Information processing method, information processing program, and information processing device

Abstract

To provide a novel solution capable of removing noise even from brain waves measured by a brain wave cap without a carbon wire loop.SOLUTION: An information processing method includes the steps of: measuring brain waves and carbon wire loop signals of a subject placed under an artificial magnetic field using a brain wave cap equipped with a carbon wire loop; removing noise on the basis of the carbon wire loop signals from the brain waves of the subject or the brain waves after removing an artifact from the brain waves of the subject; and optimizing a hyper parameter of a noise removal model by using the time-series data of the brain waves before removing noise on the basis of the carbon wire loop signals as the input data and using the time-series data of the brain waves after removing noise on the basis of the carbon wire loop signals as the correct answer data.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an information processing method, an information processing program, and an information processing device. [Background technology]

[0002] When measuring electroencephalograms in a high magnetic field such as that in an MRI (Magnetic Resonance Imaging), artifacts caused by minute vibrations (from the device itself, heartbeat, body movement, etc.) may be mixed into the measurement results.

[0003] A method has been proposed for removing noise contained in measured EEG by using an EEG cap equipped with a carbon-wire loop (hereinafter also abbreviated as "CWL") to measure the induced electromotive force signal generated during EEG measurement and inputting the measured induced electromotive force signal into a linear model based on Faraday's law of electromagnetic induction (see Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Johan N. van der Meer et al., "Carbon-wire loop based artifact correction outperforms post-processing EEG / fMRI corrections-A validation of a real-time simultaneous EEG / fMRI correction method", NeuroImage, Volume 125, 15 January 2016, Pages 880-894 Summary of the Invention [Problem to be solved by the invention]

[0005] In the above-mentioned method, an EEG cap with a CWL is used to simultaneously measure an EEG and a carbon wire loop signal, and then the carbon wire loop signal is used to remove noise contained in the EEG. The inventors of the present application have come up with a new solution that can remove noise even from EEGs measured using an EEG cap without a CWL, by focusing on the relationship between the measured EEG and the EEG after noise removal.

[0006] The present invention provides a novel solution that can remove noise even from EEG measured using an EEG cap without a CWL. [Means for solving the problem]

[0007] An information processing method according to one embodiment of the present invention includes the steps of measuring the electroencephalogram and carbon wire loop signal of a subject placed under an artificial magnetic field using an EEG cap to which a carbon wire loop is attached, removing noise from the subject's EEG or from the EEG after artifacts have been removed from the subject's EEG based on the carbon wire loop signal, and optimizing hyperparameters of a noise removal model using as input data time series data of the EEG before noise is removed based on the carbon wire loop signal and as ground truth data time series data of the EEG after noise has been removed based on the carbon wire loop signal.

[0008] The noise removal model may include a function unit that calculates a predicted value, a first regression unit that outputs time series data of a first number of samples of the input data to the function unit, and a second regression unit that outputs time series data of a second number of samples of the predicted value to the function unit. The step of optimizing hyperparameters of the noise removal model may include a step of optimizing at least one of the first number of samples and the second number of samples.

[0009] The function portion of the time series data may include at least one of a nonlinear function and a linear function. The step of optimizing the hyperparameters of the noise removal model may include a step of optimizing coefficients that define at least one of the nonlinear function and the linear function.

[0010] The noise removal model may be configured to receive an input of one channel of electroencephalogram and output one channel of electroencephalogram. The step of optimizing the hyperparameters of the noise removal model may include a step of sequentially arranging each of the multiple channels included in the input data in series to make it one-dimensional, and a step of sequentially arranging each of the multiple channels included in the ground truth data in series to make it one-dimensional.

[0011] An information processing method according to another embodiment of the present invention includes a step of measuring the electroencephalogram of a subject placed under an artificial magnetic field using an electroencephalogram cap, and a step of inputting the electroencephalogram of the subject or the electroencephalogram after artifacts have been removed from the electroencephalogram of the subject into a trained model constructed by a learning process to remove noise. The learning process includes a step of measuring the electroencephalogram of the subject placed under an artificial magnetic field and a carbon wire loop signal using an electroencephalogram cap to which a carbon wire loop is attached, a step of removing noise from the electroencephalogram of the subject or the electroencephalogram after artifacts have been removed from the electroencephalogram of the subject based on the carbon wire loop signal, and a step of optimizing a hyperparameter using the time series data of the electroencephalogram before noise is removed based on the carbon wire loop signal as input data and the time series data of the electroencephalogram after noise is removed based on the carbon wire loop signal as correct answer data.

[0012] The trained model may be configured to receive an input of one channel of electroencephalogram and output one channel of electroencephalogram. The step of removing noise may include a step of inputting one channel of electroencephalogram to the trained model and repeating a process of calculating time series data of a predicted value by the number of channels.

[0013] The trained model may include at least one of a nonlinear function and a linear function, and may include a function unit that calculates a predicted value, a first regression unit that outputs time series data for a first number of samples of the input data to the function unit, and a second regression unit that outputs time series data for a second number of samples of the predicted value to the function unit.

[0014] According to yet another embodiment of the present invention, there is provided an information processing program for causing a computer to execute the above-mentioned information processing method.

[0015] An information processing device according to yet another embodiment of the present invention includes means for storing the electroencephalogram and carbon wire loop signal of a subject placed under an artificial magnetic field measured using an electroencephalogram cap to which a carbon wire loop is attached, means for removing noise from the subject's electroencephalogram or from the electroencephalogram after artifacts have been removed from the subject's electroencephalogram based on the carbon wire loop signal, and means for optimizing hyperparameters of a noise removal model using as input data time series data of the electroencephalogram before noise is removed based on the carbon wire loop signal and as correct data time series data of the electroencephalogram after noise has been removed based on the carbon wire loop signal.

