Medical data processing device

Abstract

To provide a medical data processing device allowing image quality on medical images to be improved.SOLUTION: A medical data processing device 100 is provided with a processing circuit applying a linear layer of complex number factors and non-linear activation (activation function) on medical data with complex values. A processing circuit 110 creates a leaned model by training a neural network applying the linear layer of the complex number factors and the non-linear activation (activation function) on the medical data with the complex values using a learning function 110b, and the created learned model is stored, for instance, in a memory 132 as necessary.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The embodiments disclosed in this specification and the drawings relate to a medical data processing device.

Background Art

[0002] In machine learning using neural networks, real-valued neural networks are typically used.

[0003] However, in medical data processing devices such as magnetic resonance imaging devices and ultrasonic diagnostic devices, signal processing using complex numbers is often used. Therefore, various applications are expected by using complex-valued neural networks.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] One of the problems to be solved by the embodiments disclosed in this specification and the drawings is to improve the image quality. However, the problems to be solved by the embodiments disclosed in this specification and the drawings are not limited to the above problems. The problems corresponding to the effects of each configuration shown in the embodiments described later can also be regarded as other problems.

Means for Solving the Problems

[0006] The medical data processing device according to the embodiment includes a processing unit. The processing unit applies a linear operation with complex coefficients and a non-linear activation whose gain does not change according to the complex argument to medical data having complex values.

Brief Description of the Drawings

[0007]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Embodiments for Carrying Out the Invention

[0008] (Embodiment) Hereinafter, embodiments of a medical data processing device and a data processing device will be described in detail with reference to the drawings.

[0009] The configuration of the data processing device 100 according to the embodiment will be described with reference to FIG. 1.

[0010] The data processing device 100 is a device that generates data using machine learning. As an example, the data processing device 100 is connected to various medical image diagnostic devices such as a magnetic resonance imaging device shown in FIG. 2 and an ultrasonic diagnostic device shown in FIG. 3, and processes signals received from the medical image diagnostic device, generates a trained model, executes the trained model, and the like. Note that examples of the medical image diagnostic device to which the data processing device 100 is connected are not limited to a magnetic resonance imaging device and an ultrasonic diagnostic device, and other devices such as an X-ray CT device and a PET device may also be used.

[0011] The data processing device 100 is typically a medical data processing device that processes medical data. However, the embodiment is not limited to the case where the data processing device 100 is a medical data processing device. As an example, the data processing device 100 may be a device that processes magnetic resonance data that is not medical data.

[0012] The medical image processing device 100 includes a processing circuit 110, a memory 132, an input device 134, and a display 135. The processing circuit 110 includes a training data creation function 110a, a learning function 110b, an interface function 110c, a control function 110d, an application function 110e, and an acquisition function 110f.

[0013] In the embodiment, each processing function and the learned model (e.g., neural network) performed by the training data creation function 110a, learning function 110b, interface function 110c, control function 110d, application function 110e, and acquisition function 110f are stored in the memory 132 in the form of a program executable by a computer. The processing circuit 110 is a processor that reads out and executes the program to realize the functions corresponding to the respective programs. In other words, the processing circuit 110 in the state of having read out each program has each function shown in the processing circuit 110 of FIG. 1. Further, the processing circuit 110 in the state of having read out the program corresponding to the learned model (neural network) can perform processing according to the learned model. In FIG. 1, the function of the processing circuit 110 is described as being realized by a single processing circuit. However, the processing circuit 110 may be configured by combining a plurality of independent processors, and each processor may realize the function by executing a program. In other words, each of the above-described functions may be configured as a program, and even when one processing circuit executes each program, it may be the case. Further, two or more of the functions of the processing circuit 110 may be realized by a single processing circuit. As another example, a specific function may be implemented in a dedicated independent program execution circuit.

[0014] Note that in FIG. 1, the processing circuit 110, training data creation function 110a, learning function 110b, interface function 110c, control function 110d, application function 110e, and acquisition function 110f are each an example of a processing unit, creation unit, input unit (learning unit), reception unit, control unit, application unit, and acquisition unit, respectively.

[0015] As used in the above description, the term "processor" means, for example, a CPU (Central Processing Unit), a GPU (Graphical Processing Unit), or a circuit such as an application specific integrated circuit (ASIC), a programmable logic device (e.g., a simple programmable logic device (SPLD), a complex programmable logic device (CPLD), and a field programmable gate array (FPGA)). The processor realizes its functions by reading and executing programs stored in the memory 132.

[0016] Alternatively, instead of storing the program in the memory 132, the program may be directly incorporated into the circuit of the processor. In this case, the processor realizes its functions by reading and executing the program incorporated in the circuit. Therefore, for example, instead of storing the learned model in the memory 132, a program related to the learned model may be directly incorporated into the circuit of the processor.

[0017] In addition, when the processing circuit 110 is incorporated into various medical image diagnostic devices, or when processing is performed in cooperation with various medical image diagnostic devices, it may have a function of executing related processes together.

[0018] The processing circuit 110 generates training data for learning based on the data and images acquired by the interface function 110c by the training data generation function 110a.

[0019] The processing circuit 110 performs learning using the training data generated by the training data creation function 110a by the learning function 110b, and generates a learned model.

[0020] The processing circuit 110 acquires, via the interface function 110c, data, images, etc. for image generation by the application function 110e from the memory 132.

[0021] The processing circuit 110 controls the overall processing of the data processing device 100 by the control function 110d. Specifically, the processing circuit 110 controls its own processing based on various setting requests input by the operator via the input device 134 and various control programs and various data read from the memory 132 by the control function 110d.

[0022] Also, the processing circuit 110 generates an image based on the results of the processing performed using the training data generation function 110a and the learning function 110b by the application function 110e. Further, the processing circuit 110 applies the learned model generated by the learning function 110b to the input image and generates an image based on the application result of the learned model by the application function 110e.

[0023] The memory 132 is composed of a semiconductor memory element such as a RAM (Random Access Memory), a flash memory, a hard disk, an optical disk, etc. The memory 132 is a memory that stores data such as display image data and training image data generated by the processing circuit 110.

[0024] The memory 132 stores various data such as control programs for performing image processing and display processing as needed.

[0025] The input device 134 receives various instructions and information inputs from the operator. The input device 134 is, for example, a pointing device such as a mouse or a trackball, a selection device such as a mode switching switch, or an input device such as a keyboard.

[0026] The display 135 receives an input of imaging conditions through a GUI (Graphical User Interface) under the control of the control function 110d or the like, and displays an image or the like generated by the control function 110d or the like. The display 135 is, for example, a display device such as a liquid crystal display. The display 135 is an example of a display unit. The display 135 has a mouse, a keyboard, buttons, a panel switch, a touch command screen, a foot switch, a trackball, a joystick, and the like.

[0027] FIG. 2 is an example of a magnetic resonance imaging apparatus 200 incorporating the data processing apparatus 100 according to the embodiment.

