Magnetic resonance image generation method and magnetic resonance image generation device

Abstract

To provide a magnetic resonance image generation method and a magnetic resonance image generation device capable of generating a reconfiguration image with a high degree of precision on the basis of simultaneous imaging of a plurality of cross-sectional images.SOLUTION: A magnetic resonance image generation method includes: an excitation step of selectively exciting each cross section constituting a plurality of cross sections of a subject; a phase modulation step of giving different phase modulations to each cross section; a two-dimensional image signal generation step of generating a two-dimensional image signal S1 in which image signals imaged in each cross section are superimposed; a primary reproduction image signal generation step of generating a primary reproduction image signal S2 including an image signal focused on a specific cross section of the plurality of cross sections; and a reconfiguration image signal generation step of generating a reconfiguration image signal S3 focused on the specific cross section from the primary reproduction image signal S2 using a machine learning model ML constructed with the primary reproduction image signals S2 acquired beforehand and a target image signal SG acquired beforehand as the image signal focused on the specific cross section as teacher data.SELECTED DRAWING: Figure 8

Description

[Technical Field]

[0001] The present disclosure relates to a magnetic resonance image generating method and a magnetic resonance image generating apparatus. [Background technology]

[0002] The magnetic resonance imaging (MRI) method described in Patent Document 1 includes the steps of: inputting a two-dimensional measurement image signal in which information on each cross section of a subject is visualized in a superimposed manner; obtaining a first focused image based on the measurement image signal, the first focused image having a desired depth position coordinate as a focal plane for the two-dimensional measurement image signal; obtaining a point spread function indicating a blurring effect caused by image components other than the focal plane being superimposed on the focused image; and performing an inverse Fourier transform on an absolute value image of the point spread function and an absolute value amplitude image of the obtained first focused image to obtain an incoherent imaging approximation image in which some or all of the image components other than the focal plane have been removed. As a result, the magnetic resonance imaging method described in Patent Document 1 can generate a clear reconstructed image by removing the image components other than the focal plane. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2005-253702 Summary of the Invention [Problem to be solved by the invention]

[0004] The above-mentioned magnetic resonance image generation method uses the maximum likelihood method as an example of a method for removing image components outside the focal plane. Specifically, an iterative deconvolution integral is used to perform image correction by maximizing the likelihood, and the original image components are asymptotically estimated. However, while this method is effective for estimating images of linear structures such as blood vessels, it has been difficult to estimate images of cross-sectional images with filled interiors.

[0005] The present disclosure provides a magnetic resonance image generating method and a magnetic resonance image generating device that can generate a highly accurate reconstructed image based on simultaneous imaging of a plurality of cross-sectional images. [Means for solving the problem]

[0006] The gist of the present disclosure is as follows.

[0007] [1] A magnetic resonance image generating method comprising: an excitation step for selectively exciting each cross section constituting a plurality of cross sections of a subject; a phase modulation step for applying different phase modulation to each of the cross sections; a two-dimensional image signal generating step for generating a two-dimensional image signal in which image signals captured at each of the cross sections are superimposed; a primary reproduction image signal generating step for generating a primary reproduction image signal including an image signal focused on a specific cross section among the plurality of cross sections; and a reconstruction image signal generating step for generating a reconstructed image signal focused on the specific cross section from the primary reproduction image signal using a machine learning model constructed using as training data the primary reproduction image signal acquired in advance and a target image signal obtained in advance as an image signal focused on the specific cross section.

[0008] In the magnetic resonance image generating method, the excitation step selectively excites each cross section, and the phase modulation step imparts a different phase modulation to each cross section. As a result, the amount of phase modulation for each cross section differs for each cross section. Then, in the two-dimensional image signal generating step, a two-dimensional image signal in which image signals for each cross section are superimposed is acquired. In the primary reconstruction image signal generating step, a primary reconstruction image signal including an image signal focused on a specific cross section is generated, for example, by canceling out only the phase modulation amount for a specific cross section. Furthermore, in the reconstruction image signal generating step, a machine learning model can be used to efficiently and reliably extract only an image focused on a specific cross section. As described above, the magnetic resonance image generating method efficiently generates a clear reconstruction image, thereby enabling high-precision images to be generated based on simultaneous imaging of multiple cross-sectional images.

[0009] [2] The magnetic resonance image generating method according to [1], wherein the reconstructed image signal generating step uses a plurality of machine learning models corresponding to the cross sections constituting the plurality of cross sections to generate a plurality of reconstructed image signals focused on the respective cross sections. In this case, by utilizing a plurality of machine learning models, reconstructed image signals focused on the respective cross sections can be generated simultaneously, and reconstructed images can be generated more efficiently.

[0010] [3] The magnetic resonance image generating method according to [1] or [2], wherein the machine learning model is a deep learning model. In this case, the deep learning model can capture complex patterns and hierarchical features of the primary reconstructed image signal, thereby improving the accuracy of generating the reconstructed image signal.

