JP2022172417A - Generation method of learned model, deduction apparatus, medical apparatus and program - Google Patents

Abstract

To provide a technique for highly accurately distinguishing between an area where a contrast agent is taken and an area where the contrast agent is not taken when contrast imaging is executed.SOLUTION: There is provided a learned model generation method for generating a learned model 91a which executes deduction for distinguishing between an area where a contrast agent is taken and an area where the contrast agent is not taken in the chest of an imaging object person by receiving input of a non-contrast CT image including the chest of the imaging object person and a contrast CT image including the chest of the imaging object person. A computer generates a learned model 91a by causing a neural network 91 to execute learning using a plurality of non-contrast CT images CN1 to CNn including the chest of a person, contrast CT images CP1 to CPn in a first time phase including the chest of the person, contrast CT images CQ1 to CQn in a second time phase, contrast CT images CQ1 to CQn in a third time phase, and a plurality of pieces of correct answer data GP1 to GPn, GQ1 to GQn and GR1 to GRn.SELECTED DRAWING: Figure 8

Description

The present invention relates to a method of generating a trained model, an inference device that performs inference using the trained model, a medical device having the inference device, and a program for performing inference using the trained model.

2. Description of the Related Art An X-ray CT apparatus is known as a medical apparatus that noninvasively captures images of the inside of a subject's body. The X-ray CT apparatus is widely used in medical facilities such as hospitals because it can image a region to be imaged in a short period of time.

When imaging a subject using an X-ray CT apparatus, various CT images can be acquired according to clinical purposes by scanning the subject under various scanning conditions. A doctor such as a radiologist can interpret the acquired CT image and make a diagnosis based on the result of the interpretation.

JP 2016-87469 A

Single energy CT is known as an imaging method using a CT apparatus. Single-energy CT is a method of obtaining CT value information by imaging with one tube voltage (usually 120 kVp), but this method has artifacts ( beam hardening artifact) is likely to occur. Dual-energy CT technology has been researched and developed as a technology for coping with this problem (see Patent Document 1), and dual-energy CT is expected as an imaging method that overcomes the shortcomings of single-energy CT.

For example, when distinguishing between benign and malignant tumors by contrast imaging, it is important to quantitatively evaluate the contrast agent taken into the tissue. This method is suitable for Therefore, dual-energy CT can improve the accuracy of distinguishing between benign and malignant tumors. However, some CT systems cannot use dual energy CT. In order to deal with this problem, a DNN (Deep Neural Network) performs learning using dual-energy CT data to generate a learned DNN, and inputs single-energy CT data to the generated trained DNN. Research has been conducted to infer dual-energy information. However, with this method, there are cases where a portion of the imaged region appears brighter than other tissues even though the contrast agent is not incorporated. For example, if the imaged region includes food and calcified parts in the stomach, the food and calcified parts may be depicted brightly. Therefore, it may be difficult to discriminate between areas in which the contrast agent has been incorporated and areas in which the contrast agent has not been incorporated.

Therefore, there is a demand for a technique capable of distinguishing with high accuracy a region in which a contrast medium has been taken and a region in which no contrast medium has been taken, when contrast imaging is performed.

According to a first aspect of the present invention, a non-contrast-enhanced CT image including a predetermined portion of an imaging subject and a contrast-enhanced CT image including a predetermined portion of the imaging subject are input, whereby the predetermined A trained model generation method for generating a trained model that performs inference for discriminating between a region in which a contrast agent has been taken and a region in which no contrast agent has been taken in a part of
the computer
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

is a learned model generation method for generating the learned model by executing learning using

According to a second aspect of the present invention, by inputting a non-contrast CT image including a predetermined portion of a subject and a contrast-enhanced CT image including a predetermined portion of the subject, the predetermined An inference device that includes a trained model that performs inference for discriminating between a region in which a contrast agent is taken up and a region in which a contrast agent is not taken up in a part of
The learned model is a computer,
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

is an inference device generated by executing learning using

In a third aspect of the present invention, a non-contrast-enhanced CT image including a predetermined portion of a subject and a contrast-enhanced CT image including a predetermined portion of the subject are input to obtain the predetermined portion of the subject. A medical device that includes a trained model that performs inference to discriminate between contrast agent uptake regions and contrast agent non-uptake regions at a site of
The learned model is a computer,
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

is a medical device generated by performing learning using

A fourth aspect of the present invention provides one or more processors with:
A non-contrast-enhanced CT image including a predetermined part of an imaging subject and a contrast-enhanced CT image including a predetermined part of the imaging subject are input to a trained model, and the predetermined part of the imaging subject is input to the learned model. A program for executing an operation including executing inference for distinguishing between a region in which a contrast agent is taken and a region in which a contrast agent is not taken in,
The learned model is a computer,
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

