KR102090690B1 - Apparatus and method for selecting imaging protocol of magnetic resonance imaging by using artificial neural network, and computer-readable recording medium storing related program - Google Patents

Abstract

The present invention relates to a computer readable recording medium in which an imaging protocol selection apparatus and method and program for magnetic resonance imaging using an artificial neural network are recorded. According to the present invention, an artificial neural network is trained using learning data, and when the learning is completed, an imaging protocol suitable for a patient is determined using the artificial neural network. Thereafter, a magnetic resonance image is generated according to the determined imaging protocol for the patient.

Description

Apparatus and method for selecting imaging protocol of magnetic resonance imaging by using artificial neural network, and computer-readable recording medium storing related program}

The present invention relates to magnetic resonance imaging technology, and more specifically, an apparatus and method for selecting an imaging protocol for magnetic resonance imaging based on patient information using an artificial neural network, and a program performing the method are recorded Computer readable recording media.

Magnetic Resonance Imaging (MRI) is one of the imaging technologies and uses the nuclear magnetic resonance principle. When a high frequency is generated while a human body is located in a magnetic resonance imaging apparatus that generates a magnetic field, the hydrogen atom nucleus of the human body is resonated. The magnetic resonance image is obtained by measuring the difference between the signals and then reconstructing the image through a computer.

Magnetic resonance imaging (MRI) is characterized by being harmless to the human body, unlike X-ray computed tomography (CT), which is harmful to the human body using X-rays. In addition, CT is primarily a cross-sectional image, while MRI is free in orientation. Using MRI, you can measure the oxygen content in your blood, and through this you can get information about blood flow in your brain.

The MRI device is a diagnostic device that can not only clarify the location of a lesion (tumor, etc.) in the body, but also reveal whether it is malignant, such as cancer. In these MRI devices, it is most important to determine the optimal MRI protocol for patient-specific diagnosis and treatment. In general, the MRI protocol is determined according to the MRI imaging reason, clinical information, patient information, referral information, and the like.

However, the decision of the MRI protocol requires a lot of time and puts a considerable burden on the radiology department.

An object of the present invention is to determine an imaging protocol of a magnetic resonance image suitable for a patient through an artificial neural network, and to read a device and method capable of photographing a magnetic resonance image according to the determined imaging protocol, and a computer in which a program performing the method is recorded. To provide a possible recording medium.

The apparatus for selecting an imaging protocol of a magnetic resonance image according to the present invention may include an artificial neural network and a learning module. The artificial neural network may be composed of a plurality of layers including an input layer, a convolutional layer, a pooling layer, and an output layer, each of the plurality of layers including a plurality of operations, and a result of any one operation of the plurality of layers The weight is applied to the input to the next layer. The learning module inputs learning data including a plurality of words describing patient information and an expected value representing an image protocol suitable for the patient into the artificial neural network, and outputs through a plurality of calculations of the artificial neural network according to the input. The weight may be corrected according to a backpropagation algorithm so that the difference between the output value and the expected value is minimized by comparing the output value with the expected value.

The device may further include a selection module. The selection module inputs discrimination data including a plurality of words describing basic information of a test object to the artificial neural network, and performs an imaging protocol of the test object according to the output of the artificial neural network through a plurality of calculations of the artificial neural network. You can choose.

The device may further include a scan unit and a control unit. The scan unit may form a magnetic field around an inspection object through a magnet and a coil, and acquire a magnetic resonance signal generated from the inspection object due to the magnetic field. The control unit may generate a magnetic resonance image by acquiring a magnetic resonance signal generated in a region corresponding to the imaging protocol through the scan unit due to the magnetic field.

The convolutional layer may include a plurality of first feature maps generated by performing a convolution operation using a first filter, which is a matrix in which each element is a weighted matrix, for the input layer.

The pooling layer may include a second feature map generated by performing a pooling operation using a second filter of the plurality of first feature map sizes on each of the plurality of first feature maps.

The output layer may include the output value calculated according to a softmax operation on an element of the matrix of the second feature map.

In addition, a method for selecting an imaging protocol of a magnetic resonance image according to the present invention comprises: a learning module inputting learning data including a plurality of words describing patient information and an expected value representing an imaging protocol suitable for the patient into an artificial neural network; And, the artificial neural network calculating an output value through a plurality of operations to which a weight is applied based on the learning data, and the weighting according to a backpropagation algorithm so that the learning module minimizes the difference between the output value and the expected value. It may be configured to include the step of modifying.

The method includes inputting, by the selection module, discrimination data including a plurality of words describing basic information of a test object into the artificial neural network, and performing a plurality of operations in which the artificial neural network is weighted based on the discrimination data. The method may further include outputting an output value through the selection module, and selecting an image protocol of the inspection target according to the output of the artificial neural network through a plurality of calculations of the artificial neural network.

