JP2022061484A - MR signal processing device - Google Patents

Abstract

To improve accuracy of MRS signals.SOLUTION: An MR signal processing device comprises a parameter output unit. The parameter output unit applies a learned model to a plurality of MRS signals collected via MR spectroscopy with respect to the same object to output a plurality of parameters for MRS reconfiguration.SELECTED DRAWING: Figure 1

Description

Embodiments disclosed herein and in the drawings relate to MR signal processing equipment.

Data acquisition by MR Spectroscopy (MRS: Magnetic Resonance Spectroscopy) is known to be extremely unstable. Therefore, the number of integrations is set to about 50 to 100, and the MRS signals are averaged by adding the MRS signals collected at each time. However, the average MRS signal is also distorted due to the body movement of the subject, the disturbance of the static magnetic field, and the like during a plurality of data acquisitions.

A technique for generating an MRS signal by least squares fitting is known. An artificial spectrum of about 20 metabolites or the like is generated by a physical model, and a spectrum parameter such as a signal intensity value is added to each artificial spectrum and fitted to generate an MRS spectrum to be measured. A technique is also known in which a deep neural network expresses a regression process for obtaining spectral parameters for each artificial spectrum from a single MRS spectrum.

Nima Hatami, Michael Sdika, Helene Ratiney, “Magnetic Resonance Spectoroscopy Quantification using Deep Learning”, arXiv: 1806.07237v1 [cs.CV] 19 Jun 2018.

One of the problems to be solved by the embodiments disclosed in the present specification and the drawings is to improve the accuracy of the MRS signal. However, the problems to be solved by the embodiments disclosed in the present specification and the drawings are not limited to the above problems. The problem corresponding to each effect by each configuration shown in the embodiment described later can be positioned as another problem.

The MR signal processing apparatus according to the embodiment has a parameter output unit. The parameter output unit applies the trained model to a plurality of MRS signals collected by MR spectroscopy for the same object, and outputs a plurality of parameters for MRS reconstruction.

FIG. 1 is a diagram showing a configuration example of the magnetic resonance imaging apparatus according to the first embodiment. FIG. 2 is a diagram schematically showing the input / output relationship of the trained model (material amount estimation NN) according to the first embodiment. FIG. 3 is a diagram schematically showing an example of an MRS spectrum. FIG. 4 is a diagram showing an example of a signal processing flow by the MR signal processing apparatus according to the first embodiment. FIG. 5 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the first embodiment. FIG. 6 is a diagram schematically showing the generation process of the synthetic MRS spectrum in step SA3 of FIG. FIG. 7 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the modification 1 of the first embodiment. FIG. 8 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the second embodiment of the first embodiment. FIG. 9 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the modification 3 of the first embodiment. FIG. 10 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the modified example 4 of the first embodiment. FIG. 11 is a diagram showing a network configuration example of the first integrated model according to the modified example 5 of the first embodiment. FIG. 12 is a diagram showing a network configuration example of another first integrated model according to the modified example 5 of the first embodiment. FIG. 13 is a diagram showing a configuration example of the magnetic resonance imaging apparatus according to the second embodiment. FIG. 14 is a diagram schematically showing the input / output relationship of the trained model (basal component amount estimation NN) according to the second embodiment. FIG. 15 is a diagram showing an example of a flow of signal processing by the MR signal processing apparatus according to the second embodiment. FIG. 16 is a diagram schematically showing the input / output relationship of the basal component amount estimation NN according to the second embodiment. FIG. 17 is a diagram schematically showing a normal or abnormal determination process (determination result = normal) performed in step SB4 of FIG. 15 based on the basal component amount information. FIG. 18 is a diagram schematically showing a normal or abnormal determination process (determination result = abnormality) performed in step SB4 of FIG. 15 based on the basal component amount information. FIG. 19 is a diagram showing an example of a display screen of the synthesized MRS spectrum and the determination result displayed in step SB5 of FIG. FIG. 20 is a diagram showing a configuration example of the magnetic resonance imaging apparatus according to the third embodiment. FIG. 21 is a diagram schematically showing the input / output relationship of the trained model (modulation amount estimation NN) according to the third embodiment. FIG. 22 is a diagram showing an example of a flow of signal processing by the MR signal processing apparatus according to the third embodiment. FIG. 23 is a diagram schematically showing a specific example of the input / output relationship of the modulation amount estimation NN according to the third embodiment. FIG. 24 is a diagram schematically showing the correction process and the spectrum generation process in step SC3 and step SC4 of FIG. 22. FIG. 25 is a diagram showing a configuration example of the magnetic resonance imaging apparatus according to the fourth embodiment. FIG. 26 is a diagram showing a network configuration example of the second integrated model according to the fourth embodiment. FIG. 27 is a diagram showing a network configuration example of another second integrated model according to the fourth embodiment. FIG. 28 is a diagram schematically showing the input / output relationship of the trained model (modulation amount estimation NN) according to the fifth embodiment. FIG. 29 is a diagram schematically showing a specific example of the input / output relationship of the modulation amount estimation NN according to the fifth embodiment. FIG. 30 is a diagram schematically showing a further specific example of the input / output relationship of the modulation amount estimation NN shown in FIG. 29. FIG. 31 is a diagram schematically showing the input / output relationship of the trained model NN of the sixth embodiment. FIG. 32 is a diagram schematically showing a processing example according to the seventh embodiment. FIG. 33 is a diagram schematically showing a processing example according to the eighth embodiment. FIG. 34 is a diagram schematically showing a processing example according to the ninth embodiment. FIG. 35 is a diagram showing a configuration example of a magnetic resonance imaging system including the MR signal processing device according to the tenth embodiment.

Hereinafter, embodiments of the MR signal processing apparatus will be described in detail with reference to the drawings.

The MR signal processing device according to the present embodiment is a computer that processes the MR signal collected by the magnetic resonance imaging device. The MR signal processing device may be a computer incorporated in the magnetic resonance imaging device, or may be a computer separate from the magnetic resonance imaging device. Hereinafter, the MR signal processing apparatus will be described separately in several embodiments.

(First Embodiment)
FIG. 1 is a diagram showing a configuration example of the magnetic resonance imaging device 1 according to the first embodiment. As shown in FIG. 1, the magnetic resonance imaging device 1 includes a gantry 11, a sleeper 13, a gradient magnetic field power supply 21, a transmission circuit 23, a reception circuit 25, a sleeper drive device 27, a sequence control circuit 29, and an MR signal processing device (host computer). ) Has 50.

The gantry 11 has a static magnetic field magnet 41 and a gradient magnetic field coil 43. The static magnetic field magnet 41 and the gradient magnetic field coil 43 are housed in the housing of the gantry 11. A hollow bore is formed in the housing of the gantry 11. A transmission coil 45 and a reception coil 47 are arranged in the bore of the gantry 11.

The static magnetic field magnet 41 has a hollow substantially cylindrical shape, and generates a static magnetic field inside the substantially cylindrical shape. As the static magnetic field magnet 41, for example, a permanent magnet, a superconducting magnet, a normal conducting magnet, or the like is used. Here, the central axis of the static magnetic field magnet 41 is defined as the Z axis, the axis perpendicular to the Z axis is defined as the Y axis, and the axis horizontally orthogonal to the Z axis is defined as the X axis. The X-axis, Y-axis, and Z-axis form an orthogonal three-dimensional coordinate system.

The gradient magnetic field coil 43 is a coil unit attached to the inside of the static magnetic field magnet 41 and formed in a hollow substantially cylindrical shape. The gradient magnetic field coil 43 generates a gradient magnetic field by receiving a current supplied from the gradient magnetic field power supply 21. More specifically, the gradient magnetic field coil 43 has three coils corresponding to X-axis, Y-axis, and Z-axis orthogonal to each other. The three coils form a gradient magnetic field in which the magnetic field strength changes along each of the X-axis, Y-axis, and Z-axis. The gradient magnetic fields along the X-axis, Y-axis, and Z-axis are combined to form slice-selective gradient magnetic fields Gs, phase-encoded gradient magnetic fields Gp, and frequency-encoded gradient magnetic fields Gr that are orthogonal to each other in desired directions. The slice selection gradient magnetic field Gs is arbitrarily used to determine the imaging cross section (slice). The phase-encoded gradient magnetic field Gp is used to change the phase of a magnetic resonance signal (hereinafter referred to as MR signal) according to a spatial position. The frequency-encoded gradient magnetic field Gr is used to change the frequency of the MR signal depending on the spatial position. In the following description, it is assumed that the inclination direction of the slice selection gradient magnetic field Gs is the Z axis, the inclination direction of the phase-encoded gradient magnetic field Gp is the Y-axis, and the inclination direction of the frequency-encoded gradient magnetic field Gr is the X-axis.

The gradient magnetic field power supply 21 supplies a current to the gradient magnetic field coil 43 according to the sequence control signal from the sequence control circuit 29. The gradient magnetic field power supply 21 supplies a current to the gradient magnetic field coil 43 to generate a gradient magnetic field along each of the X-axis, Y-axis, and Z-axis by the gradient magnetic field coil 43. The gradient magnetic field is superimposed on the static magnetic field formed by the static magnetic field magnet 41 and applied to the subject P.

The transmission coil 45 is arranged inside the gradient magnetic field coil 43, for example, and receives a current supply from the transmission circuit 23 to generate a high frequency pulse (hereinafter referred to as an RF pulse).

The transmission circuit 23 supplies a current to the transmission coil 45 in order to apply an RF pulse for exciting the target proton existing in the subject P to the subject P via the transmission coil 45. The RF pulse vibrates at a resonance frequency inherent in the target proton, exciting the target proton. An MR signal is generated from the excited target proton and detected by the receiving coil 47. The transmission coil 45 is, for example, a whole body coil (WB coil). The whole body coil may be used as a transmission / reception coil.

The receiving coil 47 receives the MR signal emitted from the target proton existing in the subject P under the action of the RF pulse. The receiving coil 47 has a plurality of receiving coil elements capable of receiving MR signals. The received MR signal is supplied to the receiving circuit 25 via wired or wireless. Although not shown in FIG. 1, the receiving coil 47 has a plurality of receiving channels mounted in parallel. The receiving channel includes a receiving coil element that receives the MR signal, an amplifier that amplifies the MR signal, and the like. The MR signal is output for each reception channel. The total number of receive channels and the total number of receive coil elements may be the same, and the total number of receive channels may be larger or smaller than the total number of receive coil elements.

The receiving circuit 25 receives the MR signal generated from the excited target proton via the receiving coil 47. The receiving circuit 25 processes the received MR signal to generate a digital MR signal. The digital MR signal can be expressed in k-space defined by the spatial frequency. Therefore, hereinafter, the digital MR signal will be referred to as k-space data. The k-space data is an example of an MR collected signal. The k-space data is supplied to the host computer 50 via wired or wireless.

The transmission coil 45 and the reception coil 47 are merely examples. Instead of the transmitting coil 45 and the receiving coil 47, a transmitting / receiving coil having a transmitting function and a receiving function may be used. Further, the transmission coil 45, the reception coil 47, and the transmission / reception coil may be combined.

A sleeper 13 is installed adjacent to the gantry 11. The sleeper 13 has a top plate 131 and a base 133. The subject P is placed on the top plate 131. The base 133 slidably supports the top plate 131 along each of the X-axis, Y-axis, and Z-axis. The sleeper drive device 27 is housed in the base 133. The sleeper drive device 27 moves the top plate 131 under the control of the sequence control circuit 29. The sleeper drive device 27 may include, for example, any motor such as a servo motor or a stepping motor.

The sequence control circuit 29 has a processor of a CPU (Central Processing Unit) or an MPU (Micro Processing Unit) and a memory such as a ROM (Read Only Memory) or a RAM (Random Access Memory) as hardware resources. The sequence control circuit 29 synchronously controls the gradient magnetic field power supply 21, the transmission circuit 23, and the reception circuit 25 based on the data collection conditions set by the condition setting function 512 of the processing circuit 51A, and corresponds to the data collection conditions. Data collection is applied to the subject P, and k-space data related to the subject P is collected. The sequence control circuit 29 is an example of a sequence control unit.

The sequence control circuit 29 according to the present embodiment executes data acquisition for MR spectroscopy, which is a kind of chemical shift measurement. Chemical shift measurement is a technique for measuring chemical shift, which is a minute difference in the resonance frequency of a target proton such as a hydrogen nucleus, which occurs in response to a difference in the chemical environment. MR spectroscopy includes a single voxel method for collecting data for a single voxel and a multi-voxel method for collecting data for a plurality of voxels, and this embodiment can be applied to any method. The multi-voxel method is also called chemical shift imaging (CSI: Chemical Shift Imaging) or MRS imaging (MRSI: MRS Imaging). The voxel to be measured is called a voxel of interest.

The sequence control circuit 29 executes data acquisition for MR spectroscopy for the subject P. By performing data acquisition for MR spectroscopy, a free induction decay (FID) signal or spin echo signal is generated from the voxel of interest of subject P. The receiving circuit 25 receives the FID signal or the spin echo signal via the receiving coil 47, processes the received FID signal or the spin echo signal, and collects k-space data regarding the boxel of interest. It is assumed that the collected k-space data is digital data representing the signal strength value emitted from the voxel of interest as a time function. The pulse sequence for MR spectroscopy is repeated for the number of integrations (NEX: number of excitation), and k-space data for the number of integrations is collected. Hereinafter, the k-space data collected by MR spectroscopy will be referred to as MRSk data. The MRSk data is an example of an MRS signal.

