阅读说明：本技术 借助于人工神经网络的功能性磁共振成像伪影移除 (Functional magnetic resonance imaging artifact removal by means of artificial neural networks ) 是由 A·埃瓦尔德 N·弗莱施纳 G·伯恩哈德 I·格雷斯林 P·博尔纳特 I·施马勒 J·A 于 2019-11-26 设计创作，主要内容包括：本发明提供了一种医学成像系统(100、400),其包括存储器(110),所述存储器存储机器可执行指令(120)和经配置的人工神经网络(122)。所述医学成像系统还包括处理器(104),所述处理器被配置用于控制所述医学成像系统。对所述机器可执行指令的执行使所述处理器接收(200)磁共振成像数据(124),其中,所述磁共振成像数据是描述针对体素的集合中的每个体素的时间相关性BOLD信号(1100)的BOLD功能性磁共振成像数据。对所述机器可执行指令的执行还使所述处理器通过使用所述磁共振成像数据重建针对体素的集合中的每个体素的所述时间相关性BOLD信号来构建(202)初始信号的集合(126)。对所述机器可执行指令的执行还使所述处理器响应于将初始信号的集合输入到所述经配置的人工神经网络中而接收(204)经修改的信号的集合(128)。所述经配置的人工神经网络被配置用于从初始信号的集合移除生理伪影。(A medical imaging system (100, 400) is provided that includes a memory (110) storing machine executable instructions (120) and a configured artificial neural network (122). The medical imaging system further comprises a processor (104) configured for controlling the medical imaging system. Execution of the machine executable instructions causes the processor to receive (200) magnetic resonance imaging data (124), wherein the magnetic resonance imaging data is BOLD functional magnetic resonance imaging data that describes a time-dependent BOLD signal (1100) for each voxel in a set of voxels. Execution of the machine-executable instructions further causes the processor to construct (202) a set of initial signals (126) by reconstructing the time-dependent BOLD signal for each voxel in the set of voxels using the magnetic resonance imaging data. Execution of the machine-executable instructions further causes the processor to receive (204) a set of modified signals (128) in response to inputting a set of initial signals into the configured artificial neural network. The configured artificial neural network is configured to remove physiological artifacts from a set of initial signals.)

1. A medical imaging system (100, 400) comprising:

-a memory (110) storing machine executable instructions (120) and a configured artificial neural network (122);

a processor (104) configured for controlling the medical imaging system by executing the machine executable instructions, wherein execution of the machine executable instructions causes the processor to:

-receiving (200) magnetic resonance imaging data (124), wherein the magnetic resonance imaging data is BOLD functional magnetic resonance imaging data describing a time-dependent BOLD signal (1100) for each voxel in a set of voxels, wherein the magnetic resonance imaging data is BOLD functional magnetic resonance imaging data describing a time-dependent BOLD signal for each voxel in a set of voxels, wherein the configured artificial neural network comprises a recurrent neural network (600) for each voxel in the set of voxels;

-construct (202) a set (126) of initial signals by reconstructing the time-dependent BOLD signal for each voxel in the set of voxels using the magnetic resonance imaging data; and is

-receiving (204) a set of modified signals (128) from the configured artificial neural network in response to inputting the set of initial signals into the configured artificial neural network, wherein the configured artificial neural network is configured for removing physiological artifacts from the set of initial signals.

2. The medical imaging system of claim 1, wherein the medical imaging system further comprises a magnetic resonance imaging system (402) configured for acquiring the magnetic resonance imaging data according to a time-dependent functional magnetic resonance imaging protocol, wherein execution of the machine executable instructions further causes the processor to control (500) the magnetic resonance imaging system to acquire the magnetic resonance imaging data.

3. The medical imaging system of claim 1 or 2, wherein the neural network comprises a plurality of recurrent neural networks, and wherein the recurrent neural network of each voxel in the set of voxels forms the plurality of recurrent neural networks.

4. The medical imaging system of any one of the preceding claims, wherein the configured artificial neural network comprises an input computation layer (700), wherein the recurrent neural network of each voxel comprises an input (602) connected to the input computation layer.

5. The medical imaging system of claim 4, wherein the input computational layer is any one of:

-fully connected;

-connected by convolution; and

6. the medical imaging system of any one of the preceding claims, wherein the configured artificial neural network comprises an output computation layer (800), wherein the recurrent neural network of each voxel comprises an output (604) connected to the output computation layer.

7. The medical imaging system of claim 6, wherein the output computational layer is any one of:

-fully connected;

-connected by convolution; and

-partially connected.

8. The medical imaging system of any one of the preceding claims, wherein the recurrent neural network of each voxel comprises direct feedback (606), wherein the configured artificial neural network comprises a hidden feedback layer (900), wherein the direct feedback of the recurrent neural network of each voxel is provided via the hidden feedback layer.

9. The medical imaging system of any one of the preceding claims, wherein execution of the machine executable instructions further causes the processor to:

-receiving (300) a set of training signals for each voxel in the set of voxels;

-receiving (302) a set of clean signals for each voxel in the set of voxels, wherein the set of training signals comprises the set of clean signals plus physiological artifacts;

-receiving (304) physiological artifact data descriptive of subject motion, wherein the physiological artifact data is temporally correlated with the set of cleaning signals;

-training (306) the configured artificial neural network using the set of training signals, the cleaning signals and the physiological artifact data.

10. The medical imaging system of any one of the preceding claims, wherein execution of the machine executable instructions further causes the processor to:

-receiving a noise reduction value; and is

-using the set of initial signals, the set of modified signals and the noise reduction value to construct a set of controllably cleanable signals.

11. The medical imaging system of claim 10, wherein any one of the following is performed:

-wherein, if the set of modified signals is a set of noise signals, constructing the set of controllably cleaned signals comprises subtracting from the set of initial signals a multiple of the noise reduction value multiplied by the set of modified signals; and

-wherein, if the set of modified noise signals is a set of cleaned signals, constructing the set of controllably cleaned signals comprises:

-calculating a set of noise signals by subtracting the set of modified signals from the set of initial signals; and is

-constructing the set of controllably cleanable signals by subtracting from the set of initial signals the set of noise signals multiplied by a multiple of the noise reduction value.

