JP7185242B2 - Program and diagnostic imaging aid - Google Patents

Description

The present invention relates to a storage medium, an image diagnosis assisting device, a learning device, and a method of generating a trained model.

2. Description of the Related Art Conventionally, there has been known an image diagnosis assisting device for a doctor to diagnose whether or not there is a tumor in the breast of a human body.

In this regard, for example, in Patent Document 1, a breast medical image showing a breast is input to a learned learning device, a tumor candidate region (tumor candidate region) and a mammary gland region are detected, and a tumor is detected. A technique has been disclosed in which tumor candidate regions in regions other than mammary gland regions are removed from candidate regions and the remaining tumor candidate regions are determined to be tumors.

WO2018/180386

However, with the technique described in Patent Document 1, for example, if the size or position of a mammary gland region is detected incorrectly, even a tumor candidate region that should actually be a tumor region may be removed. , the accuracy of tumor determination was not sufficient.

The present invention has been made in view of such problems, and an object of the present invention is to provide a storage medium, an image diagnosis assisting device, a learning device, and a trained model that can improve the accuracy of tumor determination in breast medical images. To provide a method for generating

In order to solve the above problems, a storage medium according to a first aspect of the present invention causes a computer to input a breast medical image of a human breast to a learner constructed of a learned model in which machine learning has been performed. By doing so, a detection means for detecting a first region that is a tumor candidate in the breast medical image and a second region that is a tissue candidate other than the tumor in the breast medical image, and the detection means A program is stored that functions as determination means for determining whether or not the first region is a tumor based on the positional relationship of the plurality of detected regions, and output means for outputting the determination result of the determination means.

Further, in the storage medium according to the second aspect of the present invention, the detection means detects, as the second regions, a bone candidate region and a muscle candidate region.

Further, in the storage medium according to the third aspect of the present invention, the detection means may include, as the first regions, a region associated with a probability of being a malignant tumor and a region associated with a probability of being a benign tumor. and the determination means determines that the first region is a benign tumor based on the probability associated with the region detected by the detection means when the first region is detected as a tumor or malignant tumor.

Further, in the storage medium according to the fourth aspect of the present invention, the detection means has a probability that the first region is a tumor belonging to each category defined according to the BI-RADS (Breast Imaging Reporting And Data System) category. The associated region is detected, and the determining means, when the first region is detected as a tumor, based on the probability associated with the region detected by the detecting means, determines the first It is determined which category a region belongs to.

Further, in the storage medium according to the fifth aspect of the present invention, the trained models include a first trained model in which machine learning is performed from a plurality of breast medical images showing tumors, and a first trained model in which the other tissues are shown. a second trained model machine-learned from a plurality of breast medical images; Detecting a region and detecting the second region by inputting the breast medical image into a learner constructed with the second trained model.

Further, in the storage medium according to the sixth aspect of the present invention, the first region is a region associated with a probability that the probability of being a tumor is equal to or greater than a first threshold, and the second region is a region other than the tumor. This is an area associated with a probability that the probability of being other tissue is equal to or greater than the second threshold.

Also, in the storage medium according to the seventh aspect of the present invention, the second threshold is lower than the first threshold.

Also, in the storage medium according to the eighth aspect of the present invention, the output means outputs support information for supporting a doctor based on the determination result of the determination means.

Further, in the image diagnosis assisting device according to the ninth aspect of the present invention, by inputting a breast medical image showing a human breast to a learner constructed by a learned model on which machine learning has been performed, the breast medical image A detection means for detecting a first region that is a tumor candidate in an image and a second region that is a tissue candidate other than the tumor in the breast medical image; and a plurality of regions detected by the detection means. determining means for determining whether or not the first region is a tumor based on; and output means for outputting a determination result of the determining means.

Further, the learning device according to the tenth aspect of the present invention acquires a plurality of first breast medical images showing a tumor and a plurality of second breast medical images showing tissues other than the tumor. acquiring localization information of the tumor in the plurality of first breast medical images as first training data, and localizing the other tissue in the plurality of second breast medical images, based on a means and a user's operation; a teacher data acquisition means for acquiring information as teacher data; and machine learning by inputting a set of the plurality of first breast medical images and the first teacher data to obtain a first tumor candidate in a predetermined breast medical image. generating a first trained model for detecting a region, and machine learning by inputting the plurality of second breast medical images and the second teacher data set, and performing machine learning to detect a non-tumor in a predetermined breast medical image; learning processing means for generating a second trained model for detecting a second region that is a candidate for another tissue.

Further, the method for generating a trained model according to the eleventh aspect of the present invention includes a plurality of first breast medical images showing a tumor, and a plurality of second breast medical images showing tissues other than the tumor. and obtaining localization information of the tumor in the plurality of first breast medical images as first training data based on a user's operation, and the other tissue in the plurality of second breast medical images a step of acquiring position specifying information of as training data; and performing machine learning by inputting a set of the plurality of first breast medical images and the first training data to obtain a first tumor candidate in a predetermined breast medical image generating a first trained model for detecting a region, and machine learning by inputting the plurality of second breast medical images and the second teacher data set, and performing machine learning to detect a non-tumor in a predetermined breast medical image; generating a second trained model for detecting a second region that is a candidate for other tissue.

ADVANTAGE OF THE INVENTION According to this invention, the determination accuracy of the tumor can be improved in a breast medical image.

1 is a schematic diagram showing an example of the overall configuration of a diagnostic imaging system according to an embodiment of the present invention; FIG. 2 is a block diagram showing an example of the hardware configuration of the learning device shown in FIG. 1; FIG. 2 is a block diagram showing an example of the hardware configuration of the image diagnosis assisting device shown in FIG. 1; FIG. 3 is a block diagram showing an example of a functional configuration (software configuration) of the learning device shown in FIG. 2; FIG. FIG. 4 is a block diagram showing an example of a functional configuration (software configuration) of the image diagnosis assisting device shown in FIG. 3; 4 is a flow chart showing the processing procedure of each functional configuration of the learning device; 4 is a flow chart showing a processing procedure of each functional configuration of the image diagnosis assisting device; It is a figure which shows an example of the breast medical image which reflected the determination result. 2 is a table showing results of comparison between actual diseases and determination results of a tumor determination program when benign tumors and malignant tumors are regarded as "tumors". FIG. 10 is a table showing results of comparison between actual diseases and determination results of a tumor determination program when normal without benign tumors and when benign tumors are regarded as “normal”. FIG. 2 is a graph showing changes in sensitivity and specificity depending on the number of times of learning when benign tumors and malignant tumors are regarded as "tumors". FIG. 10 is a graph showing changes in sensitivity and specificity depending on the number of times of learning when normal without benign tumors and when benign tumors are regarded as “normal”. It is a schematic diagram showing an example of a report screen.