As shown in FIG. 1, the MR signal processing device 50 is a computer having a processing circuit 51A, a memory 53, a display 55, an input interface 57, and a communication interface 59. The processing circuit 51A is an example of a processing unit, the memory 53 is an example of a storage unit, the display 55 is an example of a display unit, the input interface 57 is an example of an input unit, and the communication interface 59 is an example of a communication unit. Is.

The processing circuit 51A has a processor such as a CPU as a hardware resource. The processing circuit 51A functions as the center of the magnetic resonance imaging device 1. For example, the processing circuit 51A realizes an acquisition function 511, a condition setting function 512, a signal processing function 513, a substance amount output function 514, a synthesis function 515, a learning function 516, and a display control function 517 by executing various programs. The acquisition function 511 is an example of an acquisition unit, the condition setting function 512 is an example of a setting unit, the signal processing function 513 is an example of a spectrum generation unit, and the substance amount output function 514 is an example of a parameter output unit. The synthesis function 515 is an example of the synthesis unit, the learning function 516 is an example of the learning unit, and the display control function 517 is an example of the display unit.

In the acquisition function 511, the processing circuit 51A acquires various MRS signals. For example, the processing circuit 51A acquires MRSk data collected by the sequence control circuit 29. The processing circuit 51A may directly acquire MRSk data from the sequence control circuit 29 or the reception circuit 25, or may acquire MRSk data once stored in the memory 53 from the memory 53.

In the condition setting function 512, the processing circuit 51A automatically or manually sets the data acquisition condition. Specifically, in the present embodiment, data collection conditions related to MR spectroscopy are set. Data acquisition conditions related to MR spectroscopy include, for example, condition items such as pulse sequence, repetition time (TR), echo time (TE), number of integrations, spectrum width, number of samples, data acquisition method, and region selection pulse. As the pulse sequence, for example, PRESS (point resolved spectroscopy), STEAM (stimulated echo acquisition mode) and the like are known. If the TR is a long TR, it may be set to, for example, 5000 ms or more, and if it is a short TR, it may be set to, for example, about 1000 to 3000 ms. The longer the TR, the closer the obtained MR signal strength value is to the true value, but the longer the data acquisition time. The TE may be set to about 100 to 300 ms for a long TE and about 20 to 100 ms for a short TE. The shorter the TE, the larger the number of peaks and the higher the accuracy of the MRS spectrum, and the longer the TE, the smaller the number of peaks and the better the visibility of the MRS spectrum.

The number of integrations is not particularly limited and may be set to 1 or more. The number of spectral widths and the number of samplings are conditional items related to spectral resolution. The number of spectrum widths and the number of samplings may be set to arbitrary values. As described above, as the data collection method, there are a single voxel method in which an MRS spectrum of one voxel is obtained by one data collection and a multi-voxel method in which an MRS spectrum of each of a plurality of voxels is obtained by one data collection. The region selection pulse includes a pulse for exciting a hydrogen nucleus only in a set region and a pulse for not exciting a hydrogen nucleus limited to a set region. As the data collection conditions, the presence / absence of application of the region selection pulse, the frequency information of the selected region, and the like are set.

In the signal processing function 513, the processing circuit 51A performs various signal processing on the MRS signal acquired by the acquisition function 511. For example, the processing circuit 51A generates a spectrum (hereinafter, referred to as an MRS spectrum) showing the signal intensity for each chemical shift based on the MRSk data. The MRS spectrum is an example of an MRS signal.

In the material amount output function 514, the processing circuit 51A applies a trained model to a plurality of MRS signals obtained by MR spectroscopy with respect to the same measurement target site of the subject P, and sets a plurality of parameters for MRS reconstruction. Output. The plurality of parameters according to the first embodiment are spectral parameters for fitting based on a plurality of spectral models corresponding to the plurality of parameters. The plurality of spectrum models are a plurality of artificial spectra corresponding to the plurality of substances contained in the measurement target site of the subject P. The plurality of spectral parameters include the amount of substance information of the plurality of substances regarding the measurement target site of the subject P. Amount of substance information means information about the amount of substance of the substance. Hereinafter, the parameter according to the first embodiment will be referred to as substance amount information. The trained model is a machine learning model trained to output substance amount information by inputting a plurality of MRS signals. Neural networks and deep neural networks are used as machine learning models. Hereinafter, the trained model according to the first embodiment will be referred to as a substance amount estimation NN.

The processing circuit 51A in the synthesis function 515 generates an MRS spectrum for the subject P by fitting based on the plurality of spectral parameters output in the substance amount output function 514 and the plurality of spectral models. The MRS spectrum generated by the synthesis function 515 will be referred to as a synthetic MRS spectrum.

In the learning function 516, the processing circuit 51A generates the substance amount estimation NN used in the substance amount output function 514 by machine learning based on the learning data.

In the display control function 517, the processing circuit 51A displays various information on the display 55. For example, the processing circuit 51A displays the MRS spectrum, the substance amount information, the synthetic MRS spectrum, the data acquisition condition setting screen, and the like on the display 55.

The memory 53 is a storage device such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), or an integrated circuit storage device that stores various information. Further, the memory 53 may be a drive device or the like that reads and writes various information to and from a portable storage medium such as a CD-ROM drive, a DVD drive, and a flash memory. For example, the memory 53 stores a substance amount estimation NN, data acquisition conditions, an MRS signal, a control program, and the like.

The display 55 displays various information by the display control function 517. As the display 55, for example, a CRT display, a liquid crystal display, an organic EL display, an LED display, a plasma display, or any other display known in the art can be appropriately used.

The input interface 57 includes an input device that receives various commands from the user. As input devices, keyboards, mice, various switches, touch screens, touch pads, etc. can be used. The input device is not limited to a device equipped with physical operation parts such as a mouse and a keyboard. For example, an electric signal processing circuit that receives an electric signal corresponding to an input operation from an external input device provided separately from the magnetic resonance imaging device 1 and outputs the received electric signal to various circuits is also input. Included in the example of interface 57. Further, the input interface 57 may be a voice recognition device that converts a voice signal collected by the microphone into an instruction signal.

The communication interface 59 connects the magnetic resonance imaging device 1 with a workstation, a PACS (Picture Archiving and Communication System), a HIS (Hospital Information System), a RIS (Radiology Information System), etc. via a LAN (Local Area Network) or the like. It is an interface to connect. The network IF sends and receives various information to and from the connected workstation, PACS, HIS, and RIS.

The above configuration is an example and is not limited to this. For example, the sequence control circuit 29 may be incorporated in the MR signal processing device 50. Further, the sequence control circuit 29 and the processing circuit 51A may be mounted on the same board. Further, the condition setting function 512 does not necessarily have to be mounted on the processing circuit 51A of the magnetic resonance imaging device 1. For example, the condition setting function 512 may be mounted on a computer for setting data acquisition conditions, which is separate from the magnetic resonance imaging device 1. In this case, the data acquisition conditions generated by the computer are supplied to the magnetic resonance imaging apparatus 1 via a network, a portable recording medium, or the like. Further, the storage area of the data acquisition condition in the memory 53 does not need to be mounted on the magnetic resonance imaging device 1, and is mounted on, for example, a storage device connected to the magnetic resonance imaging device 1 via a network. May be good.

Hereinafter, an operation example of the MR signal processing apparatus 50 according to the first embodiment will be described in detail.

As described above, the processing circuit 51A applies the substance amount estimation NN to the plurality of MRS signals by the substance amount output function 514, and outputs the substance amount information of each substance contained in the measurement target portion of the plurality of MRS signals.

FIG. 2 is a diagram schematically showing the input / output relationship of the trained model (material amount estimation NN) according to the first embodiment. As shown in FIG. 2, the substance amount estimation NN takes the first MRS signal and the second MRS signal as inputs, and each substance included in the measurement target portion of the first MRS signal and the second MRS signal. It is a machine learning model in which the parameters are trained to output the substance amount information about. The first MRS signal and the second MRS signal are MRS signals targeting the same subject P. The data acquisition conditions of the first MRS signal and the second MRS signal may be the same or different. As the MRS signal input to the material amount estimation NN, MRSk data or MRS spectrum can be used. The number of MRS signals input to the substance amount estimation NN shown in FIG. 2 is assumed to be two, but this is an example, and any number may be used as long as it is two or more.

Amount of substance information is information on the amount of substance of a substance discriminated by MR spectroscopy. The substance to be discriminated is not particularly limited, and any substance that can be contained in the measurement target site can be discriminated. The substance to be discriminated varies depending on the measurement target site, but various molecules such as propylene glycol, ethanol, acetate, and acetone are known. The amount of substance may be a numerical value indicating the signal strength, a class according to the signal strength, or an arbitrary analysis value calculated based on the signal strength. The amount of substance takes a discrete value including 0 or a continuous value. The amount of substance corresponds to the spectral parameters given to the artificial spectrum for the fitting of the artificial spectrum, which is the spectral model corresponding to the substance.

FIG. 3 is a diagram schematically showing an example of an MRS spectrum. The MRS spectrum shown in FIG. 3 is an example of the MRS spectrum of the head. As shown in FIG. 3, in the MRS spectrum, the vertical axis is defined by the MR signal intensity value [AU (Arbitrary Unit)], and the horizontal axis is the difference from the reference frequency, that is, the chemical shift [ppm (parts per million)]. It is stipulated. The reference frequency is set to the frequency of an arbitrarily selected reference material. The reference substance is not particularly limited, but is set to, for example, tetramethylsilane (TMS). According to the MRS spectrum, it is possible to visualize the amount of substance of various substances such as propylene glycol, ethanol, acetate and acetone present in the voxel of interest.

The MRS spectrum according to the present embodiment may be waveform data of the MRS spectrum, or may be numerical data of a combination of a substance identifier (name or symbol of the substance) corresponding to the peak and the amount of substance. Numerical data of the combination of the frequency difference corresponding to the peak and the amount of substance may be used. Further, the substance amount information may include reliability information regarding the substance amount in addition to the substance amount. The confidence information corresponds to the full width at half maximum of the peak contained in the MRS spectrum. As the half-value width, for example, a half-value full width, a half-value half-width, or the like may be used.

The substance amount estimation NN is generated by the learning function 516 of the processing circuit 51A. The processing circuit 51A trains a machine learning model based on a plurality of training samples and trains the model to generate an amount of substance estimation NN. The training sample is a combination of the first MRS signal and the second MRS signal, which are input data, and the substance amount information (hereinafter, referred to as correct substance amount information), which is the correct answer data. The first MRS signal and the second MRS signal, which are input data, are generated by the magnetic resonance imaging device 1 or another magnetic resonance imaging device. As the correct substance amount information, the substance amount information obtained by signal analysis based on the input data is used. The processing circuit 51A applies a machine learning model to the first MRS signal and the second MRS signal to perform forward propagation processing, and outputs substance amount information (hereinafter referred to as estimated substance amount information). Next, the processing circuit 51A applies the difference (error) between the estimated substance amount information and the correct substance amount information to the machine learning model to perform back propagation processing, and performs back propagation processing, and the gradient which is the differential coefficient of the error function which is a function of the parameter. Calculate the vector. Next, the processing circuit 51A updates the parameters of the machine learning model based on the gradient vector. These forward propagation processing, back propagation processing, and parameter update processing are repeated while changing the training sample, and the parameters that minimize the error function are determined according to a predetermined optimization method. As a result, the amount of substance estimation NN is generated.

Next, the signal processing by the MR signal processing apparatus 50 according to the first embodiment will be described with reference to FIG. In the following description, it is assumed that the MRS signal input to the substance amount estimation NN is an MRS spectrum.

FIG. 4 is a diagram showing an example of the flow of signal processing by the MR signal processing device 50 according to the first embodiment.

As shown in FIG. 4, the processing circuit 51A acquires the first MRS spectrum and the second MRS spectrum by realizing the acquisition function 511 (step SA1). It is assumed that the first MRS spectrum and the second MRS spectrum are, for example, MRS spectra related to the same measurement target site collected under the same data acquisition conditions by the single voxel method. In MR spectroscopy, for example, an integration number of about 50 to 150 times is set as a data acquisition condition, and an MRS spectrum corresponding to the integration number is generated by the processing circuit 51A. The processing circuit 51A generates one MRS spectrum by performing addition processing such as addition or addition averaging on the MRS spectrum for the number of integrations. In step SA1, the processing circuit 51A selects and acquires two arbitrary MRS spectra as the first MRS spectrum and the second MRS spectrum from the MRS spectra corresponding to the number of integrations before the addition processing. The two MRS spectra are not particularly limited. It may be two MRS spectra having consecutive data acquisition times, or two MRS spectra separated in time. The two MRS spectra may be selected manually by the user according to instructions via the input interface 57 or automatically according to a predetermined algorithm.

The addition process may be applied to the MRSk data for the number of integrations. In this case, the processing circuit 51A selects two MRSk data from the MRSk data for the number of integrations before the addition process, generates two MRS spectra from the two selected MRSk data, and generates 2 The MRS spectra may be acquired as a first MRS spectrum and a second MRS spectrum.