The method includes the steps of forming a magnetic field around the inspection target through a magnet and a coil, and the control unit acquires a magnetic resonance signal generated in a region corresponding to the imaging protocol due to the magnetic field through the scanning unit to obtain a magnetic resonance image. It may further include the step of generating.

In the calculating of the output value, the artificial neural network performs a convolution operation using a first filter that is a matrix in which each element is a weighted matrix with respect to the input layer of the artificial neural network to obtain a convolution layer including a plurality of first feature maps. And generating.

In the calculating of the output value, after the step of generating the convolution layer, the artificial neural network performs a pooling operation using a second filter of the plurality of first feature map sizes on each of the plurality of first feature maps. The method may further include generating a pooling layer including the feature map.

The calculating of the output value may further include generating an output layer having an output value according to a softmax operation on elements of the matrix of the second feature map after the artificial neural network generates the pooling layer.

Meanwhile, the present invention can provide a computer-readable recording medium in which a program for performing the method is recorded.

According to the present invention, simply by inputting patient information, it is possible to quickly and automatically determine an appropriate imaging protocol for a patient through an artificial neural network, thereby accurately diagnosing and treating a patient. That is, the present invention can minimize the time required for the determination of the MRI protocol and reduce the workload of the radiology department.

1 is a schematic diagram showing a magnetic resonance imaging apparatus according to an embodiment of the present invention.
2 is a block diagram showing a detailed configuration of a protocol selection unit including an artificial neural network according to an embodiment of the present invention.
3 is a view for explaining the configuration of an artificial neural network according to an embodiment of the present invention.
4 to 6 are diagrams for explaining an operation including the weight of the artificial neural network shown in FIG. 3.
7 is a flowchart illustrating an artificial neural network learning method according to an embodiment of the present invention.
8 is a flowchart illustrating an imaging protocol selection and a magnetic resonance imaging method of an MR image using an artificial neural network according to an embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention may be implemented in various forms and should not be interpreted as being limited to the embodiments disclosed herein. The disclosed embodiments are provided to sufficiently convey the scope of the present invention to those of ordinary skill in the art. The principles and features of the invention may be applied in a wide variety of embodiments without departing from the scope of the invention.

In addition, in describing the embodiments, for the matters well known in the technical field to which the present invention pertains or not directly related to the present invention, descriptions may be omitted so as not to obscure and clearly convey the core of the present invention. In the accompanying drawings, some components are exaggerated, omitted, or schematically illustrated, and the size of each component does not entirely reflect the actual size. The same reference numbers are assigned to the same or corresponding elements throughout the accompanying drawings.

All terms used herein, including technical terms and scientific terms, have the same meaning as understood by a person skilled in the art to which the present invention pertains, unless otherwise defined. On the other hand, even if described in the singular form, unless the context clearly indicates otherwise, the plural form is included.

First, a magnetic resonance imaging (MRI) device according to an embodiment of the present invention will be described.

1 is a schematic diagram showing a magnetic resonance imaging apparatus according to an embodiment of the present invention. Referring to FIG. 1, a magnetic resonance imaging device 10 (hereinafter referred to as an MRI device) includes a scan unit 110, an input unit 130, a display unit 140, a control unit 150, and a protocol selection unit 160.

The scan unit 110 corresponds to the main body of the MRI device 10. Basically, the scanning unit 110 forms a magnetic field around an inspection object (1: for example, a patient's human body, hereinafter referred to as an inspection object or an imaging object) through a magnet and a coil, and the magnetic field from the inspection object 1 Generate a magnetic resonance signal. Then, the control unit 150 may receive a magnetic resonance signal and generate a magnetic resonance image using the received magnetic resonance signal.

The scan unit 110 includes a plurality of coils 111 composed of a main magnetic field coil, a gradient coil, and an RF coil. Each of the plurality of coils 111 has a cylindrical shape as a whole, and is coaxially arranged in the cylindrical bore 112.

The inspection object 1 is moved into the bore 112 on the cradle 113, that is, into the plurality of coils 111. At this time, the main magnetic field coil generates a static magnetic field in the bore 112. The direction of the static magnetic field is generally parallel to the direction of the body axis of the object 1 to be photographed, and thus a horizontal magnetic field. The main magnetic field coil is usually constructed using a superconducting magnet. However, the main magnetic field coil may be configured using a resistive magnet or the like in addition to the superconducting magnet.

