Artificial intelligence-based magnetic resonance sequence

Abstract

For artificial intelligence-based optimization of MR sequences, MR pulse sequences are generated for the patient using a proxy machine of reinforcement learning training. The agent may generate values for a plurality or all of the parameters defining the MR pulse sequence. The agents are trained using a final objective or task (e.g., MR maps or partitions) as a reward function, so MR pulse sequences generated by the agents provide good quality MR imaging. The agent quickly generates MR pulse sequences without requiring multiple sequences to be used on the patient.

Description

Background Art

[0001] Magnetic resonance imaging (MRI) scanners are complex machines that require specific MR pulse sequences to scan patients to generate images. Practitioners use a variety of sequences, including spin echo, fast spin echo, inversion recovery, short tau inversion recovery, fluid attenuated inversion recovery, gradient echo sequences, steady-state free precession and others. Each sequence includes many options. These sequences are determined by empirical methods, where practitioners perform trial and error experiments to optimize parameters such as echo time, repetition time, echo sequence length or flip angle to achieve desired contrast and reduce artifacts. It may be difficult to achieve good quality. Multiple MRI scans are often required to compensate for the sub-optimal quality of each individual scan, resulting in lengthy inspection times. Using multiple sequences during the inspection is time-consuming, expensive, and increases the workload of radiologists who must analyze all the resulting images. Although artificial intelligence (AI) has allowed MR reconstruction with higher acceleration, thereby reducing overall inspection time, MRI sequences are typically still manually designed. Summary of the invention

[0002] By way of introduction, the preferred embodiments described below include methods, systems, instructions, and non-transitory computer-readable media for artificial intelligence-based optimization of MR sequences. An agent machine trained using reinforcement learning generates MR pulse sequences for a patient. The agent can generate values ​​for multiple or all parameters that define an MR pulse sequence. The agent is trained using an ultimate goal or task (e.g., MR map or segmentation) as a reward function so that the MR pulse sequences generated by the agent provide good quality MR imaging. The agent generates MR pulse sequences quickly and without the need to use multiple sequences on a patient.

[0003] In a first aspect, a method for establishing an MR pulse sequence is provided for an MR scanner. An indication of an environment for an MR scan of a patient is received. A reinforcement-learned AI establishes an MR pulse sequence for the environment based on the input of the environment and a digital twin of the patient. The reinforcement-learned AI establishes the MR pulse sequence in the environment of the digital twin of the patient by optimization. The MR scanner is configured using the MR pulse sequence, and imaging is performed as the MR scanner configured using the MR pulse sequence.

[0004] In a second aspect, a method for reinforcement machine learning to establish an MR pulse sequence is provided. A state space representing an MR pulse sequence, an MR scanner, and a virtual object to be scanned is defined. An action space of the MR pulse sequence is parameterized using a plurality of objects corresponding to actions that change a characteristic of the MR pulse sequence. A processor performs reinforcement learning of an agent. Reinforcement learning simulates different MR pulse sequences, different MR scanners, and different virtual objects of the state space in different combinations, wherein each combination is optimized in the action space. The agent is stored.

[0005] In a third aspect, a magnetic resonance (MR) system includes an MR scanner configured to scan a region of a patient via a setting of controls. A processor is configured by a machine-learned agent to configure the MR scanner. The machine-learned agent uses a simulation of an MR scan performed by the MR scanner using different values ​​of the setting by simulating a sequence of actions. The different values ​​are optimized by the machine-learned agent, wherein the optimized values ​​are used to set the configuration of the MR scanner.

[0006] Other aspects are summarized below in the illustrative embodiments. Aspects noted for one type of illustrative embodiment (method or system) may be used in another type. The present invention is defined by the following claims, and nothing in this section should be construed as limiting those claims. Further aspects and advantages of the present invention are discussed below in conjunction with the preferred embodiments, and may be claimed thereafter independently or in combination. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a block diagram of an embodiment of an MR system for MR imaging using a pulse sequence established by artificial intelligence;

[0008] Figure 2 is a flow chart of one embodiment of a method for enhanced machine learning for MR pulse sequence generation;

[0009] Figure 3 Illustration of reinforcement machine learning in an MR environment;

[0010] Figure 4 Illustration of a simulation of reinforcement training of an agent using the final task generated by the graph;

[0011] Figure 5 illustrates a simulation of augmented training of an agent using an annotated (e.g., segmentation) final task; and

[0012] Figure 6 is a flow chart of one embodiment of a method for establishing an MR pulse sequence for an MR scanner by artificial intelligence. DETAILED DESCRIPTION

[0013] Some attempts have been made to optimize MRI pulse sequences. Two methods have been used: maximizing Fisher information, i.e. minimizing Cramer-Rao bounds, and minimizing the correlation of the Bloch magnetization responses of different tissue parameters (T1, T2). These methods are limited to using local criteria to optimize only one or two parameters, such as flip angle or repetition time. These methods do not capture the ultimate goal of MRI sequences. Additionally, the methods ignore the influence of sampling patterns, which may generate artifacts with more serious effects compared to the Gaussian noise assumed in these models. The sequences optimized using these criteria have improved the targeting criteria, but ultimately deteriorated the reconstruction results. When using gradient-based optimization, the presence of multiple local minima in the optimization problem leads to suboptimal solutions close to the initial solution.

[0014] Instead of optimizing the MR pulse sequence, the sampling pattern of the already acquired scan data is optimized in another approach. The MR pulse sequence is not optimized. Full sampling acquisition is used, and the k-space data is retrospectively downsampled to simulate different sampling patterns. The losses are propagated back to the sampling pattern parameters in an end-to-end manner. The framework cannot be extended to change other parameters of the MRI sequence, because changes in other parameters will change the contrast of the target and require new acquisition of the data to generate the baseline truth. It is also expected that because of the non-differentiability problem, the gradient information cannot be used to optimize most of the sequence parameters. The method is limited to sequences where full sampling acquisition is possible, and for many types of sequences (such as HASTE, MRF, DWI EPI and GRASP), this method is not feasible.

[0015] To overcome the limitations of the above methods, a reinforcement learning (RL) agent is trained to optimize the parameters of MRI sequences. The RL agent is trained to optimize MRI sequences for different scanners with varying specifications and datasets of different objects with various body regions, where the goal is to solve specific tasks, such as image reconstruction of different tissue maps, organ segmentation, or tumor detection, while minimizing acquisition time. Large-scale cluster computing and AI are used to find the optimal MR sequence that can achieve the desired imaging goals while minimizing both the number of scans and acquisition time required in the MR scanner. The time and cost of the examination are reduced while improving efficiency and accuracy.

