Ultrasonic diagnostic device, diagnostic assistance method, and computer program product

Abstract

The present invention relates to an ultrasonic diagnostic apparatus and a diagnostic support method. A detection unit (33) uses a machine learning detection model to detect lesion candidates contained in display frame data (tomographic images). An exclusion processing unit (34) compares the feature vector of the detected lesion candidate with a feature vector group in an exclusion database to determine whether the detected lesion candidate corresponds to an exclusion object. A mark display control unit (30) limits the display of a mark notifying the lesion candidate when the detected lesion candidate corresponds to an exclusion object.

Description

Technical Field

[0001] The present disclosure relates to an ultrasonic diagnostic apparatus and a diagnosis support method, and more particularly to a technique for notifying an examiner of a candidate lesion. Background Art

[0002] During an ultrasound examination, a probe is placed in contact with the patient's surface and scanned along it. During this scanning process, the examiner observes the real-time tomographic images displayed on a monitor to determine the presence of lesions. If a lesion is found, it is then examined in detail.

[0003] Visually identifying a temporarily appearing lesion on a dynamically changing tomographic image is not easy. One technology that assists in lesion identification is CADe (Computer Aided Detection). This technology detects potential lesions within a tomographic image and notifies the examiner of the detected lesions. For example, a marker surrounding the potential lesion is displayed on the tomographic image. CADe is used in conjunction with or is included in CAD (Computer Aided Diagnosis). CAD is also denoted as CADx.

[0004] Document 1 (Japanese Patent No. 5982726) discloses a data analysis system with CAD functionality. The automatic detection device within this system performs feature vector calculations and comparisons with learning data. If a lesion candidate is misidentified, relearning is performed. Document 1 does not disclose a technique for preventing misidentification without relearning.

[0005] In this specification, a lesion refers to a site that may be a disease site or a site that requires detailed examination. A lesion candidate refers to a site that is detected to assist the examiner in identifying and diagnosing a lesion site.

[0006] In ultrasonic diagnostic apparatuses that detect lesion candidates using a machine-learning detection unit and display a marker indicating the lesion candidate along with the ultrasound image, if the marker display target needs to be modified, the machine-learning detection unit is typically relearned. However, relearning typically requires considerable time. Since relearning is not easily performed, it is difficult to meet the need for flexible modification of marker display targets based on the patient, examination purpose, medical institution, and other factors. Summary of the Invention

[0007] The present disclosure aims to flexibly modify the lesion candidates that are the subject of marker display in an ultrasonic diagnostic apparatus that detects lesion candidates using a machine learning-based detection unit. Alternatively, the present disclosure aims to provide a novel mechanism that reduces the need for relearning in the machine learning-based detection unit.

[0008] The ultrasonic diagnostic device involved in the present disclosure is characterized in that it includes: a detection unit, which has a machine learning detection model and detects lesion candidates in an ultrasonic image; a database, which registers the feature value of each lesion candidate set as an exclusion object; a judgment unit, which judges whether the detected lesion candidate corresponds to an exclusion object by comparing the feature value of the lesion candidate detected by the detection unit with the feature value group registered in the database; and a display control unit, which displays a mark on the ultrasonic image to notify the detected lesion candidate when the detected lesion candidate does not correspond to the exclusion object, and limits the display of the mark when the detected lesion candidate corresponds to the exclusion object.

[0009] The diagnostic assistance method involved in the present disclosure is characterized in that it includes the following steps: using a machine learning completed detection model to detect lesion candidates in an ultrasonic image; registering the characteristic value of each lesion candidate set as an exclusion object into a database; determining whether the detected lesion candidate is equivalent to an exclusion object by comparing the characteristic value of the detected lesion candidate with the characteristic value group registered in the database; and displaying a mark notifying the detected lesion candidate on the ultrasonic image when the detected lesion candidate is not equivalent to an exclusion object, and restricting the display of the mark when the detected lesion candidate is equivalent to an exclusion object.