The gradient coil generates three magnetic field gradients that are used to cause the intensity of the static magnetic field generated by the main magnetic field coil to undergo a gradient along three axes orthogonal to each other, that is, a slice axis, a phase axis, and a frequency axis. To create a magnetic field gradient, the gradient coil consists of three. Under the control of the controller 150, a gradient coil is driven to generate a magnetic field gradient. The magnetic field gradient whose direction corresponds to the direction of the slice axis is called a slicing magnetic field gradient, and the magnetic field gradient whose direction corresponds to the direction of the phase axis is called a phase encoding magnetic field gradient, and the magnetic field gradient whose direction corresponds to the direction of the frequency axis is read from the magnetic field gradient ( Alternatively, it may be referred to as a frequency encoding magnetic field gradient). If the coordinate axes in the Cartesian coordinate system defined in three-dimensional space are related to mutually orthogonal axes in the space of a static magnetic field, and it is the X, Y, and Z axes, any of the X, Y, and Z axes can be regarded as a slice axis. have. In this embodiment, the slice axis can be aligned with the body axis of the subject 1, and is regarded as the Z axis. Of the other two axes, one is the phase axis, and the other is the frequency axis. In addition, the slice axis, the phase axis, and the frequency axis can be inclined with respect to the X, Y, and Z axes while maintaining orthogonality with each other.

Under the control of the control unit 150, the RF coil applies an RF pulse. The RF coil forms a high-frequency magnetic field that is used to excite the spin in the imaging object 1 in the static magnetic field space. Forming a high-frequency magnetic field is called transmission of an RF excitation signal, and an RF excitation signal is called an RF pulse. The electromagnetic wave generated by the excited spin, that is, the magnetic resonance signal is received by the RF coil and transmitted to the control unit 150.

The input unit 130 receives a user's key manipulation for controlling various functions and operations of the MRI device 10, generates an input signal, and transmits the input signal to the control unit 150 or the protocol selection unit 160. The input unit 130 may be a keyboard, mouse, or the like. The input unit 130 may include at least one of a power key for power on / off, a character key, a number key, and a direction key. When the display unit 140 is implemented as a touch screen, the function of the input unit 130 may be performed in the display unit 140. In addition, when all functions can be performed only by the display unit 140, the input unit 130 may be omitted.

The display unit 140 may receive data for screen display from the control unit 150, for example, a magnetic resonance image and display the received data on a screen. In addition, the display 140 may visually provide a menu, data, function setting information, and various other information of the MRI device 10 to the user. When formed as a touch screen, the display unit 140 may perform some or all of the functions of the input unit 130 instead. The display unit 140 may be formed of a liquid crystal display (LCD), organic light emitting diodes (OLED), active matrix organic light emitting diodes (AMOLED), or the like.

The protocol selector 160 receives learning data including basic information from the input unit 130 and trains the artificial neural network. The basic information may include the patient's personal information, clinical information, and the reason for the imaging. In an embodiment, the basic information includes the referral department (e.g. orthopedic surgeon), patient gender (e.g. man), patient age (e.g. 56 years old), site (e.g. hip), contrast medium (e.g. contrast agent use), etc. It can contain. In addition, the protocol selector 160 receives discrimination data including basic information, and determines an imaging protocol of a magnetic resonance image suitable for a patient by using the learned artificial neural network. Here, the imaging protocol includes a routine protocol and a tumor protocol. The determined video protocol is provided to the control unit 150. On the other hand, in another embodiment, the video protocol may further include at least one of a stroke protocol, an inspiration protocol, a trauma protocol, and other protocols.

The control unit 150 may control the overall operation of the MRI apparatus 10 and the signal flow between internal blocks, and perform a data processing function for processing data. The control unit 150 may be a central processing unit (CPU), an application processor (AP), or a graphic processing unit (GPU).

The control unit 150 generates a magnetic resonance image from the magnetic resonance signal received from the RF coil. The magnetic resonance images are sagittal T1-weighted image, sagittal T2-weighted image, axis T1-weighted image, and axis T2-weighted image (Axial T2- weighted image). To this end, the controller 150 may convert the magnetic resonance signal into a magnetic resonance image in the form of digital data. The obtained magnetic resonance signal may be a signal defined in a frequency domain, such as Fourier space. A magnetic field gradient whose direction corresponds to the phase axis direction and the frequency axis direction is applied to encode the distribution of the source of the magnetic resonance signal along the two axes. For example, when the Fourier space is adopted as the frequency domain, the magnetic resonance signal is provided as a signal defined in the two-dimensional Fourier space. The two-dimensional Fourier space can be referred to as k-space. The phase-encoded magnetic field gradient and frequency-encoded magnetic field gradient determine the location of the sampled signal in a two-dimensional Fourier space.