It is a program generated by executing learning using

According to a fifth aspect of the present invention, a plurality of contrast-enhanced CT images obtained by contrast-enhanced imaging of a predetermined part of a subject person in a plurality of time phases are input, thereby obtaining the predetermined region of the subject person. A trained model generation method for generating a trained model that performs inference for discriminating between a region in which a contrast agent is taken and a region in which no contrast agent is taken in a part,
the computer
(1) a plurality of contrast-enhanced CT images obtained by contrast-enhanced imaging of a predetermined part of a human at the plurality of time phases; and (2) a plurality of correct data labeled on the plurality of contrast-enhanced CT images, , each correct data is a difference between a contrast-enhanced CT image captured in one of the plurality of time phases and a contrast-enhanced CT image captured in another time phase of the plurality of time phases. A plurality of correct data generated based on the difference image generated by

It is a trained model generation method that is generated by executing learning using

When a tumor is included in the imaging region, there is a large difference between the brightness of the tumor depicted in the non-contrast CT image and the brightness of the tumor depicted in the contrast-enhanced CT image C. On the other hand, when focusing on food and calcified parts in the stomach, the food and calcified parts visualized on non-contrast CT images are brighter than the food and calcified parts visualized on contrast-enhanced CT images. does not show much difference. Therefore, when the non-contrast CT image and the contrast CT image are subtracted, the difference value in the tumor region (the region in which the contrast agent is taken) becomes a relatively large value, but the difference value in the food and calcification region (the region where the contrast agent is taken in) is relatively large. The difference value of the non-captured area) is a relatively small value. For this reason, by subtracting the non-contrast CT image and the contrast CT image, regions such as food and calcified areas that are brightly rendered even though the contrast agent is not taken in and tumors in which the contrast agent is taken in are detected. Thus, a difference image can be obtained in which the difference from the brightly drawn area is emphasized.

In the present invention, such difference images are used to generate a trained model. Therefore, even if the non-contrast-enhanced CT image and the contrast-enhanced CT image include areas such as food and calcified areas that are brightly rendered even though the contrast agent is not incorporated, the non-contrast-enhanced CT image and the contrast-enhanced CT image is input into the trained model, it is possible to discriminate between regions in which the contrast agent is incorporated and regions in which the contrast agent is not incorporated.

1 is an explanatory diagram of an X-ray CT apparatus of this embodiment; FIG. 4 is an explanatory diagram of the gantry 2, the table 4, and the operation console 8; FIG. 3 is a diagram showing main functional blocks of a processing unit 84; FIG. FIG. 10 is a diagram showing a flow chart of a learning phase; FIG. 4 is an explanatory diagram of CT images used in the learning phase; FIG. 2 schematically shows an example of a non-contrast CT image CNi and contrast CT images CPi, CQi, and CRi; It is explanatory drawing of an example of the preparation method of correct data. It is explanatory drawing of step ST2. It is a figure which shows a contrast examination flow. FIG. 4 is a diagram showing an example of a reconstructed image; FIG. FIG. 4 is an explanatory diagram of an inference phase; FIG. 9 is a diagram showing functional blocks of a processing unit 84 in a second form; FIG. 10 is a diagram showing a contrast imaging examination flow in the second embodiment; It is explanatory drawing of step ST30.

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments for carrying out the invention will be described below, but the invention is not limited to the following embodiments.

FIG. 1 is an explanatory diagram of an example of the X-ray CT apparatus of the first embodiment.
As shown in FIG. 1, the X-ray CT apparatus 1 includes a gantry 2, a table 4, a camera 6, and an operation console 8.

A gantry 2 and a table 4 are installed in the scan room 100 . A camera 6 is installed on the ceiling 101 of the scan room 100 . An operation console 8 is installed in the operation room 200 .

Next, the gantry 2, table 4, and operation console 8 will be described with reference to FIG.

FIG. 2 is an explanatory diagram of the gantry 2, the table 4, and the operation console 8. FIG.
The gantry 2 has inner walls defining a bore 21 in which the patient 40 can move.

