049] Figure 8 shows a single slice (SBS) 2- and 3-spoke flip angle map comparing FTA and BOGAT in a phantom test object;
[0050] Figure 9 shows improvements in uniformity and energy deposition that may be achieved by using the BOGAT over an FTA approach measured in vivo single slice (SBS);
[0051] Figures 10 shows flip angle maps and pulse energies for multiband factor 3 pulse designs in a phantom test object;
[0052] Figures 11a to 11c show retrospective in vivo results of the present techniques;
[0053] Figure 12 is a diagram of an apparatus for identifying parameters for a radiofrequency, RF waveform for a magnetic resonance imaging, MRI, process;
[0054] Figure 13 is a diagram of an apparatus 1200 for identifying optimised parameters for a radiofrequency, RF, waveform for a magnetic resonance imaging, MRI, process;
[0055] Figure 14 shows the FA RMSE, pulse energy, and cost function landscapes plotted as a function of spoke position in kT-space for a two-spoke 30° pulse;
[0056] Figures 15a to 15c show how the BOGAT algorithm converges for two-spoke pulses
[0057] Figure 16 shows prospective in vivo validation of the present techniques;
[0058] Figures 17a to 17c show how BOGAT improves the image homogeneity in frontal lobe and temporal lobe, while also reducing the energy deposition; and
[0059] Figure 18 shows retrospective kT-point pulse design for 1 subject with a target flip angle of 30°. Detailed description of the drawings
[0060] Broadly speaking, embodiments of the present techniques provide a method for identifying optimised parameters for a radiofrequency, RF, waveform in a magnetic resonance imaging, MRI, process using a Bayesian optimisation process. Advantageously, the present techniques enable quick and accurate determination of the RF parameters by optimising k - space locations of a sequence of subpulses in a spokes pulse using the Bayesian optimisation process, which enables a fast exploration of kT-space.
[0061] Parallel transmit (pTx) can in principle solve the above-mentioned challenges in ultra- high field MRI (signal dropout areas due to transmit field Br inhomogeneity), but the design of parallel transmit pulses for slice-selective sequences, required by many neuroimaging protocols, is still not fully solved. A popular approach for achieving slice-selective pTx excitation is the “spokes pulse”. Spokes pulses are comprised of several slice-selective subpulses separated by in-plane gradient blips, as is, for example, shown in Figure 1a and 1b. Spokes pulses are efficient to design, as the RF optimization variables consist of only one complex weight per channel per subpulse. Several algorithms exist to optimise the subpulse coefficients to improve excitation flip-angle homogeneity and minimise specific absorption rate, SAR, including the popular Tikhonov-regularised Magnitude Least Squares, MLS, algorithm (“Magnitude least squares optimization for parallel radio frequency excitation design demonstrated at 7 Tesla with eight channels”, Setsompop et al., Magnetic Resonance in Medicine, 2008. 59(4): p. 908-915). The specific absorption rate describes the amount of heating on the target object due to the RF waveform, and is therefore an important indicator of the safety of a given MRI process when this is, for example, used to obtain an image of a subject’s brain.
[0062] At present, spokes pulses are normally optimised online per-subject. This optimisation strategy requires that the optimisation completes while the subject lies still in the scanner, which is especially challenging when designing pulses for whole-brain slice-by-slice (SBS) imaging. Therefore, it is common to select gradient blips using fixed values or through heuristics such as the iterative Fourier transform approach (FTA), shown in Figure 2. Finding a robust and fast optimisation technique for this problem, is challenging. For example, grid searches are not practical for online optimisation due to amount of time and resources necessary to run a grid search algorithm. Local optimisation approaches such as greedy joint optimization, as described in: “Joint design of trajectory and RF pulses for parallel excitation” by Yip et al. published in Magnetic Resonance in Medicine, 2007. 58(3): p. 598-604, are susceptible to getting stuck in local minima with suboptimal results in terms of Bi+homogeneity or energy deposition.
[0063] Furthermore, many neuroimaging studies use simultaneous multi-slice (SMS) acceleration. This means that while the RF weights for each slice can be optimized individually and modulated according to the slice positions, all slices must share a single common set of gradient blips. RF coefficient optimization for simultaneous multi-slice (SMS) excitations has been studied, but so far, no work has been dedicated to investigating the spoke positioning choices for SMS.
[0064] The present techniques provide a way to investigate the optimization landscape for spokes pulse design in, for example, the human head and assess the implications for heuristic, local and global optimisation approaches. Thus, the present techniques introduce a new global optimization method called “Bayesian Optimisation of GrAdient Trajectory” (BOGAT) for spokes pulse design. It will be clear that while the present techniques are described with using the human head as an example, they may be applied to any other target object, or any other area of the body of a target human / animal subject.
[0065] Figure 3 is a flowchart of example steps for using the present techniques for identifying optimised parameters for a radiofrequency, RF, waveform in a magnetic resonance imaging, MRI, process. The method may be performed during a calibration stage of an MRI process. The method comprises, receiving at least one calibration dataset that comprises initial parameters of an RF waveform for imaging a target object using the MRI process (step S100). The target object may be a patient’s head, such that the patient’s brain may be imaged using the present techniques. The target object may be any other part of the body or the whole body. The patient may be a human or animal.
[0066] Turning briefly to Figure 4, this shows an example of calibration data that may be contained in the calibration dataset. Figure 4 shows examples of Boand B2+ maps for a slice in brain with a pTx capable head coil. The set of images on the left side of the Figure show an example of the magnitude of Bx+ field in 3 out of 8 channel, while the set of images in the middle of the Figure show the phase of the B + field of these channels. The image on the right side of the Figure shows the Bofrequency offset in this slice. Thus, the at least one calibration dataset may take a variety of forms. Preferably, the calibration dataset includes at least a Boand a B^ map of the target object. That is, the calibration dataset may be or may comprise magnetic field information, such as the AB0and per-channel B / fields. In general, the AB0and per-channel B field data may be in the form of a 3D volume, from which a slice is interpolated to aid the design of the RF waveform. Additionally or alternatively, the calibration dataset may be or may comprise at least one image obtained via the MRI process, for example, by using one or more specialist MRI imaging sequences. The calibration dataset may be acquired via the MRI process, or be obtained by any other suitable method. For example, the calibration dataset may be estimated using machine learning techniques.
[0067] Returning to Figure 3, in some cases receiving the at least one calibration dataset (step S100) may comprise receiving at least one calibration dataset corresponding to any of the following images acquired from an MRI process: multiple images; multiple images for each slice; simultaneous multi slice images; and a 3D volumetric RF image. It will be understood these are non-limiting and non-exhaustive techniques that may be used to obtain the calibration dataset (and in particular, the a Boand map of the target object). Thus, the calibration dataset may be obtained using the data and images captured during the calibration M I of the target object.
[0068] Alternatively, the step (S100) of receiving the at least one calibration dataset may comprise receiving at least one generated calibration dataset that is generated using a trained machine learning, ML, model. That is, instead of using data from a full MRI scan of the target object, a calibration dataset may be generated using other data, and then this generated calibration dataset may be used to optimise the parameters of the RF waveform needed to image the target object using MRI.