12. The medical imaging system of claim 10 or 11, wherein execution of the machine executable instructions further causes the processor to reconstruct a functional magnetic resonance image (130) from the set of controllably cleaned signals.

13. The medical imaging system of any one of the preceding claims, wherein the recurrent neural network for each voxel in the set of voxels is a long-short term memory neural network (600').

14. A method of operating a medical imaging system, wherein the method comprises:

-receiving (200) magnetic resonance imaging data (124), wherein the magnetic resonance imaging data is BOLD functional magnetic resonance imaging data describing a time-dependent BOLD signal for each voxel of a set of voxels;

-construct (202) a set (126) of initial signals by reconstructing the time-dependent BOLD signal for each voxel in the set of voxels using the magnetic resonance imaging data; and is

-receiving (204) a set of modified signals (128) from a configured artificial neural network (122) in response to inputting the set of initial signals into the configured artificial neural network, wherein the configured artificial neural network is configured for removing physiological artifacts from the set of initial signals, wherein the configured artificial neural network comprises a recurrent neural network (600) for each voxel in the set of voxels.

15. A computer program product comprising machine executable instructions (120) for execution by a processor (104) of an embodiment of a configured artificial neural network (122), wherein execution of the machine executable instructions causes the processor to:

-receiving (200) magnetic resonance imaging data (124), wherein the magnetic resonance imaging data is BOLD functional magnetic resonance imaging data describing a time-dependent BOLD signal for each voxel of a set of voxels;

-construct (202) a set (126) of initial signals by reconstructing the time-dependent BOLD signal for each voxel in the set of voxels using the magnetic resonance imaging data; and is

-receiving (204) a set of modified signals (128) from the configured artificial neural network in response to inputting the set of initial signals into the configured artificial neural network, wherein the configured artificial neural network is configured for removing physiological artifacts from the set of initial signals, wherein the configured artificial neural network comprises a recurrent neural network (600) for each voxel in the set of voxels.

Technical Field

The present invention relates to magnetic resonance imaging, in particular to functional magnetic resonance imaging.

Background

As part of the procedure for generating images within the body of a subject, large static magnetic fields are used by Magnetic Resonance Imaging (MRI) scanners to align the nuclear spins of atoms. This large static magnetic field is referred to as the B0 field or main magnetic field. Various quantities or properties of the object may be spatially measured using MRI. Various imaging protocols may be implemented using pulse sequences to control the acquisition of magnetic resonance data and may be used to measure various properties of the subject.

For example, in functional magnetic resonance imaging (fMRI), magnetic resonance imaging is used to measure brain activity. A common type of functional magnetic resonance imaging is Blood Oxygen Level Dependent (BOLD) contrast. BOLD imaging relies on the properties of oxygenated and deoxygenated hemoglobin. Oxyhemoglobin is paramagnetic, while deoxyhemoglobin is diamagnetic. Thus, the T2-weighted pulse sequence can detect changes in the oxygenation of blood in the brain. However, the BOLD effect is small and the contrast of the T2 x weighted image changes by only a few percent. Patient motion caused by respiratory or cardiac motion can be, at a minimum, a change in contrast, also on the order of a few percent. In order to perform BOLD magnetic resonance imaging, this physiological noise needs to be removed from the signal.

A journal article by Anderson et al, "Common component classification: What can be seen from a machine learning? "(NeuroImage, volume 56, No. 2, page 517-524 (2011, 5 months and 15 days)) discloses that machine learning methods have been applied to classify fMRI scans by studying the location of temporal intensity variations between presentation groups in the brain, often reporting classification accuracies of 90% or better. Machine learning classifiers are created and then used to deconstruct the classifiers to check their sensitivity to physiological noise, task reordering, and cross-scan classification capabilities. The model is trained and tested both intra-and cross-run to assess stability and reproducibility across conditions. The use of independent component analysis for both feature extraction and artifact removal is illustrated.

Disclosure of Invention

The invention provides a medical imaging system, a computer program product and a method.

As mentioned above, functional magnetic resonance imaging signals can be easily blurred by physiological artifacts. A typical way to remove this noise from functional magnetic resonance imaging data is to monitor physiological processes of the subject, such as respiration and cardiac motion. This enables removal of physiological artifacts. Embodiments may simplify this process by automatically removing physiological artifacts using a configured artificial neural network. The physiological neural network can be trained to remove physiological artifacts without the need for sensors to monitor respiratory or cardiac motion.

In one aspect, the invention provides a medical imaging system comprising a memory storing machine executable instructions. The medical imaging system further comprises a processor configured to control the medical imaging system by executing the machine executable instructions. Execution of the machine executable instructions causes the processor to receive magnetic resonance imaging data. The magnetic resonance imaging data is functional magnetic resonance imaging data describing a time-dependent signal for each voxel in a set of voxels. For example, the magnetic resonance imaging data comprises data individually describing each voxel and the data contains a time-dependent signal. Execution of the machine-executable instructions further cause the processor to construct a set of initial signals by reconstructing the time-dependent signal for each voxel in the set of voxels using the magnetic resonance imaging data.

The set of initial signals has a temporal correlation signal for each voxel in the set of voxels. Execution of the machine-executable instructions further causes the processor to receive a set of modified signals in response to inputting the set of initial signals into a configured artificial neural network. The configured artificial neural network is configured to remove physiological artifacts from the set of initial signals. This embodiment may be beneficial because functional magnetic resonance imaging is particularly susceptible to physiological artifacts, which are artifacts caused by physiological changes or motion of the subject.

It should be noted that the set of modified signals may be pure noise signals from the set of initial signals, or it may be noise-removed signals. Both of which are substantially equivalent.

In another embodiment, the medical imaging system further comprises a magnetic resonance imaging system configured for acquiring the magnetic resonance imaging data according to a time-dependent functional magnetic resonance imaging protocol. Execution of the machine executable instructions further causes the processor to control the magnetic resonance imaging system to acquire the magnetic resonance imaging data.

In another embodiment, the configured artificial neural network comprises a recurrent neural network for each voxel in the set of voxels. The use of a recurrent neural network for each voxel may be beneficial because the signal of each voxel in the set of voxels is time dependent. The use of a recurrent neural network enables the signal to be processed in the time domain.