[0010] The program involved in the present disclosure is used to execute a diagnosis assistance method in an information processing device, and is characterized in that it includes the following functions: using a machine learning completed detection model to detect lesion candidates in an ultrasonic image; registering the characteristic value of each lesion candidate set as an exclusion object into a database; determining whether the detected lesion candidate is equivalent to an exclusion object by comparing the characteristic value of the detected lesion candidate with the characteristic value group registered in the database; and displaying a mark notifying the detected lesion candidate on the ultrasonic image when the detected lesion candidate is not equivalent to an exclusion object, and limiting the display of the mark when the detected lesion candidate is equivalent to an exclusion object. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a block diagram showing the configuration of an ultrasonic diagnostic apparatus according to an embodiment.

[0012] Figure 2 This is a diagram for explaining a method for generating a logo.

[0013] Figure 3 This is a diagram showing a first configuration example of the image analysis unit.

[0014] Figure 4 It is a diagram schematically showing the first operation example.

[0015] Figure 5 This is a flowchart showing the first operation example.

[0016] Figure 6 It is a diagram schematically showing the second operation example.

[0017] Figure 7 This is a flowchart showing the second operation example.

[0018] Figure 8 This is a diagram showing a second configuration example of the image analysis unit.

[0019] Figure 9 It is a diagram showing a modified example. DETAILED DESCRIPTION

[0020] The following describes the embodiments based on the drawings.

[0021] (1) Overview of implementation methods

[0022] The ultrasonic diagnostic apparatus according to the embodiment includes a detection unit, a database, a determination unit, and a display control unit. The detection unit includes a machine-learned detection model for detecting lesion candidates in an ultrasonic image. In the database, the feature value of each lesion candidate set as an exclusion object is registered. The determination unit determines whether the detected lesion candidate corresponds to an exclusion object by comparing the feature value of the lesion candidate detected by the detection unit with the feature value group registered in the database. If the detected lesion candidate does not correspond to an exclusion object, the display control unit displays a mark notifying the detected lesion candidate on the ultrasonic image, and if the detected lesion candidate corresponds to an exclusion object, the display control unit restricts the display of the mark. The detection unit corresponds to a detector. The determination unit corresponds to a determiner. The display control unit corresponds to a controller.

[0023] The above structure allows a portion of the detection object to be excluded from the mark display target after the fact so that the portion does not become the mark display target. According to the above structure, the mark display target can be easily restricted without modifying the machine-learned detection model. The removal of this restriction is also easy.

[0024] The concept of a candidate lesion feature includes: features extracted from the candidate lesion; features extracted from the image portion containing the candidate lesion and its surroundings; and features extracted from the ultrasound image containing the candidate lesion. Features also include feature vectors and image patterns. A marker notifying a candidate lesion is a display element that identifies the candidate lesion on an ultrasound image.

[0025] The ultrasonic diagnostic apparatus according to the embodiment further includes an input unit and a registration unit. The input unit receives exclusion instructions from a user observing lesion candidates in an ultrasonic image and observing markers on the ultrasonic image. The registration unit registers the feature values ​​of the lesion candidates subject to the exclusion instructions in a database. This allows the user to modify the group of lesion candidates subject to marker display without retraining the machine-learned detection model. The input unit serves as an input device, and the registration unit serves as a register.

[0026] In the embodiment, the input unit receives an exclusion instruction in the frozen state. The registration unit registers the feature value of the lesion candidate subject to the exclusion instruction in the database to update the database in the frozen state. The frozen state is a state in which transmission and reception are stopped and a still image is continuously displayed.

[0027] In an embodiment, the input unit receives an exclusion instruction in real time. The registration unit, in real time, identifies one or more frames of data that have been displayed based on the timing of receiving the exclusion instruction. Furthermore, the registration unit, in real time, extracts feature quantities of candidate lesions from the identified one or more frames of data and registers the feature quantities in a database to update the database.

[0028] The above configuration enables database updates in real time. For example, it is possible to limit the display of flags caused by false detections. Real time operation is a state where ultrasound images are displayed as dynamic images while transmitting and receiving simultaneously. For example, based on the timing of receiving an exclusion designation, multiple frames of data within a predetermined period of time are determined, and multiple features or representative features are extracted based on these frames.

[0029] In an embodiment, the detection unit has a function of obtaining a characteristic value of a lesion candidate. Alternatively, a calculation unit is provided separately from the detection unit. The calculation unit calculates the characteristic value of the lesion candidate. The calculation unit corresponds to a calculator.