[0069] When a trained ML model is used, preferably, the method may further comprise generating the at least one generated calibration dataset by using the trained machine learning model to process a set of localiser images. This may involve using the method described in, for example: “Rigid motion-resolved B1+ prediction using deep learning for real-time paralleltransmission pulse design” by Plumley et al. published in Magn Reson Med. 2022 May;87(5):2254-2270. doi: 10.1002 / mrm.29132. Epub 2021 Dec 27. PMID: 34958134; PMCID: PMC7613077. Localiser images (also known as scout images) may be used to ensure accurate positioning of the target object within an MRI machine. The localiser images may not be of diagnostic quality, but they may provide a useful starting point to estimate BO and perchannel B1+ calibration data that can then be used for optimising the parameters of the RF pulse used in an MRI process as described above.
[0070] Alternatively, the method may further comprise: selecting, using the trained machine learning model, from a database of pre-computed pulses for an MRI process, at least one pulse most suitable for imaging the target object. The database of possible pre-computed pulses may be a database of pre-computed pulses that have been designed using the present techniques, but which need to be further optimised for the specific target object being imaged using MRI.
[0071] Generally speaking, the calibration data may be 3D data, and may be sampled for each of the slices that needs designing. For slice-by-slice imaging, one slice at a time may sampled, while for simultaneous-multi-slice (SMS) imaging, more than one slice may be sampled at a time. Thus, in some cases, step S100 may comprise receiving at least two calibration images (or calibration datasets) corresponding to different slices, and determining the parameters (step S104) may comprise determining the parameters for each of the at least two calibration images (datasets) separately. That is, one dataset may be received for each slice. The at least two calibration datasets may correspond to the Boand map of the target object for different slices (i.e. as obtained using slice-by-slice MRI). The parameters may be determined for a target slice. Then, the separately determined parameters may be combined in some way.
[0072] Alternatively, step S100 may comprise receiving at least two calibration images (or calibration datasets) corresponding to different slices, and determining the parameters (step S104) may comprise determining the parameters for each of the at least two calibration images (datasets) simultaneously. That is, one dataset may be received for each slice. The at least two calibration datasets may correspond to the Boand B + map of the target object for different slices (i.e. as obtained using simultaneous multi-slice MRI). The parameters may be determined for a target slice. Then, the simultaneously determined parameters may be combined in some way.
[0073] The method comprises processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of the RF waveform to be used during the MRI process (steps S102 and S104). Using a Bayesian optimisation process to optimise the subpulses may involve selecting a cost function to be optimised. For example, the root-mean-squared error in excitation flipangle may be used, as well as a nominal prescribed flip angle and a Tikhonov regularisation factor that favours low-energy deposition. However, other things could be used, such as: magnitude of a desired echo in the extended phase graph formalism; predicted image fidelity (such as predictions of the uncertainty (Cramer Rao bounds) for derived parameters like the apparent diffusion coefficient or diffusion tensor, or T1 or T2 or T2* or tissue susceptibility); robustness to subject motion for the pulse; minimised acoustic noise from the pulse; cost functions that minimise excitation of unwanted signals such as fat by considering off-resonance performance at the appropriate frequency; and robustness to flow. It will be understood that these are non-limiting and non-exhaustive examples of cost functions that may be used during the optimisation process.
[0074] The method may comprise outputting the determined parameters of the optimised gradient trajectory and optimised radiofrequency waveform for the target object (step S108).
[0075] Figure 5 is a flowchart of an iterative process that may be used to determine the optimised parameters. The iterative process may comprise: in a first step, picking the location in kyspace of one subpulse (in the sequence of RF subpulses); in a second step, using an RF optimisation algorithm to obtain the RF parameters of the RF waveform given the fixed location (in kyspace) of one subpulse; and repeating the first and second steps for all subpulses. Thus, in each iteration, a different subpulse is used for the first step, and the process continues until convergence.
[0076] In the second step of the iterative process, using the RF optimisation algorithm may comprise using one of: a magnitude least squares algorithm; and a large tip angle RF optimisation algorithm. It will be understood that these are exemplary, non-limiting optimisation algorithms and other suitable algorithms may be used.
[0077] In the first step of the iterative process, there are many ways the initial location in ky space may be chosen for a subpulse. For example, it may be advantageous to pick a location around a circle in k-space for each subsequent preceding subpulse. This is because minima in the optimization landscape tend to be located around the origin, where the location of the final subpulse is fixed. That is, the location of the last subpulse in the sequence of subpulses is fixed at the centre of k-space, and it is advantageous to pick a location of any of the preceding subpulses to be in a circle that has the centre of k-space as its centre.
[0078] Determining the parameters may comprise starting with the last subpulse in the sequence of subpulses. That is, the location of the last subpulse in the sequence of subpulses may be determined. In this case, location refers to a location in / cT-space, which is an inverse Fourier representation of physical space defined in: “A k-space analysis of small-tip-angle excitation” by Pauly et al. in J Magn Reson 1989;213:544-57.
[0079] . The location of the last subpulse is often simply set at the origin kyspace, which is advantageous because it tends to make the amplitude of this final pulse be largest, reducing the effective echo time (TE) of the scans which reduces unwanted T2* signal loss and minimises unwanted phase accrual that could otherwise lead to signal loss. Setting the location of the last subpulse provides initial conditions for solving the present optimisation problem, i.e. the location of the last spokes subpulse is used as the starting value for an optimizer. Using the initial conditions, and a Bayesian optimisation approach, the kyspace location of each subpulse in the sequence is optimised. This optimisation is done “backwards”, i.e. starting with the second-to-last subpulse, since this tends to make the amplitude of the later subpulses be larger, reducing undesired signal loss due to T2* relaxation or due to unwanted phase accrual. .
[0080] As shown in Figure 5, at the end of the optimisation process, the kyspace location of each subpulse is fixed. Then, an RF optimisation algorithm is used to obtain the parameters of the RF waveform corresponding to the fixed location subpulses. That is, two optimisation processes are run - the first one finds an optimised location of the subpulses in kyspace. However, what is needed in order to complete the MRI process, are the RF coefficients, which is why a second optimisation process is run to obtain RF coefficients that best correspond to the determined k-space location of the subpulses. Any suitable optimisation technique may be used, such as, for example, the MLS algorithm by Setsompop et al. which is described in more detail below. Alternatively, a large tip angle RF optimisation algorithm may be used, as, for example, described in: "Ultra-high field MRI: parallel-transmit arrays and RF pulse design" by Williams et al. published in Physics in Medicine & Biology (2022).
[0081] Transmit kyspace Parallel transmit (pTx) radio frequency (RF) pulse design is generally based on the small-tip-angle (STA) approximation of the Bloch equation: dt, where Mxyis the transverse magnetization, y is the gyromagnetic ratio, Mois the initial magnetization, T is the total length of the pulse, x is the spatial position, G is the gradient vector and is the RF transmit field due to Pauly ‘Pauly J, Nishimura D, Macovski A. A k-space analysis of small-tip-angle excitation. J Magn Reson 1989;213:544-57’.
[0082] If a time-varying gradient field is applied during the excitation process, we can understand the integral term in the exponent by introducing the spatial wavenumber fT
[0083] Kt) = -Y G(0 K-
[0084] Jt which can be thought of as a point in a “transmit kT-space” . Note that the integration limits mean that at the time T defined as the end of the pulse, then k(T) = 0, by definition, i.e. all pulses end at the origin in kT-space.