The gantry 2 also includes an X-ray tube 22, an aperture 23, a collimator 24, an X-ray detector 25, a data acquisition system 26, a rotating part 27, a high voltage power supply 28, an aperture drive device 29, It has a rotating part driving device 30, a GT (Gantry Table) control part 31, and the like.

The X-ray tube 22 , aperture 23 , collimator 24 , X-ray detector 25 and data acquisition section 26 are mounted on rotating section 27 .

X-ray tube 22 irradiates patient 40 with X-rays. The X-ray detector 25 detects X-rays emitted from the X-ray tube 22 . The X-ray detector 25 is provided on the opposite side of the bore 21 from the X-ray tube 22 .

Aperture 23 is arranged between X-ray tube 22 and bore 21 . The aperture 23 shapes the X-rays emitted from the X-ray focus of the X-ray tube 22 toward the X-ray detector 25 into a fan beam or a cone beam.

X-ray detector 25 detects X-rays that have passed through patient 40 .
The collimator 24 is arranged on the X-ray incident side with respect to the X-ray detector 25 and removes scattered X-rays.

A high voltage power supply 28 supplies high voltage and current to the x-ray tube 22 .
Aperture driver 29 drives aperture 23 to deform its opening.
The rotating portion driving device 30 rotates the rotating portion 27 .

The table 4 has a cradle 41 , a cradle support base 42 and a driving device 43 . A cradle 41 supports a patient 40 to be imaged. The cradle support base 42 supports the cradle 41 so as to be movable in the y and z directions. The driving device 43 drives the cradle 41 and the cradle support base 42 . Here, the longitudinal direction of the cradle 41 is the z-direction, the height direction of the table 4 is the y-direction, and the horizontal direction perpendicular to the z-direction and the y-direction is the x-direction.

The GT control unit 31 controls each device and each part in the gantry 2, the driving device 43 of the table 4, and the like.

The operation console 8 has an input section 81, a display section 82, a storage section 83, a processing section 84, a console control section 85, and the like.

The input unit 81 includes a keyboard, a pointing device, and the like for receiving instructions and information input from the operator and for performing various operations. The display unit 82 displays a setting screen for setting scanning conditions, a CT image, and the like, and is, for example, an LCD (Liquid Crystal Display) or an organic EL (Electro-Luminescence) display.

The storage unit 83 stores programs for causing the processor to execute various processes. The storage unit 83 also stores various data and various files. The storage unit 83 has a HDD (Hard Disk Drive), an SSD (Solid State Drive), a DRAM (Dynamic Random Access Memory), a ROM (Read Only Memory), and the like. The storage unit 83 may also include a portable storage medium 90 such as a CD (Compact Disk) or a DVD (Digital Versatile Disk).

The processing unit 84 performs image reconstruction processing based on data of the patient 40 acquired by the gantry 2, and performs various other calculations. The processing unit 84 has one or more processors, and the one or more processors execute various processes of programs stored in the storage unit 83 . Note that the processing unit 84 corresponds to the inference device of the present invention.

FIG. 3 is a diagram showing main functional blocks of the processing unit 84. As shown in FIG.
The processing unit 84 has a reconstruction unit 841 and an inference unit 842 .

The reconstruction unit 841 reconstructs a CT image (for example, a scout image, a non-contrast CT image, and a contrast CT image) based on projection data obtained by scanning.

The inference unit 842 inputs a CT image to the trained model, and performs inference for distinguishing regions in which the contrast medium is taken in and regions in which the contrast medium is not taken in the imaging region of the patient.

The storage unit 83 also stores one or more instructions that can be executed by one or more processors. The one or more instructions cause one or more processors to perform the following actions (a1) and (a2).
(a1) Reconstruct a CT image based on the projection data (reconstruction unit 841).
(a2) Input a CT image to the trained model, and perform inference for discriminating between a region in which the contrast agent is taken and a region in which the contrast medium is not taken in the imaging region of the patient 40 (inference unit 842). .

The processing unit 84 of the operation console 8 can read the commands stored in the storage unit 83 and execute the above operations (a1) to (a2).

The console control section 85 controls the display section 82 and the processing section 84 based on the input from the input section 81 .
The X-ray CT apparatus 1 is configured as described above.

