Real-time design of high-frequency pulses and gradient pulses in magnetic resonance imaging.

Abstract

Disclosed herein is a medical system 100, 300 comprising a memory 110 storing machine-executable instructions 120 and a convolutional neural network 122. The convolutional neural network is configured to receive as input a composite array 128 encoding a selection of at least one excitation field 324, 900, and in response output a radio frequency waveform 130 and a plurality of spatially selective gradient pulse waveforms 132. The convolutional neural network is a multitasking convolutional neural network. Execution of the machine-executable instructions causes the computing system 104 to receive (200) a selection 124 of at least one excitation field, receive (202) an initial pulse sequence command 126, encode (204) a composite array with the at least one excitation field, receive (206) a radio frequency waveform and a plurality of spatially selective gradient pulse waveforms in response to inputting the composite array into a convolutional neural network, and construct (208) a modified pulse sequence command 134 by modifying the initial pulse sequence command with the radio frequency waveform and the plurality of spatially selective gradient pulse waveforms.

Description

Technical Field

[0001] The present invention relates to magnetic resonance imaging, and more particularly, to the design of pulse sequences for controlling a magnetic resonance imaging system.

Background Art

[0002] A large static magnetic field is used by a magnetic resonance imaging (MRI) scanner to align the nuclear spins of atoms as part of a procedure for generating an image of a patient's body. This large static magnetic field is called the B0 field or the main magnetic field. Various quantities or properties of a subject can be spatially measured using MRI. Spatial encoding in magnetic resonance imaging is performed using a combination of a high-frequency (RF) waveform (or RF pulse) used to control the transmission coil of the MRI scanner and a plurality of spatially selective gradient pulse waveforms (gradient pulses).

[0003] The journal publication "Ultrafast (milliseconds), multidimensional RF pulse design with deep learning" by Vinding et al., Magn. Reson. Med. 2019, 82:586-599 discloses the use of a neural network that outputs a single-channel multi-dimensional RF pulse considering an input map of a desired excitation region of interest.

[0004] The academic paper "Joint Design of RF and gradient waveforms via auto-differentiation for tailored excitation in MRI" by Tianrui Luo et al., retrieved from ARXIV.org, Cornell University Library on August 24, 2020, discloses a method for the joint design of RF waveforms and gradient waveforms based on an auto-differentiable Bloch simulator.

Summary of the Invention

[0005] The present invention provides a medical system, a computer program, and a method as described in the independent claims. Embodiments are shown in the dependent claims.

[0006] Embodiments provide improved selective RF excitation. Embodiments achieve this by using a convolutional neural network trained to output both a high-frequency waveform and an accompanying spatially selective gradient pulse waveform used to generate the B1 field. A variety of different types of neural network architectures were investigated, and a multitask neural network yielded the best results. A multitask neural network provides multiple outputs in response to receiving a single input. In the example described herein, the convolutional neural network takes a complete composite array encoded by the desired RF excitation field and outputs a complete high-frequency waveform and a complete spatially selective gradient pulse waveform. If a two-dimensional slice is acquired, there are two spatially selective gradient pulse waveforms, and if the acquisition is three-dimensional, there are three spatially selective gradient pulse waveforms output.

[0007] In one embodiment, the present invention provides a medical system comprising a memory for storing machine-executable instructions and a convolutional neural network. The convolutional neural network is configured to receive as input a composite array encoding a selection of at least one excitation field of view. The composite array represents, for example, a two-dimensional slice or a three-dimensional volume acquired by magnetic resonance imaging.

[0008] The convolutional neural network is configured to output a high-frequency waveform and multiple spatially selected gradient pulse waveforms in response to receiving a composite array. The high-frequency waveform is used to control the transmitter in the magnetic resonance imaging system to generate a B1 magnetic field. The multiple spatially selected gradient pulse waveforms are each gradient pulse waveforms that are fitted to the high-frequency waveform to acquire magnetic resonance data for at least one excitation field of view.

[0009] The convolutional neural network is a multitask convolutional neural network having a first output for high-frequency waveforms and separate output for each of multiple spatially selected gradient pulse waveforms. For a two-dimensional magnetic resonance imaging protocol, there are two spatially selected gradient pulse waveforms, and for a three-dimensional magnetic resonance imaging protocol, there are three spatially selected gradient pulse waveforms.

[0010] The medical system further includes a computing system. The computing system receives a selection of at least one excitation field of view by executing a machine-executable instruction. This excitation field of view is, for example, a field of view of interest to the operator or physician for imaging. The computing system further receives an initial pulse sequence command configured to control the magnetic resonance imaging system to acquire k-space data describing the subject by executing a machine-executable instruction. The computing system further receives a high-frequency waveform and multiple spatially selected gradient pulse waveforms in response to inputting a composite array into a convolutional neural network by executing a machine-executable instruction.

[0011] The execution of machine-executable instructions further configures the computational system to configure the pulse sequence command to control the magnetic resonance imaging system so that the pulse sequence command acquires k-space data from at least one excitation field of view, thereby configuring the system to construct a modified pulse sequence command by modifying the initial pulse sequence command with a high-frequency waveform and multiple spatially selective gradient pulse waveforms. This embodiment is beneficial because it allows the excitation field of view to be easily specified by the operator of the magnetic resonance imaging system. This embodiment also allows for the specification of an excitation field of view that is arbitrarily sized or shaped. In other cases, this embodiment allows for the specification of multiple excitation fields of view for a single acquisition.

[0012] The initial pulse sequence command may be a preliminary or templated pulse sequence command that includes, for example, a region where RF pulses and gradient pulses for slice selection need to be specified. Thus, embodiments provide customization of pulse sequence commands for specific acquisitions by modifying those pulse sequence commands with high-frequency waveforms and multiple spatially selective gradient pulse waveforms.

[0013] This invention relates to the automatic generation of a high-frequency (RF) waveform and multiple (two or more) spatially selected gradient waveforms from an input excitation field, and by modifying the initial pulse sequence command, insofar as it relates to the RF waveform and gradient waveform. The automatically generated RF waveform and gradient waveform are returned in response to the input excitation field by a multitask convolutional neural network. This multitask convolutional neural network has several parallel output sections (for each waveform) and is not fully connected. This invention achieves very high accuracy in the resulting waveforms, i.e., deviates only slightly from ground truth information while using a relatively small number of nodes and weights compared to the use of a fully connected convolutional network.

