JP2020121032A - Medical image processing apparatus, magnetic resonance imaging apparatus, medical image processing method, and program - Google Patents

Abstract

To improve image quality of a medical image.SOLUTION: A medical image processing apparatus of an embodiment comprises an acquisition unit and a generation unit. The acquisition unit acquires a medical image obtained on the basis of a magnetic resonance signal obtained by exciting an analyte using a high frequency magnetic field transmitted by an RF coil. The generation unit inputs the medical image acquired by the acquisition unit to a model on which learning is performed so that when an image is input, the model outputs a correction image obtained by reducing influence of ununiformity of the high frequency magnetic field from the input image, to generate a correction image obtained by reducing influence of the ununiformity from the medical image acquired by the acquisition unit.SELECTED DRAWING: Figure 3

Description

Embodiments of the present invention relate to a medical image processing apparatus, a magnetic resonance imaging apparatus, a medical image processing method, and a program.

An MRI apparatus using a magnet that generates a high magnetic field of 3 Tesla or higher can obtain a high SNR (Signal to Noise Ratio), but the high frequency magnetic field (RF magnetic field) used to excite the subject is high. As the magnetic field is applied, unevenness is likely to occur, which may affect the image quality. The non-uniformity of the transmission of the high-frequency magnetic field changes depending on the physique of the subject, so it is necessary to deal with it.

JP, 2013-176553, A

The problem to be solved by the present invention is to improve the image quality of medical images.

The medical image processing apparatus according to the embodiment includes an acquisition unit and a generation unit. The acquisition unit acquires a medical image obtained based on a magnetic resonance signal obtained by exciting the subject with a high-frequency magnetic field transmitted by an RF (Radio Frequency) coil. When the image is input, the generation unit acquires, from the input image, a model learned to output a corrected image in which the influence of the inhomogeneity of the high frequency magnetic field is reduced, by the acquisition unit. By inputting the medical image, a corrected image in which the influence of the non-uniformity is reduced is generated from the medical image acquired by the acquisition unit.

FIG. 1 is a diagram showing an example of a configuration of a medical image processing system 1 including a medical image processing apparatus 200 according to an embodiment. FIG. 1 is a diagram showing an example of an MRI apparatus 100 according to an embodiment. FIG. 1 is a diagram showing an example of a medical image processing apparatus 200 according to an embodiment. The figure which shows an example of the teacher information 234. 6 is a flowchart showing a flow of a series of processing at runtime of the processing circuit 210 on the medical image processing apparatus 200 side according to the embodiment. The figure which shows an example of a structure of the correction model MDL at the time of runtime concerning embodiment. 6 is a flowchart showing a flow of a series of processing at the time of training of the processing circuit 210 on the medical image processing apparatus 200 side according to the embodiment. The figure which shows an example of a structure of the correction model MDL at the time of training which concerns on embodiment. The figure which shows the other example of the MRI apparatus 100 which concerns on embodiment.

Hereinafter, embodiments of a medical image processing apparatus, a magnetic resonance imaging apparatus, a medical image processing method, and a program will be described in detail with reference to the drawings.

FIG. 1 is a diagram showing an example of the configuration of a medical image processing system 1 including a medical image processing apparatus 200 according to the embodiment. For example, as shown in FIG. 1, the medical image processing system 1 includes an MRI apparatus 100 and a medical image processing apparatus 200. The MRI apparatus 100 and the medical image processing apparatus 200 are connected via the network NW. The network NW includes, for example, a WAN (Wide Area Network), a LAN (Local Area Network), the Internet, a dedicated line, a wireless base station, a provider, and the like.

The MRI apparatus 100 is, for example, an apparatus that generates an MR image, which is one of medical images, from an MR (Magnetic Resonance) signal obtained by a nuclear magnetic resonance phenomenon generated by applying a magnetic field to a subject (for example, a human body).

The medical image processing apparatus 200 is realized by one or a plurality of processors. For example, the medical image processing apparatus 200 may be a computer included in a cloud computing system or a computer (standalone computer) that operates independently without depending on other devices.

[Configuration example of MRI apparatus]
FIG. 2 is a diagram illustrating an example of the MRI apparatus 100 according to the embodiment. As shown in FIG. 2, the MRI apparatus 100 includes a static magnetic field magnet 101, a gradient magnetic field coil 102, a gradient magnetic field power source 103, a bed 104, a bed control circuit 105, one or more transmission coils 106, and a transmission circuit. 107, one or more receiving coils 108, a receiving circuit 109, a sequence control circuit 110, and a console device 120. The transmitter coil 106 and the receiver coil 108 are also called “RF coils”.

The static magnetic field magnet 101 is a magnet formed in a hollow and substantially cylindrical shape, and generates a uniform static magnetic field in the internal space. The static magnetic field magnet 101 is, for example, a permanent magnet or a superconducting magnet. The gradient magnetic field coil 102 is a coil formed in a hollow and substantially cylindrical shape, and is arranged inside the static magnetic field magnet 101. The gradient magnetic field coil 102 is formed by combining three coils corresponding to x, y, and z axes that are orthogonal to each other. The z-axis direction represents the longitudinal direction of the top plate 104a of the bed 104, and the x-axis direction represents the axial direction orthogonal to the z-axis direction and parallel to the floor surface of the room in which the MRI apparatus 100 is installed. , Y-axis direction represents an axial direction that is a direction perpendicular to the floor surface. The three coils corresponding to the respective axial directions individually receive currents from the gradient magnetic field power source 103 and generate gradient magnetic fields whose magnetic field strength changes along the x, y, and z axes. The z-axis direction is the same as the static magnetic field.