[0030] An ultrasonic diagnostic apparatus according to an embodiment includes a management unit that clears all or part of a database. In the embodiment, the database includes multiple tables. Each table registers at least one feature value. The management unit selects a table to be cleared from the multiple tables. For example, tables may be prepared for each patient, each ultrasonic examination, or each examiner, and used tables may be cleared at a predetermined time. The management unit serves as a manager.

[0031] The diagnostic assistance method involved in the embodiment includes a detection process, a registration process, a judgment process, and a display control process. In the detection process, a machine learning detection model is used to detect lesion candidates in an ultrasonic image. In the registration process, the feature value of each lesion candidate set as an exclusion object is registered in a database. In the judgment process, the feature value of the detected lesion candidate is compared with the feature value group registered in the database, thereby determining whether the detected lesion candidate corresponds to an exclusion object. In the display control process, when the detected lesion candidate does not correspond to an exclusion object, a mark notifying the detected lesion candidate is displayed on the ultrasonic image. When the detected lesion candidate corresponds to an exclusion object, the display of the mark is restricted.

[0032] The above-mentioned diagnostic assistance method is implemented as a hardware function or a software function. In the latter case, the program for executing the diagnostic assistance method is installed on an information processing device via a removable storage medium or a network. The concept of information processing device includes ultrasonic diagnostic equipment, image processing equipment, computers, etc. The information processing device includes a non-transitory storage medium storing the above-mentioned program.

[0033] (2) Details of implementation methods

[0034] exist Figure 1 The structure of an ultrasonic diagnostic device according to an embodiment is shown as a block diagram. An ultrasonic diagnostic device is a medical device installed in a hospital or other medical institution. It generates ultrasonic images based on the received signals obtained by transmitting and receiving ultrasonic waves to and from a living body (examination subject). An example of an organ to be examined by ultrasonic examination is the breast.

[0035] During group breast examinations, it is essential to identify lesions quickly and without missing any. To assist the examiner in identifying lesions, the ultrasound diagnostic apparatus according to the present embodiment includes a CADe function that automatically detects lesion candidates (e.g., low-brightness areas that indicate a potential tumor) contained in ultrasound images. This function will be described in detail later.

[0036] The probe 10 functions as a unit for transmitting and receiving ultrasonic waves. Specifically, the probe 10 is a portable transceiver that is held and operated by the examiner (physician, technician, etc.). During breast ultrasound diagnosis, the transceiver surface of the probe 10 (specifically, the acoustic lens surface) is placed in contact with the patient's chest surface. While observing the real-time tomographic image, the probe 10 is manually scanned along the chest surface. Once a potential lesion is identified on the tomographic image, the position and posture of the probe 10 are gradually adjusted. Afterwards, the position and posture of the probe 10 are fixed, and the tomographic image is carefully observed.

[0037] In the illustrated configuration, the probe 10 includes a transducer array comprised of a plurality of transducers arranged one-dimensionally. The transducer array forms an ultrasonic beam (a transmit beam and a receive beam) 12, and electronic scanning of the ultrasonic beam 12 forms a scanning plane 14. The scanning plane 14 is an observation plane, that is, a two-dimensional data acquisition area. Known electronic scanning methods for the ultrasonic beam 12 include electronic sector scanning and electronic linear scanning. Convex scanning of the ultrasonic beam 12 is also possible. Alternatively, a 2D transducer array may be provided within the probe 10 to acquire volume data from within a living body through two-dimensional scanning of the ultrasonic beam.

[0038] A positioning system may also be provided to determine the position of the probe 10. The positioning system, for example, consists of a magnetic sensor and a magnetic field generator. In this case, the probe 10 (specifically, the probe head portion of the probe 10) is equipped with a magnetic sensor. The magnetic sensor detects the magnetic field generated by the magnetic field generator. This allows the three-dimensional coordinate information of the magnetic sensor to be obtained. The position and posture of the probe 10 can be determined based on this three-dimensional coordinate information.

[0039] The transmitting circuit 22 functions as a transmit beamformer. Specifically, during transmission, it supplies multiple transmit signals in parallel to the transducer array, thereby forming a transmit beam. During reception, when reflected waves from within the body reach the transducer array, multiple transducers output multiple receive signals in parallel. The receiving circuit 24 functions as a receive beamformer, generating beam data by phase-aligning and summing (also called delay and summing) the multiple receive signals.