[0016] In one implementation, all parameters of the MR pulse sequence, including the sampling pattern and other parameters, can be optimized, resulting in a more optimal MR pulse sequence. In another implementation, the reward (loss function) used in reinforcement learning is global, because, in contrast to local heuristics, the reward is based on the final task. As a result, the effects of different sampling pattern artifacts, coil sensitivity, spectral noise or other defects or characteristics of the system, or reconstruction and segmentation algorithms are taken into account. The sequence is adapted to the final goal. In contrast to gradient-based optimization, AI-based RL methods are used for optimization. These gradient-based methods are limited to continuous optimization that converges only to local minima, i.e., often close to the initial conditions. RL methods are better adapted to this type of optimization problem.

[0017] Figure 1 An embodiment of an MR system for performing MR scanning with an MR scanner 90 is shown. A processor 160 configures the MR scanner 90 for scanning. An RL agent implemented by the processor 160 optimizes the MR pulse sequence used by the MR scanner 90 to scan a patient 140. Multiple different parameters can be optimized for a particular patient's MR imaging environment. Because the RL agent is trained using a final task (e.g., a map (e.g., image) and / or annotation (e.g., segmentation or detection)) that is used for a reward, the resulting final task of MR imaging using the MR pulse sequence can be better for the final task (e.g., fewer artifacts and / or limited rescanning). The MR system performs Figure 5 The MR scanner 90 scans a given patient 140 using the optimized MR pulse sequence created by the agent.

[0018] The MR scanner 90 includes a main field magnet 100, a gradient coil 110, a whole body coil 120, a local coil 130, and a patient support (e.g., a bed) 150. The MR system includes the MR scanner 90, a processor 160, a memory 170, and a display 180. Additional, different, or fewer components may be provided for the MR scanner 90 and / or the MR system. For example, the local coil 130 or the whole body coil 120 is not used. In another example, a processor 160, a memory 170, and a display 180 are provided without the coils 100-120 and the patient support 150, such as a workstation that communicates with one or more MR scanners 90. In yet another example, the processor 160, the memory 170, and / or the display 180 are part of the MR scanner 90.

[0019] The MR scanner 90 is configured by the settings of the control to scan the area of ​​the patient 140. The values ​​of the settings configure the MR scanner 90 to scan the patient. The settings define the MR pulse sequence for pulse transmission and / or response measurement. The scan provides scan data in the scan domain. The MR scanner 90 scans the patient 140 to provide raw measurements. For scanning, the main field magnet 100 creates a static base magnetic field B0 in the body of the patient 140 positioned on the patient support 150. The gradient coil 110 generates a positioning-related magnetic field gradient superimposed on the static magnetic field. The gradient coil 110 generates positioning-related and shimmed magnetic field gradients in three orthogonal directions and generates a magnetic field pulse sequence. The whole body coil 120 and / or the local coil 130 receive radio frequency (RF) transmit pulses, thereby generating a magnetic field pulse (B1) that rotates the spins of protons in the imaging region of the patient 140. The MR pulse sequence includes gradient field pulses and / or RF transmit pulses.

[0020] In response to the applied RF pulse signals, the whole body coil 120 and / or the local coil 130 receive MR signals, i.e., signals from excited protons in the body when the protons return to an equilibrium position established by the static and gradient magnetic fields. The MR signals are detected and processed by the detector, thereby providing an MR data set of raw data (e.g., k-space data). The raw storage array of the memory 170 stores the corresponding individual measurements that form the MR data set.

[0021] The MR scanner 90 is configured by a processor 160 for scanning. Any of a variety of scanner controls may be set, such as k-space coordinates, TR, TE, flip angle, pulse envelope, carrier frequency, timing, duration, and / or raw transmit pulses. A protocol, with or without user input or modification, may at least initially establish the settings used for a particular scan. Any level of generality may be provided for the settings, such as actual variables or abstractions of actual variables for a particular hardware. A memory 170 stores the configuration (e.g., a predetermined MR pulse sequence for imaging) and the resulting raw data or measurements.

[0022] The patient support 150 is a flat or contoured board (eg, a bed) on which the patient 140 lies or is supported. In an open hole, given a larger hole, the patient support 150 can be formed as a recliner or chair.

[0023] The processor 160 configures the MR scanner 90 and / or determines the values ​​of one or more (e.g., multiple or all) settings for the MR pulse sequence. The processor 160 is a general purpose processor, a digital signal processor, a graphics processing unit, an application specific integrated circuit, a field programmable gate array, an artificial intelligence processor, a tensor processor, a digital circuit, an analog circuit, a combination thereof, or another device known or developed later for applying AI. The processor 160 is a single device, multiple devices, or a network. For more than one device, parallel or sequential processing division can be used. Different devices that make up the image processor 160 can perform different functions, such as optimizing the MR pulse sequence by one device and configuring the MR scanner 90 for scanning by another device. In one embodiment, the processor 160 is a control processor or other processor of the MR scanner 90. Other processors of the MR scanner 90 or outside the MR scanner 90 can be used. In another embodiment, a server, a workstation, or a computer implements the processor 160.

[0024] The processor 160 is configured by software, firmware, and / or hardware to perform its tasks. The processor 160 operates in accordance with instructions stored on a non-transitory medium (e.g., memory 170) to perform the various actions described herein. For AI, the processor 160 operates in accordance with the architecture and values ​​of learnable parameters learned during machine training.

[0025] The processor 160 is configured by a machine-learned agent to configure the MR scanner 90. Different values ​​of the MR pulse sequence are optimized by the machine-learned agent 175 through a sequence of actions taken by the agent to establish an MR pulse sequence that is optimal or sufficient for the environment. For example, the machine-learned agent uses a simulation of an MR scan performed by the MR scanner 90 using a sequence of simulated actions using different setting values. A digital twin or other virtual representation of the patient is used in the simulation. The optimized values ​​are used for the settings of the configuration of the MR scanner 90. In one implementation, the machine-learned agent 175 is configured by past training and / or user selection to establish the values ​​of multiple settings in the settings in the optimization. In another implementation, the past RL training and / or simulation of the agent 175 enables the optimization to be based on the final task of MR image generation and / or annotation (e.g., segmentation, detection, identification, classification and / or diagnosis).

[0026] The agent 175 of enhanced machine learning can be a neural network that accepts inputs of the environment (e.g., the patient, the MR scanner, and / or other state information) and outputs an MR pulse sequence (e.g., parameter values ​​of the MR pulse sequence). The agent 175 of enhanced machine learning is trained to control or act to establish an MR pulse sequence. Different actions (e.g., changing parameter values ​​of the MR pulse sequence) are performed by the agent 175 (e.g., a network of enhanced deep learning or a machine-trained actor) to find a combination of values ​​of settings for scanning the patient by the MR scanner 90. The agent can use simulation to determine actions, or provide actions without simulation.