[0014] In another embodiment, a convolutional neural network is trained by a computational system that repeatedly performs the following steps, the first iteration of which is to generate a training high-frequency waveform and a set of training spatially selective gradient pulse waveforms using a selective excitation pulse design algorithm. Another iteration of which is to compute a model excitation field by inputting the training high-frequency waveform and the set of training spatially selective gradient pulse waveforms into a magnetic resonance imaging signal model. Another iteration of which is to receive a forward-propagated high-frequency waveform and a set of forward-propagated training spatially selective gradient pulse waveforms by inputting the model excitation field into a convolutional neural network.

[0015] Another iterative step is to update the parameters of the convolutional neural network by backpropagating with a training high-frequency waveform and a forward-propagated high-frequency waveform, the backpropagation also being performed with multiple training spatial selection gradient pulse waveforms and multiple forward-propagated training spatial selection gradient pulse waveforms of the fitted pair. This embodiment is useful because it provides a convolutional neural network that can specify any shaped, spaced, and sized patient field of view. These steps may be repeated in a loop, but are typically performed in parallel, and it should be noted that backpropagation is for training the convolutional neural network and is performed as a vector computation process.

[0016] In another embodiment, the selective excitation pulse design algorithm is the Shinnar-Le Roux algorithm.

[0017] In another embodiment, the selective excitation pulse design algorithm is a small flip angle approximation algorithm.

[0018] In another embodiment, the selective excitation pulse design algorithm is a numerically optimized control algorithm.

[0019] In another embodiment, the magnetic resonance imaging signal model is a numerical solution of the Bloch equation.

[0020] In another embodiment, the pulse sequence command is configured to acquire k-space data according to a parallel imaging magnetic resonance imaging protocol. This embodiment is particularly beneficial because, when combined with a custom excitation field, it enables the acquisition of highly accurate and rapid magnetic resonance images.

[0021] In another embodiment, the computing system further receives the investigation magnetic resonance image by executing a machine-executable instruction. The computing system further renders the investigation magnetic resonance image on a display by executing a machine-executable instruction. In response to displaying the investigation magnetic resonance image, the computing system further receives a selection of at least one excitation field of view. The selection of at least one excitation field of view is located within the investigation magnetic resonance image. This embodiment is beneficial because it allows for customization of the excitation field of view for specific subjects and magnetic resonance imaging protocols.

[0022] In another embodiment, the selection of at least one field of view is received from an automated image segmentation algorithm. For example, when a radiologist prescribes a cardiac magnetic resonance imaging protocol, automated segmentation can be used to very precisely limit and specify the imaging of the heart.

[0023] In another embodiment, the selection of at least one field of view is received from the user interface. For example, a system operator can precisely trace out the desired field of view. This provides a more flexible means of performing magnetic resonance imaging.

[0024] In another embodiment, the selection of at least one field of view is received from an automated segmentation algorithm and a user interface. For example, some parts of the excitation field of view selection may be received from the automated imaging segmentation algorithm, and others from the user interface. The operator may also use the user interface to correct the segmentation from the algorithm.

[0025] In another embodiment, the multiple spatially selective gradient pulse waveforms are two spatially selective gradient pulse waveforms, each having at least one excitation field, a two-dimensional excitation field. For example, two gradient pulse waveforms are specified and used to take two-dimensional slices. However, they may be configured so that multiple slices are obtained.

[0026] In another embodiment, the plurality of spatial selection gradient pulse waveforms are three spatial selection gradient pulse waveforms. Each of at least one excitation field is a three-dimensional excitation field. This is advantageous because any three-dimensional excitation field may be specified.

[0027] In another embodiment, each of at least one excitation field has an arbitrary shape. The excitation field is specified by entering a composite value into a composite array used to encode the selection of at least one excitation field. This enables the specification of excitation fields having arbitrary shapes.

[0028] In another embodiment, each of at least one excitation field is spatially separated. For example, since the excitation fields can be arbitrarily specified for three-dimensional acquisition, multiple excitation fields may be specified. One example is a magnetic resonance imaging scan in which magnetic resonance images of breast tissue for two different breasts are acquired simultaneously but for different high-frequency excitation fields.

[0029] In another embodiment, the neural network is a multi-scale convolutional neural network.

[0030] In another embodiment, the convolutional neural network is a U-net convolutional neural network.

[0031] In another embodiment, the medical system includes a magnetic resonance imaging system. The computing system further controls the magnetic resonance imaging system by a modified pulse sequence command to acquire k-space data by executing machine-executable instructions. The computing system further reconstructs magnetic resonance imaging data from the k-space data by executing machine-executable instructions. This embodiment is beneficial because the excitation field is customized for a particular magnetic resonance imaging acquisition. This provides, for example, better imaging quality compared to a situation where pre-calculated excitation fields were used.

[0032] In another embodiment, the present invention provides a computer program comprising machine-executable instructions for execution by a computing system controlling a medical system. The computer program further comprises a convolutional neural network configured to receive as input a composite array encoding a selection of at least one excitation field of view. The convolutional neural network is configured to output a high-frequency waveform and a plurality of spatially selected gradient pulse waveforms in response to receiving the composite array. The convolutional neural network is a multitask convolutional neural network having a first output unit for the high-frequency waveform and separate output units for each of the plurality of spatially selected gradient pulse waveforms.

[0033] The execution of a machine-executable instruction causes the computing system to receive a selection of at least one excitation field of view. The execution of a machine-executable instruction further causes the computing system to receive an initial pulse sequence command configured to control the magnetic resonance imaging system to acquire k-space data describing the subject. The execution of a machine-executable instruction further causes the computing system to encode a composite array using at least one excitation field of view.

[0034] The execution of machine-executable instructions further enables the computing system to receive a high-frequency waveform and multiple spatially selective gradient pulse waveforms in response to inputting a composite array into a convolutional neural network. The execution of machine-executable instructions further enables the computing system to construct a modified pulse sequence command by modifying the initial pulse sequence command with the high-frequency waveform and multiple spatially selective gradient pulse waveforms configured to control the magnetic resonance imaging system so that the pulse sequence command acquires k-space data from at least one excitation field of view.