The gradient magnetic field power supply 103 supplies a current to the gradient magnetic field coil 102. Here, the gradient magnetic fields of the x, y, and z axes generated by the gradient magnetic field coil 102 respectively correspond to the slice selection gradient magnetic field Gs, the phase encoding gradient magnetic field Ge, and the readout gradient magnetic field Gr, respectively. .. The slice selection gradient magnetic field Gs is used to arbitrarily determine an imaging cross section. The phase encoding gradient magnetic field Ge is used to change the phase of the magnetic resonance signal according to the spatial position. The readout gradient magnetic field Gr is used to change the frequency of the magnetic resonance signal depending on the spatial position.

The bed 104 includes a top plate 104a on which the subject OB is placed, and under the control of the bed control circuit 105, the top plate 104a is placed in the cavity of the gradient magnetic field coil 102 in a state where the subject OB is placed ( Insert into the imaging port). Usually, the bed 104 is installed so that the longitudinal direction thereof is parallel to the central axis of the static magnetic field magnet 101. Under the control of the console device 120, the bed control circuit 105 drives the bed 104 to move the top 104a in the longitudinal direction and the vertical direction.

The transmission coil 106 is arranged inside the gradient magnetic field coil 102. The transmission coil 106 is, for example, a whole body coil (Whole Body Coil) housed in a gantry including the static magnetic field magnet 101. The transmission coil 106 receives the supply of current from the transmission circuit 107 and generates a high frequency magnetic field for exciting the nuclear spins inside the object OB.

The transmission circuit 107 supplies the transmission coil 106 with a current corresponding to the Larmor frequency determined by the type of target nuclei and the strength of the magnetic field, and causes the transmission coil 106 to generate a high-frequency magnetic field. Further, the transmission circuit 107 may independently control the currents supplied to the plurality of feeding points of the transmission coil 106 to correct the spatial nonuniformity of the high frequency magnetic field. Spatial homogenization of such a magnetic field is called B1 shimming (RF shimming). In B1 shimming, the magnetic field applied to the object OB is spatially uniformed by adjusting the amplitude and phase of the high-frequency magnetic field generated from each transmission coil 106. Hereinafter, the high frequency magnetic field generated from the transmission coil will be referred to as “transmission RF”.

The receiving coil 108 is arranged inside the gradient magnetic field coil 102. The reception coil 108 receives a magnetic resonance signal emitted from the subject OB under the influence of the transmission RF. The magnetic resonance signal includes, for example, a signal intensity component and a phase component. Upon receiving the magnetic resonance signal, the receiving coil 108 outputs the received magnetic resonance signal to the receiving circuit 109. The receiving coil 108 may be realized by a phased array coil in which at least two or more coil elements are arranged in an array. The receiving coil 108 is not limited to the phased array coil, and may be realized by a single RF coil having both transmission and reception.

The receiving circuit 109 generates magnetic resonance data based on the magnetic resonance signal output from the receiving coil 108. Specifically, the receiving circuit 109 generates magnetic resonance data, which is a digital signal, by performing analog-digital conversion on the magnetic resonance signal of the analog signal output from the receiving coil 108. The reception circuit 109 transmits the generated magnetic resonance data to the sequence control circuit 110. Further, the receiving circuit 109 may be provided on the gantry device side including the static magnetic field magnet 101, the gradient magnetic field coil 102, and the like. The magnetic resonance signals output from the coil elements of the receiving coil 108 are appropriately distributed and combined and output to the receiving circuit 109.

The sequence control circuit 110 drives the gradient magnetic field power supply 103, the transmission circuit 107, and the reception circuit 109 based on the sequence information transmitted from the console device 120 to image (scan) the subject OB. The sequence information is information that defines the procedure for performing the imaging process. The sequence information includes the intensity of the current supplied by the gradient magnetic field power supply 103 to the gradient magnetic field coil 102, the timing of supplying the current, the amplitude of the transmission RF generated by the transmission coil 107 from the transmission coil 106, the phase of the transmission RF, and the reception circuit 109. Contains information defining the timing at which the magnetic resonance signal is detected.

When the sequence control circuit 110 drives the gradient magnetic field power supply 103, the transmission circuit 107, and the reception circuit 109 to image the subject OB and receives the magnetic resonance data from the reception circuit 109, the received magnetic resonance data is sent to the console device. Transfer to 120.

The console device 120 controls the entire MRI apparatus 100 and collects magnetic resonance data. For example, the console device 120 includes a communication interface 122, an input interface 124, a display 126, a processing circuit 130, and a memory (storage) 150.

The communication interface 122 includes a communication interface such as a NIC (Network Interface Card). The communication interface 122 communicates with the medical image processing apparatus 200 via the network NW and receives information from the medical image processing apparatus 200. The communication interface 122 outputs the received information to the processing circuit 130. Further, the communication interface 122 may be controlled by the processing circuit 130 to transmit information to another device connected via the network NW.

