Robust segmentation using high-level image understanding

Abstract

The system identifies anatomical objects visualized by medical imaging images. The system applies two machine learning models to the images. For each of a list of views, the system trains a first machine learning model to predict a view probability vector attributing the probability that the image was captured from the view, and for each of a list of anatomical objects, the system trains a second machine learning model to predict an object probability vector attributing the probability that the object is visualized by the image. For each object, the system (1) accesses a list of views in which the object is permitted, (2) multiplies the predicted probability that the object is visualized by the image by the sum of predicted probabilities that the accessed images were captured from views in which the object is permitted, and (3) if the resulting probability exceeds a threshold, determines that the object is visualized by the accessed images.

Description

Technical Field

[0001] (Cross - Reference to Related Applications) This application claims the benefit of U.S. Patent Application No. 17 / 091,263, filed on November 6, 2020, entitled "ROBUST SEGMENTATION THROUGH HIGH - LEVEL IMAGE UNDERSTANDING", which is hereby incorporated by reference in its entirety.

[0002] In case of conflict between this application and the documents incorporated by reference, this application shall prevail.

Background Art

[0003] Ultrasonic imaging is a useful medical imaging modality. For example, the internal structure of a patient's body can be imaged before, during, or after a therapeutic intervention. Medical professionals typically hold a portable ultrasonic probe, sometimes called a "transducer", in close proximity to the patient and move the transducer as needed to visualize one or more target structures within the patient's region of interest. The transducer may be placed on the body surface or, in some procedures, inserted into the patient's body. Medical professionals coordinate the movement of the transducer to obtain a desired representation on a screen, such as a two - dimensional cross - section of a three - dimensional volume.

[0004] A particular view of an organ or other tissue or body feature (such as body fluid, bone, joint, etc.) may be clinically important. Such views may be defined by clinical criteria as views that an ultrasound technician should capture, depending on the target organ, diagnostic purpose, etc.

[0005] In some ultrasonic images, it is useful to identify the anatomical structures visualized within the image. For example, in an ultrasonic image view showing a particular region of the body, it can be useful to identify the organs shown and / or the constituent structures within the organs. As an example, a right upper quadrant view may show parts of the liver, kidney, and diaphragm organs.

[0006] Existing software solutions have sought to automatically identify organs or other anatomical structures within ultrasound images. These existing solutions either "detect" structures by specifying bounding boxes in which each is visible, or "segment" structures by, for example, tracking the boundaries of each structure within the image. [Brief explanation of the drawing]

[0007] A patent or application file must include at least one drawing made in color. A copy of the published patent or patent application containing the color drawing will be provided by the Office upon request and payment of the necessary fees.

[0008] [Figure 1] This is a schematic diagram of a physiological sensing device. [Figure 2] This block shows some of the components that are typically incorporated into at least some of the other devices on which computer systems and equipment operate. [Figure 3] This flowchart illustrates the processes performed by the equipment in some embodiments to train the machine learning models used by the equipment. [Figure 4] This is a model architecture diagram showing the design details of the machine learning model used by the equipment. [Figure 5] This is a data flow diagram showing the data flow used by the equipment. [Figure 6] This is a flowchart showing the processes performed by the equipment. [Figure 7] This image shows sample patient video frames received by the equipment. [Figure 8] This is a table diagram showing sample contents of the acceptable object table. [Figure 9] This is a display diagram showing a version of the sample patient video frame with expanded equipment to visually represent objects identified as being visualized in the current frame. [Modes for carrying out the invention]

[0009] The inventors recognize that conventional approaches to automatically identify anatomical objects shown in ultrasound images have significant drawbacks. In particular, these conventional approaches often produce inaccurate results. Adjusting the confidence threshold may replace a high false positive rate with a high false negative rate, but the overall error is often unacceptably high.