[0011] [4] The magnetic resonance image generating method according to any one of [1] to [3], wherein in the excitation step, a gradient magnetic field is applied in a depth direction, which is the direction in which the multiple cross sections are arranged, and an excitation pulse corresponding to each of the multiple cross sections is applied. In this case, by applying a gradient magnetic field, the rotation frequency of protons varies for each cross section along the depth direction. By applying an excitation pulse corresponding to the rotation frequency in each cross section, it is possible to easily excite each cross section selectively.

[0012] [5] A magnetic resonance image generating device comprising: an excitation unit that selectively excites each cross section constituting a plurality of cross sections of a subject; a phase modulation unit that imparts different phase modulation to each of the cross sections; a two-dimensional image signal generating unit that generates a two-dimensional image signal in which image signals captured at each of the cross sections are superimposed; a primary reproduction image signal generating unit that generates a primary reproduction image signal including an image signal focused on a specific cross section among the plurality of cross sections; and a reconstruction image signal generating unit that generates a reconstructed image signal focused on the specific cross section from the primary reproduction image signal using a machine learning model constructed using as training data the primary reproduction image signal that has been acquired in advance and a target image signal that has been obtained in advance as an image signal focused on the specific cross section.

[0013] In the magnetic resonance imaging device, the excitation unit selectively excites each cross section, and the phase modulation step imparts different phase modulation to each cross section. As a result, the amount of phase modulation in each cross section varies. The two-dimensional image signal generation unit then acquires a two-dimensional image signal in which the image signals of each cross section are superimposed, and the primary reconstruction image signal generation unit generates a primary reconstruction image signal including an image signal focused on the specific cross section, for example, by canceling out only the phase modulation amount of the specific cross section. Furthermore, the reconstructed image signal generation unit can efficiently and reliably extract only an image focused on the specific cross section using a machine learning model. As a result, the magnetic resonance imaging device efficiently generates clear reconstructed images, thereby enabling high-precision images to be generated based on the simultaneous capture of multiple cross-sectional images. [Effects of the Invention]

[0014] According to one exemplary embodiment, a technique is provided that allows for easily generating a clear reconstructed image. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a diagram showing an example of the configuration of a magnetic resonance image generating device. [Figure 2] FIG. 2 is a flowchart showing an example of a magnetic resonance image generating method using the magnetic resonance image generating device. [Figure 3] Fig. 3(a) is a diagram showing an example of a sequence of excitation pulses. Fig. 3(b) is a diagram showing an example of a sequence of gradient magnetic fields in the z-axis direction. Fig. 3(c) is a diagram showing an example of a sequence of gradient magnetic fields in the y-axis direction. Fig. 3(d) is a diagram showing an example of a sequence of gradient magnetic fields in the x-axis direction. Fig. 3(e) is a diagram showing an example of a sequence of composite echo signals. [Figure 4] FIG. 4 is a diagram showing an example of generation of a two-dimensional image signal. [Figure 5] FIG. 5 is a diagram showing an example of an image signal generated by the magnetic resonance image generating device. [Figure 6]FIG. 6 is a flowchart illustrating an example of a method for generating a machine learning model. [Figure 7] FIG. 7 is a diagram illustrating an example of a method for generating a machine learning model. [Figure 8] FIG. 8 is a diagram illustrating an example of generation of a reconstructed image signal using a machine learning model. [Figure 9] FIG. 9 is a diagram illustrating an example of images used in a plurality of machine learning models and images generated by the plurality of machine learning models. DETAILED DESCRIPTION OF THE INVENTION

[0016] Various exemplary embodiments will be described in detail below with reference to the drawings, in which the same or equivalent parts are designated by the same reference numerals.

[0017] [Configuration of magnetic resonance imaging device] A magnetic resonance image generating device is a device that generates a primary reconstructed image including an image focused on a specific cross section from a two-dimensional image in which image signals from each cross section of a subject are superimposed, and extracts only the image focused on the specific cross section from the primary reconstructed image using a machine learning model. Figure 1 is a diagram showing an example of the configuration of the magnetic resonance image generating device 1. In the following example, the multiple cross sections of subject A will be described as consisting of three cross sections: cross section D1, cross section D2, and cross section D3.

[0018] A subject A is placed in a predetermined space in a magnetic resonance image generating apparatus 1. The magnetic resonance image generating apparatus 1 includes a static magnetic field coil 2, a gradient magnetic field coil 3, an irradiation coil 4, a receiving coil 5, a highly stable DC power supply unit 6, a gradient magnetic field generating unit 7, an excitation unit 8, a receiving circuit 9, a control unit 10, and a display unit 20.

[0019] The static magnetic field coil 2 generates a uniform static magnetic field. The static magnetic field coil 2 is configured by, for example, an electromagnet, an air-core coil magnet, a superconducting magnet, or a permanent magnet, and forms a uniform static magnetic field in a predetermined space in which the subject A is placed.

