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 1 is a diagram showing an example of a data processing apparatus according to an embodiment. [Figure 2] Figure 2 is a diagram showing an example of a magnetic resonance imaging apparatus according to an embodiment. [Figure 3] Figure 3 is a diagram showing an example of an ultrasonic diagnostic apparatus according to an embodiment. [Figure 4] Figure 4 is a diagram showing an example of a neural network according to an embodiment. [Figure 5] Figure 5 is a diagram for explaining a neural network according to an embodiment. [Figure 6] Figure 6 is a diagram for explaining an example of a neural network according to an embodiment. [Figure 7] Figure 7 is a diagram for explaining an example of a neural network according to an embodiment. [Figure 8] Figure 8 is a diagram for explaining the background according to an embodiment. [Figure 9] Figure 9 is a diagram for explaining an example of a neural network according to an embodiment. [Figure 10] Figure 10 is a diagram for explaining an example of the processing performed by a neural network according to an embodiment. [Figure 11] Figure 11 is a diagram for explaining an example of a neural network according to an embodiment. [Figure 12] Figure 12 is a diagram for explaining an example of a neural network according to an embodiment. [Figure 13] Figure 13 is a diagram for explaining an example of a neural network according to an embodiment. [Figure 14] Figure 14 is a diagram for explaining an example of a neural network according to an embodiment. [Figure 15] Figure 15 is a diagram for explaining an example of a neural network according to an embodiment.

Mode for Carrying Out the Invention

[0008] (Embodiment) The following describes in detail the medical data processing device and embodiments of the data processing device with reference to the drawings.

[0009] The configuration of the data processing device 100 according to this embodiment will be explained using Figure 1.

[0010] The data processing device 100 is a device that generates data using machine learning. For example, the data processing device 100 is connected to various medical imaging diagnostic devices such as the magnetic resonance imaging device shown in Figure 2 and the ultrasound diagnostic device shown in Figure 3, and performs tasks such as processing signals received from the medical imaging diagnostic device, generating trained models, and executing trained models. The medical imaging diagnostic device to which the data processing device 100 is connected is not limited to magnetic resonance imaging devices and ultrasound diagnostic devices, but may also be other devices such as X-ray CT scanners or PET scanners.

[0011] The data processing device 100 is typically a medical data processing device that processes medical data. However, the embodiments are not limited to the case where the data processing device 100 is a medical data processing device. For 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 comprises 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 this embodiment, each processing function performed by the training data creation function 110a, learning function 110b, interface function 110c, control function 110d, application function 110e, and acquisition function 110f, as well as the trained model (e.g., a neural network), are stored in memory 132 in the form of a program that can be executed by a computer. The processing circuit 110 is a processor that reads the program from memory 132 and executes it to realize the function corresponding to each program. In other words, the processing circuit 110, when it has read each program, will have the functions shown in the processing circuit 110 in Figure 1. Furthermore, the processing circuit 110, when it has read the program corresponding to the trained model (neural network), can perform processing according to that trained model. In Figure 1, the functions of the processing circuit 110 are described as being realized by a single processing circuit, but it is also possible to configure the processing circuit 110 by combining multiple independent processors, with each processor realizing the functions by executing a program. In other words, each of the above functions may be configured as a program, and one processing circuit may execute each program. Furthermore, a single processing circuit may implement two or more of the functions of the processing circuit 110. As another example, a specific function may be implemented in a dedicated, independent program execution circuit.

[0014] In Figure 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 examples of a processing unit, creation unit, input unit (learning unit), reception unit, control unit, application unit, and acquisition unit, respectively.

[0015] In the above explanation, the term "processor" refers to circuits such as a CPU (Central Processing Unit), a GPU (Graphical Processing Unit), an Application Specific Integrated Circuit (ASIC), or 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 functions by reading and executing programs stored in memory 132.

[0016] Alternatively, instead of saving the program in memory 132, the program may be directly incorporated into the processor's circuitry. In this case, the processor performs its functions by reading and executing the program incorporated into the circuitry. Therefore, for example, instead of saving the trained model in memory 132, the program related to the trained model may be directly incorporated into the processor's circuitry.

[0017] Furthermore, if the processing circuit 110 is incorporated into various medical imaging diagnostic devices, or if it performs processing in conjunction with various medical imaging diagnostic devices, it may have a function to execute related processing together.

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

[0019] The processing circuit 110 uses the training data generated by the training data creation function 110a to perform training using the learning function 110b and generate a trained model.

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

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

[0022] Furthermore, the processing circuit 110 generates an image based on the results of processing performed using the training data generation function 110a and the learning function 110b, via the application function 110e. Additionally, the processing circuit 110 applies the trained model generated by the learning function 110b to the input image via the application function 110e, and generates an image based on the results of applying the trained model.

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

[0024] Memory 132 stores various data, such as control programs for 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 trackball, a selection device such as a mode switch, or an input device such as a keyboard.

[0026] The display 135, under the control of the control function 110d, etc., displays a GUI (Graphical User Interface) for accepting input of imaging conditions, and images generated by the control function 110d, etc. The display 135 is 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, keyboard, buttons, panel switches, touch command screen, foot switch, trackball, joystick, etc.