[0028] As shown in FIG. 2, the magnetic resonance imaging apparatus 200 includes a static magnetic field magnet 201, a static magnetic field power supply (not shown), a gradient magnetic field coil 203, a gradient magnetic field power supply 204, a bed 205, a bed control circuit 206, a transmission coil 207, a transmission circuit 208, a reception coil 209, a reception circuit 210, a sequence control circuit 220 (sequence control unit), and the data processing apparatus 100 described in FIG. 1. Note that the magnetic resonance imaging apparatus 200 does not include a subject P (for example, a human body). Also, the configuration shown in FIG. 2 is merely an example.

[0029] The static magnetic field magnet 201 is a magnet formed in a hollow substantially cylindrical shape, and generates a static magnetic field in the internal space. The static magnetic field magnet 201 is, for example, a superconducting magnet or the like, and is excited by receiving a current supply from a static magnetic field power supply. The static magnetic field power supply supplies a current to the static magnetic field magnet 201. As another example, the static magnetic field magnet 201 may be a permanent magnet, and in this case, the magnetic resonance imaging apparatus 200 may not include a static magnetic field power supply. Also, the static magnetic field power supply may be provided separately from the magnetic resonance imaging apparatus 100.

[0030] The gradient magnetic field coil 203 is a coil formed in a hollow substantially cylindrical shape and is disposed inside the static magnetic field magnet 201. The gradient magnetic field coil 203 is formed by combining three coils corresponding to the X, Y, and Z axes orthogonal to each other, and these three coils are individually supplied with current from the gradient magnetic field power supply 204 to generate a gradient magnetic field in which the magnetic field strength changes along the X, Y, and Z axes. The gradient magnetic fields of the X, Y, and Z axes generated by the gradient magnetic field coil 203 are, for example, the slice gradient magnetic field Gs, the phase encoding gradient magnetic field Ge, and the readout gradient magnetic field Gr. The gradient magnetic field power supply 204 supplies current to the gradient magnetic field coil 203.

[0031] The bed 205 includes a top plate 205a on which the subject P is placed, and under the control of the bed control circuit 206, the top plate 205a is inserted into the cavity (imaging opening) of the gradient magnetic field coil 203 with the subject P placed thereon. Usually, the bed 205 is installed such that the longitudinal direction is parallel to the central axis of the static magnetic field magnet 201. The bed control circuit 206 drives the bed 205 under the control of the data acquisition device 100 to move the top plate 205a in the longitudinal direction and the vertical direction.

[0032] The transmission coil 207 is disposed inside the gradient magnetic field coil 203, receives an RF pulse from the transmission circuit 208, and generates a high-frequency magnetic field. The transmission circuit 208 supplies an RF pulse corresponding to the Larmor frequency determined by the type of the target atom and the magnetic field strength to the transmission coil 207.

[0033] The reception coil 209 is disposed inside the gradient magnetic field coil 203 and receives a magnetic resonance signal (hereinafter, referred to as "MR signal" as necessary) emitted from the subject P due to the influence of the high-frequency magnetic field. When the reception coil 209 receives the magnetic resonance signal, it outputs the received magnetic resonance signal to the reception circuit 210.

[0034] Note that the above-described transmission coil 207 and reception coil 209 are merely examples. It may be configured by combining one or more of a coil having only a transmission function, a coil having only a reception function, or a coil having both transmission and reception functions.

[0035] The reception circuit 210 detects the magnetic resonance signal output from the reception coil 209 and generates magnetic resonance data based on the detected magnetic resonance signal. Specifically, the reception circuit 210 generates magnetic resonance data by digitally converting the magnetic resonance signal output from the reception coil 209. Further, the reception circuit 210 transmits the generated magnetic resonance data to the sequence control circuit 220. Note that the reception circuit 210 may be provided on the gantry device side including the static magnetic field magnet 201, the gradient magnetic field coil 203, etc.

[0036] The sequence control circuit 220 performs imaging of the subject P by driving the gradient magnetic field power supply 204, the transmission circuit 208, and the reception circuit 210 based on the sequence information. Here, the sequence information is information that defines the procedure for performing imaging. The sequence information defines the strength of the current supplied by the gradient magnetic field power supply 204 to the gradient magnetic field coil 203 and the timing of supplying the current, the strength of the RF pulse supplied by the transmission circuit 208 to the transmission coil 207 and the timing of applying the RF pulse, the timing at which the reception circuit 210 detects the magnetic resonance signal, etc. For example, the sequence control circuit 220 is an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array), or an electronic circuit such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). The sequence control circuit 220 is an example of the scan unit.

[0037] Furthermore, when the sequence control circuit 220 drives the gradient magnetic field power supply 204, the transmission circuit 208, and the reception circuit 210 to image the subject P and receives magnetic resonance data from the reception circuit 210, the sequence control circuit 220 transfers the received magnetic resonance data to the data processing device 100. In addition to the processes described in FIG. 1, the data processing device 100 performs overall control of the magnetic resonance imaging device 200.

[0038] Returning to FIG. 1, regarding the processes performed by the data processing device 100 other than the processes described in FIG. 1, the processing circuit 110 transmits sequence information to the sequence control circuit 220 through the interface function 110c and receives magnetic resonance data from the sequence control circuit 220. Further, when receiving the magnetic resonance data, the processing circuit 110 having the interface function 110c stores the received magnetic resonance data in the memory 132.

[0039] The magnetic resonance data stored in the memory 132 is arranged in the k-space by the control function 110d. As a result, the memory 132 stores the k-space data.

[0040] The memory 132 stores magnetic resonance data received by the processing circuit 110 having the interface function 110c, k-space data arranged in the k-space by the processing circuit 110 having the control function 110d, image data generated by the processing circuit 110 having the generation function (or application function 110e), and the like.

[0041] The processing circuit 110 performs overall control of the magnetic resonance imaging device 200 by the control function 110d, and controls imaging, image generation, image display, and the like. For example, the processing circuit 110 having the control function 110d receives input of imaging conditions (imaging parameters, etc.) on the GUI and generates sequence information according to the received imaging conditions. Further, the processing circuit 200 having the control function 110d transmits the generated sequence information to the sequence control circuit 220.

[0042] The processing circuit 110 reads k-space data from the memory 132 by a generation function (or application function 110e) not shown in FIG. 1, and generates a magnetic resonance image by performing a reconstruction process such as a Fourier transform on the read k-space data.

[0043] FIG. 3 is a configuration example of an ultrasonic diagnostic apparatus 300 incorporating the data processing apparatus 100 according to the embodiment. The ultrasonic diagnostic apparatus according to the embodiment includes an ultrasonic probe 305 and an ultrasonic diagnostic apparatus main body 300. The ultrasonic diagnostic apparatus main body 300 includes a transmission circuit 309, a reception circuit 311, and the above-described data processing apparatus 100.