[0035] In another embodiment, the present invention provides a medical imaging method using a convolutional neural network. The convolutional neural network is configured to receive as input a composite array encoding a selection of at least one excitation field of view. The convolutional neural network is configured to output a high-frequency waveform and a plurality of spatially selected gradient pulse waveforms in response to receiving the composite array as input. The convolutional neural network is a multitask convolutional neural network having a first output unit for the high-frequency waveform and separate output units for each of the plurality of spatially selected gradient pulse waveforms.

[0036] The method comprises the step of receiving a selection of at least one excitation field of view. The method further comprises the step of receiving an initial pulse sequence command configured to control a magnetic resonance imaging system to acquire k-space data describing a subject. The method further comprises the step of encoding a composite array using at least one excitation field of view. The method further comprises the step of receiving a high-frequency waveform and a plurality of spatially selective gradient pulse waveforms in response to inputting the composite array into a convolutional neural network. Finally, the method further comprises the step of configuring a modified pulse sequence command by modifying the initial pulse sequence command with the high-frequency waveform and the plurality of spatially selective gradient pulse waveforms so that the pulse sequence command is configured to control a magnetic resonance imaging system to acquire k-space data from at least one excitation field of view.

[0037] In another embodiment, the present invention provides a method for training a convolutional neural network. The method comprises repeatedly performing the following steps. It should be understood that these steps may be repeated multiple times sequentially or multiple times in parallel. The method includes the step of generating a training high-frequency waveform and a plurality of training spatially selective gradient pulse waveforms using a selective excitation pulse design algorithm. The method further includes the step of calculating a model excitation field by inputting the training high-frequency waveform and the plurality of training spatially selective gradient pulse waveforms into a magnetic resonance imaging signal model.

[0038] The method further comprises the step of receiving a forward-propagated high-frequency waveform and a plurality of forward-propagated training spatial selection gradient pulse waveforms by inputting a model excitation field into a convolutional neural network. The method further comprises the step of updating the parameters of the convolutional neural network by backpropagating with the training high-frequency waveform and the forward-propagated high-frequency waveform. Then, backpropagation is similarly performed with a plurality of training spatial selection gradient pulse waveforms and a plurality of forward-propagated training spatial selection gradient pulse waveforms of the fitted pair.

[0039] It is understood that one or more of the above-described embodiments of the present invention can be combined, provided that the combined embodiments are not mutually exclusive.

[0040] As will be understood by those skilled in the art, aspects of the present invention are embodied as apparatus, methods, or computer program products. Accordingly, aspects of the present invention may take the form of entirely hardware embodiments, entirely software embodiments (including firmware, resident software, microcode, etc.), or embodiments combining software and hardware embodiments, which may all be comprehensively referred to herein as “circuits,” “modules,” or “systems.” Furthermore, aspects of the present invention may take the form of computer program products embodied in one or more computer-readable media in which computer executable code is embodied.

[0041] Any combination of one or more computer-readable media may be used. The computer-readable media may be computer-readable signal media or computer-readable storage media. As used herein, “computer-readable storage media” encompasses any tangible storage media that stores instructions executable by the processor or computing system of a computing device. Computer-readable storage media may also be referred to as computer-readable non-temporary storage media. Computer-readable storage media may also be referred to as tangible computer-readable media. In some embodiments, computer-readable storage media may also be capable of storing data that can be accessed by the computing system of a computing device. Examples of computer-readable storage media include, but are not limited to, floppy disks, magnetic hard disk drives, solid-state hard disks, flash memory, USB thumb drives, random access memory (RAM), read-only memory (ROM), optical disks, magneto-optical disks, and register files of computing systems. Examples of optical disks include compact disks (CDs) and digital multi-purpose disks (DVDs), such as CD-ROMs, CD-RWs, CD-Rs, DVD-ROMs, DVD-RWs, or DVD-R discs. The term "computer-readable storage medium" also refers to various types of recording media that can be accessed by computer devices via a network or communication link. For example, data may be retrieved via a modem, the Internet, or a local area network. Computer executable code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, fiber optic cable, RF, or any suitable combination thereof.

[0042] A computer-readable signal medium may include, for example, a propagating data signal in which computer-executable code is embodied, either in the baseband or as part of a carrier wave. Such a propagating signal may take any of a variety of forms, including but not limited to electromagnetic, optical, or any suitable combination thereof. A computer-readable signal medium may be any computer-readable medium that is not a computer-readable storage medium, on which a program for use by or in connection with an instruction execution system, apparatus, or device can communicate, propagate, or transport.

[0043] "Computer memory" or "memory" is an example of a computer-readable storage medium. Computer memory is any memory that is directly accessible to the computing system. "Computer storage" or "storage" is another example of a computer-readable storage medium. Computer storage is any non-volatile computer-readable storage medium. In some embodiments, computer storage may also be computer memory, or vice versa.

[0044] As used herein, “computational system” includes electronic components capable of executing programs, machine-executable instructions, or computer-executable code. References to computational systems, including examples of “computational systems,” should be interpreted as including, where applicable, two or more computational systems or processing cores. A computational system is, for example, a multi-core processor. A computational system also refers to a collection of computational systems within a single computer system or distributed across multiple computer systems. The term computational system should also be interpreted as referring to a collection or network of computing devices, each comprising a processor or computational system. Machine-executable code or instructions are executed by multiple computational systems or processors within the same computing device or even distributed across multiple computing devices.

[0045] Machine-executable instructions or computer-executable code include instructions or programs that cause a processor or other computing system to perform aspects of the present invention. Computer-executable code for performing operations according to aspects of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, Smalltalk®, C++ and conventional procedural programming languages ​​such as the C programming language or similar languages, and may be compiled into machine-executable instructions. In some cases, computer-executable code may be in the form of a high-level language or in a pre-compiled form and may be used with an interpreter that generates machine-executable instructions on a case-by-case basis. In other cases, machine-executable instructions or computer-executable code may be in the form of programming for a programmable logic gate array.

[0046] Computer executable code may run entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or wide area network (WAN), or the connection may be made to an external computer (for example, through the internet using an internet service provider).