[0027] The machine-learned agent 175 is trained with training data having ground truth. Losses based on simulated outputs compared to ground truth for one or more final tasks are used in optimization to train the agent 175. Any optimization, such as Adam, can be used. Any loss, such as cross entropy, L1 loss, or L2 loss can be used. Pre-training, cross-training, and / or continuous training can be used. Training data is collected from a database of examples executed under expert control and / or based on simulations. Ground truth can be curated, created through expert review, or based on simulations.

[0028] The memory 170 is a cache, a buffer, a RAM, a removable medium, a hard drive, or another computer-readable storage medium. Non-transitory computer-readable storage media include various types of volatile and non-volatile storage media.

[0029] The memory 170 stores settings for controls (e.g., settings for MR pulse sequences), machine-learned agents 175, scan data, simulation data, and / or outputs (e.g., MR images and / or annotations). The memory 170 may alternatively or additionally store instructions for the processor 160. The functions, actions, or tasks illustrated in the various figures or described herein are performed by the processor 160 in response to one or more instruction sets stored in or on a non-transitory computer-readable storage medium of the memory 170. The functions, actions, or tasks are independent of a specific type of instruction set, storage medium, processor, or processing strategy, and may be performed by software, hardware, integrated circuits, firmware, microcode, and the like operating alone or in combination.

[0030] In one embodiment, the instruction is stored on a removable media device for reading by a local or remote system. In other embodiments, the instruction is stored in a remote location for transmission over a computer network. In yet other embodiments, the instruction is stored in a given computer, CPU, GPU, or system. Because some of the constituent system components and method steps depicted in the accompanying figures can be implemented in software, the actual connection between the system components (or process steps) can be different depending on the way the present embodiment is programmed.

[0031] The display 180 is a CRT, LCD, plasma, projector, printer or other display device. The display 180 is configured by loading an image to a display plane or buffer. The display 180 is configured to display the settings of the MR pulse sequence, the MR pulse sequence and / or the output of the MR scanner 90 (e.g., MR images and / or annotations).

[0032] Figure 2 One embodiment of a method for reinforcement machine learning to establish an MR pulse sequence is illustrated. Deep reinforcement learning trains an agent (e.g., a neural network) to implement a policy for determining actions in setting an MR pulse sequence. The agent can be trained to set multiple or all settings that define an MR pulse sequence for a given situation. The agent can be trained using rewards based on a final task (e.g., a reconstructed map or segmentation).

[0033] Figure 2 The method by Figure 1 The system or another system is executed. For example, the image processor simulates MR scanning and reconstruction in different environments as part of the reinforcement training of the agent to establish an MR pulse sequence optimized for the environment through a sequence of actions taken based on the policy. In act 230, the memory stores the resulting network (i.e., the RL trained agent).

[0034] The actions are performed in the order shown (numerical or from top to bottom) or in other orders. For example, actions 200 and 210 are performed in the order shown, in the opposite order, or in reverse order. Additional, different, or fewer actions may be used. For example, actions are provided for scanning, for building an architecture for a machine learning network, and / or for selecting input metrics for reinforcement learning. As another example, actions for defining (action 200) and / or parameterizing (action 210) are not performed, such as in the case of using an existing environment and / or parameterization.

[0035] The definition of actions 200 and the parameterization of actions 210 provide the action space of the agent and the state of the agent. The simulation is used to provide a reward function that is employed to train the RL agent to optimize the MR sequence.

[0036] In act 200, based on user input and / or programming, the processor defines a state space representing the MR pulse sequence, the MR scanner, and the virtual object to be scanned. During the deployment or testing phase, the user will provide specific details of an instance of this actual environment, such as providing a scanner model, an object or object type to be scanned, a body region, orientation, size, and / or a final task (e.g., T1 and / or T2 map reconstruction or organ segmentation). The user will also specify the sequence parameters they wish to optimize, which may include all sequence parameters. This information provides the state space. For training, various instances (states) of these inputs are used with simulations.

[0037] In one implementation, the state representation includes several components: (1) the current sequence description, (2) the hardware specifications of the scanner, such as B0, maximum gradient Gmax, and slope Smax, (3) information about the virtual object (e.g., many samples of a digital twin of a patient or phantom), such as proton density M0, T1, T2, T2*, off-resonance and / or non-rigid motion field, (4) the current prediction, i.e., the current image reconstruction and / or annotation output, (5) the target that can be extracted from the virtual object, if it is an organ, one or more tissue maps and / or annotations (e.g., which can be automatic or manual segmentations), and / or (6) a constraint mask that indicates which parameters are optimized and which are not optimized, constraints (limits) set, and / or filters for actions. The constraint mask can assist in developing an agent to optimize any subset of the action space.

[0038] Other state spaces may be created, such as MR scanner information, patient information, and MR pulse sequence information. Figure 3 An example is shown. An environment 310 includes an MRI scanner description 312, an object description 314 (e.g., a virtual patient), and an MRI pulse sequence description 316. The environment 310 also includes an MR simulator 318 to simulate an MR imaging scanner 312 as described that scans an object 314 as described using an MRI sequence 316 as described. The environment 310 includes an MR reconstruction 320 for generating output from the MR simulator 318. The environment 310 provides states St and rewards Rt to the agent 175, and the agent 175 outputs actions that change or update the MRI sequence description 316. Through reinforcement machine training using many samples, the agent 175 (e.g., a neural network with various learnable parameters) is trained to establish or optimize MR pulse sequences through actions in various environments 310.

[0039] exist Figure 2In act 210, the processor parameterizes the action space of the MR pulse sequence based on user input and / or programming. Parameterization defines variables used to create or establish the MR pulse sequence. Any level of abstraction can be used for parameterization. For example, settings used by a user of the MR scanner are used, such as flip angle and repetition time. As another example, variables of circuit, processing and / or hardware settings at the MR scanner hardware level are used, such as analog-to-digital sampling rate and amplifier amplitude. In yet another example, a level of abstraction between user settings and hardware settings is used, such as variables that define different characteristics or aspects of the MR pulse sequence.

[0040] The parameterization of the action space provides the agent with settings and corresponding values ​​that define MR pulse sequences that can be changed as actions. The parameterization defines actions that can be taken by the agent in the environment to optimize the MR pulse sequence.

[0041] Any format for parameterization can be used, such as a spreadsheet or data table. For different settings associated with an object, a hierarchical data structure can be used, such as an object with nested or sub-object variables (e.g., properties). In one implementation, parameterization provides multiple objects corresponding to actions that change the properties of an MR pulse sequence.