[0044] The ultrasonic probe 305 has a plurality of piezoelectric vibrators, and these plurality of piezoelectric vibrators generate ultrasonic waves based on a drive signal supplied from a transmission circuit 309 included in the ultrasonic diagnostic apparatus main body 300 described later. Further, the plurality of piezoelectric vibrators included in the ultrasonic probe 305 receive a reflected wave from the subject P and convert it into an electrical signal (reflected wave signal). The ultrasonic probe 305 also includes a matching layer provided on the piezoelectric vibrator and a backing material or the like that prevents the propagation of ultrasonic waves backward from the piezoelectric vibrator. Note that the ultrasonic probe 305 is detachably connected to the ultrasonic diagnostic apparatus main body 300. The ultrasonic probe 305 is an example of a scan unit.

[0045] When ultrasonic waves are transmitted from the ultrasonic probe 305 to the subject P, the transmitted ultrasonic waves are successively reflected at discontinuous surfaces of acoustic impedance in the body tissues of the subject P, received by the plurality of piezoelectric vibrators included in the ultrasonic probe 305 as reflected waves, and converted into reflected wave signals. The amplitude of the reflected wave signal depends on the difference in acoustic impedance at the discontinuous surface where the ultrasonic wave is reflected. When the transmitted ultrasonic pulse is reflected at the surface of a moving blood flow or a heart wall or the like, the reflected wave signal undergoes a frequency shift depending on the velocity component of the moving object with respect to the ultrasonic wave transmission direction due to the Doppler effect.

[0046] The ultrasonic diagnostic apparatus main body 300 is an apparatus that generates ultrasonic image data based on the reflected wave signals received from the ultrasonic probe 305. The ultrasonic diagnostic apparatus main body 300 is an apparatus capable of generating two-dimensional ultrasonic image data based on two-dimensional reflected wave signals and generating three-dimensional ultrasonic image data based on three-dimensional reflected wave signals. However, the embodiment is applicable even when the ultrasonic diagnostic apparatus 10 is an apparatus dedicated to two-dimensional data.

[0047] As illustrated in FIG. 3, the ultrasonic diagnostic apparatus 10 includes a transmission circuit 309, a reception circuit 311, and a medical image processing apparatus 100.

[0048] The transmission circuit 309 and the reception circuit 311 control the ultrasonic transmission and reception performed by the ultrasonic probe 305 based on an instruction from a data processing apparatus 110 having a control function. The transmission circuit 309 includes a pulse generator, a transmission delay unit, a pulsar, etc., and supplies a drive signal to the ultrasonic probe 305. The pulse generator repeatedly generates rate pulses for forming transmitted ultrasonic waves at a predetermined pulse repetition frequency (PRF: Pulse Repetition Frequency). Further, the transmission delay unit applies a delay time for each piezoelectric vibrator necessary for focusing the ultrasonic waves generated from the ultrasonic probe 305 into a beam shape and determining the transmission directivity to each rate pulse generated by the pulse generator. Also, the pulsar applies a drive signal (drive pulse) to the ultrasonic probe 305 at a timing based on the rate pulse.

[0049] That is, the transmission delay unit arbitrarily adjusts the transmission direction of the ultrasonic waves transmitted from the piezoelectric vibrator surface by changing the delay time given to each rate pulse. Also, the transmission delay unit controls the position of the focusing point (transmission focus) in the depth direction of the ultrasonic transmission by changing the delay time given to each rate pulse.

[0050] In addition, the receiving circuit 311 includes an amplifier circuit, an A / D (Analog / Digital) converter, a reception delay circuit, an adder, a quadrature demodulation circuit, etc., and performs various processes on the reflected wave signal received from the ultrasonic probe 305 to generate a received signal (reflected wave data). The amplifier circuit amplifies the reflected wave signal for each channel and performs gain correction processing. The A / D converter performs A / D conversion on the gain-corrected reflected wave signal. The reception delay circuit gives the reception delay time necessary to determine the reception directivity to the digital data. The adder performs an addition process on the reflected wave signals given the reception delay time by the reception delay circuit. By the addition process of the adder, the reflection component from the direction corresponding to the reception directivity of the reflected wave signal is emphasized. Then, the quadrature demodulation circuit converts the output signal of the adder into an in-phase signal (I signal, I: In-phase) and a quadrature signal (Q signal, Q: Quadrature-phase) in the baseband band. Then, the quadrature demodulation circuit transmits the I signal and the Q signal (hereinafter referred to as IQ signals) as received signals (reflected wave data) to the processing circuit 110. Note that the quadrature demodulation circuit may convert the output signal of the adder into an RF (Radio Frequency) signal and then transmit it to the processing circuit 110. The IQ signals and the RF signal are received signals having phase information.

[0051] When scanning a two-dimensional region in the subject P, the transmission circuit 309 causes the ultrasonic probe 305 to transmit an ultrasonic beam for scanning the two-dimensional region. Then, the receiving circuit 311 generates a two-dimensional received signal from the two-dimensional reflected wave signal received from the ultrasonic probe 305. Also, when scanning a three-dimensional region in the subject P, the transmission circuit 309 causes the ultrasonic probe 305 to transmit an ultrasonic beam for scanning the three-dimensional region. Then, the receiving circuit 311 generates a three-dimensional received signal from the three-dimensional reflected wave signal received from the ultrasonic probe 305. The receiving circuit 311 generates a received signal based on the reflected wave signal and transmits the generated received signal to the processing circuit 110.

[0052] The transmission circuit 309 causes the ultrasonic probe 305 to transmit an ultrasonic beam from a predetermined transmission position (transmission scanning line). The reception circuit 311 receives, from the ultrasonic probe 305, a signal due to the reflected wave of the ultrasonic beam transmitted by the transmission circuit 309 at a predetermined reception position (reception scanning line). When parallel simultaneous reception is not performed, the transmission scanning line and the reception scanning line are the same scanning line. On the other hand, when parallel simultaneous reception is performed, when the transmission circuit 309 causes the ultrasonic probe 305 to transmit one ultrasonic beam on one transmission scanning line, the reception circuit 311 simultaneously receives, as a plurality of reception beams, signals due to the reflected waves derived from the ultrasonic beam transmitted by the transmission circuit 309 to the ultrasonic probe 1, through the ultrasonic probe 305 at a plurality of predetermined reception positions (reception scanning lines).

[0053] The data processing device 100 is connected to the transmission circuit 309 and the reception circuit 311, and in addition to the functions already shown in FIG. 1, performs processing of the signal received from the reception circuit 311, control of the transmission circuit 309, generation of a learned model, execution of the learned model, and various image processes. The processing circuit 110 includes a B-mode processing function, a Doppler processing function, a generation function, etc., in addition to the functions already shown in FIG. 1. Hereinafter, returning to FIG. 1, a configuration that the data processing device 100 incorporated in the ultrasonic diagnostic device 10 may have in addition to the configuration already shown in FIG. 1 will be described.

[0054] Each processing function and learned model performed by the B-mode processing function, the Doppler processing function, and the generation function are stored in the memory 132 in the form of a program executable by a computer. The processing circuit 110 is a processor that reads out the program from the memory 132 and executes it to realize the functions corresponding to the respective programs. In other words, the processing circuit 110 in the state of having read out each program has these respective functions.