[0016] An information processing device according to yet another embodiment of the present invention includes a means for storing the electroencephalogram of a subject placed under an artificial magnetic field, measured using an electroencephalogram cap, and a means for inputting the electroencephalogram of the subject or the electroencephalogram after artifacts have been removed from the electroencephalogram of the subject into a trained model to remove noise. The trained model is configured by optimizing hyperparameters using the electroencephalogram of the subject placed under an artificial magnetic field, measured using an electroencephalogram cap with a carbon wire loop attached, and a carbon wire loop signal, with time series data of the electroencephalogram before noise is removed based on the carbon wire loop signal as input data, and time series data of the electroencephalogram after noise is removed based on the carbon wire loop signal as correct answer data. Effect of the Invention

[0017] According to the present invention, noise can be removed even from electroencephalograms measured using an electroencephalogram cap without a CWL. [Brief description of the drawings]

[0018] [Figure 1] 1 is a diagram for explaining an overview of a noise removal technique according to an embodiment of the present invention; [Diagram 2] 1 is a schematic diagram showing a configuration example of a measurement system according to an embodiment of the present invention. [Diagram 3] 1 is a schematic diagram showing an example of a hardware configuration of an information processing device according to an embodiment of the present invention; [Figure 4] 13 is a flowchart showing the procedure for pre-processing of electroencephalograms measured using an electroencephalogram cap without a CWL. [Diagram 5] 13 is a flowchart showing the procedure for pre-processing of electroencephalograms measured using an electroencephalogram cap with CWL. [Figure 6] 10 is a flowchart showing a processing procedure of a noise removal method according to the present embodiment. [Figure 7] 10 is a flowchart showing a processing procedure of a learning process for configuring a noise removal model used in the noise removal method according to the present embodiment. [Figure 8] 11 is a diagram for explaining an example of serialization used in the noise removal technique according to the present embodiment. FIG. [Figure 9] FIG. 2 is a schematic diagram showing an example of a functional configuration of an information processing device according to the present embodiment. [Figure 10] FIG. 2 is a schematic diagram showing an example of the configuration of a noise removal model used in the noise removal technique according to the present embodiment. [Figure 11] 11 is a graph showing an example of noise removal performance of a noise removal model according to the present embodiment. [Figure 12] 13 is a graph showing a case where the noise removal model according to the present embodiment is applied to electroencephalograms measured on different days of the same subject. [Figure 13] 13 is a graph showing a case where the noise removal model according to the present embodiment is applied to electroencephalograms measured from different subjects. [Figure 14] 11 is a graph showing an example of channel generalization performance of a noise removal model according to the present embodiment. [Figure 15] 13 is a graph showing another example of channel generalization performance of the noise removal model according to the present embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will now be described in detail with reference to the accompanying drawings, in which the same or corresponding parts in the drawings are designated by the same reference characters and will not be described repeatedly.

[0020] [A. Overview] First, an overview of the noise removal technique according to the present embodiment will be described.

[0021] Fig. 1 is a diagram for explaining an outline of a noise removal method according to the present embodiment. Referring to Fig. 1, an electroencephalogram 10 and a carbon wire loop signal (hereinafter also referred to as a "CWL signal 16") measured from a subject placed under an artificial magnetic field using an electroencephalogram cap 1 with CWL are input to a noise removal library 14, whereby a noise-removed electroencephalogram 12 is generated. The CWL signal 16 is a signal of an induced electromotive force generated during electroencephalogram measurement.

[0022] The noise removal library 14 is disclosed as, for example, CWL regression. The noise removal library 14 mainly removes noise caused by placing a subject under an artificial magnetic field. The artificial magnetic field means an environment in which a magnetic field generated artificially using an arbitrary device exists, not a magnetic field existing in nature. The artificial magnetic field may be, for example, a high magnetic field of 10,000 times or more than the geomagnetic field. As a typical example of placing a subject under an artificial magnetic field, a situation in which a magnetic resonance image or a functional magnetic resonance image of the subject is measured using an MRI measurement device is assumed. In the following description, a case in which an electroencephalogram is measured from a subject in an MRI will be mainly described, but the present invention is not limited to this and can be applied to a process for removing noise from an electroencephalogram measured from a subject in a situation affected by a magnetic field generated by an arbitrary device.

[0023] In the noise removal method according to the present embodiment, a noise removal model 24, which is a trained model, is constructed using an EEG 10 measured from a subject in an MRI using an EEG cap 1 with CWL and an EEG 12 from which noise has been removed. It is considered that the EEG 10 measured using the EEG cap 1 with CWL has some correlation with the CWL signal 16 measured at the same time. In other words, it is considered that the EEG 10 contains a component reflecting the characteristics of the CWL signal 16. Therefore, a noise removal model 24 for removing noise can be constructed by learning using the EEG 10, which is raw data, as input data and the EEG 12 from which noise has been removed as a teacher signal. In other words, the EEG 12 from which noise has been removed becomes ground truth data.

[0024] The noise removal technique according to the present embodiment predicts electroencephalogram 22 from which noise has been removed using electroencephalogram 12, which has been noise removed using a CWL signal, as a teacher signal.

[0025] By using the noise removal model 24, it is possible to generate noise-removed electroencephalograms 22 from electroencephalograms 20 measured using an electroencephalogram cap 2 without a CWL.