[0042] Example parameterization defines various actions as sequences, where each action in the sequence is associated with an object (i.e., a sequence of changes to an object). A sequence object defines an operation taken on an object. A sequence object may refer to an object such as: a block, an RF event, a gradient event, an analog-to-digital conversion (ADC) event, a shape (e.g., a pulse shape), and / or a delay (e.g., a delay of a pulse from a start time). These sequence objects have a type (block, RF event, gradient event, ADC event, shape, or delay) and an identification number (ID) for referring to the object in the sequence description. In each of these eight types of actions, there are two or three sequential steps labeled by letters (a, b, c). Examples are provided as follows:

[0043] 1) Create a new object

[0044] a. Specify the object type among: Block, RF Event, Gradient Event, ADC Event, Shape, Delay

[0045] b. Specify the ID number of the object before the newly created object

[0046] 2) Delete existing objects

[0047] a. Specify the object type among: Block, RF Event, Gradient Event, ADC Event, Shape, Delay

[0048] b. Specify the ID number of the object to be deleted

[0049] 3) Modify the block

[0050] a. Specify the block index

[0051] b. Specify the tag to be modified among the following 6: <duration> <rf> <gx> <gy> <gz> <adc>,in <duration>is the duration of the current block, <rf>is the ID of the RF event, <gx>is the ID of the gradient event on the X channel, <gy>is the ID of the gradient event on the Y channel, <gz>is the ID of the gradient event on the Z channel, <adc>Is the ID of the ADC event.

[0052] c. Apply incremental changes (e.g., + / - 1) to the selected events

[0053] 4) Modify RF events

[0054] a. Select the event ID

[0055] b. Specify one of the RF parameters (tags): <amp><mag id><phase id> <time id> <delay> <freq> <phase>,in <amp>is the peak amplitude,<mag_id> is the shape ID of the amplitude profile,<phase_id> is the shape ID of the phase profile,<time_id> is the shape ID of the time sampling point, <delay>is the delay before starting the RF pulse, <freq>is the frequency offset, and <phase>is the phase shift.

[0056] c. Apply incremental changes to the selected events (eg, + / - 1 depending on the unit).

[0057] 5) Modify gradient events

[0058] a. Select the event ID

[0059] b. Specify one of the gradient parameters: <amp> <rise> <flat> <fall> <delay>,in <amp>is the peak amplitude,<shape_id> is the shape ID of the arbitrary gradient waveform,<time_id> is the shape ID of the time sampling point, <rish>is the rise time of the trapezoid, <flat>is the flat top time of the trapezoid, <fall>is the fall time of the trapezoid, and <delay>is the delay before starting a gradient event.

[0060] c. Apply incremental changes to the selected events (eg, + / - 1 depending on the units).

[0061] 6) Modify ADC events

[0062] a. Select the event ID

[0063] b. Specify one of the ADC parameters: <num> <dwell> <delay> <freq> <phase>,in <num>is the sample size, <dwell>is the dwell time of the ADC, <delay>is the delay between the start of the block and the first sample, <freq>is the frequency offset of the ADC receiver, <phase>is the phase offset of the ADC receiver.

[0064] c. Apply incremental changes to the selected events (eg, + / - 1 depending on the units).

[0065] 7) Modify the shape

[0066] In a first embodiment, a pre-computed shape dictionary is utilized, which does not require any modification of the shapes themselves. Instead, the shape's ID is used to specify which shape should be used in a given block.

[0067] In the second embodiment, each sample of each shape is customized. In this scenario, the following actions can be performed:

[0068] a. Select the shape's ID.

[0069] b. Select the ID of the sample in the shape

[0070] c. Apply incremental changes (e.g., + / - 1) to the selected events.

[0071] 8) Modify delay events

[0072] a. Select the event ID

[0073] b. Apply incremental changes (eg, + / - 1) to the selected events.

[0074] Other combinations or sets of actions may be provided, such as in the above example using sequence objects (eg, actions) 1-6 and 8, rather than 7. Other actions, objects, and / or sub-objects may be provided.

[0075] To represent the action space, one approach is to start with an ID dimension of, for example, 256 (which can be adjusted and corresponds to the maximum number of events per object type). The second dimension is used to select pairs of (action, label), such as {(1, block), (1, RF event), ..., (2, block), (2, RF event), ..., (3, <duration> )，(3， <rf>), ..., (8)}. The action numbers correspond to the eight types of actions mentioned earlier, and the final dimension of the action space labels the action step, i.e., the value of the incrementing or decrementing step. Other formats for parameterization may be used.

[0076] If each sample is optimized, the action space dimensionality can increase significantly for shape objects, since the ID of each sample is encoded in an additional dimension of the action space along with the shape ID. One way to reduce this dimensionality is to instead utilize a fixed dictionary of shapes. Alternatively, fixing other optimization parameters and training a separate RL agent separately for optimizing shapes can also help slow down the growth of dimensionality. These two optimization techniques can be applied alternatively in a coordinate descent strategy.

[0077] In act 220, the processor performs reinforcement learning to train the agent. Figure 3 As shown in , reinforcement learning uses an MR simulator 318 and MR reconstruction 320 in an environment 310 to train from various states provided by an MR imaging scanner description 312, an object description 314, and an MR imaging sequence description 316. The states are provided by the simulation of an MR scanner of the MRI scanner description 312 performing an MR scan on an object of the object description 314 using an MRI sequence description 316. Many different states can be provided by different combinations of virtual objects, MR scanners, and MRI pulse sequences. The agent 175 is machine trained to act to change the MRI sequence descriptor 316 to optimize the MR pulse sequence for each or many possible combinations in the environment 310. Reinforcement learning simulates different MR pulse sequences, different MR scanners, and different virtual objects in the state space in different combinations, where the MR pulse sequence descriptor 316 for each combination is optimized in the action space.

[0078] The image processor applies deep reinforcement machine learning. The machine learns from training data (simulations). A wide range of multiple examples or combinations of environments are used for learning. The RL agent is trained using the agent's action space, the agent's state, and the reward function to optimize the MR sequence.

[0079] The learning is deep reinforcement learning. Deep learning uses neural networks. Reinforcement learning learns to decide the next action in the action space that defines the MR pulse sequence. Any architecture or layer structure for deep reinforcement learning can be used. Training is performed on a dataset of phantom objects and simulated scanners, where the RL agent policy is updated using an RL algorithm, such as Proximal Policy Optimization (PPO), Advantage Actor-Critic (A2C / A3C), Deep Deterministic Policy Gradient (DDPG), Trust Region Policy Optimization (TRPO), Soft Actor-Critic (SAC), Deep Q Network (DQN) or REINFORCE.

[0080] The simulation provides training data. Since the training is to adapt to changes in the environment, the training data includes many combinations or states. The simulation provides many examples that can be generated by different MR scanners, objects, and MR pulse sequences. Using an MR simulator 318 that mimics the behavior of a specified scanner, an MR sequence is simulated using virtual or phantom objects and the sequence is evaluated on a tissue map reconstruction task (which can involve generating a tissue atlas, such as T1, T2, and PD maps) or a target task (such as organ segmentation). It is also possible to combine these two tasks together.