[0010] In response to recognizing these shortcomings, the inventors have devised and implemented software and / or hardware equipment for automatically identifying organs, sub-organ structures, or other anatomical objects in medical images such as ultrasound images, by leveraging machine learning predictions regarding the specific view and visualized object ("Equipment") from which each image is captured. In doing so, the Equipment provides an efficient way to control the appearance of objects in an image, as the object vector directly estimates the probability of object occurrence and the view vector directly estimates the probability of view. These two auxiliary probability output vectors are correlated and can be used to generate a more robust estimate of the object occurrence likelihood using an object refinement module based on their relationship. Thus, the Equipment minimizes false positive prediction masking without compromising many false negative rates.

[0011] In some embodiments, the equipment applies a machine learning model trained to predict three characteristics of an image, such as a frame of ultrasound video: (1) a view probability vector for each of several possible views, indicating the likelihood that the image was captured from the view; (2) an object probability vector for each of several possible objects, indicating the likelihood that the image visualizes the object; and (3) for each pixel of the image, a set of segmentation masks, or "cubes," indicating the probability that the pixel is part of the visualization of the object for each of the possible objects. In some embodiments, the equipment averages or otherwise aggregates these predicted characteristics over a number of recent frames, such as 30.

[0012] For each possible object, to determine the probability that the object is visualized, the system examines a list of views that the object is allowed to identify. It sums the probabilities for those views included by the view probability vector and multiplies some of that by the probability for that object included by the object probability vector.

[0013] In some embodiments, the equipment uses the determined object probabilities to select a segmentation mask to use to extend the image with the shapes and positions of objects visualized in the image. In some embodiments, the equipment selects a segmentation mask for objects where the determined object probabilities exceed an object probability threshold. In various embodiments, the equipment uses the determined object probabilities and / or selected segmentation masks in various other ways.

[0014] By operating using some or all of the methods described above, the equipment can automatically identify anatomical structures in medical images with greater accuracy than before.

[0015] Furthermore, the system enhances the capabilities of computers or other hardware by reducing the dynamic display area, processing, storage, and / or data transmission resources required to perform specific tasks, thereby enabling tasks to be performed by less powerful, less cumbersome, and / or less expensive hardware devices, and / or enabling tasks to be performed with shorter latency, and / or allowing more of the stored resources to be used for other tasks. For example, by maximizing the usefulness of initial medical images by more frequently identifying all structures visualized with significantly fewer false positives, the system avoids many cases that require re-imaging. By reducing the need for re-imaging, the system captures additional images and performs additional automated structure identification rounds with less overall memory and processing resources.

[0016] Figure 1 is a schematic diagram of a physiological sensing device 10 according to one or more embodiments of the present disclosure. In the illustrated embodiments, the device 10 includes a probe 12 electrically coupled to a handheld computing device 14 by a cable 17. The cable 17 includes a connector 18 that detachably connects the probe 12 to the computing device 14. The handheld computing device 14 may be any portable computing device having a display, such as a tablet computer or a smartphone. In some embodiments, the probe 12 does not need to be electrically coupled to the handheld computing device 14, but may operate independently of the handheld computing device 14, and the probe 12 may communicate with the handheld computing device 14 via a wireless communication channel.

[0017] The probe 12 is configured to transmit an ultrasonic signal toward a target structure and to receive an echo signal returned from the target structure in response to the transmission of the ultrasonic signal. In various embodiments, the probe 12 includes an ultrasonic sensor 20 which may include an array of transducer elements (e.g., a transducer array) capable of transmitting an ultrasonic signal and receiving a subsequent echo signal.

[0018] The device 10 further includes a processing circuit and a drive circuit. Partially, the processing circuit controls the transmission of ultrasonic signals from the ultrasonic sensor 20. The drive circuit is operably coupled to the ultrasonic sensor 20 for driving the transmission of ultrasonic signals in response to control signals received from the processing circuit, for example. The drive circuit and processor circuit may be included in one or both of the probe 12 and the handheld computing device 14. The device 10 also includes a power supply that provides power to the drive circuit for transmitting ultrasonic signals in, for example, pulse wave mode or continuous wave mode.

