WO2022080327A1 - Estimator learning device, estimator learning method, and estimator learning program - Google Patents

Estimator learning device, estimator learning method, and estimator learning program Download PDF

Info

Publication number
WO2022080327A1
WO2022080327A1 PCT/JP2021/037615 JP2021037615W WO2022080327A1 WO 2022080327 A1 WO2022080327 A1 WO 2022080327A1 JP 2021037615 W JP2021037615 W JP 2021037615W WO 2022080327 A1 WO2022080327 A1 WO 2022080327A1
Authority
WO
WIPO (PCT)
Prior art keywords
kidney
image
subject
learning
feature amount
Prior art date
Application number
PCT/JP2021/037615
Other languages
French (fr)
Japanese (ja)
Inventor
勉 井上
雅浩 石川
栄人 小澤
浩一 岡田
直樹 小林
守 新津
Original Assignee
学校法人埼玉医科大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 学校法人埼玉医科大学 filed Critical 学校法人埼玉医科大学
Priority to JP2022556974A priority Critical patent/JPWO2022080327A1/ja
Publication of WO2022080327A1 publication Critical patent/WO2022080327A1/en

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves 
    • A61B5/055—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves  involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • the present invention relates to an estimator learning device, an estimator learning method, and an estimator learning program.
  • eGFR glomerular filtration rate
  • the eGFR Slope is obtained from, for example, the rate of change between eGFR at a certain point in time and eGFR several months to several years after a certain point in time (the amount of change in eGFR per hour).
  • eGFR Slope deterioration rate of renal disease
  • MR Magnetic Resonance
  • An object of the present invention is to provide a technique capable of estimating the rate of exacerbation of renal disease based on renal MR images.
  • the first aspect is A learning MR (Magnetic Resonance) image of the part including the kidney of the subject is acquired, an image of the kidney extracted from the part of the kidney included in the learning MR image is generated, and the image of the kidney is made into a non-rigid body. It is converted to a standard kidney image by conversion, and the learning feature amount based on the converted kidney image is calculated.
  • An estimator that estimates the deterioration rate of the kidney disease of the subject by machine learning using the data set including the calculated feature amount for learning and the deterioration rate of the kidney disease of the subject as teacher data. Processor to build, It is an estimation device equipped with.
  • the aspect of disclosure may be realized by executing the program by the information processing device. That is, the configuration of the disclosure can be specified as a program for causing the information processing apparatus to execute the process executed by each means in the above-described embodiment, or as a computer-readable recording medium on which the program is recorded. Further, the configuration of the disclosure may be specified by a method in which the information processing apparatus executes the processing executed by each of the above-mentioned means. The configuration of the disclosure may be specified as a system including an information processing apparatus that performs processing executed by each of the above-mentioned means.
  • FIG. 1 is a diagram showing a configuration example of an estimation device according to an embodiment.
  • FIG. 2 is a diagram showing an example of a T2 * map image.
  • FIG. 3 is a diagram showing an example of an operation flow when constructing an estimator in an estimator.
  • FIG. 4 is an MR image of a site including the kidney of the subject.
  • FIG. 5 is an image obtained by extracting a portion corresponding to the kidney from the MR image of FIG.
  • FIG. 6 is a diagram showing an example of the non-rigid body conversion process of the kidney.
  • FIG. 7 is a diagram showing an example of a histogram generated by the feature amount calculation process.
  • FIG. 8 is a diagram showing an example of an operation flow when estimating the exacerbation rate of renal disease in the estimation device.
  • FIG. 1 is a diagram showing a configuration example of the estimation device of the present embodiment.
  • the estimation device 100 shown in FIG. 1 has a general computer configuration.
  • the estimation device 100 of the present embodiment includes a processor 101, a memory 102, a storage unit 103, an input unit 104, an output unit 105, and a communication control unit 106. These are connected to each other by a bus.
  • the memory 102 and the storage unit 103 are computer-readable recording media.
  • the hardware configuration of the estimation device 100 is not limited to the example shown in FIG. 1, and components may be omitted, replaced, or added as appropriate.
  • the estimation device 100 of the present embodiment includes clinical information (gender, age, urinary protein, serum creatinine level, etc.) of the subject, an MR image of a site including the kidney of the subject, and an MR image of the subject.
  • the rate of exacerbation of renal disease (eGFR slope) calculated after the imaging of the image is obtained.
  • the estimation device 100 performs a predetermined process on the MR image, and constructs an estimator that estimates the exacerbation rate of the renal disease from the clinical information of the subject and the MR image.
  • the estimation device 100 estimates the rate of deterioration of renal disease (eGFR slope) from clinical information and MR images of the subject using the constructed estimator.
  • the MR image is an image of a part of the body taken by a nuclear magnetic resonance apparatus.
  • MR images of the kidney reflect renal oxygen status, renal fibrosis, and renal perfusion. Kidney oxygen status, renal fibrosis, and renal perfusion are associated with renal disease.
  • the estimation device 100 can be realized by using a dedicated or general-purpose computer such as a workstation (WS, WorkStation), a PC (Personal Computer), a smartphone, a tablet terminal, or an electronic device equipped with a computer. ..
  • the estimation device 100 can be realized by using a computer (server device) that provides services through a network.
  • the estimation device 100 can be realized by a computer that executes parallelization by MPI (Message Passing Interface) in which CPUs or GPUs are parallelized on a large scale.
  • the estimation device 100 loads a program stored in the recording medium by the processor 101 into a work area of the memory 102 and executes the program, and each component or the like is controlled through the execution of the program, so that the function that meets a predetermined purpose is met. Can be realized.
  • MPI Message Passing Interface
  • the processor 101 is, for example, a CPU (Central Processing Unit) or the like.
  • the processor 101 loads and executes a program stored in the memory 102 to execute an image acquisition process, a kidney region extraction process, a non-rigid body conversion process, a feature amount calculation process, an estimator construction process, an estimation process, and the like. .. Further, the processor 101 can acquire data and the like used in each process from other devices via the storage unit 103 and the communication control unit 106.
  • a CPU Central Processing Unit
  • the image acquisition process is a process of acquiring an MR image (MRI) of a part including the kidney of a subject from another information processing device or the like. Further, in the image acquisition process, information on the subject such as clinical information on the subject is acquired together with the MR image.
  • the clinical information of the subject includes, for example, gender, age, mean blood pressure, urinary protein level, uric acid level, serum creatinine level, eGFR, and the presence or absence of a history of diabetes. Other information may be adopted as the clinical information of the subject.
  • the subject corresponding to the MR image used when constructing the estimator can perform eGFR Slope (deterioration rate of renal disease) after the MR image is taken (for example, several months to several years later). ).
  • the eGFR Slope is calculated based on the time change of the eGFR since the MR image of the subject was taken. For example, it is assumed that the subject calculates eGFR Slope (deterioration rate of renal disease) based on the eGFR when the MR image is taken and the eGFR one year after the MR image is taken. ..
  • a T2 * map image, an R2 * map image, and an ADC (Apparent Diffusion Coefficient) map image can be used.
  • the MR image used here is not limited to these.
  • the T2 * map image used here is created based on 12 T2 * emphasized images taken at different echo times (echo time: TE).
  • the T2 * map image is not limited to the one described here.
  • FIG. 2 is a diagram showing an example of a T2 * map image.
  • the image on the left in FIG. 2 is an example of a T2 * map image containing a normal kidney.
  • the image on the right in FIG. 2 is an example of a T2 * map image containing a kidney with chronic kidney disease with impaired renal function.
  • the kidney region extraction process extracts the portion corresponding to the kidney from the MR image acquired by the image acquisition process.
  • an estimator that estimates the kidney portion from MR images constructed in advance by deep learning or the like is used.
  • the MR image having the largest size of the extracted kidney may be used in the subsequent process.
  • the MR image with the largest kidney size is likely to be included near the center of the kidney, which is thought to contribute to improving the performance of the estimator in constructing the estimator.
  • the non-rigid body conversion process is a process of converting the shape of the kidney extracted by the kidney region extraction process into a standard kidney shape.
  • a well-known non-rigid conversion process is used.
  • the standard kidney shape is, for example, the normal kidney shape. Converting to a standard kidney shape improves the performance of the estimator in estimator construction.
  • the feature amount calculation process is a process of extracting the feature amount of the kidney from the image of the kidney converted into the standard kidney shape by the non-rigid body conversion process.
  • the kidney is divided into 12 layers from the cortex to the medulla, and the pixel values of the pixels included in each layer are aggregated to obtain a feature amount.
  • the aggregation of the pixel values for example, the frequency of the pixel values of the pixels and the average of the pixel values of the pixels can be obtained for each layer.
  • the estimator construction process uses a data set including clinical information of the subject, the feature amount of the kidney obtained from the MR image of the subject, and the rate of deterioration of the renal disease of the subject as teacher data. This is a process for constructing an estimator that estimates the rate of deterioration of renal disease in a subject from clinical information of the examiner and the feature amount of the kidney.
  • the estimation process uses the estimator constructed by the estimator construction process to obtain the clinical information of the subject and the features of the kidney obtained from the MR image of the subject. This is a process for estimating the rate of deterioration.
  • the memory 102 is composed of, for example, a RAM (RandomAccessMemory), a RAM, and a ROM (ReadOnlyMemory).
  • the memory 102 is also called a main storage device.
  • the storage unit 103 is, for example, an EPROM (ErasableProgrammableROM), a hard disk drive (HDD, HardDiskDrive), or the like. Further, the storage unit 103 can include a removable medium, that is, a portable recording medium.
  • the removable media is, for example, a USB (Universal Serial Bus) memory or a disc recording medium such as a CD (Compact Disc) or a DVD (Digital Versatile Disc).
  • the storage unit 103 is also called a secondary storage device.
  • the storage unit 103 stores various programs, various data, and various tables used in the estimation device 100 in a readable / writable recording medium.
  • the storage unit 103 stores an operating system (Operating System: OS), various application programs, various tables, and the like.
  • OS Operating System
  • the information stored in the storage unit 103 may be stored in the memory 102. Further, the information stored in the memory 102 may be stored in the storage unit 103.
  • a program for executing image acquisition processing, kidney region extraction processing, non-rigid body conversion processing, feature amount calculation processing, estimator construction processing, estimation processing, etc. is installed in the storage unit 103. Further, the storage unit 103 stores various data and the like used in each process and the like in the estimation device 100. The storage unit 103 stores MR images of subjects, clinical information, eGFR Slope, and the like.
  • the operating system is software that mediates between software and hardware, manages memory space, manages files, manages processes and tasks, and so on.
  • the operating system includes a communication interface.
  • the communication interface is a program for exchanging data with other external devices and the like connected via the communication control unit 106.
  • the external device and the like include, for example, another information processing device, an external storage device, and the like.
  • the input unit 104 includes a keyboard, a pointing device, a wireless remote controller, a touch panel, and the like. Further, the input unit 104 can include a video or image input device such as a camera, or an audio input device such as a microphone.
  • the output unit 105 includes a display device such as an LCD (Liquid Crystal Display), an EL (Electroluminescence) panel, a CRT (Cathode Ray Tube) display, a PDP (Plasma Display Panel), and an output device such as a printer. Further, the output unit 105 can include an audio output device such as a speaker.
  • a display device such as an LCD (Liquid Crystal Display), an EL (Electroluminescence) panel, a CRT (Cathode Ray Tube) display, a PDP (Plasma Display Panel), and an output device such as a printer.
  • the output unit 105 can include an audio output device such as a speaker.
  • the communication control unit 106 is connected to another device and controls communication between the estimation device 100 and the other device.
  • the communication control unit 106 is, for example, a LAN (Local Area Network) interface board, a wireless communication circuit for wireless communication, and a communication circuit for wired communication.
  • the LAN interface board and the wireless communication circuit are connected to a network such as the Internet.
  • the step of writing a program includes not only the processes performed in chronological order in the described order but also the processes executed in parallel or individually even if they are not necessarily processed in chronological order. Some of the steps in writing the program may be omitted.
  • the series of processes executed by the processor 101 can be executed by hardware or software.
  • Hardware components are hardware circuits, such as FPGAs (Field Programmable Gate Arrays), application-specific integrated circuits (ASICs), gate arrays, logic gate combinations, analog circuits, and the like. ..
  • FIG. 3 is a diagram showing an example of an operation flow when constructing an estimator in an estimator.
  • the estimation device 100 uses the clinical information of the subject, the feature amount based on the MR image of the subject, and the deterioration rate of the renal disease of the subject as teacher data, and obtains the deterioration rate of the renal disease from the clinical information and the feature amount. Build an estimator to estimate.
  • the processor 101 of the estimation device 100 acquires an MR image or the like of the subject, which is the basis of the teacher data used for constructing the estimator.
  • the processor 101 acquires the MR image of the subject stored in the storage unit 103.
