Automatically identifying anatomical structures in medical images in a manner sensitive to the particular view in which each image was captured

Abstract

A mechanism for processing medical imaging images is described. The mechanism applies a first machine learning model trained to recognize a view to which the image corresponds and a second machine learning model trained to identify any of a set of anatomical features visualized in the image. The mechanism accesses a list of allowed anatomical features of the image corresponding to the recognized view and filters the identified anatomical features to exclude those not on the accessed list. The mechanism displays the accessed image, overlaid with a visual representation of each of the filtered identified anatomical features.

Description

[Technical Field]

[0001] (CROSS-REFERENCE TO RELATED APPLICATIONS) This application claims the benefit of U.S. Provisional Patent Application No. 63 / 022,986, filed May 11, 2020, entitled "CLASSIFYING OUT-OF-DISTRIBUTION RESULTS FROM OBJECT DETECTION OR SEGMENTATION OF ULTRASOUND IMAGES," and U.S. Non-Provisional Patent Application No. 16 / 913,322, filed June 26, 2020, entitled "AUTOMATICALLY IDENTIFYING ANATOMICAL STRUCTURES IN MEDICAL IMAGES IN A MANNER THAT IS SENSITIVE TO THE PARTICULAR VIEW IN WHICH EACH IMAGE IS CAPTURED," both of which are incorporated herein by reference in their entireties.

[0002] In the event that this application conflicts with a document incorporated by reference, the present application will control. [Background technology]

[0003] Ultrasound imaging is a useful medical imaging modality. For example, internal structures of a patient's body can be imaged before, during, or after a therapeutic intervention. A medical professional typically holds a handheld ultrasound probe, called a "transducer," in close proximity to the patient and can 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 surface of the body, or in some procedures, the transducer is inserted into the patient's body. The medical professional coordinates 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] Certain views of organs or other tissues or body features (such as fluids, bones, joints, etc.) may be clinically significant. Such views may be defined by clinical criteria as views that the ultrasound operator should acquire depending on the target organ, diagnostic objectives, etc.

[0005] In some ultrasound images, it is useful to identify anatomical structures visualized within the image. For example, in ultrasound image views showing a particular organ, it may be useful to identify constituent structures within the organ. As an example, in some views of the heart, constituent structures such as the left and right atria, left and right ventricles, and the aorta, mitral valve, pulmonary, and tricuspid valves are visible.

[0006] Existing software solutions have attempted to automatically identify such structures, and these existing solutions seek to either "detect" the structures by specifying a bounding box within which each is visible, or to "segment" the structures by tracing the boundaries of each structure within the image. [Brief explanation of the drawings]

[0007] [Figure 1] 1 is a schematic diagram of a physiological sensing device 10 in accordance with one or more embodiments of the present disclosure. [Figure 2] FIG. 1 is a block diagram illustrating some of the components typically incorporated into at least some of the computer systems and other devices in which the mechanism operates. [Figure 3] FIG. 1 is a flow diagram illustrating a process performed by the facility in some embodiments to train one or more machine learning models to predict views and detect structures in medical images. [Figure 4] FIG. 1 is a model diagram illustrating the model architecture used by the facility in some embodiments. [Figure 5] FIG. 1 is a flow diagram illustrating a process performed by the mechanism in some embodiments to process patient images during production. [Figure 6]FIG. 1 is a medical image diagram showing sample patient images accessed by the facility. [Figure 7] FIG. 1 is a medical image diagram showing a sample patient image annotated to show structures identified by the mechanism. [Figure 8] FIG. 10 is a table diagram showing sample contents of a table used by the facility in some embodiments to store a list of allowed structures for various views. [Figure 9] FIG. 10 is a medical image diagram showing a sample patient image annotated to show structures identified by the mechanism and filtered to structures allowed for predictive views. [Figure 10] FIG. 2 is a data flow diagram illustrating the data flow that occurs within the mechanism in some embodiments. DETAILED DESCRIPTION OF THE INVENTION