[0081] Figure 4 and Figure 5 An example simulation of an environment using state space is shown. The MR simulator 318 uses the current state of the MRI scanner description 312 (e.g., currently selected MR scanner characteristics), the object descriptors 314 (e.g., the currently selected virtual object to be scanned), and the MRI sequence descriptors 316 (e.g., the currently established MR pulse sequence) to create MR raw data 410 (e.g., k-space data). The simulation utilizes MR reconstruction 320 for reconstruction. The reconstruction can be fixed or can form part of the environment 310, which can be altered in state space for training purposes. The reconstruction provides a reconstructed tissue map 430 (e.g., an image). In Figure 4 In , the ultimate goal or task is the organizational diagram 430, so the loss function 440 determines the difference between the simulated organizational diagram 430 and the ground truth. Figure 5 , the final goal is the annotation of the tissue map, so this downstream task 500 is performed on the tissue map 430 created in simulation. The loss function 440 compares the annotation (the output of the downstream task) with the ground truth provided from the dataset annotation 510 (i.e., from the object description 314 that defines the virtual annotation (e.g., segmentation). The loss from the loss function 440 is used to reward the agent 175. The agent uses the reward to learn the values ​​of the learnable parameters of the agent 175 for better selection of changes to the MRI sequence description 316 in reinforcement learning.

[0082] The training data includes ground truth information. Since simulation is used, the ground truth of the annotations is provided by the object description. For example, the organ or tumor boundaries are known from the object description 314. The ground truth of the tissue map 430 is generated by modeling. The loss function 440 can be a combination of losses, such as both annotations and tissue maps. The loss function 440 can include other terms, such as a penalty for scan time (e.g., discouraging longer scan times).

[0083] The agent learns to perform actions (eg, sequence objects) that optimize an MR pulse sequence for a given state. Through a sequence of changes in the value of one or more parameters (eg, objects), the agent alters the MR pulse sequence to optimize MR imaging. The agent learns to optimize for various states.

[0084] The parameter values ​​of one or more (e.g., multiple or all) objects are optimized. The parameters are optimized by finding values ​​that result in good or sufficient MR imaging for the final task. It is possible to optimize only a subset space of parameters by constraining other parameters to remain fixed. For example, the user decides to optimize a list of flip angles (FA) and repetition times (RT). The corresponding objects and variables of FA and RT can be changed by the actions of the agent, while other parameters are fixed in value using constraint masking. Constraints are implemented by masking out actions that are impossible or violate hard constraints. This can be achieved by assigning infinite values ​​to the logic associated with such actions. In order to calculate the objective function, the proposed sequence undergoes MR reconstruction, which can be implemented using an algorithm or network that takes the k-space calculated from the MR simulator as input.

[0085] For parameters that can be changed, the changes or actions can be constrained. MR pulse sequences are constrained based on input from the user and / or hardware. These constraints limit the value of the parameter. For example, the amplitude can be constrained so that the action can change the amplitude within the constraint, but not exceed the constraint. Similarly, the type of action or sequence object can be restricted, such as only allowing specific types of changes in parameter values.

[0086] To train the agent, any reward can be used. For example, if the final goal or ultimate task of the MR pulse sequence is to reconstruct a tissue map (see Figure 4 ) or comments (see Figure 5 ), then the reward is based on the final task. The reward is greater in the case where the final result of the simulation using the MR pulse sequence minimizes the distance to the reference truth. Based on the final goal of MR imaging, a loss function is used to learn the optimization of the MR pulse sequence. In one embodiment, the optimization uses a loss function based on 1) a comparison of the reconstructed MR map from the simulation with (2) the reference truth MR map. Typically, the contrast changes at each time point unless there are constraints in the MR sequence action space that prevent it. In such a case, a pre-computed fingerprint dictionary based on a physical model (such as Bloch simulation or pharmacokinetic modeling) can be provided to the reconstruction task, and the reference truth can be generated using an MRF algorithm, a network, or another method. The reward function is a weighted sum of 1) a loss function based on the target task (tissue map) and 2) a scan time that is as short as possible. In the case where the final task is to reconstruct the situation (tissue map), the loss includes the error (e.g., L1 and / or L2 norm error) on any compilation of the tissue map (such as T1 map, T2 map and / or PD map).

[0087] Even in cases where T1, T2 and PD maps are the ultimate task, it is possible to reconstruct any contrast image based on these tissue parameters. One approach is to use a dictionary of magnetization responses to convert tissue parameters in contrast at different echo times. Another solution is to use a trained neural network to convert tissue maps to a specific contrast.

[0088] In another embodiment, the optimization uses a loss function based on a comparison of (1) the segmentation from the simulation with (2) a ground truth segmentation from a virtual object. If the target task is a different task than reconstruction, such as an organ segmentation task, a detection task, or a classification task, then the loss on the target task is used instead or as well. For reconstruction, the model can be a single network from the MR sequence to the target task, or decomposed into two modules: an intermediate reconstruction and a target task module (e.g., image segmentation).

[0089] The reward may be determined as a loss or from the loss of the loss function 440. A combination of reconstruction (tissue map) and annotation final tasks may be used. This combined loss as a reward may be used to enable few-shot adaptation to new scenes using a multi-objective RL approach (MORL). The reward and corresponding loss function may include multiple criteria, such as rewards for multiple final tasks (e.g., tissue map reconstruction and segmentation or segmentation and classification). In one approach, the reward and corresponding loss function include a reward for minimizing the MR pulse sequence length or a penalty for the MR pulse sequence length. Other reward functions may be used.

[0090] In one embodiment, the reward decays over iterations. For example, the reward decays exponentially over time to incentivize fast convergence or a limited number of actions. Any amount of exponential decay can be used. Linear or nonlinear reward changes can be used. Over time, training maximizes the sum of discounted or decayed rewards.

[0091] To improve the optimization process, a multi-resolution approach can be employed. Use coarser action steps (e.g., change amounts) and / or lower resolution state images / graphs at the beginning of training or inference. For later actions in the sequence, use finer action steps or resolutions, such as gradually increasing the resolution near the end. During training, different agents can be trained at various scales or resolutions. Then, during inference, optimization can be performed first at a coarser scale before proceeding to a finer scale. Alternatively, the agent is trained to operate at different resolutions in sequence. The agent can be trained to select different scales at different times, such as including the scale as a sequence object or action in the parameterization.

[0092] After training, the agent architecture is represented as a matrix or tensor. The learned convolution kernels, weights, connections and / or layers of the neural network(s) used for the agent are provided. The matrix or tensor represents the trained policy estimation network or agent.

[0093] In act 230, the machine-learned neural network generated by the machine learning (agent) is stored. The matrices / tensors and the operators or functional blocks of the architecture are saved in the memory. The machine-learned neural network can be stored locally, or transferred over the network, or by moving the memory to other computers, workstations and / or MR scanners.