[0040] Each electronic scan generates multiple beam data arrayed in the electronic scanning direction, which form received frame data corresponding to the scan plane 14. Each beam data array consists of multiple echo data arrayed in the depth direction. A beam data processing unit is provided after the receiving circuit 24, but this unit is not shown in the figure.

[0041] The image forming unit 26 is an electronic circuit that forms a tomographic image (B-mode tomographic image) based on received frame data. It includes a DSC (Digital Scan Converter). The DSC has functions such as coordinate transformation, pixel interpolation, and frame rate conversion. More specifically, the image forming unit 26 generates a display frame data string based on the received frame data string. Each display frame data constituting the display frame data string is tomographic image data. A real-time dynamic image is formed from multiple tomographic image data. Ultrasonic images other than tomographic images can also be generated. For example, a color blood flow imaging image can be formed, or a three-dimensional image that stereoscopically represents tissue can be formed. In the illustrated structural example, the display frame data string is sent to the display processing unit 32 and the image analysis unit 28.

[0042] The image analysis unit 28 is a module that performs the CADe function and includes a detection unit 33 and an exclusion processing unit 34 .

[0043] The detection unit 33 detects lesion candidates for each display frame, that is, for each tomographic image. The detection unit 33 includes a machine learning detector with a machine learning detection model. This detector, for example, is comprised of a CNN (Convolutional Neural Network), and detects low-brightness, enclosed areas as lesion candidates. Prior to detecting lesion candidates in the detection unit 33, binarization processing and noise removal may be performed on the tomographic images. After detecting lesion candidates, the detection unit 33 outputs lesion candidate information.

[0044] The lesion candidate information includes the candidate's location information, the candidate's size information, and the reliability of detection. If the reliability exceeds a certain threshold, the candidate lesion is determined to have been detected. The detection unit 33 has the function of calculating a feature vector as a feature quantity of the candidate lesion.

[0045] The position information of the lesion candidate is, for example, information indicating the coordinates of the center point of the lesion candidate itself, or information indicating the coordinates of the center point of a figure that is tangent to and surrounds the lesion candidate. The center point is a representative point. As the center point, the geometric center point of the figure or the center of gravity of the figure can be used. The size information of the lesion candidate is, for example, information indicating the size of the lesion candidate itself, or information indicating the size of a figure that is tangent to and surrounds the lesion candidate. For example, the size of the lesion candidate is determined based on the coordinates of the center point of the figure and the coordinates of the upper left corner point of the figure. On the premise that the coordinates of the center point are determined, the coordinates of the upper left corner point can be regarded as the size information of the lesion candidate. As the size information of the lesion candidate, the area of ​​the lesion candidate can also be obtained. It is also possible to detect multiple lesion candidates in parallel.

[0046] A feature vector is composed of multiple vector elements. One or more known feature vector calculation methods can be used. A multivariate analysis method can be used as a feature vector calculation method. An image pattern can be used as a feature quantity. Examples of feature quantities for a candidate lesion include feature quantities extracted from the candidate lesion, feature quantities extracted from an image portion containing the candidate lesion and its surroundings, and feature quantities extracted from an ultrasound image containing the candidate lesion.

[0047] The exclusion processing unit 34 includes an exclusion database. The exclusion processing unit 34 compares the feature vector of the lesion candidate currently detected by the detection unit 33 (the current lesion candidate) with the feature vector group in the exclusion database. If a feature vector with a certain degree of similarity or greater is found, that is, if the detected lesion candidate corresponds to an exclusion candidate, the exclusion processing unit 34 outputs an exclusion instruction signal to the mark display control unit 30. If a feature vector with a certain degree of similarity or greater is not found, that is, if the detected lesion candidate does not correspond to an exclusion candidate, no exclusion instruction signal is output.

[0048] The marker display control unit 30 displays a marker indicating the detected candidate lesion overlaid on the ultrasound image. However, if the exclusion processing unit 34 outputs an exclusion instruction signal, the marker display control unit 30 does not display the marker. In this way, by distinguishing between the detection target and the marker display target, the marker display target can be customized later. Customization eliminates the need for retraining the machine learning detector within the detection unit 33. Graphic data containing the marker generated by the marker display control unit 30 is output to the display processing unit 32.