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

[0028] As shown in Figure 2, the magnetic resonance imaging apparatus 200 comprises 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 transmitting coil 207, a transmitting circuit 208, a receiving coil 209, a receiving circuit 210, a sequence control circuit 220 (sequence control unit), and the data processing device 100 described in Figure 1. Note that the magnetic resonance imaging apparatus 200 does not include a subject P (e.g., a human body). Also, the configuration shown in Figure 2 is merely an example.

[0029] The static magnetic field magnet 201 is a hollow, substantially cylindrical magnet that generates a static magnetic field in its internal space. The static magnetic field magnet 201 is, for example, a superconducting magnet and is excited by the supply of current from the static magnetic field power supply. The static magnetic field power supply supplies current to the static magnetic field magnet 201. As an alternative example, the static magnetic field magnet 201 may be a permanent magnet, in which case the magnetic resonance imaging apparatus 200 does not need to have 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 coil 203 is a hollow, substantially cylindrical coil placed inside the static magnetic field magnet 201. The gradient coil 203 is formed by combining three coils corresponding to the mutually orthogonal X, Y, and Z axes. These three coils receive current individually from the gradient power supply 204, generating gradient magnetic fields whose magnetic field strength changes along the X, Y, and Z axes. The gradient magnetic fields in the X, Y, and Z axes generated by the gradient coil 203 are, for example, a slicing gradient magnetic field Gs, a phase encoding gradient magnetic field Ge, and a readout gradient magnetic field Gr. The gradient power supply 204 supplies current to the gradient coil 203.

[0031] The bed 205 is equipped with 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 port) of the gradient magnetic field coil 203 with the subject P placed on it. Normally, the bed 205 is installed so that its longitudinal direction is parallel to the central axis of the static magnetic field magnet 201. Under the control of the data acquisition device 100, the bed control circuit 206 drives the bed 205 to move the top plate 205a in the longitudinal and vertical directions.

[0032] The transmitting coil 207 is positioned inside the gradient coil 203 and receives RF pulses from the transmitting circuit 208 to generate a high-frequency magnetic field. The transmitting circuit 208 supplies RF pulses to the transmitting coil 207 that correspond to the Larmor frequency, which is determined by the type of atom being targeted and the magnetic field strength.

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

[0034] The transmitting coil 207 and receiving coil 209 described above are merely examples. The system can be constructed by combining one or more coils, such as a coil with only a transmitting function, a coil with only a receiving function, or a coil with both transmitting and receiving functions.

[0035] The receiving circuit 210 detects the magnetic resonance signal output from the receiving coil 209 and generates magnetic resonance data based on the detected magnetic resonance signal. Specifically, the receiving circuit 210 generates magnetic resonance data by digitally converting the magnetic resonance signal output from the receiving coil 209. The receiving circuit 210 also transmits the generated magnetic resonance data to the sequence control circuit 220. The receiving circuit 210 may also be provided on the mounting device side, which includes the static magnetic field magnet 201 and the gradient magnetic field coil 203.

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

[0037] Furthermore, the sequence control circuit 220 drives the gradient power supply 204, the transmitting circuit 208, and the receiving circuit 210 to image the subject P. Upon receiving magnetic resonance data from the receiving circuit 210, the received magnetic resonance data is transferred to the data processing device 100. In addition to the processing described in Figure 1, the data processing device 100 performs overall control of the magnetic resonance imaging apparatus 200.

[0038] Returning to Figure 1, the processing performed by the data processing device 100, other than the processing described in Figure 1, will be explained. The processing circuit 110 transmits sequence information to the sequence control circuit 220 via the interface function 110c and receives magnetic resonance data from the sequence control circuit 220. Upon receiving the magnetic resonance data, the processing circuit 110, which has the interface function 110c, stores the received magnetic resonance data in the memory 132.

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

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

[0041] The processing circuit 110, through its control function 110d, performs overall control of the magnetic resonance imaging apparatus 200, controlling imaging, image generation, image display, etc. For example, the processing circuit 110 with the control function 110d receives input of imaging conditions (imaging parameters, etc.) via a GUI and generates sequence information according to the received imaging conditions. The processing circuit 200 with the control function 110d then transmits the generated sequence information to the sequence control circuit 220.

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

[0043] Figure 3 shows an example configuration of an ultrasound diagnostic apparatus 300 incorporating the data processing device 100 according to the embodiment. The ultrasound diagnostic apparatus according to the embodiment includes an ultrasound probe 305 and an ultrasound diagnostic apparatus body 300. The ultrasound diagnostic apparatus body 300 includes a transmitting circuit 309, a receiving circuit 311, and the data processing device 100 described above.

[0044] The ultrasound probe 305 has multiple piezoelectric transducers, which generate ultrasound based on a drive signal supplied from a transmission circuit 309 of the ultrasound diagnostic device body 300, which will be described later. The multiple piezoelectric transducers of the ultrasound probe 305 also receive reflected waves from the subject P and convert them into electrical signals (reflected wave signals). The ultrasound probe 305 also has a matching layer provided on the piezoelectric transducers and a backing material to prevent the propagation of ultrasound backward from the piezoelectric transducers. The ultrasound probe 305 is detachably connected to the ultrasound diagnostic device body 300. The ultrasound probe 305 is also an example of a scanning unit.