When step SA1 is performed, the processing circuit 51A applies the substance amount estimation NN to the first MRS spectrum and the second MRS spectrum acquired in step SA1 by realizing the substance amount output function 514, and the substance amount information. Is output (step SA2).

FIG. 5 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN. As shown in FIG. 5, the substance amount estimation NN relates to a substance contained in the measurement target portion of the first MRS spectrum and the second MRS spectrum by inputting the first MRS spectrum and the second MRS spectrum. It is a machine learning model in which parameters are trained to output substance amount information. As for the amount of substance information, specifically, the amount of substance M1 and half price width W1 of propylene glycol, the amount of substance M2 and half price width W2 of ethanol, the amount of substance M3 and half price width W3 of acetate, the amount of substance M4 and half price width of acetone. W4 is exemplified. If there are multiple peaks for one substance, the amount of substance and half width are output for each peak. The peak may be identified by the frequency difference.

In the processing circuit 51A, the substance amount estimation NN is applied to the first MRS spectrum and the second MRS spectrum acquired in step SA1, and the substance amount M1 and half price width W1 of propylene glycol, the substance amount M2 of ethanol, and the substance amount M2 of ethanol are applied. The half price width W2, the amount of substance M3 of acetate and the half price width W3, the amount of substance M4 of acetone and the half price width W4 are output. According to the first embodiment, since the substance amount information can be estimated from a plurality of MRS spectra related to the same measurement target site, the substance amount information of the substance contained in the measurement target site can be estimated accurately.

When step SA2 is performed, the processing circuit 51A generates a synthetic MRS spectrum based on the substance amount information output in step SA2 by realizing the synthesis function 515 (step SA3).

FIG. 6 is a diagram schematically showing the generation process of the synthetic MRS spectrum in step SA3. As shown in FIG. 6, it is assumed that the substance amount information IM1 of propylene glycol, the substance amount information IM2 of ethanol, the substance amount information IM3 of acetate, and the substance amount information IM4 of acetone are output in step SA2. On the other hand, an artificial spectrum corresponding to each substance to be discriminated by the substance amount estimation NN is generated in advance and stored in the memory 53 or the like. The artificial spectrum is a theoretical MRS spectrum related to the substance, and is generated by a prediction calculation or the like using a physical model or the like. It can be said that the artificial spectrum is an ideal MRS spectrum without noise caused by unstable data acquisition of MR spectroscopy, body movement of subject P, disturbance of static magnetic field, and the like. In the example of FIG. 6, the artificial spectrum ES1 of propylene glycol, the artificial spectrum ES2 of ethanol, the artificial spectrum ES3 of acetate, and the artificial spectrum ES4 of acetone are generated in advance.

In step SA3, the processing circuit 51A generates a synthetic MRS spectrum based on the amount of substance and half width of each of the plurality of substances and the plurality of artificial spectra corresponding to the plurality of substances. More specifically, the processing circuit 51A applies the amount of substance and the full width at half maximum of each of the plurality of substances to the artificial spectrum of the substance to generate a plurality of corrected artificial spectra, and fitting based on the generated plurality of corrected artificial spectra. Generates a synthetic MRS spectrum.

Specifically, the processing circuit 51A applies the amount of substance M1 of propylene glycol and the half width W1 to the artificial spectrum ES1 to generate a corrected artificial spectrum of propylene glycol (step SA31). Similarly, a corrected artificial spectrum is generated for ethanol, acetate and acetone (step SA31). Next, the processing circuit 51A fits the corrected artificial spectrum of propylene glycol, the corrected artificial spectrum of ethanol, the corrected artificial spectrum of acetate, and the corrected artificial spectrum of acetone to generate a synthetic MRS spectrum FS1 (step SA32). As the fitting method, any fitting method such as least squares fitting may be used. Since the synthetic MRS spectrum FS1 is generated based on the ideal artificial spectra ES1 to ES4, the data collection of MR spectroscopy is unstable compared to the input first MRS spectrum and the second MRS spectrum. Noise caused by the body movement of the subject P, disturbance of the static magnetic field, and the like is reduced. In other words, the substance amount estimation by the substance amount estimation NN (step SA2) and the generation of the synthetic MRS spectrum based on the substance amount information (step SA3) are equated with the denoise processing of the first MRS spectrum and the second MRS spectrum. can do.

When step SA3 is performed, the processing circuit 51A displays the synthetic MRS spectrum generated in step SA3 by realizing the display control function 517 (step SA4). The synthetic MRS spectrum is displayed on the display 55. This allows the user to observe the synthetic MRS spectrum, which is the denoised MRS spectrum.

As a result, the signal processing by the MR signal processing device 50 according to the first embodiment is completed.

The first embodiment can be modified in various ways.

(Modification 1)
FIG. 7 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the modification 1 of the first embodiment. The substance amount estimation NN according to the first modification takes the first MRS spectrum and the second MRS spectrum as inputs, and the substance amount related to the substance contained in the measurement target portion of the first MRS spectrum and the second MRS spectrum. It is a machine learning model whose parameters are trained to output information. The first MRS spectrum is an MRS spectrum collected by MR spectroscopy by an unapplied pulse sequence of region selection pulses. The second MRS spectrum is an MRS spectrum collected by MR spectroscopy by a pulse sequence with a region selection pulse applied.

The application of the region selection pulse excites the hydrogen nuclei of the substance belonging to the specific frequency band, and a second MRS spectrum showing the chemical shift distribution of the signal intensity from the excited hydrogen nuclei is generated. Alternatively, the application of the region selection pulse does not excite the hydrogen nuclei of the substance belonging to a specific frequency band, and a second MRS spectrum showing the chemical shift distribution of the signal intensity from the hydrogen nuclei of the substance belonging to another frequency band is generated. Will be done. On the other hand, when the region selection pulse is not applied, the hydrogen nuclei of the substances belonging to all frequency bands are excited, and the first MRS spectrum showing the chemical shift distribution of the signal intensity from the hydrogen nuclei of the substances belonging to all frequency bands is generated. Will be done. Since the region selection pulse separates multiple peaks of overlapping chemical shifts (frequency differences) on the MRS spectrum due to different substances, the substance belonging to the above specific frequency band among these multiple substances is selectively excited or not selected. Used to excite.

By inputting the first MRS spectrum and the second MRS spectrum into the substance amount estimation NN, the substance amount estimation NN can recognize the specific substance belonging to the specific frequency band separately from other substances. It is possible to improve the estimation accuracy of the substance amount information of a specific substance.

(Modification 2)
FIG. 8 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the second embodiment of the first embodiment. The substance amount estimation NN according to the second modification is a substance included in the measurement target portion of the first MRS spectrum and the second MRS spectrum by inputting the first MRS spectrum, the second MRS spectrum, and the selection information. It is a machine learning model whose parameters are trained to output the amount of substance information about. The first MRS spectrum is an MRS spectrum collected by MR spectroscopy by an unapplied pulse sequence of region selection pulses. The second MRS spectrum is an MRS spectrum collected by MR spectroscopy by a pulse sequence with a region selection pulse applied. The selection information is information for identifying the region selected by the region selection pulse, and includes, for example, frequency information of a frequency band that is selectively excited or de-excited.

By inputting the first MRS spectrum, the second MRS spectrum, and the selection information into the substance amount estimation NN, it is possible to facilitate the separation of the specific substance belonging to the specific frequency band by the substance amount estimation NN, and the substance amount information of the specific substance. It is expected that the estimation accuracy of will be further improved.

(Modification 3)
FIG. 9 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the modification 3 of the first embodiment. The substance amount estimation NN according to the third modification takes the first MRS spectrum and the second MRS spectrum as inputs, and the substance amount related to the substance contained in the measurement target portion of the first MRS spectrum and the second MRS spectrum. It is a machine learning model whose parameters are trained to output information. The first MRS spectrum and the second MRS spectrum are MRS spectra in which the combinations of TR and TE, which are data acquisition conditions, are different from each other. TR and TE are largely involved in the SNR (Signal to Noise Ratio) of MR spectroscopy. The longer the TR, the closer the obtained signal strength value is to the true value, but the longer the data acquisition time. The shorter the TE, the larger the number of observed peaks and the higher the accuracy of the MRS spectrum, and the longer the TE, the smaller the number of observed peaks and the better the visibility of the MRS spectrum.

The first MRS spectrum is collected under the combination of the first TR and TE, and the second MRS spectrum is collected under the combination of the second TR and TE. In the first MRS spectrum and the second MRS spectrum, TR may be the same and TE may be different, TR may be different and TE may be the same, or both TR and TE may be different. An example of inputting two types of MRS spectra has been shown. For example, the MRS spectra are collected about 3 to 1000 times while changing the TE or TR each time, and a part or all of the obtained MRS spectra is a substance amount. It may be used as an input of an estimated NN.

By inputting the first MRS spectrum and the second MRS spectrum having different combinations of TR and TE into the substance amount estimation NN, the peak recognition performance by the substance amount estimation NN is improved, so that the estimation accuracy of the substance amount information is improved. Is expected to improve.

(Modification example 4)
FIG. 10 is a diagram schematically showing a specific example of the input / output relationship of the substance amount estimation NN according to the modified example 4 of the first embodiment. In the substance amount estimation NN according to the modified example 4, parameters are learned so as to output the substance amount information regarding the substance contained in the measurement target site of the MRS spectrum and the synthetic MRS spectrum by inputting the MRS spectrum and the synthetic MRS spectrum. It is a machine learning model. The MRS spectrum is an MRS spectrum generated based on the MRSk data by the signal processing function 513. The synthetic MRS spectrum is an MRS spectrum generated by the synthesis function 515. The synthetic MRS spectrum may or may not be generated based on the other input, the MRS spectrum.

By inputting the MRS spectrum and the synthetic MRS spectrum into the substance amount estimation NN, it is possible to improve the discrimination ability between the peak and noise of the MRS spectrum by the substance amount estimation NN, and it is expected that the estimation accuracy of the substance amount information will be improved. To.

(Modification 5)
In the above embodiment, it is assumed that only the substance amount estimation NN by the substance amount output function 514 is created by the machine learning model. However, the substance amount output function 514 and the synthesis function 515 may be created by a single machine learning model (hereinafter referred to as a first integrated model).

FIG. 11 is a diagram showing a network configuration example of the first integrated model NN10 according to the modified example 5 of the first embodiment. As shown in FIG. 11, the first integrated model NN10 is a deep neural network having a substance amount information output layer NN11, a synthetic layer NN12, a substance amount information output layer NN13, and a synthetic layer NN14. The substance amount information output layer NN11 and the substance amount information output layer NN13 are neural network layers corresponding to the substance amount estimation NN described in some of the above embodiments. The substance amount information output layer NN11 and the substance amount information output layer NN13 are examples of the parameter output unit. The synthetic layer NN12 and the synthetic layer NN14 are neural network layers that are processed by the synthetic function 515 to generate a synthetic MRS spectrum based on the amount of substance information. The synthetic layer NN12 and the synthetic layer NN14 are examples of the synthetic unit.

Specifically, as shown in FIG. 11, the substance amount information output layer NN11 inputs the first MRS spectrum and the second MRS spectrum, and outputs the substance amount information of propylene glycol, ethanol, acetate, and acetone. do. The synthetic layer NN12 outputs a synthetic MRS spectrum by inputting the substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN11. The substance amount information output layer NN13 inputs the first MRS spectrum and the synthetic MRS spectrum output from the synthetic layer NN12, and outputs the substance amount information of propylene glycol, ethanol, acetate and acetone. The synthetic layer NN14 outputs a synthetic MRS spectrum by inputting the substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN13.

The first integrated model NN10 shown in FIG. 11 has two unit layers in which a substance amount information output layer and a synthetic layer are connected in series. However, the integrated model NN10 may have only one unit layer or may have three or more unit layers. Further, although the number of input channels of the substance amount information output layer is two, the number is not limited to this, and one or three may be used.

FIG. 12 is a diagram showing a network configuration example of another first integrated model NN20 according to the modified example 5 of the first embodiment. As shown in FIG. 12, the first integrated model NN20 is a deep neural network having a substance amount information output layer NN21, a synthetic layer NN22, a substance amount information output layer NN23, and a synthetic layer NN24. The substance amount information output layer NN21 and the substance amount information output layer NN23 are examples of the parameter output unit. The synthetic layer NN22 and the synthetic layer NN24 are examples of the synthetic unit.

The substance amount information output layer NN21 inputs the first MRS spectrum and the second MRS spectrum, and outputs the substance amount information of propylene glycol, ethanol, acetate and acetone. The synthetic layer NN22 outputs a synthetic MRS spectrum by inputting the substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN21. The substance amount information output layer NN23 inputs the first MRS spectrum, the second MRS spectrum, and the synthetic MRS spectrum output from the synthetic layer NN12, and outputs the substance amount information of propylene glycol, ethanol, acetate, and acetone. .. The synthetic layer NN24 outputs the synthetic MRS spectrum by inputting the substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN23.

As described above, according to the modification 5, the substance amount output function 514 and the synthesis function 515 are combined with a plurality of unit layers (material amount information output layer and synthetic layer) as in the first integrated models 10 and 20. It is possible to create a deep neural network in which (combinations of) are connected in series. In this way, since the process of generating the synthetic MRS spectrum from a plurality of MRS spectra can be created by the deep neural network, the synthetic MRS spectrum can be generated at high speed and with high accuracy.