[0008] The present inventors have recognized that conventional approaches for automatically identifying constituent structures of organs shown in ultrasound images have significant drawbacks. In particular, these conventional approaches often generate false positives, i.e., they identify structures that are not actually present in the image. This is particularly common when an organ contains multiple structures of similar size and shape, such as the four valves of the heart. In some cases, conventional approaches attempt to reduce or eliminate the false positive problem by significantly increasing a confidence threshold operating parameter that limits identified structures to those with a confidence value higher than the threshold. To eliminate most or all false positives, this confidence threshold must be set so high that it creates the opposite problem, i.e., false negatives in which actually visible structures are excluded from the identification results.

[0009] In response to recognizing these shortcomings, the inventors have devised and implemented a software and / or hardware mechanism that automatically identifies anatomical structures in medical images, such as ultrasound and MRI, in a manner (the "mechanism") that is sensitive to the particular view in which each image was captured. By doing so, the mechanism can set a confidence threshold operating parameter relatively low to minimize the occurrence of false negatives. At the same time, the mechanism minimizes the occurrence of false positives by limiting the identified structures to those that typically appear in the view in which the image was captured. In some embodiments, the mechanism automatically determines the view in which each image was captured using image recognition machine learning techniques.

[0010] In one example of the mechanism's operation, described further below, the mechanism uses image recognition machine learning techniques to identify five candidate structures in an image of a human heart. The mechanism uses image recognition machine learning techniques to classify the image as being from an "apical four-chamber" view of the heart. Based on this classification, the mechanism accesses a list of structures that can be identified in an image of the heart captured from this view. By comparing this list of allowed structures to the five identified candidate structures, the mechanism determines that only four of the identified candidate structures are present in the list. In response, the mechanism identifies only four of the five identified candidate structures that are present in the list.

[0011] By operating in some or all of the above-described ways, the system automatically identifies anatomical structures in medical images with greater accuracy than previously possible.

[0012] Additionally, the mechanism improves the functionality of computers or other hardware, such as by reducing the dynamic display area, processing, storage, and / or data transmission resources required to perform a particular task, thereby allowing the task to be performed with less powerful, capacitive, and / or expensive hardware devices and / or with lower latency, and / or more conserving resources for use in performing other tasks. For example, by maximizing the usefulness of initial medical images by more frequently identifying all visualized structures with significantly fewer false positives, the mechanism avoids many cases in which reimaging is necessary. Reducing the need for reimaging generally results in the mechanism consuming fewer memory and processing resources to capture additional images and perform additional rounds of automatic structure identification.

[0013] 1 is a schematic diagram of a physiological sensing device 10 in accordance with one or more embodiments of the present disclosure. Device 10, in the illustrated embodiment, includes a probe 12 electrically coupled to a handheld computing device 14 by a cable 17. Cable 17 includes a connector 18 that removably connects probe 12 to computing device 14. Handheld computing device 14 may be any portable computing device with a display, such as a tablet computer, a smartphone, or the like. In some embodiments, probe 12 need not be electrically coupled to handheld computing device 14, but may operate independently of handheld computing device 14, and probe 12 may communicate with handheld computing device 14 via a wireless communication channel.

[0014] The probe 12 is configured to transmit ultrasonic signals toward a target structure and to receive echo signals returned from the target structure in response to the transmission of the ultrasonic signals. 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 ultrasonic signals and receiving subsequent echo signals.

[0015] The device 10 further includes a processing circuit and a driver circuit. In part, the processing circuit controls the transmission of ultrasonic signals from the ultrasonic sensor 20. The driver circuit is operatively coupled to the ultrasonic sensor 20 for driving the transmission of the ultrasonic signals, for example, in response to control signals received from the processing circuit. The driver 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 driver circuit for transmitting ultrasonic signals, for example, in a pulsed wave or continuous wave mode of operation.