The recurrent neural network for each of the voxels may have several or more layers. These multiple layers may include different types, such as partial connection layers and convolutional layers.

In another embodiment, the configured artificial neural network includes an input computational layer. The recurrent neural network of each of the voxels includes an input connected to the input computation layer. In this embodiment, rather than inputting data directly into the individual recurrent neural networks, the inputs are first input into the input computation layer and then the inputs of the recurrent neural networks are connected thereto. This may be beneficial because the effects of collaborative behavior such as blood flow and movement of other voxels, and others, may be better modeled using the input computational layer.

In another embodiment, the set of modified signals is one of a set of noise signals and a set of clean signals. The set of initial signals is a measured time-dependent BOLD signal for each of the voxels. For each voxel there is a time dependent BOLD signal. The set of initial signals also has noise due to physiological artifacts. The set of noise signals is a noise component of the set of initial signals, and the set of cleaned signals is a set of time-dependent signals without a noise component. The neural network is configured to remove physiological artifacts or noise. Having a configured neural network that produces a set of cleaned signals is equivalent to a configured neural network that produces a set of noise signals.

In another embodiment, the neural network comprises a plurality of recurrent neural networks, and wherein the recurrent neural network of each voxel in the set of voxels forms the plurality of recurrent neural networks. For each voxel in the set of voxels, there is an individual or separate recurrent neural network.

In another embodiment, the use of partially connected layers may be used to specify the effect of voxels within a particular distance or radius of a particular voxel. This may be useful when representing local effects.

In another embodiment, the configured artificial neural network includes an output computational layer. The recurrent neural network for each voxel includes an output connected to the output computation layer. This embodiment may also be beneficial because the output of neighboring voxels may be taken into account when providing the output. This may provide more efficient or better removal of noise signals.

In another embodiment, the output computation layer is fully connected.

In another embodiment, the output layers are convolutionally connected.

In another embodiment, the output computation layers are partially connected.

In another embodiment, the recurrent neural network of each voxel comprises a direct feedback or a direct feedback loop. The direct feedback may be individual signals fed back into the recurrent neural network as feedback for the next iteration or via feedback provided by a hidden layer within the recurrent neural network. The configured artificial neural network includes a hidden feedback layer. The direct feedback of the recurrent neural network for each voxel is provided via the hidden feedback layer.

In this embodiment, instead of the direct feedback of the recurrent neural network being fed directly to itself, the direct feedback of the recurrent neural network passes through an intermediate layer that is capable of combining and/or distributing the feedback from multiple recurrent neural networks. This may help to model the effect of affecting more than one voxel.

In another embodiment, the hidden feedback layer comprises at least one fully connected layer.

The hidden feedback layer includes at least one convolutional layer.

In another embodiment, execution of the machine-executable instructions further causes the processor to receive a set of training signals for each voxel in the set of voxels. Execution of the machine-executable instructions further causes the processor to receive a set of cleaning signals for each voxel in the set of voxels. The set of training signals includes the set of clean signals plus physiological artifacts. Execution of the machine-executable instructions further causes the processor to receive physiological artifact data describing motion of a subject. The physiological artifact data is movement data. The movement data is time-dependent on the set of cleaning signals.

Execution of the machine-executable instructions further cause the processor to train the configured artificial neural network using the set of training signals, the cleaning signals, and the physiological artifact data. These training steps may be beneficial because they may be able to train a configured artificial neural network to remove physiological artifact data in the absence of physiological artifact data. For example, if a respiratory or heart rate sensor provides physiological artifact data, the resulting configured artificial neural network may be able to remove this noise from the set of initial signals in the absence of physiological artifact data or sensor data.

In another embodiment, execution of the machine-executable instructions further causes the processor to receive a noise reduction value. Execution of the machine-executable instructions further causes the processor to construct a set of controllably cleanable signals using the set of initial signals, the modified signals, and the noise reduction values. As mentioned previously, the set of modified signals may contain clean signals or signals with noise. Either way, using it in conjunction with the set of initial signals, the amount of noise removed can be dynamically or controllably adjusted. For example, if the set of modified signals contains only noise signals, a set of controllably cleanable signals may be constructed by multiplying the noise reduction value by the set of modified signals and then subtracting it from the set of initial signals. Conversely, if the set of modified signals are signals that have been cleaned of noise, a similar operation may be performed with an equivalent result.

In another embodiment, if the set of modified signals is a set of noise signals, constructing the set of controllably cleanable signals comprises subtracting a multiple of the noise reduction value multiplied by the set of modified signals from the set of initial signals.

In another embodiment, if the set of modified signals is a set of cleaned signals, constructing the set of controllably cleaned signals comprises: calculating a set of noise signals by subtracting the set of modified signals from the set of initial signals; and constructing the set of controllably cleanable signals by subtracting from the set of initial signals the set of noise signals multiplied by a multiple of the noise reduction value.

In another embodiment execution of the machine executable instructions further causes the processor to reconstruct a BOLD magnetic resonance image from the set of modified signals. If the set of modified signals is a set of cleaned signals, this may be performed directly using the set of modified signals. If the set of modified signals is a set of noise signals, a set of cleaned signals may be calculated by subtracting the set of modified signals from the set of modified signals.

In another embodiment execution of the machine executable instructions further causes the processor to reconstruct a functional magnetic resonance image from the set of controllably cleaned signals. This may be beneficial as it may enable the operator to adjust the optimal signal in the resulting functional magnetic resonance image.

In another embodiment execution of the machine executable instructions further causes the processor to reconstruct a functional magnetic resonance image from the set of modified signals. For example, if the set of modified signals includes signals that have all noise removed by the configured artificial neural network.

In another embodiment, the recurrent neural network for each voxel in the set of voxels is a long-short term memory neural network. The use of long-short term memory neural networks may be beneficial because it may provide better removal of temporally related artifacts.