  • the MR image is an image taken of a site including the kidney of the subject.
  • the processor 101 acquires the information of the subject such as the clinical information of the subject together with the MR image of the subject from the storage unit 103.
  • the processor 101 acquires the exacerbation rate (eGFR Slope) of the renal disease of the subject, which is stored in the storage unit 103. It is assumed that the subject calculates the rate of exacerbation of renal disease after the MR image is taken (for example, months to years later).
  • the processor 101 may acquire these data from other devices via the communication control unit 106, the network, and the like.
  • the MR image, clinical information, and deterioration rate of renal disease acquired here are MR images for learning, clinical information, and deterioration rate of renal disease.
  • the processor 101 extracts a portion corresponding to the kidney from the MR image acquired in S101.
  • the processor 101 uses an estimator that estimates the kidney portion from the MR image constructed in advance by deep learning or the like, and extracts the portion corresponding to the kidney from the MR image.
  • the estimator is stored in the storage unit 103 in advance.
  • the processor 101 stores an image (image of the kidney) of a portion (region) corresponding to the kidney from the extracted MR image in the storage unit 103. Further, the processor 101 may store the MR image and the information of the portion (region) corresponding to the kidney included in the image in the storage unit 103 in association with each other.
  • the processor 101 extracts a portion corresponding to the kidney for each MR image. Further, the processor 101 uses the MR image having the largest portion corresponding to the extracted kidney in the subsequent processing. Extraction of the portion corresponding to the kidney may be performed by another method. Extraction of the portion corresponding to the kidney may be performed, for example, by designating the region of the kidney portion by the input unit 104 by the user.
  • FIG. 4 is an MR image of the site including the kidney of the subject.
  • the MR image of FIG. 4 is an image of a person's abdomen viewed from the front.
  • kidneys appear on the left and right.
  • FIG. 5 is an image obtained by extracting a portion corresponding to the kidney from the MR image of FIG.
  • a portion corresponding to a kidney (right kidney and left kidney) is extracted by an estimator that estimates the kidney portion from the MR image of FIG.
  • the processor 101 converts the image of the portion corresponding to the kidney extracted in S102 into a standard kidney shape by a non-rigid body conversion process.
  • a well-known method is used as the non-rigid body conversion process.
  • the non-rigid body conversion process is a process of converting the shape of the kidney in the image of the portion corresponding to the kidney extracted in S102 into the shape of a standard kidney.
  • the standard kidney shape is, for example, the normal kidney shape.
  • the shape of the kidney may be deformed or atrophied due to the influence of renal disease or the like.
  • the non-rigid conversion process improves the performance of the estimator by aligning the shape of the kidneys.
  • the processor 101 stores the image of the kidney converted by the non-rigid body conversion process in the storage unit 103.
  • FIG. 6 is a diagram showing an example of the non-rigid body conversion process of the kidney.
  • the example of FIG. 6 shows an example in which the kidney of the first subject and the kidney of the second subject are transformed into a standard kidney.
  • Points A1, B1, and C1 in the kidney of the first subject in FIG. 6 correspond to points A0, B0, and C0 in the standard kidney, respectively.
  • points A2, B2, and C2 in the kidney of the second subject in FIG. 6 correspond to points A0, B0, and C0 in the standard kidney, respectively.
  • each pixel (point) in the image of the kidney before conversion is associated with any pixel (point) in the image of the standard kidney.
  • each pixel value (color, shading) of each corresponding pixel is maintained.
  • points A0, A1 and A2 are points corresponding to each other. That is, for example, the pixel value of the kidney of the first subject corresponding to the pixel value of the point A0 in the standard kidney is the pixel value of the point A1 in the image of the kidney of the first subject.
  • the processor 101 calculates the feature amount of the kidney from the image of the kidney of the standard shape converted in S103.
  • the processor 101 divides the kidney into a plurality of layers (for example, 12 layers) from the cortex to the medulla, aggregates the pixel values of the pixels included in each layer, and calculates the feature amount.
  • a plurality of layers from the cortex to the medulla for example, a normal kidney that has been previously divided into a plurality of layers by a specialist or the like is used. By dividing into a plurality of layers from the cortex to the medulla, it is possible to calculate the feature amount according to the internal part of the kidney.
  • the feature amount of each layer is, for example, a histogram showing the appearance frequency of the pixel values of each layer, an average value of the pixel values of each layer, and the like.
  • the pixel values are aggregated for each layer here, the pixel values of the pixels included in the kidney image may be aggregated without being divided into a plurality of layers. That is, the processor 101 may use a histogram showing the appearance frequency of the pixel values of all the pixels included in the image of the kidney as the feature amount, or may use the average value of the pixel values of all the pixels as the feature amount.
  • the processor 101 may use the pixel value of one or more specific pixels (positions) of the image of the kidney as a feature amount.
  • a particular location of the kidney is, for example, a location that is medically considered to be highly correlated with the rate of exacerbation of renal disease.
  • the designation of the position is facilitated by aligning the shape of the kidney with the standard shape.
  • the processor 101 stores the calculated feature amount in the storage unit 103.
  • the feature amount is stored in association with the clinical information of the subject and the rate of deterioration of renal disease.
  • As the feature amount another amount based on the image of the kidney may be adopted.
  • a histogram of pixel values for each layer an average value of pixel values for each layer, a histogram of pixel values for kidney images, an average pixel value for kidney images, one or more specific values for kidney images.
  • a plurality of quantities may be adopted as feature quantities.
  • the feature amount calculated here is an amount indicating the feature of the kidney included in the MR image.
  • the feature amount calculated here is a feature amount for learning.
  • FIG. 7 is a diagram showing an example of a histogram generated by the feature amount calculation process.
  • the horizontal axis of the histogram in FIG. 7 indicates the pixel value, and the vertical axis indicates the appearance frequency (number of appearances) of each pixel value.
  • the histogram shows the appearance frequency (number of appearances) of the pixel values of the pixels included in the image of each layer.
  • the vertical axis represents the appearance frequency (number of appearances) of the pixel value, but it may be the appearance ratio of the pixel value.
  • the pixel value takes a value from 0 to 255.
  • the processor 101 performs processing from S101 to S104 using various MR images of the subject, calculates a feature amount for each MR image, and stores it in the storage unit 103.
  • the processor 101 includes a plurality of features calculated in S104 for a plurality of subjects, clinical information of the subject acquired in S101, and a plurality of exacerbation rates of renal disease (eGFR EGFR) of the subject.
  • eGFR EGFR renal disease
  • a deep learning model of machine learning is used to construct an estimator that estimates the rate of deterioration of renal disease (eGFR Slope) from features and clinical information. Any model may be used as the deep learning model used here.
  • the processor 101 stores the constructed estimator in the storage unit 103.
  • a method using a learning space such as deep learning by a neural network, Regression SVM, Regression Random Forest, multiple regression analysis, Look Up Table, etc. can be used.
  • the data set as the teacher data may not include a part or all of the clinical information. That is, an estimator may be constructed using a plurality of data sets including the feature amount calculated in S104 and the rate of deterioration of renal disease as teacher data. If the data set does not include part or all of the clinical information, it is not necessary to acquire the clinical information that is not included in S101.
  • FIG. 8 is a diagram showing an example of an operation flow when estimating the exacerbation rate of renal disease in the estimation device.
  • the estimation device 100 uses the clinical information of the subject to be estimated, the feature amount based on the MR image of the subject, and the estimator constructed by the operation flow of FIG. 3, and worsens the renal disease of the subject. Estimate the speed.
  • the processor 101 acquires MR images and clinical information of the subject to be estimated from other devices via the storage unit 103, the communication control unit 106, the network, or the like. If the data set used as the teacher data when the estimator is constructed in the operation flow of FIG. 3 does not include a part or all of the clinical information, the processor 101 does not have to acquire the clinical information that is not included.
  • the processing from S202 to S204 is the same as the processing from S102 to S104 in FIG.
  • the processor 101 calculates the feature amount based on the MR image of the subject to be estimated by processing from S202 to S204, and stores it in the storage unit 103.
  • the feature amount calculated here by the processor 101 is the same type of feature amount as the feature amount (feature amount included in the teacher data) used when the estimator was constructed in the operation flow of FIG. That is, for example, when the feature amount is used as a histogram of the pixel values of each layer of the kidney image when constructing the estimator, the feature amount calculated here is for each layer of the image of the kidney of the subject to be estimated. It is a histogram of the pixel value of.
  • the processor 101 uses the estimator for estimating the exacerbation rate of the renal disease constructed by the operation flow of FIG. 3, and based on the clinical information acquired in S201 and the feature amount calculated in S204, the renal disease. Estimate the rate of deterioration (eGFR Slope).
  • the processor 101 stores the estimated exacerbation rate of the renal disease in the storage unit 103 in association with the information identifying the subject to be estimated.
  • the estimation device 100 can estimate the exacerbation rate of renal disease (eGFR Slope) from the MR image or the like of the subject to be estimated.
  • the processor 101 estimates the exacerbation rate of renal disease (eGFR Slope) without using the acquired clinical information.
  • the estimation device 100 can estimate the deterioration rate of the kidney disease by using the MR image of the site including the kidney of the subject and the estimator for estimating the deterioration rate of the kidney disease.
  • kidney is targeted, but it can be applied to other organs as well.
  • the teacher data includes (1) clinical information (urinary protein amount), (2) average pixel value of kidney image, and (3) kidney image layer.
  • Each estimator using the histogram (appearance frequency) of each pixel value, (4) (2) and (3), (5) (1), (2) and (3) was examined.
  • Each estimator is constructed based on the above.
  • the correlation coefficient between the eGFR slope estimated by each estimator and the actual eGFR slope was calculated.
  • each correlation coefficient was (1) 0.56, (2) 0.61, (3) 0.79, (4) 0.85, and (5) 0.87. The higher the correlation coefficient, the higher the performance of the estimator.
  • the performance of the estimator is improved when the feature amount based on the kidney image (MR image) is used as the teacher data than when the clinical information (urinary protein amount) which is non-image information is used as the teacher data. do. Further, it is useful when the histogram of the pixel value for each layer of the image of the kidney is used for the teacher data ((3), (4), (5)). In addition to the histogram of the pixel values for each layer of the kidney image, clinical information and the average value of the pixel values of the kidney image are added to the teacher data to further improve the performance of the estimator.
  • the estimation device 100 acquires an MR image including the kidney of the subject, clinical information of the subject, and eGFR slope (exacerbation rate of renal disease) of the subject.
  • the estimation device 100 extracts an image of the kidney from the acquired MR image.
  • the estimation device 100 non-rigidly transforms the kidney shape of the extracted kidney image into a standard shape, and then calculates the feature amount based on the kidney image. For example, the estimation device 100 divides the kidney into a plurality of layers from the cortex to the medulla, and calculates the appearance frequency of the pixel value of the image for each layer as a feature amount.
  • the estimation device 100 constructs an estimator that estimates eGFR slope from at least the feature amount based on the MR image, using at least a plurality of data sets including the feature amount based on the MR image and the eGFR slope as teacher data. According to the estimation device 100, a higher performance estimator can be constructed by using the feature amount based on the MR image. In addition, by converting the shape of the kidney of the subject into a standard shape in a non-rigid body, the shape of the kidney becomes uniform, and the performance of the estimator is improved.
  • the estimation device 100 by dividing the image of the kidney into a plurality of layers (for example, 12 layers) from the cortex to the medulla, it is possible to calculate the feature amount according to the internal part of the kidney.
  • the performance of the estimator is improved by using the features according to the internal part of the kidney.
  • the estimation device 100 acquires an MR image including the kidney of the subject to be estimated, and calculates a feature amount based on the MR image.
  • the estimation device 100 estimates the rate of deterioration of the renal disease of the subject to be estimated by using the calculated feature amount and the constructed estimator. According to the estimation device 100, the rate of deterioration of renal disease can be estimated more accurately by using the feature amount based on the MR image including the kidney.
  • Computer readable recording medium A program that realizes any of the above functions in a computer or other machine or device (hereinafter referred to as a computer or the like) can be recorded on a recording medium that can be read by a computer or the like. Then, by having a computer or the like read and execute the program of this recording medium, the function can be provided.
  • a recording medium that can be read by a computer or the like is a recording medium that can store information such as data and programs by electrical, magnetic, optical, mechanical, or chemical action and can be read from a computer or the like.
  • elements constituting a computer such as a CPU and a memory may be provided, and the CPU may execute a program.
  • recording media those that can be removed from a computer or the like include, for example, flexible disks, magneto-optical disks, CD-ROMs, CD-R / Ws, DVDs, DATs, 8 mm tapes, memory cards, and the like.