The input interface 124 is an interface that receives various input operations from an operator. When receiving the input operation, the input interface 124 converts the received input operation into an electric signal and outputs the electric signal to the processing circuit 130. For example, the input interface 124 is realized by a mouse, keyboard, trackball, switch, button, joystick, touch panel, or the like. The input interface 124 may be realized by, for example, a user interface such as a microphone that receives a voice input. When the input interface 124 is a touch panel, the display 126 described below may be formed integrally with the input interface 124.

The display 126 displays various information. For example, the display 126 displays an image generated by the processing circuit 130, a GUI (Graphical User Interface) for receiving various input operations from an operator, and the like. For example, the display 126 is an LCD (Liquid Crystal Display), a CRT (Cathode Ray Tube) display, an organic EL (Electroluminescence) display, or the like.

The processing circuit 130 executes, for example, an acquisition function 132, a generation function 134, a communication control function 136, and a display control function 138. The processing circuit 130 realizes these functions by, for example, a hardware processor included in a computer executing a program stored in a memory 150 that is a storage device (storage circuit).

The hardware processor that realizes each function of the processing circuit 130 is, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an application-specific integrated circuit (ASIC), a programmable logic device, or the like. Means a circuit. The programmable logic device is, for example, a simple programmable logic device (SPLD), a complex programmable logic device (CPLD), a field programmable gate array (FPGA), or the like. Instead of storing the program in the memory 150, the program may be directly incorporated in the circuit of the hardware processor. In this case, the hardware processor realizes each of the above functions by reading and executing the program incorporated in the circuit. The hardware processor is not limited to the one configured as a single circuit, and may be configured as one hardware processor by combining a plurality of independent circuits to realize each function. Also, a plurality of constituent elements may be integrated into one hardware processor to realize each function.

The memory 150 is realized by, for example, a RAM (Random Access Memory), a semiconductor memory device such as a flash memory, a hard disk, an optical disk, or the like. These non-transitory storage media may be realized by other storage devices connected via the network NW, such as NAS (Network Attached Storage) and external storage server devices. The non-transitory storage medium that realizes the memory 150 may include a storage medium such as a ROM (Read Only Memory) or a register.

The acquisition function 132 acquires magnetic resonance data from the sequence control circuit 110. As described above, the magnetic resonance data is data obtained by analog-digital conversion of the electromagnetic wave signal (magnetic resonance signal) generated in the object OB by the nuclear magnetic resonance phenomenon. The data arranged according to the amount of phase encoding and the amount of frequency encoding provided by the gradient magnetic field described above is also referred to as k-space data. The k-space represents a frequency space in which the one-dimensional waveform is collected when the magnetic resonance signal is repeatedly collected by the receiving coil 108 as the one-dimensional waveform.

The generation function 134 performs a reconstruction process including a process such as a Fourier transform (for example, an inverse Fourier transform) on the k-space data acquired by the acquisition function 132, so that the MR image reconstructed from the k-space data. To generate. The generation function 134 is an example of a “first generation unit”. These medical images may be two-dimensional images or three-dimensional images.

When the generation function 134 generates an MR image by reconstruction, the communication control function 136 transmits the reconstructed MR image to the medical image processing apparatus 200 via the communication interface 122. Further, the communication control function 136 may receive various information from the medical image processing apparatus 200 via the communication interface 122.

The display control function 138 causes the display 126 to display the MR image generated by the generation function 134. Further, when the communication interface 122 receives a medical image such as an MR image from the medical image processing apparatus 200, the display control function 138 may display the medical image received by the communication interface 122 on the display 126.

[Configuration example of medical image processing apparatus]
FIG. 3 is a diagram illustrating an example of the medical image processing apparatus 200 according to the embodiment. As shown in FIG. 3, the medical image processing apparatus 200 includes, for example, a communication interface 202, an input interface 204, a display 206, a processing circuit 210, and a memory 230.

The communication interface 202 includes, for example, a communication interface such as NIC. For example, the communication interface 202 communicates with the MRI apparatus 100 via the network NW, and receives the reconstructed MR image and the like from the MRI apparatus 100. The communication interface 202 outputs the received medical image to the processing circuit 210. Further, the communication interface 202 may be controlled by the processing circuit 210 to transmit information to the MRI apparatus 100 and other devices connected via the network NW. The other device may be, for example, a terminal device that can be used by an image reader such as a doctor.

The input interface 204 receives various input operations from the operator, converts the received input operations into electric signals, and outputs the electric signals to the processing circuit 210. For example, the input interface 204 is realized by a mouse, keyboard, trackball, switch, button, joystick, touch panel, or the like. Further, the input interface 204 may be realized by, for example, a user interface such as a microphone that receives a voice input. When the input interface 204 is a touch panel, the display 206 described later may be formed integrally with the input interface 204.

The display 206 displays various information. For example, the display 206 displays an image generated by the processing circuit 210, a GUI for receiving various input operations from the operator, and the like. For example, the display 206 is an LCD, a CRT display, an organic EL display, or the like.

The processing circuit 210 executes, for example, an acquisition function 212, a correction function 214, an output control function 216, and a learning function 218. The acquisition function 212 is an example of an “acquisition unit”, the correction function 214 is an example of a “generation unit” or a “second generation unit”, and the learning function 218 is an example of a “learning unit”.

The processing circuit 210 realizes these functions, for example, by a hardware processor included in a computer executing a program stored in a memory 230 that is a storage device (storage circuit).