[0055] The B-mode processing function and the Doppler processing function are examples of a B-mode processing unit and a Doppler processing unit.

[0056] The processing circuit 110 performs various signal processes on the reception signal received from the reception circuit 311.

[0057] The processing circuit 110 receives data from the receiving circuit 311 by the B-mode processing function, performs logarithmic amplification processing, envelope detection processing, logarithmic compression processing, etc., and generates data (B-mode data) in which the signal strength is expressed by the brightness of luminance (Brightness).

[0058] Also, the processing circuit 110 frequency-analyzes velocity information from the received signal (reflected wave data) received from the receiving circuit 311 by the Doppler processing function, and generates data (Doppler data) in which moving body information such as velocity, variance, and power due to the Doppler effect is extracted for multiple points.

[0059] Note that in the B-mode processing function and the Doppler processing function, processing can be performed on both two-dimensional reflected wave data and three-dimensional reflected wave data.

[0060] Also, the processing circuit 110 controls the entire processing of the ultrasonic diagnostic apparatus by the control function 110d. Specifically, the processing circuit 110 controls the processing of the transmission circuit 309, the receiving circuit 311, and the processing circuit 110 based on various setting requests input from the operator via the input device 134 and various control programs and various data read from the memory 132 by the control function 110d. Also, the processing circuit 110 controls the display of the ultrasonic image data for display stored in the memory 132 on the display 135 by the control function 110d.

[0061] Also, the processing circuit 110 generates ultrasonic image data from the data generated by the B-mode processing function and the Doppler processing function by a generation function (not shown). The processing circuit 110 generates two-dimensional B-mode image data in which the intensity of the reflected wave is represented by luminance from the two-dimensional B-mode data generated by the B-mode processing function by the generation function. Also, the processing circuit 110 generates two-dimensional Doppler image data representing moving body information from the two-dimensional Doppler data generated by the Doppler processing function 110b by the generation function. The two-dimensional Doppler image data is velocity image data, variance image data, power image data, or image data combining these.

[0062] Also, the processing circuit 110, by its generation function, converts (scan-converts) the scan line signal sequence of the ultrasonic scan into a scan line signal sequence in a video format typified by a television or the like, and generates ultrasonic image data for display. Further, by its generation function, the processing circuit 110 performs, as various image processes other than scan conversion, for example, an image process (smoothing process) of regenerating an average value image of luminance using a plurality of image frames after scan conversion, an image process (edge enhancement process) using a differential filter within the image, and the like. Also, the processing circuit 110 performs various rendering processes on the volume data in order to generate two-dimensional image data for displaying the volume data on the display 135 by its generation function.

[0063] The memory 132 can also store the data generated in the B-mode processing function and the Doppler processing function. The B-mode data and Doppler data stored in the memory 132 can be called by the operator, for example, after diagnosis, and become ultrasonic image data for display via the processing circuit 110. Also, the memory 132 can store the reception signal (reflected wave data) output by the reception circuit 311.

[0064] In addition, the memory 132 stores, as necessary, a control program for performing ultrasonic transmission / reception, image processing, and display processing, diagnostic information (for example, patient ID, doctor's findings, etc.), diagnostic protocols, and various data such as various body marks.

[0065] Subsequently, with reference to FIGS. 4 to 6, the configuration of the neural network according to the embodiment will be described.

[0066] FIG. 4 shows an example of the interconnection between layers in a neural network 7 used for machine learning by a processing circuit 110 having a learning function 110b. The neural network 7 is composed of an input layer 1, an output layer 2, and intermediate layers 3, 4, 5, etc. between the input layer 1 and the output layer 2. Each intermediate layer is composed of a layer related to the respective input (hereinafter referred to as the input layer in each layer), a linear layer, and a layer related to processing using an activation function (hereinafter referred to as the activation layer). For example, the intermediate layer 3 is composed of an input layer 3a, a linear layer 3b, and an activation layer 3c, the intermediate layer 4 is composed of an input layer 4a, a linear layer 4b, and an activation layer 4c, and the intermediate layer 5 is composed of an input layer 5a, a linear layer 5b, and an activation layer 5c. Also, each layer is composed of a plurality of nodes (neurons).

[0067] Here, the data processing apparatus 100 according to the embodiment applies a linear layer with complex coefficients and a non-linear activation (activation function) to medical data having complex values. That is, the processing circuit 110 generates a learned model by training a neural network 7 that applies a linear layer with complex coefficients and a non-linear activation (activation function) to medical data having complex values by means of the learning function 110b. The processing circuit 100 stores the generated learned model in, for example, the memory 132 as necessary.

[0068] Note that the data input to the input layer 1 is typically a medical image or medical image data acquired from a medical imaging device. When the medical imaging device is a magnetic resonance imaging device 200, the data input to the input layer 1 is, for example, a magnetic resonance image. Also, when the medical imaging device is, for example, an ultrasonic diagnostic device 300, the data input to the input layer 1 is, for example, an ultrasonic image.

[0069] Also, the input data input to the input layer 1 may be a medical image, or may be various image data, projection data, intermediate data, or raw data at the stage before the medical image is generated. For example, when the medical image diagnostic apparatus is a PET apparatus, the input data input to the input layer 10 may be a PET image, or may be various data before the reconstruction of the PET image, such as time-series data of coincidence coefficient information.

[0070] Also, the data output from the output layer 2 is, for example, a medical image or medical image data, and similar to the data input to the input layer 1, may be various projection data, intermediate data, or raw data at the stage before the medical image is generated. When the purpose of the neural network 7 is noise processing, the data output from the output layer 2 is, for example, an image with noise removed and high image quality compared to the input image.

[0071] In addition, when the neural network 7 according to the embodiment is, for example, a convolutional neural network (CNN: Convolutional Neural Network), the data input to the input layer 1 is data represented by a two-dimensional array with a size of 32×32 or the like, for example, and the data output from the output layer 2 is data represented by a two-dimensional array with a size of 32×32 or the like, for example. The size of the data input to the input layer 1 and the size of the data output from the output layer 2 may be the same or different. Similarly, the number of nodes in the intermediate layer may be the same as or different from the number of nodes in the layers before and after it.

[0072] Next, the generation of the learned model according to the embodiment, that is, the learning step, will be described. The processing circuit 110 generates a learned model by performing, for example, machine learning on the neural network 7 by the learning function 110b. Here, performing machine learning means, for example, determining the weightings in the neural network 7 composed of the input layer 1, the intermediate layers 3, 4, 5, and the output layer 2. Specifically, it means determining a set of coefficients characterizing the connection between the input layer 1 and the intermediate layer 3, a set of coefficients characterizing the connection between the intermediate layer 3 and the intermediate layer 4, …, and a set of coefficients characterizing the connection between the intermediate layer 5 and the output layer 2. The processing circuit 150 determines these sets of coefficients by the learning function 110b, for example, by the backpropagation method.