Abstract

Provided is an estimating device comprising a processor for constructing an estimator for estimating the rate of deterioration of a subject's kidney disease by machine learning in which a learning magnetic resonance (MR) image capturing a kidney-including area of the subject is acquired, a kidney image extracting the kidney portion included in the learning MR image is generated, the kidney image is converted by non-rigid conversion into a standard kidney image, a learning feature amount based on the kidney image obtained by the conversion is calculated, and a data set comprising the calculated learning feature amount and the rate of deterioration of the subject's kidney disease is used as training data.

Description

推定器学習装置、推定器学習方法、及び、推定器学習プログラムEstimator learning device, estimator learning method, and estimator learning program
　本発明は、推定器学習装置、推定器学習方法、及び、推定器学習プログラムに関する。 The present invention relates to an estimator learning device, an estimator learning method, and an estimator learning program.
　腎臓の働きを調べる指標として推算糸球体濾過量（ｅＧＦＲ）がある。ｅＧＦＲは、年齢、性別、血清クレアチニン値によって、算出される。血清クレアチニン値は、血液検査によって検出され得る。また、ｅＧＦＲの時間変化（ｅＧＦＲ　Ｓｌｏｐｅ）は、腎疾患の悪化速度の指標として使用される。ｅＧＦＲの時間変化が少ないと、腎機能が安定しているとされる。また、ｅＧＦＲが時間に伴って低下すると、腎疾患が悪化しているとされる。 There is an estimated glomerular filtration rate (eGFR) as an index for investigating the function of the kidney. eGFR is calculated by age, gender, and serum creatinine levels. Serum creatinine levels can be detected by blood tests. In addition, the time change of eGFR (egFR Slope) is used as an index of the exacerbation rate of renal disease. When the time change of eGFR is small, renal function is considered to be stable. In addition, when eGFR decreases with time, renal disease is said to be exacerbated.
特開2019-204484号公報Japanese Unexamined Patent Publication No. 2019-204484 特表2017-502439号公報Special Table 2017-502439 Gazette
　ｅＧＦＲ　Ｓｌｏｐｅは、例えば、ある時点のｅＧＦＲと、ある時点から数か月から数年後のｅＧＦＲとの変化率（時間あたりのｅＧＦＲの変化量）から求められる。しかし、ｅＧＦＲ　Ｓｌｏｐｅを求めるには、時間をおいて少なくとも２回、ｅＧＦＲを算出するため、ある時点で、ｅＧＦＲ　Ｓｌｏｐｅ（腎疾患の悪化速度）を求めることは困難である。また、腎臓の状態を非侵襲的に検出する臨床検査として、腎臓ＭＲ（Magnetic Resonance）画像がある。しかし、腎臓ＭＲ画像により、腎疾患の悪化速度を推定することは困難である。 The eGFR Slope is obtained from, for example, the rate of change between eGFR at a certain point in time and eGFR several months to several years after a certain point in time (the amount of change in eGFR per hour). However, in order to obtain eGFR Slope, it is difficult to obtain eGFR Slope (deterioration rate of renal disease) at a certain point in time because eGFR is calculated at least twice at intervals. In addition, as a clinical test for non-invasively detecting the state of the kidney, there is a kidney MR (Magnetic Resonance) image. However, it is difficult to estimate the rate of exacerbation of renal disease from renal MR images.
　本発明は、腎臓ＭＲ画像に基づいて、腎疾患の悪化速度を推定できる技術を提供することを目的とする。 An object of the present invention is to provide a technique capable of estimating the rate of exacerbation of renal disease based on renal MR images.
　上記課題を解決するために、以下の手段を採用する。
　即ち、第１の態様は、
　被検者の腎臓を含む部位を撮影した学習用ＭＲ（Magnetic Resonance）画像を取得し、学習用ＭＲ画像に含まれる腎臓の部分を抽出した腎臓の画像を生成し、当該腎臓の画像を非剛体変換により標準の腎臓の画像に変換し、変換された腎臓の画像に基づく学習用特徴量を算出し、
　算出された前記学習用特徴量と、前記被検者の腎疾患の悪化速度とを含むデータセットを教師データとして利用した機械学習によって、被検者の腎疾患の悪化速度を推定する推定器を構築するプロセッサ、
を備える推定装置とする。 In order to solve the above problems, the following means will be adopted.
That is, the first aspect is
A learning MR (Magnetic Resonance) image of the part including the kidney of the subject is acquired, an image of the kidney extracted from the part of the kidney included in the learning MR image is generated, and the image of the kidney is made into a non-rigid body. It is converted to a standard kidney image by conversion, and the learning feature amount based on the converted kidney image is calculated.
An estimator that estimates the deterioration rate of the kidney disease of the subject by machine learning using the data set including the calculated feature amount for learning and the deterioration rate of the kidney disease of the subject as teacher data. Processor to build,
It is an estimation device equipped with.
　開示の態様は、プログラムが情報処理装置によって実行されることによって実現されてもよい。即ち、開示の構成は、上記した態様における各手段が実行する処理を、情報処理装置に対して実行させるためのプログラム、或いは当該プログラムを記録したコンピュータ読み取り可能な記録媒体として特定することができる。また、開示の構成は、上記した各手段が実行する処理を情報処理装置が実行する方法をもって特定されてもよい。開示の構成は、上記した各手段が実行する処理を行う情報処理装置を含むシステムとして特定されてもよい。 The aspect of disclosure may be realized by executing the program by the information processing device. That is, the configuration of the disclosure can be specified as a program for causing the information processing apparatus to execute the process executed by each means in the above-described embodiment, or as a computer-readable recording medium on which the program is recorded. Further, the configuration of the disclosure may be specified by a method in which the information processing apparatus executes the processing executed by each of the above-mentioned means. The configuration of the disclosure may be specified as a system including an information processing apparatus that performs processing executed by each of the above-mentioned means.
　本発明によれば、腎臓ＭＲ画像に基づいて、腎疾患の悪化速度を推定できる技術を提供することができる。 According to the present invention, it is possible to provide a technique capable of estimating the exacerbation rate of renal disease based on renal MR images.
図１は、実施形態の推定装置の構成例を示す図である。FIG. 1 is a diagram showing a configuration example of an estimation device according to an embodiment. 図２は、Ｔ２＊ｍａｐ画像の例を示す図である。FIG. 2 is a diagram showing an example of a T2 * map image. 図３は、推定装置における推定器構築の際の動作フローの例を示す図である。FIG. 3 is a diagram showing an example of an operation flow when constructing an estimator in an estimator. 図４は、被検者の腎臓を含む部位のＭＲ画像である。FIG. 4 is an MR image of a site including the kidney of the subject. 図５は、図４のＭＲ画像から腎臓に相当する部分を抽出した画像である。FIG. 5 is an image obtained by extracting a portion corresponding to the kidney from the MR image of FIG. 図６は、腎臓の非剛体変換処理の例を示す図である。FIG. 6 is a diagram showing an example of the non-rigid body conversion process of the kidney. 図７は、特徴量算出処理によって生成されるヒストグラムの例を示す図である。FIG. 7 is a diagram showing an example of a histogram generated by the feature amount calculation process. 図８は、推定装置における腎疾患の悪化速度の推定の際の動作フローの例を示す図である。FIG. 8 is a diagram showing an example of an operation flow when estimating the exacerbation rate of renal disease in the estimation device.
　以下、図面を参照して実施形態について説明する。実施形態の構成は例示であり、発明の構成は、開示の実施形態の具体的構成に限定されない。発明の実施にあたって、実施形態に応じた具体的構成が適宜採用されてもよい。 Hereinafter, embodiments will be described with reference to the drawings. The configuration of the embodiment is an example, and the configuration of the invention is not limited to the specific configuration of the disclosed embodiment. In carrying out the invention, a specific configuration according to the embodiment may be appropriately adopted.
　〔実施形態〕
　（構成例）
　図１は、本実施形態の推定装置の構成例を示す図である。図１に示す推定装置１００は、一般的なコンピュータの構成を有している。本実施形態の推定装置１００は、プロセッサ１０１、メモリ１０２、記憶部１０３、入力部１０４、出力部１０５、通信制御部１０６を有する。これらは、互いにバスによって接続される。メモリ１０２及び記憶部１０３は、コンピュータ読み取り可能な記録媒体である。推定装置１００のハードウェア構成は、図１に示される例に限らず、適宜構成要素の省略、置換、追加が行われてもよい。 [Embodiment]
(Configuration example)
FIG. 1 is a diagram showing a configuration example of the estimation device of the present embodiment. The estimation device 100 shown in FIG. 1 has a general computer configuration. The estimation device 100 of the present embodiment includes a processor 101, a memory 102, a storage unit 103, an input unit 104, an output unit 105, and a communication control unit 106. These are connected to each other by a bus. The memory 102 and the storage unit 103 are computer-readable recording media. The hardware configuration of the estimation device 100 is not limited to the example shown in FIG. 1, and components may be omitted, replaced, or added as appropriate.
　本実施形態の推定装置１００は、被検者の臨床情報（性別、年齢、尿蛋白、血清クレアチニン値など）等と、被検者の腎臓を含む部位のＭＲ画像と、被検者のＭＲ画像の撮影後に算出された腎疾患の悪化速度（ｅＧＦＲ　ｓｌｏｐｅ）と取得する。推定装置１００は、ＭＲ画像に所定の処理を行い、被検者の臨床情報、ＭＲ画像から、腎疾患の悪化速度を推定する推定器を構築する。また、推定装置１００は、構築した推定器を用いて、被検者の臨床情報、ＭＲ画像から、腎疾患の悪化速度（ｅＧＦＲ　ｓｌｏｐｅ）を推定する。ＭＲ画像は、核磁気共鳴装置により身体の部位を撮影した画像である。腎臓のＭＲ画像は、腎臓の酸素状態、腎臓の線維化、腎臓の灌流量を反映する。腎臓の酸素状態、腎臓の線維化、腎臓の灌流量は、腎疾患と関係する。 The estimation device 100 of the present embodiment includes clinical information (gender, age, urinary protein, serum creatinine level, etc.) of the subject, an MR image of a site including the kidney of the subject, and an MR image of the subject. The rate of exacerbation of renal disease (eGFR slope) calculated after the imaging of the image is obtained. The estimation device 100 performs a predetermined process on the MR image, and constructs an estimator that estimates the exacerbation rate of the renal disease from the clinical information of the subject and the MR image. In addition, the estimation device 100 estimates the rate of deterioration of renal disease (eGFR slope) from clinical information and MR images of the subject using the constructed estimator. The MR image is an image of a part of the body taken by a nuclear magnetic resonance apparatus. MR images of the kidney reflect renal oxygen status, renal fibrosis, and renal perfusion. Kidney oxygen status, renal fibrosis, and renal perfusion are associated with renal disease.
　推定装置１００は、ワークステーション（ＷＳ、Work Station）のような専用または汎用のコンピュータ、ＰＣ（Personal Computer）、スマートフォン、タブレット型端末、あるいは、コンピュータを搭載した電子機器を使用して実現可能である。推定装置１００は、ネットワークを通じてサービスを提供するコンピュータ（サーバ機器）を使用して、実現可能である。推定装置１００は、ＣＰＵまたはＧＰＵを大規模に並列させたＭＰＩ（Message Passing Interface）による並列を実行する計算機によって実現可能である。推定装置１００は、プロセッサ１０１が記録媒体に記憶されたプログラムをメモリ１０２の作業領域にロードして実行し、プログラムの実行を通じて各構成部等が制御されることによって、所定の目的に合致した機能を実現することができる。 The estimation device 100 can be realized by using a dedicated or general-purpose computer such as a workstation (WS, WorkStation), a PC (Personal Computer), a smartphone, a tablet terminal, or an electronic device equipped with a computer. .. The estimation device 100 can be realized by using a computer (server device) that provides services through a network. The estimation device 100 can be realized by a computer that executes parallelization by MPI (Message Passing Interface) in which CPUs or GPUs are parallelized on a large scale. The estimation device 100 loads a program stored in the recording medium by the processor 101 into a work area of the memory 102 and executes the program, and each component or the like is controlled through the execution of the program, so that the function that meets a predetermined purpose is met. Can be realized.
　プロセッサ１０１は、例えば、ＣＰＵ（Central Processing Unit）などである。プロセッサ１０１は、メモリ１０２に記憶されたプログラムをロードし、実行することによって、画像取得処理、腎臓領域抽出処理、非剛体変換処理、特徴量算出処理、推定器構築処理、推定処理等を実行する。また、プロセッサ１０１は、各処理で使用されるデータ等を、記憶部１０３や、通信制御部１０６を介して他の装置から取得しうる。 The processor 101 is, for example, a CPU (Central Processing Unit) or the like. The processor 101 loads and executes a program stored in the memory 102 to execute an image acquisition process, a kidney region extraction process, a non-rigid body conversion process, a feature amount calculation process, an estimator construction process, an estimation process, and the like. .. Further, the processor 101 can acquire data and the like used in each process from other devices via the storage unit 103 and the communication control unit 106.