The hardware processor that realizes each function of the processing circuit 210 means, for example, a circuit such as a CPU, a GPU, an application-specific integrated circuit, or a programmable logic device. Instead of storing the program in the memory 230, the program may be directly incorporated in the circuit of the hardware processor. In this case, the hardware processor realizes each of the above functions by reading and executing the program incorporated in the circuit. The hardware processor is not limited to the one configured as a single circuit, and may be configured as one hardware processor by combining a plurality of independent circuits to realize each function. Also, a plurality of constituent elements may be integrated into one hardware processor to realize each function.

The memory 230 is realized by, for example, a RAM, a semiconductor memory device such as a flash memory, a hard disk, an optical disk, or the like. These non-transitory storage media may be realized by another storage device connected via the network NW such as a NAS or an external storage server device. Further, the memory 230 may include a non-transitory storage medium such as a ROM or a register. The memory 230 stores, for example, correction model information 232, teacher information 234, a sensitivity map 236, and the like.

The correction model information 232 is information (program or data structure) defining a correction model MDL described later. The correction model MDL is a model learned to output an image in which the influence of the non-uniformity of the transmitted high frequency magnetic field is reduced with respect to the input image. Hereinafter, the non-uniformity of the transmitted high-frequency magnetic field will be referred to as “Tx unevenness”, and the image with reduced Tx unevenness will be referred to as “corrected image”.

The sensitivity map is data of spatial distribution of sensitivity, which represents the sensitivity nonuniformity of the receiving coil 108.

The correction model MDL may be implemented by a DNN (Deep Neural Network(s)) such as a CNN (Convolutional Neural Network).

When the correction model MDL is realized by the DNN, the correction model information 232 includes, for example, an input layer, one or more hidden layers (intermediate layers), and an output layer that configure each DNN included in the correction model MDL. Includes connection information about how the included neurons (units or nodes) are connected to each other, and weight information about how many connection coefficients are given to the data input/output between the connected neurons. Be done. The connection information is, for example, information such as the number of neurons included in each layer, information specifying the type of neuron to which each neuron is connected, an activation function that realizes each neuron, and a gate provided between neurons in the hidden layer. including. The activation function that realizes the neuron may be, for example, a function (ReLU (Rectified Linear Unit) function, ELU (Exponential Linear Units) function, clipping function) that switches operation according to an input code, or a sigmoid function or , A step function, a hyperpolic tangent function, or an identity function. The gate selectively passes or weights data transmitted between neurons, for example, depending on the value returned by the activation function (eg, 1 or 0). The coupling coefficient is a parameter of the activation function. For example, in a hidden layer of a neural network, when data is output from a neuron of a certain layer to a neuron of a deeper layer, a weight given to the output data. including. Further, the coupling coefficient may include a bias component unique to each layer.

The teacher information 234 is information used for learning the above-described correction model MDL with a teacher. FIG. 4 is a diagram showing an example of the teacher information 234. As shown in the figure, the teacher information 234 is information in which a learning image to be input to the correction model MDL is associated with a teacher image as a model to be output by the correction model MDL as a teacher label. The learning image is a medical image generated when the transmission RF is not spatially uniform in the object OB, and is, for example, a medical image including Tx unevenness. The teacher image is a medical image generated under the condition that the transmission RF is spatially and uniformly irradiated in the subject OB, and is, for example, Tx as compared with a medical image not including Tx unevenness or a learning image. It is a medical image with less unevenness. The medical image used as the teacher image is, for example, a medical image acquired in a low magnetic field MRI system having a magnetic field strength of about 1.0 Tesla, or a medical image obtained by using an MP2RAGE (Magnetization Prepared 2 Rapid Acquisition Gradient Echoes) sequence. Is.

Rx unevenness is a magnetic resonance signal generated in the subject OB when the transmitting coil 106 realized by the whole-body coil is caused to transmit the transmitting RF by the receiving coil 108 realized by the phased array coil. This is the unevenness of pixel values on the sensitivity map (hereinafter referred to as the reception sensitivity map) obtained when

The sensitivity map 236 is a sensitivity map showing the spatial distribution of Rx unevenness. The sensitivity map 236 is the reception sensitivity map obtained when the reception coil 108 realized by the phased array coil receives the magnetic resonance signal, and the transmission coil 106 realized by the whole body coil receives the magnetic resonance signal. It is generated by dividing by the transmission sensitivity map obtained when The purpose of division by the transmission sensitivity map is to eliminate the influence of Tx unevenness from the reception sensitivity map. The sensitivity map 236 is an example of “reception sensitivity distribution”.

Returning to the description of FIG. 3, the acquisition function 212 acquires a medical image, that is, an MR image from the MRI apparatus 100 via the communication interface 202.

The correction function 214 uses the correction model MDL indicated by the correction model information 232 to determine the Tx unevenness caused by the transmission coil 106 and the Rx unevenness caused by the reception coil 108 from the MR image acquired by the acquisition function 212. Remove. For example, the correction model MDL may be implemented as part of the correction function 214 by the processor executing the correction model MDL.

The output control function 220 controls the communication interface 202 to transmit the corrected image from which the Tx unevenness and the Rx unevenness are removed by the correction function 214 to the MRI apparatus 100. Further, the output control function 220 may display the corrected image from which the Tx unevenness and the Rx unevenness are removed by the correction function 214 on the display 206.