EMBODIMENT OF THE INVENTION Hereinafter, embodiment (it is also described as "this embodiment" hereafter.) of this invention is described, referring an accompanying drawing. In order to facilitate understanding of the description, the same components and steps are denoted by the same reference numerals as much as possible in each drawing, and overlapping descriptions are omitted.

<Overall composition>
FIG. 1 is a schematic diagram showing an example of the overall configuration of a diagnostic imaging system 1 according to an embodiment of the present invention.

As shown in FIG. 1 , the diagnostic imaging system 1 includes a learning device 10 and an auxiliary diagnostic imaging device 30 .

The learning device 10 is a computer that performs machine learning of a learner so as to generate a trained model for constructing a learner (program) such as a neural network, which is used in the image diagnosis assisting device 30 . Specifically, the learning device 10 then includes a plurality of first-breast medical images showing (the mammary glands of) a tumor and a set of tumor localization information in the plurality of first-breast medical images, A set of a plurality of second breast medical images showing tissues other than the first and a set of position specifying information of the other tissues in the plurality of second breast medical images are acquired as learning data. Among these learning data, the first breast medical image and the second breast medical image are used as input data, and the position specifying information of tumors and the position specifying information of other tissues are used as teacher data. That is, the learning device 10 inputs a set of a plurality of first breast medical images and tumor localization information and a set of a plurality of second breast medical images and other tissue localization information to the learning device. Do machine learning.

"Machine learning" is to use a computer to find patterns hidden in learning data, and "learning device" is a program that can acquire the ability to identify a predetermined pattern through such machine learning. Competencies are built by learning models. The type of this learning device is not particularly limited and may be a learning device using a deep learning technique or another learning device. Deep learning techniques include, for example, convolutional neural networks, adversarial generative networks, and recursive neural networks, but are not particularly limited. Below, the case where the “learning device” is a learning device using a convolutional neural network will be described. A "learner" may also be referred to as a "classifier" or a "classifier." "Position specifying information" is information for specifying where a tumor or other tissue is located in a breast medical image, such as one or more pieces of coordinate information of a tumor or other tissue. "Other tissues" include bones, muscles, mammary glands, fat, and the like. A “learning model” is a set of data including various parameters for setting (constructing) a learning device. Further, the “breast medical image” includes breast echo images, CT (Computed Tomography) images, X-ray images, MRI (Magnetic Resonance Imaging) images, and the like, which are used for medical purposes.

As described above, the trained model generated by the learning device 10 is installed in the image diagnosis assisting device 30 via a portable storage medium such as a compact disc or USB memory, or a communication network NT or the like. The type of communication network NT may be appropriately selected from, for example, the Internet, wireless communication network, mobile communication network, telephone network, dedicated network, and the like.

On the other hand, the image diagnosis assisting device 30 is a device for imaging the internal tissues of the human body, particularly the structure of the breast (mammary gland). In this embodiment, the image diagnosis assisting device 30 not only images the breast, but also determines the presence or absence of a breast tumor based on the breast medical image after imaging. Specifically, the image diagnosis assisting device 30 inputs a breast medical image to a learning device constructed from a trained model generated by the learning device 10, thereby obtaining a tumor candidate in the breast medical image. Detect one region and a second region that is a candidate for tissue other than the tumor in the breast medical image. Then, the image diagnosis assisting device 30 determines whether the first region is a tumor based on the detected positional relationship of the plurality of regions. Examples of the diagnostic imaging assisting device 30 include an ultrasonic diagnostic imaging assisting device, an X-ray imaging device, and the like, but are not particularly limited. In this embodiment, the image diagnosis assisting device 30 transmits ultrasonic waves to the breast of the human body and images the reflection (echo) from the tissue of the breast, thereby obtaining a medical image of the breast of the human body through ultrasound imaging diagnosis. A case of an auxiliary device will be described.

<Hardware configuration>
FIG. 2 is a block diagram showing an example of the hardware configuration of the learning device 10 shown in FIG. 1. As shown in FIG.

As shown in FIG. 2, the learning device 10 includes a control unit 12, a storage device 14, an input device 16, an output device 18, a communication interface 20, and a drive 22, which are electrically connected to each other. Prepare. In addition, in FIG. 2, the communication interface is described as "communication I/F."

The control unit 12 includes a CPU (Central Processing Unit) 12A, which is a hardware processor, a GPU (Graphics Processing Unit) 12B, a RAM (Random Access Memory) 12C, and a ROM (Read Only Memory) 12D. It is configured to execute various information processing based on the data.

The storage device 14 is composed of, for example, a hard disk, a ROM, or the like. This storage device 14 stores a learning program 14A, learning data 14B, a first trained model 14C, and a second trained model 14D. The learning program 14A is a program including a learning device that machine-learns the learning data 14B and generates learned models (first trained model 14C and second trained model 14D) as learning result data. The training data 14B may be, for example, a plurality of first breast medical images showing the breast and the tumor, a set of tumor localization information in the plurality of first breast medical images, and a plurality of images showing the other tissues other than the breast and the tumor. second breast medical image and a set of other tissue localization information in the plurality of second breast medical images. The first trained model 14C is data for a learning device to detect a first region that is a tumor candidate in a breast medical image. The second trained model 14D is data for detecting a second region that is a candidate for tissue other than the tumor in the breast medical image.

The input device 16 is, for example, a device such as a mouse and a keyboard for performing an input operation to the learning device 10 .

The output device 18 is, for example, a device for outputting such as a display and a speaker. A user can operate the learning device 10 via the input device 16 and the output device 18 .

The communication interface 20 is, for example, a wired LAN (Local Area Network) module, a wireless LAN module, or the like, and is an interface for performing wired or wireless communication via the communication network NT. The learning device 10 may distribute the generated trained model to an external device via the communication interface 20 .

The drive 22 is, for example, a CD drive, a DVD drive, or the like, and is a drive device for reading programs and data stored in the storage medium 24 . The type of drive 22 may be appropriately selected according to the type of storage medium 24 .

The storage medium 24 stores information such as programs by electrical, magnetic, optical, mechanical or chemical action so that information such as programs can be read by computers, other devices, machines, etc. It is a non-transitory medium for The learning device 10 may acquire the learning program 14A and/or the learning data 14B from this storage medium 24 . Here, in FIG. 2, as an example of the storage medium 24, a disk-type storage medium such as a CD or DVD is illustrated. However, the type of storage medium 24 is not limited to the disc type, and may be other than the disc type. As a storage medium other than the disk type, for example, a semiconductor memory such as a flash memory can be cited.