　画像取得処理は、他の情報処理装置などから、被検者の腎臓を含む部位を撮影したＭＲ画像（ＭＲＩ）を取得する処理である。また、画像取得処理では、ＭＲ画像とともに、被検者の臨床情報などの被検者の情報を取得する。被検者の臨床情報として、例えば、性別、年齢、平均血圧、尿蛋白量、尿酸値、血清クレアチニン値、ｅＧＦＲ、糖尿病歴の有無などが挙げられる。被検者の臨床情報として、他の情報が採用されてもよい。また、推定器構築の際に使用されるＭＲ画像に対応する被検者は、ｅＧＦＲ　Ｓｌｏｐｅ（腎疾患の悪化速度）を、当該ＭＲ画像を撮影された後（例えば、数か月から数年後）に算出されているものとする。即ち、当該被検者のＭＲ画像を撮影されてからのｅＧＦＲの時間変化に基づくｅＧＦＲ　Ｓｌｏｐｅが算出されているものとする。例えば、被検者は、ＭＲ画像を撮影されたときｅＧＦＲと、ＭＲ画像を撮影されてから１年後のｅＧＦＲとに基づいて、ｅＧＦＲ　Ｓｌｏｐｅ（腎疾患の悪化速度）を算出されているとする。ＭＲ画像として、例えば、Ｔ２＊ｍａｐ画像、Ｒ２＊ｍａｐ画像、ＡＤＣ（Apparent Diffusion Coefficient）ｍａｐ画像が使用され得る。ここで使用されるＭＲ画像は、これらに限定されるものではない。ここで使用されるＴ２＊ｍａｐ画像は、異なるエコー時間（echo time: ＴＥ）で撮影された１２枚のＴ２＊強調画像を元に作成される。腎門部から腎実質内部に入り込む脂肪成分の影響を抑制する為に、ＴＥ時間はすべてin-phaseに揃えて設定されている。Ｔ２＊ｍａｐ画像は、ここで説明したものに限定されるものではない。また、Ｒ２＊ｍａｐ画像は、Ｔ２＊ｍａｐ画像で使用されるＴ２＊値の逆数であるＲ２＊値（Ｒ２＊＝１／Ｔ２＊）に基づいて作成される画像である。 The image acquisition process is a process of acquiring an MR image (MRI) of a part including the kidney of a subject from another information processing device or the like. Further, in the image acquisition process, information on the subject such as clinical information on the subject is acquired together with the MR image. The clinical information of the subject includes, for example, gender, age, mean blood pressure, urinary protein level, uric acid level, serum creatinine level, eGFR, and the presence or absence of a history of diabetes. Other information may be adopted as the clinical information of the subject. In addition, the subject corresponding to the MR image used when constructing the estimator can perform eGFR Slope (deterioration rate of renal disease) after the MR image is taken (for example, several months to several years later). ). That is, it is assumed that the eGFR Slope is calculated based on the time change of the eGFR since the MR image of the subject was taken. For example, it is assumed that the subject calculates eGFR Slope (deterioration rate of renal disease) based on the eGFR when the MR image is taken and the eGFR one year after the MR image is taken. .. As the MR image, for example, a T2 * map image, an R2 * map image, and an ADC (Apparent Diffusion Coefficient) map image can be used. The MR image used here is not limited to these. The T2 * map image used here is created based on 12 T2 * emphasized images taken at different echo times (echo time: TE). All TE times are set in-phase in order to suppress the influence of the fat component that enters the renal parenchyma from the hilar region. The T2 * map image is not limited to the one described here. The R2 * map image is an image created based on the R2 * value (R2 * = 1 / T2 *), which is the reciprocal of the T2 * value used in the T2 * map image.
　図２は、Ｔ２＊ｍａｐ画像の例を示す図である。図２の左の画像は、正常の腎臓を含むＴ２＊ｍａｐ画像の例である。図２の右の画像は、腎臓機能が低下した慢性腎臓病の腎臓を含むＴ２＊ｍａｐ画像の例である。 FIG. 2 is a diagram showing an example of a T2 * map image. The image on the left in FIG. 2 is an example of a T2 * map image containing a normal kidney. The image on the right in FIG. 2 is an example of a T2 * map image containing a kidney with chronic kidney disease with impaired renal function.
　腎臓領域抽出処理は、画像取得処理で取得されたＭＲ画像から、腎臓に相当する部分を抽出する。腎臓領域抽出処理では、あらかじめ深層学習などにより構築されたＭＲ画像から腎臓部分を推定する推定器が使用される。腎臓領域抽出処理では、１人の被験者に複数のＭＲ画像が存在する場合には、抽出された腎臓の大きさが一番大きいＭＲ画像を以後の処理で使用するものとしてもよい。腎臓の大きさが一番大きいＭＲ画像は、腎臓の中心付近まで含まれている可能性が高く、推定器構築において、推定器の性能向上に寄与すると考えられる。 The kidney region extraction process extracts the portion corresponding to the kidney from the MR image acquired by the image acquisition process. In the kidney region extraction process, an estimator that estimates the kidney portion from MR images constructed in advance by deep learning or the like is used. In the kidney region extraction process, when a plurality of MR images are present in one subject, the MR image having the largest size of the extracted kidney may be used in the subsequent process. The MR image with the largest kidney size is likely to be included near the center of the kidney, which is thought to contribute to improving the performance of the estimator in constructing the estimator.
　非剛体変換処理は、腎臓領域抽出処理で抽出された腎臓の形状を、標準の腎臓の形状に変換する処理である。ここでは、周知の非剛体変換処理が使用される。標準の腎臓の形状は、例えば、正常の腎臓の形状である。標準の腎臓の形状に変換することで、推定器構築において、推定器の性能が向上する。 The non-rigid body conversion process is a process of converting the shape of the kidney extracted by the kidney region extraction process into a standard kidney shape. Here, a well-known non-rigid conversion process is used. The standard kidney shape is, for example, the normal kidney shape. Converting to a standard kidney shape improves the performance of the estimator in estimator construction.
　特徴量算出処理は、非剛体変換処理で標準の腎臓の形状に変換された腎臓の画像から、腎臓の特徴量を抽出する処理である。例えば、腎臓を皮質から髄質までの１２層に分割し、各層に含まれる画素の画素値を集計して、特徴量とする。画素値の集計として、例えば、層ごとに、画素の画素値の頻度、画素の画素値の平均を求めることなどが挙げられる。 The feature amount calculation process is a process of extracting the feature amount of the kidney from the image of the kidney converted into the standard kidney shape by the non-rigid body conversion process. For example, the kidney is divided into 12 layers from the cortex to the medulla, and the pixel values of the pixels included in each layer are aggregated to obtain a feature amount. As the aggregation of the pixel values, for example, the frequency of the pixel values of the pixels and the average of the pixel values of the pixels can be obtained for each layer.
　推定器構築処理は、被検者の臨床情報と、被検者のＭＲ画像から求められた腎臓の特徴量と、被検者の腎疾患の悪化速度とを含むデータセットを教師データとして、被検者の臨床情報と腎臓の特徴量とから被検者の腎疾患の悪化速度を推定する推定器を構築する処理である。 The estimator construction process uses a data set including clinical information of the subject, the feature amount of the kidney obtained from the MR image of the subject, and the rate of deterioration of the renal disease of the subject as teacher data. This is a process for constructing an estimator that estimates the rate of deterioration of renal disease in a subject from clinical information of the examiner and the feature amount of the kidney.
　推定処理は、推定器構築処理で構築された推定器を用いて、被検者の臨床情報と、被検者のＭＲ画像から求められた腎臓の特徴量とから、被検者の腎疾患の悪化速度を推定する処理である。 The estimation process uses the estimator constructed by the estimator construction process to obtain the clinical information of the subject and the features of the kidney obtained from the MR image of the subject. This is a process for estimating the rate of deterioration.
　メモリ１０２は、例えば、ＲＡＭ（Random Access Memory）、ＲＡＭ及びＲＯＭ（Read Only Memory）によって構成される。メモリ１０２は、主記憶装置とも呼ばれる。 The memory 102 is composed of, for example, a RAM (RandomAccessMemory), a RAM, and a ROM (ReadOnlyMemory). The memory 102 is also called a main storage device.
　記憶部１０３は、例えば、ＥＰＲＯＭ（Erasable Programmable ROM）、ハードディスクドライブ（ＨＤＤ、Hard Disk Drive）などである。また、記憶部１０３は、リムーバブルメディア、即ち可搬記録媒体を含むことができる。リムーバブルメディアは、例えば、ＵＳＢ（Universal Serial Bus）メモリ、あるいは、ＣＤ（Compact Disc）やＤＶＤ（Digital Versatile Disc）のようなディスク記録媒体である。記憶部１０３は、二次記憶装置とも呼ばれる。 The storage unit 103 is, for example, an EPROM (ErasableProgrammableROM), a hard disk drive (HDD, HardDiskDrive), or the like. Further, the storage unit 103 can include a removable medium, that is, a portable recording medium. The removable media is, for example, a USB (Universal Serial Bus) memory or a disc recording medium such as a CD (Compact Disc) or a DVD (Digital Versatile Disc). The storage unit 103 is also called a secondary storage device.
　記憶部１０３は、推定装置１００で使用される、各種のプログラム、各種のデータ及び各種のテーブルを読み書き自在に記録媒体に格納する。記憶部１０３には、オペレーティングシステム（Operating System :ＯＳ）、各種のアプリケーションプログラム、各種テーブル等が格納される。記憶部１０３に格納される情報は、メモリ１０２に格納されてもよい。また、メモリ１０２に格納される情報は、記憶部１０３に格納されてもよい。 The storage unit 103 stores various programs, various data, and various tables used in the estimation device 100 in a readable / writable recording medium. The storage unit 103 stores an operating system (Operating System: OS), various application programs, various tables, and the like. The information stored in the storage unit 103 may be stored in the memory 102. Further, the information stored in the memory 102 may be stored in the storage unit 103.
　記憶部１０３には、画像取得処理、腎臓領域抽出処理、非剛体変換処理、特徴量算出処理、推定器構築処理、推定処理等を実行するためのプログラムがインストールされている。また、記憶部１０３は、推定装置１００における各処理等で使用される各種データ等を格納する。記憶部１０３は、被検者のＭＲ画像、臨床情報、ｅＧＦＲ　Ｓｌｏｐｅ等が格納されている。 A program for executing image acquisition processing, kidney region extraction processing, non-rigid body conversion processing, feature amount calculation processing, estimator construction processing, estimation processing, etc. is installed in the storage unit 103. Further, the storage unit 103 stores various data and the like used in each process and the like in the estimation device 100. The storage unit 103 stores MR images of subjects, clinical information, eGFR Slope, and the like.
　オペレーティングシステムは、ソフトウェアとハードウェアとの仲介、メモリ空間の管理、ファイル管理、プロセスやタスクの管理等を行うソフトウェアである。オペレーティングシステムは、通信インタフェースを含む。通信インタフェースは、通信制御部１０６を介して接続される他の外部装置等とデータのやり取りを行うプログラムである。外部装置等には、例えば、他の情報処理装置、外部記憶装置等が含まれる。 The operating system is software that mediates between software and hardware, manages memory space, manages files, manages processes and tasks, and so on. The operating system includes a communication interface. The communication interface is a program for exchanging data with other external devices and the like connected via the communication control unit 106. The external device and the like include, for example, another information processing device, an external storage device, and the like.
　入力部１０４は、キーボード、ポインティングデバイス、ワイヤレスリモコン、タッチパネル等を含む。また、入力部１０４は、カメラのような映像や画像の入力装置や、マイクロフォンのような音声の入力装置を含むことができる。 The input unit 104 includes a keyboard, a pointing device, a wireless remote controller, a touch panel, and the like. Further, the input unit 104 can include a video or image input device such as a camera, or an audio input device such as a microphone.
　出力部１０５は、ＬＣＤ（Liquid Crystal Display）、ＥＬ（Electroluminescence）パネル、ＣＲＴ（Cathode Ray Tube）ディスプレイ、ＰＤＰ（Plasma Display Panel）等の表示装置、プリンタ等の出力装置を含む。また、出力部１０５は、スピーカのような音声の出力装置を含むことができる。 The output unit 105 includes a display device such as an LCD (Liquid Crystal Display), an EL (Electroluminescence) panel, a CRT (Cathode Ray Tube) display, a PDP (Plasma Display Panel), and an output device such as a printer. Further, the output unit 105 can include an audio output device such as a speaker.
　通信制御部１０６は、他の装置と接続し、推定装置１００と他の装置との間の通信を制御する。通信制御部１０６は、例えば、ＬＡＮ（Local Area Network）インタフェースボード、無線通信のための無線通信回路、有線通信のための通信回路である。ＬＡＮインタフェースボードや無線通信回路は、インターネット等のネットワークに接続される。 The communication control unit 106 is connected to another device and controls communication between the estimation device 100 and the other device. The communication control unit 106 is, for example, a LAN (Local Area Network) interface board, a wireless communication circuit for wireless communication, and a communication circuit for wired communication. The LAN interface board and the wireless communication circuit are connected to a network such as the Internet.
　プログラムを記述するステップは、記載された順序に沿って時系列的に行われる処理はもちろん、必ずしも時系列的に処理されなくても、並列的または個別に実行される処理を含む。プログラムを記述するステップの一部が省略されてもよい。 The step of writing a program includes not only the processes performed in chronological order in the described order but also the processes executed in parallel or individually even if they are not necessarily processed in chronological order. Some of the steps in writing the program may be omitted.
　本実施形態において、プロセッサ１０１によって実行される一連の処理は、ハードウェアにより実行させることも、ソフトウェアにより実行させることもできる。ハードウェアの構成要素は、ハードウェア回路であり、例えば、ＦＰＧＡ（Field Programmable Gate Array）、特定用途向け集積回路（ＡＳＩＣ：Application Specific Integrated Circuit）、ゲートアレイ、論理ゲートの組み合わせ、アナログ回路等がある。 In the present embodiment, the series of processes executed by the processor 101 can be executed by hardware or software. Hardware components are hardware circuits, such as FPGAs (Field Programmable Gate Arrays), application-specific integrated circuits (ASICs), gate arrays, logic gate combinations, analog circuits, and the like. ..
　（動作例）
　〈推定器構築〉
　図３は、推定装置における推定器構築の際の動作フローの例を示す図である。推定装置１００は、被検者の臨床情報、被検者のＭＲ画像に基づく特徴量、被検者の腎疾患の悪化速度を教師データとして、臨床情報、特徴量から、腎疾患の悪化速度を推定する推定器を構築する。 (Operation example)
<Estimator construction>
FIG. 3 is a diagram showing an example of an operation flow when constructing an estimator in an estimator. The estimation device 100 uses the clinical information of the subject, the feature amount based on the MR image of the subject, and the deterioration rate of the renal disease of the subject as teacher data, and obtains the deterioration rate of the renal disease from the clinical information and the feature amount. Build an estimator to estimate.