[0085] Spokes trajectory
[0086] The spokes form (also known as the “fast kz” form in early literature) is one of the simplest types of kT-space trajectory. It combines several slice-selective pulses in succession, with small in-plane gradient blips in between. Despite the simplicity, however, there is currently not a set of standard or optimal configurations for spokes pTx pulse design.
[0087] MLS Optimisation
[0088] The core algorithm for spokes pulse optimisation is shown in Figures 3 and 5. To understand why this approach works, let us consider the points r across an imaging volume. The magnetisation profile m(r) immediately following a spokes pTx pulse can be written using the small tip angle (STA) approximation, as shown in: “Magnitude least squares optimization for parallel radio frequency excitation design demonstrated at 7 Tesla with eight channels” by Setsompop et al. published in Magnetic Resonance in Medicine, 2008. 59(4): p. 908-915, as where i = -l, y is the gyromagnetic ratio, m0is the equilibrium magnetisation, Sris the transmit scaling (sometimes called “sensitivity” due to the principle of reciprocity) of the rthcoil at position x, Ncis the number of transmit coils, B1(t,r) is the Bi+field of the rthcoil at time t, k is the location in transmit k-space, AB0(x) is the static field offset at position x, and T is the total length of the RF pulse.
[0089] For a spokes pulse, it is possible to separate the waveform of each subpulse from its complex coefficients c(n,r). Furthermore, for the ease of computation, the time steps are discretised, giving where B1;Subdenotes the unit Bi+field of the subpulses.
[0090] This equation can be re-written in the matrix form following Equation 5 in ‘Grissom W, Yip CY, Zhang ZH, Stenger VA, Fessler JA, Noll DC. Spatial domain method for the design of RF pulses in multicoil parallel excitation. Magnetic Resonance in Medicine 2006;56:620-29.’ to give m - Ab where m is a complex vector whose elements mpare the complex transverse magnetisation immediately following the pulse at each of the Npdiscretised positions xp, A is an Npx Nsx Nc) flattened system matrix defined as where the indexes r and n are flattened as in a Kroenecker product into the second dimension. The pulse design vector b has dimensions (Nsx Nc) x 1 and elements
[0091] If the spokes positions K are chosen, the optimal RF coefficients b can be determined by, for example, regularised Magnitude Least Squares b = argmin || |A(K)6| - mtarget112+ A| |b| | ]. where X is a Tikhonov regularisation parameter that penalises high energy pulse designs.
[0092] Hence, the A matrix depends directly on the spoke positions K and the optimum pulse coefficients b depend on K implicitly.
[0093] Because the MLS problem is a nonconvex optimization problem, to ensure the consistency of the b values in the following step, the RF optimizations should keep the same initialization conditions. For example, one can set initial values to a CP mode excitation, and in the case of a variable exchange MLS optimizer, set the initial target phases to 0° for the DC spoke. It is possible to optimise for the spoke positions K in an outer loop, updating the RF waveform design at each step:
[0094] It = argmin
[0095] Similarly, in simultaneous multislices, SMS (or multiband, MB) applications, the cost function can be a sum of the cost of all slices. Therefore, the algorithm optimises for
[0096] K = argmin and b = argmin where the weight vv, can be set to adjust the contribution of each slice to the total cost. For example, they can be determined by the average cost of one or more test samples, w(= to avoid the total cost being dominated by some difficult slices. Different Tikhonov regularization factors ALare allowed for each slice.
[0097] Note that the MLS RF optimizer used to illustrate this derivation is merely one possible RF optimization method. In general, there are many existing RF optimizers using various forms of cost functions, and we can write b = argmin [y(b; K)] where y is the generic RF optimizer cost function. Then the optimization with respect to the spoke positions K becomes
[0098] K = argmin[yg(K)] where yg(K) = y(b K), the RF cost function value at the optimum b. Regardless of the exact form of y, the formulation of our additional outer optimization problem for the spoke positions K is the same.
[0099] Bayesian approach
[0100] Bayesian optimization, as, for example, described in: “Efficient Global Optimization of Expensive Black-Box Functions” by Jones et al. published in Journal of Global Optimization, 1998. 13(4): p. 455-492, is a global optimization approach for object functions that are difficult to evaluate. It is applied to fields including robotics, reinforcement learning and drug discovery. Bayesian Optimization approximates the target function (in this case the cost function of spoke positions K) with a Gaussian process prior. The method uses a Gaussian process model for the target function which describes the potential values of the function f(x) at any given point x through the posterior distribution. At each iteration, the function is evaluated at a point, and the distribution is updated with the information of the true value. Then, the algorithm automatically decides the next point to evaluate through an acquisition function, which usually measures the expected improvement of the target function value. By picking the point with highest expected improvement (or a similar metric), the method can find the global optimum of the target function in minimum number of steps (and hence time).
[0101] As new observations are acquired (K points sampled), the mean function can be easily updated. The cost function ysfor K is non-convex and slow to evaluate, so Bayesian Optimization is a suitable approach to find the global minimum.
[0102] Without explicit knowledge about yg, Bayesian optimization uses a Gaussian process (GP) to model the cost function of spoke positions. A Gaussian process is a statistical model that describes a probability distribution over possible functions that fit a set of observed points, where any finite samples of the model are jointly Gaussian distributed. It is characterized by a mean function and a kernel that describes the expected correlation between samples. As new observations are acquired (i.e., as gradient k-space points are sampled and the corresponding cost function values are evaluated), the GP model is updated by an efficient analytical step. This provides an evolving best estimate (mean function) of the cost function and its uncertainty (uncertainty function) at each step in the optimization. The GP acts as a surrogate to the underlying cost function, which can be evaluated rapidly without time-consuming RF solver steps.
[0103] Bayesian optimization is robust to noise in the computed cost function samples, which arises due to small inconsistencies in the result of the RF optimization step due to the RF optimizer stopping criteria, because these get absorbed into the uncertainty function of the Gaussian process.
[0104] Instead of sampling K randomly, a key feature of Bayesian optimization is that it uses an “acquisition function” to help automatically determine the most efficient point where the cost function should be evaluated next:
[0105] Knext = argmax «(K)
[0106] Bayesian Optimisation uses an acquisition function to express accumulating latent knowledge of the optimum gradient blips. Thus, Bayesian optimisation automatically and efficiently balances exploration and exploitation. The acquisition function is evaluated every time a new observation is made. This function determines which point should be next evaluated. The acquisition function is designed to have large values either where an area in the kT-space is unvisited (promotes “exploration”) or around the known potential minima (promotes “exploitation”). This behavior helps Bayesian optimization achieve rapid convergence. Although Bayesian optimization is not mathematically guaranteed to find the true global minimum, in many cases it often does. A common choice is the expected improvement (El) function, which is chosen to maximize the expected improvement in the target function value over the current predicted best. The acquisition function balances exploration in unknown areas in the function domain, with the exploitation around known potential minima, and thus helps Bayesian Optimization achieve rapid convergence to the global minimum. In this implementation, a variation of El, termed “expected-improvement-plus” and described in: “Convergence Rates of Efficient Global Optimization Algorithms” by Bull published in Journal of Machine Learning Research, 2011. 12: p. 2879-2904, is used that avoids overexploitation behaviour.