[0019] The ultrasonic sensor 20 of the probe 12 may include one or more transmitting transducer elements that transmit ultrasonic signals, and one or more receiving transducer elements that receive echo signals returning from a target structure in response to the transmission of ultrasonic signals. In some embodiments, some or all of the transducer elements of the ultrasonic sensor 20 may function as transmitting transducer elements during a first period and as receiving transducer elements during a second period different from the first period (i.e., the same transducer elements may be available to transmit ultrasonic signals and receive echo signals at different times).

[0020] The computing device 14 shown in FIG. 1 includes a display screen 22 and a user interface 24. The display screen 22 may be a display incorporating any type of display technology, including but not limited to LCD or LED display technology. The display screen 22 is used to display one or more images generated from echo data obtained from echo signals received in response to the transmission of ultrasonic signals. In some embodiments, the display screen 22 may be used to display color flow image information, such as may be provided in a Color Doppler imaging (CDI) mode. Further, in some embodiments, the display screen 22 may be used to display audio waveforms, such as waveforms representing acquired or adjusted auscultation signals.

[0021] In some embodiments, the display screen 22 may be a touch screen capable of receiving input from a user touching the screen. In such embodiments, the user interface 24 may include a portion or the entire display screen 22 capable of receiving user input via touch. In some embodiments, the user interface 24 may include one or more buttons, knobs, switches, etc. capable of receiving input from a user of the ultrasonic device 10. In some embodiments, the user interface 24 may include a microphone 30 capable of receiving audible input, such as voice commands.

[0022] The computing device 14 may further include one or more audio speakers 28 that may be used to output an audible representation of an acquired or adjusted auscultation signal, or an echo signal, blood flow during Doppler ultrasound imaging, or other characteristics derived from the operation of the device 10.

[0023] Probe 12 includes a housing that forms an outer portion of the probe 12. The housing includes a sensor portion located near the distal end of the housing and a handle portion between the proximal end and the distal end of the housing. The handle portion is located proximal to the sensor portion.

[0024] The handle portion is a part of the housing that is gripped by a user to hold, control, and manipulate the probe 12 during use. The handle portion may include a gripping mechanism such as one or more anti-slip features, and in some embodiments, the handle portion may have the same general shape as a portion of the housing on the distal or proximal side of the handle portion.

[0025] The housing surrounds the internal electronic components and / or circuits of the probe 12, including, for example, electronic devices such as drive circuits, processing circuits, oscillators, beamforming circuits, filtering circuits, etc. The housing may be formed to surround or at least partially surround portions of the probe 12 that are disposed externally, such as a sensing surface. The housing may be a sealed housing such that moisture, liquid, or other fluids are prevented from entering the housing. The housing may be formed of any suitable material, and in some embodiments, the housing is formed of a plastic material. The housing may be formed as a single piece (e.g., a single material molded to surround the internal components). Alternatively, it may be formed of two or more parts (e.g., an upper half and a lower half) that are joined to each other or attached to each other in another way.

[0026] In some embodiments, the probe 12 includes a motion sensor. The motion sensor is operable to sense the movement of the probe 12. The motion sensor is contained within or on the probe 12 and may include, for example, one or more accelerometers, magnetometers, or gyroscopes for sensing the movement of the probe 12. For example, the motion sensor may be any piezoelectric, piezoresistive, or capacitive accelerometer capable of sensing the movement of the probe 12, and may include any of these accelerometers. In some embodiments, the motion sensor is a three-axis motion sensor capable of sensing movement around any of the three axes. In some embodiments, two or more motion sensors 16 are contained within or on the probe 12. In some embodiments, the motion sensor includes at least one accelerometer and at least one gyroscope.