As described above, the MR signal processing device 50 according to the first embodiment has a processing circuit 51A. The processing circuit 51A applies a trained model to a plurality of MRS signals collected by MR spectroscopy for the same object, and outputs physical quantity information of a plurality of substances for MRS reconstruction. The processing circuit 51A generates an MRS spectrum for a measurement target site by fitting based on physical quantity information of a plurality of substances and a plurality of artificial spectra.

According to the above configuration, since the physical quantity information is obtained based on a plurality of MRS signals, the accuracy of the physical quantity information can be improved. Further, since the MRS spectrum is generated based on such physical quantity information, the accuracy of the MRS spectrum can be improved.

(Second Embodiment)
Next, the MR signal processing apparatus according to the second embodiment will be described. In the following description, components having substantially the same functions as those of the first embodiment are designated by the same reference numerals and will be duplicated and described only when necessary.

FIG. 13 is a diagram showing a configuration example of the magnetic resonance imaging device 1 according to the second embodiment. As shown in FIG. 13, the magnetic resonance imaging device 1 includes a gantry 11, a sleeper 13, a gradient magnetic field power supply 21, a transmission circuit 23, a reception circuit 25, a sleeper drive device 27, a sequence control circuit 29, and an MR signal processing device (host computer). ) Has 50. The MR signal processing device 50 is a computer having a processing circuit 51B, a memory 53, a display 55, an input interface 57, and a communication interface 59.

The processing circuit 51B has a processor such as a CPU as a hardware resource. The processing circuit 51B functions as the center of the magnetic resonance imaging device 1. For example, the processing circuit 51B realizes an acquisition function 521, a condition setting function 522, a signal processing function 523, a component amount output function 524, a synthesis function 525, a determination function 526, a learning function 527, and a display control function 528 by executing various programs. do. The acquisition function 521 is an example of an acquisition unit, the condition setting function 522 is an example of a setting unit, the signal processing function 523 is an example of a spectrum generation unit, and the substance amount output function 524 is an example of a parameter output unit. The synthesis function 525 is an example of the synthesis unit, the determination function 526 is an example of the determination unit, the learning function 527 is an example of the learning unit, and the display control function 528 is an example of the display unit.

The acquisition function 521, the condition setting function 522, and the signal processing function 523 are functions that substantially match the acquisition function 511, the condition setting function 512, and the signal processing function 513 according to the first embodiment, respectively.

In the component amount output function 524, the processing circuit 51B applies the trained model to a plurality of MRS signals collected by MR spectroscopy for the same target P, and outputs a plurality of parameters related to MRS reconstruction. The plurality of parameters according to the second embodiment are spectral parameters for fitting based on a plurality of spectral models corresponding to the plurality of parameters. The plurality of spectrum models are a plurality of basis spectra corresponding to a plurality of basis for the measurement target site. The plurality of bases are obtained by performing data conversion on the MRS spectrum with respect to the measurement target site. The plurality of spectral parameters include a plurality of basis component amount information. The component amount information means information regarding the base component amount. Hereinafter, the parameter according to the second embodiment will be referred to as basal component amount information. As an example, the basis is a first basis based on the MRS spectrum for a healthy person (hereinafter referred to as a healthy person basis) and a second basis based on the MRS spectrum for an unhealthy person (hereinafter referred to as an unhealthy person basis). The trained model including (called) is a machine learning model trained to output component amount information by inputting a plurality of MRS signals. Neural networks and deep neural networks are used as machine learning models. Hereinafter, the trained model according to the second embodiment will be referred to as a basal component amount estimation NN.

In the synthesis function 525, the processing circuit 51B generates a synthetic MRS signal relating to the target P, which has noise reduced as compared with a plurality of MRS signals, based on the base component amount information output in the component amount output function 524.

In the determination function 526, the processing circuit 51B determines whether the subject P is normal or abnormal based on the basal component amount information regarding the healthy subject base and the basal component amount information regarding the unhealthy subject base.

In the learning function 527, the processing circuit 51B generates the basal component amount estimation NN used in the component amount output function 524 by machine learning based on the learning data.

In the display control function 528, the processing circuit 51B displays various information on the display 55. For example, the processing circuit 51B displays an MRS signal, basal component amount information, a synthetic MRS signal, a normal / abnormal determination result, a data acquisition condition setting screen, and the like on the display 55.

Hereinafter, an operation example of the MR signal processing apparatus 50 according to the second embodiment will be described in detail.

As described above, the processing circuit 51B applies the basal component amount estimation NN to the plurality of MRS signals by the component amount output function 524, and outputs the basal component amount information of the basal.

FIG. 14 is a diagram schematically showing the input / output relationship of the trained model (basal component amount estimation NN) according to the second embodiment. As shown in FIG. 14, the basal component amount estimation NN is a machine learning model in which parameters are trained so as to output basal component amount information by inputting a first MRS signal and a second MRS signal. be. The first MRS signal and the second MRS signal are MRS signals relating to the same measurement target site of the subject P. The data acquisition conditions of the first MRS signal and the second MRS signal may be the same or different. As the MRS signal input to the basal component amount estimation NN, MRSk data or MRS spectrum can be used. Although the number of MRS signals input to the basal component amount estimation NN shown in FIG. 14 is two, this is an example, and any number may be used as long as it is two or more.

The basal component amount information is information on the component amount of each basal when the MRS spectrum representing the measurement target portion of the first MRS signal and the second MRS signal is mathematically converted into a plurality of basals. The basis can be obtained by applying data compression to the MRS spectrum representing the measurement target site. As the data compression method, for example, low-rank approximation, principal component analysis, singular value decomposition, self-encoder and the like can be used. The MRS spectrum representing the measurement target site may be an MRS spectrum measured in advance for the measurement target site of the subject P, or may be an MRS spectrum for the same measurement target site of a person different from the subject P. The obtained plurality of bases are used for the basal component amount estimation NN. Further, a large number of MRS spectra related to the same measurement target site may be data-compressed and converted into a large number of bases, and the base used for the basal component amount estimation NN may be selected from these many bases. For example, a large number of bases may be divided into a plurality of clusters by clustering or the like, and the base used for the base component amount estimation NN may be selected from each cluster.

The basal component amount estimation NN is generated by the learning function 527 of the processing circuit 51B. The processing circuit 51B trains a machine learning model based on a plurality of training samples to generate a basal component amount estimation NN. The training sample is a combination of the first MRS signal and the second MRS signal, which are input data, and the basal component amount information (hereinafter, referred to as correct basal component amount information), which is the correct answer data. The first MRS signal and the second MRS signal, which are input data, are generated by the magnetic resonance imaging device 1 or another magnetic resonance imaging device. The correct basis component amount information can be obtained as the component amounts of a plurality of bases when the MRS spectrum corresponding to the input data is converted into a plurality of previously obtained bases. The processing circuit 51B applies a machine learning model to the first MRS signal and the second MRS signal to perform forward propagation processing, and outputs basal component amount information (hereinafter referred to as estimated basal component amount information). Next, the processing circuit 51B applies the difference (error) between the estimated basal component amount information and the correct basal component amount information to the machine learning model to perform back propagation processing, and uses the differential coefficient of the error function, which is a function of the parameters. Compute a gradient vector. Next, the processing circuit 51B updates the parameters of the machine learning model based on the gradient vector. The forward propagation process, the back propagation process, and the parameter update process are repeated while changing the training sample, and the parameters that minimize the error function are determined according to a predetermined optimization method. As a result, the basal component amount estimation NN is generated.

Next, the signal processing by the MR signal processing apparatus 50 according to the second embodiment will be described with reference to FIG. In the following description, it is assumed that the MRS signal input to the basal component amount estimation NN is an MRS spectrum.

FIG. 15 is a diagram showing an example of a flow of signal processing by the MR signal processing device 50 according to the second embodiment.

As shown in FIG. 15, the processing circuit 51B acquires the first MRS spectrum and the second MRS spectrum by realizing the acquisition function 521 (step SB1). Step SB1 is performed in the same manner as step SA1.

When step SB1 is performed, the processing circuit 51B applies the basal component amount estimation NN to the first MRS spectrum and the second MRS spectrum acquired in step SB1 by realizing the component amount output function 524, and the basal component. Output the quantity information (step SB2).

FIG. 16 is a diagram schematically showing the input / output relationship of the basal component amount estimation NN. As shown in FIG. 16, the basal component amount estimation NN is a machine learning model in which parameters are learned so as to output basal component amount information by inputting a first MRS spectrum and a second MRS spectrum. By converting the first MRS spectrum and the second MRS spectrum into the basis component amount information, the deterioration process of the first MRS spectrum and the second MRS spectrum can be woven into the basis.

The basal component amount information includes the component amount information of each basal and the reliability information. In FIG. 16, as specific examples, the component amount C1 and the half-value width W1 of the healthy person base # 1, the component amount C2 and the half-value width W2 of the healthy person base # 2, and the component amount C3 and the half-value width of the unhealthy person base # 3 are shown. W3, the component amount C4 of the unhealthy person basis # 4, and the half width W4 are exemplified. Healthy person bases # 1 and # 2 are bases obtained by applying data compression to the MRS spectrum of a healthy person. A healthy person is a person who is evaluated as "no abnormality" with respect to the measurement target sites of the first MRS spectrum and the second MRS spectrum. The unhealthy person bases # 3 and # 4 are the bases obtained by applying data compression to the MRS spectrum of the unhealthy person. An unhealthy person is a person who is evaluated as "abnormal" with respect to the measurement target sites of the first MRS spectrum and the second MRS spectrum. When a plurality of peaks exist for one basis, the component amount and the half width are output for each peak. The peak may be identified by the frequency difference.

The processing circuit 51B applies the basis component amount estimation NN to the first MRS spectrum and the second MRS spectrum acquired in step SB1, and the component amount C1 and the half width at half maximum W1 of the healthy person base # 1 and the healthy person. The component amount C2 and half width W2 of the basis # 2, the component amount C3 and the half width W3 of the unhealthy person base # 3, and the component amount C4 and the half width W4 of the unhealthy person base # 4 are output. According to the second embodiment, since the basal component amount information can be estimated from a plurality of MRS spectra related to the same measurement target part, it is possible to accurately estimate the basal component amount information included in the measurement target part. can.

When step SB2 is performed, the processing circuit 51B generates a synthetic MRS spectrum based on the basal component amount information output in step SB2 by realizing the synthesis function 515 (step SB3). Step SB3 is performed in the same manner as in step SA3. That is, a plurality of basis spectra corresponding to each of the plurality of bases are stored in advance. In step SB3, the processing circuit 51B applies the component amount and full width at half maximum of each of the plurality of bases to the base spectrum to generate a plurality of corrected base spectra, and fits the generated multiple corrected base spectra to synthesize a synthesized MRS spectrum. To generate. Since the synthetic MRS spectrum is generated by fitting based on the basal spectrum, the data collection of the MR spectroscopy is unstable and the body of the subject P is compared with the input first MRS spectrum and the second MRS spectrum. Noise caused by turbulence of dynamic and static magnetic fields is reduced. In other words, the basal component amount estimation by the basal component amount estimation NN (step SB2) and the generation of the synthetic MRS spectrum based on the basal component amount information (step SB3) are the denoises of the first MRS spectrum and the second MRS spectrum. It can be equated with processing.

When step SB3 is performed, the processing circuit 51B determines normality or abnormality based on the basal component amount information output in step SB2 by realizing the determination function 526 (step SB4).

17 and 18 are diagrams schematically showing normality or abnormality determination processing based on basal component amount information, FIG. 17 is related to normality determination, and FIG. 18 is related to abnormality determination. As shown in FIGS. 17 and 18, in step SB2, as the basal component amount information, the component amounts C11 and C12 of the healthy person base # 1, the component amounts C21 and C22 of the healthy person base # 2, and the unhealthy person base # 3 It is assumed that the component amounts C31 and C32 of the above and the component amounts C41 and C42 of the unhealthy person base # 4 are output. The processing circuit 51B comprises the component amounts C11, C12 and the component amounts C21, C22 of the healthy person bases # 1 and # 2, and the component amounts C31, C32 and the component amounts C41, C42 of the unhealthy person bases # 3 and # 4. Judge whether it is normal or abnormal according to the ratio. In the processing circuit 51B, as shown in FIG. 17, the component amounts C11 and the component amounts C21 of the healthy person bases # 1 and # 2 are larger than the component amounts C31 and the component amounts C41 of the unhealthy person bases # 3 and # 4. If so, it is determined to be normal. On the other hand, in the processing circuit 51B, as shown in FIG. 18, the component amounts C32 and the component amounts C42 of the unhealthy person bases # 3 and # 4 are based on the component amounts C12 and the component amounts C22 of the healthy person bases # 1 and # 2. If there are many, it is judged to be abnormal.

When step SB4 is performed, the processing circuit 51B displays the synthetic MRS spectrum generated in step SB3 and the determination result obtained in step SB4 by realizing the display control function 528 (step SB5). The synthesized MRS spectrum and the determination result are displayed on the display 55 in a predetermined layout.