[0073] The processing circuit 110 performs machine learning based on the training data, which is teacher data composed of the data input to the input layer 1 and the data output to the output layer 12 by the learning function 110b, determines the weightings between the layers, and generates a learned model with the weightings determined.

[0074] Note that in deep learning, autoencoding (autoencoder) can be used. In this case, the data required for machine learning does not have to be data with a teacher.

[0075] Next, the process when applying the learned model according to the embodiment will be described. First, the processing circuit 110 inputs an input medical image, which is a clinical image for example, into the learned model by the application function 110e. For example, the processing circuit 110 inputs an input medical image, which is a clinical image, into the input layer 1 of the neural network 7, which is a learned model, by the application function 110e. Subsequently, the processing circuit 110 acquires, as an output medical image, the data output from the output layer 2 of the neural network 7, which is a learned model, by the application function 110e. The output medical image is a medical image on which predetermined processing such as noise removal has been performed. In this way, the processing circuit 150 generates an output medical image on which predetermined processing such as noise removal has been performed by the application function 150e. If necessary, the processing circuit 150 may display the obtained output medical image on the display 135 by the control function 110d.

[0076] Returning to the description of the activation function and the activation layer, the activation function in the neural network 7 will be described with reference to FIG. 5. In FIG. 5, nodes 10a, 10b, 10c, and 10d are shown by cutting out a part of the nodes of the input layer in a certain layer. On the other hand, node 11 is one of the nodes of the linear layer, and node 12 is one of the nodes of the activation layer, which is a layer related to the process (activation) using the activation function. In the embodiment, a complex neural network is used, but here the case of a real neural network will be described first.

[0077] Here, considering the case where the output values of nodes 10a, 10b, 10c, and 10d are real numbers x1, x 2、 x 3、 x4 respectively, the output result to node 11 in the linear layer is Σ i=1 m (ω i x i +b). Here, ω iLet \(w_{i}\) be the weight coefficient between the \(i\)-th input layer and node 11, \(m\) be the number of nodes connected to node 11, and \(b\) be a predetermined constant. Subsequently, assuming the output result output to node 12, which is the activation layer, is \(y\), \(y\) is expressed as in the following formula (1) using the activation function \(f\).

[0078] [Number]

[0079] Here, the activation function \(f\) is usually a non-linear function. For example, the sigmoid function, tanh function, ReLu (Rectified Linear Unit), etc. are selected as the activation function \(f\).

[0080] Figure 6 shows the processing using such an activation function. In Figure 6, the intermediate layer 5 is the \(n\)-th layer of the neural network 7 and consists of an input layer 5a, a linear layer 5b, and an activation layer 5c. The input layer 6a is the \((n + 1)\)-th layer of the neural network. Also, the input layer 5a has nodes 20a, 20b, 20c, 20d, etc., the linear layer 5b has nodes 21a, 21b, 21c, 21d, etc., and the linear layer 5c has nodes 22a, 22b, 22c, etc. Further, Figure 6 is a real-number neural network in which each node has a real value, and the input result \(x\) to the input layer 5a n,i and the output result \(x\) of the input layer 6a n+1、i are real values.

[0081] Here, by performing predetermined weighted addition for each node in the input layer 5a, the output result to the linear layer 5b is calculated. For example, the output result to the \(j\)-th node 21b in the linear layer 5b is \(\sum\) i=1 m \(\omega\) i,j \(x\) n,i + \(b\) n,j and is given by. Here, \(\omega\) i,j is the weight coefficient between the \(i\)-th input layer and the \(j\)-th linear layer, and \(b\) n,j is a predetermined constant known as the bias term. Note that \(\omega\) i,j is the first-order coefficient in the linear operation in the linear layer, and \(b\)n,j can be referred to as the coefficient of the 0th order in the linear operation in the linear layer. Subsequently, by applying the activation function f to each node of the linear layer 5b, the output result to the activation layer 5c is calculated. For example, the output to the j-th node 22b in the activation layer 5c is the activation function f n,j using n,j (Σ i=1 m ω i,j x n,i +b n,j ) is given. Note that the processing in the linear layer is not limited to the above processing, and various known methods may be used. For example, a convolutional layer may be used as the linear layer.

[0082] Subsequently, based on the values output by the nodes of the activation layer 5c, the values of each node of the input layer 6a of the n-th layer are determined. As an example, the values of each node of the activation layer 5c are input directly to each node of the input layer 6a. Also, as another example, a further non-linear function may be applied to the activation layer 5c to determine each node of the input layer 6a.

[0083] Subsequently, a complex neural network will be described. FIG. 7 shows the configuration when the neural network 7 used for machine learning by the processing circuit 110 having the learning function 110b is a complex neural network. The complex neural network shown in FIG. 7 has the same configuration as the real neural network shown in FIG. 6, but the value z n,j of each node is a complex number. Also, the activation function f n,j is also defined in the complex number domain and becomes a function that takes complex number values. If the input value input to the j-th node of the j-th input layer 6a of the n + 1-th input layer is z n+1、j , for example, the following equation (2) holds. Thus, for example, by making each node of the neural network 7 a complex number node, a neural network 7 having complex number values can be generated.

[0084]

Equation

[0085] Next, the background related to the embodiment will be described.

[0086] In machine learning using neural networks, real-valued neural networks are typically used. However, in medical data processing devices such as magnetic resonance imaging (MRI) devices and ultrasonic diagnostic devices, signal processing using complex numbers is often employed. Therefore, various applications are expected by using complex-valued neural networks.

[0087] Here, as a method for handling complex numbers in a neural network, for example, a complex number can be divided into a real part and an imaginary part, and each can be considered as a node of a standard real-valued neural network, thereby providing a method for handling complex numbers in a neural network. As an example, as an activation function, a method of applying ReLU to each of the real part and the imaginary part of a complex number can be considered for handling complex numbers in a neural network.

[0088] Also, as another example, a complex number can be expressed using an absolute value (or a signed absolute value) and a phase, and each can be considered as a node of a standard real-valued neural network, thereby providing a method for handling complex numbers in a neural network.

[0089] Here, in medical images such as magnetic resonance images and ultrasonic images, for example, the phase information of the image, such as the phase gradient, is often important, while there are relatively few cases where the absolute value of the phase has a dominant meaning. For example, in a magnetic resonance imaging device, a slight difference in the central frequency appears as a phase modulation of the entire image, but in many cases, the importance of the absolute values of those phases themselves is relatively low. Therefore, in the application of complex-valued neural networks to medical images, such as noise removal and region extraction, it is desirable to configure a neural network that does not ignore the phase information of the input image and, on the other hand, does not significantly fluctuate the output result with respect to the phase modulation of the entire image.

[0090] As an example, when considering denoising processing using Complex-convolution and ReLU for the real and imaginary parts respectively, if the distribution of the teacher image in the complex neural network is biased towards the real part, for example, coefficients that emphasize the real part of the input image and abstract the imaginary part can be learned.