In another embodiment, the magnetic resonance imaging data is BOLD functional magnetic resonance imaging data describing a time-dependent BOLD signal for each voxel of the set of voxels. This may be particularly beneficial as BOLD functional magnetic resonance imaging is particularly susceptible to physiological artifacts. For example, movement of the subject due to breathing or the heart may result in distortion of the same order of magnitude as the BOLD signal. The BOLD signal is typically a few percent of the signal. This is also of the same order of magnitude as the noise caused by physiological movement or artifacts.

In another aspect, the invention provides a method of operating a medical imaging system. The method includes receiving magnetic resonance imaging data. The magnetic resonance imaging data is functional magnetic resonance imaging data describing a time-dependent signal for each voxel in a set of voxels. The method further comprises constructing a set of initial signals by reconstructing a time-dependent signal for each voxel in the set of voxels using the magnetic resonance imaging data. The method also includes receiving a set of modified signals in response to inputting the set of initial signals into a configured artificial neural network. The configured artificial neural network is configured to remove physiological artifacts from the set of initial signals.

In another aspect, the invention provides a method of training a configured artificial neural network. The method first includes receiving a set of training signals for each voxel in the set of voxels. The method also includes receiving a set of cleaning signals for each voxel in the set of voxels. The set of training signals includes the set of clean signals plus physiological artifacts. The method further includes receiving physiological artifact data describing motion of the subject. The physiological artifact data is movement data. The movement data is time-dependent on the set of cleaning signals. The method also includes training the configured artificial neural network using the set of training signals, the cleaning signal, and the physiological artifact data.

In another embodiment, the set of training signals comprises sequential data points. Execution of the machine-executable instructions causes the processor to train the configured artificial neural network such that the set of cleaning signals are sequentially offset by a predetermined number of iterations. This training causes the recurrent neural network to delay outputting the set of modified signals for a predetermined number of iterations. This may have the following advantages: the recurrent neural network uses several samples of the set of initial signals to produce a set of modified signals.

In another aspect, the invention provides a computer program product comprising machine executable instructions for execution by a processor. Execution of the machine executable instructions causes the processor to receive magnetic resonance imaging data. The magnetic resonance imaging data is BOLD functional magnetic resonance imaging data describing a time-dependent BOLD signal for each voxel in the set of voxels. Execution of the machine-executable instructions further cause the processor to construct a set of initial signals by reconstructing the time-dependent BOLD signal for each voxel in the set of voxels using the magnetic resonance imaging data. The method also includes receiving a set of modified signals in response to inputting the set of initial signals into the configured artificial neural network. The configured artificial neural network is configured to remove physiological artifacts from the set of initial signals.

It should be understood that one or more of the foregoing embodiments of the invention may be combined, as long as the combined embodiments are not mutually exclusive.

As will be appreciated by one skilled in the art, aspects of the present invention may be embodied as an apparatus, method or computer program product. Accordingly, aspects of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software and hardware aspects that may all generally be referred to herein as a "circuit," module "or" system. Furthermore, various aspects of the invention may take the form of a computer program product embodied in one or more computer-readable media having computer-executable code embodied thereon.

Any combination of one or more computer-readable media may be used. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. A "computer-readable storage medium," as used herein, includes any tangible storage medium that can store instructions that are executable by a processor of a computing device. The computer-readable storage medium may be referred to as a "computer-readable non-transitory storage medium". The computer readable storage medium may also be referred to as a tangible computer readable medium. In some embodiments, the computer-readable storage medium may also be capable of storing data that is accessible by a processor of the computing device. Examples of computer-readable storage media include, but are not limited to: a floppy disk, a magnetic hard drive, a solid state drive, flash memory, a USB thumb drive, Random Access Memory (RAM), Read Only Memory (ROM), an optical disk, a magneto-optical disk, and a register file for a processor. Examples of optical disks include Compact Disks (CDs) and Digital Versatile Disks (DVDs), such as CD-ROMs, CD-RWs, CD-R, DVD-ROMs, DVD-RWs, or DVD-R disks. The term computer-readable storage medium also refers to various types of recording media that can be accessed by the computer device via a network or a communication link. For example, the data may be retrieved via a modem, via the internet, or via a local area network. Computer executable code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

A computer readable signal medium may include a propagated data signal with computer executable code embodied therein, for example, within a baseband or as part of a carrier wave. Such a propagated signal may take any of a variety of forms, including, but not limited to, electro-magnetic, 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 and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

"computer memory" or "memory" is an example of computer-readable storage media. Computer memory is any memory that can be directly accessed by a processor. A "computer storage device" or "storage device" is another example of a computer-readable storage medium. The computer storage may be any volatile or non-volatile computer-readable storage medium.

"processor" as used herein encompasses an electronic component capable of executing a program or machine-executable instructions or computer-executable code. References to a computing device that includes a "processor" should be interpreted as being capable of including more than one processor or processing core. The processor may be, for example, a multi-core processor. A processor may also refer to a collection of processors within a single computer system or distributed among multiple computer systems. The term computing device should also be construed to possibly refer to a collection or network of computing devices, each of which includes a processor or multiple processors. The computer executable code may be executed by multiple processors, which may be within the same computing device or which may even be distributed across multiple computing devices.

The computer executable code may include machine executable instructions or programs that cause the processor to perform various aspects of the present invention. Computer executable code for performing operations for aspects of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C + + or the like and conventional procedural programming languages, such as the C programming language or similar programming languages, and compiled as machine executable instructions. In some cases, the computer executable code may be used in a high level language or in a pre-compiled form and in conjunction with an interpreter that generates machine executable instructions in flight.

The computer executable code may run entirely on the user's computer, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server as a stand-alone software package. 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 a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider).

Aspects of the present invention are described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block or portion of a block of the flowchart, illustrations, and/or block diagrams, can be implemented by computer program instructions in the form of computer-executable code, where applicable. It will also be understood that combinations of blocks in the various flow diagrams, illustrations, and/or block diagrams can be combined, when not mutually exclusive. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable medium that can direct a computer, other programmable data processing apparatus, or other devices to function in a particular manner, such that the instructions stored in the computer-readable medium produce an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.