[0107] To further accelerate convergence, Bayesian optimization can benefit from initial seeds, which are the K sample points that are always evaluated at the start, if some prior information about the typical cost function surface is known.
[0108] BOGAT
[0109] Figure 5 shows the BOGAT approach, which uses an outer loop to perform Bayesian Optimization to determine the optimal spokes subpulse positions. The last spokes subpulse is fixed at the centre of kr-space. The preceding spokes subpulse is sampled in the kx-kyplane to perform Bayesian Optimization for a given number of iterations. The RF coefficients are optimized at each position to calculate the MLS cost. The position with the lowest cost is determined to be the best subpulse position.
[0110] If more than two spoke subpulses are used, each preceding subpulse is optimized iteratively, and the MLS target phase is conserved between each addition of subpulses to ensure solution stability. This is a greedy approach with each iteration only considering the best subpulse location for that step. Alternatively, all spoke positions could be simultaneously optimized to find a globally optimal solution, but the efficiency of Bayesian optimization, as with most sampling-based methods, generally decreases with the dimensionality of the search space. In this case, performance may be maintained by using an additive GP kernel between different spokes instead of a multiplicative kernel. For example, Mattern kernels may be added instead of multiplied.
[0111] Methods
[0112] Pulse design: All pulse designs were performed with a modified version of the PTx Pulse Design toolbox (Siemens, Erlangen, Germany) in Matlab (version 2022b, MathWorks, MA, USA). All RF optimizations were performed with a Tikhonov regularized MLS algorithm using the least-squares QR-decomposition (LSQR) solver. Simulation studies-. All simulation studies and prediction results were based on a database of in vivo adjustment data sets acquired using an 8-transmit (Tx) / 32- receive (Rx) head coil (Nova Medical, USA) on a 7T Terra MRI system (Siemens). These data were collected in previous scans under the Cambridge Human Biology Research Ethics Committee ethics approval (HBREC.2020.27).
[0113] The AB0and maps were acquired with dual-echo gradient-echo and magnetization- prepared turbo-FLASH (TFL) sequences, respectively, and they were processed, masked, and exported as MATLAB files in the vendor's PTx Pulse Design framework. Flip-angle (FA) root mean squared errors (RMSEs) were computed by numerical solution of the Bloch equations (“Bloch simulation”).
[0114] Grid search'. A grid search (Figure 6) was performed varying the kT-space position of the first subpulse in a 2-spoke pulse (i.e. a spoke pulse with two subpulses). The kx-kyplane was sampled at 51x51 locations between ±0.15 cm-1 (3 / FOV). The final subpulse was fixed at kx=ky=O. RF optimizations were performed using regularised MLS with AB0and per-channel B maps from a healthy volunteer. Transverse slices of the whole brain were divided into 8 slabs. For each slab, two target flip angles 30° and 45°, and two fixed Tikhonov regularization factors (0.03 and 0.10) were calculated. The max number of iterations was large enough to allow convergence at all points. The resulting plots reveal the typical features in a spokes optimisation landscape.
[0115] A further grid search (Figure 14) was performed in a larger database in which we varied the kT-space position of the first spoke in a two-spoke pulse. The kx-kyplane was sampled at 51 * 51 locations between ±0. 15 cm"1(±3 / FOV). The final spoke was fixed at kx= ky= 0.00 cm-1. RF optimizations were performed using regularized MLS with ABo and per-channel B? maps as described previously. T ransverse (A / = 5) and transverse oblique (A / = 9) slices of the whole brain were divided into 16-mm slabs, totaling 85 slabs. Each pulse was optimized over all slices in that slab. Two target FAs (30° and 45°) and two fixed Tikhonov regularization factors (0.03 and 0.10) were used. The maximum number of iterations for the RF solver was set to 200, which was large enough to allow convergence at all points. The resulting plots reveal the typical features in a two-spoke optimization landscape and are discussed in further detail subsequently.
[0116] Fourier Transform Approach (FTA)-. The default approach to determine subpulse positions uses what is called the Fourier transform approach (FTA). As shown in Figure 2, this iteratively searches for subpulse positions by first fixing the final subpulse at k=0, optimizing its RF waveform coefficients, computing the resulting predicted transverse magnetization using the subject’s ABo and per-channel B? maps, subtracting this from the target magnetization to give a residual, Fourier transforming the residual to / cT-space and placing the preceding subpulse at the fcT-space point with largest absolute residual. Bayesian optimisation (BOGAT)-. Bayesian Optimization of subpulse positions, as shown in Figure 5, was implemented using the ‘bayesopt’ function in Matlab. The kx-ky search space ranged between ±0.15 cm-1(3 / FOV) in each dimension. Initial seeds consisted of two rings, one with 4 points at radius 0.05 cm-1, the other with 5 points at radius 0.09 cm-1. This is based on a sum cost map of 50 slabs. All designs used a maximum of 30 Bayesian Optimization iterations unless otherwise specified.
[0117] Comparison with grid search’. To evaluate the performance of BOGAT to estimate the cost function, one 30° bipolar two-spoke pulse was designed for each slab used in the grid search described previously (total of 85 slabs, and transverse and transverse oblique slices) with the BOGAT method. RF coefficients were optimized with regularized MLS with the Tikhonov factor fixed at 0.03. Sine subpulses with time-bandwidth product of 4.0 and duration of 3.0 ms were used. The optimal kT-space location for the first (movable) spoke determined by the BOGAT method in 30 iterations was compared with the grid search results.
[0118] Additionally, we compared the BOGAT optimization performance against local optimizations with one or more random starting points based on our grid search data set. This was computed from the grid search data by assuming that each random “starting point” falling within the catchment basin (Figure 3C, E) of a local minimum will converge to the minimum value for that catchment basin. This mimicked an idealized convergence for a gradient-descent local optimizer.
[0119] The efficiency of sequential and simultaneous sampling for three-spoke designs was compared. For the sequential sampling case, one further spoke was added to this BOGAT- optimized two-spoke pulse, where all other design parameters were kept constant. For the simultaneous sampling case, the Bayesian optimization algorithm was run with four variables (kxand kyfor both first and second spokes) for 100 iterations. The RF optimizer was run as described previously.
[0120] To illustrate the advantage of BOGAT in terms of the trade-off between excitation quality and pulse energy, two-spoke pulses were designed with whole-brain transverse field maps from the same field map database described previously. We tested two-spoke 45° excitation for single-band, multiband (MB) 2, MB3, and MB4 — each with a total of 36 sets of slice groups in 9 subjects, with 8 Tikhonov factors (0.001, 0.01, 0.03, 0.05, 0.08, 0.1 , 0.2, and 0.5). We compared the vendor's default FTA method with spoke positions determined only by the central slice in each slab. The test samples used to determine the slice weightings in BOGAT were chosen to be the same as the BOGAT seed samples. The slice weightings are intended to force the optimizer to balance its effort between the slices, instead of, for example, spending almost all effort on the harder slice leading, which might lead to an unexpectedly poor solution for the easier slice. Phantom validation-. Gradient blips add a position-dependent phase shift to the magnetisation, meaning that the net effect of each gradient blip + subpulses combination has more possible variations. This means that the overall spokes pulse has more chance to achieve the desired target magnetisation pattern. When spokes pulses were originally conceived, it was envisaged that several subpulses would be used, but in practice limitations due to T2 and T2* relaxation mean that most sequences use two or three subpulses.