[0020] The gradient magnetic field coil 3 applies gradient magnetic fields in three axial directions, i.e., the x-axis, y-axis, and z-axis, to the subject A. Here, the z-axis direction corresponds to the body axis direction of the subject A. The y-axis direction corresponds to the dorsoventral axis direction of the subject A and is perpendicular to the z-axis direction. The x-axis direction is perpendicular to the y-axis and z-axis directions. The gradient magnetic field coil 3 is composed of, for example, parallel wire, rectangular coils, or circular coils. The gradient magnetic field coil 3 is controlled by the gradient magnetic field generating unit 7, and applies to the subject A a gradient magnetic field Gx having a linear gradient in the x-axis direction, a gradient magnetic field Gy having a linear gradient in the y-axis direction, and a gradient magnetic field Gz having a linear gradient in the z-axis direction.

[0021] The irradiation coil 4 irradiates the subject A with an excitation pulse. The irradiation coil 4 is placed near the subject A. The irradiation coil 4 irradiates the subject A with an excitation pulse input from the excitation unit 8. For example, the irradiation coil 4 applies a 90-degree pulse, a 180-degree pulse, or a pulse of any angle as the excitation pulse to the subject A.

[0022] The receiving coil 5 detects an echo signal (magnetic resonance signal). The receiving coil 5 is composed of a single coil or multiple coils and is placed near the subject A. The receiving coil 5 detects the echo signal (magnetic resonance signal), which is an electromagnetic wave emitted when the excitation pulse irradiated from the irradiation coil 4 magnetically resonates with the atomic nuclei that make up the tissue of the subject A.

[0023] The highly stable DC power supply unit 6 drives the static magnetic field coil 2. The highly stable DC power supply unit 6 drives the static magnetic field coil 2 to form a uniform static magnetic field within a predetermined space as described above.

[0024] The gradient magnetic field generating unit 7 applies different phase modulation to each of the multiple cross sections D1 to D3 of the subject A. The gradient magnetic field generating unit 7 may also be referred to as a phase modulation unit. The gradient magnetic field generating unit 7 drives the gradient magnetic field coil 3. The gradient magnetic field generating unit 7 generates gradient magnetic fields Gx, Gy, and Gz in three axial directions, and controls the gradient magnetic field coil 3 to apply the gradient magnetic fields Gx, Gy, and Gz within a predetermined space. As will be described in detail later, for example, by applying a gradient magnetic field Gy in the y-axis direction for a predetermined time to each cross section of the subject A, phase modulation can be applied to each cross section. In this case, by applying the gradient magnetic field Gy for a different time to each cross section, different phase modulation can be applied to each cross section.

[0025] The excitation unit 8 selectively excites each of the cross sections constituting the multiple cross sections D1 to D3 of the subject A. The excitation unit 8 includes, for example, a high-frequency oscillator and generates a high-frequency signal. Then, based on the generated high-frequency signal, the excitation unit 8 generates an excitation pulse having a predetermined frequency spectrum and a pulse form with an appropriate width and height, which has a high-frequency magnetic field equal to the Larmor frequency corresponding to the magnetic field strength of the gradient magnetic field in which the subject A is placed. The excitation unit 8 amplifies the excitation pulse to a predetermined signal level and outputs it to the irradiation coil 4. The excitation unit 8 selectively outputs, for example, an excitation pulse RF1 that excites the cross section D1, an excitation pulse RF2 that excites the cross section D2, and an excitation pulse RF3 that excites the cross section D3 to the irradiation coil 4.

[0026] The receiving circuit 9 receives the echo signals detected by the receiving coil 5. Since the excitation unit 8 selectively excites each of the slices constituting the multiple slices D1 to D3, the receiving circuit 9 receives, for example, an echo signal corresponding to the slice D1, an echo signal corresponding to the slice D2, and an echo signal corresponding to the slice D3. The receiving circuit 9 may then combine the echo signals corresponding to the slices into one combined echo signal SE. The receiving circuit 9 converts the combined echo signal SE from analog to digital and then outputs it to the control unit 10.

[0027] The control unit 10 includes a two-dimensional image signal generating unit 11, a primary reproduced image signal generating unit 12, a reconstructed image signal generating unit 13, and a system control unit .

[0028] The two-dimensional image signal generator 11 receives the composite echo signal SE from the receiving circuit 9. The two-dimensional image signal generator 11 generates a two-dimensional image signal S1 based on the composite echo signal SE. The two-dimensional image signal S1 is an image signal in which image signals captured at each of the multiple cross sections D1 to D3 are superimposed. The two-dimensional image signal generator 11 outputs the generated two-dimensional image signal S1 to the primary reconstruction image signal generator 12.

[0029] The primary reproduction image signal generator 12 generates a primary reproduction image signal S2 based on the two-dimensional image signal S1. The primary reproduction image signal S2 includes an image signal focused on a specific cross section among the multiple cross sections D1 to D3. For example, if the specific cross section is cross section D1, the primary reproduction image signal S2 includes an image signal focused on cross section D1, as well as image signals of out-of-focus cross sections D2 and D3. The primary reproduction image signal generator 12 outputs the primary reproduction image signal S2 to the reconstruction image signal generator 13.