[0016] The ultrasonic sensor 20 of the probe 12 may include one or more transmit transducer elements that transmit ultrasonic signals and one or more receive transducer elements that receive echo signals returning from the target structure in response to the transmission of the ultrasonic signals. In some embodiments, some or all of the transducer elements of the ultrasonic sensor 20 may function as transmit transducer elements during a first period of time and as receive transducer elements during a second period of time that is different from the first period of time (i.e., the same transducer element may be usable to transmit ultrasonic signals and receive echo signals at different times).

[0017] 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 ultrasound signals, and 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. Additionally, in some embodiments, the display screen 22 may be used to display audio waveforms, such as waveforms representing acquired or conditioned auscultation signals.

[0018] In some embodiments, the display screen 22 may be a touchscreen capable of receiving input from a user touching the screen. In such embodiments, the user interface 24 may include a portion or all of the 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 ultrasound device 10. In some embodiments, the user interface 24 may include a microphone 30 capable of receiving audible input, such as voice commands.

[0019] The computing device 14 may further include one or more audio speakers 28 that may be used to output an audible representation of acquired or conditioned auscultation or echo signals, blood flow during Doppler ultrasound imaging, or other features derived from the operation of the device 10.

[0020] Probe 12 includes a housing that forms the exterior portion of probe 12. The housing includes a sensor portion located near the distal end of the housing and a handle portion located between the proximal and distal ends of the housing. The handle portion is located proximal to the sensor portion.

[0021] The handle portion is the portion of the housing that is grasped by a user during use to hold, control, and manipulate the probe 12. The handle portion may include one or more gripping features, such as cleats, and in some embodiments, the handle portion may have the same general shape as the portion of the housing distal or proximal to the handle portion.

[0022] The housing encloses the internal electronic components and / or circuitry of the probe 12, including, for example, electronics such as drive circuitry, processing circuitry, oscillators, beamforming circuitry, filtering circuitry, etc. The housing may be formed to enclose or at least partially enclose an externally disposed portion of the probe 12, such as the sensing surface. The housing may be a sealed housing to prevent moisture, liquids, or other fluids 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 of a single piece (e.g., a single piece of material molded to enclose the internal components) or may be formed of two or more pieces (e.g., an upper half and a lower half) that are joined or otherwise attached to each other.

[0023] In some embodiments, the probe 12 includes a motion sensor. The motion sensor is operable to sense movement of the probe 12. The motion sensor may be included in or on the probe 12 and may include, for example, one or more accelerometers, magnetometers, or gyroscopes for sensing movement of the probe 12. For example, the motion sensor may include any of a piezoelectric, piezoresistive, or capacitive accelerometer capable of sensing movement of the probe 12. In some embodiments, the motion sensor is a three-axis motion sensor capable of sensing movement about any of three axes. In some embodiments, two or more motion sensors 16 are included in or on the probe 12. In some embodiments, the motion sensor includes at least one accelerometer and at least one gyroscope.

[0024] The motion sensor may be at least partially contained 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 the surface that is in operative contact with the patient during an examination, such as for ultrasound imaging or auscultation sensing. The ultrasound sensor 20 and one or more auscultation sensors are located on, in, or near the sensing surface.

[0025] In some embodiments, the transducer array of ultrasonic sensor 20 is a one-dimensional (1D) array or a two-dimensional (2D) array of transducer elements. The transducer array may include a piezoelectric ceramic, such as lead zirconate titanate (PZT), or may be based on a microelectromechanical system (MEMS). For example, in various embodiments, ultrasonic sensor 20 may include a piezoelectric micromachined ultrasonic transducer (PMUT), which is a microelectromechanical system (MEMS)-based piezoelectric ultrasonic transducer, or ultrasonic sensor 20 may include a capacitive micromachined ultrasonic transducer (CMUT), in which energy transduction is provided due to changes in capacitance.