[0121] BOGAT was applied in test objects (also referred to as “phantoms”) and in vivo to design single slice and SMS spokes pulses, and non-selective / cr-points pulses. AB0and perchannel B maps were acquired with dual-echo GRE and saturation prepared turboFLASH using an 8Tx32Rx head coil (Nova Medical) on a 7T MRI (Terra, Siemens, Erlangen, Germany). A 2-spoke and a 3-spoke 90° monopolar pulse were designed for a single slice with BOGAT and regularised MLS with a fixed 0.10 Tikhonov factor. For comparison, designs were repeated with FTA spoke positioning, with identical RF MLS optimization parameters.
[0122] The flip angle maps of these pulses were measured with a saturation prepared TFL sequence with readout FA=5°, TR=5000ms, TE=1.85ms, 1.6mm*1.6mm in plane resolution and the linear correction recommended in: “Fully Integrated Scanner Implementation of Direct Signal Control for 2D T2-Weighted TSE at Ultra-High Field” by Tomi-Tricot et al. published in Proc. ISMRM2021, 2021.
[0123] In vivo validation-. One volunteer gave written consent and was scanned in the same system as described above for the Phantom validations. The AB0and per-channel Br maps for one transverse slice across the brain were acquired as above. 2-spoke 90° excitation pulses were designed with both BOGAT and FTA each with four Tikhonov factors (0.01 , 0.03, 0.05 and 0.1) and the resulting flip angle maps were measured with the TFL sequence described above.
[0124] In one subject, we additionally implemented BOGAT with a Jacobian matrix large tip angle (LT A) optimizer, as described in “Joint design of large-tip-angle parallel RF pulses and blipped gradient trajectories” by Cao et al. published in Magn Reson Med. 2016; 75: 1198- 1208. Two-spoke and three-spoke 135° pulses were designed. The LTA optimizer was set only to update the RF coefficients and not alter the spoke positions. Twenty BOGAT iterations were used to reduce computation time. This was compared with FTA designs in which spoke positions were set based on the same STA calculation as previously and then fed to the LTA optimizer for RF optimization at those fixed spoke positions. FA maps were measured with the same TFL sequence.
[0125] We also applied BOGAT for spin-echo EPI in a volunteer. The methods are described in “Comparing parallel transmit 2-spoke against circularly polarized pulses for 7T whole brain diffusion MRI” by Zhang et al. published in Magnetic Resonance Imaging, except that BOGAT was used for pulse design. The scan was performed on the IMPULSE version MAGNETOM Terra 7T MRI scanner (Siemens Healthineers) at the University of California in Berkeley using a 16 Tx / 96 Rx head coil (MR CoilTech, Glasgow, UK). The excitation pulses had the same pulse design parameters as previously, and the refocusing pulse had a longer duration of 5.12 ms for each subpulse. To avoid eddy current effects, we did not increase the gradient and slew-rate limits for the pulses compared with our 7T Terra scans. Other sequence parameters were as follows: TR = 4400 ms, TE = 60 ms, 15 slices, 1.25-mm3 isotropic voxels, bandwidth = 1860 Hz / px, GRAPPA = 3, and partial Fourier = 6 / 8.
[0126] Multiband design: Field maps were acquired as above for 3 slices 40 mm apart in the agar phantom. BOGAT was applied with a cost function weighted by seed points between all three slices. This was compared to the default FTA method with spoke positions determined only by the central slice. RF coefficients were optimized with regularised MLS, with the Tikhonov factor determined by a scaled down by half, L-curve method, as described in: “Rankdeficient and discrete ill-posed problems: numerical aspects of linear inversion” by Hansen et al. published in SIAM monographs on mathematical modeling and computation. 1998, Philadelphia: SIAM, xvi, p. 247.
[0127] SMS multiband FA maps were measured in three separate scans. Each time magnetisation was prepared with the multiband pulse, but readout was for a single slice each scan. The pulse energy was calculated for the modulated and summed multiband pulse with optimized phase offset using the methods described in: “Optimized Phase Schedules for Minimizing Peak RF Power in Simultaneous Multi-Slice RF Excitation Pulses” by Wong published in Proc. ISMRM2012, 2012.
[0128] In vivo prediction: The same design methods were applied retrospectively for field maps from 9 healthy volunteers. 45° excitation for MB2, MB3 and MB4 each with 36 sets of slice groups and 8 Tikhonov factors were tested. Flip-angle root mean squared error, RMSE, was predicted using Bloch simulations. kr- Point pulse design validation: Finally, we also tested whether the BOGAT approach can usefully be generalized to nonselective excitation. To demonstrate this, we retrospectively designed 30° k-r-point pulses for whole-brain imaging for 1 subject in the same database. We compared the whole-brain FA RMSE performance for fixed 7k and 5k-r, fixed 4k? with 1 BOGAT point, fixed 3ki with 2 BOGAT points, and fixed 1k? with 4 BOGAT points. The fixed 7ky points were placed at k = ±0.05 cm-1symmetrically in all three directions; and in the latter cases, the positions with the lowest RF coefficient norms in the optimized 7k? pulse were removed. The RF optimization was performed with the same MLS optimizer as mentioned previously, with the Tikhonov factor fixed at 0.01 .
[0129] Results Figure 6 shows the cost function landscape plotted as a function of spoke position in fcT-space for a 2-spoke 30° pulse. The landscape is clearly non-convex. It tends to have minima in a ring around k=0. The landscape is generally smooth but has discontinuities which correspond to regions where the optimisation settles on a different set of subpulses. Small changes in the target flip angle or the Tikhonov regularization factor do not alter the general structure of the cost function landscape. The distribution of minima depends on the prescribed slab position. With the same optimization method, an optimal choice in the spoke position can on average decrease the RMSE by 25% or reduce the pulse energy by 42%. The minima for excitation fidelity (flip angle RMSE) do not in general coincide with the minima for pulse energy, illustrating the known trade-off between excitation fidelity and power deposition.
[0130] Figure 7 shows three iterations of the Bayesian optimization Gaussian process mean estimation of the spoke position cost function surface as iterations progress. The optimisation is calculated on one slice of in vivo field maps for a 2-spoke pulse. Figure 7 shows that Bayesian Optimization approximates the cost function closely (exploration) and converges to the global minimum (exploitation) within 50 iterations.
[0131] Figure 8 shows a single slice (SBS) 2- and 3-spoke flip angle map comparing FTA and BOGAT in a phantom test object, with both methods using a fixed Tikhonov regularization factor of 0.1. It can be seen that BOGAT achieves a more homogeneous excitation pattern in the agar phantom at a lower pulse energy. In particular, Figure 8 shows that for single slice pulse designs BOGAT achieves more homogeneous flip angle patterns (R SE=12.9° and 8.6° for 2- and 3-spoke respectively; 29% and 55% decrease) than FTA (RMSE=18.1 ° and 19.4°).