FIG. 19 is a diagram showing an example of a display screen I1 of a synthetic MRS spectrum I12 and a determination result I13. As shown in FIG. 19, the interest voxel setting image I11 is displayed on the display screen I1. The interest voxel setting image I11 is an MR image in which the interest voxel V1 of the first MRS spectrum and the second MRS spectrum is set. For example, the brain tumor region R1 is drawn on the interest voxel setting image I11, and the interest voxel V1 is drawn on the brain tumor region R1. When the voxels of interest in the first MRS spectrum and the second MRS spectrum are set at different positions, the voxels of interest in the first MRS spectrum and the voxels of interest in the second MRS spectrum are drawn at different positions. Will be. The voxel setting image I11 of interest may be an MR image collected by any imaging method.

As shown in FIG. 19, the synthetic MRS spectrum I12 generated in step SB3 is displayed on the display screen I1. This allows the user to observe the synthetic MRS spectrum, which is the denoised MRS spectrum.

As shown in FIG. 19, the normal or abnormal determination result I13 obtained in step SB4 is displayed on the display screen I1. For example, when it is determined to be abnormal in step SB4, the determination result is displayed as "suspicious of abnormality". In addition, information accompanying the determination result may be displayed, such as "detailed inspection required".

As a result, the signal processing by the MR signal processing device 50 according to the second embodiment is completed.

The second embodiment can be modified in various ways. For example, in FIG. 15, it is assumed that the normal or abnormal determination process is performed after the synthetic MRS spectrum generation process is performed. However, the synthetic MRS spectrum generation process may be performed after the normal or abnormal determination process is performed.

In the above embodiment, it is assumed that normality or abnormality is determined. However, the presence or absence of individual diseases such as brain tumors, toxic leukoencephalopathy, stroke, dementia, and trauma may be determined. In this case, as an unhealthy person's basis, a basis vector unique to the lesion (hereinafter referred to as a disease basis vector) is prepared for each disease. The disease basis vector can be obtained by applying a data compression technique to the MRS spectrum of an unhealthy person diagnosed with the disease. The processing circuit 51B compares the amount of the basal component of the disease basal vector with respect to the threshold value, and if the amount of the basal component exceeds the threshold value, determines that the patient is suffering from the disease, and the amount of the basal component does not exceed the threshold value. In that case, it may be determined that the patient does not have the disease.

As described above, the MR signal processing device 50 according to the second embodiment has a processing circuit 51B. The processing circuit 51B applies a trained model to a plurality of MRS signals collected by MR spectroscopy for the same object, and outputs information on the amount of basis components of a plurality of bases for MRS reconstruction. The processing circuit 51B generates an MRS spectrum for a measurement target site by fitting based on information on the amount of basal components of a plurality of basals and a plurality of basal spectra.

According to the above configuration, since the basal component amount information is obtained based on a plurality of MRS signals, the accuracy of the basal component amount information can be improved. Further, since the MRS spectrum is generated based on such information on the amount of basal components, the accuracy of the MRS spectrum can be improved.

(Third Embodiment)
Next, the MR signal processing apparatus according to the third embodiment will be described. In the following description, components having substantially the same functions as those of the first embodiment and the second embodiment are designated by the same reference numerals and will be duplicated and described only when necessary.

FIG. 20 is a diagram showing a configuration example of the magnetic resonance imaging device 1 according to the third embodiment. As shown in FIG. 20, the magnetic resonance imaging apparatus 1 according to the third embodiment includes a gantry 11, a sleeper 13, a gradient magnetic field power supply 21, a transmission circuit 23, a reception circuit 25, a sleeper drive device 27, a sequence control circuit 29, and an MR. It has a signal processing device (host computer) 50. The MR signal processing device 50 is a computer having a processing circuit 51C, a memory 53, a display 55, an input interface 57, and a communication interface 59.

The processing circuit 51C has a processor such as a CPU as a hardware resource. The processing circuit 51C functions as the center of the magnetic resonance imaging device 1. For example, the processing circuit 51C realizes an acquisition function 531, a condition setting function 532, a signal processing function 533, a modulation amount output function 534, a correction function 535, a learning function 536, and a display control function 537 by executing various programs. The acquisition function 531 is an example of an acquisition unit, the condition setting function 532 is an example of a setting unit, the signal processing function 533 is an example of a spectrum generation unit, and the modulation amount output function 534 is an example of a parameter output unit. The correction function 535 is an example of the correction unit, the learning function 536 is an example of the learning unit, and the display control function 537 is an example of the display unit.

The acquisition function 531 and the condition setting function 532 and the signal processing function 533 are functions that substantially match the acquisition function 511, the condition setting function 512, and the signal processing function 513 according to the first embodiment, respectively.

In the modulation amount output function 534, the processing circuit 51C applies the trained model to a plurality of MRS signals collected by MR spectroscopy with respect to the subject P, and outputs a plurality of parameters for MRS reconstruction. The plurality of parameters according to the third embodiment include information regarding k-space modulation resulting from magnetic field modulation between the plurality of MRS signals. Hereinafter, the parameter according to the third embodiment will be referred to as k-space modulation amount information. The trained model is a machine learning model trained to output k-space modulation amount information by inputting a plurality of MRS signals. Neural networks and deep neural networks are used as machine learning models. Hereinafter, the trained model according to the third embodiment will be referred to as a modulation amount estimation NN.

In the correction function 535, the processing circuit 51C corrects the MRS signal based on the k-space modulation amount information output by the modulation amount output function 534. Hereinafter, the corrected MRS signal will be referred to as a corrected MRS signal.

In the learning function 536, the processing circuit 51C generates the modulation amount estimation NN used in the modulation amount output function 534 by machine learning based on the learning data.

In the display control function 537, the processing circuit 51C displays various information on the display 55. For example, the processing circuit 51C displays an MRS signal, modulation amount information, a corrected MRS signal, a data acquisition condition setting screen, and the like on the display 55.

Hereinafter, an operation example of the MR signal processing apparatus 50 according to the third embodiment will be described in detail.

As described above, by the modulation amount output function 534, the processing circuit 51C applies the modulation amount estimation NN to a plurality of MRS signals, and k-space related to the k-space modulation caused by the magnetic field modulation between the data acquisitions of the plurality of MRS signals. Output the modulation amount information.

FIG. 21 is a diagram schematically showing the input / output relationship of the trained model (modulation amount estimation NN) according to the third embodiment. As shown in FIG. 21, the modulation amount estimation NN is caused by the magnetic field modulation between the data acquisition of the first MRS signal and the second MRS signal with the first MRS signal and the second MRS signal as inputs. This is a machine learning model in which parameters are learned so as to output k-spatial modulation amount information regarding k-spatial modulation. The first MRS signal and the second MRS signal are MRS signals targeting the same subject P. The data acquisition conditions of the first MRS signal and the second MRS signal may be the same or different. As the MRS signal input to the modulation amount estimation NN, MRSk data or MRS spectrum can be used. Although the number of MRS signals input to the modulation amount estimation NN shown in FIG. 21 is two, this is an example, and any number may be used as long as it is two or more.

The k-space modulation amount information is information regarding the modulation amount of the k-space modulation due to the magnetic field modulation between the data acquisitions of the first MRS signal and the second MRS signal. As factors of magnetic field modulation, there are various factors such as static magnetic field modulation, gradient magnetic field modulation, and noise. Factors of static magnetic field modulation include static magnetic field non-uniformity and static magnetic field offset (f0 shift). The factor of the modulation of the gradient magnetic field is the transient response (phase shift) of the gradient magnetic field. The magnetic field modulation appears as a deviation of the actual MRSk data from the ideal MRSk data without magnetic field modulation in k-space. When magnetic field modulation occurs between data acquisition for the first MRS signal and the second MRS signal, there will be a discrepancy between the first MRSk data and the second MRSk data in k-space.

The modulation amount estimation NN is generated by the learning function 536 of the processing circuit 51C. The processing circuit 51C trains a machine learning model based on a plurality of training samples to generate a modulation amount estimation NN. The training sample is a combination of the first MRS signal and the second MRS signal, which are input data, and k-space modulation amount information (hereinafter, referred to as correct k-space modulation amount information), which is correct data. The first MRS signal and the second MRS signal, which are input data, are generated by the magnetic resonance imaging device 1 or another magnetic resonance imaging device. The correct k-space modulation amount information is obtained by measuring the deviation of the second MRSk data in the k-space with respect to the first MRSk data. The processing circuit 51C applies a machine learning model to the first MRS signal and the second MRS signal to perform forward propagation processing, and outputs k-space modulation amount information (hereinafter referred to as estimated k-space modulation amount information). .. Next, the processing circuit 51C applies the difference (error) between the estimated k space modulation amount information and the correct answer k space modulation amount information to the machine learning model to perform back propagation processing, and differentiates the error function which is a function of the parameter. Calculate the gradient vector, which is a coefficient. Next, the processing circuit 51C updates the parameters of the machine learning model based on the gradient vector. The forward propagation process, the back propagation process, and the parameter update process are repeated while changing the training sample, and the parameters that minimize the error function are determined according to a predetermined optimization method. As a result, the modulation amount estimation NN is generated.

Next, the signal processing by the MR signal processing apparatus 50 according to the third embodiment will be described with reference to FIG. 22. In the following description, it is assumed that the MRS signal input to the modulation amount estimation NN is MRSk data.

FIG. 22 is a diagram showing an example of the flow of signal processing by the MR signal processing device 50 according to the third embodiment.

As shown in FIG. 22, the processing circuit 51C acquires the first MRSk data and the second MRSk data by realizing the acquisition function 531 (step SC1). It is assumed that the first MRSk data and the second MRSk data are, for example, MRSk data related to the same measurement target site collected under the same data collection conditions by the single voxel method. The processing circuit 51C selects and acquires two arbitrary MRSk data as the first MRSk data and the second MRSk data from the MRSk data for the number of integrations before the addition processing. The two MRSk data may be manually selected by the user according to instructions via the input interface 57 or automatically according to a predetermined algorithm.

When step SC1 is performed, the processing circuit 51C applies the modulation amount estimation NN to the first MRSk data and the second MRSk data acquired in step SC1 by realizing the modulation amount output function 534, and k-space modulation. Output the quantity information (step SC2).

FIG. 23 is a diagram schematically showing a specific example of the input / output relationship of the modulation amount estimation NN. As shown in FIG. 23, the modulation amount estimation NN takes the first MRSk data and the second MRSk data as inputs, and the k space modulation amount between the data acquisition of the first MRSk data and the second MRSk data. It is a machine learning model in which parameters are trained to output information. Specifically, f0 shift and phase shift are exemplified as the k-space modulation amount information. The k-space modulation amount information regarding the f0 shift is the amount of deviation of the MRSk data in the k-space due to the static magnetic field modulation, and the phase shift is the amount of deviation of the MRSk data in the k-space due to the gradient magnetic field modulation. The k-space modulation amount information may be a numerical value indicating the deviation amount of the k-space modulation, or may be a parameter such as a time constant, a gain, or a time origin of the impulse response representing the k-space modulation.

The processing circuit 51C applies the modulation amount estimation NN to the first MRSk data and the second MRSk data acquired in step SC1 during the data acquisition of the first MRSk data and the second MRSk data. k Space modulation amount information is output. According to the third embodiment, the k-space modulation amount information at the time of data acquisition of MRSk data can be estimated. Further, since the k-space modulation amount information can be estimated from a plurality of MRSk data, the k-space modulation amount information can be estimated accurately.

When step SC2 is performed, the processing circuit 51C corrects the MRSk data based on the k-space modulation amount information output in step SC2 by realizing the correction function 535 (step SC3). The corrected MRSk data is referred to as corrected MRSk data.

When step SC3 is performed, the processing circuit 51C generates an MRS spectrum based on the corrected MRSk data generated in step SC3 by realizing the signal processing function 533 (step SC4).

FIG. 24 is a diagram schematically showing the correction process and the spectrum generation process in step SC3 and step SC4. As shown in FIG. 24, the MRSk data OK1 to be corrected has been acquired. As the MRSk data OK1, the first MRSk data and / or the second MRSk data acquired in step SC1 may be selected, or other MRSk data may be selected. As the other MRSk data, for example, MRSk data other than the first MRSk data and the second MRSk data acquired in step SC1 may be selected from the MRSk data for the number of integrations. Further, as the MRSk data OK1, the average MRSk data generated by subjecting the MRSk data for the number of integrations to the addition processing may be selected. Hereinafter, it is assumed that the MRSk data OK1 is the average MRSk data.

As shown in FIG. 24, k-space modulation amount information regarding f0 shift and phase shift is output in step SC2. The processing circuit 51C corrects the average MRSk data OK1 based on the k-space modulation amount information regarding the f0 shift and the phase shift, and generates the corrected MRSk data CK1 (step SC3). In the corrected MRSk data CK1, the f0 shift and the phase shift are reduced.

Next, as shown in FIG. 24, the processing circuit 51C performs a Fourier transform on the corrected MRSk data CK1, converts the signal intensity value into digital data represented by a frequency function, and performs phase correction or baseline correction on the converted digital data. The MRS spectrum CS1 is generated by performing post-processing such as (step SC4). In the MRS spectrum CS1, the noise caused by k-space modulation is reduced as compared with the MRS spectrum generated from the average MRSk data OK1 without the correction process (step SC3). That is, according to the third embodiment, even when the average MRSk data obtained under a smaller number of integrations than in the conventional case is used, the MRS spectrum can be corrected by the correction based on the k-space modulation amount information. The accuracy can be improved. Therefore, according to the third embodiment, the number of integrations can be reduced as compared with the case where the correction process (step SC3) is not performed.