[0045] When ultrasound is transmitted from the ultrasound probe 305 to the subject P, the transmitted ultrasound is reflected one after another by discontinuities in acoustic impedance within the subject P's internal tissues. The reflected waves are received by multiple piezoelectric transducers on the ultrasound probe 305 and converted into reflected wave signals. The amplitude of the reflected wave signal depends on the difference in acoustic impedance at the discontinuities where the ultrasound is reflected. If the transmitted ultrasound pulse is reflected by a moving blood flow or the surface of the heart wall, the reflected wave signal undergoes a frequency shift due to the Doppler effect, depending on the velocity component of the moving object relative to the ultrasound transmission direction.

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

[0047] As illustrated in Figure 3, the ultrasound diagnostic device 10 includes a transmitting circuit 309, a receiving circuit 311, and a medical image processing device 100.

[0048] The transmitting circuit 309 and the receiving circuit 311 control the ultrasonic transmission and reception performed by the ultrasonic probe 305 based on instructions from the data processing device 110, which has a control function. The transmitting circuit 309 includes a pulse generator, a transmission delay unit, a pulser, etc., and supplies a drive signal to the ultrasonic probe 305. The pulse generator repeatedly generates rate pulses to form the transmitted ultrasonic waves at a predetermined pulse repetition frequency (PRF). The transmission delay unit provides a delay time for each piezoelectric transducer necessary to focus the ultrasonic waves generated from the ultrasonic probe 305 into a beam and determine the transmission directivity, to each rate pulse generated by the pulse generator. The pulser applies a drive signal (drive pulse) to the ultrasonic probe 305 at a timing based on the rate pulse.

[0049] In other words, the transmission delay unit arbitrarily adjusts the transmission direction of the ultrasonic waves transmitted from the piezoelectric transducer surface by changing the delay time applied to each rate pulse. Furthermore, the transmission delay unit controls the position of the focal point (transmission focus) in the depth direction of ultrasonic transmission by changing the delay time applied to each rate pulse.

[0050] The receiving circuit 311 also includes an amplifier circuit, an A / D (Analog / Digital) converter, a receiving delay circuit, an adder, a quadrature detection 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 of the gain-corrected reflected wave signal. The receiving delay circuit gives the digital data the receiving delay time necessary to determine the receiving directivity. The adder performs summing processing on the reflected wave signal to which the receiving delay time has been given by the receiving delay circuit. Through the summing processing of the adder, the reflected component from the direction corresponding to the receiving directivity of the reflected wave signal is emphasized. The quadrature detection circuit then converts the output signal of the adder into a baseband in-phase signal (I signal, I: In-phase) and a quadrature signal (Q signal, Q: Quadrature-phase). The quadrature detection circuit then transmits the I signal and Q signal (hereinafter referred to as the IQ signal) as received signals (reflected wave data) to the processing circuit 110. Alternatively, the quadrature detection circuit may convert the output signal of the adder into an RF (Radio Frequency) signal before transmitting it to the processing circuit 110. The IQ signal and RF signal are received signals that contain phase information.

[0051] When scanning a two-dimensional region within the subject P, the transmitting circuit 309 causes the ultrasonic probe 305 to transmit an ultrasonic beam for scanning the two-dimensional region. The receiving circuit 311 then generates a two-dimensional received signal from the two-dimensional reflected wave signal received from the ultrasonic probe 305. When scanning a three-dimensional region within the subject P, the transmitting circuit 309 causes the ultrasonic probe 305 to transmit an ultrasonic beam for scanning the three-dimensional region. The receiving circuit 311 then 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 transmitting circuit 309 causes the ultrasonic probe 305 to transmit an ultrasonic beam from a predetermined transmitting position (transmitting scan line). The receiving circuit 311 receives the signal from the ultrasonic probe 305 at a predetermined receiving position (receiving scan line) due to the reflected wave of the ultrasonic beam transmitted by the transmitting circuit 309. If parallel simultaneous reception is not performed, the transmitting scan line and the receiving scan line will be the same scan line. On the other hand, if parallel simultaneous reception is performed, when the transmitting circuit 309 transmits one ultrasonic beam to the ultrasonic probe 305 on one transmitting scan line, the receiving circuit 311 simultaneously receives the signal from the reflected wave originating from the ultrasonic beam transmitted by the transmitting circuit 309 to the ultrasonic probe 1 as multiple receiving beams at multiple predetermined receiving positions (receiving scan lines) through the ultrasonic probe 305.

[0053] The data processing unit 100 is connected to the transmitting circuit 309 and the receiving circuit 311, and in addition to the functions already shown in Figure 1, it processes the signal received from the receiving circuit 311, controls the transmitting circuit 309, generates a trained model, executes the trained model, and performs various image processing. The processing circuit 110, in addition to the functions already shown in Figure 1, includes B-mode processing, Doppler processing, generation functions, etc. Returning to Figure 1, we will now describe the configurations that the data processing unit 100 incorporated into the ultrasound diagnostic device 10 may have in addition to the configuration already shown in Figure 1.