Regarding the specific hardware configuration of the learning device 10, it is possible to omit, replace, or add components as appropriate according to the embodiment. For example, controller 12 may include multiple hardware processors. The hardware processor may consist of a microprocessor, a field-programmable gate array (FPGA), or the like. The learning device 10 may be composed of a plurality of information processing devices. The learning device 2 may be an information processing device designed exclusively for the service provided, or may be a general-purpose server device, a PC (Personal Computer), or the like.

FIG. 3 is a block diagram showing an example of the hardware configuration of the image diagnosis assisting device 30 shown in FIG.

As shown in FIG. 3, the image diagnosis assisting device 30 includes a control unit 32, a storage device 34, a display device 36, an input device 38, an ultrasonic probe 40, and a transmission/reception circuit, which are electrically connected to each other. 42 , a signal processing circuit 44 and an image processing circuit 46 .

The control unit 32 includes a CPU 32A, which is a hardware processor, a GPU 32B, a RAM 32C, and a ROM 32D, and is configured to execute various types of information processing based on programs and data.

The storage device 34 is composed of, for example, a hard disk, a ROM, or the like. This storage device 34 stores a tumor determination program 34A, a first trained model 14C, a second trained model 14D, and the like. The tumor determination program 34A inputs a breast medical image obtained from the ultrasonic probe 40 to a learning device constructed by the first trained model 14C and the second trained model 14D, thereby identifying regions that are candidate tumors. A program that detects and determines if the area is in fact a tumor.

The display device 36 is a device that displays a breast medical image obtained from the ultrasonic probe 40, a menu screen, a setting screen of the image diagnosis auxiliary device 30, and the like.

The input device 38 is a device such as a mouse, keyboard, button, panel switch, touch command screen, foot switch, trackball, joystick, etc., for performing input operations to the image diagnosis auxiliary device 30 .

The ultrasonic probe 40 has a plurality of transducers, and the plurality of transducers generate ultrasonic waves based on drive signals supplied from the transmission/reception circuit 42 . Also, the plurality of transducers included in the ultrasonic probe 40 receive reflected waves from the breasts of the human body and convert them into electrical signals. Specifically, when ultrasonic waves are transmitted from the ultrasonic probe 40 to the human body, the transmitted ultrasonic waves are reflected one after another from discontinuous surfaces of acoustic impedance in the body tissues of the human body, resulting in reflected wave signals (echo signals ) are received by a plurality of transducers of the ultrasonic probe 40 . The amplitude of the received reflected wave signal depends on the difference in acoustic impedance at the discontinuity from which the ultrasonic waves are reflected.

The transmission/reception circuit 42 supplies a driving signal to the ultrasonic probe 40 . Also, the transmission/reception circuit 42 causes the ultrasonic probe 40 to transmit an ultrasonic beam. Then, the transmission/reception circuit 42 generates reflected wave data from the reflected wave signal received by the ultrasonic probe 40 .

For example, the signal processing circuit 44 performs various processes on the reflected wave data received from the transmission/reception circuit 42 to generate intermediate data. This intermediate data is output to the image processing circuit 46 .

The image processing circuit 46 generates breast medical images for display from the intermediate data generated by the signal processing circuit 44 . A breast medical image generated by the image processing circuit 46 is displayed on the display device 36 in real time.

Regarding the specific hardware configuration of the image diagnosis assisting device 30, it is possible to omit, replace, and add components as appropriate according to the embodiment. For example, controller 12 may include multiple hardware processors. The hardware processor may consist of a microprocessor, a field-programmable gate array (FPGA), or the like. The image diagnosis assisting device 30 may be composed of a plurality of information processing devices.

<Functional configuration>
FIG. 4 is a block diagram showing an example of the functional configuration (software configuration) of the learning device 10 shown in FIG.

As shown in FIG. 4, the learning device 10 includes storage means 60, image acquisition means 62, preprocessing means 64, teacher data acquisition means 66, and learning processing means 68 as functional configurations. The storage means 60 is realized by the storage device 14 . Other functional configurations are implemented by loading the learning program 14A in the RAM 12C and having the CPU 12A interpret and execute the learning program 14A loaded in the RAM 12C.

The storage means 60 has a function of storing the learning data 14B, the first trained model 14C, and the second trained model 14D.

The image acquisition means 62 has a function of acquiring a plurality of first breast medical images showing a tumor and a plurality of second breast medical images showing tissues other than the tumor. In this embodiment, the first breast medical image includes an image showing a benign tumor and an image showing a malignant tumor. Also, the second breast medical image includes an image showing bones and an image showing muscles.

The preprocessing means 64 has a function of performing various preprocessing on the breast medical images (the first breast medical image and the second breast medical image) acquired by the image acquiring means 62 . Specifically, the preprocessing means 64 performs first preprocessing for normalizing the image contrast of the breast medical image acquired by the image acquisition means 62 in order to avoid adverse effects on machine learning. In addition, the pre-processing means 64 executes a second pre-processing for reducing the 256-step color expression to less than half in order to suppress erroneous identification. In addition, the preprocessing means 64 performs a third process of finely dividing the breast medical image while overlapping in order to accurately capture a tumor that occupies a low proportion of the entire image and to avoid dividing the tumor. Run.

For example, based on the operation of the user of the learning device 10, the teacher data acquisition means 66 acquires tumor position specifying information in a plurality of preprocessed first breast medical images as first teacher data, and preprocessed It has a function of acquiring position specifying information of other tissues in a plurality of second breast medical images as teacher data. In this embodiment, based on the doctor's judgment and operation, high-quality images are extracted from the first breast medical image and the second breast medical image, and the position of the benign tumor or malignant tumor is extracted from each of the extracted first breast medical images. Good-quality teacher data is obtained by obtaining specific information and position specific information of bones or muscles from each of the extracted second breast medical images. Note that the position specifying information is obtained by a doctor specifying a region such as a malignant tumor in a breast medical image. The shape of this area is not particularly limited and may be rectangular, circular, elliptical, etc. For example, it is rectangular. As a rectangle, for example, a circumscribing rectangle that circumscribes the contour of a malignant tumor or the like, or an inscribed rectangle that inscribes the contour of a malignant tumor or the like can be used. In this embodiment, the position specifying information is specified by a circumscribing rectangle.