In recent years, dual-energy CT has been used to distinguish between benign and malignant tumors. However, some CT systems cannot use dual energy CT. In order to deal with this problem, a DNN (Deep Neural Network) is trained using dual-energy CT data to generate a trained DNN, and the single-energy CT data is input to this trained DNN. Research has been conducted to infer dual-energy information. However, with this method, there may appear areas within the imaged region that are rendered brighter than other tissues even though the contrast medium is not incorporated. For example, if the imaged region includes food and calcified parts in the stomach, the food and calcified parts may be depicted brightly. Therefore, it may be difficult to discriminate between areas in which the contrast agent has been incorporated and areas in which the contrast agent has not been incorporated.

Therefore, in the present embodiment, in order to deal with this problem, DL (deep learning) is used to determine a region in which the contrast agent is taken and a region in which the contrast agent is not taken. is generating
A method of generating this trained model will be described below with reference to FIGS. 4 to 8. FIG.

FIG. 4 is a diagram showing a flow chart of the learning phase.
At step ST1, a plurality of CT images to be used in the learning phase are prepared. FIG. 5 is an explanatory diagram of CT images used in the learning phase. In this embodiment, the CT images used in the learning phase are as follows.
(1) Non-enhanced CT image of chest including liver (2) Contrast-enhanced CT image of chest including liver

In FIG. 5, the non-contrast CT images are denoted by CNi (1≤i≤n). Contrast-enhanced CT images are denoted by CPi (1≤i≤n), CQi (1≤i≤n), and CRi (1≤i≤n). The contrast-enhanced CT images CPi, CQi, and CRi are images obtained by multiphase contrast-enhanced imaging. The contrast-enhanced CT image CPi is a contrast-enhanced CT image of a first phase, the contrast-enhanced CT image CQi is a contrast-enhanced CT image of a second time phase, and the contrast-enhanced CT image CRi is a contrast-enhanced CT image of a third time phase. . The first phase, the second phase, and the third phase are, for example, an arterial phase, a parenchymal phase, and a venous phase. If there is a tumor in the liver, contrast-enhanced imaging shows that the degree of staining of the tumor changes over time depending on the type of tumor. Therefore, multiphase CT images are useful images for differentiating tumors.

FIG. 6 is a diagram schematically showing an example of the non-contrast CT image CNi and the contrast CT images CPi, CQi, and CRi shown in FIG.
FIG. 6 shows images with i=1, ie, non-contrast CT image CN1, contrast CT images CP1, CQ1, and CR1.

In FIG. 6, for convenience of explanation, each image CN1, CP1, CQ1, and CR1 shows only the liver, tumor, and food in the stomach in a simplified manner.

In the non-contrast-enhanced CT image CN1, the tumor is rendered dark (pixel value is small) because the tumor does not contain the contrast agent. On the other hand, in the contrast-enhanced CT images CP1, CQ1, and CR1, the tumor is depicted brightly because the contrast agent is incorporated into the tumor. On the other hand, in order to simplify the explanation, it is assumed that the regions of the liver excluding the tumor are depicted with the same brightness among the images CN1, CP1, CQ1, and CR1. In FIG. 6, the area of the liver excluding the tumor is shaded.

Food in the stomach appears brighter than the surrounding tissue. In FIG. 6, in order to simplify the explanation, it is assumed that the food is rendered with the same brightness among the images CN1, CP1, CQ1, and CR1.

In addition, in FIG. 6, an image depicting food is shown as an image used in the learning phase. However, images depicting objects other than food that may be confused with tumors (for example, calcified parts) can also be prepared as images to be used in the learning phase. In addition, the images used in the learning phase may be images in which positional displacement caused by human body motion (for example, body motion due to breathing, body motion due to heart) have not been corrected, or images in which positional deviation due to body motion has not been corrected. may be an image in which the is corrected.

Returning to FIG. 5, the description continues.
In addition, a plurality of correct data are also prepared. Correct data are denoted by GPi (1≤i≤n), GQi (1≤i≤n), and GRi (1≤i≤n). A method for generating correct data will be described below.

FIG. 7 is an explanatory diagram of an example of a method for creating correct answer data.
7, the image of i=1 shown in FIG. 6, that is, the non-contrast CT image CN1, the first phase contrast CT image CP1, the second phase contrast CT image CQ1, and the 3 phase contrast-enhanced CT image CR1 is reproduced.

In FIG. 7, from the non-contrast CT image CN1, the first phase contrast CT image CP1, the second phase contrast CT image CQ1, and the third phase contrast CT image CR1, the difference An example is shown for generating images GP1, GQ1 and GR1.