[0054] The B-mode processing function, Doppler processing function, and generation function, along with the trained models, are stored in memory 132 in the form of programs executable by the computer. The processing circuit 110 is a processor that reads programs from memory 132 and executes them to realize the functions corresponding to each program. In other words, the processing circuit 110, in the state where each program has been read, will have these functions.

[0055] The B-mode processing function and Doppler processing function are examples of B-mode processing units and Doppler processing units.

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

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

[0058] Furthermore, the processing circuit 110 uses Doppler processing to analyze the velocity information from the received signal (reflected wave data) received from the receiving circuit 311 using frequency analysis, and generates data (Doppler data) that extracts moving object information such as velocity, dispersion, and power at multiple points due to the Doppler effect.

[0059] Furthermore, the B-mode processing function and Doppler processing function can process both 2D and 3D reflected wave data.

[0060] Furthermore, the processing circuit 110 controls the entire processing of the ultrasound diagnostic device through the control function 110d. Specifically, the processing circuit 110 controls the processing of the transmission circuit 309, the reception 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 data read from the memory 132, through the control function 110d. In addition, the processing circuit 110 controls the display of ultrasound image data stored in the memory 132 on the display 135 through the control function 110d.

[0061] Furthermore, the processing circuit 110 generates ultrasonic image data from the data generated by the B-mode processing function and the Doppler processing function using a generation function (not shown). The processing circuit 110 generates two-dimensional B-mode image data, in which the intensity of reflected waves is represented by brightness, from the two-dimensional B-mode data generated by the B-mode processing function using a generation function. The processing circuit 110 also generates two-dimensional Doppler image data representing moving object information from the two-dimensional Doppler data generated by the Doppler processing function 110b using a generation function. The two-dimensional Doppler image data is velocity image data, dispersion image data, power image data, or image data combining these.

[0062] Furthermore, the processing circuit 110, through its generation function, converts the scan line signal sequence of the ultrasonic scan into a scan line signal sequence of a video format, such as that used in televisions, to generate ultrasonic image data for display. In addition to scan conversion, the processing circuit 110, through its generation function, also performs various image processing tasks, such as image processing that regenerates an average brightness image using multiple image frames after scan conversion (smoothing process), and image processing that uses a differential filter within the image (edge ​​enhancement process). Furthermore, the processing circuit 110, through its generation function, 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.

[0063] Memory 132 can also store data generated during B-mode processing and Doppler processing. The B-mode data and Doppler data stored in memory 132 can be retrieved by the operator after a diagnosis, for example, and become ultrasound image data for display via the processing circuit 110. Memory 132 can also store the received signal (reflected wave data) output by the receiving circuit 311.

[0064] In addition, memory 132 stores control programs for ultrasound transmission and reception, image processing and display processing as needed, as well as various data such as diagnostic information (e.g., patient ID, physician's findings, etc.), diagnostic protocols, and various body marks.

[0065] Next, the configuration of the neural network according to the embodiment will be explained using Figures 4 to 6.

[0066] Figure 4 shows an example of the inter-layer interconnection in a neural network 7 used for machine learning by a processing circuit 110 having a learning function 110b. The neural network 7 consists of an input layer 1, an output layer 2, and hidden layers 3, 4, 5, etc., between the input layer 1 and the output layer 2. Each hidden layer consists of a layer related to its 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, hidden layer 3 consists of an input layer 3a, a linear layer 3b, and an activation layer 3c; hidden layer 4 consists of an input layer 4a, a linear layer 4b, and an activation layer 4c; and hidden layer 5 consists of an input layer 5a, a linear layer 5b, and an activation layer 5c. Each layer also consists of multiple nodes (neurons).

[0067] Here, the data processing device 100 according to this embodiment applies a linear layer with complex coefficients and a nonlinear activation function to medical data that has complex values. That is, the processing circuit 110 generates a trained model by training a neural network 7 that applies a linear layer with complex coefficients and a nonlinear activation function to medical data that has complex values ​​using the learning function 110b. The processing circuit 100 stores the generated trained model in, for example, memory 132, as needed.

[0068] The data input to input layer 1 is typically medical images or medical image data acquired from a medical imaging device. If the medical imaging device is a magnetic resonance imaging device 200, the data input to input layer 1 is, for example, a magnetic resonance image. If the medical imaging device is, for example, an ultrasound diagnostic device 300, the data input to input layer 1 is, for example, an ultrasound image.

[0069] Furthermore, the input data input to input layer 1 may be a medical image, or it may be various image data, projection data, intermediate data, or raw data from the stage prior to the generation of a medical image. For example, if the medical imaging diagnostic device is a PET device, the input data input to input layer 10 may be a PET image, or it may be various data before the reconstruction of the PET image, such as time-series data of simultaneous coefficient information.

[0070] Furthermore, the data output from output layer 2 may be, for example, a medical image or medical image data, or, similar to the data input to input layer 1, various projection data, intermediate data, or raw data from the stage prior to the generation of a medical image. If the purpose of the neural network 7 is denoising, the data output from output layer 2 may be, for example, a high-quality image from which noise has been removed compared to the input image.