[0026] The electroencephalogram 20 to be subjected to noise removal may be one measured in the past. By applying the noise removal method according to the present embodiment, even if the electroencephalogram 20 was measured in the past using an electroencephalogram cap 2 without a CWL, noise can be appropriately removed.

[0027] The noise removal model 24 may be prepared for each model of the MRI measurement device, or may be prepared for each MRI measurement device. After constructing the noise removal model 24 using a plurality of MRI measurement devices, an integrated noise removal model 24 may be constructed using a known technique such as transfer learning.

[0028] [B. System configuration example] Next, a configuration example of a measurement system according to the present embodiment will be described.

[0029] 2 is a schematic diagram showing an example of the configuration of a measurement system 4 according to the present embodiment. With reference to FIG. 2, the measurement system 4 includes an information processing device 100, an electroencephalogram measuring device 200, and an MRI measuring device 300.

[0030] The information processing device 100 acquires electroencephalograms measured by the electroencephalogram measuring device 200. The information processing device 100 may acquire magnetic resonance images or functional magnetic resonance images measured by the MRI measuring device 300.

[0031] The electroencephalogram measuring device 200 measures a signal (electrical signal) indicating an electroencephalogram from an electroencephalogram cap 1 with CWL (or an electroencephalogram cap 2 without CWL) worn on the head of a subject. The electroencephalogram measuring device 200 includes a multiplexer 202, a noise filter 204, an A / D (Analog to Digital) converter 206, a storage unit 208, and an interface 210.

[0032] The multiplexer 202 sequentially selects a pair of electrodes (channels) from the group of electrodes of the EEG cap, and electrically connects them to the noise filter 204. The noise filter 204 includes a high-frequency cut filter and the like, and removes noise components included in the signal (electrical signal) indicating the EEG generated in the selected channel.

[0033] The A / D converter 206 samples the electrical signal (analog signal) output from the noise filter 204 at predetermined intervals, and outputs the sampled signal as a digital signal.

[0034] The storage unit 208 sequentially stores the time-series data (digital signal) output from the A / D converter 206 in association with the selected channel and information indicating timing (for example, time or counter value).

[0035] The interface 210 outputs the time-series data indicating the electroencephalogram stored in the storage unit 208 to the information processing device 100 in response to an access from the information processing device 100 or the like.

[0036] 3 is a schematic diagram showing an example of a hardware configuration of information processing device 100 according to the present embodiment. Information processing device 100 can typically employ a personal computer that follows a general-purpose architecture.

[0037] Referring to FIG. 3, the information processing device 100 includes, as main components, one or more processors 102, a main memory unit 104, a control interface 106, a network interface 108, an input unit 110, a display unit 112, and a secondary memory unit 120.

[0038] The processor 102 is composed of calculation processing circuits such as a CPU (Central Processing Unit) and a GPU (Graphical Processing Unit), and performs the processing described below by executing codes contained in various programs stored in the secondary memory unit 120 in a specified order.

[0039] The main memory unit 104 is configured with a volatile storage device such as a dynamic random access memory (DRAM) or a static random access memory (SRAM), and stores program codes executed by the processor 102 and data referenced during program execution.

[0040] The control interface 106 exchanges data with an interface 210 of the electroencephalogram measuring device 200 .

[0041] The network interface 108 exchanges data with an external device (for example, a data server device on the cloud, etc.).

[0042] The control interface 106 and the network interface 108 are configured with any communication component such as a wired LAN (Local Area Network), a wireless LAN, a USB (Universal Serial Bus), or Bluetooth (registered trademark).

[0043] The input unit 110 is configured with a mouse, a keyboard, and the like, and receives operations from the user.

[0044] The display unit 112 is configured with a display or the like, and notifies the user of various information related to the execution state of processing in the information processing device 100 and operations.

[0045] The secondary storage unit 120 is configured with a non-volatile storage device such as a hard disk or a solid state drive (SSD), and stores the programs executed by the processor 102, data necessary for executing the programs, and the like.

[0046] The secondary storage unit 120 stores raw data, that is, electroencephalogram time series data 121 (electroencephalogram 10 or electroencephalogram 20), raw data, that is, CWL time series data 122 (CWL signal 16), electroencephalogram time series data 123 from which scanning artifacts (described later) have been removed, electroencephalogram time series data 124 (electroencephalogram 12 or electroencephalogram 22) from which noise has been removed, a scanning artifact removal program 125, a noise removal program 126 (including the noise removal library 14), and a learning program 127. The secondary storage unit 120 may store a noise removal model 24 generated by the learning program 127. The secondary storage unit 120 may store a system program such as an OS (Operating System).

[0047] [C.Noise] Next, an example of noise that occurs when measuring a subject in an MRI will be described.

[0048] (1) Helium pump artifact The MRI measurement device 300 has a compressor for filling the liquid helium used to cool the coil. The vibration of the pump included in the compressor is transmitted to the EEG cap and cables, and can become noise. Although it is possible to stop the cooling system for a short period of time, the cooling system is not usually stopped due to the risk of quenching caused by temperature rise. Such noise generated by the MRI measurement device is also called helium pump artifact.

[0049] (2) Scanning artifacts Noise may occur when an MRI measurement device performs measurement (scanning). Such noise is also called scanning artifacts. Scanning artifacts are generally periodic noise.

[0050] (3) BCG artifacts The expansion and contraction of blood vessels due to heartbeat can become noise. Noise caused by heartbeat is also called ballistocardiogram (BCG) artifact. BCG artifact can be removed by using an electrocardiogram (ECG) measured simultaneously with the brainwave.