[0049] When displaying a logo, the degree of reliability can be expressed by changing the display form of the logo. For example, when the reliability is low, the logo can be displayed in a cool color, and when the reliability is high, the logo can be displayed in a warm color. When the reliability is low, the logo can be displayed with low brightness, and when the reliability is high, the logo can be displayed with high brightness. When the reliability is low, the logo can be displayed with high transparency, and when the reliability is high, the logo can be displayed with low transparency. When the reliability is low, the logo can be displayed with a thin line width, and when the reliability is high, the logo can be displayed with a thick line width. When the reliability is low, the logo can be displayed with a dotted line, and when the reliability is high, the logo can be displayed with a solid line. The type of logo itself can also be switched. For example, the display of four display elements representing the four corners and the display of a rectangular graphic can be switched. It is also possible to use several techniques at the same time.

[0050] The image forming unit 26, the image analyzing unit 28, and the logo display control unit 30 can each be configured as a processor. A single processor may function as the image forming unit 26, the image analyzing unit 28, and the logo display control unit 30. Alternatively, a CPU (Central Processing Unit) described later may function as the image forming unit 26, the image analyzing unit 28, and the logo display control unit 30.

[0051] The display processing unit 32 has color calculation functions, image synthesis functions, and the like. The output of the image forming unit 26 and the output of the marker display control unit 30 are provided to the display processing unit 32. The marker surrounding the lesion candidate is one element of the graphic image. In the embodiment, the marker display control unit 30 generates the marker, but the main control unit 38, the display processing unit 32, or the like may also generate the marker.

[0052] The display 36 is composed of an LCD, an organic EL display device, etc. The tomographic image is displayed in real time as a moving image on the display 36, and the mark is displayed as part of the graphic image. The display processing unit 32 is composed of, for example, a processor.

[0053] The main control unit 38 controls Figure 1 The operations of the components shown in the figure are as follows. In the embodiment, the main control unit 38 is composed of a CPU that executes programs. An operation panel 40 is connected to the main control unit 38. The operation panel 40 is an input unit, i.e., an input device, and includes multiple switches, multiple buttons, a trackball, a keyboard, and the like. The operation panel 40 can be used to set or change the logo display conditions. Operation of the operation panel 40 is performed to register an exclusion target in the exclusion database.

[0054] In the embodiment, the display frame data string is given to the image analysis unit 28, but the reception frame data string (see reference numeral 42) may also be given to the image analysis unit 28. In this case, another image forming unit that performs image formation simply and quickly may be provided separately from the image forming unit 26.

[0055] The cine memory 27 has a ring buffer structure. It temporarily stores a sequence of display frame data spanning a certain period from the present time to the past. In the freeze state (described later), the display frame data selectively read from the cine memory 27 is displayed on the display 36 as a tomographic image (still image). At this time, the image analysis unit 28 and the mark display control unit 30 can be activated again to register exclusion targets. The data generated by the image analysis unit 28 and the mark display control unit 30 can be temporarily stored in the buffer, and the exclusion targets can be registered using the data read from the buffer.

[0056] exist Figure 2The following describes a marker generation method. A lesion candidate 46 is included in a tomographic image 44. A binary image is generated by binarizing the tomographic image 44. Edge detection or region detection is performed on the binary image to extract the binarized lesion candidate 46A. For example, a rectangle 52 circumscribing the lesion candidate 46A is defined by the horizontal and vertical coordinates of the lesion candidate 46A. In practice, the coordinates of the center point 48 and the upper left corner 50 of the rectangle are determined.

[0057] A rectangle 54 is defined outside rectangle 52 as a graphic with certain margins 56 and 58 in the horizontal and vertical directions. This rectangle 54 is displayed as a marker 64 on tomographic image 44. Marker 64 is a graphic that encloses the lesion candidate 46 and its surroundings. In the illustrated example, marker 64 is formed by a dotted line. A marker consisting of only four elements representing the four corners can also be displayed. A circular or elliptical marker can also be displayed.