[0071] In the case where the neural network 7 according to this embodiment is, for example, a convolutional neural network (CNN), the data input to the input layer 1 is, for example, data represented by a two-dimensional array of size such as 32x32, and the data output from the output layer 2 is, for example, data represented by a two-dimensional array of size such as 32x32. 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 hidden 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 a trained model according to the embodiment, i.e., the training step, will be described. The processing circuit 110 generates a trained model by performing machine learning on the neural network 7, for example, using the training function 110b. Here, performing machine learning means, for example, determining the weights in the neural network 7, which consists of an input layer 1, hidden layers 3, 4, and 5, and an output layer 2. Specifically, it means determining sets of coefficients that characterize the connection between the input layer 1 and the hidden layer 3, sets of coefficients that characterize the connection between the hidden layer 3 and the hidden layer 4, ..., sets of coefficients that characterize the connection between the hidden layer 5 and the output layer 2. The processing circuit 150 determines these sets of coefficients by, for example, the backpropagation method using the training function 110b.

[0073] The processing circuit 110, using its learning function 110b, performs machine learning based on training data, which is the teacher data consisting of data input to the input layer 1 and data output to the output layer 12. It determines the weights between each layer and generates a trained model with the determined weights.

[0074] Furthermore, in deep learning, autoencoders can be used, and in this case, the data required for machine learning does not need to be supervised data.

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

[0076] Returning to the explanation of activation functions and activation layers, we will explain the activation function in neural network 7 using Figure 5. In Figure 5, nodes 10a, 10b, 10c, and 10d are excerpts of nodes in the input layer of a certain layer. On the other hand, node 11 is one of the nodes in the linear layer, and node 12 is one of the nodes in the activation layer, which is the layer involved in processing (activation) using the activation function. Although a complex number neural network is used in the embodiment, we will first explain the case of a real number neural network.

[0077] Here, the output values ​​of nodes 10a, 10b, 10c, and 10d are real numbers x1 and x 2、 x 3、 Considering the case where x is 4, the output result to node 11 in the linear layer is Σ i=1 m (ω i x i It is given by +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, if the output result output to node 12, which is an activation layer, is \(y\), then \(y\) 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] In FIG. 6, the processing using such an activation function is shown. In FIG. 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. Also, FIG. 6 is a real-valued 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 of 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 is given. 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 This can be said to be the zeroth-order coefficient in the linear operation in the linear layer. Next, 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 f n,j (Σ i=1 m ω i,j x n,i +b n,j The result is given by ). 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] Next, the values ​​of each node in the input layer 6a of the nth layer are determined based on the values ​​output by the nodes of the activation layer 5c. For example, the values ​​of each node in the activation layer 5c are directly input to each node in the input layer 6a. Alternatively, the nodes of the input layer 6a may be determined by applying a further nonlinear function to the activation layer 5c.

[0083] Next, we will explain complex neural networks. Figure 7 shows the configuration when the neural network 7 used for machine learning by the processing circuit 110 having a learning function 110b is a complex neural network. The complex neural network shown in Figure 7 has a similar configuration to the real-number neural network shown in Figure 6, but the value z of each node is different. n,j The value becomes a complex number. Also, the activation function f n,j This is a function defined in the complex number domain and takes complex values. 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 Given this, for example, equation (2) below holds true. In this way, for example, by making each node of neural network 7 a complex number node, it is possible to generate neural network 7 with complex number values.

[0084]

number

[0085] Next, the background of the embodiment will be explained.

[0086] In machine learning using neural networks, real-number neural networks are the standard. However, in medical data processing devices such as magnetic resonance imaging devices and ultrasound diagnostic devices, complex-number signal processing is frequently used, and therefore, using complex-number neural networks is expected to open up a variety of applications.

[0087] One way to handle complex numbers in a neural network is to separate the complex number into a real part and an imaginary part, and consider each part as a node in a standard real-number neural network. As an example, one can consider handling complex numbers in a neural network by applying the ReLU activation function to both the real and imaginary parts of the complex number.

[0088] Another approach involves representing complex numbers using their absolute value (or signed absolute value) and phase, and considering each of these as a node in a standard real-number neural network, thereby handling complex numbers within a neural network.

[0089] In medical imaging, such as magnetic resonance imaging and ultrasound imaging, the phase information of the image, such as the phase gradient, is often important, while the absolute value of the phase is relatively less significant. For example, in magnetic resonance imaging, slight differences in the center frequency appear as phase modulation of the entire image, but in most cases, the importance of the absolute values ​​of these phases themselves is relatively low. Therefore, when applying complex neural networks to medical images, such as for noise reduction or region extraction, it is desirable to configure a neural network that does not ignore the phase information of the input image, while at the same time ensuring that the output results do not fluctuate significantly with respect to the phase modulation of the entire image.

[0090] For example, when considering complex convolution and denoising using ReLU for both the real and imaginary parts, if the distribution of training images in a complex neural network is biased towards the real part, for example, coefficients that prioritize the real part of the input image and disregard the imaginary part may be learned.