A difference image GP1 is a difference image between the non-contrast CT image CN1 and the first phase contrast CT image CP1, and a difference image GQ1 is a difference image between the non-contrast CT image CN1 and the second phase contrast CT image CQ1. The difference image GR1 is a difference image between the non-contrast CT image CN1 and the contrast CT image CR1 of the third time phase.

First, the difference image GP1 will be explained.
Focusing on the tumor, the tumor in the non-contrast CT image CN1 is rendered darker than the tumor in the first phase contrast CT image CP1. Therefore, when the non-contrast CT image CN1 and the first phase contrast CT image CP1 are compared, there is a large difference in the brightness of the tumor. On the other hand, focusing on food, the food in the non-contrast CT image CN1 does not show a large difference in brightness compared to the food in the contrast CT image CP1 of the first time phase. Therefore, when the difference between the non-contrast CT image CN1 and the contrast CT image CP1 of the first time phase is obtained, the difference value of the tumor region (the region into which the contrast agent is taken) becomes a relatively large value. The difference value of the area (the area where the contrast agent is not taken) is a relatively small value. Therefore, by subtracting the non-contrast-enhanced CT image CN1 and the contrast-enhanced CT image CP1 in the first phase, it is possible to obtain a differential image GP1 in which the difference between the food and the tumor is emphasized. still,
Also, the other two difference images GQ1 and GR1 can be explained in the same manner as the difference image GP1. Therefore, by subtracting the non-contrast-enhanced CT image CN1 and the contrast-enhanced CT image CQ1 in the second time phase, a differential image GQ1 in which the difference between the food and the tumor is emphasized can be obtained. Similarly, by subtracting the non-contrast-enhanced CT image CN1 and the contrast-enhanced CT image CR1 in the third time phase, it is possible to obtain a subtraction image GR1 in which the difference between the food and the tumor is emphasized.

As described above, the difference images GP1, GQ1, and GR1 are images in which the difference between food and tumor is emphasized. can be used as correct data for

In FIG. 7, an example of generating difference images GP1, GQ1, and GR1 from a set of images CN1, CP1, CQ1, and CR1 with i=1 has been described. Generate an image.

As described above, the correct data shown in FIG. 5, that is, the difference images GP1 to GPn, the difference images GQ1 to GQn, and the difference images GR1 to GRn can be prepared. Predetermined image processing is performed on the difference images GP1 to GPn, the difference images GQ1 to GQn, and the difference images GR1 to GRn. Images GQ1' to GQn' and differential images GR1' to GRn' may be prepared as correct data. Here, to simplify the explanation, it is assumed that difference images GP1 to GPn, difference images GQ1 to GQn, and difference images GR1 to GRn are prepared as correct data.

Each differential image (correct data) is labeled as a set of non-contrast CT images and contrast-enhanced CT images. Specifically, the differential image GPi is labeled with the non-contrast CT image CNi and the contrast-enhanced CT image CPi of the first phase, and the differential image GQi is labeled with the non-contrast CT image CNi and the second phase CT image CNi. The contrast-enhanced CT image CQi is labeled, and the difference image GRi is labeled with the non-contrast-enhanced CT image CNi and the third phase contrast-enhanced CT image CRi.
After preparing these CT images and differential images, the process proceeds to step ST2.

FIG. 8 is an explanatory diagram of step ST2.
In step ST2, the neural network (NN) 91 performs learning using the non-contrast CT image CNi, the contrast CT images CPi, CQi, CRi, and the differential images (correct data) GPi, GQi, and GRi. This allows the trained model 91a to be generated.
The trained model 91a generated in this manner is stored in the storage unit 83 of the CT apparatus.

The learned model 91a obtained by the above learning phase is used when making inferences for discriminating between areas in which the contrast medium is taken and areas in which the contrast medium is not taken in the imaging examination of the patient. . A patient contrast examination flow will be described below.

FIG. 9 is a diagram showing an imaging examination flow.
In step ST10, the operator guides the patient 40, who is the subject of imaging, to the scan room, and lays the patient 40 on the table 4 as shown in FIG. A non-contrast scan of the chest of patient 40 is then performed, the patient is administered a contrast agent, and a multi-phase contrast-enhanced scan is performed.

In step ST20, the reconstruction unit 841 reconstructs a CT image based on the projection data obtained by the scanning in step ST10. FIG. 10 is a diagram showing an example of a reconstructed image.