[0026] The ultrasonic sensor 20 may further include an ultrasound focusing lens that may be disposed on the transducer array and may form part of the sensing surface. The focusing lens may be any lens operable to focus an ultrasound beam transmitted from the transducer array toward a patient and / or to focus an ultrasound beam reflected from the patient back to the transducer array. In some embodiments, the ultrasound focusing lens may have a curved surface shape. The ultrasound focusing lens may have different shapes depending on the desired application, such as the desired operating frequency. The ultrasound focusing lens may be formed of any suitable material, and in some embodiments, the ultrasound focusing lens is formed of a room-temperature vulcanizing (RTV) rubber material.

[0027] In some embodiments, the first and second membranes are positioned adjacent to each other on opposite sides of the ultrasonic sensor 20 and form part of the sensing surface. The membranes may be formed of any suitable material, and in some embodiments, the membranes are formed of a room temperature vulcanizing (RTV) rubber material. In some embodiments, the membranes are formed of the same material as the ultrasound focusing lens.

[0028] 2 is a block diagram illustrating some of the components typically incorporated in at least some of the computer systems and other devices on which the mechanisms operate. In various embodiments, these computer systems and other devices 200 may include server computer systems, other configurations of cloud computing platforms or virtual machines, desktop computer systems, laptop computer systems, netbooks, mobile phones, personal digital assistants, televisions, cameras, automobile computers, electronic media players, physiological sensing devices, and / or their associated display devices, etc. In various embodiments, the computer systems and devices include one or more of the following: a processor 201, such as a CPU, GPU, TPU, NNP, FPGA, or ASIC, for executing computer programs and / or training or applying machine learning models; computer memory 202, for storing programs and data while they are in use, including the mechanism and associated data; persistent storage 203, such as a hard drive or flash drive, for persistently storing an operating system, including a kernel, device drivers, programs, and data; a computer-readable media drive 204, such as a floppy, CD-ROM, or DVD drive, for reading programs and data stored on a computer-readable medium; and a network connection 205, for connecting the computer system to other computer systems to transmit and / or receive data, such as via the Internet or another network and its network hardware, such as switches, routers, repeaters, electrical cables and optical fibers, light emitters and receivers, wireless transmitters and receivers, etc. While a computer system configured as described above is typically used to support the operation of the mechanism, those skilled in the art will understand that the mechanism may be implemented using various types and configurations of devices and have various components.

[0029] 3 is a flow diagram illustrating a process performed by the facility in some embodiments to train one or more machine learning models to predict views and detect structures in medical images. In some embodiments, the process is performed on a server. In operation 301, the facility initializes the models.

[0030] 4 is a model diagram illustrating the model architecture used by the mechanism in some embodiments. Model architecture 400 includes a shared layer 410 used by the mechanism for both structure detection and view classification, an object detection layer 430 used by the mechanism only to perform structure detection, and a classifier layer 450 used by the mechanism only to perform view classification. For each layer, FIG. 4 indicates the layer type and layer size. The layer types are represented by the following abbreviations:

[0031] [Table 1]

[0032] For example, layer 411, designated by the type abbreviation "Conv," is a convolutional layer. Layer size is expressed in dimensions: number of output values ​​per pixel times number of horizontal pixels times number of vertical pixels. For example, layer 411 designated by dimensions 16x320x320 outputs 16 values ​​per pixel in a rectangular array of 320 pixels by 320 pixels.

[0033] Returning to FIG. 3 , in operation 302, the mechanism accesses training data. In some embodiments, the training data is obtained as follows: 3-5 second long ultrasound video clips are captured and annotated by an expert sonographer or cardiologist. A manually curated test set of 100 or more video clips, evenly distributed across all possible views, is reserved for final evaluation. Of the remaining data, 10% is retained as a validation set during training. The validation set is equally stratified by device type and view. The remaining data is used for training. For model preprocessing, input images are resized to 320x320 pixels and pixel values ​​are scaled between 0.0 and 1.0. For training data only, the input images are also randomly augmented to increase the diversity of the training input. In various embodiments, augmentation transformations include, among others, flipping horizontally or vertically, rotation, scaling (zoom in / out), translation, blurring, contrast / brightness scaling, and random pixel dropout.