When step SC4 is performed, the processing circuit 51C displays the MRS spectrum generated in step SC4 by realizing the display control function 549 (step SC5). This allows the user to observe the denoised MRS spectrum.

As a result, the signal processing by the MR signal processing device 50 according to the third embodiment is completed.

As described above, the MR signal processing device 50 according to the third embodiment has a processing circuit 51C. The processing circuit 51C applies the trained model to a plurality of MRSk data collected by MR spectroscopy for the same object, and outputs k-space modulation amount information for MRS reconstruction. The processing circuit 51C corrects the MRSk data based on the k-space modulation amount information. Further, the processing circuit 51C generates an MRS spectrum based on the corrected MRSk data.

According to the above configuration, since the k-space modulation amount information is obtained based on a plurality of MRSk data, the accuracy of the k-space modulation amount information can be improved. Further, since the MRSk data and the MRS spectrum can be generated based on such k-space modulation amount information, the accuracy of the MRSk data and the MRS spectrum can be improved.

(Fourth Embodiment)
Next, the MR signal processing apparatus according to the fourth embodiment will be described. In the following description, components having substantially the same functions as those of the first embodiment are designated by the same reference numerals and will be duplicated and described only when necessary.

FIG. 25 is a diagram showing a configuration example of the magnetic resonance imaging device 1 according to the fourth embodiment. As shown in FIG. 25, the magnetic resonance imaging device 1 includes a gantry 11, a sleeper 13, a gradient magnetic field power supply 21, a transmission circuit 23, a reception circuit 25, a sleeper drive device 27, a sequence control circuit 29, and an MR signal processing device (host computer). ) Has 50. The MR signal processing device 50 is a computer having a processing circuit 51D, a memory 53, a display 55, an input interface 57, and a communication interface 59.

The processing circuit 51D has a processor such as a CPU as a hardware resource. The processing circuit 51D functions as the center of the magnetic resonance imaging device 1. For example, the processing circuit 51D has an acquisition function 541, a condition setting function 542, a signal processing function 543, a substance amount output function 544, a synthesis function 545, a modulation amount output function 546, a correction function 547, a learning function 548, and the processing circuit 51D. The display control function 549 is realized. The acquisition function 541 is an example of an acquisition unit, the condition setting function 542 is an example of a setting unit, the signal processing function 543 is an example of a spectrum generation unit, and the substance amount output function 544 is an example of a parameter output unit. The synthesis function 545 is an example of the synthesis unit, the modulation amount output function 546 is an example of the parameter output unit, the correction function 547 is an example of the correction unit, the learning function 548 is an example of the learning unit, and the display control function. 549 is an example of a display unit.

The acquisition function 541, the condition setting function 542, the signal processing function 543, the substance amount output function 544, and the synthesis function 545 have the acquisition function 511, the condition setting function 512, the signal processing function 513, and the substance amount output function, respectively, according to the first embodiment. It is a function that substantially matches the 514 and the synthesis function 515. The modulation amount output function 546 and the correction function 547 are functions that substantially match the modulation amount output function 534 and the correction function 535 according to the third embodiment, respectively. The learning function 548 is a function that substantially matches the learning function 516 according to the first embodiment and the learning function 536 according to the third embodiment. The display control function 549 is a function that substantially matches the display control function 517 according to the first embodiment and the display control function 537 according to the third embodiment.

Hereinafter, an operation example of the MR signal processing apparatus 50 according to the fourth embodiment will be described in detail.

The fourth embodiment is a combination of the first embodiment and the third embodiment. That is, based on the synthetic MRS spectrum generated by the synthesis function 545 based on the substance amount information output by the substance amount output function 544, the k-space modulation amount information by the modulation amount output function 546 is output. As one means of this, the substance amount output function 544, the synthesis function 545 and the modulation amount output function 546 may be built by a single machine learning model (hereinafter referred to as a second integrated model).

FIG. 26 is a diagram showing a network configuration example of the second integrated model NN30 according to the fourth embodiment. As shown in FIG. 26, the second integrated model NN30 includes a substance amount information output layer NN31, a synthetic layer NN32, an inverse conversion layer NN33, a substance amount information output layer NN34, a synthetic layer NN35, an inverse conversion layer NN36, and k spatial modulation. It is a deep neural network having a quantity information output layer NN37. The substance amount information output layer NN31 and the substance amount information output layer NN34 are a kind of neural network layer that outputs the substance amount information from the MRS signal by the substance amount output function 544. The synthetic layer NN32 and the synthetic layer NN35 are neural network layers that are processed by the synthetic function 545 to generate a synthetic MRS spectrum based on the amount of substance information. The inverse conversion layer NN33 and the inverse conversion layer NN36 are neural network layers that are processed by the signal processing function 543 to convert the synthesized MRS spectrum into MRSk data. The k-space modulation amount information output layer NN37 is a neural network layer that performs processing for outputting k-space modulation amount information from two MRSk data by the modulation amount output function 546.

Specifically, as shown in FIG. 26, the substance amount information output layer NN31 receives the first MRS spectrum as an input, and the first of propylene glycol, ethanol, acetate and acetone corresponding to the first MRS spectrum. Outputs the substance amount information of. The synthetic layer NN32 receives the first substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN31 as input, and outputs the first synthetic MRS spectrum. The inverse transformation layer NN33 takes the first synthetic MRS spectrum output from the synthetic layer NN32 as an input, and generates the first MRSk data which is k-space data corresponding to the first synthetic MRS spectrum.

Similarly, the substance amount information output layer NN34 takes the second MRS spectrum as an input and outputs the second substance amount information of propylene glycol, ethanol, acetate and acetone corresponding to the second MRS spectrum. The synthetic layer NN35 receives the second substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN34 as input, and outputs the second synthetic MRS spectrum. The inverse transformation layer NN36 takes a second synthetic MRS spectrum output from the synthetic layer NN35 as an input, and generates a second MRSk data which is k-space data corresponding to the second synthetic MRS spectrum.

The k-space modulation amount information output layer NN37 receives the first MRSk data output from the inverse conversion layer NN33 and the second MRSk data output from the inverse conversion layer NN36 as inputs, and inputs the first MRS spectrum and the second MRS spectrum. Outputs k-space modulation amount information regarding f0 shift and phase shift during data acquisition of the MRS spectrum of. As described above, the combination of the substance amount information output layer NN31 and the synthetic layer NN32 and the combination of the substance amount information output layer NN34 and the synthetic layer NN35 can be equated with the denoise processing of the MRS spectrum. Since the k-space modulation amount information output layer NN37 outputs k-space modulation amount information from MRSk data based on the synthetic MRS spectrum which is a denoised MRS spectrum, improvement in the accuracy of k-space modulation amount information is expected.

The second integrated model NN30 is learned by the processing circuit 51D by realizing the learning function 548. The second integrated model NN30 can be generated, for example, by pre-train type learning or end-to-end type learning. The pre-learning type can be performed according to the following procedure, for example. First, the processing circuit 51D individually comprises the material amount information output layer NN31, the synthetic layer NN32, the inverse conversion layer NN33, the material amount information output layer NN34, the synthetic layer NN35, the inverse conversion layer NN36, and the k-space modulation amount information output layer 37. To learn. Next, the processing circuit 51D comprises a substance amount information output layer NN31, a synthetic layer NN32, an inverse conversion layer NN33, a substance amount information output layer NN34, a synthetic layer NN35, an inverse conversion layer NN36, and a k space modulation amount information output layer 37. As shown in FIG. 26, they are connected and learned as one deep neural network. This makes it possible to generate a second integrated model NN30. In the case of the end-to-end type, the processing circuit 51D learns one deep neural network to input two MRS spectra and output k-space modulation amount information of f0 shift and phase shift.

The second integrated model NN30 shown in FIG. 26 can be modified in various ways. For example, two or more unit layers of the substance amount information output layer NN31 and the synthetic layer NN32 may be connected in series. Similarly, two or more unit layers of the substance amount information output layer NN34 and the synthetic layer NN35 may be connected in series.

FIG. 27 is a diagram showing a network configuration example of another second integrated model NN40 according to the fourth embodiment. As shown in FIG. 27, the second integrated model NN40 includes a substance amount information output layer NN41, a synthetic layer NN42, an inverse conversion layer NN43, a substance amount information output layer NN44, a synthetic layer NN45, an inverse conversion layer NN46, and k spatial modulation. It is a deep neural network having a quantity information output layer NN47. The substance amount information output layer NN41 is a kind of neural network layer that outputs the substance amount information from the MRS signal by the substance amount output function 544. The substance amount information output layer NN44 is a kind of neural network layer that outputs substance amount information from two MRS signals by the substance amount output function 544. The synthetic layer NN42 and the synthetic layer NN45 are neural network layers that are processed by the synthetic function 545 to generate a synthetic MRS spectrum based on the amount of substance information. The inverse conversion layer NN43 and the inverse conversion layer NN46 are neural network layers that are processed by the signal processing function 543 to convert the synthesized MRS spectrum into MRSk data. The k-space modulation amount information output layer NN47 is a neural network layer that performs processing for outputting k-space modulation amount information from two MRsk data by the modulation amount output function 546.

Specifically, as shown in FIG. 27, the substance amount information output layer NN41 takes the first MRS spectrum as an input, and the first of propylene glycol, ethanol, acetate and acetone corresponding to the first MRS spectrum. Outputs the substance amount information of. The synthetic layer NN42 receives the first substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN41 as input, and outputs the first synthetic MRS spectrum. The inverse transformation layer NN43 takes the first synthetic MRS spectrum output from the synthetic layer NN42 as an input, and generates the first MRSk data which is k-space data corresponding to the first synthetic MRS spectrum.

Similarly, the substance amount information output layer NN44 has a first MRS spectrum and a second MRS spectrum as inputs, and propylene glycol, ethanol, and acetate corresponding to the first MRS spectrum and the second MRS spectrum. And the second substance amount information of acetone is output. The synthetic layer NN45 receives the second substance amount information of propylene glycol, ethanol, acetate and acetone output from the substance amount information output layer NN44 as input, and outputs the second synthetic MRS spectrum. The inverse transformation layer NN46 takes a second synthetic MRS spectrum output from the synthetic layer NN45 as an input, and generates a second MRSk data which is k-space data corresponding to the second synthetic MRS spectrum.

The k-space modulation amount information output layer NN47 receives the first MRSk data output from the inverse conversion layer NN43 and the second MRSk data output from the inverse conversion layer NN46 as inputs, and inputs the first MRS spectrum and the second MRS spectrum. Outputs k-space modulation amount information regarding f0 shift and phase shift during data acquisition of the MRS spectrum of. As described above, the combination of the substance amount information output layer NN41 and the synthetic layer NN42 and the combination of the substance amount information output layer NN44 and the synthetic layer NN45 can be equated with the denoise processing of the MRS spectrum. Since the k-space modulation amount information output layer NN47 outputs k-space modulation amount information from MRSk data based on the synthetic MRS spectrum which is a denoised MRS spectrum, improvement in the accuracy of k-space modulation amount information is expected. Further, the synthetic MRS spectrum, which is the output from the combination of the substance amount information output layer NN44 and the synthetic layer NN45, functions as an ideal MRS spectrum for the measurement target site. It can be said that the k-space modulation amount information output layer NN47 outputs k-space modulation amount information from the MRSk data based on the denoised first MRS spectrum and the MRSk data based on the ideal MRS spectrum. It can be said that this k-space modulation amount information represents the modulation amount of k-space modulation due to the modulation of the actual magnetic field at the time of data acquisition for the first MRS spectrum with respect to the ideal magnetic field. By correcting the MRSk data based on such k-space modulation amount information, it is possible to generate more accurate MRSk data and an MRS spectrum based on the MRSk data.

The second integrated model NN40 is learned by the processing circuit 51D by realizing the learning function 548. The second integrated model NN40 can be generated, for example, by pre-train type learning or end-to-end type learning. In the case of the pre-learning type, first, the processing circuit 51D has a substance amount information output layer NN41, a synthetic layer NN42, an inverse conversion layer NN43, a substance amount information output layer NN44, a synthetic layer NN45, an inverse conversion layer NN46, and k spatial modulation amount information. Each of the output layers NN47 is learned individually. Next, the processing circuit 51D includes a substance amount information output layer NN41, a synthetic layer NN42, an inverse conversion layer NN43, a substance amount information output layer NN44, a synthetic layer NN45, an inverse conversion layer NN46, and a k space modulation amount information output layer NN47. As shown in FIG. 27, they are connected and learned as one deep neural network. This makes it possible to generate a second integrated model NN40. In the case of the end-to-end type, the processing circuit 51D learns one deep neural network to input three MRS spectra and output k-space modulation amount information of f0 shift and phase shift.

The second integrated model NN40 shown in FIG. 27 can be modified in various ways. For example, it is assumed that the first MRS spectrum, which is the same as the input to the substance amount information output layer NN41, is input to the substance amount information output layer NN44 together with the second MRS spectrum. However, a third MRS spectrum different from the input to the substance amount information output layer NN41 may be input to the substance amount information output layer NN44 together with the second MRS spectrum. Further, instead of the 1-channel input substance amount information output layer NN44, a 2-channel input substance amount information output layer may be provided. Two or more unit layers of the material amount information output layer NN41 and the synthetic layer NN42 may be connected in series. Similarly, two or more unit layers of the substance amount information output layer NN44 and the synthetic layer NN45 may be connected in series.