The learning function 222 inputs the learning image included in the teacher information 234 to the correction model MDL, and the image output by the correction model MDL is used as a teacher label in advance for the learning image input to the correction model MDL. The correction model MDL is learned so as to approach the associated teacher image.

For example, when the correction model MDL is a neural network, the learning function 222 performs correction so that the difference between the image output by the correction model MDL and the teacher image (for example, the difference in pixel value for each pixel) becomes small. The parameters of the model MDL are learned using a stochastic gradient descent method such as SGD (Stochastic Gradient Descent), Momentum SGD, AdaGrad, RMSprop, AdaDelta, and Adam (Adaptive moment estimation).

[Processing Flow of Medical Image Processing Device at Runtime]
Hereinafter, a flow of a series of processes of the processing circuit 210 on the medical image processing apparatus 200 side according to the embodiment will be described with reference to a flowchart. FIG. 5 is a flowchart showing a flow of a series of processes at the runtime of the processing circuit 210 on the medical image processing apparatus 200 side according to the embodiment. The runtime is to execute the processing by using the sufficiently learned correction model MDL. The process of this flowchart may be repeated, for example, in a predetermined cycle. Note that, when the processor that realizes the processing circuit 210 of the medical image processing apparatus 200 is configured by combining a plurality of independent circuits, that is, when the processing circuit 210 of the medical image processing apparatus 200 is realized by a plurality of processors. A part or all of the processing of this flowchart may be processed in parallel by a plurality of processors.

First, the correction function 214 waits until the acquisition function 212 acquires a medical image from the MRI apparatus 100 via the communication interface 202 (step S100), and the acquisition function 212 receives the medical image from the MRI apparatus 100 via the communication interface 202. When is acquired, the medical image is input to the sufficiently learned correction model MDL (step S102).

Next, the correction function 214 causes the correction model MDL to which the medical image is input to output a correction image in which Tx unevenness and Rx unevenness are reduced from the input medical image (step S104).

Next, the output control function 216 transmits the corrected image output by the correction model MDL to the MRI apparatus 100 via the communication interface 202 (step S106). In response to this, the display control function 138 of the processing circuit 130 of the MRI apparatus 100 causes the display 126 to display the corrected image received from the medical image processing apparatus 200. The output control function 216 may display the corrected image output by the correction model MDL on the display 206 instead of or in addition to transmitting it to the MRI apparatus 100.

Image correction processing by the correction model MDL at runtime will be described below with reference to the configuration diagram of the correction model MDL. FIG. 6 is a diagram showing an example of the configuration of the correction model MDL at runtime according to the embodiment. As in the illustrated example, the correction model MDL at runtime according to the embodiment may include a first layer 310, a second layer 320, a third layer 330, and a CNN (Convolutional Neural Network) 340.

The first layer 310 corrects the brightness of the input medical image IMG_IN using the sensitivity map 236. As a result, Rx unevenness included in the input medical image IMG_IN is reduced. The first layer 310 outputs the medical image IMG_IN subjected to the brightness correction, that is, the medical image IMG_IN with reduced Rx unevenness, to the second layer 320 and the third layer 330.

The second layer 320 smoothes the medical image IMG_IN input from the first layer 310 and generates a medical image from which high frequency components have been removed (hereinafter referred to as a low frequency component image IMG_A). For example, the second layer 320 may be an LPF (Low-Pass Filter). The second layer 320, which is an LPF, repeats, for example, a product-sum operation on a medical image IMG_IN while sliding a linear conversion matrix called a filter or a kernel with a certain stride amount. The linear transformation matrix may be a box-type filter in which weighting factors (filtering factors) are normalized to the same value for all elements (Averaging), or a Gaussian that changes the weighting factor according to the distance to a pixel of interest. It may be a filter (Gaussian Filtering), or another filter such as a median filter or a bilateral filter. Further, the second layer 320 may perform morphological processing including various processing such as Erosion, Dilation, Opening, Closing, and Gradient on the medical image IMG_IN to smooth the medical image IMG_IN.

When the second layer 320 generates the low frequency component image IMG_A, the second layer 320 outputs the generated low frequency component image IMG_A to the CNN 340. The low frequency component image IMG_A is an example of a “low frequency component image”.

The third layer 330 sharpens the medical image IMG_IN input from the first layer 310 and generates a medical image from which low frequency components have been removed (hereinafter referred to as high frequency component image IMG_B). For example, the third layer 330 may be an HPF (High-Pass Filter). The third layer 330, which is an HPF, outputs, for example, a difference image obtained by subtracting the low frequency component image IMG_A generated by the second layer 320 from the medical image IMG_IN input from the first layer 310 as a high frequency component image IMG_B. .. The high frequency component image IMG_B is an example of a “high frequency component image”.

The CNN 340 includes, for example, an input layer 342, a plurality of convolutional layers 344, a plurality of activation layers 346, and an output layer 348.

The low frequency component image IMG_A is input to the input layer 342 from the second layer 320. For example, when the low-frequency component image IMG_A is a matrix having elements corresponding to the respective pixels, the input layer 342 adds a bias component to the matrix corresponding to the low-frequency component image IMG_A at a suitable time. Output to the convolutional layer 344.