[0030] The reconstructed image signal generation unit 13 generates a reconstructed image signal S3 based on the primary reconstructed image signal S2. The reconstructed image signal generation unit 13 uses a machine learning model to generate the reconstructed image signal S3. The reconstructed image signal generation unit 13 may use, for example, a machine learning model stored in a memory unit within the reconstructed image signal generation unit 13, or may use a machine learning model stored in a memory unit provided within or external to the control unit 10. The machine learning model is constructed using, as training data, the primary reconstructed image signal S2 acquired in advance and a target image signal previously obtained as an image signal focused on a specific cross section. The reconstructed image signal generation unit 13 generates the reconstructed image signal S3 by inputting the primary reconstructed image signal S2 newly output from the primary reconstructed image signal generation unit 12 into the machine learning model. The reconstructed image signal S3 is an image signal focused on a specific cross section and does not include out-of-focus image signals. The reconstructed image signal S3 is, for example, an image signal focused on the cross section D1. The reconstructed image signal generating unit 13 may output the generated reconstructed image signal S3 to the outside of the control unit 10. For example, the reconstructed image signal generating unit 13 causes the display unit 20 to display the reconstructed image signal S3.

[0031] The system control unit 14 controls the highly stable DC power supply unit 6, the gradient magnetic field generation unit 7, and the excitation unit 8 based on a pulse sequence (see FIG. 3) described later. For example, the system control unit 14 may instruct the excitation unit 8 to generate an excitation pulse for which cross section. The system control unit 14 may instruct the gradient magnetic field generation unit 7 as to the order in which to apply the gradient magnetic fields Gx, Gy, and Gz to the subject A. The system control unit 14 may instruct the highly stable DC power supply unit 6 as to the timing at which to drive the static magnetic field coil 2.

[0032] [Magnetic resonance imaging method] Next, the operation of the magnetic resonance image generating device 1 will be described in detail with reference to Fig. 2 and Fig. 3. Fig. 2 is a flowchart showing an example of a magnetic resonance image generating method using the magnetic resonance image generating device 1. Fig. 3 is a diagram showing an example of a pulse sequence when acquiring a composite echo signal SE using the magnetic resonance image generating device 1. The magnetic resonance image generating method of Fig. 2 starts operation when, for example, an operation instruction is received from an operator of the device.

[0033] In step ST11, the highly stable DC power supply unit 6 drives the static magnetic field coil 2 under instructions from the system control unit 14. The static magnetic field coil 2 forms a uniform static magnetic field within a predetermined space.

[0034] Next, in step ST12, the gradient magnetic field generating unit 7 generates a gradient magnetic field Gz in the z-axis direction under instructions from the system control unit 14. As shown in FIG. 3(b), the gradient magnetic field coil 3 applies the gradient magnetic field Gz to the subject A so as to superimpose it on the static magnetic field. By applying the gradient magnetic field Gz, the magnetic field strength in the z-axis direction changes in each of the cross sections D1, D2, and D3. As a result, the Larmor frequencies in each cross section differ from one another.

[0035] Next, in step ST13, the excitation unit 8 selectively excites the cross section D1 of the subject A under instructions from the system control unit 14. Step ST13, step ST16 (described later), and step ST19 (described later) correspond to excitation steps that selectively excite each cross section constituting the multiple cross sections D1 to D3 of the subject A. In the excitation steps, excitation pulses corresponding to each cross section are applied in a state in which a gradient magnetic field is applied in the Z-axis direction (depth direction), which is the direction in which the multiple cross sections D1 to D3 are arranged. As shown in FIG. 3(a), the excitation unit 8 generates an excitation pulse RF1 having a radio frequency magnetic field equal to the Larmor frequency at the cross section D1 and outputs it to the irradiation coil 4. The irradiation coil 4 applies the excitation pulse RF1 to the subject A. As shown in FIGS. 3(a) and 3(b), the period during which the excitation pulse RF is applied and the period during which the gradient magnetic field Gz is applied are synchronized. In the example of FIG. 3, the gradient magnetic field generation unit 7 stops applying the gradient magnetic field Gz every time the application of the excitation pulse RF is completed. The receiving coil 5 detects an echo signal (not shown in FIG. 3) corresponding to the cross section D1 generated by application of the excitation pulse RF1, and outputs the signal to the receiving circuitry 9.

[0036] Next, in step ST14, the gradient magnetic field generating unit 7 generates a gradient magnetic field Gy in the y-axis direction under instructions from the system control unit 14. Step ST14, step ST17 (described later), and step ST20 (described later) correspond to phase modulation steps that impart different phase modulation to each cross section. As shown in FIG. 3(c), the gradient magnetic field generating unit 7 drives the gradient magnetic field coil 3 to apply a gradient magnetic field Gy1 for an application time t1. The y-axis direction coincides with the phase encoding direction. This imparts a phase modulation corresponding to the application time t1 to the cross section D1.