In particular, the control unit 150 sets the parameters of the magnetic resonance imaging apparatus 10 according to the imaging protocol, and controls the magnetic resonance imaging apparatus 10 to photograph a magnetic resonance image of a portion corresponding to the imaging protocol. In the case of a knee magnetic resonance image, the controller 150 sets the routine protocol to 3 mm slice thickness and 0.3 mm gap. In addition, the field of view (FOV) is set to 14 or 16 cm. On the other hand, in the case of a tumor protocol, the controller 150 sets slice thickness 3 to 4 mm / gap 0.3 to 1 mm according to the tumor size. In addition, in the case of a routine protocol, the control unit 150 includes a fat suppression T2-emphasized axial image, a T1 emphasis axial image, a T2-emphasis sagittal image, and a fat-suppression PD-emphasis sagittal. The image is controlled so that the coronal image emphasizing the suppressed T2 is photographed, and, if there is a contrast agent, it is controlled to add and photograph the fat suppressed T1 highlighted sagittal image before and after the contrast agent. On the other hand, in the case of a tumor protocol, the controller 150 is a contrast-enhanced image, a T1-emphasized axial image, a T2-emphasized axial image, a fat suppression, a T2-emphasized axial image, a fat suppression T2 emphasis Coronal images, diffuse weighted images are taken, after contrast enhancement, perfusion images, fat suppression T1-emphasized axial images, fat suppression T1-emphasized coronal images, fat Control so that the suppressed T1-emphasized sagittal image is taken.

Next, the detailed configuration of the protocol selector 160 including the artificial neural network according to an embodiment of the present invention will be described.

2 is a block diagram showing a detailed configuration of a protocol selection unit including an artificial neural network according to an embodiment of the present invention. 3 is a view for explaining the configuration of an artificial neural network according to an embodiment of the present invention. 4 to 6 are diagrams for explaining an operation including the weight of the artificial neural network shown in FIG. 3.

Referring first to FIG. 2, the protocol selection unit 160 according to an embodiment of the present invention includes an artificial neural network 200, a learning module 300, and a selection module 400.

As illustrated in FIG. 3, the artificial neural network 200 is composed of a plurality of layers including an input layer 210, a convolution layer 220, a pooling layer 230, and an output layer 240. Any one layer is calculated through a predetermined operation using the previous layer as an input. In addition, each of the plurality of operations is weighted, the weight has an initial value, and the final value is derived through learning.

The input layer 210 has a predetermined size (eg, n × k) and is represented by a matrix in which each element is a binary value. Although a plurality of feature maps are illustrated in the drawings, they are all identical feature maps having the same value. In the matrix of the input layer 210, binary values corresponding to a plurality of words representing basic information are input. Therefore, if the size of the input layer 210 is n × k, n words can be input, and the length of the binary value of each word is smaller than k. For example, n words may be input, including a referral department, patient gender, patient age, region, contrast, and the like, each of which is less than k in length. It is converted to a binary value of and input to the input layer 210.

The convolution layer 220 includes a plurality of first feature maps. The plurality of first feature maps of the convolution layer 220 are all represented by a matrix having a predetermined size (eg, m × 1). The plurality of first feature maps of the convolution layer 220 are generated by performing a convolution operation using the plurality of first filters F1 on the input layer 210. The first filter F1 used in the convolution operation is a matrix in which each element has a weight w. The first filter F1 is for extracting a plurality of different features. Therefore, each of the plurality of first filters F1 may have a different size (matrix size), and each element of the plurality of first filters F1 may have a different value. Therefore, each of the plurality of first feature maps is generated through a convolution operation in which different first filters F1 are applied to the input layer 210.

4 shows an example of a convolution operation. The illustrated example represents a convolution operation that derives any one first feature map by applying any one first filter F1 to the input layer 210. Here, it is assumed that the matrix of the input layer 210 has a size of 5 × 5, and the first filter F1 has a size of 4 × 3. In addition, it is assumed that the value of each element of the illustrated first filter F1, that is, the weight w is an initial value. Accordingly, the first feature map of the 6 × 1 size convolution layer 220 is output through the convolution operation by applying the 4 × 3 first filter F1 to the 5 × 5 size input layer 210. do. The convolution operation is performed while elements and operations of the corresponding input layer 210 are performed while the first filter F1 is shifted by one row or one column, and is composed of a product of a corresponding element and a sum of the products. For example, elements (1, 1) of the first feature map of the convolution layer 220 correspond to elements in columns 1 to 3 of each of rows 1 to 4 of the input layer 210, as shown in Equation 1 below. Calculated by the product of the elements of columns 1 to 3 of each of the first to fourth rows of the first filter F1 and the sum of the products.