[0091] When an image biased towards the imaginary part is applied as the input image to the thus generated trained model, it is expected that the model will not exhibit the expected performance.

[0092] Such a situation is shown, for example, in FIG. 8. FIG. 8 is a graph showing the denoising performance of a trained model when generating a trained model for performing denoising processing using a complex neural network 7 with 6 layers of Complex-convolution and CReLU stacked. FIG. 8 shows that when generating the trained model, 38 teacher images were prepared for training, and for a certain test image applied to the generated trained model while modulating only the phase, the mean square error (MSE) of the output image after applying the trained model is plotted as a function of the phase. As can be seen from FIG. 8, the quality of the output image after applying the trained model is not constant with respect to the phase of the input image.

[0093] As one method for solving the problem that the image quality of the output image after applying the trained model is not constant with respect to the phase of the input image, a method of achieving statistical stabilization by randomly applying phase modulation to the teacher image (phase argumentation) can be considered. However, phase argumentation may not have good learning efficiency in some cases.

[0094] The medical image data device according to the embodiment is based on such a background, and constructs a neural network by combining an activation function (activation) that is angle-independent on the complex plane and a linear layer with complex numbers as coefficients.

[0095] In other words, the data processing apparatus 100 according to the embodiment has a processing circuit 110 that applies a linear layer with complex coefficients and a non-linear activation that does not depend on the complex argument (i.e., the gain does not change according to the complex argument) to medical data having complex values.

[0096] Such a situation is shown in FIG. 9. FIG. 9 shows a cut-out display of the configuration of one intermediate layer 5 shown in FIG. 4 in the neural network 7 in which the data processing apparatus 100 according to the embodiment performs training. As already described in FIG. 4, the intermediate layer 5 includes an input layer 5a, a linear layer 5b, and an activation layer 5c (activation function). In the data processing apparatus 100 according to the embodiment, the activation layer 5c is an angle-independent activation function 5c1 that is a non-linear activation that does not depend on the complex argument.

[0097] That is, the processing circuit 110 generates a learned model by training a neural network 7 that applies a linear layer 5b or the like with complex coefficients and an angle-independent activation function 5c1 that is a non-linear activation that does not depend on the complex argument to medical data having complex values by the learning function 110b.

[0098] Here, examples of the angle-independent activation function 5c1 include function systems such as equations (3) to (8). In equations (3) to (8), z represents a complex number.

[0099]

Equation

[0100]

Equation

[0101]

Equation

[0102]

Number

[0103] In Equation (3), λ is a real number. Taking x as a real number, ReLU(x) (Rectified Linear Unit) is max(0, x). The right side of Equation (3) becomes 0 when the absolute value of the complex number z is less than λ, and when the absolute value of the complex number z is greater than λ, it becomes a complex number whose absolute value is equal to |z| - λ and whose argument is equal to z. Therefore, Equation (3) becomes a non-linear activation that does not depend on the complex argument. Equation (3) can also be considered as an extension of the soft-shrink function defined for real numbers to complex numbers.

[0104] Also, in Equation (4), z * represents the complex conjugate of z. Equation (4) is a function similar to Equation (3), but while Equation (3) is a function constructed based on the first power of the absolute value of the complex number z, Equation (4) is a function constructed based on the square of the absolute value of the complex number z. Equation (4) has the advantage that the value on the right side can be evaluated quickly when performing numerical calculations.

[0105] Also, in Equation (5), β is a real number. The right side of Equation (5) also has its absolute value represented by a non-linear function of the input signal, and its argument is equal to the argument of the input signal. Therefore, Equation (5) becomes a non-linear activation that does not depend on the complex argument. Equation (5) can also be considered as an extension of the tanh-shrink function defined for real numbers to complex numbers.

[0106] Also, in Equation (6), p is a real number. Since the contribution in a specific complex argument direction does not selectively increase on the right side of Equation (6), Equation (6) becomes a non-linear activation that does not depend on the complex argument.

[0107] Also, as variations of Equation (3), for example, function systems such as the following Equation (7) or Equation (8) can also be considered.

[0108]

Mathematics

[0109]

Mathematics

[0110] The non-linear activations independent of the complex argument in the above formulas (3) to (8) are only examples, and various other function systems can be considered. As an example, a function that can be expressed as f(z) = g(|z|)z using a real function g(x), that is, a function including a real function related to the absolute value of the input complex number, for example, a function obtained by multiplying the input complex number by a real function related to the absolute value of the input complex number, is an example of a non-linear activation independent of the complex argument. Examples of the real function constituting the non-linear activation include, as described above, a soft-shrink function, a tanh-shrink function, a power function, or a function including a ReLU function, etc.

[0111] Figure 10 shows the situation when the activation according to the embodiment is applied. Figure 10 is a graph showing the denoising performance of a learned model generated by a complex neural network 7 with six layers of Complex-convolution with the bias term fixed to 0 and the complex tanh-shrink function shown in Equation (5) performing denoising processing. That is, the processing circuit 110 further applied a complex convolution process with the bias term fixed to 0. In other words, the linear operation of the complex number coefficients applied by the processing circuit 110 to the medical data having complex number values includes a process with the bias term fixed to 0 and also includes complex convolution. Note that the complex convolution process is not limited to the case where the bias term is fixed to 0. Figure 10 plots the mean square error (MSE) of the output image after applying a certain test image to the generated learned model while modulating only the phase as a function of the phase. As can be seen from Figure 10, even when the phase of the input image changes, an output image with stable image quality is obtained.

[0112] The embodiment is not limited to the above example, and the processing circuit 110 may generate a learned model using two or more non-linear activations that do not depend on the complex argument by the learning function 110b. Such a situation is shown in Figure 11. In Figure 11, the intermediate layer 5 has an input layer 5a, a linear layer 5b, a first angle-independent activation function 5c2, and a second angle-independent activation function 5c3. That is, the processing circuit 110 applies the first angle-independent activation function 5c2 to some of the nodes in the linear layer 5b to generate an output result for the next layer, and applies the second angle-independent activation function 5c3 to some other nodes in the linear layer 5b different from the some nodes to generate an output result for the next layer.

[0113] Here, as a first example of using non-linear activation that does not depend on two or more complex arguments, there is a case of using non-linear activation that does not depend on complex arguments in different function systems. For example, the processing circuit 110 performs learning by applying non-linear activation in a plurality of different function systems, such as a complex soft-shrink function represented by, for example, Equation (3) and a complex tanh-shrink function represented by Equation (5), to medical data by the learning function 110b. In the example of FIG. 10, the processing circuit 110 performs learning by applying a complex soft-shrink function as a first angle-independent activation function 5c2 to some nodes of the linear layer 5b and applying a complex tanh-shrink function as a second angle-dependent activation function 5c3 to some nodes of the linear layer 5b by the learning function 110b.