The learning processing means 68 inputs sets of a plurality of first breast medical images and first teacher data and sets of a plurality of second breast medical images and second teacher data, respectively, and performs machine learning to perform predetermined breast medical treatment. It has a function of generating a trained model for detecting a first region that is a tumor candidate in the image and a second region that is a tissue candidate other than the tumor in the predetermined breast medical image. In this embodiment, the learning processing means 68 inputs sets of a plurality of first breast medical images and first teacher data, performs machine learning (first learning processing), and becomes tumor candidates in predetermined breast medical images. A first trained model 14C is generated for detecting the first region. In addition, the learning processing means 68 inputs sets of a plurality of second breast medical images and second teacher data, performs machine learning (second learning processing), and performs tissue candidates other than the tumor in a predetermined breast medical image. A second trained model 14D is generated for detecting the second region where Although the number of machine learning is not particularly limited, it is preferably 20,000 times or more and 60,000 times or less from the viewpoint of increasing both sensitivity and specificity. More preferably, it is 40,000 times or more and 60,000 times or less. In addition, "sensitivity" is the probability that it can be detected by a test among sick people (those with tumors). "Specificity" is the probability of determining that a healthy person is not sick (no tumor).

FIG. 5 is a block diagram showing an example of the functional configuration (software configuration) of the image diagnosis assisting device 30 shown in FIG.

As shown in FIG. 5 , the image diagnosis assisting device 30 includes storage means 70 , acquisition means 72 , detection means 74 , determination means 76 and output means 78 . The storage means 70 is realized by the storage device 34 . Other functional configurations are implemented by loading the tumor determination program 34A in the RAM 32C and interpreting and executing the tumor determination program 34A loaded in the RAM 32C by the CPU 32A.

The storage means 70 has a function of storing the above-described first trained model 14C and second trained model 14D.

The acquisition unit 72 has a function of acquiring a breast medical image showing a human breast by driving and controlling the ultrasound probe 40 and the like. Note that, in practice, the acquisition unit 72 acquires the breast medical image in real time as part of the breast video.

The detection means 74 inputs the breast medical image acquired by the acquisition means 72 to one or a plurality of learners constructed with trained models that have undergone machine learning, so that the learners perform arithmetic processing, It has a function of detecting a first region that is a tumor candidate in the breast medical image and a second region that is a tissue candidate other than the tumor in the breast medical image. In this embodiment, the detection means 74 detects the first region by inputting the breast medical image to the learner constructed (set) by the first trained model 14C, and detects the breast medical image as the second trained model. The second region is detected by inputting to the learner constructed with the finished model 14D.

Here, the “first region that is a candidate for tumor” is, for example, a region associated with a probability that the probability of being a tumor is equal to or greater than the first threshold. Also, the “second region that is a candidate for another tissue” is, for example, a region associated with a probability that the probability of being a tissue other than a tumor is equal to or greater than the second threshold. In the present embodiment, the detection means 74 uses, as the first regions to be benign tumor candidates, a region associated with a probability of being a benign tumor equal to or higher than a first threshold, and a first region to be a malignant tumor candidate. , a region associated with a probability that the probability of being a malignant tumor is equal to or greater than the first threshold is detected. Further, the detecting means 74 detects, as a second region that is a bone candidate, a region associated with a probability that the probability of being a bone is equal to or higher than the second threshold, and as a second region that is a muscle candidate, a probability that it is a muscle. are associated with probabilities equal to or greater than the second threshold.

Note that the first threshold and the second threshold may be the same or different. For example, the second threshold is preferably lower than the first threshold from the viewpoint that the more candidates are detected for other tissues than the tumor, the more the tumor determination accuracy is improved. The first threshold and the second threshold are preferably 5% or more and 100% or less, more preferably 10% or more and 50% or less, from the viewpoint of increasing the detection accuracy of tumor candidates. In particular, it is even more preferable that the first threshold is 15% or more and 30% or less. Further, it is even more preferable that the second threshold is 10% or more and 20% or less. In this embodiment, the first threshold is 20% and the second threshold is 10%.

The determination means 76 has a function of determining whether or not the first region is a tumor based on the positional relationship of the multiple regions detected by the detection means 74 . For example, if the position of the first region that is a candidate for benign or malignant tumor is higher than the position of the second region that is a candidate for muscle or bone, the determination means 76 makes a positive determination that it is a tumor. Alternatively, if it is below the position of the second region that is a candidate for bone, it is negatively judged as not being a tumor. Further, for example, the determining means 76 determines that a tumor is present when the position of the first region, which is a candidate for a benign or malignant tumor, is separated from the position of the second region, which is a candidate for muscle or bone, by a threshold value or more. Affirmative determination may be made, and if the distance is less than the position of the second region, which is a candidate for muscle or bone, a negative determination may be made that it is not a tumor. The position of the first area and the position of the second area may be, for example, the position of the central point of the area or the position of the lowest point of the area. In the above positional relationship, "upper" means the skin side in the breast medical image.

Further, the determination means 76 determines whether the first region is a benign tumor or not based on the probability associated with the region detected by the detection means 74 when the first region is detected as a tumor Determine whether it is a malignant tumor. For example, when the region associated with the probability of being a malignant tumor and the region associated with the probability of being a benign tumor overlap, the determining means 76 selects the tumor with the higher probability. I judge.

Further, the determining means 76 determines that the second region, which is a candidate for another tissue detected by the detecting means 74, is another tissue as it is.

The output means 78 has a function of outputting the determination result of the determination means 76 . For example, the output means 78 reflects breast medical images displayed on the display device 36 in real time. Specifically, the output means 78 encloses the area determined to be a tumor by the determination means 76 in the breast medical image acquired and displayed in real time, and describes that the area is a tumor. . In addition, the output unit 78 encloses and displays the area determined to be other tissue in the breast medical image, and describes that the area is other tissue. In addition, the output means 78 may display a character or mark indicating a tumor in a corner of the breast medical image so as not to interfere with diagnosis by a doctor. Moreover, the output means 78 may output the determination result to sound or a paper medium.

<Processing procedure>
FIG. 6 is a flow chart showing the processing procedure of each functional configuration of the learning device 10. As shown in FIG. Note that the contents and order of the processing of the following steps can be changed as appropriate.

(Step SP10)
Image acquisition means 62 acquires a plurality of first breast medical images showing breasts and benign tumors and a plurality of first breast medical images showing breasts and malignant tumors from storage means 60 or other storage means. . Further, the image acquisition means 62 acquires a plurality of second breast medical images showing breasts and bones and a second breast medical image showing breasts and muscles from the storage means 60 or other storage means. At this time, based on the doctor's judgment and operation, only high-quality images are acquired from the first breast medical image and the second breast medical image. Then, the process shifts to the process of step SP12.