[Equation 1]

(1 × 1) + (0 × 0) + (0 × 1) +

(0 × 0) + (0 × 1) + (1 × 0) +

(0 × 1) + (1 × 1) + (0 × 0) +

(0 × 0) + (0 × 1) + (1 × 1) = 3

As another example, elements (1, 2) of the first feature map of the convolution layer 220 may be shifted by one row of the first filter F1, as shown in Equation 2 below, from 1 row to 1 of the input layer 210. It is calculated by the product of the elements of columns 1 to 3 of each of columns 1 to 4 of the first filter F1 corresponding to the elements of columns 2 to 4 of each of 4 rows and the sum of the products.

[Equation 2]

(0 × 1) + (0 × 0) + (1 × 1) +

(0 × 0) + (1 × 1) + (0 × 0) +

(1 × 1) + (0 × 1) + (1 × 0) +

(0 × 0) + (1 × 1) + (0 × 1) = 4

As such, a first feature map of any one of the convolution layer 220 may be generated through a convolution operation in which any one of the plurality of first filters F1 is applied to the matrix of the input layer 210. In addition, a plurality of first feature maps of the convolution layer 220 may be generated through a convolution operation by applying a plurality of first filters F1 having different sizes and weights to the matrix of the input layer 210. .

The pooling layer 230 includes a second feature map. The second feature map of the pooling layer 230 is represented by a matrix having a predetermined size (eg, L × 1). The second feature map of the pooling layer 230 is generated by performing a pooling or subsampling operation using a plurality of second filters F1 on each of the plurality of first feature maps of the convolution layer 220. The second filter F2 used for the pooling operation is a matrix in which each element has a weight w. The second filter F2 is for extracting the most dominant feature through downsampling. Accordingly, each of the plurality of second filters F1 may have a different size (matrix size), and elements of each of the plurality of second filters F1 may have different values. The second feature map is generated through a pooling operation in which different second filters F1 are applied to different first feature maps of the convolution layer 220.

5 shows an example of a pooling operation. In the illustrated example, a pooling operation is performed to derive any one element of the second feature map of the pooling layer 230 by applying any second filter F1 to any one feature map of the convolution layer 220. Shows. Here, it is assumed that the first feature map of the convolution layer 220 is 6 × 1 in size, and the second filter F2 is also 6 × 1 in size. In addition, it is assumed that the value of each element of the illustrated second filter F2, that is, the weight w is an initial value. Accordingly, any one of the second feature maps of the pooling layer 230 through the pooling operation by applying the 6 × 1 second filter F1 to the first feature map of the 6 × 1 size convolution layer 220 The element of is output.

In the pooling operation, the second filter F2 is shifted by the size of the filter F2, and the corresponding element and operation are performed. In particular, the pooling operation may be any one of operations for selecting a maximum value, an average value, a median value, and a norm value. In this embodiment, it is assumed that the maximum value is a pooling operation (max pooling).

For example, in the max pooling, a product of a corresponding element and a maximum value among the products are selected. For example, any one element of the second feature map of the pooling layer 230 may correspond to a second filter element corresponding to each column of each row 1 to 6 of the convolution layer 220 as shown in Equation 3 below. It is calculated by selecting the maximum value among the product of the elements in each column of F1).

[Equation 3]

3 × 0.21 = 0.63

4 × 0.05 = 0.20

2 × 0.10 = 0.20

2 × 0.12 = 0.24

3 × 0.15 = 0.45

5 × 0.01 = 0.05

Max {0.63, 0.20, 0.20, 0.24, 0.45, 0.05} = 0.63

As such, any one element of the second feature map of the pooling layer 230 through a pooling operation in which any one of the plurality of second filters F2 is applied to any one first feature map of the convolution layer 220 Can be calculated. In addition, the pooling layer 230 through a pooling operation in which the plurality of second filters F2 having different sizes and weights are applied to each of the plurality of first feature maps of the convolution layer 220 in the same manner as the illustrated method. All elements of the second feature map of can be calculated.

The output layer 240 includes a plurality of nodes corresponding to the video protocol. The plurality of nodes will be referred to as output nodes, and the values of the output nodes will be referred to as output values. In an embodiment of the present invention, the output layer 240 has two output nodes, each of the two output nodes corresponds to a routine and a tumor protocol. In the embodiment of the present invention, it is described that the output layer 240 is composed of only two output nodes, but is not limited thereto, and output nodes corresponding to more various video protocols may be further included. The output value of each output node of the output layer 240 represents the probability that the imaging protocol corresponding to the output node is suitable for the patient. Here, the patient means a patient who inputs basic information input to the input layer 210.