As used herein, a "user interface" is an interface that allows a user or operator to interact with a computer or computer system. The "user interface" may also be referred to as a "human interface device". The user interface may provide information or data to and/or receive information or data from an operator. The user interface may enable input from an operator to be received by the computer and may provide output to the user from the computer. In other words, the user interface may allow an operator to control or manipulate the computer, and the interface may allow the computer to indicate the effect of the operator's control or manipulation. The display of data or information on a display or graphical user interface is an example of providing information to an operator. Receiving data via a keyboard, mouse, trackball, touchpad, pointing stick, tablet, joystick, game pad, web camera, helmet, pedal, wired glove, remote control, and accelerometer are examples of user interface components that enable receiving information or data from an operator.

As used herein, "hardware interface" encompasses an interface that enables a processor of a computer system to interact with or control external computing devices and/or appliances. The hardware interface may allow the processor to send control signals or instructions to an external computing device and/or apparatus. The hardware interface may also enable the processor to exchange data with external computing devices and/or apparatus. Examples of hardware interfaces include, but are not limited to: a universal serial bus, an IEEE1394 port, a parallel port, an IEEE1284 port, a serial port, an RS-232 port, an IEEE-488 port, a Bluetooth connection, a wireless local area network connection, a TCP/IP connection, an Ethernet connection, a control voltage interface, a MIDI interface, an analog input interface, and a digital input interface.

"display" or "display device" as used herein encompasses an output device or user interface suitable for displaying images or data. The display may output visual, audio, and tactile data. Examples of displays include, but are not limited to: computer monitors, television screens, touch screens, tactile electronic displays, braille screens, Cathode Ray Tubes (CRTs), memory tubes, bi-stable 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 displays (OLEDs), projectors, and head mounted displays.

Magnetic resonance imaging data is defined herein as the recorded measurements of radio frequency signals emitted by atomic spins using an antenna of a magnetic resonance apparatus during a magnetic resonance imaging scan. A Magnetic Resonance Imaging (MRI) image or MR image is defined herein as a reconstructed two-dimensional or three-dimensional visualization of anatomical data contained within the magnetic resonance imaging data. Such visualization can be performed using a computer.

Drawings

Preferred embodiments of the present invention will be described hereinafter, by way of example only, and with reference to the accompanying drawings, in which:

fig. 1 illustrates an example of a medical imaging system;

FIG. 2 shows a flow chart illustrating a method of operating the medical imaging system of FIG. 1;

FIG. 3 shows a flow diagram illustrating a method of training a configured artificial neural network;

FIG. 4 illustrates yet another example of a medical imaging system;

FIG. 5 shows a flow chart illustrating a method of operating the medical imaging system of FIG. 4;

FIG. 6 illustrates an example of a configured artificial neural network;

FIG. 7 illustrates yet another example of a configured artificial neural network;

FIG. 8 illustrates yet another example of a configured artificial neural network;

FIG. 9 illustrates yet another example of a configured artificial neural network;

figure 10 illustrates an example of a time-dependent functional magnetic resonance imaging signal with noise;

figure 11 illustrates an idealized time-dependent functional magnetic resonance imaging signal without noise;

FIG. 12 illustrates an example of an unrolled recurrent neural network;

FIG. 13 illustrates an example of an expanded LSTM neural network; and is

FIG. 14 illustrates yet another example of a configured artificial neural network.

List of reference numerals

100 medical imaging system

102 computer

104 processor

106 hardware interface

108 user interface

110 memory

120 machine executable instructions

122 configured artificial neural network

124 magnetic resonance imaging data

126 set of initial signals

128 set of modified signals

130 functional magnetic resonance image

200 receive magnetic resonance imaging data

202 construct a set of initial signals by reconstructing a time-dependent signal for each of a set of voxels using magnetic resonance imaging data

204 receive the set of modified signals in response to inputting the set of initial signals into a configured artificial neural network, wherein the configured artificial neural network is configured to remove physiological artifacts from the set of initial signals

300 receive a set of training signals for each of a set of voxels

302 receives a set of cleaning signals for each of a set of voxels

304 receive physiological artifact data describing motion of a subject

306 train the configured artificial neural network using the set of training signals, the cleaning signal, and the physiological artifact data

400 medical imaging system

402 magnetic resonance imaging system

404 magnet

406 bore of magnet

408 imaging zone

409 region of interest

410 magnetic field gradient coil

412 magnetic field gradient coil power supply

414 radio frequency coil

416 transceiver

418 object

420 object support

430 pulse sequence commands

500 control a magnetic resonance imaging system to acquire magnetic resonance imaging data

600 recurrent neural network

600 LSTM

602 input

604 output

606 direct feedback or hidden layer neurons

700 input computing layer

702 input of layer input

800 output computing layer

802 output of the computing layer

900 connected hidden layer

BOLD signal of 1000 miscellaneous artifacts

1002 time

1004 BOLD response

1100 idealized artifact-free bold signal

1102 initial descent

1104 Peak

1106 post-stimulation undershoot

Detailed Description

In which like numbered elements are either equivalent elements or perform the same function. Elements that have been previously discussed will not necessarily be discussed in later figures if they are functionally equivalent.

Fig. 1 illustrates an example of a medical imaging system 100. The medical imaging system 100 includes a computer 102. The computer includes a processor 104. Processor 104 is intended to represent one or more processors or processing cores, and may also be distributed among multiple computers. The processor is shown connected to an optional hardware interface 106, an optional user interface 108, and a memory 110. The hardware interface 106 may enable the processor 104 to send and receive signals or information to and control other components of the medical imaging system 100.

The hardware interface 106 may also be a network interface and may enable the processor 104 to exchange data and/or instructions with other computer systems. The user interface 108 may enable display of data and/or rendering of images. The user interface 108 may also be used by an operator to control the operation and functions of the medical imaging system. The memory 110 is any memory accessible to the processor 104. The memory 110 may include volatile and non-volatile memory. The entries in memory 110 may be copied or duplicated within multiple modalities of memory, such as in main memory and stored on a hard disk drive or other computer storage medium.

The memory 110 is shown as containing machine-executable instructions 120. Execution of the machine-executable instructions 120 by the processor 104 enables the processor 104 to control various components of the medical imaging system. Execution of the machine-executable instructions may also enable the processor 104 to perform various data and numerical calculations and data processing.