(Step SP12)
The preprocessing means 64 performs preprocessing on each breast medical image acquired by the image acquiring means 62 . Specifically, the preprocessing means 64 performs a first preprocessing for normalizing the image contrast of the breast medical image acquired by the image acquiring means 62 . The pre-processing means 64 also performs a second pre-processing that reduces the 256-step color representation to less than half for each breast medical image. Also, a third preprocessing is executed to finely divide each breast medical image while overlapping. Note that the order of executing the first pre-processing, the second pre-processing, and the third pre-processing can be changed as appropriate. Also, all or part of these pretreatments may be omitted. Then, the process shifts to the process of step SP14.

(Step SP14)
A teacher data acquiring means 66 acquires, as first teacher data, tumor position specifying information in a plurality of preprocessed first breast medical images based on a user's operation, and acquires a plurality of preprocessed second breast medical images. The position specifying information of other tissues in the medical image is obtained as second training data. Specifically, based on the judgment and operation of a doctor as a user, the teacher data acquisition means 66 acquires identification information indicating that each first breast medical image is a benign tumor or malignant tumor, and its position specifying information (rectangular area information) as the first teacher data. In addition, based on the doctor's judgment and operation, the teacher data acquisition means 66 acquires identification information indicating that it is a bone or muscle from each second breast medical image and its position specifying information (area information of a rectangular area) to the second breast medical image. Acquire as teacher data. As a result, the learning device 10 acquires high-quality learning data 14B including a plurality of sets of first breast medical images and first teacher data and a plurality of sets of second breast medical images and second teacher data. This learning data 14 B is stored in the storage means 60 . Then, the process shifts to the process of step SP16.

(Step SP16)
The learning processing means 68 accepts the setting of one or more first parameters for generating the first trained model 14C based on the user's operation. For example, the learning processing means 68 receives settings such as the number of first learning times and the first threshold value for the first learning process for generating the first trained model 14C as the first parameter. Then, the process shifts to the process of step SP18.

(Step SP18)
The learning processing means 68 accepts the setting of one or more second parameters for generating the second trained model based on the user's operation. For example, the learning processing means 68 receives settings such as the second learning count and the second threshold value for the second learning process for generating the second trained model 14D as the second parameters. Then, the process shifts to the process of step SP20.

(Step SP20)
The learning processing means 68 receives from the user an execution operation for executing two learning processes, that is, the first learning process and the second learning process. Then, the process moves to the process of step SP22 and the process of step SP34. Each process of steps SP22 to SP32 and each process of steps SP34 to SP44 below are executed in parallel, but after one of the processes is executed, the other process is executed. may be

(Step SP22)
The learning processing means 68 repeats the processing of steps SP24 to SP30 for the set first learning times a (loop processing). That is, the learning processing means 68 repeats the processing of steps SP24 to SP30 while incrementing the variable i from 1 until the variable i reaches the first number of times of learning a.

(Step SP24)
The learning processing means 68 inputs the learning data 14B related to the tumor to the learning device in the initial state. Specifically, the learning processing means 68 uses, as the learning data 14B, a first breast medical image showing a breast and a tumor, and one of tumor position specifying information (including tumor identification information) in the first breast medical image. Input the tuples into the learner in the initial state. Then, the process shifts to the process of step SP26.

(Step SP26)
The learning processing means 68 causes the learning device to perform a first learning process (machine learning) based on the input learning data 14B. Note that, in this first learning process, for example, the convolution process and the pooling process are repeated for the learning data 14B. Then, the process shifts to the process of step SP28.

(Step SP28)
The learning processing means 68 updates various parameters of the learning device based on the learning result data generated by the learning device. Then, the process shifts to the process of step SP30.

(Step SP30)
If the variable i is less than the first learning frequency a, the learning processing means 68 returns to step SP22. The process proceeds to step SP32.

(Step SP32)
The learning processing means 68 stores various parameters of the learning device finally updated by the loop processing in the storage means 60 as the first trained model 14C. Then, the series of processes shown in FIG. 6 ends.

(Step SP34)
After moving from step SP20, the learning processing means 68 repeats the processing of steps SP36 to SP42 for the set second learning times b (loop processing). That is, the learning processing means 68 repeats the processing of steps SP36 to SP42 while incrementing the variable j from 1 until the variable j reaches the second number of times of learning b.

(Step SP36)
The learning processing means 68 inputs the learning data 14B related to other tissues to the learning device in the initial state. Specifically, the learning processing means 68 uses, as the learning data 14B, a second breast medical image in which the breast and other tissues appear and position specifying information of other tissues in the second breast medical image (identification of other tissues information) is input to the learner in the initial state. Then, the process shifts to the process of step SP38.

(Step SP38)
The learning processing means 68 causes the learning device to execute a second learning process (machine learning) based on the input learning data 14B. In addition, in the second learning process, for example, the convolution process and the pooling process are repeated for the learning data 14B. Then, the process shifts to the process of step SP40.

(Step SP40)
The learning processing means 68 updates various parameters of the learning device based on the learning result data generated by the learning device. Then, the process moves to the process of step SP42.

(Step SP42)
The learning processing means 68 returns to the processing of step SP34 when the variable j is less than the second learning count b, and exits the loop processing when the variable j is the second learning count b. The process proceeds to step SP44.

(Step SP44)
The learning processing means 68 stores various parameters of the learning device finally updated by the loop processing in the storage means 60 as the second trained model 14D. Then, the series of processes shown in FIG. 6 ends.

FIG. 7 is a flow chart showing a processing procedure of each functional configuration of the image diagnosis assisting device 30. As shown in FIG. Note that the contents and order of the processing of the following steps can be changed as appropriate.

(Step SP50)
Acquisition means 72 drives and controls the ultrasound probe 40 to acquire a breast medical image. Then, the process shifts to the process of step SP52 and the process of step SP56. Each process of steps SP52 to SP54 and each process of steps SP56 to SP58 below are executed in parallel, but after one of the processes is executed, the other process is executed. may be

(Step SP52)
The detection means 74 reads the first trained model 14C from the storage means 70, and builds a learning device with the first trained model 14C. Note that a learner may be constructed with the first trained model 14C before the process of step SP50. In this case, step SP52 is omitted. Then, the process shifts to the process of step SP54.

(Step SP54)
The detection means 74 detects a first tumor candidate region from the breast medical image. Specifically, the detection means 74 inputs the breast medical image to the learner constructed by the first trained model 14C, and the detection means 74 detects that the probability of being a benign tumor is the first region as the first region. A region associated with a probability equal to or higher than a threshold and a region associated with a probability of being a malignant tumor equal to or higher than the first threshold are detected. Then, the process shifts to the process of step SP60.

(Step SP56)
After moving from step SP50, the detection means 74 reads the second trained model 14D from the storage means 70, and builds a learning device with the second trained model 14D. Note that a learning device may be constructed with the second trained model 14D before the process of step SP50. In this case, this step SP56 is omitted. Then, the process shifts to the process of step SP58.