[0047] Aspects of the present invention will be described with reference to flowcharts, diagrams, and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It will be understood that, where applicable, each block or part of a set of blocks in a flowchart, diagram, and / or block diagram can be implemented by computer program instructions in the form of computer executable code. It will be further understood that, where not mutually exclusive, combinations of blocks in various flowcharts, diagrams, and / or block diagrams may be combined. These computer program instructions may be provided to a general-purpose computer, a special-purpose computer, or a computer system of other programmable data processing equipment to create a machine such that instructions executed via the computer system or other programmable data processing equipment will produce means for performing one or more functions / actions specified in the flowchart and / or block diagrams.

[0048] These machine-executable instructions or computer program instructions may also be stored on a computer-readable medium that can instruct a computer, other programmable data processing device, or other device to function in a particular way, so as to produce a product containing instructions that perform a function / action specified in one or more blocks of a flowchart and / or block diagram.

[0049] Machine-executable instructions or computer program instructions may also be loaded onto a computer, another programmable data processing device, or another device to create a computer-executable process by causing a series of operational steps on the computer, another programmable device, or another device to be performed so that the instructions executed on the computer or other programmable device provide a process for performing a function / action specified in one or more blocks of a flowchart and / or block diagram.

[0050] As used herein, “user interface” refers to an interface that enables a user or operator to interact with a computer or computer system. A “user interface” may also be called a “human interface device.” A user interface provides information or data to an operator and / or receives information or data from an operator. A user interface may also enable input from an operator to be received by the computer, and may provide output from the computer to the user. In other words, a user interface may enable an operator to control or operate a computer, and may enable the computer to display the results of the operator's control or operation. Displaying data or information on a display or graphical user interface is one example of providing information to an operator. Receiving data via a keyboard, mouse, trackball, touchpad, pointer, graphic tablet, joystick, gamepad, webcam, headset, pedal, wired glove, remote control, and accelerometer are all examples of user interface components that enable the reception of information or data from an operator.

[0051] As used herein, “hardware interface” encompasses interfaces that enable a computer system’s computing system to interact with and / or control external computing devices and / or equipment. Hardware interfaces enable a computing system to send control signals or commands to external computing devices and / or equipment. Hardware interfaces also enable a computing system to exchange data with external computing devices and / or equipment. Examples of hardware interfaces include, but are not limited to, universal serial buses, IEEE 1394 ports, parallel ports, IEEE 1284 ports, serial ports, RS-232 ports, IEEE 488 ports, Bluetooth® connections, wireless local area network connections, TCP / IP connections, Ethernet connections, control voltage interfaces, MIDI interfaces, analog input interfaces, and digital input interfaces.

[0052] As used herein, “display” or “display device” encompasses output devices or user interfaces designed to display images or data. Displays output visual, auditory, and / or tactile data. Examples of displays include, but are not limited to, computer monitors, television screens, touchscreens, tactile electronic displays, braille screens, cathode ray tubes (CRTs), storage tubes, bistable displays, electronic paper, vector displays, flat panel displays, vacuum fluorescent displays (VFs), light-emitting diode (LED) displays, electroluminescent displays (ELDs), plasma display panels (PDPs), liquid crystal displays (LCDs), organic light-emitting diode (OLED) displays, projectors, and head-mounted displays.

[0053] k-space data is defined herein as recorded measurements of high-frequency signals emitted by atomic spins using the antenna of a magnetic resonance apparatus during a magnetic resonance imaging scan. Magnetic resonance data is an example of tomographic medical image data.

[0054] Magnetic resonance imaging (MRI) image data or magnetic resonance (MR) images are defined herein as reconstructed two-dimensional or three-dimensional visualizations of anatomical data contained within k-space data. This visualization can be performed using a computer.

[0055] Hereinafter, preferred embodiments of the present invention will be described as merely examples with reference to the drawings. [Brief explanation of the drawing]

[0056] [Figure 1] Here is an example of a healthcare system. [Figure 2] A flowchart illustrating how to use the medical system shown in Figure 1 is provided. [Figure 3] Further examples of healthcare systems are presented. [Figure 4] Figure 3 shows a flowchart illustrating how to use the medical system. [Figure 5] A flowchart illustrating how to train a convolutional neural network is shown. [Figure 6] An example of a convolutional neural network is shown. [Figure 7] This demonstrates the effectiveness of convolutional neural networks in providing a desired excitation field of view. [Figure 8] Show the desired field of view. [Figure 9] This shows the field of view obtained using a convolutional neural network and the field of view obtained using the LSR algorithm. [Modes for carrying out the invention]

[0057] Elements with the same numbering in the diagram are either equivalent elements or perform the same function. Elements discussed previously are not necessarily discussed in later diagrams if their function is equivalent.

[0058] Figure 1 shows a diagram illustrating an example of a medical system 100. In Figure 1, the medical system 100 is shown to include a computer 102, which includes a computing system 104. The computing system 104 is intended to represent one or more processing cores or computing systems, which may be located in the same place or distributed. The computer 102 is shown to include an optional hardware interface 106 and an optional user interface connected to the computing system 104. The hardware interface 106 allows, for example, the computing system 104 to exchange data with and control other components of the medical system 100. The user interface 108 allows an operator of the medical system 100 to control and interact with the medical system. The user interface 108 may include, for example, a graphical user interface or other graphical control device.

[0059] The medical system 100 is further shown to include a memory 110 that communicates with a computing system 104. The memory 110 is intended to represent any combination of memories accessible by the computing system 104.

[0060] Memory 110 is shown as containing machine-executable instructions 120. Machine-executable instructions 120 are instructions that enable the computing system to perform tasks such as controlling other components and performing various numerical and image processing tasks. Memory 110 is further shown as containing a convolutional neural network 122. This convolutional neural network is a multitask convolutional neural network having a first output unit for outputting high-frequency waveforms and separate output units for each of a plurality of spatially selected gradient pulse waveforms. The convolutional neural network outputs these in response to receiving a composite array encoding the selection of at least one excitation field of view for magnetic resonance imaging acquisition.

[0061] Memory 110 is further shown to include a selection 124 of at least one excitation field of view. Memory is further shown to include an initial pulse sequence command 126. These initial pulse sequence commands are, for example, a set of pulse sequence commands that are retrieved from a database or selected when it is desired that a particular protocol be executed.