[0051] (4) Motion artifacts Vibrations caused by movements of the head, face, limbs, etc. can become motion artifacts. Noise parts are excluded from the analysis, but since there is no regularity in the timing of occurrence, it is difficult to extract the noise parts.

[0052] [D. Pre-processing procedure] Next, a procedure for pre-processing the measured electroencephalogram will be described. The main purpose of pre-processing is to remove noise before various analyses are performed.

[0053] 4 is a flowchart showing a procedure for pre-processing the electroencephalogram 20 measured using the CWL-free electroencephalogram cap 2. Each step shown in FIG. 4 may be realized by the processor 102 of the information processing device 100 executing the scanning artifact removal program 125.

[0054] 4, the information processing device 100 acquires raw data measured using the EEG cap 2 without a CWL (step S10). The raw data is measured from the subject by the EEG measuring device 200. The raw data includes, for example, time-series data of 63 channels of electroencephalogram (EEG) and 1 channel of electrocardiogram (ECG).

[0055] The information processing device 100 removes scanning artifacts included in the raw data of the electroencephalogram by AAS (average artifact subtraction) (step S12). When the MRI measurement device 300 is not performing an operation of measuring a magnetic resonance image or a functional magnetic resonance image of the subject (hereinafter, also referred to as "fMRI scan"), the process of step S12 may be omitted.

[0056] The information processing device 100 detects R waves appearing in the time-series data of the electrocardiogram (ECG) (step S14), and further removes BCG artifacts included in the electroencephalogram according to the appearance position of the R waves (step S16). The AAS method can be used to remove the BCG artifacts.

[0057] The information processing device 100 outputs the time-series data in which the scanning artifacts and the BCG artifacts have been removed from the raw electroencephalogram data as a preprocessing result (step S18).

[0058] 5 is a flowchart showing a procedure for pre-processing the electroencephalogram 10 measured by the CWL-enabled electroencephalogram cap 1. Each step shown in FIG. 5 may be realized by the processor 102 of the information processing device 100 executing the scanning artifact removal program 125 and the noise removal program 126.

[0059] 5, the information processing device 100 acquires raw data measured using the EEG cap 1 with CWL (step S20). The raw data is measured from the subject by the EEG measuring device 200. The raw data includes, for example, time-series data of 63 channels of electroencephalograms (EEG) and 4 channels of CWL signals. The raw data may also include 1 channel of electrocardiogram (ECG).

[0060] In this manner, the information processing device 100 measures or acquires the electroencephalogram and CWL signal of the subject using the electroencephalogram cap 1 to which the CWL is added.

[0061] The information processing device 100 removes scanning artifacts contained in the raw data of the electroencephalogram by the AAS method (step S22). If an fMRI scan is not performed, the process of step S22 may be omitted.

[0062] The information processing device 100 inputs the time series data of the electroencephalogram (or the raw electroencephalogram signal) and the CWL signal after removing the scanning artifact into a noise removal library (CWL regression) to remove noise including the helium pump artifact, the BCG artifact, and the body movement artifact (step S24). The information processing device 100 outputs the time series data of the electroencephalogram output from the noise removal library 14 as a preprocessing result (step S26).

[0063] In this manner, the information processing device 100 removes noise from the electroencephalogram of the subject or from the electroencephalogram after artifacts have been removed from the electroencephalogram of the subject, based on the CWL signal.

[0064] According to the noise removal method of the present embodiment, even for raw EEG data without a CWL signal measured at the same time, noise removal similar to that of the noise removal library (CWL regression) can be achieved. That is, noise including helium pump artifacts, BCG artifacts, and body movement artifacts can be removed from raw EEG data without the need for a CWL signal.

[0065] Fig. 6 is a flowchart showing a processing procedure of the noise removal method according to the present embodiment. Each step shown in Fig. 6 may be realized by the processor 102 of the information processing device 100 using the noise removal model 24 and executing the scanning artifact removal program 125 and the noise removal program 126.

[0066] 6, the information processing device 100 acquires raw data measured using the EEG cap 2 without CWL (step S30). The raw data is measured from the subject by the EEG measuring device 200. The raw data includes, for example, 63 channels of electroencephalograms (EEG). The raw data may also include one channel of electrocardiogram (ECG).

[0067] In this manner, the information processing device 100 measures or acquires the electroencephalogram of the subject using the electroencephalogram cap 2.

[0068] The information processing device 100 removes scanning artifacts contained in the raw electroencephalogram data by the AAS method (step S32). If an fMRI scan is not performed, the process of step S22 may be omitted.

[0069] The information processing device 100 inputs the electroencephalogram time series data after the scanning artifacts have been removed to the noise removal model 24 to generate noise-removed electroencephalogram time series data (step S34). The information processing device 100 outputs the electroencephalogram time series data output from the noise removal model 24 as a preprocessing result (step S36).

[0070] In this way, the information processing device 100 inputs the subject's electroencephalogram or the electroencephalogram after artifacts have been removed from the subject's electroencephalogram into a trained model (noise removal model 24) constructed by a learning process, and removes noise.

[0071] By employing the noise removal technique according to this embodiment, noise including helium pump artifacts, BCG artifacts, and body movement artifacts can be removed from raw EEG data even in the absence of a simultaneously measured CWL signal.

[0072] The noise removal model 24 may be configured to be able to input electroencephalograms of 63 channels or to input electroencephalograms of only one channel. When the noise removal model 24 has an input of one channel, the process of step S34 includes a process of inputting the time series data of the electroencephalogram after removing the scanning artifact into the noise removal model 24 one channel at a time. That is, in step S34, the process of inputting the time series data of the electroencephalogram of one channel into the noise removal model 24 and calculating the time series data of the predicted value may be repeated the number of times equal to the number of channels.