[0034] In operation 303, the facility trains the model using the accessed training data. In some embodiments, model training is a two-step process. First, the model is trained only for the object detection task, and the classifier layer is not used. Then, the model is trained for the classification task, and the network weight parameters of the shared layer are frozen, and the object detection layer is not used. In some embodiments, for each step, the facility trains the network using the ADAM optimizer for 1000 epochs with a batch size of 512, with an exponentially decaying learning rate starting at 0.001 and ending at 0.0001.

[0035] In operation 304, the mechanism stores the trained model. In some embodiments, operation 304 includes saving neural network connection weights determined during training. In some embodiments, operation 304 includes distributing the trained model to devices used to evaluate generated images, such as each of several handheld imaging devices. After operation 304, the process ends. In some embodiments, the process can be repeated for various purposes, including improving model accuracy, adding new views, structures, imaging modalities, or device designs, etc.

[0036] Those skilled in the art will understand that the operations shown in Figure 3 and any flow diagrams discussed below may be modified in various ways, for example, the order of operations may be rearranged, some operations may be performed in parallel, illustrated operations may be omitted or other operations may be included, illustrated operations may be divided into sub-operations, or multiple illustrated operations may be combined into a single operation, etc.

[0037] 5 is a flow diagram illustrating a process performed by the facility in some embodiments to process patient images being created. In some embodiments, the facility performs this process for each of one or more portable imaging devices. In some embodiments, the facility performs this process for each patient image captured by one of the portable imaging devices. In some embodiments, the facility performs this process on one or more physical servers and / or one or more virtual servers, such as virtual cloud servers. In operation 501, the facility accesses patient images. In some embodiments, the patient images are ultrasound images simultaneously captured by ultrasound sensors.

[0038] Figure 6 is a medical imaging diagram showing a sample patient image accessed by the facility. Patient image 600 is an ultrasound image. This ultrasound image, also shown in Figures 7 and 9 discussed below, has been grayscale inverted to make it easier and more faithfully produced in patent drawings.

[0039] Returning to FIG. 5, in operation 502, the mechanism subjects the patient image accessed in operation 501 to one or more trained machine learning models to predict both the view represented by the patient image and which structures are visualized (i.e., visible) and their locations.

[0040] 7 is a medical imaging diagram showing a sample patient image that has been annotated to show structures identified by the mechanism. In particular, patient image 700 has been annotated to show the following structures, each shown with a central dot and a bounding rectangle: right ventricular outflow tract 701, right ventricle 702, left ventricle 703, tricuspid valve 704, and mitral valve 705. In processing the sample patient image, the mechanism also determines that it represents an apical four-chamber view (not shown).

[0041] Returning to FIG. 5, in operation 503 , the mechanism accesses a list of allowable structures for the view predicted in operation 502 .

[0042] FIG. 8 is a table diagram illustrating sample contents of an allowed structures table used by the facility in some embodiments to store a list of allowed structures for various views. Table 800 of allowed structures is composed of rows, such as rows 801-828, each corresponding to a different combination of a view and a structure allowed in that view. Each row is divided into a view column 851, which identifies the view, and an allowed structures column 852, which identifies and labels the structure in the view. For example, row 808 indicates that in the apical four-chamber view, the mitral valve is an allowed structure. In various embodiments, the list shown in the table of allowed structures is generated manually as a matter of editorial discretion and automatically compiled from labels on training images.

[0043] To access the list of allowed structures for a particular view, the mechanism selects the row of the table of allowed structures where that view occurs in the view column and extracts from the selected row the structures that occur in the allowed structures column. In the context of an example where the predicted view is an apical four-chamber view, the mechanism selects rows 809-818 to obtain the following ten allowed structures: aorta, atrial septum, interventricular septum, left atrium, left ventricle, left ventricular outflow tract, mitral valve, right atrium, right ventricle, and tricuspid valve.