The memory 110 is also shown as containing a configured artificial neural network 122. The configured artificial neural network is configured for removing physiological artifacts from the time-dependent functional magnetic resonance imaging signals. An example would be to remove noise from the so-called BOLD magnetic resonance imaging signal. The memory 110 is also shown as containing magnetic resonance imaging data 124. The magnetic resonance imaging data 124 is functional magnetic resonance imaging data that describes a time-dependent signal for each of a set of voxels. The memory 110 is also shown as containing a set 126 of initial signals reconstructed from the magnetic resonance imaging data 124. Each of the set of signals 126 is a time-dependent signal and there is a signal for each of the set of voxels. The memory 110 also includes a set 128 of modified signals that have been received in response to inputting the set 126 of initial signals into the configured artificial neural network 122.

In different examples, the set of modified signals 128 may take different forms. In one example, the set of modified signals are time-dependent functional magnetic resonance imaging signals with noise removed. In another example, the set of modified signals 128 are noise components. The memory 110 is also shown as containing a reconstructed functional magnetic resonance image 130 that has used the at least one set 128 of modified signals. In some examples, both the set of modified signals 128 and the set of initial signals 126 are used to construct the functional magnetic resonance image 130.

Fig. 3 shows a flow chart illustrating a method of operating the medical imaging system 100 of fig. 1. First, in step 200, magnetic resonance imaging data 124 is received. The magnetic resonance imaging data is functional magnetic resonance imaging data describing a time-dependent signal for each of a set of voxels. Next in step 202, a set of initial signals is constructed by reconstructing a time-dependent signal for each of the set of voxels using the magnetic resonance imaging data 124. The set of initial signals 126 is then input into the configured artificial neural network 122 in step 204. In response, the configured artificial neural network 122 then outputs a set 128 of modified signals. The method may then execute optional step number 206. In step 206, a functional magnetic resonance image 130 is constructed using at least the set of modified signals 128.

FIG. 3 shows a flow diagram illustrating a method of training or configuring the configured artificial neural network 122 of FIG. 1. First, in step 300, a set of training signals is received for each of a set of voxels. Next in step 302, a set of clean signals is received for each of the set of voxels. The set of training signals includes a set of clean signals plus physiological artifacts or noise. Then, in step 304, physiological artifact data describing the motion of the subject is received. The movement data is time-dependent with respect to the set of cleaning signals. Finally, in step 206, the configured artificial neural network is trained using the set of training signals, the cleaning signal, and the physiological artifact data. The training method illustrated in fig. 3 may be performed prior to deploying the configured artificial neural network 122 to the medical imaging system 100. In other examples, the training method illustrated in fig. 3 is performed using the medical imaging system 100. This may be prior to the initial use of the configured artificial neural network 122 and/or it may be continued training as more data and training signals become available.

Fig. 4 illustrates yet another example of a medical imaging system 400. The medical imaging system 400 illustrated in fig. 4 is similar to the medical imaging system in fig. 1 except that it additionally comprises a magnetic resonance imaging system 402. The magnetic resonance imaging system 402 includes a magnet 404. Magnet 404 is a superconducting cylindrical magnet having a bore 406 therethrough. It is also possible to use different types of magnets, for example, it is also possible to use both split cylindrical magnets and so-called open magnets. The split cylindrical magnet is similar to a standard cylindrical magnet except that the cryostat has been split into two sections to allow access to the iso-plane of the magnet, such a magnet may be used, for example, in conjunction with charged particle beam therapy. An open magnet has two magnet segments, one above the other, with sufficient space between them to receive an object: the arrangement of the two segments is similar to that of a helmholtz coil. Open magnets are common because the object is less constrained. Inside the cryostat of the cylindrical magnet there is a series of superconducting coils. Within the bore 406 of the cylindrical magnet 404 there is an imaging zone 408 in which the magnetic field is strong enough and uniform enough to perform magnetic resonance imaging. A region of interest 409 within the imaging region 408 is shown. Magnetic resonance data are typically acquired for a region of interest. The object 418 is shown supported by an object support 420 such that at least part of the object 418 is within the imaging region 408 and the region of interest 409.

Also within the bore 406 of the magnet is a set of magnetic field gradient coils 410 which are used to initially acquire magnetic resonance data to spatially encode the magnetic spins within the imaging zone 408 of the magnet 404. The magnetic field gradient coils 410 are connected to a magnetic field gradient coil power supply 412. The magnetic field gradient coils 410 are intended to be representative. Typically, the magnetic field gradient coils 410 contain three independent sets of coils for spatial encoding in three orthogonal spatial directions. A magnetic field gradient coil power supply supplies current to the magnetic field gradient coils. The current supplied to the magnetic field gradient coils 410 is controlled as a function of time, and may be ramped or pulsed.

Adjacent to the imaging zone 408 is a radio frequency coil 414 for manipulating the orientation of magnetic spins within the imaging zone 408 and for receiving radio transmissions from spins also within the imaging zone 408. The radio frequency antenna may comprise a plurality of coil elements. The radio frequency antenna may also be referred to as a channel or antenna. The radio frequency coil 414 is connected to a radio frequency transceiver 416. The radio frequency coil 414 and the radio frequency transceiver 416 may be replaced by separate transmit and receive coils and separate transmitters and receivers. It should be understood that the radio frequency coil 414 and the radio frequency transceiver 416 are representative. The radio frequency coil 414 is also intended to represent a dedicated transmit antenna and a dedicated receive antenna. Likewise, the transceiver 416 could also represent a separate transmitter and receiver. The radio frequency coil 414 may also have multiple receive/transmit elements and the radio frequency transceiver 416 may have multiple receive/transmit channels. For example, if a parallel imaging technique such as SENSE is performed, the radio frequency coil 414 would have multiple coil elements.

In this example, the subject 418 is positioned such that a head region of the subject is within the region of interest 409 for performing functional magnetic resonance imaging.