6, if the value of the output node corresponding to the routine protocol is 0.89 and the value of the output node corresponding to the tumor protocol is 0.11, the probability that the routine protocol is suitable for the patient is 89%, and the tumor protocol It indicates that the probability of being suitable for this patient is 11%. The value of each output node of the output layer 240 is calculated by performing a softmax operation in which a weight w is applied to each element of the second feature map of the pooling layer 230.

6 shows an example of a softmax operation. The example of FIG. 6 represents a softmax operation in which weight is applied to all elements of the second feature map of the pooling layer 230 to calculate the output value of the output node of the output layer 240 through the operation.

The softmax operation is linearly combined by applying weights w11 to w25 to all elements of the second feature map of the pooling layer 230 and then input to a nonlinear function. And the result of the operation of the nonlinear function becomes the output value of the output node.

For example, as shown in Equation 4 below, all elements of the second feature map of the pooling layer 230 are multiplied and multiplied by weights w11 to w25, and the summed values shown in Equation 5 below are nonlinear functions of sigmoid It is input to the sigmoid function, and the result of the operation of the nonlinear function becomes the output value of the output node corresponding to the routine protocol.

[Equation 4]

k = (0.24 × w11) + (0.49 × w12) + (0.72 × w13) + (0.29 × w14) + (0.38 × w15)

[Equation 5]

Here, f () represents a sigmoid function. However, the present invention describes the nonlinear function as using a sigmoid function, but is not limited thereto, and various nonlinear functions may be applied.

Referring back to FIG. 2, the learning module 300 trains the artificial neural network 300 using learning data. Learning data refers to data in a state in which an imaging protocol suitable for a patient is known. The training data includes a plurality of words describing basic information and an expected value representing an imaging protocol suitable for the patient. Here, the basic information includes a shooting reason, clinical information, personal information, and a referral department. For example, if it is determined that it is appropriate to photograph a magnetic resonance image of a specific patient according to the routine protocol, and when the patient is photographed according to the routine protocol, and is found to be the correct diagnosis, basic information about the patient may be used as learning data. have. In addition, the expected value may be set such that the output value corresponding to the routine protocol is higher than a predetermined value than the output value corresponding to the tumor protocol. For example, the expected value may set the routine protocol to (0.50 + A) and the tumor protocol to (0.50-A).

When the learning data is input to the learning module 300, the learning data is input to the artificial neural network 200. At this time, the learning module 300 converts each of a plurality of words describing basic information into a binary value and inputs it to the input layer 210.

When the training data is input, the artificial neural network 200 calculates an output value through a plurality of operations for applying weights based on the input training data, as described above. Then, the learning module 300 corrects the weight w applied to a plurality of operations of the artificial neural network 200 so that the difference between the output value and the expected value is minimized through a backpropagation.

The selection module 400 selects an image protocol suitable for the patient through the artificial neural network 300 sufficiently learned using the discrimination data. The discrimination data refers to data in which an imaging protocol suitable for a patient is unknown. The learning data includes basic information and expected values, while the discrimination data includes only basic information. Similarly, the basic information includes the reason for the imaging, clinical information, personal information, and referral department. That is, the discrimination data means basic information of a patient whose imaging protocol has not been determined.

When the discrimination data is input, the selection module 400 inputs the discrimination data to the artificial neural network 200. The selection module 400 converts each of a plurality of words describing basic information into a binary value and inputs it to the input layer 210 of the artificial neural network 200. When the discrimination data is input, the artificial neural network 200 calculates an output value through a plurality of calculations that apply weights based on the input discrimination data. Then, the selection module 400 selects an imaging protocol suitable for the patient according to the output value. For example, as illustrated in FIG. 6, when the routine protocol is 89% and the tumor protocol is 11%, the selection module 400 determines the routine protocol as an appropriate imaging protocol for the patient.

Next, a method of learning the artificial neural network 200 according to an embodiment of the present invention will be described.

7 is a flowchart illustrating an artificial neural network learning method according to an embodiment of the present invention.

Referring to FIG. 7, when learning data including a plurality of words describing basic information and an expected value representing an image protocol suitable for the patient are input, the learning module 300 transmits the learning data in step S110 to the artificial neural network 200. Enter in In this case, the learning module 300 may convert each of a plurality of words describing basic information into a binary value, and then input it into the input layer 210 of the artificial neural network 200.