The reconstruction unit 841 (see FIG. 3) reconstructs a non-contrast CT image CN0 including the chest of the patient 40 based on the projection data obtained by the non-contrast scanning. The reconstruction unit 841 also reconstructs multi-phase contrast-enhanced CT images CP0, CQ0, and CR0 including the chest of the patient 40 based on projection data obtained by multi-phase contrast-enhanced scanning.

In this embodiment, the multiphase contrast scan is a scan for acquiring CT images in three time phases (first time phase, second time phase, and third time phase). , CQ0, and CR0 represent contrast-enhanced CT images of the first, second, and third phases, respectively.

In FIG. 10, for convenience of explanation, images CN0, CP0, CQ0, and CR0 simply show only the liver, tumor, and food in the stomach in the same manner as in FIG.
After reconstructing the image, the process proceeds to step ST30.

In step ST30, the inference unit 842 performs inference for distinguishing regions in the chest of the patient 40 in which the contrast medium has been taken and regions in which the contrast medium has not been taken. FIG. 11 is an explanatory diagram of the inference phase.

The inference unit 842 inputs the non-contrast-enhanced CT image CN0 and the contrast-enhanced CT images CP0, CQ0, and CR0 reconstructed in step ST20 to the learned model 91a. Based on these input images CN0, CP0, CQ0, and CR0, the trained model 91a discriminates regions in the chest of the patient 40 in which the contrast agent is taken and regions in which the contrast agent is not taken. and outputs differential images GP0, GQ0, and GR0 representing the discrimination results.

Referring to FIG. 11, the contrast-enhanced CT images CP0, CQ0, and CR0 input to the trained model 91a brightly depict not only the tumor that has taken up the contrast medium, but also the food in the stomach that has not taken up the contrast medium. It is However, in the results output by the trained model 91a, the food in the stomach is depicted dark, and it can be seen that the tumor and the food can be distinguished. These differential images GP0, GQ0, and GR0 contain position data representing the positions of the areas in which the contrast medium has been taken and position data representing the positions of the areas in which the contrast medium has not been taken. Therefore, in the chest of the patient 40, it is possible to distinguish between areas in which the contrast medium has been taken and areas in which the contrast medium has not been taken.
Thus, the flow shown in FIG. 9 ends.

In this embodiment, a trained model 91a is generated for discriminating between a region (for example, a tumor) in which a contrast medium is taken and a region in which no contrast medium is taken (for example, food in the stomach). . Therefore, even if an area such as food in the stomach appears brightly in the patient's CT image obtained in the examination, even if the contrast medium is not taken in, the learned By using the model 91a, it is possible to discriminate between regions that have contrast agent uptake and regions that have no contrast agent uptake. Therefore, it is possible to improve the accuracy of distinguishing between benign and malignant tumors.

In the first mode, three time phases are considered as a multiphase contrast-enhanced CT image. However, instead of three phases, there may be two phases, or four or more phases. Also, instead of the multi-phase contrast-enhanced CT image, a single-phase contrast-enhanced CT image may be used.

In the CT images CN0, CP0, CQ0, and CR0 (see FIG. 11) input to the trained model 91a, not only the tumor but also the food are depicted brightly, so the CT images CN0, CP0, CQ0, and CR0 It is difficult to distinguish between food and tumor just by looking at it. Therefore, in order to obtain an image suitable for diagnosis, the following steps A and B can be executed after outputting the inference result in step ST30.

Step A: A step of identifying the position of the area where the contrast medium is not taken in based on the inference result of step ST30. This step A can be executed by providing the processing unit 84 with an identification unit that executes this identification.

Step B: A step of reconstructing an image based on the positions identified in step A and the projection data obtained by scanning in step ST10 so that areas where the contrast medium is not incorporated are dark. This step B can be executed by the reconstruction unit 841 (see FIG. 3) of the processing unit 84 .

By executing steps A and B above, it is possible to provide the operator and radiologist with a CT image suitable for distinguishing between benign and malignant tumors.

(Second form)
In the second embodiment, an example will be described in which the output result of the learned model created in the first embodiment is used to improve the reliability of dual energy information.

In the second form, the processing section 84 has the following functional blocks.
FIG. 12 is a diagram showing functional blocks of the processing unit 84 in the second mode.
The processing unit 84 of the second form has a reconstruction unit 841, an inference unit 842, and a determination unit 843 as main functional blocks.

The reconstruction unit 841 reconstructs a CT image (for example, a scout image, a non-contrast CT image, and a contrast CT image) based on projection data obtained by scanning.