(Step SP58)
The detection means 74 detects a second region that is a candidate for another tissue from the breast medical image. Specifically, the detection means 74 inputs the breast medical image to the learning device constructed by the second trained model 14D, and the detection means 74 detects the probability that the second region is bone with the second threshold Regions associated with the above probabilities and regions associated with probabilities of muscle being greater than or equal to the second threshold are detected. Then, the process shifts to the process of step SP60.

(Step SP60)
The determination means 76 acquires the number c of the second areas detected in step SP58. Note that the number of regions c is zero or a natural number. Then, the process shifts to the process of step SP62.

(Step SP62)
The determining means 76 repeats the processing of steps SP64 to SP66 by the acquired number of regions c (loop processing). That is, the determination means 76 repeats the processing of steps SP64 to SP66 while incrementing the variable k from 1 until the variable k reaches the number of areas c.

(Step SP64)
The determining means 76 acquires one second area out of the second areas detected by the detecting means 74, and determines that the one second area is another tissue. Note that the determining means 76 may determine that it is another tissue (bone or muscle) associated with the one region. In this case, when the second regions overlap each other, it may be determined that the tissue has a higher probability of being associated with each region. For example, when the second region associated with the probability of being bone (30%) overlaps with the second region associated with the probability of being muscle (40%), the determination means 76 Since the probability of being is higher, it may be determined that the second region is a muscle. Then, the process shifts to the process of step SP66.

(Step SP66)
If the variable k is less than the number of regions c, the determination means 76 returns to the processing of step SP62, and if the variable i is the number of regions c, exits the loop processing and proceeds to the processing of step SP68. Transition.

(Step SP68)
The determination means 76 acquires the number of regions d detected in step SP54. Note that the number of regions d is zero or a natural number. Then, the process shifts to the process of step SP70.

(Step SP70)
The determining means 76 repeats the processing from step SP72 to step SP88 by the acquired area number d (loop processing). That is, the determination means 76 repeats the processing of steps SP72 to SP88 while incrementing the variable l from 1 until the variable l reaches the number of regions d.

(Step SP72)
The determination means 76 obtains one first region from among the first regions detected by the detection means 74, and determines whether or not the one first region is positioned below the second region determined to be another tissue. determine whether Then, when the determination is affirmative, the process proceeds to step SP74, and when the determination is negative, the process proceeds to step SP78.

(Step SP74)
The determining means 76 determines that the acquired one first region is not a tumor. Then, the process shifts to the process of step SP76.

(Step SP76)
The output means 78 outputs the determination result. For example, the output means 78 reflects breast medical images displayed on the display device 36 in real time. In addition, this step SP76 may be performed after the process of step SP88, for example.

FIG. 8 is a diagram showing an example of a breast medical image 100 reflecting determination results. Note that the breast medical image 100 shown in FIG. 8 was newly created by the present inventors for explanation of the present embodiment, and may differ from the actual breast medical image.

As shown in FIG. 8, the output means 78 displays and outputs a rectangular mark 102 so as to enclose the region determined to be a tumor by the determination means 76 in the breast medical image 100, and also displays the indication that the region is a malignant tumor. is displayed. In addition, the output means 78 displays and outputs a rectangular mark 104 so as to enclose the area determined to be muscle by the determination means 76 in the breast medical image 100, and displays and outputs that the area is muscle. In addition, the output means 78 displays and outputs a rectangular mark 106 so as to enclose the area determined to be bone by the determination means 76 in the breast medical image 100, and displays and outputs that the area is bone. Note that this display output can be omitted.

Returning to FIG. 7, the process shifts to the process of step SP88.

(Step SP78)
When a negative determination is made in step SP72, the determining means 76 determines whether or not the acquired one first area overlaps with another first area. Then, when the determination is affirmative, the process proceeds to step SP82, and when the determination is negative, the process proceeds to step SP80.

(Step SP80)
The determining means 76 determines that one obtained first region is a tumor (benign tumor or malignant tumor) associated with the first region. Then, the process shifts to the process of step SP76.

(Step SP82)
If the determination in step SP78 is affirmative, the determining means 76 determines whether or not the malignant tumor has a higher probability among the probabilities associated with the overlapping first regions. For example, among the overlapping first regions, one first region is associated with the probability of being a malignant tumor (50%), and the other first region is associated with the probability of being a benign tumor (30%). If so, the determining means 76 makes an affirmative determination that the probability of being a malignant tumor is higher. Then, when the determination is affirmative, the process proceeds to step SP84, and when the determination is negative, the process proceeds to step SP86.

(Step SP84)
The determining means 76 determines that one obtained first region is a malignant tumor associated with the first region. Then, the process shifts to the process of step SP76.

(Step SP86)
The determining means 76 determines that one obtained first region is a benign tumor associated with the first region. Then, the process shifts to the process of step SP76.

(Step SP88)
If the variable l is less than the number of regions d, the judging means 76 returns to the processing of step SP70. ends.

<effect>
As described above, in the present embodiment, the tumor determination program 34A inputs a breast medical image showing a human breast to the image diagnosis assisting device 30, and inputs a breast medical image of a human breast to a learning device constructed by a learned model in which machine learning has been performed. Detecting means 74 for detecting a first region that is a tumor candidate in the breast medical image and a second region that is a tissue candidate other than the tumor in the breast medical image, and detected by the detecting means 74 Based on the positional relationship of the plurality of regions, it functions as determination means 76 for determining whether or not the first region is a tumor, and output means 78 for outputting the determination result of the determination means 76 .

According to this configuration, it is determined whether or not the first region, which is a candidate for a tumor, is a tumor based on the positional relationship of the plurality of regions detected by the detection means 74. Therefore, determination is made based on the positional relationship. Tumor determination accuracy can be improved as compared to the case where the measurement is not performed.

Further, in this embodiment, the detecting means 74 detects a bone candidate region and a muscle candidate region as the second regions.

According to this configuration, it is possible to determine whether the first region is a tumor based on the positional relationship among the first region that is a candidate for tumor, the second region that is a candidate for bone, and the second region that is a candidate for muscle. Since it is determined whether or not there is a tumor, it is possible to further improve the determination accuracy of the tumor.

Further, in this embodiment, the detecting means 74 detects, as the first regions, a region associated with the probability of being a malignant tumor and a region associated with the probability of being a benign tumor, and the determining means 76 determines whether the first region is a benign tumor or a malignant tumor based on the probability associated with the region detected by the detection means 74 when the first region is detected as a tumor judge.