The transceiver 416 and gradient controller 412 are shown connected to the hardware interface 106 of the computer system 102. The memory 110 is shown to also contain a set of pulse sequence commands 430. The pulse sequence commands 430 are configured for controlling the magnetic resonance imaging system to acquire the magnetic resonance imaging data 124 in accordance with a functional magnetic resonance imaging protocol that measures a time-dependent signal for each of the set of voxels. For example, if the functional magnetic resonance imaging technique is a BOLD technique, the pulse sequence commands 430 may be configured to make an accurate measurement of T2.

Fig. 5 shows a flow chart illustrating a method of operating the medical imaging system 400 of fig. 4. The method starts with step 500. In step 500, the processor 104 controls the magnetic resonance imaging system 402 with pulse sequence commands 430 to acquire magnetic resonance imaging data 124. The method then proceeds to steps 200, 202, 204 and 206 as described in fig. 2.

Fig. 6 illustrates one architecture of a configured artificial neural network 122. In this example, there is a recurrent neural network 600 assigned to each voxel of the magnetic resonance imaging data. The schematic representation of recurrent neural network 600 may represent a single recurrent neural network, a recurrent neural network having multiple layers, or multiple recurrent neural networks in a series. Each recurrent neural network 600 includes an input 602 and an output 604. Each recurrent neural network 600 also includes a direct feedback 606 or direct feedback loop, which may be considered a feedback signal or a connection via hidden layers of neurons. The direct feedback 606 is illustrated as a line, but may also be illustrated as a box to represent a hidden layer or neuron. The example illustrated in fig. 6 is the simplest architecture.

Fig. 7 illustrates a more complex version of the configured artificial neural network 122. In fig. 7, instead of the input of the configured artificial neural network 122 feeding directly into the individual recurrent neural network 600, there is now instead an input computation layer 700. The input computation layer 700 has an input 702 that serves as an input for the configured artificial neural network 122. The output of the input computation layer 700 is then directed to the input of the individual recurrent neural network 600. The input computation layer may be fully connected, convolutionally connected or partially connected, for example. The initial use of the input computation layer may help remove artifacts that are spatially or over multiple voxels correlated.

Fig. 8 illustrates a further improvement or refinement of the configured artificial neural network 122. The example illustrated in fig. 8 is similar to the example in fig. 7, except that an additional output computation layer 800 has been added. The output 604 of the recurrent neural network 600 is connected to an input of the output computation layer 800. The output computation layer 800 then has its own output 802, which is the output of the overall configured artificial neural network 122. The output layers may also be full-connected layers, convolutional-connected layers, and partial-connected layers. The addition of the output computation layer 800 may further improve the ability to remove noise or physiological artifacts that are correlated over multiple voxels or that are spatially correlated. The example shown in fig. 8 shows both an input computing layer 700 and an output computing layer 800. The input computing layer 700 may be removed from the example shown in fig. 8.

Fig. 9 shows yet another example of a configured artificial neural network 122. The example illustrated in fig. 9 is similar to the example illustrated in fig. 8, except that the direct feedback 606 or hidden layer neurons are implemented as a hidden feedback layer 900. The hidden feedback layer 900 provides communication between the individual recurrent neural networks 600. The hidden feedback layer 900 may be implemented, for example, as a fully connected layer, a convolutionally connected layer, or a partially connected layer.

Functional magnetic resonance imaging (fMRI) has great potential for benefit in clinical assessment of psychological disease. Pathological brain activity measured in vivo via BOLD (blood oxygen level correlation) signals is associated with psychiatric disorders such as depression, schizophrenia, autism, and the like.

However, fMRI suffers from an inherently low signal-to-noise ratio due to different signal artifacts, which are primarily of a physiological nature. Pure frequency domain filtering or other artifact removal algorithms, such as Independent Component Analysis (ICA), often fail due to the object specificity of these artifacts, or require cumbersome user interaction. On the other hand, measuring artifact signals is not always feasible due to limited hardware availability (e.g. respiratory belt, ECG).

Embodiments may provide an imaging system that maps confounding artifact fMRI temporal processes to artifact-free temporal processes based on an Artificial Neural Network (ANN). During a first training session, fMRI data is recorded. In addition, physiological artifacts are recorded simultaneously, for example, with a respiratory belt, camera, ECG, or the like. The artifact signal is used to clean the BOLD time process, for example, using a simple regression technique. Given the aliasing artifact and clean signal, the recursive deep neural network is trained to map from the first to the next. Providing a sufficiently large and clean data set, the algorithm will be able to automatically clean the data of unseen data sets.

In addition, the external parameters controlled by the user in the GUI settings control the intensity of the cleaning. The user sets "strong clean" which may also remove valid data portions or "weak clean" which may lead to residual artifacts.

Functional magnetic resonance imaging (fMRI) has great potential for benefit in clinical assessment of psychological diseases, for example for disease classification, treatment selection or disease progression prognosis. Pathological brain activity measured in vivo via BOLD (blood oxygen level dependent) signals has been associated with psychiatric disorders such as depression, schizophrenia, autism, and the like.

As mentioned previously, fMRI suffers from an inherently low signal-to-noise ratio and is therefore prone to many different signal artifacts of predominantly physiological nature. Although thermal and system noise scales linearly with static field strength, physiological artifacts scale to powers of 2.

Recent automatic artifact reduction techniques, such as frequency domain filtering or Independent Component Analysis (ICA), often fail due to the object specificity (e.g., signal frequency and shape) of these artifacts. To improve these methods, cumbersome manual work and expert knowledge are required. On the other hand, measuring artifact signals is not always feasible due to limited hardware availability in proper settings (e.g. respiratory belt, ECG).

Examples may provide an imaging system that maps confounding artifact fMRI temporal processes to artifact-free temporal processes based on an Artificial Neural Network (ANN). During a first training session, fMRI data is recorded. Furthermore, physiological artifacts are recorded simultaneously, for example, with a respiratory belt, camera, ECG or similar kind of sensor. The artifact signal is used to clean the BOLD time process, for example, using a simple linear regression technique. Given the aliasing artifact and clean signal, the recursive deep neural network is trained to map from the first to the next. Providing a sufficiently large and clean data set, the algorithm will be able to automatically clean the data of unseen data sets.

In addition, external parameters controlled by a user in a Graphical User Interface (GUI) may be considered to set the cleaning level. The user sets a "high cleaning level" that may also remove valid data portions or a weak cleaning level that may result in some residual artifacts remaining.