[0132] Figure 9 shows improvements in uniformity and energy deposition that may be achieved by using the BOGAT over an FTA approach measured in vivo single slice (SBS). The flip angle maps (2-spoke, target=90°) are measured at the same transverse slice in one subject, with pulses designed with FTA (top row) and with BOGAT (lower row). Masks were created by eroding and dilating the flip angle maps for both scans. These were intersected to make a final mask used to summarise both scans. BOGAT improves RMSE and reduces the pulse energy (except in the case where Tikhonov factor A=0.01 in the left-most column).
[0133] Figure 10 shows flip angle maps and pulse energies for multiband factor 3 pulse designs. Automatic Tikhonov regularization factors determined with L-curve were used. Note the markedly lower pulse energy (55% reduction) when designed with BOGAT. All other MLS optimization parameters for RF coefficients were kept constant in all pTx methods. Thus, Figure 10 shows that at a similar level of homogeneity, BOGAT reduces MB pulse energy in phantoms.
[0134] The mean total optimization workflow time for a two-spoke BOGAT MB3 pulse increased to 22.8 s compared to 13.7 s with FTA. As a reference, the grid searches with Bloch simulations shown in Figure 14 with 51 x 51 grid points took an average of 55.2 min for each slab on a 32-core computer. Hence, BOGAT achieved a global optimum spoke location more than 50 times faster than these grid searches.
[0135] Figures 11a to 11c show retrospective in vivo results of the present techniques. What is shown are flip angle RMSE and pulse energy scatter plots of (a) Multiband 2 (MB2) acquiring two slices simultaneously, (b) Multiband 3 (MB3) acquiring three slices simultaneously and (c) Multiband 4 (MB4) acquiring four slices simultaneously, pulse designs with BOGAT and FTA, each with 36 sets of slice groups in 9 subjects. Pulses were designed as 2-spoke 45° excitation, time-bandwidth product 4.0, repeated with 8 different Tikhonov factors between 0.001 and 0.5. The better spoke location found by BOGAT shifts the L-curve to the left, reducing RMSE by 12.8% (P<0.001) and pulse energy by 17.2% (P<0.001) on average. Thus, Figures 11a to 11c show how the present technique provide similar benefits for healthy volunteers, with 17.2% mean energy reduction (P<0.001) and 12.8% RMSE reduction (P<0.001). Of the 36 slice groups tested for each MB factor, BOGAT showed a mean energy reduction of 8.9% (single-band), 5.9% (MB2), 9.4% (MB3), and 7.0% (MB4) (p < 0.001 in all cases), while reducing FA RMSE by 11.4% (single-band), 35.6% (MB2), 19.7% (MB3), and 19.4% (MB4) (p < 0.001 in all cases).
[0136] Discussion
[0137] The present techniques, i.e. the BOGAT method take into account the cost function landscape of the spoke positions in / cr-space.
[0138] In particular, the minima in cost exist only in the low spatial frequency area (typically within 3 / FOV), and most significant local minima are also in this region of k-space. This is expected, since for a small number of spokes, the extra spoke(s) correct for the slow-varying inhomogeneity across the whole FOV due to the B+ spatial distribution for each channel and interference patterns between channels justifying the choice of the BOGAT search range ± 3 / FOV. Based on a weighted average of the cost function, a typical minimum in the two-spoke landscape occurs on a ring around the center of the kTspace. This observation has been used to initialize the seed points to accelerate BOGAT optimization.
[0139] The Gaussian process prior in the Bayesian optimization method can technically only represent smooth target functions, yet it is observed that this approach is effective at finding the global minimum even when that is in a discontinuous area. Biasing the acquisition function toward “exploration” can produce better results when the landscape is expected to contain more discontinuities, such as when optimizing single slices, in midbrain areas.
[0140] Unlike many interleaved algorithms, the BOGAT method treats the gradient optimization as a separate problem from the RF optimization. Therefore, while it is generally deemed not possible to seek a global minimum of the RF solution or the whole waveform, the present techniques are able to efficiently determine the global optimum of the gradient trajectory under a particular RF algorithm and its parameters. On the other hand, the interleaved gradient and RF optimization algorithms effectively increase the dimension of the problem, and hence create and converge to a more local optimum. Another advantage of the present techniques / BOGAT is that it is ‘modular’ as it allows various RF optimization algorithms to be inserted into BOGAT.
[0141] Because BOGAT uses an extra outer loop to optimize for the gradient positions, it may, at first glance, be deemed a very time-consuming algorithm. However, it should be noted that depending on the MR systems, a large proportion of the pTx optimization workflow time may consist of field-map conversions, calculation environment setups, visualizations, constraint checks, and output waveform conversions. Therefore, when the inner loop RF optimization is implemented efficiently, BOGAT does not significantly increase the overall workflow time. As shown in the results section, BOGAT computation times of about 10 s are suitable for online subject-specific optimizations of spoke pulses.
[0142] BOGAT can be applied for single band and multiband pulse design; the benefit appears to be greater for multiband pulse designs than single-band designs.
[0143] In summary, the present techniques, i.e. the BOGAT method was evaluated in phantoms and in vivo for single-band and multi-band excitation, and retrospectively in 9 volunteers using existing AB0and per-channel B / maps. BOGAT improves flip angle homogeneity (by 12.8% vs FTA, P<0.001) and reduces SAR (17.2%, P<0.001). Calculations take on average 9.1s extra for a set of MB3 pulses, making it feasible to use for online persubject optimisation. The present techniques wider applicability for human pTx neuroimaging were also tested. The results came from a 16 Tx / 96 Rx 7T head coil (MR Coiltech, UK), and in this case BOGAT recovered more signal in the frontal and temporal lobe areas, while reducing the total pulse energy by 14%.
[0144] The present technique, i.e. BOGAT, being an optimization outer loop, have been successfully demonstrated to be compatible with STA LS and a Jacobian matrix LTA solver for spoke pulse designs, and for k-r-point pulse design. However, detailed performance analysis and tuning for the LTA and k-r-point cases remain to be addressed. Likewise, it could be valuable to apply BOGAT with SAR-constrained RF optimizers, which would avoid the arbitrary Tikhonov regularization factors.
[0145] Furthermore, Bayesian optimization implementation has been chosen with an emphasis on computation efficiency, which is required for an online pulse optimizer. With ongoing rapid advances in Bayesian optimization algorithms, there may be better choices for the Gaussian process kernels (e.g., to better represent discontinuous functions) and acquisition functions.
[0146] Figure 12 is a diagram of an apparatus 1200 for identifying optimised parameters for a radiofrequency, RF, waveform for a magnetic resonance imaging, MRI, process. The apparatus 1200 comprises: at least one processor 1202 coupled to memory 1204, that are arranged for: receiving at least one calibration dataset comprising initial parameters of an RF waveform for imaging a target object using the MRI process; processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of the RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and the optimised parameters of the RF waveform for imaging the target object.
[0147] The apparatus 1200 may further comprise a communication module 1208 for transmitting the determined parameters to an external apparatus (not shown here) for use in an MRI process to image the target object. Thus, once the parameters have been determined, they may be sent to an external apparatus, such as an MRI machine, for use. The external apparatus may preferably be the same apparatus used during the calibration process.