[0062] Memory 110 is further shown to include a composite array used to encode a selection 124 of at least one excitation field of view. The memory is further shown to include a high-frequency waveform 130 and a plurality of spatially selective gradient pulse waveforms 132 received from the convolutional neural network 122 in response to inputting the composite array 128. Memory 110 is further shown to include a modified pulse sequence command 134. The modified pulse sequence command 134 consists of the initial pulse sequence command 126, modified to combine the high-frequency waveform 130 and the plurality of spatially selective gradient pulse waveforms 132.

[0063] Figure 2 shows a flowchart illustrating how to operate the medical system 100 of Figure 1. First, in step 200, at least one excitation field of view selection 124 is received. Next, in step 202, an initial pulse sequence command 126 is received. Next, in step 204, a composite array 128 is encoded using at least one excitation field of view 124. In step 206, a high-frequency waveform 130 and multiple spatially selected gradient pulse waveforms 132 are received in response to inputting the composite array 128 into the convolutional neural network 122. Finally, in step 208, a modified pulse sequence command is constructed by integrating the high-frequency waveform 130 and the multiple spatially selected gradient pulse waveforms 132 into the initial pulse sequence command 126.

[0064] Figure 3 shows a further example of a medical system 300. The medical system 300 is similar to the medical system 100 in Figure 1, except that it further comprises a magnetic resonance imaging system 302 controlled by a computing system 104.

[0065] The magnetic resonance imaging system 302 is equipped with a magnet 304. The magnet 304 is a cylindrical superconducting magnet through which a bore 306 passes. Various different types of magnets can also be used; for example, both segmented cylindrical magnets and so-called open magnets can be used. A segmented cylindrical magnet is similar to a standard cylindrical magnet, except that the cryostat is divided into two sections to allow equal-plane access to the magnet; such magnets are used, for example, in conjunction with charged particle beam therapy. An open magnet has two magnet sections with a space large enough to accommodate a subject between the upper and lower magnet sections; the arrangement of these two sections is similar to that of a Helmholtz coil. Open magnets are common because the subject is not confined. Inside the cylindrical magnet cryostat is an assembly of superconducting coils.

[0066] Within the bore 306 of the cylindrical magnet 304 is an imaging zone 308 where the magnetic field is strong and uniform enough to perform magnetic resonance imaging. The acquired magnetic resonance data is typically obtained for a region of interest. The subject 318 is shown supported by a subject support 320 so that at least a portion of the subject 318 is within the imaging zone 308.

[0067] Within the magnet's bore 306 is also a set of magnetic gradient coils 310, which are used to acquire preliminary magnetic resonance data for spatial coding of magnetic spins within the imaging zone 308 of the magnet 304. The magnetic gradient coils 310 are connected to a magnetic gradient coil power supply 312. The magnetic gradient coils 310 are intended to be a typical example. Typically, the magnetic gradient coils 310 include three separate sets of coils for spatial coding in three orthogonal spatial directions. The magnetic gradient power supply provides current to the magnetic gradient coils. The current supplied to the magnetic gradient coils 310 is controlled as a function of time and may be ramp-like or pulse-like.

[0068] Adjacent to the imaging zone 308 is a radio frequency coil 314 that manipulates the orientation of magnetic spins within the imaging zone 308 and also receives radio transmissions from the spins within the imaging zone 308. A radio frequency antenna includes multiple coil elements. The radio frequency antenna is sometimes called a channel or antenna. The radio frequency coil 314 is connected to a radio frequency transceiver 316. The radio frequency coil 314 and radio frequency transceiver 316 may be replaced with separate transmitting and receiving coils and separate transmitters and receivers. It is understood that the radio frequency coil 314 and radio frequency transceiver 316 are typical examples. The radio frequency coil 314 is also intended to represent a dedicated transmitting antenna and a dedicated receiving antenna. Similarly, the transceiver 316 may also represent a separate transmitter and receiver. The radio frequency coil 314 may also have multiple receiving / transmitting elements, and the radio frequency transceiver 316 may have multiple receiving / transmitting channels. For example, when parallel imaging techniques such as SENSE are used, the radio frequency coil 314 has multiple coil elements.

[0069] The transceiver 316 and the gradient controller 312 are shown as being connected to the hardware interface 106 of the computer system 102.

[0070] Within the imaging zone 308, there is an anatomical structure 322 of the subject 318. It can be seen that there is an excitation field 324 that has been modified to fit precisely to the anatomical structure 322. The excitation field 324 is designated by selection 124 of at least one excitation field and encoded in a composite array 128.

[0071] Memory 110 is further shown to include k-space data 330 acquired by the magnetic resonance imaging system 302 by controlling the magnetic resonance imaging system 302 with a modified pulse sequence command 134. The k-space data 330 describes anatomical structures 322. Memory 110 is further shown to include magnetic resonance imaging data 332 reconstructed from the k-space data 330.

[0072] Figure 4 shows a flowchart illustrating how to operate the medical system 300 shown in Figure 3. The procedure begins with steps 200-208 as shown in Figure 2. After step 208, step 400 and then step 402 are performed. In step 400, k-space data 330 is acquired by controlling the magnetic resonance imaging system 302 with a modified pulse sequence command 134. Finally, in step 402, magnetic resonance imaging data 332 is reconstructed from the k-space data 330.

[0073] Figure 5 shows a flowchart illustrating how to train the convolutional neural network 122. Steps 500, 502, 504, and 506 are shown as being performed in a loop or repeated multiple times. However, these steps may be performed in parallel. That is, all of the training data may be obtained by performing step 504, then step 506 multiple times, then step 500 multiple times, then step 502 multiple times. First, in step 500, a training high-frequency waveform and multiple training spatially selective gradient pulse waveforms are generated using a selective excitation pulse design algorithm. Next, in step 502, the model excitation field is calculated by inputting the training high-frequency waveform and training spatially selective gradient pulse waveforms into a magnetic resonance imaging signal model.