[0044] While FIG. 8 shows a table whose contents and organization are designed to be more easily understood by a human reader, those skilled in the art will understand that the actual data structures used by the mechanism to store this information may differ from the table shown in that, for example, they may be organized differently, may contain more or less information than shown, may be encoded, compressed, encrypted, and / or indexed, and may contain many more rows than shown.

[0045] 5, in operation 504, the facility filters the visible structures predicted in operation 502 to exclude those not on the list of allowed structures accessed in operation 503. In the context of the example, the facility matches the following predicted structures with the list accessed in operation 503 for the apical four-chamber view: the right ventricle, the left ventricle, the tricuspid valve, and the mitral valve. If the right ventricular outflow tract does not match, the facility excludes this structure, leaving only the four structures listed above. In operation 505, the facility augments the patient image accessed in operation 501 to identify all of the visible structures predicted in operation 502 that were not excluded by the filtering of operation 504.

[0046] FIG. 9 is a medical image diagram showing a sample patient image annotated to show structures identified by the mechanism and filtered to structures allowed for the predicted view. By comparing FIG. 9 with FIG. 6, it can be seen that the mechanism has added annotations for the right ventricle 902, left ventricle 903, tricuspid valve 904, and mitral valve 905. In each case, the annotations shown include the full name or abbreviation of the structure, a dot in the center of the structure, and a bounding rectangle. These annotations omit the left ventricular outflow tract 701 shown in FIG. 7, which is not allowed for the apical four-chamber view. In some embodiments (not shown), the mechanism traces the boundary of each identified structure as part of annotating the patient image.

[0047] 5, in operation 506, the mechanism causes the augmented image created in operation 505 to be displayed, such as on a display device integrated with or connected to the medical imaging device that produced the patient image. In some embodiments (not shown), the mechanism also persistently stores the augmented image and / or transmits it to another location for storage, review, and / or analysis. After operation 506, the process ends.

[0048] FIG. 10 is a data flow diagram illustrating the data flow that occurs within the mechanism in some embodiments. In data flow 1000, a patient image 1001 is subjected to both an object detection network 1010 and a view classifier network 1030. As described above, in various embodiments, these networks are independent, intersecting, or identifiably merged. The object detection network generates a list 1020 of structures detected in the patient image. The view classifier network generates a view classification 1040 of the patient image. From this view classification, the mechanism generates a list 1050 of allowed structures. The mechanism uses the list of allowed structures to filter the list of detected structures to obtain a filtered list 1060 of detected structures. The mechanism uses the filtered list of detected structures to generate a copy 1070 of the patient image that has been augmented to identify the list of filtered structures.

[0049] The various embodiments described above can be combined to provide further embodiments. All U.S. patents, U.S. patent application publications, U.S. patent specifications, foreign patents, foreign patent specifications, and non-patent publications mentioned herein and / or listed in the Application Data Sheet are incorporated herein by reference in their entirety. Aspects of the embodiments can be modified, and, if necessary, concepts from the various patents, specifications, and publications can be used to provide further embodiments.

[0050] These and other changes can be made to the embodiments in light of the above detailed description. Generally, in the following claims, the terms used should not be construed to limit the claims to the specific embodiments disclosed in the specification and claims, but should be construed to include all possible embodiments along with the full range of equivalents to which such claims are entitled. Accordingly, the claims are not limited by this disclosure.