[0027] The motion sensor may be at least partially housed within the housing of the probe 12. In some embodiments, the motion sensor is located on or near the sensing surface of the probe 12. In some embodiments, the sensing surface is a surface that operably contacts the patient during an examination such as ultrasound imaging or auscultation sensing. The ultrasound sensor 20 and one or more auscultation sensors are positioned on, on, or near the sensing surface.

[0028] In some embodiments, the transducer array of the ultrasonic sensor 20 is a one-dimensional (1D) or two-dimensional (2D) array of transducer elements. The transducer array may include a piezoelectric ceramic such as lead zirconate titanate (PZT), or it may be based on a microelectromechanical system (MEMS). For example, in various embodiments, the ultrasonic sensor 20 may include a piezoelectric micromachined ultrasonic transducer (PMUT), which is a MEMS-based piezoelectric ultrasonic transducer, or the ultrasonic sensor 20 may include a capacitive micromachined ultrasonic transducer (CMUT), in which energy conversion is provided due to a change in capacitance.

[0029] The ultrasonic sensor 20 may further include an ultrasonic focusing lens that can be positioned on the transducer array and can form part of the sensing surface. The focusing lens may be any lens that can operate to focus the ultrasonic beam transmitted from the transducer array toward the patient and / or to focus the ultrasonic beam reflected from the patient toward the transducer array. In some embodiments, the ultrasonic focusing lens may have a curved surface shape. The ultrasonic focusing lens may have different shapes depending on the desired application, such as a desired operating frequency. The ultrasonic focusing lens may be formed of any suitable material, and in some embodiments, the ultrasonic focusing lens is formed of a room-temperature-vulcanizing (RTV) rubber material.

[0030] In some embodiments, the first and second films may be positioned adjacent to each other on opposite sides of the ultrasonic sensor 20, forming part of the sensing surface. The films may be made of any suitable material, and in some embodiments, the films are made of room-temperature vulcanized (RTV) rubber material. In some embodiments, the films are made of the same material as the ultrasonic focusing lens.

[0031] Figure 2 is a block diagram showing some of the components typically incorporated into at least some of the other devices on which computer systems and equipment operate. In various embodiments, these computer systems and other devices 200 may include server computer systems, cloud computing platforms or virtual machines in other configurations, desktop computer systems, laptop computer systems, netbooks, mobile phones, personal digital assistants, televisions, cameras, automotive computers, electronic media players, physiological sensing devices and / or their associated display devices, etc. In various embodiments, the computer system and apparatus include zero or more of each of the following: a processor 201 for executing computer programs and / or training or applying machine learning models, such as a CPU, GPU, TPU, NNP, FPGA, or ASIC; an operating system including a kernel, and equipment and associated data; computer memory 202 for storing the program and data while it is in use, including equipment drivers; persistent storage devices 203, such as a hard drive or flash drive, for persistently storing the program and data; computer-readable media drives 204, such as a floppy disk, CD-ROM, or DVD drive, for reading the program and data stored on computer-readable media; and the Internet or another network, and network connections 205 for connecting the computer system to other computer systems to send and receive data via its networking hardware, such as switches, routers, repeaters, electrical cables and optical fibers, light emitters and optical receivers, wireless transmitters and receivers. While the computer system configured as described above is typically used to support the operation of the equipment, those skilled in the art will understand that the equipment may be implemented using various types and configurations of apparatus and may have various components.

[0032] Figure 3 is a flowchart illustrating the process performed by the equipment in several embodiments to train a machine learning model used by the equipment. In operation 301, the equipment receives training observations. In operation 302, the equipment trains a machine learning model using the training observations received in operation 301. In operation 303, the equipment stores the model trained in operation 302. After operation 303, the process terminates.

[0033] Those skilled in the art will understand that the operations in each of the flowcharts shown in Figure 3 and discussed below can be modified in various ways. For example, the order of operations may be rearranged, some operations may be performed in parallel, some illustrated operations may be omitted, other operations may be included, some illustrated operations may be divided into partial operations, or multiple illustrated operations may be combined into a single operation.