[0037] Subsequently, in step ST15, the gradient magnetic field generating unit 7 generates a gradient magnetic field Gz in the z-axis direction again under the instruction of the system control unit 14, and the gradient magnetic field coil 3 applies the gradient magnetic field Gz to the subject A.

[0038] Subsequently, in step ST16, the excitation unit 8 selectively excites a cross section D2 of the subject A under instructions from the system control unit 14. As shown in FIG. 3( a), the excitation unit 8 generates an excitation pulse RF2 having a high-frequency magnetic field equal to the Larmor frequency at the cross section D2, and outputs it to the irradiation coil 4. The irradiation coil 4 applies the excitation pulse RF2 to the subject A. The receiving coil 5 detects an echo signal (not shown in FIG. 3) corresponding to the cross section D2 generated by application of the excitation pulse RF2, and outputs it to the receiving circuitry 9.

[0039] Next, in step ST17, the gradient magnetic field generation unit 7 generates a gradient magnetic field Gy in the y-axis direction under instructions from the system control unit 14. As shown in FIG. 3(c), the gradient magnetic field generation unit 7 drives the gradient magnetic field coil 3 so as to apply a gradient magnetic field Gy2 for an application time t2. As a result, the gradient magnetic field generation unit 7 imparts a phase modulation corresponding to the application time t2 to the cross section D2. At the same time, the gradient magnetic field generation unit 7 also imparts a phase modulation corresponding to the application time t2 to the cross section D1, which has already been excited. The phase modulation for the cross section D1 is the sum of the phase modulation corresponding to the application time t1 and the phase modulation corresponding to the application time t2.

[0040] Subsequently, in step ST18, the gradient magnetic field generating unit 7 generates a gradient magnetic field Gz in the z-axis direction again under the instruction of the system control unit 14, and the gradient magnetic field coil 3 applies the gradient magnetic field Gz to the subject A.

[0041] Subsequently, in step ST19, the excitation unit 8 selectively excites a cross section D3 of the subject A under instructions from the system control unit 14. As shown in FIG. 3( a), the excitation unit 8 generates an excitation pulse RF3 having a radio frequency magnetic field equal to the Larmor frequency at the cross section D3, and outputs it to the irradiation coil 4. The irradiation coil 4 applies the excitation pulse RF3 to the subject A. The receiving coil 5 detects an echo signal (not shown in FIG. 3 ) corresponding to the cross section D3 generated by the application of the excitation pulse RF3, and outputs it to the receiving circuitry 9.

[0042] Next, in step ST20, the gradient magnetic field generator 7 generates a gradient magnetic field Gy in the y-axis direction under instructions from the system control unit 14. As shown in FIG. 3(c), the gradient magnetic field generator 7 drives the gradient magnetic field coil 3 to apply a gradient magnetic field Gy3 for application time t3. As a result, the gradient magnetic field generator 7 imparts a phase modulation corresponding to application time t3 to the cross section D3. At the same time, the gradient magnetic field generator 7 also imparts a phase modulation corresponding to application time t3 to the cross sections D1 and D2 that have already been excited. The phase modulation for the cross section D1 is the sum of the phase modulation corresponding to application time t1, the phase modulation corresponding to application time t2, and the phase modulation corresponding to application time t3. The phase modulation for the cross section D2 is the sum of the phase modulation corresponding to application time t2 and the phase modulation corresponding to application time t3.

[0043] Next, in step ST21, the gradient magnetic field generation unit 7 generates a gradient magnetic field Gx in the x-axis direction under instructions from the system control unit 14, and the gradient magnetic field coil 3 applies the gradient magnetic field Gx to the subject A. The x-axis direction coincides with the frequency encoding direction. As shown in FIGS. 3(d) and 3(e), the receiving circuitry 9 combines the echo signals corresponding to the cross sections D1, D2, and D3 into one combined echo signal SE in synchronization with the period during which the gradient magnetic field Gx is applied, and outputs the combined echo signal SE to the two-dimensional image signal generation unit 11.

[0044] The above-described steps ST12 to ST21 are operations performed during one repetition period. That is, steps ST12 to ST21 may be repeated a number of times equal to the number n of phase encodings (n is an integer equal to or greater than 1, for example). In step ST22, it is determined whether steps ST12 to ST21 have been repeated a number of times equal to the number n of phase encodings. If steps ST12 to ST21 have not been repeated a number of times equal to the number n of phase encodings (step ST22: NO), steps ST12 to ST21 are performed in the second and subsequent repetition periods.