The convolutional layer 344 repeats a product-sum operation while sliding a linear transformation matrix called a filter or a kernel with respect to the input matrix by a predetermined stride amount, and transforms the input matrix into a linear transformation matrix. A matrix including a plurality of elements in which the sum of products is associated as an element value is generated. At this time, the convolutional layer 344 performs padding (for example, zero padding) around the input matrix to interpolate elements of arbitrary values, and inputs the matrix input to the convolutional layer 344 to the input layer 342. The low frequency component image IMG_A may be converted into a matrix having the same number of rows and columns. Then, the convolutional layer 344 outputs the generated matrix to the activation layer 346.

The activation layer 346 performs activation function calculation processing on each element of the matrix input from the convolutional layer 344, and outputs the calculated matrix to the subsequent layer. The activation function of the activation layer 346 may be, for example, a ReLU function.

The output layer 348 outputs the matrix processed by the convolutional layer 344 and the activation layer 346 in the previous stage.

As described above, by inputting the low-frequency component image IMG_A to the CNN 340 in which the convolutional layer 344 and the activation layer 346 are configured in multiple layers, the low-frequency component image in which the Tx unevenness is removed or reduced from the low-frequency component image IMG_A (hereinafter , Corrected low-frequency component image IMG_A#) can be generated. The corrected low-frequency component image IMG_A# is an example of the “second corrected image”.

The correction function 214 acquires the matrix output by the output layer 348 of the CNN 340 as the corrected low frequency component image IMG_A#. Then, the correction function 214 adds the high-frequency component image IMG_B output by the third layer 330 and the corrected low-frequency component image IMG_A# output by the CNN 340 from the medical image IMG_IN input to the correction model MDL. A corrected image IMG_GEN in which Tx unevenness and Rx unevenness are removed or reduced is generated. The corrected image IMG_GEN is an example of “corrected image”.

[Processing Flow of Medical Image Processing Device During Training]
Hereinafter, a flow of a series of processes of the processing circuit 210 on the medical image processing apparatus 200 side according to the embodiment will be described with reference to a flowchart. FIG. 7 is a flowchart showing a flow of a series of processing at the time of training of the processing circuit 210 on the medical image processing apparatus 200 side according to the embodiment. Training is learning the correction model MDL. The process of this flowchart may be repeated, for example, in a predetermined cycle. Note that, when the processor that realizes the processing circuit 210 of the medical image processing apparatus 200 is configured by combining a plurality of independent circuits, that is, when the processing circuit 210 of the medical image processing apparatus 200 is realized by a plurality of processors. A part or all of the processing of this flowchart may be processed in parallel by a plurality of processors.

First, the learning function 218 selects one learning image IMG_LR from the plurality of learning images IMG_LR included in the teacher information 234, and inputs the selected learning image IMG_LR to the correction model MDL (step S200). The learning image IMG_LR is an example of “first medical image”.

Next, the learning function 218 smoothes the corrected low-frequency component image IMG_A# output by the CNN 340 of the correction model MDL into which the learning image IMG_LR is input and the teacher image IMG_TR associated with the learning image IMG_LR as a teacher label. The difference from the image (hereinafter referred to as the low frequency component teacher image IMG_TR#) is derived (step S202). The teacher image IMG_TR is an example of the “second medical image”.

Next, the learning function 218 reduces the difference between the corrected low-frequency component image IMG_A# and the low-frequency component teacher image IMG_TR# so that the element values of the linear transformation matrix of the convolutional layer 344 and the activation layer 346 are reduced. Various parameters such as the coefficient of the activation function and the bias component are determined by learning using the stochastic gradient descent method or the like (step S204).

Next, the learning function 218 selects all the learning images IMG_LR included in the teacher information 234 to determine whether or not the correction model MDL has been learned, and selects all the learning images IMG_LR to learn the correction model MDL. When it is determined that the learning image IMG_LR has not been selected, the processing returns to S200, and the learning image IMG_LR different from the previously selected learning image IMG_LR is selected.

When the learning function 218 determines that all the learning images IMG_LR have been selected and the correction model MDL has been learned, the processing of this flowchart ends.

FIG. 8 is a diagram showing an example of the configuration of the correction model MDL during training according to the embodiment. As in the illustrated example, during training, the correction model MDL includes a fourth layer 360 and a fifth layer 370 in addition to the above-described first layer 310, second layer 320, third layer 330, and CNN340. including.

At the time of training, the first layer 310 corrects the brightness of the input learning image IMG_LR using the sensitivity map 236.

The second layer 320 smoothes the learning image IMG_LR whose brightness has been corrected by the first layer 310 to generate a low frequency component image IMG_A.

The third layer 330 sharpens the learning image IMG_LR whose brightness has been corrected by the first layer 310 to generate a high frequency component image IMG_B.

The CNN 340 removes or reduces the Tx unevenness from the low frequency component image IMG_A generated by the second layer 320 to generate a corrected low frequency component image IMG_A#.

The teacher image IMG_TR associated with the learning image IMG_LR is input to the fourth layer 360. When the teacher image IMG_TR is input, the fourth layer 360 corrects the brightness of the input teacher image IMG_TR using the sensitivity map 236.

The fifth layer 370 smoothes the teacher image IMG_TR whose brightness has been corrected by the fourth layer 360, and generates a low frequency component teacher image IMG_TR# by removing high frequency components from the teacher image IMG_TR. For example, the fifth layer 370 may be an LPF, similar to the second layer 320.