According to this configuration, a tumor can be determined by distinguishing between malignant tumor and benign tumor, so that the determination accuracy of malignant tumor can be further improved.

Further, in the present embodiment, the trained models are the first trained model 14C machine-learned from a plurality of breast medical images showing tumors, and the machine learning from a plurality of breast medical images showing other tissues. is performed, and the detection means 74 detects the first region by inputting the breast medical image to the learner constructed with the first trained model 14C, and the breast The second region is detected by inputting the medical image to the learner constructed by the second trained model 14D.

According to this configuration, compared to the case where one trained model is generated by performing machine learning from a plurality of breast medical images showing tumors and a plurality of breast medical images showing other tissues, the learning device 10 The amount of memory used can be reduced, and the processing speed of the trained model generation process can be improved. In addition, since the first region and the second region are detected by a learning device built with a dedicated trained model, the region detection accuracy can be improved, thereby further improving the tumor determination accuracy. can be done.

Further, in the present embodiment, the first region is a region associated with a probability that the probability of being a tumor is equal to or higher than the first threshold, and the second region is a region that is associated with a probability of being a tissue other than the tumor, and the probability of being a tissue other than the tumor is the second threshold. This is an area associated with the above probabilities.

According to this configuration, the detecting means 74 does not detect regions whose probabilities are less than the first threshold and the second threshold, so that it is possible to suppress false detection of the first region, which is a candidate for a tumor, thereby preventing tumors from being detected. It is possible to further improve the determination accuracy.

Moreover, in this embodiment, the second threshold is lower than the first threshold.

According to this configuration, by detecting more candidate regions for other tissues than regions for tumor candidates, it is possible to compare the positional relationships of the regions more frequently. can be further improved.

Further, in the present embodiment, the learning device 10 acquires a plurality of first breast medical images showing a tumor and a plurality of second breast medical images showing tissues other than the tumor. and acquiring position specifying information of the tumor in the plurality of first breast medical images as first teacher data based on the user's operation, and position specifying information of the other tissue in the plurality of second breast medical images. as training data, machine learning is performed by inputting a set of the plurality of first breast medical images and the first training data, and the first tumor candidate in a predetermined breast medical image generating a first trained model 14C for detecting a region, and machine learning by inputting the plurality of second breast medical images and the second teacher data set, and performing machine learning to detect areas other than the tumor in a predetermined breast medical image; a learning processing means 68 for generating a second trained model 14D for detecting a second region that is a candidate for another tissue of .

According to this configuration, compared to the case where machine learning is performed from a plurality of breast medical images showing tumors and a plurality of breast medical images showing other tissues to generate one trained model, for example, two-class classification It can be configured with only a simpler trained model, and by combining trained models, it is expected to improve the judgment accuracy in breast medical images. In addition, it is possible to reduce the amount of memory used in the learning device 10 and improve the processing speed of the process of generating a trained model.

<Example 1>
Next, Example 1 to which the above embodiment is applied will be described.

In Example 1, a total of 331 breast medical images including 141 normal breast medical images, 77 breast medical images showing benign tumors, and 113 breast medical images showing malignant tumors were used as learning data 14B. Then, based on the user's operation, the learning device 10 inputs these learning data 14B to a learning device called "YOLOv2", respectively, and executes machine learning to obtain a first trained model 14C and a second trained model 14C. 14D was generated. Here, the number of times of learning was set to 60,000, the first threshold for detecting the first area was set to 20%, and the second threshold for detecting the second area was set to 10%. Subsequently, a total of 142 breast medical images, 62 normal breast medical images, 30 breast medical images showing benign tumors, and 50 breast medical images showing malignant tumors, were used as test data for tumor determination. Using the data, the first trained model 14C, and the second trained model 14D, the tumor determination program 34A was executed to obtain determination results regarding the tumor.

FIG. 9 is a table showing results of comparison between actual diseases and determination results of the tumor determination program 34A when benign tumors and malignant tumors are regarded as "tumors".

From the table shown in FIG. 9, it was found that the sensitivity of the judgment processing of the tumor judgment program 34A was 93.8% and the specificity was 95.2%. From this result, it was confirmed that the accuracy of tumor determination by the tumor determination program 34A was extremely high, and that the accuracy of determination was improved as compared with the conventional method.

FIG. 10 is a table showing the results of comparison between the actual disease and the determination result of the tumor determination program 34A when normal without benign tumors and benign tumors are regarded as "normal".

From the table shown in FIG. 10, it was found that the sensitivity of the judgment processing of the tumor judgment program 34A was 88.9% and the specificity was 86.9%. From this result, it was confirmed that the accuracy of tumor determination by the tumor determination program 34A was extremely high, and that the accuracy of determination was improved as compared with the conventional method.

<Example 2>
Next, Example 2 to which the above embodiment is applied will be described.

In Example 2, the same method as in Example 1 was performed, and the number of times of learning was varied between 10,000 and 80,000.

FIG. 11 is a graph showing changes in sensitivity and specificity depending on the number of times of learning when benign tumors and malignant tumors are regarded as “tumors”.

As shown in FIG. 11, from the viewpoint of high sensitivity and specificity, it is preferably 20,000 times or more and 60,000 times or less, more preferably 30,000 times or more and 60,000 times or less. 40,000 times or more and 60,000 times or less is more preferable from the viewpoint of high resistance and good balance.

Also, FIG. 12 is a graph showing changes in sensitivity and specificity depending on the number of times of learning when normal without benign tumors and when benign tumors are regarded as "normal".

As shown in FIG. 12, from the viewpoint of high sensitivity and specificity, it is preferably 20,000 times or more and 60,000 times or less, more preferably 30,000 times or more and 60,000 times or less. 40,000 times or more and 60,000 times or less is more preferable from the viewpoint of high resistance and good balance.

<Modification>
It should be noted that the present invention is not limited to the above specific examples. In other words, the above-described specific examples are also included in the scope of the present invention as long as they have the features of the present invention, as long as they are appropriately modified in design by those skilled in the art. In addition, each element provided in the embodiment described above and the modified example described later can be combined as long as it is technically possible, and the combination thereof is also included in the scope of the present invention as long as it includes the features of the present invention. .

For example, in the above-described embodiment, the case where the diagnostic imaging assisting device 30 is an ultrasonic diagnostic imaging assisting device has been described, but a server device, a personal computer, or the like connected to the ultrasonic diagnostic imaging assisting device via a communication network NT or the like may also be used. information processing apparatus.

Further, in the above embodiment, the case where the image acquisition unit 62 and the teacher data acquisition unit 66 are provided in the learning device 10 has been described, but they may be provided in another information processing device different from the learning device 10 .