[0094] The training uses a variety of simulations so that the trained network can be adapted to any environment. The machine-learned neural network (agent) is used to provide an optimized MR pulse sequence when applied to a specific MR scanner and patient. The agent infers actions to establish values ​​for MR pulse sequence parameters. The MR pulse sequence is used to configure the MR scanner to scan the patient. The machine-learned neural network is stored so that the network can be applied to scan the patient. Many copies can be stored at many different locations, but only one copy is needed per location to cope with many different combinations of MR scanners and patients (e.g., body regions, patient characteristics, etc.).

[0095] Figure 6 is a flow chart of one embodiment of a method for creating an MR pulse sequence for an MR scanner. The RL agent uses the current environment to act according to the learned policy to create the MR pulse sequence. Figure 3 The arrangement is used in inference or testing to create an optimized MR pulse sequence. The agent changes the environment by changing the MR pulse sequence, and then repeats the input and actions based on the changes. Through the sequence of actions or iterations, an optimized MR pulse sequence is established.

[0096] The method is Figure 1 The processor performs actions 600, 610, and 620 to create an MR pulse sequence and configure an MR scanner to scan a patient. The MR scanner is used in action 620 for configuration and is used in action 630 for MR imaging as configured. A display displays one or more final results, such as images of tissue maps reconstructed and / or rendered with or without annotations (e.g., classification or segmentation). Due to the use of reinforcement learning and / or trained agents, the patient can be scanned once with a MR pulse sequence of minimized length, thereby providing a good or sufficient final result.

[0097] During application of the machine-learned model (agent) to one or more different patients and corresponding different scan data, the same learned weights or values ​​of the machine-learned model are used. The model and values ​​of the learnable parameters of the agent do not change from one patient to the next, at least for a given time (e.g., weeks, months, or years) or a given number of uses (e.g., dozens or hundreds of times). These fixed values ​​and corresponding fixed models are applied sequentially and / or by different processors to scan data from different patients. The model can be updated, such as retrained or replaced, but new values ​​are not learned as part of the application of a given patient. In other embodiments, continuous learning is used. Since the simulation is used for application, the simulation can also be used to update the agent using continuous reinforcement learning on a patient-by-patient basis.

[0098] The method is performed in the order shown (from top to bottom or numerically) or in other orders. For example, actions 612 and 614 are performed together, simultaneously, or in any order. Additional, different, or fewer actions may be provided. For example, preset, default, or user input settings are used to provide a constraint table. As another example, the results of the scan are stored in a memory (e.g., a computerized patient medical record), transmitted over a computer network, and / or displayed on a display.

[0099] In act 600, the processor receives an indication of an environment for an MR scan of a patient. The patient's medical records, hospital information systems, user input, and / or other information sources provide or are mined for environment information for a given scan or appointment.

[0100] The processor receives state information. The indication is the value of the current instance of the patient scan. Information is received by the agent of the reinforcement learning AI in establishing an MR pulse sequence. For example, the processor receives the patient's body region, the size of the patient's body region, the output task of the MR scan, and the specifications of the MR scanner. Additional, different or fewer environmental indications can be provided. Information from which specific indications can be derived can be provided. The agent is to use the information (e.g., body region, size, output task, and specifications) when optimizing the MR pulse sequence, which is to be used for scanning a specific patient by a specific MR scanner for a specific diagnostic purpose. Information to be used by the agent for this optimization is received.

[0101] In one embodiment, during the deployment or testing phase, the user will provide specific details, including the scanner model, the object or object type they wish to scan (this can include a digital twin of a patient), the body region, orientation and size, and the desired task, such as T1 or T2 map reconstruction or organ segmentation. The user will also specify the sequence parameters they wish to optimize, which can include all sequence parameters. Any constraints can be specified. User input is received through a user interface (such as using a keyboard and mouse or trackball). The user can also indicate a suggested or initial MR pulse sequence. A preset or default pulse sequence - such as for the indicated MR scanner to provide the indicated end result - can be used as a starting point.

[0102] In action 610, the processor establishes an MR pulse sequence for the environment through a reinforcement learned AI (RL AI) based on inputs of the environment including a digital twin of the patient. The RL AI is a trained agent. The agent uses the current state of the indication from the environment to determine the next action for changing the current MR pulse sequence. After multiple iterations, the MR pulse sequence to be used to scan the patient is determined. The agent can indicate no more actions as a stopping criterion. Alternatively, the number of iterations is limited to a threshold, a reward threshold is applied, or another stopping criterion is applied.

[0103] In one embodiment, the agent (RLAI) acts to establish an MR pulse sequence without simulation. The context indications, including an initial (eg, default) MR pulse sequence, are used by the agent to sequence actions to alter the values ​​of one or more parameters of the MR pulse sequence.

[0104] In an alternative embodiment, the agent establishes the MR pulse sequence with simulation. The MR simulator 318 and MR reconstruction 320 with indications are used to reconstruct tissue maps and / or another final task (e.g., annotation). A digital twin of the patient is used in the simulation without actually scanning the patient. The agent can use the loss determined from the simulated ground truth as input when selecting the next action. The method also allows machine learning to be used to update the agent based on the current environment.

[0105] The processor applies a network (agent) trained by deep reinforcement machine learning to change the values ​​of the pulse sequence to determine the optimal MR pulse sequence. The agent as trained is applied. Any data used by the agent is input, such as environmental indications. For example, scanner model, object, body region, orientation, size, digital representation of the patient (digital twin), desired final task (e.g., T1 tissue map and segmentation), and any constraints (e.g., what parameters of the MR pulse sequence can be changed and / or restrictions on the values) are input indications from action 600 used by the agent to establish the MR pulse sequence.

[0106] The trained agent is configured by training to use the received instructions to determine actions for establishing an MR pulse sequence. For each iteration of establishment, one or more actions representing the changes to be made are generated. The network learned by deep reinforcement machine learning is trained to control the order of actions through iterations based on the learned strategy. For example, the actions are determined based on the learned Markov decision process. The current instructions are used to determine the probability distribution of parameter settings for the MR pulse sequence. The distribution provides a greater or lesser probability of being rewarded for the setting to provide an optimized MR pulse sequence. The action is determined by random or other sampling of the probability distribution. Probability distributions for more than one parameter can be used. Sampling can select which type of action and the setting of the action to be used. More than one action can be selected, such as changing the settings of two or more parameters for a given iteration.

[0107] The processor applies the agent of RL AI in the optimization of MR pulse sequence. The agent selects the action sequence with respect to the settings of different objects of parameterized MR sequence. The RL AI in the environment is used to establish the optimization of MR pulse sequence by changing it for actual scanning. The digital twin can be used to avoid multiple scans of actual patients. The action sequence adjusts the parameter values ​​of MR pulse sequence.