[0045] If the process has been repeated n times, which is the number of phase encodings (step ST22: YES), then in step ST23, the two-dimensional image signal generator 11 generates a two-dimensional image signal S1. Step ST23 corresponds to a two-dimensional image signal generating step of generating a two-dimensional image signal S1 in which image signals captured at each cross section are superimposed. FIG. 4 is a diagram showing an example of the generation of the two-dimensional image signal S1. In the example of FIG. 4, the two-dimensional image signal S1 is generated from n points (n1 to n) in the phase encoding direction (y-axis direction). n ) and m points (m1 to m) in the frequency encoding direction (x-axis direction). m The two-dimensional image signal generation unit 11 is configured from signal acquisition points (m, m is an integer equal to or greater than 1). For example, in the first repetition time, the two-dimensional image signal generation unit 11 acquires signals at m points in n1 rows from the composite echo signal SE acquired in step ST21. For example, in the second repetition time, the two-dimensional image signal generation unit 11 acquires signals at m points in n2 rows from the composite echo signal SE acquired in step ST21. For example, the two-dimensional image signal generation unit 11 repeats such signal acquisition a number of times equal to the number n of phase encodings along the phase encoding direction, thereby acquiring signals at m points x n points and generating a two-dimensional image signal S1.

[0046] When the two-dimensional image signal S1 is expressed by time t and spatial frequency v(t, Gy) of the gradient magnetic field Gy, the two-dimensional image signal S1 is expressed by the Fourier transform form of the spin density of the subject A as shown in Equation 1.

number

[0047] The equation obtained by transforming the variables of Equation 1 is expressed as Equation 2.

number

[0048] Referring again to FIG. 2, in step ST24, the primary reconstruction image signal generator 12 generates a primary reconstruction image signal S2 based on the two-dimensional image signal S1. Step ST24 corresponds to a primary reconstruction image signal generating step of generating a primary reconstruction image signal S2 including an image signal focused on a specific cross section among the multiple cross sections D1 to D3. The primary reconstruction image signal S2 includes an image signal focused on a specific cross section among the multiple cross sections D1 to D3. FIG. 5 is a diagram showing an example of an image signal generated by the magnetic resonance imaging device 1. In the example of FIG. 5, as examples of the primary reconstruction image signal S2, a primary reconstruction image signal S21 focused on the cross section D1, a primary reconstruction image signal S22 focused on the cross section D2, and a primary reconstruction image signal S23 focused on the cross section D3 are shown. For example, the primary reconstruction image signal S21 is an image signal obtained by adding together an image signal focused on the cross section D1 and image signals of the out-of-focus cross sections D2 and D3.

[0049] The primary reproduction image signal generator 12 performs an inverse Fourier transform on the two-dimensional image signal S1 and generates a primary reproduction image signal S2 by applying to the two-dimensional image signal S1 an inverse phase of the phase modulation applied to a cross section on which the user wishes to focus among the multiple cross sections D1 to D3. The primary reproduction image signal S2 when the cross section D1 is focused is expressed by Equation 3.

number

[0050] Subsequently, in step ST25, the reconstructed image signal generating unit 13 generates a reconstructed image signal S3 based on the primary reconstructed image signal S2. Step ST25 corresponds to a reconstructed image signal generating step of generating a reconstructed image signal S3 focused on a specific cross section from the primary reconstructed image signal S2 using a machine learning model. The machine learning model used in step ST25 may be generated at any time in the magnetic resonance image generating method. For example, the machine learning model may be generated before step ST11.

[0051] FIG. 6 is a flowchart showing an example of a learning method for a machine learning model. First, in step ST31, a primary reconstructed image signal S2 previously acquired for learning and a target image signal previously acquired as an image signal focused on a specific cross section are acquired as training data. The primary reconstructed image signal S2 and the target image signal may be acquired externally, or may be acquired from an image signal generated by the magnetic resonance image generation method shown in FIG. 2 and serving as input data for the machine learning model. For example, 500 primary reconstructed image signals S2 and 500 target image signals previously acquired are acquired. In this case, the primary reconstructed image signals S2 used as input data may be images different from or the same as the primary reconstructed image signals S2 previously acquired for learning. If the images are the same, a portion of the 500 primary reconstructed image signals S2 previously acquired may be used as the primary reconstructed image signal S2 used as input data.

[0052] Next, in step ST32, a learning process for the machine learning model is executed. FIG. 7 is a diagram showing an example of the learning process for the machine learning model ML. In the example of FIG. 7, multiple machine learning models ML1 to ML3 corresponding to the multiple cross sections D1 to D3 are learned as the machine learning model ML. In the example of FIG. 7, examples of the target image signal SG include a target image signal SG1 that is focused on the cross section D1 and does not include an out-of-focus image signal, a target image signal SG2 that is focused on the cross section D2 and does not include an out-of-focus image signal, and a target image signal SG3 that is focused on the cross section D3 and does not include an out-of-focus image signal. In this case, the machine learning model ML1 is a machine learning model that uses the primary reproduced image signal S21 and the target image signal SG1 as training data. The machine learning model ML2 is a machine learning model that uses the primary reproduced image signal S22 and the target image signal SG2 as training data. The machine learning model ML3 is a machine learning model that uses the primary reproduced image signal S23 and the target image signal SG3 as training data. Each of the machine learning models ML1 to ML3 generates a reconstructed image signal S3 corresponding to each of the cross sections D1 to D3. In training each of the machine learning models ML1 to ML3, for example, 500 images each of the plurality of primary reconstructed image signals S21 to S23 and the plurality of target image signals SG1 to SG3 acquired in advance are used, and training is performed for 100 epochs.