[0091] When such a pre-trained model is applied as input to an image biased towards the imaginary part, it is expected that it will not perform as expected.

[0092] This situation is illustrated, for example, in Figure 8. Figure 8 is a graph showing the denoising performance of a trained model when a trained model is generated using a complex neural network 7 with six layers of Complex-convolution and CReLU to perform denoising. Figure 8 plots the mean square error (MSE) of the output image after applying the trained model as a function of phase, when 38 training images are prepared for training during the generation of the trained model, and a test image is applied to the generated trained model while modulating only the phase. As can be seen from Figure 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] One way to address the issue of inconsistent image quality in the output image after applying a trained model to the input image's phase is to apply random phase modulation to the training image to achieve statistical stabilization (phase argumentation). However, phase argumentation may not always be efficient for learning.

[0094] The medical image data device according to this embodiment is based on the above background and constructs a neural network by combining an angle-independent activation function on the complex plane with a linear layer whose coefficients are complex numbers.

[0095] In other words, the data processing device 100 according to this embodiment has a processing circuit 110 that applies a linear layer with complex coefficients and a nonlinear activation that is independent of the complex argument (i.e., the gain does not change according to the complex argument) to medical data that has complex values.

[0096] Figure 9 illustrates this situation. Figure 9 shows an excerpt of the configuration of one hidden layer 5 shown in Figure 4 in the neural network 7 trained by the data processing device 100 according to the embodiment. As already explained in Figure 4, the hidden layer 5 consists of an input layer 5a, a linear layer 5b, and an activation layer 5c (activation function). In the data processing device 100 according to the embodiment, the activation layer 5c is an angle-independent activation function 5c1, which is a nonlinear activation that does not depend on the complex argument.

[0097] In other words, the processing circuit 110 generates a trained model by training a neural network 7 that applies a linear layer 5b with complex coefficients and an angle-independent activation function 5c1, which is a nonlinear activation that does not depend on the complex argument, to medical data that has complex values, using the learning function 110b.

[0098] Here, an example of an angle-independent activation function 5c1 is the function system shown in equations (3) to (8). In equations (3) to (8), z represents a complex number.

[0099]

number

[0100]

number

[0101]

number

[0102]

number

[0103] In equation (3), λ is a real number, and for x to be a real number, ReLU(x) (Rectified Linear Unit) is max(0, x). The right-hand side of equation (3) is 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 with an absolute value equal to |z|-λ and an argument equal to z. Therefore, equation (3) is a nonlinear 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 * This 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 of the right-hand side can be evaluated quickly when performing numerical calculations.

[0105] Furthermore, in equation (5), β is a real number. The right-hand side of equation (5) is also expressed as a nonlinear function of the input signal, and its argument is equal to the argument of the input signal. Therefore, equation (5) represents a nonlinear 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] Furthermore, in equation (6), p is a real number. Since the contribution of a particular complex argument direction does not selectively increase on the right-hand side of equation (6), equation (6) represents a nonlinear activation that does not depend on the complex argument.

[0107] Furthermore, as variations of equation (3), the following function systems, such as equation (7) or equation (8), can also be considered.

[0108]

number

[0109]

number

[0110] The nonlinear activations in equations (3) to (8) above, which do not depend on the complex argument, are merely examples, and various other function systems are possible. 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 that includes a real function relating to the absolute value of the input complex number, for example, a function obtained by multiplying the input complex number by a real function relating to the absolute value of the input complex number, is an example of a nonlinear activation that does not depend on the complex argument. Examples of real functions that constitute such a nonlinear activation include, as mentioned above, functions that include the soft-shrink function, tanh-shrink function, power function, or ReLU function.

[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 trained model when a trained model is generated that performs denoising processing with a complex neural network 7 which has six layers of Complex-convolution with the bias term fixed to 0 and the complex tanh-shrink function shown in equation (5). In other words, the processing circuit 110 further applies a complex convolution process with the bias term fixed to 0. To put it another way, the linear operation of complex coefficients that the processing circuit 110 applies to medical data with complex values ​​includes processing 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 shows the Mean Square Error (MSE) of the output image after applying the trained model, plotted as a function of phase, when a test image is applied to the generated trained model while modulating only the phase. As can be seen in Figure 10, even when the phase of the input image changes, a stable output image with consistent quality was obtained.

[0112] The embodiments are not limited to the examples described above, and the processing circuit 110 may generate a trained model using two or more nonlinear activations that do not depend on the complex argument angle by the learning function 110b. This situation is shown in Figure 11. In Figure 11, the hidden 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 nodes of the linear layer 5b to generate an output result for the next layer, and applies the second angle-independent activation function 5c3 to some nodes of the linear layer 5b other than those nodes to generate an output result for the next layer.

[0113] Here, as a first example of using nonlinear activation that does not depend on two or more complex argument angles, one can consider the case where nonlinear activation that does not depend on complex argument angles of different function systems is used. For example, the processing circuit 110 performs learning by applying nonlinear activation of multiple different function systems to medical data using the learning function 110b, such as the complex soft-shrink function shown in equation (3) and the complex tanh-shrink function shown in equation (5). In the example in Figure 10, the processing circuit 110 performs learning by applying the complex soft-shrink function as the first angle-independent activation function 5c2 to some nodes of the linear layer 5b, and the complex tanh-shrink function as the second angle-dependent activation function 5c3 to some nodes of the linear layer 5b using the learning function 110b.