[0073] [E. Learning process of noise removal model] Next, a procedure of a learning process for constructing a noise removal model for removing noise will be described.

[0074] 7 is a flowchart showing a procedure of a learning process for constructing a noise removal model 24 used in the noise removal method according to the present embodiment. Each step shown in FIG. 7 may be realized by the processor 102 of the information processing device 100 executing the noise removal program 126 and the learning program 127.

[0075] 7, the information processing device 100 acquires raw electroencephalogram data (time-series data) and electroencephalogram time-series data obtained by removing noise from the raw electroencephalogram data using a noise removal library (CWL regression) (step S50). Note that in step S50, the processes of steps S20 to S26 in FIG. 5 are executed in advance or simultaneously.

[0076] The information processing device 100 extracts time series data of one channel included in the raw electroencephalogram data as input data, and extracts time series data of a corresponding channel included in the noise-removed electroencephalogram time series data as correct answer data (step S52).

[0077] The information processing device 100 optimizes hyperparameters of the noise removal model 24 being trained, using the input data and the correct answer data extracted in step S52 (step S54).

[0078] The information processing device 100 determines whether or not the learning end condition is satisfied (step S56). If the learning end condition is not satisfied (NO in step S56), the processing from step S52 onwards is repeated.

[0079] If the learning end condition is met (YES in step S56), the information processing device 100 configures the noise removal model 24 using the current hyperparameters (step S58).

[0080] In this way, the information processing device 100 optimizes the hyperparameters of the noise removal model 24 using the time series data of the electroencephalogram before noise is removed based on the CWL signal as input data and the time series data of the electroencephalogram after noise is removed based on the CWL signal as correct answer data.

[0081] The learning end condition may be, for example, that hyperparameters are optimized using a predetermined number of channels among the channels included in the input data and the correct answer data, or that the performance of the model being learned, calculated using a portion of the correct answer data as validation data, reaches a predetermined level.

[0082] FIG. 7 shows an example of repeating hyperparameter optimization for one channel at a time. In step S52, however, input data and super-answer data for one channel may be generated by sequentially arranging all channels contained in the input data in series to make them one-dimensional (serialized), and sequentially arranging all channels contained in the super-answer data in series to make them one-dimensional (serialized).

[0083] Fig. 8 is a diagram for explaining an example of serialization used in the noise removal method according to the present embodiment. Referring to Fig. 8, each of the input data and the correct answer data includes time series data of a plurality of channels with respect to the same time axis. By sequentially arranging the time series data of each channel along the time axis, one-dimensional time series data can be generated.

[0084] In this way, the process of optimizing the hyperparameters of the noise removal model 24 (step S54 in FIG. 7) may include a process of sequentially arranging each of the multiple channels included in the input data in series to make it one-dimensional, and a process of sequentially arranging each of the multiple channels included in the ground truth data in series to make it one-dimensional.

[0085] FIG. 7 illustrates an example of a processing procedure for constructing a noise removal model 24 having one input channel. However, if the number of input channels of the noise removal model 24 is the same as the number of channels of the raw EEG data, the raw EEG data can be used as input data as is, and the noise-removed EEG time series data can be treated as ground truth data as is.

[0086] [F. Functional Configuration] Next, an example of a functional configuration of information processing device 100 according to the present embodiment will be described.

[0087] Fig. 9 is a schematic diagram showing an example of a functional configuration of information processing device 100 according to the present embodiment. Fig. 9(A) shows an example of a functional configuration for configuring noise removal model 24 (learning phase), and Fig. 9(B) shows an example of a functional configuration for performing noise removal using noise removal model 24 (operation phase).

[0088] 9(A) may be realized by the processor 102 of the information processing device 100 executing the noise removal program 126 and the learning program 127. With reference to FIG. 9(A), the information processing device 100 includes a scanning artifact removal unit 152, a noise removal model 24, and a learning unit 154 as functional configurations of the learning phase.

[0089] The scanning artifact removing section 152 removes scanning artifacts from the electroencephalogram time series data 121 (electroencephalogram 10 shown in FIG. 1), which is raw data, to generate electroencephalogram time series data 123.

[0090] The noise removal library 14 removes noise contained in the input electroencephalogram time series data 123 (or electroencephalogram time series data 121) based on the CWL time series data 122 (CWL signal 16 shown in FIG. 1). The noise removal library 14 outputs electroencephalogram time series data 124 (electroencephalogram 12 shown in FIG. 1) after noise removal. In this way, the noise removal library 14 removes noise based on the CWL signal from the electroencephalogram of the subject or from the electroencephalogram after artifacts have been removed from the electroencephalogram of the subject.

[0091] If an fMRI scan has not been performed, the electroencephalogram time-series data 121 may be input directly to the noise removal library 14.

[0092] In addition, the subject's electroencephalogram 10 (electroencephalogram time series data 121) and CWL signal 16 (CWL time series data 122) measured using the electroencephalogram cap 1 with CWL added are stored in the secondary memory unit 120 (or the main memory unit 104) of the information processing device 100.

[0093] Based on the input data and the correct answer data, the learning unit 154 optimizes hyperparameters of the noise removal model 24. More specifically, the learning unit 154 configures the noise removal model 24 using the electroencephalogram time series data 123 (or the electroencephalogram time series data 121) as input data and the electroencephalogram time series data 124 after noise removal as correct answer data.