[0034] Figure 4 is a model architecture diagram showing details of the design of a machine learning model used by the apparatus in several embodiments. As shown in architecture 400, based on an image 401 such as a patient ultrasound image, the apparatus predicts three properties of the image: a view probability vector 421 for each of several possible views, indicating the likelihood that the image was captured from that view; an object probability vector 431 for each of several possible objects, indicating the likelihood that the image visualizes the object; and a set of segmentation masks or cubes 441 for each pixel of the image, indicating the probability that the pixel is part of the visualization of the object for each of the possible objects. The model is shown as a deep learning neural network and consists of a shared section 410 and a view probability branch 420 dedicated to predicting the view probability vector property, an object probability branch 430 dedicated to predicting the object probability vector property, and a segmentation branch 440 dedicated to predicting the segmentation mask cube. Legend 402 explains the symbols used when diagramming the network architecture. Specifically, "conv" represents a convolutional layer, "BN" represents a batch normalization layer, and "ReLU" represents a rectified linear unit. The numbers below each layer represent either the number of channels or the number of dimensions. All convolutional layers are 3x3 kernels. In various embodiments, the system uses various other neural networks or other machine learning model architectures to predict the three image characteristics used by the system.

[0035] Figure 5 is a data flow diagram showing the data flow used by the equipment in several embodiments to identify anatomical objects within a patient video frame and to expand the frame for display using this information. The data flow begins with the current patient video frame 501 received from the imaging hardware.

[0036] Figure 6 is a flowchart illustrating the process performed by the equipment to identify anatomical objects in the patient video and use this information to expand the video frames, following the data flow shown in Figure 5. In operation 601, the equipment receives the latest frame from the imaging hardware.

[0037] Figure 7 is an image showing a sample of patient video frames received by the equipment. In particular, frame 700 is a frame of patient ultrasound video.

[0038] Returning to Figure 6, in operation 602, in the AI ​​prediction module 510, the equipment applies a machine learning model trained by the equipment to the latest frame 501 received in operation 601 in order to obtain the following characteristics of the image constituting the frame: namely, the view probability vector 521, the object probability vector 522, and the segmentation mask cube 523.

[0039] In operation 603, in the buffer module 530, the equipment aggregates the three image characteristics 520 generated in operation 602 with a previous set of image characteristics maintained in the buffer for previous frames, such as 4, 9, 19, or 29 previous frames. In various embodiments, the equipment uses various aggregation functions, including mean, median, mode, minimum, and maximum values. Operation 603 generates a set of aggregated image characteristics 540.

[0040] The operation of the equipment is described with respect to a specific example based on the current frame 501, which is shown at a larger magnification as frame 700 in Figure 7. Table 1 below shows the contents of the aggregated view probability vector 541 determined by the equipment for the current frame. [Table 1]

[0041] For example, this view probability vector shows the 20% likelihood that the frame preceding the current frame and the frame in which its information is included in the collection was captured from the top-left quadrant view.

[0042] The contents of the aggregated object probability vector 542 in this example are shown in Table 2 below. [Table 2]

[0043] For example, the contents of this object probability vector represent the 50% likelihood that the spleen object is visualized in the sample image and its preceding images. In this example, a segmentation mask cube 543 corresponding to the current frame and its preceding frames is also generated. For brevity, the contents of the segmentation of the mask cube are not shown.

[0044] In operation 604, in the object refinement module 550, the equipment refines the object probability vector aggregate result 542 using the view probability vector aggregate result 541 and the allowable object table, thereby generating the refined object probability vector aggregate result 562.

[0045] Figure 8 is a table diagram showing sample contents of an allowable object table used by the equipment in several embodiments. In particular, the allowable object table 800 indicates, for each of the numerous views 801-806, whether the view is permitted to visualize each of the numerous organs or other objects 811-819. For example, the intersection of column 811 and rows 801 and 802 indicates that the liver is permitted to be visualized in the upper right quadrant view, but not in the upper left quadrant view.