The inference unit 842 inputs a CT image to the trained model 91a of the first form, and makes an inference for distinguishing a region in which the contrast medium is taken in and a region in which the contrast medium is not taken in the imaging region of the patient 40. to run. In addition, the inference unit 842 inputs the CT image to the learned DNN 92a (see FIG. 14), which will be described later, and infers dual energy information.

The determination unit 843 determines whether or not the inference result of the learned DNN 92a is correct based on the inference result of the learned model 91a of the first form.

The storage unit 83 also stores one or more instructions that can be executed by one or more processors. The one or more instructions cause one or more processors to perform the following actions (b1)-(b4).

(b1) Reconstruct a CT image based on the projection data (reconstruction unit 841).
(b2) Input a CT image to the trained model 91a, and perform inference for distinguishing between regions in which the contrast medium is taken and regions in which the contrast medium is not taken in the imaging region of the patient 40 (the inference unit 842 ).
(b3) Input a CT image to the learned DNN 92a and infer dual energy information (inference unit 842).
(b4) Based on the inference result of the trained model 91a, it is judged whether or not the inference result of the trained DNN 92a (see FIG. 14) is correct (judgment section 843).

The contrast examination flow of the second mode will be described below with reference to FIGS. 13 and 14. FIG.

FIG. 13 is a diagram showing a contrast examination flow in the second embodiment.
Steps ST10 and ST20 are the same as those in the first mode, so description thereof will be omitted. After image reconstruction, the process proceeds to step ST30.

FIG. 14 is an explanatory diagram of step ST30.
In step ST30, the inference unit 842 inputs the non-contrast CT image CN0 and the contrast-enhanced CT images CP0, CQ0, and CR0 obtained in step ST20 to the trained model 91a. The trained model 91a outputs difference images GP0, GQ0, and GR0, as described in the first mode.

In step ST30, the inference unit 842 also inputs the CT images CN0, CP0, CQ0, and CR0 obtained in step ST20 to the trained DNN 92a. The trained DNN 92a is a trained model generated by a neural network learning dual-energy data. When single-energy data is input, the trained DNN 92a obtains positional information of the region in which the contrast agent is captured. It infers the dual energy information DI containing.

Since CT images CN0, CP0, CQ0, and CR0 are images obtained by single-energy CT, when these CT images CN0, CP0, CQ0, and CR0 are input to trained DNN 92a, trained DNN 92a is , outputs dual-energy information DI that includes the location information of the area in which the contrast agent is taken.

In step ST40, the determination unit 843 compares the result (difference images GP0, GQ0, and GR0) inferred by the trained model 91a and the result (dual energy information DI) inferred by the trained DNN 92a. Based on the result of this comparison, the determining unit 843 determines whether the area inferred by the learned DNN 92a that the contrast agent has been taken in is actually the area in which the contrast agent has been taken in, or whether the contrast medium has actually been taken in. determine if it is an area that was incorrectly inferred to have contrast agent uptake when it was not there.
Thus the flow ends.

In the second form, the estimation result of the learned model 91a and the estimation result of the learned DNN 92a are compared, and the area inferred that the contrast agent is taken in by the learned DNN 92a is actually taken in. It can be determined whether or not it is an area. Therefore, the inference results of the trained model 91a can be used to improve the inference accuracy of the trained DNN 92a.

In the first and second embodiments, the liver is the imaged region, but the imaged region is not necessarily limited to the liver, and a region including organs other than the liver may be the imaged region.

In addition, in the first and second embodiments, an example of generating a trained model using non-contrast CT images and contrast-enhanced CT images has been described. However, a trained model may be generated using only multi-phase contrast-enhanced CT images without using non-contrast-enhanced CT images. This trained model can be generated, for example, by executing the following steps (1)-(4).

(1) A plurality of contrast-enhanced CT images are prepared by contrast-enhanced imaging of a predetermined part of a human being at a plurality of time phases.

(2) Prepare a plurality of correct data. Each correct data is generated by subtracting a contrast-enhanced CT image captured in one of a plurality of time phases and a contrast-enhanced CT image captured in another time phase of the plurality of time phases. be done.

(3) Labeling the correct answer data on the contrast-enhanced CT image.

(4) The computer causes the neural network to perform learning using a plurality of contrast-enhanced CT images and a plurality of correct data. Thus, a trained model is generated. This trained model executes inference for discriminating regions in which the contrast agent is taken in and regions in which the contrast agent is not taken in the imaging region of the patient by inputting the contrast-enhanced CT image.