[0058] In the embodiment, detection of a candidate lesion 46 is performed in units of display frame data. If a candidate lesion 46 is detected and does not correspond to an exclusion target, a marker 64 is displayed on the tomographic image 44 containing the candidate lesion 46. Displaying the marker 64 allows the examiner to be aware of the presence of the candidate lesion 46 and prevents the candidate lesion 46 from being missed. On the other hand, if the candidate lesion 46 corresponds to an exclusion target, the marker 64 is not displayed.

[0059] exist Figure 3 Show Figure 1 The first configuration example of the image analysis unit 28 is shown. The detection unit 33 includes a machine-learned detection model 70. Detection model 70 detects lesion candidates contained in a tomographic image (display frame data) 72. Detection model 70 includes, for example, a CNN parameter set. The implementation of this parameter set is indicated by reference numeral 74. In this embodiment, the inclusion of the exclusion processing unit 34 allows correction of marker display targets without relearning the detection model 70. As indicated by reference numeral 76, position information, size information, and other information constituting lesion information are provided to the marker display control unit 30.

[0060] The detection unit 33 has a function of calculating the feature vector of the lesion candidate. As shown by the reference numeral 78, the information representing the feature vector is sent to the exclusion processing unit 34. In addition, the feature vector can also be calculated by a module different from the detection unit 33. Figure 8 Explain this.

[0061] The exclusion processing unit 34 includes an exclusion database 80, a comparison unit (determination unit) 82, and a registration unit 84. The management unit 89, described later, may constitute a portion of the exclusion processing unit 34. In the illustrated configuration example, the exclusion database 80 is composed of multiple exclusion tables 80A. In actual use, an exclusion table 80A is selected based on a selection signal 88. Of course, the exclusion database 80 may also consist of a single exclusion table. Multiple exclusion tables 80A may be provided to correspond to multiple subjects, multiple examiners, or multiple medical disciplines.

[0062] The comparison unit 82 compares the feature vector of the currently detected current lesion candidate with the feature vector group registered in the selected exclusion table 80A. When a feature vector having a similarity greater than a certain degree is found, the current lesion candidate is determined to be an exclusion object, and an exclusion indication signal 90 is output from the comparison unit 82 to the mark display control unit 30. When there is no feature vector having a similarity greater than a certain degree, the exclusion indication signal 90 is not output. In this way, according to the embodiment, the current lesion candidate can be evaluated afterward. In other words, by such a post-evaluation, there is no need to relearn the detection model 70.

[0063] The registering unit 84 performs processing to register the feature vector of the current lesion candidate in the selected exclusion table. As will be described later, the feature vector can be registered in a frozen state or a real-time operation state. Reference numeral 86 denotes a signal representing a user's registration instruction.

[0064] The management unit 89 executes a process for clearing all or part of the exclusion database 80. The contents of each exclusion table 80A can be deleted. For example, one or more exclusion tables 80A can be cleared in response to a user's clear instruction, or one or more exclusion tables 80A can be automatically cleared when certain clearing conditions are met. For example, automatic clearing can be performed when a new patient's examination begins, when the transmission and reception conditions are changed, and so on.

[0065] The marker display control unit 30 has the function of generating a marker and limiting the display of the marker (normally, it has the function of turning the marker off). Based on the lesion candidate information output from the detection unit 33, the marker display control unit 30 generates a marker as a graphic surrounding the lesion candidate. The marker display method is changed based on the reliability. If an exclusion instruction signal is input, the marker display control unit 30 does not generate the marker and turns the marker off. The marker is generated on a frame-by-frame basis, and the marker display / hidden determination is also performed on a frame-by-frame basis.

[0066] exist Figure 4, shows a first example of an action performed in a frozen state. A tomographic image 102 is displayed as a still image on screen 92. The freeze button 96 on the operation panel 40 has been operated, and the image is in a frozen state, where transmission and reception are stopped as described above. Below the tomographic image 102, a strip-shaped graphic represents the storage area of ​​the recorded image memory. Its right end corresponds to the frozen time point. For example, by moving the cursor 100 to the left, an earlier display frame data can be selected. The selected display frame data is displayed as the tomographic image 102. In the illustrated example, the tomographic image 102 includes a candidate lesion 104, and a marker 106 is displayed to surround it.