[0114] Further, as a second example of using non-linear activation that does not depend on two or more complex arguments, there is a case of using non-linear activation using a plurality of functions that are the same function system but have different function parameter values, such as λ in Equations (3) and (4), β in Equation (5), p in Equation (6), etc. In other words, the processing circuit 110 performs learning by applying non-linear activation to medical data using, for example, a plurality of functions that are the same function system but have different function parameters by the learning function 110b. As an example, the processing circuit 110 applies, by the learning function 110b, a complex soft-shrink function given by Equation (3) with λ = λ1 as a first angle-independent activation function 5c2 to some nodes of the linear layer 5b, and applies a complex soft-shrink function given by Equation (3) with λ = λ2 as a second angle-dependent activation function 5c3 to some nodes of the linear layer 5b to perform learning.

[0115] Also, as another example, an identity function may be applied as an activation function to some nodes.

[0116] Note that such function parameters may be fixed, or may vary during the process of machine learning and be regarded as parameters that vary. Further, in one neural network 7, there may be a mixture of fixed function parameters and function parameters that are parameterized to vary and vary.

[0117] Also, such function parameters may be trainable by machine learning. An example of such a configuration is shown in FIG. 12.

[0118] As shown in FIG. 12, the processing circuit 110 includes a first neural network 7 that is a neural network that outputs an output image / output data with respect to an input image / input data, and a second neural network 8 for adjusting an activation function in the first neural network 7. The second neural network 8 is connected to the activation layers 3c, 4c, 5c of the first neural network 7 and controls the parameters of the activation function in the activation layer. That is, the processing circuit 110 includes a neural network 7 that applies a non-linear activation independent of the complex argument to medical data, and a calculation unit (not shown in FIG. 1) that optimizes the function parameters related to the non-linear activation independent of the complex argument. The second neural network 8 is an example of the above-described calculation unit.

[0119] As an example, when the complex soft-shrink function given by Equation (3) is used as the activation function f, the value of the parameter λ of the complex soft-shirnk function, which is the activation function in the activation layers 3c, 4c, 5c of the first neural network 7, is defined as λ = λ i with the i-th layer being i. λ i has a constant value for each layer. The value of λ i is optimized by learning by the calculation unit.

[0120] As an example of such a parameter optimization method, the processing circuit 110 may alternately repeat the first learning, which is the learning of the weight coefficients in the first neural network 7 executed by the learning function 110b, and the second learning, which is the learning of the values of the parameters of the activation function of the first neural network 7 executed by the calculation unit.

[0121] Also, as another example, after the processing circuit 110 executes the second learning, which is the learning of the values of the parameters of the activation function of the first neural network 7 executed by the calculation unit, the processing circuit 110 may use the values of the parameters to perform the first learning, which is the learning of the weight coefficients in the first neural network 7 executed by the learning function 110b.

[0122] Also, the processing circuit 110 may execute the first learning and the second learning simultaneously.

[0123] Note that the configuration of the calculation unit is not limited to a neural network. For example, optimization of the values of the parameters of the activation function of the first neural network 7 may be performed using linear regression or the like.

[0124] Also, as the above-described embodiment, the case where different parameters are used for each layer and common parameters are used for each node has been described. However, the embodiment is not limited to this, and common parameters may be used for each layer, or different parameters may be used for each layer and each node.

[0125] In the embodiments described so far, the case where only an angle-independent activation function is used as the activation function of the neural network 7 has been described. However, the embodiment is not limited to this, and a complex phase angle-sensitive activation function (CPSAF) may be used in combination as the activation function of the neural network 7. That is, the processing circuit 110 may further apply, in the neural network 7, an activation whose gain changes according to the complex phase angle by the learning function 110b.

[0126] For example, when it is desired to remove specific complex argument components such as phase noise processing, by using an activation function whose gain changes according to the complex argument in combination with a non-linear activation that does not depend on the complex argument, that is, an activation function whose gain does not change according to the complex argument, for example, noise processing can be performed efficiently.

[0127] Specifically, for example, as shown in FIGS. 13 and 14, the processing circuit 110 sequentially applies a non-linear activation that does not depend on the complex argument and an activation whose gain changes according to the complex argument by the learning function 110b.

[0128] For example, as shown in FIG. 13, the intermediate layer 5 includes an input layer 5a, a linear layer 5b, an activation layer 5c using an angle-independent activation function, and an activation layer 5d using an activation function sensitive to the complex argument. The processing circuit 110 outputs the output result of the input layer 5a to the linear layer 5b by the learning function 110b, outputs the output result of the linear layer 5b to the activation layer 5c using the angle-independent activation function, outputs the output result of the activation layer 5c using the angle-independent activation function to the activation layer 5d using the activation function sensitive to the complex argument, and outputs the output result of the activation layer 5d using the activation function sensitive to the complex argument to the input layer of the next layer.

[0129] As another example, for example, as shown in FIG. 14, the intermediate layer 5 includes an input layer 5a, a linear layer 5b, an activation layer 5c using an angle-independent activation function, a linear layer 5e, and an activation layer 5d using an activation function sensitive to the complex argument. The processing circuit 110 outputs the output result of the input layer 5a to the linear layer 5b by the learning function 110b, outputs the output result of the linear layer 5b to the activation layer 5c using the angle-independent activation function, outputs the output result of the activation layer 5c using the angle-independent activation function to the linear layer 5e, outputs the output result of the linear layer 5e to the activation layer 5d using the activation function sensitive to the complex argument, and outputs the output result of the activation layer 5d using the activation function sensitive to the complex argument to the input layer of the next layer.

[0130] Note that in FIGS. 13 and 14, the order in which the activation layer 5c using the angle-independent activation function and the activation layer 5d using the activation function sensitive to the complex declination angle are executed is not limited to the order shown in these figures. For example, after the activation layer 5d using the activation function sensitive to the complex declination angle is executed, the activation layer 5c using the angle-independent activation function may be executed.

[0131] Also, as another example, as shown in FIG. 15, the processing circuit 110 may selectively apply non-linear activation independent of the complex declination angle and activation with a gain that changes according to the complex declination angle. In the example of FIG. 15, the intermediate layer 5 includes an input layer 5a, a linear layer 5b, an activation layer 5c1 using an angle-independent activation function, an activation layer 5c2 using an activation function sensitive to the complex declination angle, and an activation layer 5c3 using another activation function. Here, the processing circuit 110 outputs the output result of the input layer 5a to the linear layer 5b by the learning function 110b. Among the output results of the linear layer 5b, the output results of some nodes are output to the activation layer 5c1 using the angle-independent activation function, and the output result of the activation layer 5c1 using the angle-independent activation function is output to the corresponding node 6a1 in the input layer of the next layer. On the other hand, the processing circuit 110 outputs the output results of some other nodes among the output results of the linear layer 5b to the activation layer 5c2 using the activation function sensitive to the complex declination angle, and the output result of the activation layer 5c2 using the activation function sensitive to the complex declination angle is output to the corresponding node 6a2 in the input layer of the next layer. Also, the processing circuit 110 outputs the output results of some yet other nodes among the output results of the linear layer 5b to the activation layer 5c3 using another activation function, and the output result of the activation layer 5c3 using another activation function is output to the corresponding node 6a3 in the input layer of the next layer.

