Machine-learning-based object detection system
In at least one embodiment, an object detection system uses a neural network to identify and/or locate a set of organs in a medical image. In at least one embodiment, when training to identify and/or locate a particular organ, a subset of incompletely-labeled training images is used that excludes training images for which labels associated with particular organ are unavailable.
1 . One or more processors, comprising:
circuitry to cause a neural network to be trained by at least:
obtaining a set of incompletely labeled images, at least one image of the set of incompletely labeled images including at least one labeled object and at least one unlabeled object; and
training the neural network to identify a particular object using the set of incompletely labeled images by preventing loss associated with the at least one image from being used to update weights of the neural network as a result of identification, by the neural network, of the at least one unlabeled object in the at least one image.
2 . The one or more processors of claim 1 , wherein the at least one unlabeled object includes an organ within one or more three-dimensional medical images.
3 . The one or more processors of claim 1 , wherein:
the set of incompletely labeled images is a set of incompletely labeled three-dimensional images, where each image in the set of incompletely labeled three-dimensional images includes one or more labels identifying one or more objects and a bounding region for each of one or more objects, the set of incompletely labeled three-dimensional images including the at least one image that lacks labeling for the at least one unlabeled object; and
training the neural network to identify the particular object includes using the set of incompletely labeled three-dimensional images, excluding, for a purpose of training to identify the particular object, an image that lacks a label for the particular object.
4 . The one or more processors of claim 3 , wherein locating an organ of the one or more objects is accomplished by at least:
determining a bounding box around the one or more objects; and
indicating, to a user, the location of the bounding box.
5 . The one or more processors of claim 3 , wherein:
a location of an object is determined using an anchor that indicates where detection begins; and
that anchor describes a three-dimensional region that has a shape based at least in part on a type of the object.
6 . The one or more processors of claim 5 , wherein the three-dimensional region is a rectangular box.
7 . The one or more processors of claim 5 , wherein a size of the anchor is varied over a range of sizes that correspond to size variations of human anatomy.
8 . The one or more processors of claim 1 , wherein the neural network is trained using a loss based at least in part on a prediction of an Intersection over Union.
9 . The one or more processors of claim 8 , wherein the prediction of the Intersection over Union provides a quality assessment of a location of an object.
10 . The one or more processors of claim 1 , wherein the neural network is trained using a loss calculated based, at least in part, on skipping images with unlabeled objects.
11 . The one or more processors of claim 1 , wherein the neural network includes a plurality of region proposal networks associated with a plurality of objects, wherein at least one region proposal network associated with an unlabeled object is skipped in training.
12 . A system, comprising one or more processors to cause a neural network to be trained by at least:
obtaining a set of images, at least one image of the set of images including:
one or more labels that identify one or more objects in the at least one image; and
one or more unlabeled objects; and
training the neural network to identify a particular object using the set of images by preventing loss associated with the at least one image from being used to update weights of the neural network as a result of identification, by the neural network, of the one or more unlabeled objects in the at least one image.
13 . The system of claim 12 , wherein the one or more unlabeled objects comprise an organ within a three-dimensional medical image.
14 . The system of claim 12 , wherein:
the set of images includes one or more bounding regions for the one or more objects and lacks at least one bounding region for the one or more unlabeled objects; and
training the neural network to identify the particular object includes excluding images in the set of images that lack a label for the particular object.
15 . The system of claim 14 , wherein locating an object of the one or more objects is accomplished by at least:
determining a region around the object; and
presenting an indication of the region on an electronic display.
16 . The system of claim 15 , wherein:
the region is defined by a box-shaped region; and
the region is presented on the electronic display as an isometric view of the box-shaped region.
17 . The system of claim 14 , wherein:
the system uses an anchor that indicates, to the system, an estimate of a size and shape of the particular object; and
the estimate is specific to a type of the particular object.
18 . The system of claim 15 , wherein the anchor defines a three-dimensional shape, and a range of sizes for the three-dimensional shape.
19 . The system of claim 18 , wherein the range is determined using a scaling factor associated with a particular image.
20 . The system of claim 13 , wherein the three-dimensional medical image is:
a magnetic resonance imaging (MRI) image,
a computed tomography (CT) scan image, or
an X-ray image.
21 . One or more processors comprising processing circuitry to use a neural network, wherein the neural network is trained by:
obtaining a set of images, at least one image of the set of images including one or more labeled objects and one or more unlabeled objects; and
training the neural network to identify a particular object using the set of images by excluding the at least one image from the training.
22 . The one or more processors of claim 21 , wherein the neural network is trained to identify a plurality of organs within a three-dimensional medical image.
23 . The one or more processors of claim 21 , wherein:
the set of images includes a bounding region for each of the one or more labeled objects and lacks at least one bounding region for the one or more unlabeled objects; and
the neural network is trained to identify the particular object by excluding, for a purpose of training, images of the set of images from the training that lack a label for the particular object.
24 . The one or more processors of claim 21 , wherein the neural network is trained using a loss based at least in part on a prediction of an Intersection over Union.
25 . The one or more processors of claim 21 , wherein the neural network is further trained by at least removing, from a set of three-dimensional images, one or more images that lack labeling for a type of a particular object.
26 . The one or more processors of claim 25 , wherein each image in the set of three-dimensional images represents a portion of a human body with a known composition of organs.
27 . The one or more processors of claim 26 , wherein the set of images represent head, chest, or abdominal scan images.
28 . The one or more processors of claim 27 , wherein a location of individual organs is determined using estimates of a size and shape of each of the individual organs based at least in part on the type of each organ of the individual organs.
29 . The one or more processors of claim 28 , wherein the estimates include a size range around an average object size.
30 . A method comprising, training a neural network, at least in part, by at least:
obtaining a set of incompletely labeled images, at least one image in the set of incompletely labeled images including one or more labeled objects and at least one unlabeled object; and
training the neural network to identify a particular object using the set of incompletely labeled images by excluding the at least one image from the training.
31 . The method of claim 30 , wherein;
the at least one unlabeled object is an organ; and
the set of incompletely labeled images comprise one or more three-dimensional medical images.
32 . The method of claim 30 , wherein:
the set of incompletely labeled images are a set of incompletely labeled three-dimensional images that includes a bounding region for each of the one or more labeled objects and lacks at least one bounding region for the at least one unlabeled object; and
the neural network is trained to identify the particular object by excluding, for a purpose of training to identify the particular object, images of the set of images that lack a label for the particular object.
33 . The method of claim 32 , wherein locating an organ of the one or more labeled objects is accomplished by at least:
determining a bounding box around an object; and
indicating, to a user, the location of the bounding box.
34 . The method of claim 32 , wherein:
a location of an object is determined using an anchor that indicates where detection begins; and
that anchor describes a three-dimensional region that has a shape based at least in part on a type of the object.
35 . The method of claim 34 , wherein the three-dimensional region is a rectangular box.
36 . The method of claim 34 , wherein a size of the anchor is varied over a range of sizes that correspond to size variations of human anatomy.
37 . The method of claim 30 , wherein the neural network is trained using a loss based at least in part on a prediction of an Intersection over Union.