To construct the present invention, artifact-free fMRI training data is provided. The large number of acquired fMRI BOLD time-sequences used may be manually cleaned, for example. These time series originate from a large number of subjects and different brain regions and vary over their duration. Cleaning can be achieved in a previous step, for example by visual inspection of the spectrum and filtering in the appropriate frequency band using ICA or other techniques (e.g., regression). Furthermore, artifact signals may be measured, for example, using a respiratory belt and an ECG. Given the measured signal, it can be regressed from the BOLD signal. Fig. 10 and 11 show examples of a miscellaneous artifact-free signal and an ideal artifact-free BOLD signal. For data enhancement, the artifact signal from one subject may also be combined with a clean BOLD signal from another subject.

Fig. 10 shows a curve illustrating an idealized BOLD response 1000 for a single voxel. The x-axis is time 1002 and the y-axis 1004 is the BOLD response. As can be seen by examining the signal in fig. 10, it is extremely noisy and it is not possible to interpret or see the BOLD signal.

Fig. 11 illustrates an idealized artifact-free BOLD signal 1100. As can be seen, the artifact-free BOLD signal 1100 has an initial dip 1102, a peak 1104, and post-stimulation undershoot 1106. This is not visible in fig. 10.

Given a number of paired time series, a Recurrent Neural Network (RNN) may be trained to learn the mapping y ═ f (x) from the aliased artifact time series x (t) to the artifact-free time series y (t). In its simplest form, the RNN can be viewed as a fully-connected neural network that is developed in the time domain, see fig. 12 below. In this example, there is only a single neuron for each instance of time, and only two parameters are learned. Assuming hyperbolic tangent as the activation function, the transfer function becomes y_t＝tanh(wy_t-1+ux_t) Where w and u are learnable parameters.

Fig. 12 illustrates the operation of a single recurrent neural network 600 representing one voxel over a period of multiple time periods. This illustrates the so-called unfolding process, or it is an image of the recurrent neural network 600 that is unfolded in time. Various inputs 602 and 604 represent inputs into the same recurrent neural network 600 but at subsequent time intervals.

This simple scenario serves only for illustration purposes, i.e. how a neural network can be used to learn the mapping between time series. Network complexity can (and must) be increased by stacking more of these described units on top of each other, resulting in a deep artificial neural network. This will enable the network to properly learn the complex structure of multiple artifacts and eventually regress them out of the signal. Furthermore, a hierarchical architecture may be considered, where first the artifact signal itself is learned by the RNN and then fed into a second network responsible for regressing the artifact signal from the original BOLD sequence in a non-linear manner. For the second network, additional parameters that control the strength of the regression may be introduced. As such, the user has the ability to prefer to allow false positives (weak regression results in residual artifacts) or false negatives (strong regression results in false removed true BOLD signals).

One disadvantage of the method outlined above is that the time correlation is not captured until a single point in time. However, to learn long-term correlations, such as breathing and heartbeat artifacts, one needs a more complex architecture. One approach would be to use a Long Short Term Memory (LSTM) network as the RNN. Fig. 13 shows the architecture of these more complex cells with internal state variables modified by different operation gates (input gate, forget gate, output gate).

Fig. 13 illustrates the replacement of a simple recurrent neural network 600 with a so-called LSTM network. This is a long-short term memory network. The LSTM 600' may be used to replace the recurrent neural network 600 illustrated in other examples. In fig. 13, LSTM 600' is again used for a single voxel. As with fig. 12, LSTM 600' is shown spread out in time.

In yet another example, it may be considered that the range of clutter artifact removal also extends to those generated by the recording system (MRI, ECG, motion sensor, etc.) and their potential undesirable crosstalk. Suppliers typically pay attention to ensure perfect image/data quality throughout the course of an fMRI experiment, but system heating and other temporal effects may slightly affect data quality. Moreover, the interaction of strong gradients applied in MRI can have a detrimental effect on the concurrent readings of other sensors (ru ECG). This information can also be fed in a deep learning system to remove these effects.

Examples may be used as a first step in the analysis of fMRI data sets. After data acquisition, the user is presented with data and processing options in a graphical user interface. Among the processing options, the user may select "automatic artifact removal" and additionally set a parameter that controls the intensity of artifact removal.

Further examples may have one or more of the following properties or characteristics:

automatic pre-selection of optimal filter parameters based on the following without user interaction:

statistical analysis of large universal data sets (e.g., for training ANN).

Statistical analysis of specific data collected in the field, as the various collected signals may have different properties and may be operator and equipment type dependent.

Adaptive selection of parameters based on the fMRI sequence used.

Adaptive selection of parameters based on the type of equipment used (specific differences resulting from technical differences, such as ECG versus VCG versus camera system and vendor equipment differences).

Gated Repeat Units (GRUs) are used as a special class of LSTM to reduce the number of parameters required for training and to reduce training time.

The depth of the "ANN" is increased to increase expression ability (e.g., better separation of learning slow and fast changing dynamics).

A shortcut connection is used during back propagation, e.g. for faster gradient flows.

Two ANN's are stacked on top of each other to increase the expressive power and learning power of the network.

While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed embodiments.

Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. Although specific elements are recited in mutually different dependent claims, this does not indicate that a combination of these elements cannot be used to advantage. A computer program may be stored and/or distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together 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 telecommunication systems. Any reference signs in the claims shall not be construed as limiting the scope.

Fig. 14 illustrates yet another example of a configured artificial neural network 122. The example illustrated in fig. 14 is similar to the example illustrated in fig. 8, except that there is no input layer 700 and no output layer 800. The set of initial signals is input directly into the input 602 of the recurrent neural network 600. The set of modified signals is directly output by the output 604 of the recurrent neural network 600. For each voxel in the set of voxels, there is a recurrent neural network 600. The hidden feedback layer 900 provides communication between the individual recurrent neural networks 600. The hidden feedback layer 900 may be implemented, for example, as a fully connected layer, a convolutionally connected layer, or a partially connected layer. The hidden feedback layer 900 enables the configured artificial neural network to remove physiological artifacts that are spatially correlated between different voxels.