(Fifth Embodiment)
The fifth embodiment is an application example of the third embodiment. It is assumed that the magnetic resonance imaging device 1 and the MR signal processing device 50 according to the fifth embodiment have the same configuration as the magnetic resonance imaging device and the MR signal processing device 50 according to the third embodiment.

FIG. 28 is a diagram schematically showing the input / output relationship of the trained model (modulation amount estimation NN) according to the fifth embodiment. As shown in FIG. 28, the modulation amount estimation NN takes the first MR collection signal and the second MR collection signal as inputs, and during data collection for the first MR collection signal and the second MR collection signal. This is a machine learning model in which parameters are trained so as to output k-spatial modulation amount information regarding k-spatial modulation caused by the generated magnetic field modulation. The processing circuit 51C applies the modulation amount estimation NN to the first MR collection signal and the second MR collection signal of the processing target for the same imaging target, and applies the modulation amount estimation NN to the first MR collection signal and the second MR of the processing target. The k-spatial modulation amount information regarding the k-spatial modulation caused by the magnetic field modulation generated during the data collection for the collected signal is output.

The MR collected signal is a general term for signals obtained by MR imaging or chemical shift measurement by the sequence control circuit 29. Specifically, as the MR imaging according to the fifth embodiment, T1 weighted imaging, T2 weighted imaging, T2 * weighted imaging, diffusion weighted imaging, MR angiography, and all other MR imaging can be applied. The pulse sequence of MR imaging is not particularly limited, and spin echo sequences, gradient echo sequences, inversion recovery methods, echoplana imaging, parallel imaging, compressed sensing, and all other pulse sequences can be applied. The k-space filling method is not particularly limited, and any method can be applied regardless of the Cartesian scan method, the radial scan method, the spiral method, the stack of stars method, the cush ball method, and any other number of dimensions. The chemical shift measurement can be applied to CEST (Chemical Exchange Spectroscopy) or ZAPPED (Z-Spectrum Analysis Provides Proton Environment Data) in addition to MR spectroscopy and chemical shift imaging applied to the above embodiments.

The MR collected signal is a concept including, for example, k-space data and an MR image based on the k-space data. The first MR collection signal and the second MR collection signal are MR collection signals targeting the same subject P. The data collection conditions of the first MR collection signal and the second MR collection signal may be the same or different. As the MR collected signal input to the modulation amount estimation NN, k-space data or an MR image can be used. Although the number of MR collected signals input to the modulation amount estimation NN shown in FIG. 28 is two, this is an example, and any number may be used as long as it is two or more. The MR collected signal includes the MRS signal according to the first to fourth embodiments.

FIG. 29 is a diagram schematically showing a specific example of the input / output relationship of the modulation amount estimation NN according to the fifth embodiment. As shown in FIG. 29, the modulation amount estimation NN takes the first non-MRSk space data and the second non-MRSk space data as inputs, and describes the first non-MRSk space data and the second non-MRSk space data. This is a machine learning model in which parameters are trained to output k-spatial modulation amount information between data acquisitions. Specifically, f0 shift and phase shift are exemplified as the k-space modulation amount information. Non-MR Sk spatial data is k-space data collected by chemical shift measurement or MR imaging other than MR spectroscopy.

As described above, the non-MRSk spatial data may be collected by any MR imaging method, pulse sequence, or k-space filling method. As the first non-MRSk space data and the second non-MRSk space data, k-space data for each k-space locus may be used. The k-space locus corresponds to, for example, the k-space line of the Cartesian scan, the spokes of the radial scan, and the spiral of the spiral scan. Hereinafter, a case where the spokes of the radial scan are used as the non-MRSk spatial data will be described.

FIG. 30 is a diagram schematically showing a further specific example of the input / output relationship of the modulation amount estimation NN shown in FIG. 29. As shown in FIG. 30, the modulation amount estimation NN is a machine learning model in which parameters are trained so as to output k-space modulation amount information regarding f0 shift and phase shift by inputting a plurality of spokes SK1-SK4 of the radial scan. Is. The plurality of spokes SK1-SK4 is any spoke selected from the set SK0 of the plurality of spokes collected by the radial scan. FIG. 30 shows an example in which all the spokes included in the set SK0 are input to the modulation amount estimation NN as a plurality of spokes SK1-SK4. However, not all the spokes included in the set SK0 need to be input to the modulation amount estimation NN, and any two spokes from the set SK0 may be input to the modulation amount estimation NN.

The processing circuit 51C selects a plurality of spokes SK1-SK4 to be processed from the set SK0, applies a modulation amount estimation NN to the selected spokes SK1-SK4, and applies a modulation amount estimation NN to the k-space modulation amount related to f0 shift and phase shift. Output information. The processing circuit 51C corrects the set SK0 based on the output k-space modulation amount information, generates a correction set, and reconstructs the MR image based on the generated correction set. In the MR image, the f0 shift and the phase shift are reduced.

According to the fifth embodiment, k-space modulation amount information can be obtained by using an MR collected signal other than the MRS signal. As a result, it is possible to make corrections related to k-space modulation with high accuracy for various MR collected signals.

(Sixth Embodiment)
The sixth embodiment is a modification applicable to any of the above embodiments. Hereinafter, the MR signal processing apparatus according to the sixth embodiment will be described. In the following description, components having substantially the same functions as those in the above embodiment are designated by the same reference numerals and will be described in duplicate only when necessary.

FIG. 31 is a diagram schematically showing the input / output relationship of the trained model NN of the sixth embodiment. The trained model NN of the sixth embodiment is a machine learning model trained to input a plurality of MRS signals and output MRS reconstruction parameters. The trained model NN of the sixth embodiment may be any machine learning model of the above-described first to fifth embodiments. Specifically, the trained model NN may be a substance amount estimation NN that outputs substance amount information as shown in FIG. 2 or the like of the first embodiment, or may be a substance amount estimation NN as shown in FIG. 14 or the like of the second embodiment. , The base component amount estimation NN that outputs the base component amount information, or the modulation amount estimation NN that outputs the k-spatial modulation amount information as shown in FIG. 21 or the like of the third embodiment may be used. The substance amount information, the base component amount information, and the k-space modulation amount information are one of the MRS reconstruction parameters. The MRS reconstruction parameter is a parameter obtained by MRS reconstruction. MRS reconstruction refers to the general processing for producing an MRS spectrum. For example, the MRS reconstruction is a process of generating an MRS spectrum by fitting from a spectrum model as in the first and second embodiments, and an MRS spectrum is generated by a Fourier transform from MRSk data as in the third embodiment. Including processing.

As shown in FIG. 31, the trained model NN inputs a plurality of MRS signals having different addition times. The number of MRS signals may be any number as long as it is two or more, but FIG. 31 illustrates 96 MRS signals. The number of additions means the number of MRS signals to be added. It can also be expressed as resolution.

In the sixth embodiment, the processing circuit 51 collects the number of MRS signals corresponding to NEX (for example, 128). Next, the processing circuit 51 generates 96 MRS signals from the number of MRS signals corresponding to NEX. Specifically, the processing circuit 51 sets 96 patterns having different combinations of the number of MRS signals to be added (number of additions) and the number of MRS signals to be added. The 96 patterns may be set by any algorithm or may be artificially set. Then, the processing circuit 51 generates 96 MRS signals by extracting and adding MRS signals for each of the 96 patterns.

As described above, it is expected that the estimation accuracy of the MRS reconstruction parameter will be improved by inferring the MRS reconstruction parameter by the trained model NN based on a plurality of MRS signals having different addition times.

In the above embodiment, a plurality of MRS signals having different addition times are input to the trained model NN, but a plurality of MRS signals having different qualities may be input. The quality depends on the degree of body movement of the subject P at the time of collecting the MRS signal. If the degree of body movement is small, the quality is high, and if the degree of body movement is large, the quality is low. As an example, the degree of body movement can be determined based on the measurement data measured by an external measuring instrument in parallel with the acquisition of the MRS signal. As the external measuring instrument, for example, an electrocardiograph for measuring the electrocardiographic waveform of the subject P or a respiratory meter for measuring the respiratory waveform can be used. It can be said that the quality of the MRS signal collected during the period is relatively low because the body movement of the subject P is intense during the period when the peak value of the electrocardiographic waveform or the respiratory waveform or the temporal fluctuation of the peak value is large. Therefore, the processing circuit 51 monitors the peak value of the electrocardiographic waveform or the respiratory waveform or the temporal fluctuation of the peak value, and the larger the peak value or the temporal fluctuation, the higher the quality of the MRS signal collected during the period. The smaller the peak value or temporal fluctuation, the lower the quality of the MRS signal collected during the period. The degree of body movement may be determined from an image taken by an optical camera in parallel with the collection of the MRS signal, or from the output of any other measuring instrument capable of measuring the body movement of the subject P. You may decide.

The quality of the MRS signal may be calculated based on the MRS signal as well as measured by the output from the external measuring instrument. In this case, the quality of the MRS signal means the error between the MRS signal and any reference. As a reference, any MRS signal such as a pre-collected MRS signal or an artificially generated MRS signal may be used. The error is defined by the degree of frequency deviation of other MRS spectra with respect to the reference MRS spectrum when the MRS signal is an MRS spectrum. The processing circuit 51 calculates the difference between the frequency values of the signal peaks derived from the same substance between the target MRS spectrum and the reference MRS spectrum as an error. Since the chemical shift value of water changes when the subject P moves, the error can be used as an index for measuring quality. The processing circuit 51 sets the quality of the MRS signal collected during the period to a higher value as the error is smaller, and sets the quality of the MRS signal collected during the period to a lower value as the error is larger.

(7th Embodiment)
The seventh embodiment is a modification applicable to any of the above embodiments. Hereinafter, the MR signal processing apparatus according to the seventh embodiment will be described. In the following description, components having substantially the same functions as those in the above embodiment are designated by the same reference numerals and will be described in duplicate only when necessary.

FIG. 32 is a diagram schematically showing a processing example according to the seventh embodiment. As shown in FIG. 32, the processing circuit 51 collects a plurality of MRS signals having different qualities, sorts the collected plurality of MRS signals according to the quality (step SD1), and rearranges the plurality of MRS signals. Is applied to the trained model NN to estimate the MRS reconstruction parameters. The trained model NN and MRS reconstruction parameters according to the seventh embodiment are the same as those of the sixth embodiment. The number of MRS signals input to the trained model NN may be any number as long as it is two or more, but FIG. 32 illustrates 96 MRS signals. The quality according to the seventh embodiment is the same as that of the sixth embodiment. # N (1 <n ≦ N = 96) in FIG. 32 represents the collection order of the MRS signal. The collection order is the quality order No. It is assumed that there is no correlation with n (1 <n ≦ N = 96).

In the rearrangement process (step SD1), the processing circuit 51 rearranges a plurality of MRS signals in a predetermined order. As an example, the predetermined order is the order of high quality (descending order) as shown in FIG. 32. The processing circuit 51 inputs a plurality of MRS signals to the trained model NN in the sorted order. Specifically, each channel of the input layer of the trained model NN is associated with the quality order, and the processing circuit 51 inputs the MRS signal to the channel corresponding to the quality order of the input MRS signal. Since the MRS signal is input to the trained model NN after sorting in the order of quality, the input is always performed in the order of constant quality. As a result, the quality of the MRS signal input to the same channel becomes relatively constant, and it is expected that the estimation accuracy of the MRS reconstruction parameter will be improved. The predetermined order is not limited to this, and may be any order as long as it is a fixed order such as a low quality order (ascending order). Further, although the MRS signals are rearranged according to the quality, they may be rearranged according to the number of additions.

The trained model NN according to the seventh embodiment is trained as follows. First, by the above method, the number of additions or the quality is set for the input data in the training sample. On the other hand, the number of additions or the quality ranking can be associated with each channel of the input layer of the machine learning model. Then, the processing circuit 51 trains the machine learning model based on a plurality of learning samples as described in the first to fifth embodiments. At this time, the input data of each learning sample is input to the channel corresponding to the number of additions or the quality. As a result, the trained model NN according to the seventh embodiment is trained.

(8th Embodiment)
Eighth embodiment is a modification applicable to any of the above embodiments. Hereinafter, the MR signal processing apparatus according to the eighth embodiment will be described. In the following description, components having substantially the same functions as those in the above embodiment are designated by the same reference numerals and will be described in duplicate only when necessary.

FIG. 33 is a diagram schematically showing a processing example according to the eighth embodiment. As shown in FIG. 33, the processing circuit 51 collects a plurality of MRS signals having different qualities, distributes the collected plurality of MRS signals into a use group and a rejection group based on the quality (step SE1), and uses them. The MRS signals distributed to the groups are applied to the trained model NN to estimate the MRS reconstruction parameters, and the MRS signals distributed to the rejection groups are rejected (step SE2). In other words, the processing circuit 51 rejects the MRS signal according to the quality, and applies the MRS signal other than the rejected MRS signal to the trained model NN. The trained model NN and MRS reconstruction parameters according to the eighth embodiment are the same as those of the fifth embodiment. The quality according to the eighth embodiment is the same as that of the sixth embodiment. It is assumed that # n (1 <n ≦ N) represents the collection order of the MRS signal. The number N of MRS signals used for distribution may be any number as long as it is two or more, but in FIG. 33, N = 96.