[0074] Next, in step 504, the model excitation field is input to the convolutional neural network, receiving a forward-propagated high-frequency waveform and multiple forward-propagated training spatial selection gradient pulse waveforms. Then, in step 506, the convolutional neural network is trained by updating its parameters through backpropagation. Backpropagation is performed using the training high-frequency waveform and the forward-propagated high-frequency waveform. Backpropagation is also performed using multiple training spatial selection gradient pulse waveforms and multiple forward-propagated training spatial selection gradient pulse waveforms of the fitted pair.

[0075] Modern MRI systems typically apply pre-defined RF and associated gradients during clinical scans, with minimal adaptation to the specific requirements of each scan. An example is provided for a multi-task, multi-scale CNN method for real-time design of excitation RF pulses and associated gradient waveforms to achieve spatial two-dimensional selectivity. The CNN-designed RF and gradients were highly similar to their SLR-designed counterparts, with an NRMSE of 0.0075 ± 0.0038 for 400 test data points. Phantom imaging using predicted RF was also similar to the SLR-designed excitation. The algorithm can be executed within 500 ms on a commercially available workstation.

[0076] While there are case-specific requirements such as self-refocusing and adiabatic properties, the most common requirement for a two-dimensional excitation pulse is the excitation profile in both dimensions. Therefore, RF pulse design presents the inverse problem: given the desired spatial and frequency profiles, the small-angle approximation works for small excitation angles, and the Shinnar-Le Roux (SLR) algorithm has been shown to handle arbitrary flip angles from 0° to 180°. For modern MRI machines, conventional RF pulses and their associated gradients are typically designed at system release using one of these two methods, subject to hardware constraints including maximum allowable B1 field, gradient field intensity, and gradient field slew rate, and the resulting waveforms are stored in the system. For each particular scan, a portion of these waveforms are loaded into the spectrometer, either scaled in magnitude or extended over time.

[0077] While the SLR method is practical and considers conventional imaging sequences, it limits its range of applications and hinders any acquisition where highly specific signal excitation would be beneficial. For example, spatial two-dimensional excitation has been clinically employed in diffusion imaging of the spinal cord, breast, pancreas, prostate, and neck to limit the phase-encoded FOV for improved resolution and reduced geometric distortion. Beyond using the SLR algorithm to generate simple two-dimensional excitations, the ability to generate RF waveforms and gradient waveforms of entirely arbitrary shapes optimizes the balance between RF performance and duration. However, designing these arbitrary RF waveforms and gradient waveforms requires an iterative algorithm, which is time-consuming and hinders real-time waveform generation.

[0078] In the example, a neural network-based two-dimensional selective RF design is disclosed in which both the RF waveform and gradient waveform are generated by a multitasking convolutional neural network in less than one second. This method is performed at each specific scan for optimized RF performance under given imaging conditions, including hardware constraints (RF and gradient limits and number of transmit channels), system imperfections (eddy currents and B0 non-uniformity), imaging object-dependent issues (local magnetic field non-uniformity with respect to both B0 and B1), and formulation-dependent requirements (FOV of interest). The variable-speed selective excitation (VERSE) method is sometimes applied clinically, and its iterative design algorithm has recently been simplified, but it only addresses some of the problems mentioned above. What is presented here are experimental results of this comprehensive real-time RF design project, which demonstrates the usefulness of DL as a path for any optimized excitation pulse in real time.

[0079] Figure 6 shows an example topology or structure for a convolutional neural network 122. This neural network 122 receives an input array 128. There is a first layer 600 containing a convolutional kernel, ReLU, and max pooling. This first layer is then passed to a second layer containing another convolutional kernel ReLU and max pooling. This second layer is then passed to a third convolutional kernel or layer 604, which also has a ReLU and max pooling layer. This third layer is then passed to a flattening layer 606, which is then coupled to three separate fully connected layers 608, 608', and 608''. Layer 608 outputs a high-frequency waveform 130. Only the coupled layers 608, 608' output two spatially selected gradient pulse waveforms 132. The neural network 122 shown in Figure 6 is used for two-dimensional slicing. The basic structure can be reused and modified for three-dimensional excitation fields. For example, additional coupling layers may be added for a third spatially selective gradient pulse waveform, and the input may be modified to accept a composite 3D array. The exact structure of the various convolutional layers is not essential. However, the use of a separate coupling layer coupled to the flattening layer 606 showed better results than other types of neural network topologies.

[0080] An exemplary multitask CNN is constructed based on a convolutional framework (hereinafter referred to as CNN). As shown in Figure 6 above, the CNN consists of an input layer, three convolutional blocks 600, 602, and 604, a flattening layer 606, fully connected layers 608, 608', and 608'', and an output layer. The input layer takes the real and imaginary components of the desired excitation profile, and thus a three-dimensional matrix with a three-dimensional size of 2 (e.g., 101x101x2). The first convolutional block 600 contains 16 convolutional kernels and a rectifier linear unit (ReLU), followed by a max pooling step. The second convolutional block 602 and the third convolutional block 604 contain 32 and 64 convolutional kernels, respectively, followed by the same ReLU activation and max pooling, respectively. Next, a flattening layer 606 was configured to transform the data into a vector of 9261, which was then connected by a fully connected layer to yield advanced inference with the fitted dimension output. Network convergence was evaluated by minimum loss, which was measured as root mean square error (RMSE). Network parameters were tuned until the optimal network configuration was achieved.

[0081] Excitation profiles were generated by treating the RF, GRS, and GRF waveforms as three channels of the RF pulse, and a multi-task learning technique was employed in which a flattening layer that preserves shared features was connected in parallel with three fully connected layers. Furthermore, the convolutional kernel was tuned to multiply by sizes of 1x1, 3x3, 5x5, and 7x7, similar to the measures taken in the inception module. Convolution with multiple kernel sizes captured feature information at various scales compared to conventional convolutional operations, thereby avoiding the potential problem of overfitting when deeper networks are used.

[0082] The example includes one or more of the following characteristics: 1) A sufficiently large set of 2D RF pulses for training the CNN was generated using a conventional SLR algorithm. 2) To train the CNN, the excitation profile of each pulse was generated using Bloch simulation. 3) Custom convolutional neural network architecture with optimized layers and connections 4) The network was trained using the generated excitation profile as input and the corresponding RF pulse as output. 5) The desired excitation profile was input to the trained network to obtain the CNN design RF.