[0094] In this way, the learning unit 154 uses the EEG time series data before noise is removed based on the CWL signal 16 (electroencephalogram time series data 123 or electroencephalogram time series data 121) as input data, and the EEG time series data after noise is removed based on the CWL signal 16 (electroencephalogram time series data 124) as correct data, and optimizes the hyperparameters of the noise removal model 24.

[0095] 9(B) may be realized by the processor 102 of the information processing device 100 executing the scanning artifact removal program 125 and the noise removal program 126 while using the noise removal model 24. With reference to Fig. 9(B), the information processing device 100 includes, as a functional configuration in the operation phase, a scanning artifact removal unit 152 and the noise removal model 24 which is a trained model.

[0096] The scanning artifact removing section 152 removes scanning artifacts from the electroencephalogram time series data 121 (electroencephalogram 20 shown in FIG. 1), which is raw data, to generate electroencephalogram time series data 123.

[0097] The electroencephalogram 20 (electroencephalogram time series data 121) of the subject measured using the electroencephalogram cap 2 is stored in the secondary storage unit 120 (or the main storage unit 104) of the information processing device 100.

[0098] The noise removal model 24 predicts electroencephalograms from which noise contained in the input electroencephalogram time series data 123 (or electroencephalogram time series data 121) has been removed. The noise removal model 24 outputs electroencephalogram time series data 124 after noise removal (electroencephalograms 22 shown in FIG. 1). In this way, the electroencephalograms of the subject or electroencephalograms from which artifacts have been removed are input to the noise removal model 24, and noise is removed.

[0099] When an fMRI scan is not performed, the electroencephalogram time-series data 121 may be input directly to the noise removal model 24 .

[0100] 9(A) and 9(B) may be implemented, or only one of them may be implemented, in the same information processing device 100. That is, there may be an information processing device 100 in which only the functional configuration example shown in Fig. 9(A) is implemented, and an information processing device 100 in which only the functional configuration example shown in Fig. 9(B) is implemented.

[0101] [G. Noise reduction model] Next, a configuration example of a noise removal model for removing noise will be described.

[0102] For example, a time series model capable of statistically processing time series data can be used as the noise removal model. An example of the time series model may be an ARX (Auto-Regressive with eXo-genous) model. The ARX model may be a nonlinear ARX model whose output function includes a nonlinear function, or a linear ARX model whose output function includes only a linear function.

[0103] Fig. 10 is a schematic diagram showing an example of the configuration of a noise removal model 24 used in the noise removal technique according to the present embodiment. With reference to Fig. 10, the noise removal model 24 predicts time series data of predicted values ​​y (for one channel) indicating electroencephalograms after noise removal, for time series data of input values ​​u indicating electroencephalograms (for one channel) before noise removal. That is, the noise removal model 24 is configured to accept input of electroencephalograms of one channel and to output electroencephalograms of one channel.

[0104] The noise removal model 24 includes a regression unit 241 to which an input value u is input, a regression unit 242 to which a predicted value y is input, and a function unit 243 to which outputs of the regression unit 241 and the regression unit 242 are given. The function unit 243 calculates the predicted value y.

[0105] The regression unit 241 corresponds to a first regression unit, and outputs time series data including a current value and a past value defined by the hyperparameter 244 for the input value u. For example, when the hyperparameter 244 includes a dimension n (or the number of samples n), the regression unit 241 outputs time series data (u(t), u(t-1), u(t-2), . . . , u(tn-1)) at time t. In this way, the regression unit 241 outputs time series data for the number of samples n of the input value u (input data) to the function unit 243.

[0106] The regression unit 242 corresponds to a second regression unit, and outputs time series data including past values ​​defined by the hyperparameters 245 for the predicted value y. For example, when the hyperparameters 245 include a dimension m (or the number of samples m), the regression unit 242 outputs time series data (y(t-1), y(t-2), . . . , y(tm)) at time t. In this way, the regression unit 242 outputs time series data for the number m of samples of the predicted value y to the function unit 243.

[0107] The hyperparameter 244 (dimension m / number of samples m) and the hyperparameter 245 (dimension n / number of samples n) are optimized in the learning. That is, the process of optimizing the hyperparameters of the noise removal model 24 (step S54 in FIG. 7) may include a process of optimizing at least one of the dimension m (number of samples m) and the dimension n (number of samples n).

[0108] The function unit 243 inputs the time series data for the input value u and the time series data for the predicted value y to an output function defined in accordance with the hyperparameters 246, and calculates the predicted value y at each time t.

[0109] In the nonlinear ARX model, the function section 243 may include any function f as shown below.

[0110]

number

[0111] The hyperparameters 246 include any parameters (function formula and coefficients) that define the function f. The hyperparameters 246 are optimized during training.

[0112] In a linear ARX model, the function portion 243 may include a linear function such as:

[0113]

number

[0114] The hyperparameters 246 are the coefficients a1, a2,...a that define the linear function m ,b1,b2,···b n The coefficients included in the hyperparameters 246 are optimized so that the following relationship holds:

[0115]

number

[0116] Here, e(t) represents an arbitrary offset (error). In this way, the function unit 243 of the noise removal model 24 may include at least one of a nonlinear function and a linear function. The process of optimizing hyperparameters of the noise removal model 24 (step S54 in FIG. 7) may include a process of optimizing coefficients that define at least one of the nonlinear function and the linear function.

[0117] As shown in FIG. 10, by adopting a noise removal model 24 with one channel of input, the time required for the learning process (optimization of hyperparameters) can be reduced.

[0118] [H. Performance Evaluation] Next, an example of performance evaluation of the noise removal model 24 used in the noise removal technique according to the present embodiment will be described.