[0148] Alternatively, the apparatus 1200 may itself be the MRI machine which is being used to image the target object or may be used to control the MRI machine. In this case, the apparatus may comprise MRI components 1206 for performing an MRI process. In this case, the determined parameters may be output for use by the apparatus itself (or an operator of the apparatus). Thus, the processor may be further arranged for generating instructions for the MRI components 1206 to emit an RF waveform using the determined parameters of the optimised RF waveform for the target object. The MRI components 1206 may be capable of producing a magnetic field with a large magnitude, such as, for example, at least 7T.
[0149] The at least one processor 1202 may comprise one or more of: a microprocessor, a microcontroller, and an integrated circuit. The apparatus 1200 comprises memory 1204 coupled to the at least one processor 1202. The memory 1204 may comprise volatile memory, such as random access memory (RAM), for use as temporary memory, and / or non-volatile memory such as Flash, read only memory (ROM), or electrically erasable programmable ROM (EEPROM), for storing data, programs, or instructions, for example.
[0150] Figure 13 is a diagram of a system 1400 for identifying optimised parameters for a radiofrequency, RF, waveform of a magnetic resonance imaging, MRI, process, the system comprising: an MRI machine 1312; and an apparatus 1200 comprising: at least one processor coupled to memory, and arranged for: receiving at least one calibration dataset comprising initial parameters of an RF waveform for imaging a target object using the MRI process; processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of the RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and the optimised parameters of the RF waveform for imaging the target object.
[0151] In the system 1400, the apparatus 1200 may not be able to control the MRI machine 1312 itself. Thus, the apparatus may further comprise a communication module 1208 for transmitting the determined parameters to the MRI machine for use in an MRI process to image the target object.
[0152] Alternatively, the processor 1202 of the apparatus 1200 may be used to control the MRI machine 1312 itself. Thus, the processor of the apparatus may further be arranged for generating instructions for the MRI machine to emit an RF waveform using the determined parameters of the optimised RF waveform for the target object; and the apparatus may further comprise a communication module for transmitting the generated instructions to the MRI machine for use in an MRI process to image the target object.
[0153] Figure 14 shows the FA RMSE, pulse energy, and cost function landscapes plotted as a function of spoke position in kT-space for a two-spoke 30° pulse. Figure 14a shows examples of grid search results for the flip angle (FA) RMS error. Figure 14b shows examples of grid search results for pulse energy. Figure 14c shows examples of grid search results for magnitude least squares (MLS) cost. All grid search results are with respect to the kx-ky spoke position of the first spoke in a two-spoke parallel-transmit (pTx) pulse. The second spoke was fixed at the kT space origin (kx = ky = 0.00 cm"1). Three slab positions (superior, middle, inferior) in one subject are shown. The pulse was designed for FA = 30° with Tikhonov regularization factor 0.1 in Figures 14a to c, FA 30° with Tikhonov factor 0.03 in Figure 14d, and FA 45° with Tikhonov factor 0.1 in Figure 14e. In Figures 14a to e, the white lines mark the watersheds of the cost function (i.e., they divide kx-kyspace into sections where a local optimization algorithm will converge to a single local minimum). The second spoke is always fixed at the origin (kx= ky= 0.00 cm-1). The landscape is clearly nonconvex with an average of 18 local minima in each slab. The cost function is generally smooth but has discontinuities that correspond to regions where the RF MLS optimization settles on a different set of subpulse coefficients. Note how adding more spokes gives a smoother cost function surface, albeit one that still contains many local minima. Small changes in the target FA or the Tikhonov regularization factor do not alter the general structure of the cost function landscape, comparing columns Figure 14c to e. With the same RF optimization method, the choice of spoke position alone can produce on average a 25% spread in the optimized FA RMSE and 42% spread in the pulse energy. The minima for excitation fidelity (FA RMSE) do not generally coincide with the minima for pulse energy, illustrating the known trade-off between excitation fidelity and power deposition. The distribution of minima depends on the prescribed slab position.
[0154] Figures 15a to 15c show how the BOGAT algorithm converges for two-spoke pulses. Figure 15a shows an error descent curve of Bayesian optimization of gradient trajectory (BOGAT) iterations. Convergence is plotted with respect to the global minimum determined by a grid search, for all 85 slab positions over 14 subjects calculated in the grid search. Figure 15b is a histogram of the relative cost when BOGAT was terminated at the 30th iteration.
[0155] Figure 15c compares the cost distribution of BOGAT results. The comparison is with respect to all local minima in the above mentioned slabs, where we see that BOGAT on average finds a better minimum than a gradient descent-based local optimization algorithm starting from a random point. Ninety-three percent of the BOGAT calculations reached a cost less than 1.1 times the global minimum within 30 iterations, whereas only 44% of local minima achieved this. The closest whole number of random starting points required for a local optimization to have a similar convergence performance as BOGAT is 7, with 91 % of cases reaching a cost less than 1.1-times the global minimum.
[0156] Figure 16 shows prospective in vivo validation of the present techniques. The flip-angle maps (two-spoke, target = 90°) measured at the same transverse slice in one subject, with pulses designed with Fourier transform approach (FTA; top row) and with Bayesian optimization of gradient trajectory (BOGAT; lower row). BOGAT improves RMS error (RMSE) and reduces the pulse energy (except in the case where Tikhonov factor A = 0.01 in the leftmost column). Prospective in vivo validation also showed improved FA homogeneity and / or reduced pulse energy at different Tikhonov factors; the measured FA maps for the STA designs for one subject are plotted in Figure 16. The largest improvement is with the highest Tikhonov factor tested (0.10), where BOGAT achieved 30% reduction in RMSE (10.3° vs. 14.7° for FTA, target 90°) and 29% reduction in pulse energy (0.30 J vs. 0.42 J for FTA) across all subjects. The mean extra computation time was 7.8 s.
[0157] For the LTA designs, the BOGAT two-spoke pulse achieved better FA uniformity with a lower pulse energy than the FTA three-spoke design (Supporting Information Figure S7). The BOGAT extra computation time was 75 and 159 s for two-spoke and three-spoke LTA designs, respectively.
[0158] Figures 17a to 17c show how BOGAT improves the image homogeneity in frontal lobe and temporal lobe, while also reducing the energy deposition. The images shown in Figure 17 are spin-echo single-shot EPI acquisitions acquired on an IMPULSE 7T scanner. Energy deposition is 1.86 J for BOGAT vs. 2.16 J for FTA for a 90°-180° pair. The images are T2 spinecho EPI images with a 16 Tx / 96 Rx head coil. Raw images with Fourier transform approach (FTA) pulse design are shown in Figure 17a. Images taken with the Bayesian optimization of gradient trajectory (BOGAT) pulse design approach are shown in Figure 17b. The percentage signal difference between BOGAT and FTA images is shown in Figure 17c. BOGAT images recover more signal in the frontal lobe and temporal lobes, while reducing the pulse energy. SAR, specific absorption rate.