[0132] Thus, in the example of FIG. 15, the processing circuit 110 varies the non-selective activation applied for each node of the linear layer 5b. Thereby, machine learning reflecting the properties of the nodes of the linear layer 5b can be efficiently performed, and the image quality is improved.

[0133] Returning to the description of the activation function sensitive to the complex argument, as a specific example of the above-described activation function sensitive to the complex argument (CPSFA), for example, the function f1 given by the following equation (9) can be cited.

[0134] [Number]

[0135] Here, z represents a complex number, phase(z) represents the complex argument of the complex number z, and α and β represent real parameters. The gain control function W β (x) is a function defined on the real number x. For example, it is a function that extracts the angle near x = 0 in a way characterized by the parameter β. Hereinafter, for example, the gain control function W β (x) will be described by taking an example of a function that has a maximum value at x = 0 and whose value decreases as it moves away from x = 0. Note that since angles that differ by a constant multiple of 2π can be regarded as the same, for example, as the gain control function W β , a periodic function with a period of 2π can be selected, and W β (x + 2nπ) = W β (x) holds.

[0136] The activation function f1(z) is obtained by multiplying the complex number z by the gain control function W β (phase(z) - α). Therefore, the activation function f1(z) can obtain a large gain (signal value) when the complex argument of z is close to α to a certain extent, and the magnitude of the gain is controlled by the parameter β. Therefore, the activation function f1 represented by equation (9) αβ can be considered as a function represented by the product of a gain control function that extracts a signal component in a predetermined angular direction and the input complex number, and is an example of an activation function sensitive to the complex argument.

[0137] As another example of the activation function sensitive to the complex argument, the activation function f2(z) given by the following equation (10) can be cited.

[0138] [Number]

[0139] Here, the activation function f2(z) is a special case in Equation (9) when the gain control function W β (x) is given by the following Equation (11).

[0140]

Number

[0141] Here, the wrap function on the right side of Equation (11) is given by the following Equation (12) where n is a natural number.

[0142]

Number

[0143] That is, the gain control function W β (x) returns 1 if the angle x is within the range of β with respect to 0, and 0 otherwise. That is, the activation function f2 αβ (z) is a function that extracts the complex number region within the range of angle β from the angle α direction. In other words, the activation function f2 αβ (z) represented by Equation (10) can be considered as a function that extracts the signal components within the range from a predetermined angle α to a predetermined angle β, and is an example of an activation function sensitive to the complex argument.

[0144] Also, as another example of an activation function sensitive to the complex argument, activation functions f3(z) to f5(z) given by the following Equations (13) to (15) can be cited.

[0145]

Number

[0146]

Number

[0147] [Number]

[0148] Here, the activation function f3 given by Equation (13) αβ (z) is the activation function f1 αβ (z) rotated counterclockwise by an angle α and then taking the real part, and then rotated by an angle α in the opposite direction of the previous rotation operation. That is, the activation function f3 αβ (z) is a function corresponding to an operation including a rotation operation around the origin, an operation of taking the real part of a complex number, and a rotation operation in the opposite direction of the rotation operation.

[0149] Also, the activation function f4 given by Equation (14) αβ (z) is obtained by adding to the activation function f3 αβ (z) the result of rotating the complex number z counterclockwise by an angle α and then taking the imaginary part, and then rotating by an angle α in the opposite direction of the previous rotation operation.

[0150] Also, in Equation (15), A legacy is a standard activation function that returns a real value for a given real value. Examples of A legacy include, for example, the sigmoid function, softsign function, softplus function, tanh function, ReLU, clipped power function, polynomial, radial basis function, wavelet, etc. The activation function f5 given by Equation (15) αβ (z) is basically a function similar to the activation function f3 αβ (z), but additionally includes an operation of applying the activation function A legacy defined by a real number after taking the real part.

[0151] Summarizing the above, f1 represented by Equation (9) αβ (z), f2 represented by Equation (10) αβ (z), f3 represented by Equation (13) αβ (z), f4 represented by Equation (14) αβ (z), f5 represented by Equation (15) αβ(z) serves as a specific example of the activation function used in the activation layer 5d in FIGS. 13 and 14 and the activation layer 5c2 in FIG. 15.

[0152] According to at least one of the embodiments described above, the image quality can be improved.

[0153] Although several embodiments have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, replacements, changes, and combinations of embodiments can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope and gist of the invention, as well as in the invention described in the claims and its equivalent scope.

Description of Reference Numerals

[0154] 110 Processing circuit 132 Memory 134 Input device 135 Display

Claims

1. A medical data processing device having a processing unit that applies a linear operation with complex coefficients and a non-linear activation whose gain does not change according to the complex argument to medical data having complex values.

2. The medical data processing device according to claim 1, wherein the processing unit applies the result of the linear operation to the non-linear activation whose gain does not change according to the complex argument.

3. The medical data processing device according to claim 1, wherein the processing unit applies data having complex values to the non-linear activation whose gain does not change according to the complex argument.

4. The medical data processing device according to claim 1, wherein the processing unit applies the non-linear activation whose gain does not change according to the complex argument to output data having complex values.

5. The medical data processing device according to claim 1, wherein the medical data having complex values is magnetic resonance data or ultrasonic data.

6. The medical data processing device according to claim 1, wherein the gain is a value indicating the magnitude of the output signal with respect to the magnitude of the input signal.

7. The medical data processing device according to claim 1, wherein the non-linear activation is a function including a real function regarding the absolute value of the input complex number.

8. The medical data processing device according to claim 7, wherein the non-linear activation is a function obtained by multiplying the real function by the input complex number.

9. The medical data processing device according to claim 7 or 8, wherein the real function constituting the non-linear activation is any one of a function including a soft-shrink function, a tanh-shrink function, a power function, or a ReLU function.

10. The medical data processing device according to claim 1, wherein the processing unit applies non-linear activations of a plurality of different function systems to the medical data.

11. The medical data processing device according to claim 1, wherein the processing unit applies the non-linear activation to the medical data using a plurality of functions that are the same function system but have different function parameters.

12. The processing unit includes a neural network that applies the non-linear activation to the medical data, and a calculation unit that optimizes function parameters related to the non-linear activation. The medical data processing device according to claim 1.

13. The medical data processing device according to claim 1, wherein the processing unit further applies an activation whose gain changes according to the complex argument angle.

14. The medical data processing device according to claim 13, wherein the processing unit sequentially applies a non-linear activation that does not depend on the complex argument angle and an activation whose gain changes according to the complex argument angle.

15. The medical data processing device according to claim 13, wherein the processing unit selectively applies a non-linear activation that does not depend on the complex argument angle and an activation whose gain changes according to the complex argument angle.

16. The medical data processing device according to claim 1, wherein the linear operation in the processing unit includes a process of fixing a bias term to 0.

17. The medical data processing device according to claim 1, wherein the linear operation in the processing unit includes complex convolution.

Images