[0108] Any number of parameters of the MR pulse sequence may be set. In act 612, a plurality (e.g., three or more) of the different parameters defining or parameterizing the MR pulse sequence may be changed by the action. For example, all parameters defining the MR pulse sequence may be set (values ​​selected by the action). In other embodiments, only some of the parameters that may be changed are changed. In yet another embodiment, the initial MR pulse sequence is all null values ​​or pulses are absent. The baseline is changed to create the MR pulse sequence.

[0109] In cases where only the values ​​of a limited subset of parameters can be changed, then all or a subset of those parameters can be changed. For example, a user can indicate that some parameters are to be established and other parameters are to maintain fixed values. User input, scanning regulations (e.g., from the FDA), and / or the hardware capabilities of the MR scanner can constrain the MR pulse sequence. Based on past training and constraints, in the optimization of the MR pulse sequence, the agent can change all or only some of the parameters to be established while maintaining the values ​​of other parameters constant. Based on past training and constraints, the agent can change the parameters to values ​​within the constraints. The agent optimizes without violating the constraints.

[0110] In act 614, the RL AI (agent) considers the end goal of imaging. The end goal may be an input to the agent, so the agent considers the end goal as part of the environment when determining what actions to take. The MR pulse sequence is optimized for the end goal.

[0111] In another approach, the end goal determines the loss to be used. In the case of simulations performed when setting up an MR pulse sequence, the end goal determines the data to be compared for the loss. For example, a tissue map as simulated with the current MR pulse sequence is compared to a baseline truth. The tissue map is the end goal, so the loss is based on the tissue map. In other examples, the end goal is segmentation, multiple tissue maps, detection, and / or classification. The loss can incorporate data comparisons for each of the end goals, so the end goal is considered when determining what changes to make to the MR pulse sequence.

[0112] In another approach, an agent or group of agents determines actions at different resolutions. A sequence of actions is generated. The sequence begins with a coarser action (e.g., a larger step size in a change) and / or a limited set of actions, and ends with a finer action and / or a less or unlimited set of actions. For example, one agent operates with a coarser action. Another agent that operates with a finer action is then applied. Any number of scales or resolutions may be used. Multiple scales may additionally or alternatively use different resolutions in the generated output (map or annotation) and / or input environment.

[0113] In act 620, the processor configures the MR scanner with the MR pulse sequence. The optimized MR pulse sequence generated by the RL AI is loaded into the MR scanner. The MR scanner is configured to use the MR pulse sequence as established.

[0114] In act 630, the MR scanner images the patient. The MR scanner, as configured using an MR pulse sequence, transmits pulses and receives responses. Based on the configuration of the MR scanner, a pulse sequence is created. The pulse sequence is transmitted from the coil into the patient.

[0115] The resulting response of the tissue is measured by receiving the radio frequency signal at the same or different coils. The scan produces raw measurements as scan data. A diagnostic output is generated from the raw measurements. The response is reconstructed into one or more tissue maps and / or used for annotation (e.g., applying segmentation).

[0116] The display (display screen or device) displays the output (e.g., final task information). For example, one or more organizational diagrams are displayed. As another example, one or more annotations (such as segmentation, detection, or classification) are displayed with the organizational diagram. Figure 1 With or without organization Figure 1 The display may be at the MR scanner. The display may be part of a patient's report, a pop-up window, a lab result, or part of the patient's electronic health record.

[0117] Various illustrative embodiments are listed below.

[0118] In illustrative embodiment 1, a method for establishing a magnetic resonance (MR) pulse sequence for an MR scanner is provided. An indication of an environment for an MR scan of a patient is received. An MR pulse sequence for the environment is established by reinforcement-learned artificial intelligence based on input of the environment and a digital twin of the patient. Establishing the MR pulse sequence includes optimizing by reinforcement-learned artificial intelligence of the MR pulse sequence in the environment of the digital twin of the patient. The MR scanner is configured with the MR pulse sequence, and the patient is imaged by the MR scanner as configured with the MR pulse sequence.

[0119] Illustrative embodiment 2. The method of illustrative embodiment 1, wherein receiving the indication comprises receiving state information of the environment, the state information being used by the reinforcement-learned artificially intelligent agent in establishing the MR pulse sequence.

[0120] Illustrative embodiment 3. A method according to any one of illustrative embodiments 1 or 2, wherein receiving an indication includes receiving a body region of the patient, a size of the body region of the patient, an output task of the MR scan, and specifications of the MR scanner, the body region, size, output task, and specifications being used in optimization by an agent of artificial intelligence that has undergone reinforcement learning.

[0121] Illustrative embodiment 4. The method according to illustrative embodiments 1-3, wherein an agent including artificial intelligence through reinforcement learning is established to perform optimization using a sequence of actions to adjust parameter values ​​of an MR pulse sequence.

[0122] Illustrative embodiment 5. The method of illustrative embodiments 1-4, wherein establishing includes establishing values ​​for a plurality of different parameters of the MR pulse sequence.

[0123] Illustrative embodiment 6. The method of illustrative embodiment 5, wherein establishing comprises establishing values ​​for all different parameters that can be set in the MR scanner for the MR scan of the patient.

[0124] Illustrative embodiment 7. A method according to illustrative embodiment 5 or 6, wherein the multiple different parameters are parameters indicated by a user as parameters to be established, and other parameters are fixed, wherein establishing includes establishing values ​​of the multiple different parameters while maintaining the values ​​of other parameters constant during optimization.

[0125] Illustrative embodiment 8. A method according to any one of illustrative embodiments 1-7, wherein establishing includes establishing wherein at least one of the multiple different parameters is constrained by hardware of the MR scanner and / or by user input, and the optimization establishes the values ​​of the multiple different parameters as constrained.

[0126] Illustrative embodiment 9. The method according to any one of illustrative embodiments 1-8, wherein establishing includes establishing an MR pulse sequence optimized for an end goal by artificial intelligence through reinforcement learning, wherein the optimization takes into account the end goal of imaging.

[0127] Illustrative embodiment 10. The method of illustrative embodiment 9, wherein establishing includes considering the final target as a type of map, segmentation, and / or detection.

[0128] Illustrative embodiment 11. A method according to any of illustrative embodiments 1-10, wherein establishing includes applying an agent of artificial intelligence through reinforcement learning in optimization, wherein the agent selects a sequence of actions with respect to settings of different objects for parameterizing an MR sequence.

[0129] Illustrative embodiment 12. The method of any of illustrative embodiments 1-11, wherein building includes building on a sequence of different resolutions in steps of motion from coarser to finer.

[0130] Illustrative embodiment 13. A method for reinforcement machine learning for establishing a magnetic resonance (MR) pulse sequence, the method comprising: defining a state space representing the MR pulse sequence, an MR scanner, and a virtual object to be scanned; parameterizing the action space of the MR pulse sequence using multiple objects corresponding to actions that change the characteristics of the MR pulse sequence; performing reinforcement learning on an agent by a processor, wherein the reinforcement learning simulates different MR pulse sequences, different MR scanners, and different virtual objects in the state space in different combinations, wherein each combination is optimized in the action space; and storing the agent.