In addition, in the above embodiment, the case where the attribute classification of the tumor is either benign or malignant has been described, but other classification criteria, for example, classification based on BI-RADS (Breast Imaging Reporting And Data System) guidelines may be used. In this case, the detection means 74 detects, as the first region, a region associated with the probability of being a tumor belonging to each category defined according to the BI-RADS category, and the determination means 76 detects that the first region is a tumor. If it is detected, it is determined which category the first region belongs to, based on the probability associated with the region detected by the detection means 74 . According to this configuration, attribute determination can be performed according to the BI-RADS category, so the accuracy of tumor determination can be further improved.

The above-described classification may be, for example, BI-RADS categories 1 to 5, subcategories 4A/4B/4C, or a combination of both. Also, the output label of the learning device may be binarized, such as assigning "0" when the category is 3 or less, and assigning "1" when the category is 4 or more.

In the above embodiment, the detecting means 74 detects a candidate malignant tumor region from a breast medical image. Alternatively, a region that is a candidate for a malignant tumor may be detected for each degree of progression of the malignant tumor. For example, the detection means 74 may detect a non-invasive cancer candidate area, an invasive cancer candidate area, and a Paget's disease candidate area from the breast medical image. Further, the detection means 74 detects a stage 1 malignant tumor candidate area, a stage 2 malignant tumor candidate area, a stage 3 malignant tumor candidate area, and a stage 4 malignant tumor candidate area. Regions may be detected. Thereby, the determining means 76 can determine the type and progress of the malignant tumor based on the probability associated with each region. Here, the determining means 76 determines whether the first region is a Determining whether or not the tumor is stage 1 malignant tumor can improve the accuracy of tumor stage determination.

In the above embodiment, the output means 78 outputs the judgment result of the judgment means 76, but the position information of the area detected by the detection means 74, the probability associated with the area, The type of the probability (probability of malignant tumor, probability of benign tumor, etc.) may be output to a report screen or the like as report information for a doctor or the like to make a judgment, for example. Further, the output means 78 may generate support information for supporting the doctor based on the region detected by the detection means 74 and/or the determination result of the determination means 76, and may propose (output) it to the user. . This allows doctors to efficiently determine diagnostic results and treatment methods. The support information may include treatment content and methods, locations of hospitals suitable for treatment, names of doctors, etc., according to the area detected by the detection means 74 and/or the determination result of the determination means 76 . In addition, the output means 78 may output report information and support information in real time together with the determination result of the determination means 76 .

FIG. 13 is a schematic diagram showing an example of the report screen 110. As shown in FIG.

As shown in FIG. 13, the report screen 110 includes, for example, the breast medical image 100, the location of the first region detected by the detection means 74, the probability associated with the first region, the size of the first region, the determination Report information 114 including the determination results determined by means 76 is described. Further, the report screen 110 describes support information 116 generated based on the determination result, the position information of the area detected by the detection means 74, the probability associated with the area, and the like.

Further, in the present embodiment, the case where the determining means 76 determines whether or not the first region is a tumor from one breast medical image 100 has been described. may be finally determined whether is a tumor. For example, out of five breast medical images 100, if it is determined that the first region is a tumor in one breast medical image 100 and that the first region is not a tumor in all the remaining breast medical images 100, , the determining means 76 may finally determine (determine) that the first region is a tumor. In this case, it is preferable that the output means 78 output only the final judgment result.

Further, in the above embodiment, the case where the detection means 74 detects the first region and the second region using a plurality of trained models has been described. A second region may be detected.

<Description of symbols>
DESCRIPTION OF SYMBOLS 10... Learning apparatus 24... Storage medium 30... Image diagnosis assistance apparatus (computer) 62... Image acquisition means 66... Teacher data acquisition means 68... Learning processing means 72... Acquisition means 74... Detection means 76 ... determination means, 78 ... output means

Claims (9)

the computer,
By inputting a breast medical image showing a human breast to a learner built with a trained model that has undergone machine learning, a first region that is a candidate for a tumor in the breast medical image and the breast medical image a detection means for detecting a second region that is a candidate for tissue other than a tumor in
determination means for determining whether the first region is a tumor based on the positional relationship between the first region and the second region or the positional relationship between the first regions detected by the detection means;
output means for outputting the determination result of the determination means;
function as
The first region is associated with a probability that the probability of being the tumor is a first threshold or more, and the second region is associated with a probability that the probability of being a tissue other than the tumor is a second threshold or more. is the associated region, and
The program, wherein the second threshold is lower than the first threshold. The detection means detects, as the first region, a region associated with a probability of being a malignant tumor and a region associated with a probability of being a benign tumor,
When the first region is detected as a tumor, the determining means determines whether the first region is a benign tumor based on the probability associated with the region detected by the detecting means, or determine whether it is a malignant tumor,
A program according to claim 1. When the region associated with the probability of being a malignant tumor and the region associated with the probability of being a benign tumor overlap, the determining means determines that the tumor is the tumor with the higher probability.
3. A program according to claim 2. The detection means detects, as the first region, a region associated with a probability of being a tumor belonging to each category defined according to the BI-RADS (Breast Imaging Reporting And Data System) category,
When the first region is detected as a tumor, the determination means determines which category the first region belongs to, based on the probability associated with the region detected by the detection means. determine whether there is
A program according to any one of claims 1 to 3. The trained model is a first trained model machine-learned from a plurality of breast medical images showing tumors, and a second trained model machine-learned from a plurality of breast medical images showing other tissues including a trained model and
The detection means detects the first region by inputting the breast medical image to a learner constructed by the first trained model, and constructs the breast medical image by the second trained model. Detecting the second region by inputting to the learned learner,
A program according to any one of claims 1 to 4. The detection means detects, as the second regions, a bone candidate region and a muscle candidate region.
A program according to any one of claims 1 to 5. The number of learning times of the trained model is 20,000 or more and 60,000 or less,
A program according to any one of claims 1 to 6. The output means outputs support information for supporting a doctor based on the determination result of the determination means.
A program according to any one of claims 1 to 7. By inputting a breast medical image showing a human breast to a learner built with a trained model that has undergone machine learning, a first region that is a candidate for a tumor in the breast medical image and the breast medical image a detection means for detecting a second region that is a candidate for tissue other than a tumor in
determination means for determining whether the first region is a tumor based on the plurality of regions detected by the detection means;
an output means for outputting the determination result of the determination means;
with
The first region is associated with a probability that the probability of being the tumor is a first threshold or more, and the second region is associated with a probability that the probability of being a tissue other than the tumor is a second threshold or more. is the associated region, and
The diagnostic imaging assisting apparatus, wherein the second threshold is lower than the first threshold.

Images