[0083] Figure 7 shows the effectiveness of the high-frequency waveform 130 and multiple spatially selective gradient pulse waveforms 132 provided by the convolutional neural network 122. In this figure, exemplary high-frequency waveform 130 and multiple spatially selective gradient pulse waveforms 132 are shown at the top. Using these waveforms 130, 132, an excitation profile 700 was provided in the center, which is very similar to the ground truth excitation field 702 shown in the box below.

[0084] Furthermore, the predicted RF can actually provide an excitation profile that is very similar to the desired FOV excitation profile. This was tested by phantom imaging with a clinical MR3.0T scanner.

[0085] The effectiveness of the convolutional neural network is further demonstrated using Figures 8 and 9. Figure 8 shows the desired excitation field of view 800. In Figure 9, 900 shows the excitation field of view using the convolutional neural network. On the right, box 902 shows the excitation field of view using the LSR-designed RF pulse. It can be seen that the convolutional neural network performs very well compared to the LSR-designed RF pulse.

[0086] The example includes one or more of the following characteristics: 1) CNN-based RF design for both RF waveforms and gradient waveforms. 2) Multitask techniques are used in conjunction with a shared layer to couple the RF waveform and the two gradients. 3) Using a multiscale kernel in a CNN, we captured details of various scales in the input FOV image domain. 4) Echo-planar trajectories were selected for RF generation of the training dataset. This technique may, in some cases, be detected experimentally by scanning and acquiring images.

[0087] It has applications for spatial two-dimensional or three-dimensional excitation in imaging of the brain, spinal cord, head and neck tumors, breast, pancreas, and prostate. This considers improved spatial resolution and reduced geometric distortion for diffusion-weighted imaging.

[0088] Although the present invention is illustrated and described in detail in the drawings and the above description, such illustrations and descriptions should be considered descriptive or illustrative and not limiting, and the present invention is not limited to the embodiments disclosed.

[0089] Other variations of the disclosed embodiments can be understood and achieved by those skilled in the art in carrying out the claimed invention, based on a review of the drawings, disclosures, and appended claims. In the claims, the words “equipment, includes, and have” do not exclude other elements or steps, and singular elements do not exclude plural elements. A single processor or other unit can perform the functions of several of the matters described in the claims. The mere fact that certain means are described in different dependent claims does not imply that combinations of these means cannot be used advantageously. Computer programs may be stored / distributed on suitable media such as optical storage media or solid-state media supplied with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunications systems. No reference numerals in the claims should be construed as limiting in scope. [Explanation of Symbols]

[0090] 100 Healthcare Systems 102 Computer 104 Computing Systems 106 Optional Hardware Interfaces 108 Optional User Interfaces 110 memory 120 Machine-Executable Instructions 122 Convolutional Neural Networks 124 Selection of at least one excitation field 126 Initial pulse sequence command 128 composite array 130 High-frequency waveforms 132 Multiple spatially selected gradient pulse waveforms 134 Modified pulse sequence commands 200 Receive at least one selection of excitation field. 202 Receive an initial pulse sequence command configured to control the magnetic resonance imaging system to acquire k-space data describing the subject. 204 Encode the composite array using at least one excitation field. 206 In response to inputting a composite array into a convolutional neural network, it receives high-frequency waveforms and multiple spatially selected gradient pulse waveforms. 208 A modified pulse sequence command is configured by modifying the initial pulse sequence command with a high-frequency waveform and multiple spatially selective gradient pulse waveforms so that the pulse sequence command controls the magnetic resonance imaging system to acquire k-space data from at least one excitation field of view. 300 healthcare systems 302 Magnetic Resonance Imaging System 304 Magnet 306 Magnet Bore 308 Imaging Zone 310 Magnetic field gradient coil 312 Magnetic Field Gradient Coil Power Supply 314 Radio frequency coil 316 Transceiver 318 Target individuals 320 Support stand for the person 322 Anatomical Structure 324 Excitation field 330k spatial data 332 Magnetic Resonance Imaging Data 400 k-space data is acquired by controlling the magnetic resonance imaging system with modified pulse sequence commands. Reconstruct magnetic resonance imaging data from 402 k spatial data. 500 Selective Excitation Pulse Design Algorithms are used to generate training high-frequency waveforms and multiple training spatially selective gradient pulse waveforms. 502 The model excitation field is calculated by inputting a training high-frequency waveform and multiple training spatially selective gradient pulse waveforms into the magnetic resonance imaging signal model. By inputting the 504 model excitation field into a convolutional neural network, the forward-propagated high-frequency waveform and multiple forward-propagated training spatially selected gradient pulse waveforms are received. 506 Update the parameters of the convolutional neural network by backpropagating using the training high-frequency waveform, the forward-propagated high-frequency waveform, and multiple training spatial selection gradient pulse waveforms and multiple forward-propagated training spatial selection gradient pulse waveforms of the fitted pair. Convolutional kernel with 600 ReLU layers and maximum pooling layer Convolutional kernel with 602 ReLU layers and maximum pooling layer Convolutional kernel with 604 ReLU layers and maximum pooling layer 606 Planarization layer 608 Fully connected layer 608' fully connected layer 608'' fully connected layer 700 Generated excitation field 702 Ground Truth Excitation Field 800 Desired excitation field Excitation field using 900 CNNs Excitation field using 902 LSR-designed RF pulses

Claims

1. A memory for storing machine-executable instructions and a convolutional neural network, wherein the convolutional neural network receives a composite array encoded using at least one excitation field as input, outputs a high-frequency waveform and a plurality of spatially selective gradient pulse waveforms, and the convolutional neural network is a multitask convolutional neural network having a first output unit for the high-frequency waveform and separate output units for each of the plurality of spatially selective gradient pulse waveforms, A medical system comprising a computing system, Upon execution of the aforementioned machine-executable instruction, the computing system Receive a selection of at least one excitation field, To control the magnetic resonance imaging system to acquire k-space data describing the subject, an initial pulse sequence command is received that specifies the high-frequency waveform and the plurality of spatially selected gradient pulse waveforms. The composite array is encoded using the aforementioned at least one excitation field, In response to inputting the composite array into the convolutional neural network, the high-frequency waveform and the plurality of spatially selected gradient pulse waveforms are received, and A medical system comprising a modified pulse sequence command, which is configured by modifying the high-frequency waveform and the plurality of spatially selective gradient pulse waveforms of the initial pulse sequence command, so that the pulse sequence command controls the magnetic resonance imaging system to acquire the k-space data from the at least one excitation field of view.