[0119] (h1: Noise removal performance) 11 is a graph showing an example of the noise removal performance of the noise removal model 24 according to the present embodiment. The graph shown in FIG. 11 shows an example of performance evaluation of the noise removal model 24 having a nonlinear ARX model network structure. The number of hidden units in the target noise removal model 24 is 5, and the dimension of the regression part is 20.

[0120] 11, the predicted value y of the noise removal model 24 predicts the correct data sufficiently. Note that the prediction error (the difference between the correct data and the predicted value y: residual) follows a normal distribution.

[0121] (h2: Generalization performance of the model) Fig. 12 is a graph showing a case where the noise removal model 24 according to the present embodiment is applied to electroencephalograms measured on different days from the same subject. Fig. 12 shows the result of applying the noise removal model 24 constructed using electroencephalograms measured from a given subject to electroencephalograms measured from the same subject on different days.

[0122] 12, it can be said that the predicted value y of the noise removal model 24 sufficiently predicts the correct data. In other words, it can be seen that the noise removal model 24 configured based on the electroencephalogram measured from a given subject can obtain sufficient prediction accuracy (noise removal performance) for another electroencephalogram measured from the same subject.

[0123] Fig. 13 is a graph showing a case where the noise removal model 24 according to the present embodiment is applied to electroencephalograms measured from different subjects. Fig. 13 shows the result of applying the noise removal model 24 configured using electroencephalograms measured from one subject to electroencephalograms measured from another subject.

[0124] 13, the predicted value y of the noise removal model 24 sufficiently predicts the correct data. That is, it is found that the noise removal model 24 configured based on the electroencephalogram measured from one subject can obtain sufficient prediction accuracy (noise removal performance) for the electroencephalogram measured from another subject.

[0125] (h3: Channel generalization performance) In the above explanation, an example was described in which the noise removal model 24 was constructed using all channels included in the correct answer data. However, an example of performance evaluation of the generalization performance when the noise removal model 24 is constructed using only one channel will be shown.

[0126] Fig. 14 is a graph showing an example of the generalization performance of the channel of the noise removal model 24 according to the present embodiment. Fig. 14 shows electroencephalograms measured from a phantom (pseudo model) placed in an MRI using two types of MRI measurement devices.

[0127] 14(A) and 14(B) show example results when the first MRI measurement device is used, and FIG. 14(C) and 14(D) show example results when the second MRI measurement device is used.

[0128] In addition, Figures 14(A) and 14(C) show example results when no fMRI scan is performed using an MRI measurement device, and Figures 14(B) and 14(D) show example results when an fMRI scan is performed using an MRI measurement device.

[0129] 14(A) to 14(D), the vertical axis indicates the channel number of the ground truth data used to configure the noise removal model 24, and the horizontal axis indicates the channel number of the test data for performance evaluation. Each graph also indicates the prediction error (the difference between the ground truth data and the predicted value y) for the test data with shading, with lighter areas indicating smaller prediction errors.

[0130] According to the graphs shown in each of Figures 14(A) to 14(D), although there are some channel combinations with large errors, there are many channel combinations that can generally tolerate errors, which shows the possibility of generating a highly generalizable noise removal model 24 even with a small number of channels.

[0131] Fig. 15 is a graph showing another example of the generalization performance of the channels of the noise removal model 24 according to the present embodiment. Fig. 15(A) and Fig. 15(B) show the prediction error (the difference between the correct data and the predicted value y) when the noise removal model 24 is constructed using all channels included in the correct data and then one channel's worth of test data is input. The horizontal axis of Fig. 15(A) and Fig. 15(B) indicates the channel number of the test data for performance evaluation. In Fig. 15(A) and Fig. 15(B), the lighter the color of the area, the smaller the prediction error.

[0132] FIG. 15(A) shows an example of the results when the first MRI measurement device is used, and FIG. 15(B) shows an example of the results when the second MRI measurement device is used.

[0133] 15(A) and 15(B) show examples of evaluation of whether or not to perform fMRI scanning using an MRI measurement device, for combinations of four types of ground truth data and test data, as shown below.

[0134] (1) Training data (without fMRI scan), Test data (with fMRI scan) (2) Training data (no fMRI scans performed), Test data (no fMRI scans performed) (3) Training data (with fMRI scan), Test data (with fMRI scan) (4) Training data (with fMRI scans), test data (without fMRI scans) As shown in Figures 15(A) and 15(B), for both MRI measurement devices, the prediction error is smallest in the combination of (2). In other words, it is considered preferable to generate correct data from EEG measured in a state where fMRI scanning is not performed by the MRI measurement device.

[0135] As shown in Figures 15(A) and 15(B), for combination (2), the prediction error is sufficiently small for all channels included in the test data, and it can be said that the noise removal model 24 has high generalization even when the noise removal model 24 is constructed using all channels of the ground truth data.

[0136] (h4: metastasis) A noise removal model 24 may be constructed using a plurality of MRI measurement devices, and then an integrated noise removal model 24 may be constructed using a known technique such as transfer learning. For example, a noise removal model 24 constructed for a first MRI measurement device and a noise removal model 24 constructed for a second MRI measurement device may be combined to generate a new noise removal model 24 that can be applied to both MRI measurement devices.

[0137] [I. Modifications] For example, when the noise removal model 24 is prepared for each MRI measurement device, the noise removal model 24 may be stored in an arbitrary distribution server and may be made downloadable as necessary. By providing a mechanism for arbitrarily distributing the noise removal model 24, noise removal similar to that of the noise removal library (CWL regression) can be realized even at a site that does not have an EEG cap with CWL and necessary measurement devices.