[0159] Figure 18 shows retrospective kT-point pulse design for 1 subject with a target flip angle of 30°. As illustrated in Figure 18, each extra BOGAT-optimized kTpoint increases the whole-brain excitation homogeneity compared with fixed symmetrically placed points, with FA NRMSE values of 10.2% (5 fixed), 9.8% (7 fixed), 9.9% (4 fixed +1 BOGAT), 9.7% (3 fixed +2 BOGAT), and 9.4% (1 fixed +4 BOGAT). With 3 fixed points + 2 BOGAT points, the FA RMSE is already better than for 7 kT points. In this situation, BOGAT helps achieve a shorter pulse duration while still improving FA RMSE. Figure 18a shows the kT-point locations in the kT- space for pulses with and without Bayesian optimization of gradient trajectory (BOGAT) points (fixed points in yellow, BOGAT points in blue). Figure 18b shows simulated flip-angle (FA) maps for the center sagittal slice. Figure 18c shows the mean normalized FA RMS error (RMSE) for all voxels in the brain mask. For five fixed and even fixed kT-points, the gray lines were drawn to guide the eyes. From left to right, the whole-brain mean FA NRMSE values are 10.2% (5 fixed), 9.8% (7 fixed), 9.9% (4 fixed +1 BOGAT), 9.7% (3 fixed +2 BOGAT), and 9.4% (1 fixed +4 BOGAT).
[0160] Those skilled in the art will appreciate that while the foregoing has described what is considered to be the best mode and where appropriate other modes of performing present techniques, the present techniques should not be limited to the specific configurations and methods disclosed in this description of the preferred embodiment. Those skilled in the art will recognise that present techniques have a broad range of applications, and that the embodiments may take a wide range of modifications without departing from any inventive concept as defined in the appended claims.
Claims
CLAIMS1. A computer-implemented method for identifying optimised parameters for a radiofrequency, RF, waveform in a magnetic resonance imaging, MRI, process, the method comprising: receiving at least one calibration dataset comprising magnetic field information for imaging a target object using the MRI process; processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of an RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and the optimised parameters of the RF waveform for imaging the target object.
2. The method as claimed in claim 1, wherein the RF pulse is a spokes pulse comprising a sequence of RF subpulses interspersed with gradient blips, and wherein determining the parameters comprises using the Bayesian optimisation process to optimise a location, in kT- space, of each subpulse.
3. The method as claimed in claim 2, wherein determining the parameters further comprises using an iterative process that comprises: in a first step, picking the location in kT-space of one subpulse in the sequence of RF subpulses; in a second step, using an RF optimisation algorithm to obtain the RF parameters of the RF waveform given the fixed location of the subpulse; and repeating the first and second steps for all subpulses.
4. The method as claimed in claim 3 wherein using the RF optimisation algorithm comprises using one of: a magnitude least squares algorithm, or a large tip angle RF optimisation algorithm.
5. The method as claimed in claim 3 or 4, wherein determining the parameters comprises starting the iterative process using the last subpulse in the sequence of subpulses.
6. The method as claimed in any of claims 1 to 5, wherein receiving the at least one calibration dataset comprises receiving at least one calibration dataset corresponding to any of the following images acquired from an MRI process: multiple images; multiple images for each slice; simultaneous multi slice images; and a 3D volumetric RF image.
7. The method as claimed in any of claims 1 to 5, wherein receiving the at least one calibration dataset comprises receiving at least one generated calibration dataset that is generated using a trained machine learning model.
8. The method as claimed in claim 7 further comprising generating the at least one generated calibration dataset by using the trained machine learning model to process a set of localiser images.
9. The method as claimed in claim 7 further comprising generating the at least one generated calibration dataset by: selecting, using the trained machine learning model, at least one pre-computed pulse most suitable for imaging the target object from a database of pre-computed pulses for an MRI process.
10. The method as claimed in any one of claims 1 to 9 wherein: receiving at least one calibration dataset comprises receiving at least two calibration datasets for at least two slices; and determining the parameters comprises determining the parameters for each of the at least two calibration datasets separately.
11. The method as claimed in any one of claims 1 to 9 wherein: receiving at least one calibration dataset comprises receiving at least two calibration datasets for at least two slices, and determining the parameters comprises determining the parameters for each of the at least two calibration datasets jointly.
12. The method as claimed in any preceding claim wherein determining the parameters comprises using the Bayesian optimisation process to optimise a location, in kyspace, of gradient blips in a / cr-points pulse, wherein each / cr-points pulse comprises a sequence of RF subpulses interspersed with gradient blips.
13. An apparatus for identifying optimised parameters for a radiofrequency, RF, pulse for a magnetic resonance imaging, MRI, process, the apparatus comprising: at least one processor coupled to memory, and arranged for: receiving at least one calibration dataset comprising magnetic field information for imaging a target object using the RI process;processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of an RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and the optimised parameters of the RF waveform for imaging the target object.
14. The apparatus as claimed in claim 13 further comprising a communication module for transmitting the determined parameters to an external apparatus for use in an MRI process to image the target object.
15. The apparatus as claimed in claim 13 or 14, wherein the RF pulse is a spokes pulse comprising a sequence of RF subpulses interspersed with gradient blips, and wherein determining the parameters comprises using the Bayesian optimisation process to optimise a location, in kT-space, of each subpulse.
16. The apparatus as claimed in claim 15, wherein determining the parameters further comprises using an iterative process that comprises: in a first step, picking the location in kT-space of one subpulse in the sequence of RF subpulses; in a second step, using an RF optimisation algorithm to obtain the RF parameters of the RF waveform given the fixed location of the subpulse; and repeating the first and second steps for all subpulses.
17. The apparatus as claimed in claim 16, wherein determining the parameters comprises starting the iterative process using the last subpulse in the sequence of subpulses.
18. The apparatus as claimed in any one of claims 13 to 17 wherein: receiving at least one calibration dataset comprises receiving at least two calibration datasets for at least two slices; and determining the parameters comprises determining the parameters for each of the at least two calibration datasets separately.
19. The apparatus as claimed in any one of claims 13 to 17 wherein: receiving at least one calibration dataset comprises receiving at least two calibration datasets for at least two slices, and determining the parameters comprises determining the parameters for each of the at least two calibration datasets jointly.
20. The apparatus as claimed in any of claims 13 to 19, wherein the processor is further arranged for generating instructions for an MRI machine to emit RF and gradient signals using the determined parameters of the optimised gradient trajectory and RF waveform for the target object.
21. A system for identifying optimised parameters for a radiofrequency, RF, waveform of a magnetic resonance imaging, MRI, process, the system comprising: an RI machine; and an apparatus comprising: at least one processor coupled to memory, and arranged for: receiving at least one calibration dataset comprising magnetic field information for imaging a target object using the MRI process; processing, using a Bayesian optimisation process, the at least one calibration dataset to determine parameters of an optimised gradient trajectory and optimised parameters of an RF waveform to be used during the MRI process; and outputting the determined parameters of the optimised gradient trajectory and the optimised parameters of the RF waveform for imaging the target object.
22. The system as claimed in claim 21 wherein the apparatus further comprises a communication module for transmitting the determined parameters to the MRI machine for use in an MRI process to image the target object.
23. The system as claimed in claim 21 wherein: the processor of the apparatus is further arranged for generating instructions for the MRI machine to emit RF and gradient signals using the determined parameters of the optimised gradient trajectory and RF waveform for the target object; and the apparatus further comprises a communication module for transmitting the generated instructions to the MRI machine for use in an MRI process to image the target object.
24. A computer-readable storage medium comprising instructions which, when executed by a processor, causes the processor to carry out the method of any of claims 1 to 12.
Citation Information
Cited By
Radio frequency interference optimization method based on compliance constraint
CN121124963A