2. The aforementioned convolutional neural network is A step of generating a training high-frequency waveform and multiple training spatially selected gradient pulse waveforms using a selective excitation pulse design algorithm, The steps include: calculating a model excitation field by inputting the training high-frequency waveform and the plurality of training spatially selective gradient pulse waveforms into a magnetic resonance imaging signal model; The steps include inputting a composite array encoded using the model excitation field into the convolutional neural network to receive a forward-propagated high-frequency waveform and a plurality of forward-propagated training spatial selection gradient pulse waveforms, The medical system according to claim 1, which is trained by a computing system that repeatedly performs the step of updating the parameters of the convolutional neural network by performing backpropagation using the training high-frequency waveform, the forward-propagated high-frequency waveform, and the plurality of training spatially selective gradient pulse waveforms and the plurality of forward-propagated training spatially selective gradient pulse waveforms of a matching pair.

3. The medical system according to claim 2, wherein the selected excitation pulse design algorithm is one of the Shinnar-Le Roux algorithm and the small flip angle approximation algorithm.

4. The medical system according to any one of claims 1 to 3, wherein the pulse sequence command acquires the k-space data in accordance with a parallel imaging magnetic resonance imaging protocol.

5. The execution of the aforementioned machine-executable instructions further enables the computing system to The investigation received magnetic resonance images. The aforementioned survey magnetic resonance image is rendered on a display, A medical system according to any one of claims 1 to 4, which receives the selection of the at least one excitation field within the investigational magnetic resonance image in response to displaying the investigational magnetic resonance image.

6. The medical system according to any one of claims 1 to 5, wherein the selection of the at least one excitation field is received from any one of an automated image segmentation algorithm, a user interface, and a combination thereof.

7. The medical system according to any one of claims 1 to 6, wherein the plurality of spatially selective gradient pulse waveforms are two spatially selective gradient pulse waveforms, and each of the at least one excitation field is a two-dimensional excitation field.

8. The medical system according to any one of claims 1 to 6, wherein the plurality of spatially selective gradient pulse waveforms are three partially selective gradient pulse waveforms, and each of at least one excitation field is a three-dimensional excitation field.

9. The medical system according to any one of claims 1 to 8, wherein each of the at least one excitation field has an arbitrary shape and / or is spatially separated.

10. The medical system according to any one of claims 1 to 9, wherein the convolutional neural network is one of a multiscale convolutional neural network and a U-net convolutional neural network.

11. The medical system includes the magnetic resonance imaging system, and the computing system further includes the execution of the machine-executable instructions. The k-space data is acquired by controlling the magnetic resonance imaging system with the modified pulse sequence command. A medical system according to any one of claims 1 to 10, wherein magnetic resonance imaging data is reconstructed from the k-space data.

12. A computer program comprising machine-executable instructions for execution by a computing system controlling a medical system, the computer program further comprising a convolutional neural network that receives a composite array encoded using at least one excitation field as input, the convolutional neural network outputs a high-frequency waveform and a plurality of spatially selective gradient pulse waveforms, and the convolutional neural network is a multitask convolutional neural network having a first output unit for the high-frequency waveform and separate output units for each of the plurality of spatially selective gradient pulse waveforms, By executing the aforementioned machine-executable instructions, the computing system Upon receiving the selection of at least one excitation field, To control the magnetic resonance imaging system to acquire k-space data describing the subject, an initial pulse sequence command is received that specifies the high-frequency waveform and the plurality of spatially selected gradient pulse waveforms. The composite array is encoded using the aforementioned at least one excitation field, In response to inputting the composite array into the convolutional neural network, the high-frequency waveform and the plurality of spatially selected gradient pulse waveforms are received. A computer program that configures a modified pulse sequence command by modifying the high-frequency waveform of the initial pulse sequence command and the plurality of spatially selective gradient pulse waveforms so that the pulse sequence command controls the magnetic resonance imaging system to acquire the k-space data from the at least one excitation field of view.

13. A medical imaging method using a convolutional neural network, wherein the convolutional neural network receives a composite array encoded using at least one excitation field as input, outputs a high-frequency waveform and a plurality of spatially selective gradient pulse waveforms, and the convolutional neural network is a multitask convolutional neural network having a first output unit for the high-frequency waveform and separate output units for each of the plurality of spatially selective gradient pulse waveforms. The aforementioned method, The steps include receiving the selection of the at least one excitation field, The steps include receiving an initial pulse sequence command that specifies the high-frequency waveform and the plurality of spatially selective gradient pulse waveforms in order to control the magnetic resonance imaging system to acquire k-space data describing the subject, The steps include encoding the composite array using the at least one excitation field, The steps include receiving the high-frequency waveform and the plurality of spatially selected gradient pulse waveforms in response to inputting the composite array into the convolutional neural network, A method comprising the steps of configuring a modified pulse sequence command by modifying the high-frequency waveform and the plurality of spatially selective gradient pulse waveforms of the initial pulse sequence command so that the pulse sequence command controls the magnetic resonance imaging system to acquire the k-space data from the at least one excitation field of view.

14. A method for training a convolutional neural network, wherein the method is A step of generating a training high-frequency waveform and multiple training spatially selected gradient pulse waveforms using a selective excitation pulse design algorithm, The steps include: calculating a model excitation field by inputting the training high-frequency waveform and the plurality of training spatially selective gradient pulse waveforms into a magnetic resonance imaging signal model; The steps include inputting the model excitation field into the convolutional neural network to receive a forward-propagated high-frequency waveform and a plurality of forward-propagated training spatial selection gradient pulse waveforms, A method comprising repeatedly performing the step of updating the parameters of the convolutional neural network by performing backpropagation using the training high-frequency waveform, the forward-propagated high-frequency waveform, and a plurality of the training spatial selection gradient pulse waveforms and the plurality of forward-propagated training spatial selection gradient pulse waveforms of a fitted pair.

Images