The learning function 218 obtains a difference between the low-frequency component teacher image IMG_TR# generated by the fifth layer 370 and the corrected low-frequency component image IMG_A# generated by the CNN 340, and stochastically reduces the difference. The parameters of the CNN 340 are learned by using the error back propagation method based on the gradient descent method or the like.

According to the embodiment described above, the processing circuit 210 of the medical image processing apparatus 200 acquires the medical image generated by transmitting the transmission RF from the transmission coil 106 toward the subject OB. When an image is input, the processing circuit 210 outputs to the correction model MDL that is learned to output a correction image in which the image unevenness caused by the inhomogeneity of the high frequency magnetic field is removed or reduced from the input image. By inputting the acquired medical image, a corrected image in which image unevenness is removed or reduced from the acquired medical image is generated. This can improve the image quality of the medical image.

Further, according to the above-described embodiment, by using the low-frequency component teacher image IMG_TR# obtained by removing the high-frequency component from the teacher image IMG_TR, the correction model MDL is learned with high accuracy without depending on the resolution of the MRI apparatus 100. can do. Further, by using the low-frequency component teacher image IMG_TR# obtained by removing the high-frequency component from the teacher image IMG_TR, the medical images to be used as the teacher image and the learning image IMG_LR input to the CNN 340 are scanned for different objects OB. Even with the obtained medical image, the correction model MDL can be learned with high accuracy.

Generally, when the objects OB are different, there are individual differences in the structure of the site even if the scan target is the same site. Therefore, the unevenness that can be included in the medical image also changes according to the individual difference. Therefore, the teacher image IMG_TR and the learning image IMG_LR input to the correction model MDL have the same subject OB and the same site. Is preferred. However, when both the teacher image IMG_TR and the learning image IMG_LR input to the correction model MDL are obtained from the same subject OB, the same subject OB is scanned many times in order to prepare many teacher images IMG_TR. Therefore, the burden on the subject OB is likely to increase. Further, since the same medical image of the subject OB is used for learning the correction model MDL, the generalization property of the correction model MDL is likely to decrease.

On the other hand, in the present embodiment, even if the teacher image IMG_TR and the learning image IMG_LR input to the CNN 340 are medical images obtained by scanning different subjects OB, smoothing is performed on both images. Since the high-frequency component relating to a detailed structure such as an edge is removed by performing the conversion processing, the correction model MDL can be learned with high accuracy without considering the individual difference of the object OB, and the correction model MDL The generalization can be improved.

(Modification of the embodiment)
Hereinafter, modified examples of the embodiment will be described. In the above-described embodiment, the MRI apparatus 100 and the medical image processing apparatus 200 are described as different apparatuses, but the present invention is not limited to this. For example, the medical image processing apparatus 200 may be realized by a function of the console device 120 of the MRI apparatus 100. That is, the medical image processing apparatus 200 may be a virtual machine virtually realized by the console device 120 of the MRI apparatus 100.

FIG. 9 is a diagram showing another example of the MRI apparatus 100 according to the embodiment. As shown in FIG. 9, the processing circuit 130 of the console device 120 has a correction function 214 and a learning function 218 in addition to the acquisition function 132, the generation function 134, the communication control function 136, and the display control function 138 described above. You may perform and.

In addition, the memory 150 of the console device 120 may store the correction model information 232, the teacher information 234, and the sensitivity map 236.

With such a configuration, the image quality of the medical image can be improved by the MRI apparatus 100 alone.

Moreover, in the above-described embodiment, the correction model MDL is learned based on the low-frequency component teacher image IMG_TR# obtained by removing the high-frequency component from the teacher image IMG_TR, but the present invention is not limited to this. For example, the learning function 218 is based on the sequence parameter, the statistical value indicating the degree of unevenness included in the medical image, the B1 map, and the like in addition to the low-frequency component teacher image IMG_TR# obtained by removing the high-frequency component from the teacher image IMG_TR. , The correction model MDL may be learned.

The sequence parameters are parameters that are dependent on the transmission RF, such as the type of sequence in the main scan, the type of the transmission coil 106, the calibration value of the transmission RF, and the adjustment values such as the phase and amplitude of the transmission RF in B1 shimming. ..

The statistical value indicating the degree of unevenness includes, for example, standard deviation, RMS value, PP value and the like.

The B1 map is, for example, a sensitivity map based on the signal intensity of the magnetic resonance signal obtained by a sequence in which the flip angle of the transmission RF is α degrees (for example, 45 degrees) and the flip angle of the transmission RF is 2α degrees (for example, 90 degrees). The sensitivity map based on the signal intensity of the magnetic resonance signal obtained by the sequence (1) is obtained by substituting it into a certain theoretical formula.

As described above, in addition to the teacher image IMG_TR, the correction model MDL is learned using the sequence parameter, the statistical value indicating the degree of unevenness, the B1 map, and the like, so that the Tx unevenness that may be included in the low-frequency component image is accurately measured. Can be removed. In particular, when the correction model MDL is learned using the B1 map, if there is a region where the signal intensity is locally small or large on the input image IMG_IN at runtime, whether the region is a change in brightness due to a lesion or not. , Or the change in brightness due to Tx unevenness can be discriminated by the correction model MDL.