[0067] If the mark 106 is displayed due to an erroneous detection, or if the display of the mark 106 is for a lesion candidate for which it is undesirable, the exclusion button 108 is pressed. The feature vector of the displayed lesion candidate is then registered in the exclusion database (specifically, the selected exclusion table). This updates the exclusion database. As needed, the registration of excluded objects is repeated in the frozen state. Pressing the freeze button 96 again releases the frozen state, returning to real-time operation. At this stage, the updated database is used to determine whether the lesion candidate is an exclusion candidate.

[0068] exist Figure 5 , the first action example is shown as a flowchart. In S10, the operation of the freeze button is detected, and the ultrasonic diagnostic apparatus becomes frozen. In S12, the user performs a reverse search of the recording screen memory to select specific display frame data. In S14, it is determined whether there is an operation to register an instruction. If there is such an operation, in S16, the feature vector of the displayed lesion candidate is registered in the exclusion database. That is, the exclusion database is updated. Until the operation of releasing the freeze is determined in S18, a series of steps after S10 are repeated. In the case where the operation of releasing the freeze is determined in S18, the updated exclusion database is used to determine the lesion candidate, as shown by the reference numeral 110.

[0069] exist Figure 6 The diagram schematically illustrates the second example of action executed in real-time. (A) represents the timeline, with t1 representing the time (current time) when the exclusion operation occurs. (B) represents the recorded image memory, which stores the display data sequence 112. Reference numeral 112A indicates the first display frame data, that is, the most recent display frame data in the recorded image memory. In the second example of action, if an exclusion operation occurs, the process of registering the excluded object continues from t1 for a certain period ta in the past, as shown in (C). The certain period ta can be specified, for example, in N seconds. N can be, for example, 1, 2, 3, etc.

[0070] Each display frame data within a certain period ta is checked for the presence of a lesion candidate. If a lesion candidate is detected, its feature vector is automatically registered in an exclusion database. A single feature vector representing multiple lesion candidates contained in multiple display frames can be calculated and registered in the database. For example, if a false positive occurs during an ultrasound examination, the examiner performs an exclusion operation (e.g., pressing an exclusion button). Thus, even in real-time operation, the feature vector is registered in the database and updated. From this point on, marker display control is executed based on the updated database.

[0071] exist Figure 7 The second operation is illustrated as a flowchart in FIG. At S20, it is determined whether an exclusion operation has occurred. If an exclusion operation has occurred, a reference period is determined at S22. At S24, each display frame data within the reference period is examined. If a lesion candidate is included, the feature vector is registered in the exclusion database. The exclusion database is updated sequentially. Specifically, as indicated by reference numeral 114, the updated exclusion database is immediately validated and utilized. At S28, it is determined whether the registration process is complete.

[0072] exist Figure 8 The second structural example of the image analysis unit is shown. Figure 8 , for Figure 3 The same structures as shown are denoted by the same reference numerals, and their description is omitted.

[0073] The image analysis unit 28A has a detection unit 33A and an exclusion processing unit 34A, and also has a feature vector calculation unit 116. The feature vector calculation unit 116 is provided separately from the detection unit 33A, and calculates the feature vector of the candidate lesion part. The feature vector calculation unit 116 is not a machine learning type calculation unit, but a general calculation unit. The detection unit 33A has a machine learning completed detection model 70A. The detection unit 33A does not have the function of calculating the feature vector. In the second structural example, the exclusion processing using the exclusion database 80 is also performed in the exclusion processing unit 34A. In addition, in Figure 8 In the figure, the management department is omitted.

[0074] exist Figure 9 A modified example is shown. In real-time operation, a tomographic image 120 is displayed as a dynamic image. It includes a candidate lesion to be detected, but a marker 126 is not displayed. In this case, an indicator 128 indicating that the exclusion processing mode is enabled may be displayed. Alternatively, an indicator 130 indicating that the marker display is restricted may be displayed.

[0075] According to the above embodiment, the lesion site candidates to be displayed as markers can be flexibly customized without requiring the machine learning type detection unit to perform re-learning. An additional database may be provided in addition to or instead of the excluded database.