　Ｓ１０１では、推定装置１００のプロセッサ１０１は、推定器構築に使用される教師データの基となる被検者のＭＲ画像等を取得する。プロセッサ１０１は、記憶部１０３に格納される、被検者のＭＲ画像を取得する。ＭＲ画像は、被検者の腎臓を含む部位を撮影した画像である。また、プロセッサ１０１は、被検者のＭＲ画像とともに、当該被検者の臨床情報などの被検者の情報を、記憶部１０３から取得する。また、プロセッサ１０１は、記憶部１０３に格納される、被検者の腎疾患の悪化速度（ｅＧＦＲ　Ｓｌｏｐｅ）を取得する。当該被検者は、腎疾患の悪化速度を、当該ＭＲ画像を撮影された後（例えば、数か月から数年後）に算出されているとする。被検者の腎疾患の悪化速度は、あらかじめ、記憶部１０３に格納されているものとする。プロセッサ１０１は、これらのデータを、通信制御部１０６、ネットワーク等を介して、他の装置から取得してもよい。ここで取得されるＭＲ画像、臨床情報、腎疾患の悪化速度は、学習用のＭＲ画像、臨床情報、腎疾患の悪化速度である。 In S101, the processor 101 of the estimation device 100 acquires an MR image or the like of the subject, which is the basis of the teacher data used for constructing the estimator. The processor 101 acquires the MR image of the subject stored in the storage unit 103. The MR image is an image taken of a site including the kidney of the subject. Further, the processor 101 acquires the information of the subject such as the clinical information of the subject together with the MR image of the subject from the storage unit 103. In addition, the processor 101 acquires the exacerbation rate (eGFR Slope) of the renal disease of the subject, which is stored in the storage unit 103. It is assumed that the subject calculates the rate of exacerbation of renal disease after the MR image is taken (for example, months to years later). It is assumed that the exacerbation rate of the renal disease of the subject is stored in the storage unit 103 in advance. The processor 101 may acquire these data from other devices via the communication control unit 106, the network, and the like. The MR image, clinical information, and deterioration rate of renal disease acquired here are MR images for learning, clinical information, and deterioration rate of renal disease.
　Ｓ１０２では、プロセッサ１０１は、Ｓ１０１で取得されたＭＲ画像から、腎臓に相当する部分を抽出する。プロセッサ１０１は、あらかじめ深層学習などにより構築されたＭＲ画像から腎臓部分を推定する推定器を使用して、ＭＲ画像から腎臓に相当する部分を抽出する。当該推定器は、あらかじめ、記憶部１０３に格納されている。プロセッサ１０１は、抽出したＭＲ画像から腎臓に相当する部分（領域）の画像（腎臓の画像）を記憶部１０３に格納する。また、プロセッサ１０１は、ＭＲ画像と当該画像に含まれる腎臓に相当する部分（領域）の情報とを対応付けて、記憶部１０３に格納してもよい。プロセッサ１０１は、１人の被験者に複数のＭＲ画像が存在する場合には、それぞれのＭＲ画像に対して、腎臓に相当する部分を抽出する。さらに、プロセッサ１０１は、抽出した腎臓に相当する部分が最も大きいＭＲ画像を、以後の処理で使用する。腎臓に相当する部分の抽出は他の方法により行われてもよい。腎臓に相当する部分の抽出は、例えば、利用者によって腎臓部分の領域を入力部１０４により指定することにより行われてもよい。 In S102, the processor 101 extracts a portion corresponding to the kidney from the MR image acquired in S101. The processor 101 uses an estimator that estimates the kidney portion from the MR image constructed in advance by deep learning or the like, and extracts the portion corresponding to the kidney from the MR image. The estimator is stored in the storage unit 103 in advance. The processor 101 stores an image (image of the kidney) of a portion (region) corresponding to the kidney from the extracted MR image in the storage unit 103. Further, the processor 101 may store the MR image and the information of the portion (region) corresponding to the kidney included in the image in the storage unit 103 in association with each other. When a plurality of MR images are present in one subject, the processor 101 extracts a portion corresponding to the kidney for each MR image. Further, the processor 101 uses the MR image having the largest portion corresponding to the extracted kidney in the subsequent processing. Extraction of the portion corresponding to the kidney may be performed by another method. Extraction of the portion corresponding to the kidney may be performed, for example, by designating the region of the kidney portion by the input unit 104 by the user.
　図４は、被検者の腎臓を含む部位のＭＲ画像である。図４のＭＲ画像は、人の腹部を正面から見た画像である。図４のＭＲ画像では、左右に腎臓が現れている。 FIG. 4 is an MR image of the site including the kidney of the subject. The MR image of FIG. 4 is an image of a person's abdomen viewed from the front. In the MR image of FIG. 4, kidneys appear on the left and right.
　図５は、図４のＭＲ画像から腎臓に相当する部分を抽出した画像である。図５では、図４のＭＲ画像から腎臓部分を推定する推定器により、腎臓（右の腎臓及び左の腎臓）に相当する部分を抽出されている。 FIG. 5 is an image obtained by extracting a portion corresponding to the kidney from the MR image of FIG. In FIG. 5, a portion corresponding to a kidney (right kidney and left kidney) is extracted by an estimator that estimates the kidney portion from the MR image of FIG.
　Ｓ１０３では、プロセッサ１０１は、Ｓ１０２で抽出された腎臓に相当する部分の画像を、標準の腎臓の形状に、非剛体変換処理により、変換する。非剛体変換処理として、周知の方法が使用される。ここでは、非剛体変換処理は、Ｓ１０２で抽出された腎臓に相当する部分の画像の腎臓の形状を、標準の腎臓の形状に変換する処理である。標準の腎臓の形状は、例えば、正常の腎臓の形状である。腎臓の形状は腎疾患等の影響により、変形、萎縮することがある。非剛体変換処理により、標準の腎臓の形状に変換することで、腎臓の形状が揃えられ、変形、萎縮等の影響が低減される。さらに、非剛体変換処理により、腎臓の形状が揃うことで、推定器の性能が向上する。プロセッサ１０１は、非剛体変換処理により変換した腎臓の画像を、記憶部１０３に格納する。 In S103, the processor 101 converts the image of the portion corresponding to the kidney extracted in S102 into a standard kidney shape by a non-rigid body conversion process. A well-known method is used as the non-rigid body conversion process. Here, the non-rigid body conversion process is a process of converting the shape of the kidney in the image of the portion corresponding to the kidney extracted in S102 into the shape of a standard kidney. The standard kidney shape is, for example, the normal kidney shape. The shape of the kidney may be deformed or atrophied due to the influence of renal disease or the like. By converting to a standard kidney shape by the non-rigid body conversion process, the shape of the kidney is aligned and the effects of deformation, atrophy, etc. are reduced. In addition, the non-rigid conversion process improves the performance of the estimator by aligning the shape of the kidneys. The processor 101 stores the image of the kidney converted by the non-rigid body conversion process in the storage unit 103.
　図６は、腎臓の非剛体変換処理の例を示す図である。図６の例では、第１の被検者の腎臓、第２の被検者の腎臓を、標準の腎臓に変形する例を示す。図６の第１の被検者の腎臓における点Ａ１、点Ｂ１、点Ｃ１は、それぞれ、標準の腎臓における点Ａ０、点Ｂ０、点Ｃ０に対応する。また、図６の第２の被検者の腎臓における点Ａ２、点Ｂ２、点Ｃ２は、それぞれ、標準の腎臓における点Ａ０、点Ｂ０、点Ｃ０に対応する。非剛体変換処理により、変換前の腎臓の画像内の各画素（各点）は、標準の腎臓の画像内のいずれかの画素（点）に対応付けられる。変換の前後において、対応する各画素（各点）の画素値（色、濃淡）は、維持される。例えば、点Ａ０、点Ａ１、点Ａ２は、互いに対応する点である。即ち、例えば、標準の腎臓における点Ａ０の画素値に対応する第１被検者の腎臓の画素値は、第１被検者の腎臓の画像における点Ａ１の画素値となる。 FIG. 6 is a diagram showing an example of the non-rigid body conversion process of the kidney. The example of FIG. 6 shows an example in which the kidney of the first subject and the kidney of the second subject are transformed into a standard kidney. Points A1, B1, and C1 in the kidney of the first subject in FIG. 6 correspond to points A0, B0, and C0 in the standard kidney, respectively. In addition, points A2, B2, and C2 in the kidney of the second subject in FIG. 6 correspond to points A0, B0, and C0 in the standard kidney, respectively. By the non-rigid body conversion process, each pixel (point) in the image of the kidney before conversion is associated with any pixel (point) in the image of the standard kidney. Before and after the conversion, the pixel value (color, shading) of each corresponding pixel (each point) is maintained. For example, points A0, A1 and A2 are points corresponding to each other. That is, for example, the pixel value of the kidney of the first subject corresponding to the pixel value of the point A0 in the standard kidney is the pixel value of the point A1 in the image of the kidney of the first subject.
　Ｓ１０４では、プロセッサ１０１は、Ｓ１０３で変換された標準の形状の腎臓の画像から、腎臓の特徴量を算出する。プロセッサ１０１は、例えば、腎臓を皮質から髄質までの複数の層（例えば、１２層）に分割し、各層に含まれる画素の画素値を集計して、特徴量として算出する。皮質から髄質までの複数の層として、例えば、あらかじめ、正常の腎臓に対して、専門医などにより複数の層に分割されているものが使用される。皮質から髄質までの複数の層に分割することで、腎臓の内部の部位に応じた特徴量を算出することができる。各層の特徴量は、例えば、各層の画素値の出現頻度を表すヒストグラム、各層の画素値の平均値等である。また、ここでは、層毎に画素値を集計しているが、複数の層に分割せずに、腎臓の画像に含まれる画素の画素値を集計してもよい。即ち、プロセッサ１０１は、腎臓の画像に含まれるすべての画素の画素値の出現頻度を表すヒストグラムを特徴量としたり、すべての画素の画素値の平均値を特徴量としたりしてもよい。プロセッサ１０１は、腎臓の画像の１以上の特定の画素（位置）の画素値を特徴量としてもよい。腎臓の特定の位置は、例えば、医学的に腎疾患の悪化速度と相関が高いと考えられる位置である。当該位置の指定は、腎臓の形状を標準の形状に揃えていることによって、しやすくなっている。プロセッサ１０１は、算出した特徴量を記憶部１０３に格納する。特徴量は、被検者の臨床情報及び腎疾患の悪化速度と対応付けられて格納される。特徴量として、腎臓の画像に基づく他の量が採用されてもよい。ここに挙げた、層毎の画素値のヒストグラム、層毎の画素値の平均値、腎臓の画像の画素値のヒストグラム、腎臓の画像の画素値の平均値、腎臓の画像の１以上の特定の画素（位置）の画素値などのうち、特徴量として複数の量が採用されてもよい。ここで算出される特徴量は、ＭＲ画像に含まれる腎臓の特徴を示す量である。ここで算出される特徴量は、学習用の特徴量である。 In S104, the processor 101 calculates the feature amount of the kidney from the image of the kidney of the standard shape converted in S103. For example, the processor 101 divides the kidney into a plurality of layers (for example, 12 layers) from the cortex to the medulla, aggregates the pixel values of the pixels included in each layer, and calculates the feature amount. As a plurality of layers from the cortex to the medulla, for example, a normal kidney that has been previously divided into a plurality of layers by a specialist or the like is used. By dividing into a plurality of layers from the cortex to the medulla, it is possible to calculate the feature amount according to the internal part of the kidney. The feature amount of each layer is, for example, a histogram showing the appearance frequency of the pixel values of each layer, an average value of the pixel values of each layer, and the like. Further, although the pixel values are aggregated for each layer here, the pixel values of the pixels included in the kidney image may be aggregated without being divided into a plurality of layers. That is, the processor 101 may use a histogram showing the appearance frequency of the pixel values of all the pixels included in the image of the kidney as the feature amount, or may use the average value of the pixel values of all the pixels as the feature amount. The processor 101 may use the pixel value of one or more specific pixels (positions) of the image of the kidney as a feature amount. A particular location of the kidney is, for example, a location that is medically considered to be highly correlated with the rate of exacerbation of renal disease. The designation of the position is facilitated by aligning the shape of the kidney with the standard shape. The processor 101 stores the calculated feature amount in the storage unit 103. The feature amount is stored in association with the clinical information of the subject and the rate of deterioration of renal disease. As the feature amount, another amount based on the image of the kidney may be adopted. Listed here, a histogram of pixel values for each layer, an average value of pixel values for each layer, a histogram of pixel values for kidney images, an average pixel value for kidney images, one or more specific values for kidney images. Of the pixel values of pixels (positions), a plurality of quantities may be adopted as feature quantities. The feature amount calculated here is an amount indicating the feature of the kidney included in the MR image. The feature amount calculated here is a feature amount for learning.
　図７は、特徴量算出処理によって生成されるヒストグラムの例を示す図である。図７のヒストグラムの横軸は画素値、縦軸は各画素値の出現頻度（出現回数）を示す。ヒストグラムは、各層の画像に含まれる画素の画素値の出現頻度（出現回数）を表したものである。ここでは、縦軸は画素値の出現頻度（出現回数）としているが、画素値の出現割合であってもよい。ここでは、画素値は、０から２５５までの値をとるとする。 FIG. 7 is a diagram showing an example of a histogram generated by the feature amount calculation process. The horizontal axis of the histogram in FIG. 7 indicates the pixel value, and the vertical axis indicates the appearance frequency (number of appearances) of each pixel value. The histogram shows the appearance frequency (number of appearances) of the pixel values of the pixels included in the image of each layer. Here, the vertical axis represents the appearance frequency (number of appearances) of the pixel value, but it may be the appearance ratio of the pixel value. Here, it is assumed that the pixel value takes a value from 0 to 255.
　プロセッサ１０１は、様々な被検者のＭＲ画像等を用いて、Ｓ１０１からＳ１０４までの処理を行い、各ＭＲ画像に対する特徴量を算出し、記憶部１０３に格納する。 The processor 101 performs processing from S101 to S104 using various MR images of the subject, calculates a feature amount for each MR image, and stores it in the storage unit 103.
　Ｓ１０５では、プロセッサ１０１は、複数の被検者についての、Ｓ１０４で算出した特徴量、Ｓ１０１で取得した被検者の臨床情報、被検者の腎疾患の悪化速度（ｅＧＦＲ　Ｓｌｏｐｅ）を含む複数のデータセットを教師データとして、機械学習の深層学習モデルを使用して、特徴量、臨床情報から腎疾患の悪化速度（ｅＧＦＲ　Ｓｌｏｐｅ）を推定する推定器を構築する。ここで使用される深層学習モデルは、どのようなモデルが使用されてもよい。プロセッサ１０１は、構築した推定器を記憶部１０３に格納する。推定器の構築には、ニューラルネットワークによるディープラーニング、Regression SVM、Regression Random Forest、多重回帰分析、Look Up Table等の学習空間を利用する手法等が使用され得る。推定器の構築の際に、機械学習以外の方法が使用されてもよい。より多くの教師データを使用することで、より性能の高い推定器を構築することができる。ここで、教師データとするデータセットには、臨床情報の一部または全部が含まれなくてもよい。即ち、Ｓ１０４で算出した特徴量、腎疾患の悪化速度を含む複数のデータセットを教師データとして、推定器を構築してもよい。データセットに臨床情報の一部または全部が含まれない場合、Ｓ１０１において、含まれない臨床情報を取得しなくてもよい。 In S105, the processor 101 includes a plurality of features calculated in S104 for a plurality of subjects, clinical information of the subject acquired in S101, and a plurality of exacerbation rates of renal disease (eGFR EGFR) of the subject. Using the data set as teacher data, a deep learning model of machine learning is used to construct an estimator that estimates the rate of deterioration of renal disease (eGFR Slope) from features and clinical information. Any model may be used as the deep learning model used here. The processor 101 stores the constructed estimator in the storage unit 103. For the construction of the estimator, a method using a learning space such as deep learning by a neural network, Regression SVM, Regression Random Forest, multiple regression analysis, Look Up Table, etc. can be used. Methods other than machine learning may be used in the construction of the estimator. By using more teacher data, it is possible to build a higher performance estimator. Here, the data set as the teacher data may not include a part or all of the clinical information. That is, an estimator may be constructed using a plurality of data sets including the feature amount calculated in S104 and the rate of deterioration of renal disease as teacher data. If the data set does not include part or all of the clinical information, it is not necessary to acquire the clinical information that is not included in S101.