[0053] The machine learning model ML may be a deep learning model. The machine learning model ML may be, for example, a convolutional neural network. In the following example, the machine learning model ML is configured using a convolutional neural network. In this case, the machine learning model ML may include a convolutional layer and a pooling layer. The convolutional layer performs a "convolution" operation on the input data, the primary reconstructed image signal S2. For example, the convolutional layer slides a small window (filter or kernel) over the image, calculating the sum of the products of the pixels in the window and each element of the filter. The pooling layer performs a "pooling" operation to reduce the dimensionality of the output of the convolutional layer. The pooling layer compresses information by, for example, taking the maximum or average value within a 2x2 window. The machine learning model ML is designed, for example, using a complex-valued U-Net to enable each component of the network, such as the convolutional layer, to handle complex numbers. This enables more effective learning of data features, including complex phase information, and more accurate image reconstruction.

[0054] The design parameters of machine learning models ML1, ML2, and ML3 may all be the same or different. When the design parameters are different, for example, the size of the filter used in the convolution layer may differ depending on the machine learning model. Alternatively, whether to take the maximum value or the average value within a window in the pooling layer may differ depending on the machine learning model.

[0055] Referring again to FIG. 2, in step ST25, the reconstructed image signal generation unit 13 generates a reconstructed image signal S3 based on the primary reconstructed image signal S2. FIG. 8 is a diagram showing an example of generation of the reconstructed image signal S3 using multiple machine learning models ML1 to ML3. The reconstructed image signal generation unit 13 generates multiple reconstructed image signals S31 to S33 focused on each cross section as the reconstructed image signal S3. In the example of FIG. 8, the reconstructed image signal generation unit 13 uses machine learning model ML1 to generate a reconstructed image signal S31 that is focused on the cross section D1 and does not include out-of-focus image signals from the primary reconstructed image signal S21. The reconstructed image signal generation unit 13 uses machine learning model ML2 to generate a reconstructed image signal S32 that is focused on the cross section D2 and does not include out-of-focus image signals from the primary reconstructed image signal S22. The reconstructed image signal generation unit 13 uses machine learning model ML3 to generate a reconstructed image signal S33 that is focused on the cross section D3 and does not include out-of-focus image signals from the primary reconstructed image signal S23.

[0056] The generation of the reconstructed image signal S31, the generation of the reconstructed image signal S32, and the generation of the reconstructed image signal S33 may be performed simultaneously or sequentially. If they are performed sequentially, any order is acceptable.

[0057] FIG. 9 shows an example of images used in multiple machine learning models ML1 to ML3 and images generated by the multiple machine learning models ML1 to ML3. In the example of FIG. 9, multiple target image signals SG1 to SG3 are shown for comparison with multiple reconstructed image signals S31 to S33. In the example of FIG. 9, the primary reconstructed image signals S2 used as input data are all different images from the primary reconstructed image signal S2 acquired in advance for learning. As shown in FIG. 9, out-of-focus image components are appropriately removed from each of the multiple reconstructed image signals S31 to S33. The multiple reconstructed image signals S31 to S33 achieve image quality with the same accuracy as the multiple target image signals SG1 to SG3.

[0058] [Action and effect] As described above, in a magnetic resonance image generating method using a magnetic resonance image generating device 1 according to one aspect of the present disclosure, in the excitation step, each of the slices constituting the multiple slices D1 to D3 is selectively excited, and in the phase modulation step, a different phase modulation is applied to each slice. As a result, the amount of phase modulation for each slice varies. Subsequently, in the two-dimensional image signal generating step, a two-dimensional image signal S1 in which image signals for each slice are superimposed is acquired. In the primary reconstructed image signal generating step, a primary reconstructed image signal S2 including an image signal focused on the specific slice is generated, for example, by canceling out only the phase modulation amount for a specific slice. Furthermore, in the reconstructed image signal generating step, a machine learning model ML can be used to efficiently and reliably extract only an image focused on the specific slice. As described above, the magnetic resonance image generating method efficiently generates a clear reconstructed image signal S3, thereby generating a highly accurate reconstructed image based on the simultaneous imaging of multiple slice images for the multiple slices D1 to D3. In addition, the magnetic resonance image generating method does not require a dedicated magnetic field encoding coil, making it an imaging method that can be readily applied to general-purpose MRI. Furthermore, the magnetic resonance image generating device 1 can simultaneously image a plurality of cross sections D1 to D3, thereby enabling the imaging of a plurality of cross sections D1 to D3 at a higher speed than with a general-purpose MRI.

[0059] In the reconstructed image signal generating step, a plurality of machine learning models ML1 to ML3 corresponding to the respective cross sections constituting the plurality of cross sections D1 to D3 may be used to generate a plurality of reconstructed image signals S31 to S33 focused on each cross section. In this case, by utilizing the plurality of machine learning models ML1 to ML3, a plurality of reconstructed image signals S31 to S33 focused on each cross section can be simultaneously generated, and the reconstructed image signal S3 can be generated more efficiently.