When the learning data is input to the input layer 210, the artificial neural network 200 performs a plurality of operations including weights in step S120 to derive an output value. In more detail, the artificial neural network 200 performs a convolution operation using a first filter F1 having a plurality of weights (w) as an element on the input layer 210 to have a convolutional layer having a plurality of first feature maps (220). In addition, the artificial neural network 200 includes a second feature map by performing a pooling operation using a second filter F2 having a plurality of weights (w) as elements in the plurality of first feature maps of the convolution layer 220. To form a pooling layer 230. Then, the artificial neural network 200 configures the output layer 240 having a plurality of output nodes by performing a softmax operation in which a plurality of weights w are applied to the pooling layer 230, and the value of the output node becomes an output value.

Before learning is completed, the output value of the artificial neural network 200 is different from the expected value. Therefore, in step S130, the learning module 300 uses a weight (w) applied to a plurality of calculations of the artificial neural network 200 so that the difference between the output value and the input expected value of the artificial neural network 200 is minimized through a backpropagation algorithm. ). For example, the expected value may set the routine protocol to (0.50 + A) and the tumor protocol to (0.50-A). However, it is assumed that the output value is a routine protocol of 0.49 and a tumor protocol of 0.51. Then, the learning module 300 modifies the weight w applied to a plurality of calculations of the artificial neural network 200 so that the routine protocol is 0.50 + A.

Subsequently, a method of photographing a magnetic resonance image using the artificial neural network 200 learned through the above-described method will be described.

8 is a flowchart illustrating an imaging protocol selection and a magnetic resonance imaging method of an MR image using an artificial neural network according to an embodiment of the present invention.

Referring to FIG. 8, when learning data including a plurality of words describing basic information is input through the input unit 130, the selection module 400 inputs discrimination data to the artificial neural network 200 in step S210. At this time, the selection module 400 may convert each of a plurality of words describing basic information into a binary value, and input it to the input layer 210 of the artificial neural network 200.

The artificial neural network 200 derives an output value by performing a plurality of operations including the weight w learned in step S220 when discrimination data is input to the input layer 210. That is, the artificial neural network 200 performs a convolution operation using a first filter F1 having a plurality of weights (w) as an element on the input layer 210 to have a convolution layer having a plurality of first feature maps ( 220). In addition, the artificial neural network 200 includes a second feature map by performing a pooling operation using a second filter F2 having a plurality of weights (w) as elements in the plurality of first feature maps of the convolution layer 220. To form a pooling layer 230. In addition, the artificial neural network 200 configures the output layer 240 having a plurality of output nodes by performing a softmax operation by applying a plurality of weights w to the pooling layer 230, and the value of the output node becomes an output value. .

The selection module 400 determines the image protocol according to the output value of the artificial neural network 200 in step S230 and inputs it to the control unit 150. For example, as illustrated in FIG. 6, when the routine protocol is 89% and the tumor protocol is 11%, the selection module 400 determines the routine protocol as an appropriate imaging protocol for the patient.

Then, the control unit 150 sets the parameters according to the image protocol determined in step S240, and controls the scan unit 110 to obtain a magnetic resonance image for a region corresponding to the determined image protocol. That is, the controller 150 sets parameters according to the determined imaging protocol, forms a magnetic field around the inspection object through a magnet and a coil, and scans the magnetic resonance signal generated in a region corresponding to the imaging protocol due to the magnetic field. ) To generate a magnetic resonance image. In the knee MRI, for example, routine protocols are all set to 3 mm slice thickness and 0.3 mm gap. Also, the field of view (FOV) is fixed at 14 or 16 cm. In addition, a fat-suppressed T2-emphasized axial image, a T1-emphasized axial image, a T2-emphasized sagittal image, a fat-suppressed PD-emphasized sagittal image, and a fat-suppressed T2-emphasized coronal image are photographed. If there is a contrast agent, a fat suppression T1-emphasized sagittal image is taken before and after the contrast agent. On the other hand, in the case of a tumor protocol, a T1-weighted axial image, a T2-weighted axial image, a fat suppression, a T2-weighted axial image, a fat-suppressed T2-weighted coronal image as contrast enhancement images , Diffusion weighted image is taken, after contrast enhancement, perfusion image, fat suppression T1 emphasis axial image, fat suppression T1 emphasis coronal image, fat suppression T1 emphasis award ( sagittal) video. The parameter is adjusted to 3 ~ 4mm / gap 0.3 ~ 1mm according to the tumor size, and the imaging range (FOV) is adjusted.