[0114] Furthermore, as a second example of using nonlinear activation that does not depend on two or more complex argument angles, one can use multiple functions of the same function system but with different function parameter values, such as λ in equations (3) and (4), β in equation (5), and p in equation (6). In other words, the processing circuit 110 performs learning by applying nonlinear activation to medical data using, for example, multiple functions of the same function system but with different function parameters, via the learning function 110b. As an example, the processing circuit 110 performs learning by applying the complex soft-shrink function given by equation (3) with λ=λ1 as the first angle-independent activation function 5c2 to some nodes of the linear layer 5b, and applying the complex soft-shrink function given by equation (3) with λ=λ2 as the second angle-dependent activation function 5c3 to some nodes of the linear layer 5b.

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

[0116] These function parameters may be fixed, or they may change during the machine learning process and be treated as variable parameters. Furthermore, a single neural network 7 may contain a mixture of fixed function parameters and variable, parameterized, and variable function parameters.

[0117] Furthermore, such function parameters may be trainable by machine learning. Figure 12 shows an example of such a configuration.

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

[0119] As an example, let's consider the case where 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-shrink function, which is the activation function in the activation layers 3c, 4c, and 5c of the first neural network 7, is λ = λ, where i is the i-th layer. i It is determined as follows: λ i λ has a constant value for each layer. i The value is optimized by the calculation unit through learning.

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

[0121] As another example, the processing circuit 110 may, after performing a second learning process, which is the learning of the parameter values ​​of the activation function of the first neural network 7, executed by the calculation unit, use those parameter values ​​to perform a first learning process, which is the learning of the weight coefficients in the first neural network 7, executed by the learning function 110b.

[0122] Furthermore, the processing circuit 110 may perform the first learning and the second learning simultaneously.

[0123] Furthermore, the configuration of the calculation unit is not limited to a neural network; for example, the values ​​of the activation function parameters of the first neural network 7 may be optimized using linear regression or the like.

[0124] Furthermore, although the above-described embodiment uses parameters that differ for each layer and are common to each node, the embodiment is not limited to this, and the parameters may be common to each layer, or they may be different for each layer and node.

[0125] In the embodiments described so far, only angle-independent activation functions have been used as the activation function of the neural network 7. However, the embodiments are not limited to this, and a complex-sensitive activation function (CPSAF) may also be used as the activation function of the neural network 7. That is, the processing circuit 110 may further apply an activation in the neural network 7, in which the gain changes according to the complex angle, using the learning function 110b.

[0126] For example, when you want to remove a specific complex argument component, such as in phase denoising, you can efficiently perform denoising by using an activation function whose gain changes according to the complex argument in combination with a nonlinear activation that does not depend on the complex argument, i.e., an activation function whose gain does not change according to the complex argument.

[0127] Specifically, as shown in Figures 13 and 14, for example, the processing circuit 110, through its learning function 110b, sequentially applies a nonlinear activation independent of the complex argument angle and an activation whose gain changes according to the complex argument angle.

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

[0129] As another example, as shown in Figure 14, the intermediate layer 5 consists of 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, through its learning function 110b, outputs the output result of the input layer 5a to the linear layer 5b, the output result of the linear layer 5b to the activation layer 5c using an angle-independent activation function, the output result of the activation layer 5c using an angle-independent activation function to the linear layer 5e, the output result of the linear layer 5e to the activation layer 5d using an activation function sensitive to the complex argument, and the output result of the activation layer 5d using an activation function sensitive to the complex argument to the input layer of the next layer.

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

[0131] As another example, as shown in Figure 15, the processing circuit 110 may selectively apply a nonlinear activation that does not depend on the complex argument angle and an activation whose gain changes according to the complex argument angle. In the example in Figure 15, the hidden layer 5 has 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 argument angle, and an activation layer 5c3 using other activation functions. Here, the processing circuit 110, through its learning function 110b, outputs the output result of the input layer 5a to the linear layer 5b, outputs the output result of some nodes from the output result of the linear layer 5b to the activation layer 5c1 using an angle-independent activation function, and outputs the output result of the activation layer 5c1 using an angle-independent activation function to the corresponding node 6a1 in the input layer of the next layer. Meanwhile, the processing circuit 110 outputs the output results of some other nodes from the output results of the linear layer 5b to an activation layer 5c2 that uses an activation function sensitive to the complex argument angle, and outputs the output results of the activation layer 5c2 that uses an activation function sensitive to the complex argument angle to the corresponding node 6a2 in the input layer of the next layer. Furthermore, the processing circuit 110 outputs the output results of some other nodes from the output results of the linear layer 5b to an activation layer 5c3 that uses another activation function, and outputs the output results of the activation layer 5c3 that uses another activation function to the corresponding node 6a3 in the input layer of the next layer.