For example, the learning function 218, for a certain area, when the B1 value is significantly larger than the reference value (for example, 1.0) on the B1 map and the brightness is high on the input image IMG_IN, or on the B1 map, When the B1 value is significantly smaller than the reference value and the brightness is low on the input image IMG_IN, the correction model MDL is learned so as to identify that region as a region in which unevenness has occurred.

On the other hand, when the B1 value is almost the same as the reference value on the B1 map for a certain area and the brightness is high or low on the input image IMG_IN, the learning function 218 determines that the area is a region where a lesion exists. The correction model MDL is learned so that it is identified as being present. By learning the correction model MDL using the B1 map in this way, it is possible to appropriately correct the brightness of a local area having different brightness as compared with other areas as Tx unevenness, and thus it is likely that a lesion exists. Areas with high values can be left on the image as they are without being removed as Tx unevenness.

The embodiment described above can be expressed as follows.
Storage for storing programs,
And a processor,
The processor, by executing the program,
Acquiring a medical image obtained based on a magnetic resonance signal obtained by exciting the subject with a high-frequency magnetic field transmitted by the RF coil,
When an image is input, from the input image, with respect to the model learned to output a corrected image in which the influence of the inhomogeneity of the high frequency magnetic field is reduced, by inputting the acquired medical image, Generating a corrected image in which the influence of the non-uniformity is reduced from the acquired medical image,
A medical image processing apparatus configured as described above.

According to at least one embodiment described above, the processing circuit 210 of the medical image processing apparatus 200 acquires the medical image generated by transmitting the transmission RF from the transmission coil 106 toward the subject OB, When an image is input, the acquired medical image is input to the correction model MDL learned to output a correction image in which image unevenness caused by the inhomogeneity of the high frequency magnetic field is removed or reduced from the input image. By inputting, a corrected image in which image unevenness is removed or reduced is generated from the acquired medical image, so that the image quality of the medical image can be improved.

While some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. The embodiments and their modifications are included in the scope of the invention and the scope thereof, as well as in the invention described in the claims and the scope of equivalents thereof.

DESCRIPTION OF SYMBOLS 1... Medical image processing system, 100... MRI apparatus, 101... Static magnetic field magnet, 102... Gradient magnetic field coil, 104... Bed, 105... Bed control circuit, 106... Transmission coil, 107... Transmission circuit, 108... Reception coil, 109 ...Reception circuit, 110... Sequence control circuit, 120... Console device, 200... Medical image processing device, 202... Communication interface, 204... Input interface, 206... Display, 210... Processing circuit, 212... Acquisition function, 214... Correction function 216... Output control function, 218... Learning function, 230... Memory

Claims (7)

An acquisition unit for acquiring a medical image obtained based on a magnetic resonance signal obtained by exciting the subject with a high-frequency magnetic field transmitted by the RF coil;
When an image is input, the medical image acquired by the acquisition unit is acquired from the input image with respect to a model learned to output a corrected image in which the influence of the inhomogeneity of the high frequency magnetic field is reduced. By inputting, a generation unit that generates a corrected image in which the influence of the nonuniformity is reduced from the medical image acquired by the acquisition unit,
A medical image processing apparatus comprising: The image input to the model is a medical image whose brightness is corrected using the reception sensitivity distribution of the magnetic resonance signal in the RF coil,
The medical image processing apparatus according to claim 1. The generator is
From the medical image acquired by the acquisition unit, a high frequency component image in which a high frequency component is extracted, and a low frequency component image in which a low frequency component is extracted are generated,
By inputting the low frequency component image to the model, a second corrected image in which the influence of the non-uniformity is reduced is generated from the low frequency component image,
Generate a corrected image in which the influence of the non-uniformity is reduced from the medical image acquired by the acquisition unit based on the second corrected image and the high-frequency component image.
The medical image processing apparatus according to claim 1. The low-frequency component image of the low-frequency component of the second medical image generated when the high-frequency magnetic field is uniform is associated with the first medical image generated when the high-frequency magnetic field is non-uniform. Further comprising a learning unit for learning the model based on the obtained information,
The medical image processing apparatus according to claim 3. A first generation unit that generates a medical image based on a magnetic resonance signal obtained by exciting a subject with a high-frequency magnetic field transmitted by an RF coil;
When an image is input, the medical image generated by the first generation unit is applied to the model learned from the input image to output a corrected image in which the influence of the inhomogeneity of the high frequency magnetic field is reduced. A second generation unit for generating a corrected image in which the influence of the non-uniformity is reduced from the medical image generated by the first generation unit by inputting an image;
And a magnetic resonance imaging apparatus. Computer
Acquiring a medical image obtained based on a magnetic resonance signal obtained by exciting the subject with a high-frequency magnetic field transmitted by the RF coil,
When an image is input, from the input image, with respect to the model learned to output a corrected image in which the influence of the inhomogeneity of the high frequency magnetic field is reduced, by inputting the acquired medical image, Generating a corrected image in which the influence of the non-uniformity is reduced from the acquired medical image,
Medical image processing method. On the computer,
A process of acquiring a medical image obtained based on a magnetic resonance signal obtained by exciting a subject with a high-frequency magnetic field transmitted by an RF coil;
When an image is input, from the input image, with respect to the model learned to output a corrected image in which the influence of the inhomogeneity of the high frequency magnetic field is reduced, by inputting the acquired medical image, A process of generating a corrected image in which the influence of the non-uniformity is reduced from the acquired medical image;
A program to execute.