[0138] [J.Advantages] According to this embodiment, even in the case of electroencephalograms measured using an electroencephalogram cap without a CWL, noise caused by the MRI measurement device or the movement of the subject during measurement can be removed.

[0139] The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present invention is defined by the claims, not by the description of the embodiments described above, and is intended to include all modifications within the meaning and scope of the claims. [Explanation of symbols]

[0140] 1,2 EEG cap, 4 Measurement system, 10,12,20,22 EEG, 14 Noise removal library, 16 CWL signal, 24 Noise removal model, 100 Information processing device, 102 Processor, 104 Main memory unit, 106 Control interface, 108 Network interface, 110 Input unit, 112 Display unit, 120 Secondary memory unit, 121,123,124 EEG time series data, 122 CWL time series data, 125 Scanning artifact removal program, 126 Noise removal program, 127 Learning program, 152 Scanning artifact removal unit, 154 Learning unit, 200 EEG measuring device, 202 Multiplexer, 204 Noise filter, 206 A / D converter, 208 Memory unit, 210 Interface, 241,242 Regression unit, 243 Function part, 244, 245, 246 Hyperparameters, 300 MRI measurement device.

Claims

1. A step of measuring the electroencephalogram and the carbon wire loop signal of a subject placed under an artificial magnetic field using an electroencephalogram cap to which a carbon wire loop is attached; removing noise from the electroencephalogram of the subject or from the electroencephalogram after removing artifacts from the electroencephalogram of the subject based on the carbon wire loop signal; and optimizing hyperparameters of a noise removal model using, as input data, time series data of electroencephalograms before noise is removed based on the carbon wire loop signal, and using, as answer data, time series data of electroencephalograms after noise is removed based on the carbon wire loop signal.

2. The noise removal model is A function part for calculating a predicted value; a first regression unit that outputs time series data of a first number of samples of the input data to the function unit; a second regression unit that outputs to the function unit time series data for a second number of samples of the predicted value; The information processing method according to claim 1 , wherein optimizing the hyperparameters of the denoising model includes optimizing at least one of the first number of samples and the second number of samples.

3. The function portion of the time series data includes at least one of a nonlinear function and a linear function, The information processing method according to claim 2 , wherein optimizing hyper-parameters of the noise removal model includes optimizing coefficients that define at least one of the nonlinear function and the linear function.

4. the noise removal model is configured to receive an input of one channel of electroencephalogram and to output one channel of electroencephalogram; Optimizing the hyper-parameters of the denoising model includes: a step of sequentially arranging each of a plurality of channels included in the input data in series to make it one-dimensional; The information processing method according to any one of claims 1 to 3, further comprising the step of sequentially arranging each of a plurality of channels included in the supervised data in series to make it one-dimensional.

5. A step of measuring the electroencephalogram of a subject placed under an artificial magnetic field using an electroencephalogram cap; and inputting the electroencephalogram of the subject or the electroencephalogram after removing artifacts from the electroencephalogram of the subject into a trained model constructed by a training process to remove noise. The learning process includes: A step of measuring the electroencephalogram and the carbon wire loop signal of a subject placed under an artificial magnetic field using an electroencephalogram cap to which a carbon wire loop is attached; removing noise from the electroencephalogram of the subject or from the electroencephalogram after removing artifacts from the electroencephalogram of the subject based on the carbon wire loop signal; and optimizing hyperparameters using, as input data, time series data of electroencephalograms before noise is removed based on the carbon wire loop signal, and using, as answer data, time series data of electroencephalograms after noise is removed based on the carbon wire loop signal.

6. The trained model is configured to receive an input of one channel of electroencephalogram and output one channel of electroencephalogram, 6. The information processing method according to claim 5, wherein the step of removing noise includes a step of repeating a process of inputting one channel of electroencephalogram to the trained model and calculating time series data of a predicted value a number of times equal to the number of channels.

7. The trained model is a function unit including at least one of a nonlinear function and a linear function and calculating a predicted value; a first regression unit that outputs time series data of a first number of samples of the input data to the function unit; The information processing method according to claim 5 , further comprising: a second regression section that outputs time series data for a second number of samples of the predicted value to the function section.

8. An information processing program for causing a computer to execute the information processing method according to any one of claims 1 to 7.

9. A means for storing the EEG and carbon wire loop signals of a subject placed under an artificial magnetic field, the EEG and carbon wire loop signals being measured using an EEG cap with a carbon wire loop attached thereto; means for removing noise from the electroencephalogram of the subject or from the electroencephalogram after removing artifacts from the electroencephalogram of the subject based on the carbon wire loop signal; and means for optimizing hyperparameters of a noise removal model using, as input data, time series data of electroencephalograms before noise is removed based on the carbon wire loop signal, and using, as answer data, time series data of electroencephalograms after noise is removed based on the carbon wire loop signal.

10. A means for storing the electroencephalogram of a subject placed under an artificial magnetic field, the electroencephalogram being measured using an electroencephalogram cap; a means for inputting the electroencephalogram of the subject or the electroencephalogram after removing artifacts from the electroencephalogram of the subject into a trained model to remove noise; The trained model is an information processing device configured by optimizing hyperparameters using electroencephalograms and carbon wire loop signals of a subject placed under an artificial magnetic field, measured using an electroencephalogram cap with a carbon wire loop attached, with time series data of electroencephalograms before noise is removed based on the carbon wire loop signal as input data and time series data of electroencephalograms after noise has been removed based on the carbon wire loop signal as correct answer data.

Images