　〈腎疾患の悪化速度の推定〉
　図８は、推定装置における腎疾患の悪化速度の推定の際の動作フローの例を示す図である。推定装置１００は、推定対象となる被検者の臨床情報、被検者のＭＲ画像に基づく特徴量、図３の動作フローで構築した推定器を使用して、被検者の腎疾患の悪化速度を推定する。 <Estimation of the rate of deterioration of renal disease>
FIG. 8 is a diagram showing an example of an operation flow when estimating the exacerbation rate of renal disease in the estimation device. The estimation device 100 uses the clinical information of the subject to be estimated, the feature amount based on the MR image of the subject, and the estimator constructed by the operation flow of FIG. 3, and worsens the renal disease of the subject. Estimate the speed.
　Ｓ２０１では、プロセッサ１０１は、記憶部１０３、もしくは、通信制御部１０６、ネットワーク等を介して他の装置から、推定対象となる被検者のＭＲ画像および臨床情報を取得する。図３の動作フローで推定器を構築した際の教師データとするデータセットに臨床情報の一部または全部が含まれない場合、プロセッサ１０１は、含まれない臨床情報を取得しなくてもよい。 In S201, the processor 101 acquires MR images and clinical information of the subject to be estimated from other devices via the storage unit 103, the communication control unit 106, the network, or the like. If the data set used as the teacher data when the estimator is constructed in the operation flow of FIG. 3 does not include a part or all of the clinical information, the processor 101 does not have to acquire the clinical information that is not included.
　Ｓ２０２からＳ２０４まで処理は、図３のＳ１０２からＳ１０４までの処理と同様である。プロセッサ１０１は、Ｓ２０２からＳ２０４まで処理により、推定対象となる被検者のＭＲ画像に基づく特徴量を算出し、記憶部１０３に格納する。プロセッサ１０１がここで算出する特徴量は、図３の動作フローで推定器を構築した際に使用した特徴量（教師データに含まれる特徴量）と同じ種類の特徴量である。即ち、例えば、推定器構築の際に特徴量を腎臓の画像の層毎の画素値のヒストグラムとした場合、ここで算出される特徴量は、推定対象の被検者の腎臓の画像の層毎の画素値のヒストグラムである。 The processing from S202 to S204 is the same as the processing from S102 to S104 in FIG. The processor 101 calculates the feature amount based on the MR image of the subject to be estimated by processing from S202 to S204, and stores it in the storage unit 103. The feature amount calculated here by the processor 101 is the same type of feature amount as the feature amount (feature amount included in the teacher data) used when the estimator was constructed in the operation flow of FIG. That is, for example, when the feature amount is used as a histogram of the pixel values of each layer of the kidney image when constructing the estimator, the feature amount calculated here is for each layer of the image of the kidney of the subject to be estimated. It is a histogram of the pixel value of.
　Ｓ２０５では、プロセッサ１０１は、図３の動作フローで構築された腎疾患の悪化速度を推定する推定器を使用して、Ｓ２０１で取得した臨床情報、Ｓ２０４で算出した特徴量に基づいて、腎疾患の悪化速度（ｅＧＦＲ　Ｓｌｏｐｅ）を推定する。プロセッサ１０１は、推定した腎疾患の悪化速度を推定対象の被検者の識別する情報に対応づけて、記憶部１０３に格納する。これにより、推定装置１００は、推定対象の被検者のＭＲ画像等により、腎疾患の悪化速度（ｅＧＦＲ　Ｓｌｏｐｅ）を推定することができる。また、Ｓ２０１で臨床情報の一部または全部を取得していない場合、プロセッサ１０１は、取得していない臨床情報を使用せずに、腎疾患の悪化速度（ｅＧＦＲ　Ｓｌｏｐｅ）を推定する。 In S205, the processor 101 uses the estimator for estimating the exacerbation rate of the renal disease constructed by the operation flow of FIG. 3, and based on the clinical information acquired in S201 and the feature amount calculated in S204, the renal disease. Estimate the rate of deterioration (eGFR Slope). The processor 101 stores the estimated exacerbation rate of the renal disease in the storage unit 103 in association with the information identifying the subject to be estimated. As a result, the estimation device 100 can estimate the exacerbation rate of renal disease (eGFR Slope) from the MR image or the like of the subject to be estimated. Further, when a part or all of the clinical information is not acquired in S201, the processor 101 estimates the exacerbation rate of renal disease (eGFR Slope) without using the acquired clinical information.
　これにより、推定装置１００は、被検者の腎臓を含む部位のＭＲ画像、腎疾患の悪化速度を推定する推定器を用いて、腎疾患の悪化速度の推定を行うことができる。 Thereby, the estimation device 100 can estimate the deterioration rate of the kidney disease by using the MR image of the site including the kidney of the subject and the estimator for estimating the deterioration rate of the kidney disease.
　ここでは、腎臓を対象としているが、他の臓器等に対しても同様に適用することができる。 Here, the kidney is targeted, but it can be applied to other organs as well.
　（推定器による腎疾患の悪化速度の推定の性能）
　ここでは、実際の被検者１６４例を対象として、教師データに、（１）臨床情報（尿蛋白量）、（２）腎臓の画像の画素値の平均値、（３）腎臓の画像の層毎の画素値のヒストグラム（出現頻度）、（４）（２）及び（３）、（５）（１）、（２）及び（３）を使用した各推定器について、検討した。各推定器は、上記に基づいて構築される。このとき、各推定器が推定したｅＧＦＲ　ｓｌｏｐｅと、実際のｅＧＦＲ　ｓｌｏｐｅとの相関係数を算出した。その結果、各相関係数は、（１）０．５６、（２）０．６１、（３）０．７９、（４）０．８５、（５）０．８７であった。相関係数が高いほど、推定器の性能が高いことを示す。一般に、相関係数が０．７０以上であれば、有用とされる。この結果から、非画像情報である臨床情報（尿蛋白量）を教師データとした場合よりも、腎臓の画像（ＭＲ画像）に基づく特徴量を教師データとした場合に、推定器の性能が向上する。また、腎臓の画像の層毎の画素値のヒストグラムを教師データに用いた場合（（３）、（４）、（５））に、有用となる。また、教師データに、腎臓の画像の層毎の画素値のヒストグラムの他に、臨床情報、腎臓の画像の画素値の平均値を加えるとより推定器の性能が向上する。 (Performance of estimation of the rate of deterioration of renal disease by an estimator)
Here, for 164 actual subjects, the teacher data includes (1) clinical information (urinary protein amount), (2) average pixel value of kidney image, and (3) kidney image layer. Each estimator using the histogram (appearance frequency) of each pixel value, (4) (2) and (3), (5) (1), (2) and (3) was examined. Each estimator is constructed based on the above. At this time, the correlation coefficient between the eGFR slope estimated by each estimator and the actual eGFR slope was calculated. As a result, each correlation coefficient was (1) 0.56, (2) 0.61, (3) 0.79, (4) 0.85, and (5) 0.87. The higher the correlation coefficient, the higher the performance of the estimator. Generally, if the correlation coefficient is 0.70 or more, it is considered useful. From this result, the performance of the estimator is improved when the feature amount based on the kidney image (MR image) is used as the teacher data than when the clinical information (urinary protein amount) which is non-image information is used as the teacher data. do. Further, it is useful when the histogram of the pixel value for each layer of the image of the kidney is used for the teacher data ((3), (4), (5)). In addition to the histogram of the pixel values for each layer of the kidney image, clinical information and the average value of the pixel values of the kidney image are added to the teacher data to further improve the performance of the estimator.
　（実施形態の作用、効果）
　推定装置１００は、被検者の腎臓を含むＭＲ画像、被検者の臨床情報、被検者のｅＧＦＲ　ｓｌｏｐｅ（腎疾患の悪化速度）を取得する。推定装置１００は、取得されたＭＲ画像から、腎臓の画像を抽出する。推定装置１００は、抽出した腎臓の画像の腎臓の形状を標準の形状に非剛体変換してから、腎臓の画像に基づく特徴量を算出する。推定装置１００は、例えば、腎臓を皮質から髄質までの複数の層に分割し、層毎に画像の画素値の出現頻度を特徴量として、算出する。推定装置１００は、少なくとも、ＭＲ画像に基づく特徴量、ｅＧＦＲ　ｓｌｏｐｅを含む複数のデータセットを教師データとして、少なくとも、ＭＲ画像に基づく特徴量からｅＧＦＲ　ｓｌｏｐｅを推定する推定器を構築する。推定装置１００によれば、ＭＲ画像に基づく特徴量を用いることで、より性能の高い推定器を構築できる。また、被検者の腎臓の形状を標準の形状に非剛体変換することにより、腎臓の形状が揃うため、推定器の性能が向上する。また、推定装置１００によれば、腎臓の画像を皮質から髄質までの複数の層（例えば１２層）に分割することで、腎臓の内部の部位に応じた特徴量を算出することができる。腎臓の内部の部位に応じた特徴量を使用することで推定器の性能が向上する。 (Actions and effects of embodiments)
The estimation device 100 acquires an MR image including the kidney of the subject, clinical information of the subject, and eGFR slope (exacerbation rate of renal disease) of the subject. The estimation device 100 extracts an image of the kidney from the acquired MR image. The estimation device 100 non-rigidly transforms the kidney shape of the extracted kidney image into a standard shape, and then calculates the feature amount based on the kidney image. For example, the estimation device 100 divides the kidney into a plurality of layers from the cortex to the medulla, and calculates the appearance frequency of the pixel value of the image for each layer as a feature amount. The estimation device 100 constructs an estimator that estimates eGFR slope from at least the feature amount based on the MR image, using at least a plurality of data sets including the feature amount based on the MR image and the eGFR slope as teacher data. According to the estimation device 100, a higher performance estimator can be constructed by using the feature amount based on the MR image. In addition, by converting the shape of the kidney of the subject into a standard shape in a non-rigid body, the shape of the kidney becomes uniform, and the performance of the estimator is improved. Further, according to the estimation device 100, by dividing the image of the kidney into a plurality of layers (for example, 12 layers) from the cortex to the medulla, it is possible to calculate the feature amount according to the internal part of the kidney. The performance of the estimator is improved by using the features according to the internal part of the kidney.
　推定装置１００は、推定対象の被検者の腎臓を含むＭＲ画像を取得し、ＭＲ画像に基づく特徴量を算出する。推定装置１００は、算出した特徴量及び構築された推定器を用いて、推定対象の被検者の腎疾患の悪化速度を推定する。推定装置１００によれば、腎臓を含むＭＲ画像に基づく特徴量を使用することで、より精度よく腎疾患の悪化速度を推定することができる。 The estimation device 100 acquires an MR image including the kidney of the subject to be estimated, and calculates a feature amount based on the MR image. The estimation device 100 estimates the rate of deterioration of the renal disease of the subject to be estimated by using the calculated feature amount and the constructed estimator. According to the estimation device 100, the rate of deterioration of renal disease can be estimated more accurately by using the feature amount based on the MR image including the kidney.
　〈コンピュータ読み取り可能な記録媒体〉
　コンピュータその他の機械、装置（以下、コンピュータ等）に上記いずれかの機能を実現させるプログラムをコンピュータ等が読み取り可能な記録媒体に記録することができる。そして、コンピュータ等に、この記録媒体のプログラムを読み込ませて実行させることにより、その機能を提供させることができる。 <Computer readable recording medium>
A program that realizes any of the above functions in a computer or other machine or device (hereinafter referred to as a computer or the like) can be recorded on a recording medium that can be read by a computer or the like. Then, by having a computer or the like read and execute the program of this recording medium, the function can be provided.
　ここで、コンピュータ等が読み取り可能な記録媒体とは、データやプログラム等の情報を電気的、磁気的、光学的、機械的、または化学的作用によって蓄積し、コンピュータ等から読み取ることができる記録媒体をいう。このような記録媒体内には、ＣＰＵ、メモリ等のコンピュータを構成する要素を設け、そのＣＰＵにプログラムを実行させてもよい。 Here, a recording medium that can be read by a computer or the like is a recording medium that can store information such as data and programs by electrical, magnetic, optical, mechanical, or chemical action and can be read from a computer or the like. To say. In such a recording medium, elements constituting a computer such as a CPU and a memory may be provided, and the CPU may execute a program.
　また、このような記録媒体のうちコンピュータ等から取り外し可能なものとしては、例えばフレキシブルディスク、光磁気ディスク、ＣＤ－ＲＯＭ、ＣＤ－Ｒ／Ｗ、ＤＶＤ、ＤＡＴ、８mmテープ、メモリカード等がある。 Among such recording media, those that can be removed from a computer or the like include, for example, flexible disks, magneto-optical disks, CD-ROMs, CD-R / Ws, DVDs, DATs, 8 mm tapes, memory cards, and the like.
　また、コンピュータ等に固定された記録媒体としてハードディスクやＲＯＭ等がある。 In addition, there are hard disks, ROMs, etc. as recording media fixed to computers and the like.
　（その他）
　以上、本発明の実施形態を説明したが、これらはあくまで例示にすぎず、本発明はこれらに限定されるものではなく、特許請求の範囲の趣旨を逸脱しない限りにおいて、各構成の組み合わせなど、当業者の知識に基づく種々の変更が可能である。 (others)
Although the embodiments of the present invention have been described above, these are merely examples, and the present invention is not limited thereto, and as long as the gist of the claims is not deviated, combinations of each configuration and the like may be used. Various changes can be made based on the knowledge of those skilled in the art.
１００　　推定装置
１０１　　　プロセッサ
１０２　　　メモリ
１０３　　　記憶部
１０４　　　入力部
１０５　　　出力部
１０６　　　通信制御部
  100 Estimator 101 Processor 102 Memory 103 Storage unit 104 Input unit 105 Output unit 106 Communication control unit