[0131] Illustrative embodiment 14. The method according to illustrative embodiment 13, wherein reinforcement learning includes optimization using a loss function based on the final goal of MR imaging.

[0132] Illustrative embodiment 15. The method according to illustrative embodiment 14, wherein the optimization uses a loss function based on a comparison of a reconstructed MR map from a simulation with a reference truth MR map.

[0133] Illustrative embodiment 16. The method of any of illustrative embodiments 14 or 15, wherein the optimization uses a loss function based on a comparison of a segmentation from a simulation with a ground truth segmentation from a virtual object.

[0134] Illustrative embodiment 17. The method of any one of illustrative embodiments 13-16, wherein the optimization in the action space includes optimization of parameters of multiple objects.

[0135] Illustrative embodiment 18. The method of illustrative embodiment 17, wherein defining includes defining constraints limiting parameter values ​​using an MR pulse sequence that is constrained based on input from a user and / or hardware.

[0136] Illustrative embodiment 19. A magnetic resonance (MR) system comprising: an MR scanner configured by a setting of controls to scan a region of a patient, the scanning providing scan data; a processor configured by a machine-learned agent to configure the MR scanner, the machine-learned agent using a simulation of an MR scan performed by the MR scanner using different setting values ​​by simulating a sequence of actions, the different values ​​being optimized by the machine-learned agent, wherein the optimized values ​​are used for the settings of the configuration of the MR scanner.

[0137] Illustrative embodiment 20. The MR system of illustrative embodiment 19, wherein the machine-learned agent is configured to establish values ​​for a plurality of settings in an optimization, wherein the optimization is based on a final task of MR map generation and / or segmentation.

[0138] Although the subject matter has been described in terms of exemplary embodiments, it is not limited thereto. Rather, the appended claims should be construed broadly to include other variants and embodiments that may be made by those skilled in the art.< / rf> < / duration> < / phase> < / freq> < / delay> < / dwell> < / num> < / phase> < / freq> < / delay> < / dwell> < / num> < / delay> < / fall> < / flat> < / rish> < / amp> < / delay> < / fall> < / flat> < / rise> < / amp> < / phase> < / freq> < / delay> < / amp> < / phase> < / freq> < / delay> < / time> < / amp> < / adc> < / gz> < / gy> < / gx> < / rf> < / duration> < / adc> < / gz> < / gy> < / gx> < / rf> < / duration>

Claims

1. A method for establishing a magnetic resonance (MR) pulse sequence for an MR scanner, the method comprising: receiving an indication of an environment for an MR scan of a patient; Establishing an MR pulse sequence for the environment based on input from the environment and the digital twin of the patient through reinforcement-learned artificial intelligence, including optimizing the MR pulse sequence through reinforcement-learned artificial intelligence in the environment of the digital twin of the patient; configuring the MR scanner using an MR pulse sequence; as well as The patient is imaged by an MR scanner configured, for example, using an MR pulse sequence.

2. The method of claim 1, wherein receiving an indication comprises receiving state information of an environment, the state information being used by the reinforcement-learned artificially intelligent agent in establishing an MR pulse sequence.

3. The method of claim 1 , wherein receiving an indication comprises receiving a body region of the patient, a size of the body region of the patient, an output task of the MR scan, and specifications of the MR scanner, the body region, size, output task, and specifications being used in optimization by an agent of artificial intelligence that has undergone reinforcement learning. 4 . The method of claim 1 , wherein an agent comprising artificial intelligence via reinforcement learning is established to perform optimization using a sequence of actions to adjust parameter values ​​of the MR pulse sequence. The method of claim 1 , wherein establishing comprises establishing values ​​for a plurality of different parameters of the MR pulse sequence. 6 . The method according to claim 5 , wherein establishing comprises establishing values ​​for all different parameters that can be set in the MR scanner for the MR scan of the patient.

7. The method of claim 5, wherein the plurality of different parameters are parameters indicated by a user as parameters to be established, and other parameters are fixed, wherein establishing comprises establishing values ​​of the plurality of different parameters while maintaining values ​​of other parameters constant in the optimization.

8. The method of claim 1, wherein establishing comprises establishing wherein at least one of the plurality of different parameters is constrained by hardware of the MR scanner and / or by input from a user, the optimizing establishing values ​​of the plurality of different parameters as constrained.

9. The method of claim 1, wherein establishing comprises establishing an MR pulse sequence optimized for an end goal by reinforcement-learned artificial intelligence, wherein the optimization takes into account an end goal of imaging.

10. The method of claim 9, wherein establishing comprises considering the final target as a type of map, segmentation and / or detection.

11. The method of claim 1, wherein establishing comprises applying an agent of artificial intelligence via reinforcement learning in optimization, wherein the agent selects a sequence of actions with respect to settings of different objects for parameterizing an MR sequence.

12. The method of claim 1, wherein building includes building on a sequence of different resolutions in steps from coarser to finer motions.

13. A method for enhanced machine learning for establishing a magnetic resonance (MR) pulse sequence, the method comprising: defining a state space representing an MR pulse sequence, an MR scanner, and a virtual object to be scanned; parameterizing an action space of the MR pulse sequence using a plurality of objects corresponding to actions that change a characteristic of the MR pulse sequence; performing reinforcement learning on the agent by the processor, the reinforcement learning simulating different MR pulse sequences, different MR scanners, and different virtual objects in a state space in different combinations, wherein each combination is optimized in an action space; and Storage agent.

14. The method of claim 13, wherein reinforcement learning comprises optimization using a loss function based on a final goal of MR imaging.

15. The method of claim 14, wherein the optimization uses a loss function based on a comparison of a reconstructed MR map from a simulation with a ground truth MR map.

16. The method of claim 14, wherein the optimizing uses a loss function based on a comparison of a segmentation from a simulation to a ground truth segmentation from a virtual object.

17. The method of claim 13, wherein the optimization in the action space comprises optimization of parameters of a plurality of objects.

18. The method of claim 17, wherein defining comprises defining constraints limiting parameter values ​​using an MR pulse sequence that is constrained based on input from a user and / or hardware.

19. A magnetic resonance (MR) system comprising: an MR scanner configured by the settings of the controls to scan a region of the patient, said scanning providing scan data; A processor configured by a machine-learned agent to configure an MR scanner, wherein the machine-learned agent uses a simulation of an MR scan performed by the MR scanner using different setting values ​​by simulating a sequence of actions, wherein the different values ​​are optimized by the machine-learned agent, wherein the optimized values ​​are used for the settings of the configuration of the MR scanner.

20. The MR system of claim 19, wherein the machine-learned agent is configured to establish values ​​for a plurality of settings in an optimization, wherein the optimization is based on a final task of MR map generation and / or segmentation.

Images