[0046] Figure 8 and each of the tables described below show tables whose content and structure are designed to be better understood by human readers; however, those skilled in the art will understand that the actual data structures used by the equipment to store this information may differ from the illustrated tables in that they may be configured differently, contain more or less information than those illustrated, be compressed, encrypted, and / or indexed, and contain far more rows than those illustrated.

[0047] When performing step 604, the equipment refines each value in the object probability vector aggregate results shown in Table 2 as follows: First, the equipment uses the allowed objects table to determine which views are allowed to visualize the object for the object corresponding to the vector value. Using the kidney object as an example, column 812 of the allowed objects table indicates that the kidney is allowed to be visualized in only two views: the upper right quadrant view and the upper left quadrant view. In response, the equipment sums the view probabilities in the view probability vector aggregate results for these objects. For the kidney object, the equipment sums the view probability of the upper right quadrant view (0.7) and the view probability of the upper left quadrant view (0.2) to obtain a total of 0.9. Next, the equipment multiplies this sum of selected view probabilities by the object probability for the object. In the example of the kidney object, the equipment multiplies the total of 0.9 by the initial object probability of 0.8 to obtain a refined object probability of 0.72 for the kidney. The equipment performs this process for each object in order to obtain the refined object probability vector aggregate result 562, and the details in this example are shown in Table 3 below. [Table 3]

[0048] In operation 605, in the segment refinement module 570, the equipment uses the refined object probability vector aggregate result 562 obtained in operation 604 to filter the masks contained in the object segmentation mask aggregate result 543. Specifically, the equipment applies a minimum probability, such as 60%, to the value of each object in the refined object probability vector aggregate result. In this example, applying a minimum probability of 0.6, the equipment selects segmentation masks only for the liver, whose refined probability is 0.63, and for the kidney and diaphragm, whose refined probabilities are both 0.72. Thus, the equipment includes segmentation masks in cube 543 for the liver, kidney, and diaphragm objects in the filtered segmentation mask cube 580. In operation 606, the equipment extends the current frame 501 with visual indications of the identified objects based on the filtered object segmentation mask aggregate result 580 generated in operation 605 to obtain the displayed frame 590.

[0049] Figure 9 is a diagram showing a version of a sample patient video frame in which the equipment has been extended to visually display objects identified as being visualized in the current frame. It can be seen that the extended frame 900 includes a visual display of a liver object 901, a display of a kidney object 902, and a display of a diaphragm object 903. In various embodiments, these visual displays are of various characteristics. For example, in various embodiments, they are unfilled shapes, filled shapes whose brightness and intensity change based on probabilities in the corresponding segmentation mask and / or visual data in the original image, filled shapes with constant intensity, etc.

[0050] Returning to Figure 6, in operation 607, the equipment displays the expanded latest frame 590 generated in operation 606 on a display device physically connected to the imaging sensor, a local display device wirelessly connected to the imaging sensor, a remote display device, etc. After operation 607, the equipment processes the next frame of the patient video, following operation 601.

[0051] Further embodiments can be provided by combining the various embodiments described above. All U.S. patents, U.S. patent application publications, U.S. patent applications, foreign patents, foreign patent applications, and non-patent publications referred to herein and / or enumerated in the application data sheet are incorporated herein by reference in their entirety. Further embodiments can be provided by modifying aspects of the embodiments and, where necessary, by using concepts from various patents, specifications, and publications.

[0052] In consideration of the details described above, these and other modifications can be made to the embodiments. In general, the terms used in the following claims should not be interpreted as limiting the claims to the specific embodiments disclosed in the specification and claims, but rather as including all possible embodiments along the entire scope of equivalents to which such claims are granted. Accordingly, the claims are not limited by the disclosure herein.