Claims

1. An ultrasonic diagnostic device, characterized in that Include: A detection unit (33) having a machine learning-completed detection model (70) for detecting lesion candidates in an ultrasonic image; A database (80) registers the characteristic value of each lesion candidate set as an exclusion object; a determination unit (82) for determining whether the detected lesion candidate corresponds to an exclusion target by comparing a feature value of the lesion candidate detected by the detection unit (33) with a feature value group registered in the database (80); a display control unit (30) for displaying a mark on the ultrasonic image to inform the detected lesion candidate if the detected lesion candidate does not correspond to an exclusion target, and for limiting display of the mark if the detected lesion candidate corresponds to an exclusion target; an input unit (40) for receiving an instruction to exclude a lesion from a user who observes the lesion candidate in the ultrasonic image and a mark on the ultrasonic image; and A registration unit (84) registers the characteristic amount of the lesion candidate to be the subject of the exclusion instruction in the database, The user can correct the candidate group of lesion parts to be displayed as a mark without relearning the machine-learned detection model.

2. The ultrasonic diagnostic apparatus according to claim 1, wherein The input unit (40) receives the exclusion instruction in a frozen state, In the frozen state, the registration unit (84) registers the feature amount of the lesion site candidate to be the subject of the exclusion instruction in the database (80) to update the database (80).

3. The ultrasonic diagnostic apparatus according to claim 1, wherein The input unit (40) receives the exclusion instruction in a real-time operation state, In the real-time operation state, the registration unit (84) determines one or more frame data that have been displayed based on the timing of receiving the exclusion instruction, extracts feature quantities of lesion part candidates from the determined one or more frame data in the real-time operation state, and registers the feature quantities in the database (80) to update the database (80).

4. The ultrasonic diagnostic apparatus according to claim 1, wherein The detection unit (33) has a function of obtaining a feature value of the lesion site candidate.

5. The ultrasonic diagnostic apparatus according to claim 1, wherein The ultrasonic diagnostic device comprises: The calculation unit (116) is provided separately from the detection unit (33) and calculates the feature quantity of the lesion part candidate.

6. The ultrasonic diagnostic apparatus according to claim 1, wherein The ultrasonic diagnostic device comprises: The management unit (89) clears all or part of the database (80).

7. The ultrasonic diagnostic apparatus according to claim 6, wherein The database (80) comprises a plurality of tables (80A), At least one feature quantity is registered in each of the tables (80A), The management unit (89) selects a table to be cleared from the plurality of tables (80A).

8. A diagnosis-assisted method, characterized in that: The process includes the following steps: Step (33), using the machine learning detection model (70) to detect lesion candidates in the ultrasound image; A step of registering the characteristic value of each lesion candidate set as an exclusion object into a database (80); Step (82) of comparing the detected feature quantity of the lesion candidate with the feature quantity group registered in the database (80) to determine whether the detected lesion candidate corresponds to an exclusion object; Step (30), displaying a mark on the ultrasonic image to notify the detected lesion candidate if the detected lesion candidate does not correspond to an exclusion target, and limiting display of the mark if the detected lesion candidate corresponds to an exclusion target; Step (40), receiving an instruction to exclude the lesion from a user who observes the lesion candidate in the ultrasonic image and observes a mark on the ultrasonic image; and Step (84) is to register the characteristic value of the candidate lesion part to be the subject of the exclusion instruction in the database. The user can correct the candidate group of lesion parts to be displayed as a mark without relearning the machine-learned detection model.

9. A computer program product comprising a program for executing a diagnosis support method in an information processing device, wherein the program comprises the following functions: Function (33) uses a machine learning detection model (70) to detect lesion candidates in an ultrasound image; A function of registering the characteristic value of each lesion candidate set as an exclusion object into a database (80); Function (82) of comparing the detected feature quantity of the lesion candidate with the feature quantity group registered in the database (80) to determine whether the detected lesion candidate corresponds to an exclusion object; Function (30) of displaying a mark on the ultrasonic image to notify the detected lesion candidate when the detected lesion candidate does not correspond to an exclusion target, and restricting display of the mark when the detected lesion candidate corresponds to an exclusion target; Function (40) of receiving an instruction to exclude from the user who observes the lesion candidate in the ultrasonic image and the mark on the ultrasonic image; and Function (84) is to register the characteristic value of the candidate lesion part to be the subject of the exclusion instruction in the database, The user can correct the candidate group of lesion parts to be displayed as a mark without relearning the machine-learned detection model.

Images