[0060] The machine learning model ML may be a deep learning model, which can capture complex patterns and hierarchical features of the primary reconstructed image signal S2 and improve the accuracy of generating the reconstructed image signal S3.

[0061] In the excitation step, a gradient magnetic field Gz may be applied in the depth direction (z-axis direction), which is the direction in which the multiple cross sections D1 to D3 are arranged, and excitation pulses RF1 to RF3 corresponding to each of the multiple cross sections D1 to D3 may be applied. In this case, by applying the gradient magnetic field Gz, the Larmor frequency differs for each cross section along the z-axis direction. By applying excitation pulses RF1 to RF3 corresponding to the Larmor frequency of each cross section, it is possible to easily excite each cross section selectively.

[0062] [Variations] Although the embodiments of the present disclosure have been described above, the present disclosure is not necessarily limited to the above-described embodiments, and various modifications are possible without departing from the spirit of the present disclosure.

[0063] In the reconstructed image signal generating step, a single machine learning model may be used to generate the multiple reconstructed image signals S31-S33 instead of the multiple machine learning models ML1-ML3. In this case, in step ST32 shown in Fig. 6, the multiple primary reconstructed image signals S21-S23 and the multiple target image signals SG1-SG3 may all be used as training data for the single machine learning model. In this case, the multiple reconstructed image signals S31-S33 focused on each cross section can be generated using the single machine learning model, and the reconstructed image signal S3 can be generated efficiently.

[0064] The machine learning model ML may be a model other than a deep learning model, such as a decision tree model, a support vector machine, a gradient boosting machine, or a k-NN algorithm.

[0065] The machine learning model ML may be a deep learning model other than a convolutional neural network, such as a recurrent neural network, a long short-term memory, a gated recurrent unit, or an autoencoder.

[0066] The machine learning model ML may be trained by using 20,000 images each of the primary reconstructed image signal S2 and the target image signal S4 acquired in advance and performing 200 epochs of training. Instead of combining the echo signals corresponding to each cross section into one composite echo signal SE and outputting the combined echo signal SE to the two-dimensional image signal generator 11, the receiving circuit 9 may output the echo signals corresponding to each cross section individually to the two-dimensional image signal generator 11. The two-dimensional image signal generator 11 may combine the echo signals corresponding to each input cross section into one composite echo signal SE. [Explanation of symbols]

[0067] 8...excitation unit, 11...two-dimensional image signal generation unit, 12...primary reconstructed image signal generation unit, 13...reconstructed image signal generation unit, A...object, D1, D2, D3...cross section, Gx, Gy, Gy1, Gy2, Gy3, Gz...gradient magnetic field, ML, ML1, ML2, ML3...machine learning model, S1...two-dimensional image signal, S2, S21, S22, S23...primary reconstructed image signal, S3, S31, S32, S33...reconstructed image signal, SG, SG1, SG2, SG3...target image signal.

Claims

1. an excitation step of selectively exciting each of a plurality of cross sections of the object; a phase modulation step of providing different phase modulations to each of the cross sections; a two-dimensional image signal generating step of generating a two-dimensional image signal in which image signals captured at each of the cross sections are superimposed; a primary reproduction image signal generating step of generating a primary reproduction image signal including an image signal focused on a specific cross section among the plurality of cross sections; a reconstructed image signal generation step of generating a reconstructed image signal focused on the specific cross section from the primary reconstructed image signal using a machine learning model constructed using as training data the primary reconstructed image signal acquired in advance and a target image signal obtained in advance as an image signal focused on the specific cross section.

2. 2. The magnetic resonance image generating method according to claim 1, wherein the reconstructed image signal generating step generates a plurality of reconstructed image signals focused on each of the cross sections by using a plurality of machine learning models corresponding to each of the cross sections constituting the plurality of cross sections.

3. The magnetic resonance image generating method according to claim 1 or 2, wherein the machine learning model is a deep learning model.

4. 3. The magnetic resonance image generating method according to claim 1, wherein the excitation step applies a gradient magnetic field in a depth direction, which is a direction in which the plurality of cross sections are arranged, and applies an excitation pulse corresponding to each of the plurality of cross sections.

5. an excitation unit that selectively excites each of a plurality of cross sections of the subject; a phase modulation unit that applies different phase modulation to each of the cross sections; a two-dimensional image signal generating unit that generates a two-dimensional image signal in which image signals captured at each of the cross sections are superimposed; a primary reproduction image signal generating unit that generates a primary reproduction image signal including an image signal focused on a specific cross section among the plurality of cross sections; a reconstructed image signal generation unit that generates a reconstructed image signal focused on the specific cross section from the primary reconstructed image signal using a machine learning model constructed using as training data the primary reconstructed image signal that has been acquired in advance and a target image signal that has been obtained in advance as an image signal focused on the specific cross section.

Images