[0132] Thus, in the example shown in Figure 15, the processing circuit 110 applies different non-selective activations to each node in the linear layer 5b. This allows for efficient machine learning that reflects the properties of the nodes in the linear layer 5b, resulting in improved image quality.

[0133] Returning to the explanation of activation functions sensitive to the complex argument, a concrete example of the complex argument sensitive activation function (CPSFA) mentioned above is the function f1 given by equation (9) below.

[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. Gain control function W β (x) is a function defined on a real number x, for example, a function that extracts angles near x=0 in a way characterized by the parameter β. Below, for example, the gain control function W β (x) is explained using an example of a function that has a maximum value at x=0 and decreases in value as x moves away from x=0. Note that angles that differ by a constant multiple of 2π can be considered the same, for example, a gain control function W β As such, it has a periodicity of 2π, W β (x+2nπ)=W β We can also choose a periodic function such that (x) holds true.

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

[0137] Another example of an activation function sensitive to the complex argument is the activation function f2(z) given by equation (10) below.

[0138]

number

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

[0140]

number

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

[0142]

number

[0143] That is, the gain control function W on the left side of equation (11) β (x) is a function that returns 1 if the angle x is within the range β relative to 0, and 0 otherwise. That is, the activation function f2 αβ (z) is a function that extracts the complex number region in the range of angle β from angle α. In other words, it is the activation function f2 expressed by equation (10). αβ (z) can be considered a function that extracts signal components within a range from a given angle α to a given angle β, and is an example of an activation function that is sensitive to complex argument angles.

[0144] Another example of an activation function sensitive to the complex argument is the activation functions f3(z) to f5(z) given by equations (13) to (15) below.

[0145]

number

[0146]

number

[0147]

number

[0148] Here, the activation function f3 is given by equation (13). αβ (z) is the activation function f1 αβ (z) is rotated clockwise by an angle α, the real part is taken, and then it is rotated by an angle α in the opposite direction to the previous rotation. That is, the activation function f3 αβ (z) is a function that corresponds to an operation that includes a rotation operation around the origin, an operation that takes the real part of a complex number, and an operation that includes a rotation operation in the opposite direction to the said rotation operation.

[0149] Furthermore, the activation function f4 given by equation (14) αβ (z) is the activation function f3 αβ (z) is obtained by rotating the complex number z clockwise by an angle α, removing the imaginary part, and then rotating it by an angle α in the opposite direction to the previous rotation.

[0150] Also, in equation (15), A legacy A is a standard activation function that returns a real value for a given real value. legacy Examples include the sigmoid function, soft sine function, soft plus function, tanh function, ReLU, power-section functions, polynomials, radial basis functions, and wavelets. The activation function f5 is given by equation (15). αβ (z) is basically the activation function f3 αβ (z) is a similar function, but after removing the real part, it becomes an activation function A defined in real numbers. legacy The operation to apply it is additionally included.

[0151] To summarize, f1 can be expressed by equation (9). αβ (z), f2 represented by equation (10) αβ (z), f3 expressed by equation (13) αβ (z), f4 represented by equation (14) αβ (z), f5 expressed by equation (15) αβ(z) is a specific example of the activation function used in the activation layer 5d in Figures 13 and 14 and the activation layer 5c2 in Figure 15.

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

[0153] While several embodiments have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be implemented in a variety of other forms, and various omissions, substitutions, modifications, and combinations of embodiments are possible without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of symbols]

[0154] 110 Processing Circuit 132 memory 134 Input device 135 displays

Claims

1. The processing unit has a mechanism that applies linear calculations of complex coefficients and nonlinear activation, where the gain does not change according to the complex argument, to medical data that has complex values. Medical data processing device.

2. The medical data processing device according to claim 1, wherein the processing unit applies the result of the linear calculation to the nonlinear activation, the gain of which does not change according to the complex argument angle.

3. The medical data processing apparatus according to claim 1, wherein the processing unit applies data having complex values ​​to the nonlinear activation, in which the gain does not change according to the complex deviation angle.

4. The medical data processing apparatus according to claim 1, wherein the processing unit applies the nonlinear activation, in which the gain does not change according to the complex argument, to output data having a complex value.

5. The medical data processing apparatus according to claim 1, wherein the medical data having a complex value is magnetic resonance data or ultrasound data.

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

7. The medical data processing device according to claim 1, wherein the nonlinear activation is a function that includes a real function relating to the absolute value of the input complex number.

8. The medical data processing device according to claim 7, wherein the nonlinear 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 nonlinear activation is any 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 the nonlinear activation 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 nonlinear 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 nonlinear activation to the medical data, A calculation unit that optimizes the function parameters related to the nonlinear activation. A medical data processing device according to claim 1, comprising:

13. The medical data processing apparatus according to claim 1, wherein the processing unit further applies an activation in which the gain changes according to the complex angle.

14. The medical data processing device according to claim 13, wherein the processing unit sequentially applies a nonlinear activation independent of the complex argument and an activation whose gain changes according to the complex argument.

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

16. The medical data processing apparatus according to claim 1, wherein the linear operation in the processing unit includes a process in which the bias term is fixed to 0.

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

Images