Robust segmentation through high-level image understanding

One or more instances of computer-readable media collectively having contents configured to cause a computing system to perform a method, the method comprising:

• accessing a medical imaging image;
• accessing a first machine learning model trained to predict for an image a view probability vector that, for each of a plurality of views, assigns a probability that the image was captured from the view;
• applying the first machine learning model to the accessed image to predict a view probability vector for the accessed image;
• accessing a second machine learning model trained to predict for an image an object probability vector that, for each of a plurality of anatomical objects, assigns a probability that the anatomical object is visualized by the image;
• applying the second machine learning model to the accessed image to predict an object probability vector for the accessed image;
• for each respective anatomical object of the plurality of anatomical objects:
• accessing a stored list of views in which the respective anatomical object is permitted,
• multiplying the predicted probability that the respective anatomical object is visualized by the accessed image by the sum of the predicted probabilities that the accessed image was captured from views in which the respective anatomical object is permitted to obtain a refined probability that the respective anatomical object is visualized by the accessed image;
• and
• where the refined probability that the respective anatomical object is visualized by the accessed image exceeds a probability threshold, determining that the object is visualized by the accessed image.

A method in a computing system for training a machine learning model, the method comprising:

• accessing a body of annotated ultrasound image training observations;
• using at least a portion of the body of annotated training observations to train a first machine learning model to predict, based on an ultrasound image, for each respective view of a plurality of respective views, a probability that the image was captured from the respective view;
• using at least a portion of the body of annotated training observations to train a second machine learning model to predict, based on an ultrasound image, for each respective anatomical object of a plurality of anatomical objects, a probability that the respective anatomical object is visualized by the image;
• persistently storing the trained first and second machine learning models;
• and
• persistently storing, for each respective view of a plurality of views represented by ultrasound images, indications of one or more anatomical objects that are expected to be visualizable in ultrasound images captured from the respective view, wherein:
• a production ultrasound image can be subjected to the first machine learning model to predict, for each respective view of the plurality of views, a probability that the production ultrasound image was captured from the respective view,
• the production ultrasound image can be subjected to the second machine learning model to predict, for each respective anatomical object of the plurality of objects, a probability that the respective object is visualized by the product ultrasound image, and
• for each respective anatomical object of the plurality of anatomical objects, the probability that the anatomical object is visualized by the production ultrasound image predicted by subjecting the production ultrasound image to the second machine learning model can be refined by multiplying it by the sum of the probabilities predicted by subjecting the production ultrasound image to the first machine learning model for the views in which the anatomical object is indicated to be expected to be visualizable.