Claims

1. It is a system, Ultrasonic sensing device, A computing device is provided, and the computing device is A communication interface configured to directly receive ultrasonic echo data detected from a person by the ultrasonic sensing device, wherein the received ultrasonic echo data includes an ultrasonic video comprising a series of frames, each frame consisting of pixels, and the communication interface It is memory, For ultrasonic video frames, one or more trained neural networks are memorized. For each of the multiple ultrasound views, the probability that the frame was captured from the view is assigned, For each of the multiple anatomical objects, The probability that the frame visualizes the object is assigned, For each of the pixels in the frame, assign a probability that the object is visualized within the pixel. For each of the aforementioned multiple objects, store a list of views that the object is permitted to view, A memory configured to store object probability thresholds and pixel probability thresholds, It is a processor, For each of the aforementioned series of frames, Applying the one or more trained neural networks to the frame, For each of the aforementioned multiple views, the probability that the frame was captured from the view is assigned to it. For each of the aforementioned multiple objects, The probability that the frame visualizes the object is assigned, For each of the pixels in the frame, assign a probability that the object is visualized within the pixel. For each of the aforementioned multiple objects, Accessing the stored list of views in which the object is permitted, The sum of the probabilities assigned to the views on the accessed list, The above sum is multiplied by the object probability to obtain the refined object probability, A processor configured to identify objects visualized in a frame, by extending the frame to indicate that the object is visualized in the frame at the pixels where the pixel probability for the object exceeds the pixel probability threshold, if the refined object probability exceeds the stored object probability threshold, A display device, For each of the aforementioned series of frames, A system including a display device configured to display the aforementioned frame as an extended version.

2. The system according to claim 1, wherein the ultrasonic sensing device and the computing device are physically connected.

3. The system according to claim 1, wherein the ultrasonic sensing device includes a transducer.

4. The system according to claim 1, wherein the processor is further configured to aggregate, for each frame, the probability of belonging to one or more frames preceding the series of frames into the probability of belonging to the frame, before the identification.

5. The system according to claim 1, wherein each of the plurality of objects is an organ.

6. The system according to claim 1, wherein each of the plurality of objects is a sub-organ structure.

7. One or more computer programs stored on a computer-readable medium, which collectively contain content configured to cause a computing system to execute a method, wherein the method is Accessing medical imaging images, With respect to an image, access a first machine learning model trained to predict a view probability vector to which the image belongs in terms of the probability that it was captured from a view, for each of multiple views. The first machine learning model is applied to the accessed image to predict the view probability vector of the accessed image, With respect to an image, access to a second machine learning model trained to predict an object probability vector to which each of several anatomical objects is assigned the probability that the object is visualized in the image, The second machine learning model is applied to the accessed image to predict the object probability vector for the accessed image, For each of the aforementioned multiple objects, The aforementioned object accesses a stored list of views that are permitted, The method allows obtaining a refined probability that the object is visualized by the accessed image by multiplying the predicted object probability vector for the accessed image being a captured view by the sum of the predicted view probability vectors, One or more computer programs stored on a computer-readable medium, comprising determining that the refined probability that the object is visualized by the accessed image exceeds a probability threshold, and that the object is visualized by the accessed image.

8. The aforementioned method, One or more computer programs stored on a computer-readable medium according to claim 7, further comprising storing instructions for an object that has been determined to be visualized by the accessed image, in relation to the accessed image.

9. The aforementioned method, With respect to the image, access to a third machine learning model trained to predict a segmentation mask that identifies the pixels of the image in which the object is visualized, The third machine learning model is applied to the accessed image to predict the segmentation mask of the accessed image for each of the multiple objects, One or more computer programs stored on a computer-readable medium according to claim 7, further comprising: using the segmentation mask with respect to the object determined to be visualized by the accessed image to obtain an extended image by extending the accessed image to indicate the location of the object determined to be visualized by the accessed image.