Meanwhile, the method according to the embodiment of the present invention described above may be implemented in a program readable form through various computer means and recorded in a computer-readable recording medium. Here, the recording medium may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded on the recording medium may be specially designed and configured for the present invention or may be known and usable by those skilled in computer software. For example, the recording medium includes magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and magnetic-optical media such as floptical disks ( magneto-optical media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions may include high-level language wires that can be executed by a computer using an interpreter, as well as machine language wires such as those produced by a compiler. Such hardware devices may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

The embodiments of the present invention disclosed in the present specification and drawings are merely intended to easily describe the technical contents of the present invention and to provide specific examples to help understanding of the present invention, and are not intended to limit the scope of the present invention. It is obvious to those skilled in the art to which the present invention pertains that other modifications based on the technical spirit of the present invention can be implemented in addition to the embodiments disclosed herein.

10: magnetic resonance imaging device 110: scan unit
111: coil 112: bore
113: cradle 130: input
140: display unit 150: control unit
160: protocol selection unit 200: artificial neural network
210: input layer 220: convolution layer
230: pooling layer 240: output layer
300: learning module 400: selection module

Claims (13)

It is composed of a plurality of layers including an input layer, a convolutional layer, a pooling layer, and an output layer, each of the plurality of layers includes a plurality of operations, and a weight is applied to an operation result of any one of the plurality of layers Artificial neural network that is the input of the next layer; And
The learning data including a plurality of words describing patient information and an expected value representing an image protocol suitable for the patient are input to the artificial neural network, and the output value and the output value through a plurality of calculations of the artificial neural network according to the input. A learning module that compares the expected value and corrects the weight according to a back propagation algorithm so that the difference between the output value and the expected value is minimal.
An apparatus for selecting a video protocol of a magnetic resonance image. According to claim 1,
A selection module for inputting discrimination data including a plurality of words describing basic information of a test object into the artificial neural network and selecting an imaging protocol of the test object according to the output of the artificial neural network by a plurality of calculations of the artificial neural network
An apparatus for selecting a video protocol of a magnetic resonance image, further comprising a. According to claim 2,
A scanning unit that forms a magnetic field around the inspection object through a magnet and a coil, and acquires a magnetic resonance signal generated from the inspection object due to the magnetic field; And
And a control unit for generating a magnetic resonance image by acquiring a magnetic resonance signal generated at a region corresponding to the imaging protocol due to the magnetic field through the scan unit. According to claim 1,
The convolution layer includes a plurality of first feature maps generated by performing a convolution operation using a first filter, which is a matrix in which each element is a weighted matrix, for the input layer. Device. According to claim 4,
The pooling layer includes a second feature map generated by performing a pooling operation using a plurality of first feature map-sized second filters on each of the plurality of first feature maps. Protocol selection device. The method of claim 5,
And the output layer includes the output value calculated according to a softmax operation on an element of the matrix of the second feature map. Inputting, by the learning module, learning data including a plurality of words describing patient information and an expected value representing an image protocol suitable for the patient into an artificial neural network;
Calculating, by the artificial neural network, an output value through a plurality of calculations to which a weight is applied based on the learning data; And
Correcting the weight according to a backpropagation algorithm so that the learning module minimizes the difference between the output value and the expected value.
Method for selecting an imaging protocol of a magnetic resonance image comprising a. The method of claim 7,
Inputting, to the artificial neural network, discrimination data including a plurality of words describing the basic information of the examination object by the selection module;
The artificial neural network outputting an output value through a plurality of calculations to which a weight is applied based on the discrimination data; And
Selecting, by the selection module, an imaging protocol of the inspection target according to the output of the artificial neural network through a plurality of calculations of the artificial neural network.
A method of selecting a video protocol of a magnetic resonance image, further comprising a. The method of claim 8,
Forming a magnetic field around the inspection object through the magnet and coil; And
Generating a magnetic resonance image by a control unit obtaining a magnetic resonance signal generated in a region corresponding to the imaging protocol due to the magnetic field through the scan unit
A method of selecting a video protocol of a magnetic resonance image, further comprising a. The method of claim 7,
The step of calculating the output value is
And the artificial neural network generating a convolution layer including a plurality of first feature maps by performing a convolution operation using a first filter that is a matrix in which each element is a weighted matrix with respect to the input layer of the artificial neural network. A method of selecting an imaging protocol for a magnetic resonance image. The method of claim 10,
The step of calculating the output value is
After the step of generating the convolution layer, the artificial neural network performs a pooling operation using a second filter of the plurality of first feature map sizes on each of the plurality of first feature maps, and the pooling includes a second feature map. The method of claim 1, further comprising the step of generating a layer. The method of claim 11,
The step of calculating the output value is
After generating the pooling layer, the artificial neural network further includes generating an output layer having an output value according to a softmax operation on an element of the matrix of the second feature map. Way. A computer-readable recording medium in which a program for executing a method for selecting an imaging protocol of a magnetic resonance image according to any one of claims 7 to 12 is recorded on a computer.

Images