Claims (5)

  1. 　被検者の腎臓を含む部位を撮影した学習用ＭＲ（Magnetic Resonance）画像を取得し、前記学習用ＭＲ画像に含まれる腎臓の部分を抽出した腎臓の画像を生成し、当該腎臓の画像を非剛体変換により標準の腎臓の画像に変換し、変換された腎臓の画像に基づく学習用特徴量を算出し、
    　算出された前記学習用特徴量と、前記被検者の腎疾患の悪化速度とを含むデータセットを教師データとして利用した機械学習によって、被検者の腎疾患の悪化速度を推定する推定器を構築するプロセッサ、
    を備える推定装置。     An MR (Magnetic Resonance) image for learning was obtained by photographing a part including the kidney of the subject, and an image of the kidney obtained by extracting the part of the kidney included in the MR image for learning was generated, and the image of the kidney was not used. It is converted to a standard kidney image by rigid body conversion, and the learning feature amount based on the converted kidney image is calculated.
    An estimator that estimates the deterioration rate of the kidney disease of the subject by machine learning using the data set including the calculated feature amount for learning and the deterioration rate of the kidney disease of the subject as teacher data. Processor to build,
    Estimator equipped with.
  2. 　前記プロセッサは、前記学習用特徴量を、前記変換された腎臓の画像を複数の層に分割した各層に含まれる画素の画素値の出現頻度として算出する、
    請求項１に記載の推定装置。     The processor calculates the learning feature amount as the frequency of appearance of pixel values of pixels included in each layer obtained by dividing the converted kidney image into a plurality of layers.
    The estimation device according to claim 1.
  3. 　前記プロセッサは、
    　推定対象の被検者の腎臓を含む部位を撮影したＭＲ画像を取得し、当該ＭＲ画像に含まれる腎臓の部分を抽出した腎臓の画像を生成し、前記腎臓の画像を非剛体変換により標準の腎臓の画像に変換し、変換された腎臓の画像に基づく特徴量を算出し、
    　前記特徴量について、前記推定器を使用して、前記推定対象の被検者の腎疾患の悪化速度を推定する、
    請求項１または２に記載の推定装置。     The processor
    An MR image of the part including the kidney of the subject to be estimated is acquired, an image of the kidney obtained by extracting the part of the kidney included in the MR image is generated, and the image of the kidney is standardized by non-rigid conversion. Convert to an image of the kidney, calculate the feature amount based on the converted image of the kidney,
    With respect to the feature amount, the rate of deterioration of the renal disease of the subject to be estimated is estimated by using the estimator.
    The estimation device according to claim 1 or 2.
  4. 　コンピュータが、
    　被検者の腎臓を含む部位を撮影した学習用ＭＲ画像を取得し、前記学習用ＭＲ画像に含まれる腎臓の部分を抽出した腎臓の画像を生成し、当該腎臓の画像を非剛体変換により標準の腎臓の画像に変換し、変換された腎臓の画像に基づく学習用特徴量を算出し、
    　算出された前記学習用特徴量と、前記被検者の腎疾患の悪化速度とを含むデータセットを教師データとして利用した機械学習によって、被検者の腎疾患の悪化速度を推定する推定器を構築する
    ことを実行する推定方法。     The computer
    A learning MR image of the part including the kidney of the subject is acquired, an image of the kidney obtained by extracting the part of the kidney included in the learning MR image is generated, and the image of the kidney is standardized by non-rigid conversion. Converted to an image of the kidney of the kidney, calculated the learning feature amount based on the converted image of the kidney,
    An estimator that estimates the rate of deterioration of the subject's kidney disease by machine learning using the data set including the calculated feature amount for learning and the rate of deterioration of the subject's kidney disease as teacher data. An estimation method to perform what you build.
  5. 　コンピュータが、
    　被検者の腎臓を含む部位を撮影した学習用ＭＲ画像を取得し、前記学習用ＭＲ画像に含まれる腎臓の部分を抽出した腎臓の画像を生成し、当該腎臓の画像を非剛体変換により標準の腎臓の画像に変換し、変換された腎臓の画像に基づく学習用特徴量を算出し、
    　算出された前記学習用特徴量と、前記被検者の腎疾患の悪化速度とを含むデータセットを教師データとして利用した機械学習によって、被検者の腎疾患の悪化速度を推定する推定器を構築する
    ことを実行するための推定プログラム。     The computer
    A learning MR image of the part including the kidney of the subject is acquired, an image of the kidney obtained by extracting the part of the kidney included in the learning MR image is generated, and the image of the kidney is standardized by non-rigid conversion. Converted to an image of the kidney of the kidney, calculated the learning feature amount based on the converted image of the kidney,
    An estimator for estimating the deterioration rate of the kidney disease of the subject by machine learning using the data set including the calculated feature amount for learning and the deterioration rate of the kidney disease of the subject as teacher data. An estimation program to perform what you build.
PCT/JP2021/037615 2020-10-13 2021-10-11 Estimator learning device, estimator learning method, and estimator learning program WO2022080327A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2022556974A JPWO2022080327A1 (en) 2020-10-13 2021-10-11