In the inference phase, the patient's contrast-enhanced CT images taken at multiple time phases are input to the above-described trained model, and a method for discriminating between areas in which the contrast agent is incorporated and areas in which the contrast agent is not incorporated is performed. perform inference. This method makes it possible to differentiate tumors without using non-enhanced CT images.

In addition, in the first and second forms, the learned model is created by DL (deep learning), but the learned model may be created by machine learning other than DL.

1 CT device 4 Table 6 Camera 8 Operation console 21 Bore 22 X-ray tube 23 Aperture 24 Collimator 25 X-ray detector 26 Data acquisition part 27 Rotating part 28 High voltage power supply 29 Aperture driving device 30 Rotating part driving device 31 GT control part 40 Patient 41 Cradle 42 Cradle support base 43 Driving device 81 Input unit 82 Display unit 83 Storage unit 84 Processing unit 85 Console control unit 90 Storage medium 91 Neural network 91a Trained model 92a Trained DNN
100 Scan room 101 Ceiling 200 Operation room 841 Reconstruction unit 842 Inference unit 843 Determination unit

Claims (8)

By inputting a non-contrast-enhanced CT image including a predetermined part of an imaging subject and a contrast-enhanced CT image including a predetermined part of the imaging subject, a contrast agent is taken in the predetermined part of the imaging subject. A trained model generation method for generating a trained model that performs inference to discriminate between regions and non-contrast loaded regions, comprising:
the computer
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

A trained model generation method, wherein the trained model is generated by performing learning using 2. The method of generating a learned model according to claim 1, wherein said plurality of contrast-enhanced CT images include a plurality of contrast-enhanced CT images taken at a plurality of time phases. 3. The method of generating a trained model according to claim 1, wherein said computer generates said trained model by causing a neural network to perform said learning. By inputting a non-contrast-enhanced CT image including a predetermined part of an imaging subject and a contrast-enhanced CT image including a predetermined part of the imaging subject, a contrast agent is taken in the predetermined part of the imaging subject. An inference device that includes a trained model that performs inference to discriminate between regions and non-contrast uptake regions,
The learned model is a computer,
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

A reasoning device generated by executing learning using By inputting a non-contrast-enhanced CT image including a predetermined part of an imaging subject and a contrast-enhanced CT image including a predetermined part of the imaging subject, a contrast agent is taken in the predetermined part of the imaging subject. A medical device that includes a trained model that performs inference to discriminate between regions and non-contrast uptake regions,
The learned model is a computer,
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

A medical device that has been generated by performing learning using A learned DNN generated by performing learning using dual-energy CT data, wherein the DNN infers dual-energy information from single-energy CT data;
The medical apparatus according to any one of claims 1 to 5, further comprising a determination unit that determines whether the inference result of the learned DNN is correct based on the inference result of the learned model. to one or more processors,
A non-contrast-enhanced CT image including a predetermined part of an imaging subject and a contrast-enhanced CT image including a predetermined part of the imaging subject are input to a trained model, and the predetermined part of the imaging subject is input to the learned model. A program for executing an operation including executing inference for distinguishing between a region in which a contrast agent is taken and a region in which a contrast agent is not taken in,
The learned model is a computer,
(1) a plurality of non-contrast CT images containing the predetermined part of a human;
(2) a plurality of contrast-enhanced CT images including the predetermined part of a human; and (3) a plurality of correct data, each correct data being one non-contrast-enhanced CT image among the plurality of non-contrast-enhanced CT images. and one contrast-enhanced CT image out of the plurality of contrast-enhanced CT images. Multiple correct data, labeled with

A program generated by performing learning using A region in which a contrast agent is incorporated in the predetermined part of the subject by inputting a plurality of contrast-enhanced CT images obtained by contrast-enhanced imaging of the predetermined part of the subject in a plurality of time phases. A trained model generation method for generating a trained model that performs inference for discriminating between a region and a region where no contrast agent is captured,
the computer
(1) a plurality of contrast-enhanced CT images obtained by contrast-enhanced imaging of a predetermined part of a human at the plurality of time phases; and (2) a plurality of correct data labeled on the plurality of contrast-enhanced CT images, , each correct data is a difference between a contrast-enhanced CT image captured in one of the plurality of time phases and a contrast-enhanced CT image captured in another time phase of the plurality of time phases. A plurality of correct data generated based on the difference image generated by

A trained model generation method that is generated by performing learning using

Images