Claims

1. 1. A system comprising: an ultrasonic sensing device; 1. A computing device comprising: a communications interface configured to receive ultrasonic echo data sensed by the ultrasonic sensing device directly from a person, the received ultrasonic echo data including an ultrasound image; and A memory, storing one or more first neural networks trained to identify physiological structures in ultrasound images; storing one or more second neural networks trained to classify ultrasound images as being captured in particular views; a memory configured to store, for each of a set of ultrasound views, a list of anatomical structures allowed for identification within the ultrasound view; 1. A processor, comprising: applying the one or more trained first neural networks to the received ultrasound images to identify a set of physiological structures within the received ultrasound images; applying the one or more trained second neural networks to the received ultrasound images to classify the received ultrasound images as having been captured in a particular view; accessing a list of allowed anatomical structures for the classified view; a processor configured to select, from the set of physiological structures identified in the received ultrasound image, only those present in the accessed list; A display device, Displaying the received ultrasound image; and a display device configured to display the ultrasound images together with annotations visually indicating the selected physiological structures of the set.

2. The system of claim 1 , wherein the ultrasonic sensing device comprises a transducer.

3. 2. The system of claim 1, wherein each of the set of ultrasound views is a view of a particular organ, and the allowed anatomical structures listed for each of the set of ultrasound views are constituent structures of that organ.

4. A computer-readable medium having collectively thereon a program configured to cause a computing system to perform a method, the method comprising: accessing medical imaging images; accessing a first machine learning model trained to recognize a view to which an image corresponds; applying the first machine learning model to the accessed image to recognize a view to which the accessed image corresponds; accessing a list of allowed anatomical features of an image corresponding to the recognized view; accessing a second machine learning model trained to identify any of a set of anatomical features visualized in the image; and applying the second machine learning model to the accessed images to identify any of the set of anatomical features visualized in the accessed images; and filtering the identified anatomical features to exclude those not on the accessed list; and displaying the accessed image overlaid with a visual representation of each of the filtered identified anatomical features.

5. 5. The computer-readable medium of claim 4, wherein each superimposed visual representation of a filtered identified anatomical feature identifies at least one point in a visualization in the accessed image of the filtered identified anatomical feature.

6. 5. The computer-readable medium of claim 4, wherein each superimposed visual representation of a filtered identified anatomical feature is a bounding box that includes a visualization of the filtered identified anatomical feature within the accessed image.

7. 5. The computer-readable medium of claim 4, wherein each superimposed visual representation of a filtered identified anatomical feature is a tracing of a boundary of a visualization in the accessed image of the filtered identified anatomical feature.

8. 5. The computer-readable medium of claim 4, wherein each superimposed visual representation of a filtered identified anatomical feature is a covermap overlay of a visualization in the accessed image of the filtered identified anatomical feature.

9. 5. The computer-readable medium of claim 4, wherein each superimposed visual representation of a filtered identified anatomical feature is a name by which the filtered identified anatomical feature is known.

10. 5. The computer-readable medium of claim 4, wherein the recognized view is directed to a particular organ, and the allowed anatomical features enumerated for the recognized view comprise a constituent structure of that organ.

11. 5. The computer-readable medium of claim 4, wherein the recognized view is directed to a particular organ, and the allowed anatomical features enumerated for the recognized view comprise landmarks of that organ.

12. The method comprises: training the first machine learning model; and training the second machine learning model.

13. The computer-readable medium of claim 4 , wherein the accessed medical imaging images are ultrasound images.

14. 1. A method in a computing system for training a machine learning model, comprising: accessing a body of annotated ultrasound image training observations; training a first machine learning model using at least a portion of the body of annotated training observations to predict, based on an ultrasound image, a view that it represents; training a second machine learning model using at least a portion of the body of annotated training observations to predict the anatomical structures it visualizes based on ultrasound images; persistently storing the trained first and second machine learning models, whereby a generated ultrasound image can be subjected to the first machine learning model to predict the view it represents, and to the second machine learning model to predict the anatomical structure it visualizes; and persistently storing, for each of a plurality of views represented by ultrasound images, a representation of one or more anatomical structures predicted to be visualized in the ultrasound image representing the view, such that the anatomical structures predicted by subjecting the generated ultrasound image to the second machine learning model can be filtered to exclude anatomical structures not shown by the stored representation for the view predicted by subjecting the generated ultrasound image to the first machine learning model; A method for providing

Images