In the distribution process (step SE1), the processing circuit 51 divides each MRS signal into a use group and a rejection group according to the quality. The use group is a group to which the MRS signal having a quality exceeding the threshold value belongs. The rejection group is a group to which the MRS signal having a quality below the threshold value belongs. The processing circuit 51 compares the quality of the MRS signal with respect to the threshold value for each MRS signal, and determines whether the quality exceeds or falls below the threshold value. When the quality exceeds the threshold value, the processing circuit 51 distributes the MRS signal to the use group, and when the quality falls below the threshold value, the processing circuit 51 distributes the MRS signal to the rejection group. Then, the processing circuit 51 applies the MRS signal distributed to the use group to the trained model NN to estimate the MRS reconstruction parameter. By limiting the MRS signal used for estimating the MRS reconstruction parameter to the MRS signal whose quality exceeds the threshold value in this way, it is expected that the estimation accuracy of the MRS reconstruction parameter will be improved. Further, although the MRS signal is used or rejected according to the quality, it may be used or rejected according to the number of additions.

The trained model NN according to the eighth embodiment is trained as follows. First, by the above method, the number of additions or the quality is set for the input data in the training sample. The processing circuit 51 trains a machine learning model based on a plurality of training samples as described in the first to fifth embodiments. At this time, each learning sample is limited to those having the number of additions or quality equal to or higher than the threshold value. As a result, the trained model NN according to the eighth embodiment is learned.

(9th Embodiment)
The ninth embodiment is a modification applicable to any of the above embodiments. Hereinafter, the MR signal processing apparatus according to the ninth embodiment will be described. In the following description, components having substantially the same functions as those in the above embodiment are designated by the same reference numerals and will be described in duplicate only when necessary.

FIG. 34 is a diagram schematically showing a processing example according to the ninth embodiment. As shown in FIG. 34, the processing circuit 51 collects a plurality of MRS signals having different qualities, binning the collected plurality of MRS signals according to the quality (step SF1), and quality level (bin). The MRS signal is applied to the trained model NN in units to estimate the MRS reconstruction parameters. The trained model NN and MRS reconstruction parameters according to the ninth embodiment are the same as those of the sixth embodiment. The quality according to the ninth embodiment is the same as that of the sixth embodiment. It is assumed that # n (1 <n ≦ N) represents the collection order of the MRS signal. The number N of MRS signals used for binning may be any number as long as it is two or more, but in FIG. 34, N = 96. The quality level may be any number of stages as long as it is 2 or more stages, but in FIG. 34, it is assumed that the quality level is 3 stages as an example. It is assumed that the quality deteriorates in the order of quality levels "1", "2" and "3".

As shown in FIG. 34, a trained model NN is prepared for each quality level. Each trained model NN is a machine learning model trained to input an MRS signal of the quality level and output an MRS reconstruction parameter. The processing circuit 51 applies the MRS signal of each quality level to the trained model NN corresponding to the quality level to estimate the MRS reconstruction parameter. The processing circuit 51 calculates the final MRS reconstruction parameters based on the plurality of MRS reconstruction parameters corresponding to the plurality of quality levels. For example, the processing circuit 51 may calculate the average value of a plurality of MRS reconstruction parameters as the final MRS reconstruction parameter, or the final addition average value of the MRS reconstruction parameters weighted according to the quality level. It may be calculated as a typical MRS reconstruction parameter. Alternatively, the processing circuit 51 may use the MRS reconstruction parameter corresponding to the user-specified quality level as the final MRS reconstruction parameter and reject the MRS reconstruction parameter corresponding to another quality level. Further, although the MRS signal is binned according to the quality, it may be binned according to the number of additions.

The trained model NN according to the ninth embodiment is trained as follows. First, by the above method, the number of additions or the quality is set for the input data in the training sample. In addition, a plurality of machine learning models corresponding to a plurality of addition count levels or quality levels are prepared. The processing circuit 51 trains each machine learning model based on a plurality of training samples as described in the first to fifth embodiments. At this time, the machine learning model corresponding to each addition count level or quality level is trained based on the learning sample belonging to the level. As a result, the trained model NN according to the ninth embodiment is generated.

(10th Embodiment)
In the above embodiment, the MR signal processing device 50 is incorporated in the magnetic resonance imaging device 1. However, the MR signal processing device 50 may not be incorporated in the magnetic resonance imaging device 1. Hereinafter, the MR signal processing device 50 according to the tenth embodiment will be described. In the following description, components having substantially the same functions as those of the first to ninth embodiments are designated by the same reference numerals and will be duplicated and described only when necessary.

FIG. 35 is a diagram showing a configuration example of the magnetic resonance imaging system 100 including the MR signal processing device according to the tenth embodiment. As shown in FIG. 35, the magnetic resonance imaging system 100 includes a magnetic resonance imaging device 1, an MR signal processing device 50, and a PACS server 90 that are communicably connected to each other via a network. The MR signal processing device 50 can be applied to any of the MR signal processing devices of the first to ninth embodiments.

The magnetic resonance imaging device 1 collects MR collection signals such as MRS signals and transmits the MR collection signals to the PACS server 90. The PACS server 90 stores the MR collection signal. The PACS server 90 stores the MR collected signal and the medical signal collected by another medical diagnostic imaging apparatus such as an X-ray computer tomography apparatus so as to be searchable. The MR signal processing device 50 acquires the MR collection signal to be processed from the MR collection signal stored in the PACS server 90, and executes various processes described in the first to ninth embodiments.

According to the tenth embodiment, even when an imaging mechanism such as MR spectroscopy or MR imaging is not provided, the MRS signal or the like can be obtained by executing the various processes described in the first to ninth embodiments. The accuracy of MR collected signals can be improved.

(Addition)
According to at least one embodiment described above, the MR signal processing device 50 includes a processing circuit 51. The processing circuit 51 applies the trained model to a plurality of MRS signals collected by MR spectroscopy for the same object, and outputs parameters for MRS reconstruction. MRS reconstruction refers to the general processing for producing an MRS spectrum. For example, the MRS reconstruction is a process of generating an MRS spectrum by fitting from a spectrum model as in the first and second embodiments, and an MRS spectrum is generated by a Fourier transform from MRSk data as in the third embodiment. Including processing. The parameter for MRS reconstruction is a spectral parameter in the case of the first and second embodiments, and is a k-space modulation amount such as f0 shift and phase shift in the case of the third embodiment.

According to the above configuration, it is possible to obtain an accurate parameter, and as a result, it is possible to obtain an accurate MRS signal.

According to at least one embodiment described above, the accuracy of the MRS signal can be improved.

The term "processor" used in the above description refers to, for example, a CPU, a GPU, or an application specific integrated circuit (ASIC), a programmable logic device (for example, a simple programmable logic device (Simple Programmable Logic Device)). : SPLD), a complex programmable logic device (CPLD), and a field programmable gate array (FPGA)). The processor realizes the function by reading and executing the program stored in the storage circuit. Instead of storing the program in the storage circuit, the program may be directly embedded in the circuit of the processor. In this case, the processor realizes the function by reading and executing the program embedded in the circuit. Further, instead of executing the program, a function corresponding to the program may be realized by a combination of logic circuits. It should be noted that each processor of the present embodiment is not limited to the case where each processor is configured as a single circuit, and a plurality of independent circuits may be combined to form one processor to realize its function. good. Further, a plurality of components in FIGS. 1, 13, 20, 25 and 35 may be integrated into one processor to realize the function.

Although some embodiments have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other embodiments, and various omissions, replacements, changes, and combinations of embodiments can be made without departing from the gist of the invention. These embodiments and variations thereof are included in the scope of the invention described in the claims and the equivalent scope thereof, as are included in the scope and gist of the invention.

1 Magnetic resonance imaging device 11 Stand 13 Sleeper 21 Tilt magnetic field power supply 23 Transmission circuit 25 Reception circuit 27 Sleep drive device 29 Sequence control circuit 41 Static magnetic field magnet 43 Tilt magnetic field coil 45 Transmission coil 47 Reception coil 50 MR signal processing device 51A, 51B, 51C, 51D Processing circuit 53 Memory 55 Display 57 Input interface 58 Communication interface 131 Top plate 133 base

Claims (20)

A parameter output unit that applies a trained model to multiple MRS signals collected by MR spectroscopy for the same object and outputs parameters for MRS reconstruction.
MR signal processing device comprising. With more synthesis
The parameters include a plurality of parameters.
The plurality of parameters are spectral parameters for fitting based on a plurality of spectral models corresponding to the plurality of parameters, respectively.
The synthesizer generates an MRS spectrum for the object by fitting based on the plurality of spectral parameters and the plurality of spectral models.
The MR signal processing apparatus according to claim 1. The plurality of spectrum models are a plurality of spectra corresponding to the plurality of substances contained in the object.
The plurality of spectral parameters include the amount of substance information of the plurality of substances with respect to the subject.
The MR signal processing apparatus according to claim 2. The first MRS signal among the plurality of MRS signals is collected by performing data acquisition including application of a region selection pulse.
The second MRS signal out of the plurality of MRS signals is collected by performing a data acquisition that does not include the application of a region selection pulse.
The MR signal processing apparatus according to claim 2. The MR signal processing device according to claim 4, wherein the parameter output unit applies the plurality of MRS signals and frequency information of a frequency band selected by applying the region selection pulse to the trained model. The first MRS signal among the plurality of MRS signals is collected by performing data acquisition under the combination of the first TR and TE.
The second MRS signal among the plurality of MRS signals is collected by performing data acquisition under a second TR and TE combination different from the first TR and TE combination.
The MR signal processing apparatus according to claim 2. The plurality of spectrum models are a plurality of basis spectra corresponding to a plurality of bases obtained by applying data compression to a specific MRS spectrum.
The plurality of parameters include the plurality of basis component amount information regarding the object.
The MR signal processing apparatus according to claim 2. The MR signal processing apparatus according to claim 6, wherein the plurality of bases include a first base based on an MRS spectrum for a healthy person and a second base based on an MRS spectrum for an unhealthy person. Further equipped with a judgment unit,
The parameter output unit applies the trained model to the plurality of MRS signals to provide information on the first substance amount or component amount regarding the first basis and the second substance amount or component amount regarding the second basis. Output information and
The determination unit determines whether the target is normal or abnormal based on the first substance amount or component amount information and the second substance amount or component amount information.
The MR signal processing apparatus according to claim 8. The MR signal processing apparatus according to claim 9, further comprising a display unit for displaying a determination result by the determination unit. The MR signal processing device according to claim 2, wherein one or more MRS signals among the plurality of MRS signals are MRS signals generated by the synthesizer. The MR signal processing apparatus according to claim 1, wherein the parameter further includes reliability information. The MR signal processing apparatus according to claim 1, wherein the parameter is modulation amount information of k-spatial modulation caused by magnetic field modulation generated during data acquisition for the plurality of MRS signals. The plurality of MRS signals are a plurality of k-space data collected by MR spectroscopy.
Further, a correction unit for correcting the average k-space data based on at least one k-space data among the plurality of k-space data based on the modulation amount information is provided.
13. The MR signal processing apparatus according to claim 13. The MR signal processing apparatus according to claim 14, further comprising a generation unit that generates an MRS spectrum based on the corrected average k-space data. With multiple compositing units
The parameter output unit has a plurality of first parameter output units and a second parameter output unit.
The plurality of first parameter output units each apply the first trained model to the plurality of MRS signals to output a plurality of quantity information regarding a substance or a basis.
The plurality of synthesis units generate a plurality of synthetic MRS signals for the target based on the quantity information, respectively.
The second parameter output unit applies the second trained model to the plurality of synthetic MRS signals, and modulates k-space modulation due to magnetic field modulation that occurs during data acquisition for the plurality of MRS signals. Output quantity information,
The MR signal processing apparatus according to claim 1. Further equipped with a first synthesis part and a second synthesis part,
The parameter output unit has a first parameter output unit, a second parameter output unit, and a third parameter output unit.
The first parameter output unit applies the first trained model to one of the plurality of MRS signals and outputs the first quantity information regarding the substance or the basis.
The second parameter output unit applies the first trained model to the plurality of MRS signals and outputs a second quantity information regarding a substance or a basis.
The first synthesis unit generates a first synthetic MRS signal for the subject based on the first quantity information.
The second synthesis unit generates a second synthetic MRS signal for the object based on the second quantity information.
The third parameter output unit is generated during data acquisition for the plurality of MRS signals by applying a second trained model to the first synthetic MRS signal and the second synthetic MRS signal. Outputs modulation amount information of k-space modulation caused by magnetic field modulation,
The MR signal processing apparatus according to claim 1. The parameter output unit is
The plurality of MRS signals sorted according to the number of additions or quality are applied to the trained model, or
An MRS signal other than the MRS signal rejected according to the number of additions or the quality of the plurality of MRS signals is applied to the trained model.
The MR signal processing apparatus according to claim 1. The trained model has a plurality of models corresponding to a plurality of quality levels, respectively.
The parameter output unit binning each of the plurality of MRS signals according to the quality, and applies the MRS signal to the model in units of quality level.
The MR signal processing apparatus according to claim 1. A parameter output unit that applies a trained model to a plurality of MR collection signals for the same object and outputs a modulation amount related to k-space modulation caused by magnetic field modulation generated during data collection for the plurality of MR collection signals.
MR signal processing device comprising.

Images