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2020172356 2020-10-13
JP2020-172356 2020-10-13

Publications (1)

Publication Number Publication Date
WO2022080327A1 true WO2022080327A1 (en) 2022-04-21

Family

ID=81209137

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2021/037615 WO2022080327A1 (en) 2020-10-13 2021-10-11 Estimator learning device, estimator learning method, and estimator learning program

Country Status (2)

Country Link
JP (1) JPWO2022080327A1 (en)
WO (1) WO2022080327A1 (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150286786A1 (en) * 2014-04-02 2015-10-08 University Of Louisville Research Foundation, Inc. Computer aided diagnostic system for classifying kidneys
US20160166209A1 (en) * 2014-12-16 2016-06-16 Siemens Healthcare Gmbh Method and System for Personalized Non-Invasive Hemodynamic Assessment of Renal Artery Stenosis from Medical Images

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150286786A1 (en) * 2014-04-02 2015-10-08 University Of Louisville Research Foundation, Inc. Computer aided diagnostic system for classifying kidneys
US20160166209A1 (en) * 2014-12-16 2016-06-16 Siemens Healthcare Gmbh Method and System for Personalized Non-Invasive Hemodynamic Assessment of Renal Artery Stenosis from Medical Images

Also Published As

Publication number Publication date
JPWO2022080327A1 (en) 2022-04-21

Similar Documents

Publication Publication Date Title
KR101818074B1 (en) Artificial intelligence based medical auto diagnosis auxiliary method and system therefor
WO2020253886A1 (en) Supplementary diagnosis method for pathology
US9858665B2 (en) Medical imaging device rendering predictive prostate cancer visualizations using quantitative multiparametric MRI models
US8682051B2 (en) Smoothing of dynamic data sets
US10290084B1 (en) Correcting image blur in medical image
JP2022180466A (en) Diagnosis support device, learning device, diagnosis support method, learning method, and program
CN111814768B (en) Image recognition method, device, medium and equipment based on AI composite model
US20200134823A1 (en) Image processing apparatus, image processing method, and storage medium
US11449210B2 (en) Method for providing an image base on a reconstructed image group and an apparatus using the same
De Jong et al. Ensemble landmarking of 3D facial surface scans
CN112926537A (en) Image processing method, image processing apparatus, electronic device, and storage medium
CN111782529A (en) Test method and device for auxiliary diagnosis system, computer equipment and storage medium
Piras et al. Morphologically normalized left ventricular motion indicators from MRI feature tracking characterize myocardial infarction
Singla et al. The open kidney ultrasound data set
Backhaus et al. Impact of fully automated assessment on interstudy reproducibility of biventricular volumes and function in cardiac magnetic resonance imaging
Hadler et al. Introduction of Lazy Luna an automatic software-driven multilevel comparison of ventricular function quantification in cardiovascular magnetic resonance imaging
WO2022080327A1 (en) Estimator learning device, estimator learning method, and estimator learning program
US20170235913A1 (en) Image data ingestion applciation of a medical imaging data processing and retrieval system
WO2024051018A1 (en) Pet parameter image enhancement method and apparatus, device, and storage medium
JP2019535472A (en) Inter-patient brain registration
Zopes et al. Multi-modal segmentation of 3D brain scans using neural networks
Bartolomé et al. Accuracy and time-efficiency of an automated software tool to assess left ventricular parameters in cardiac magnetic resonance imaging
CN113902983B (en) Laparoscopic surgery tissue and organ identification method and device based on target detection model
JP2019149124A (en) Conversion device, conversion method, and program
Morales et al. Present and Future Innovations in AI and Cardiac MRI

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21880066

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2022556974

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21880066

Country of ref document: EP

Kind code of ref document: A1