10. One or more computer programs stored on a computer-readable medium according to claim 9, wherein the first machine learning model, the second machine learning model, and the third machine learning model share a common set of neural network layers.

11. The aforementioned method, One or more computer programs stored on a computer-readable medium according to claim 9, further comprising storing the extended image.

12. The aforementioned method, One or more computer programs stored on a computer-readable medium according to claim 9, further comprising displaying the extended image.

13. The aforementioned method, One or more computer programs stored on a computer-readable medium according to claim 9, further comprising displaying a list of the objects which are determined to be visualized by the accessed image.

14. The aforementioned method, Training the first machine learning model described above, One or more computer programs stored on a computer-readable medium according to claim 7, further comprising training the second machine learning model.

15. One or more computer programs stored on a computer-readable medium according to claim 7, wherein the accessed image is an ultrasound image.

16. The above content is one or more computer programs stored on a computer-readable medium according to claim 7, configured to cause a computing system to perform the method for each of a series of accessed images taken from the same patient.

17. A method in a computing system for training a machine learning model, wherein the method is Accessing the main body of the annotated ultrasound image training observation, Using at least a portion of the body of an annotated ultrasound image training observation to train a first machine learning model to predict the probability that an image was captured from a view, for each of a plurality of views based on ultrasound images, Using at least a portion of the main body of the annotated ultrasound image training observation, a second machine learning model is trained to predict, based on ultrasound images, the probability that each of a plurality of anatomical objects will be visualized in the image. This includes persistently storing the trained first machine learning model and the second machine learning model, As a result, the generated ultrasound images can be subjected to the first machine learning model to predict the probability that each of the plurality of views was captured from the view, and can be subjected to the second machine learning model to predict the probability that each of the plurality of objects was visualized by the image. For each of the multiple views represented by the ultrasound image, the method further includes permanently storing the indication of one or more anatomical objects that are expected to be visualized in the ultrasound image captured from the view. A method wherein, for each of the plurality of objects, the probability that the object will be visualized by the image predicted by applying the generated ultrasound image to the second machine learning model can be refined by multiplying the sum of the probabilities predicted by applying the generated ultrasound image to the first machine learning model for the view shown in such a way that the object is expected to be visualized.

18. The method according to claim 17, wherein a portion of the second machine learning model is shared with the first machine learning model.

19. One or more computer programs stored in a computer-readable medium that collectively stores data structures, wherein the data structures are The machine learning model includes, A first part trained to predict the probability that, for each of a plurality of views, the ultrasound image was captured from the view, based on the ultrasound image, The system includes, based on ultrasound images, a second part trained to predict the probability that each of a plurality of anatomical objects will be visualized by the anatomical structures in the ultrasound image that visualizes it, The data structure further includes, for each of one or more views, a list of anatomical objects that are permitted to be identified in the ultrasound images captured from the view. This allows the model to be applied to identified ultrasound images to predict the view from which they were captured and the probability of the anatomical objects they visualize. Furthermore, one or more computer programs stored on a computer-readable medium, wherein, for each of the plurality of objects, the probability that the object is visualized by the identified ultrasound image, predicted by multiplying the identified ultrasound image by the first part, can be refined by multiplying the sum of the probabilities predicted by multiplying the identified ultrasound image by the first part for the view in which the object is in the list of anatomical objects that are permitted to be identified in the ultrasound image supplemented from the view.

20. One or more computer programs stored on a computer-readable medium according to claim 19, wherein the first and second parts of the model are separate.

21. The first and second parts of the aforementioned model intersect at a common portion, comprising one or more computer programs stored on a computer-readable medium according to claim 19.

22. The first and second parts are parallel branches that deviate from the extended path of U-net, one or more computer programs stored on a computer-readable medium according to claim 21.

23. Each of the first and second parts of the model comprises one or more computer programs stored in a computer-readable medium according to claim 19.

Images